text
stringlengths 4
2.78M
| meta
dict |
|---|---|
harvmac amssym epsf
Shigenori Seki[^1][[email protected]]{}
*Center for Quantum Spacetime (CQUeST)*
*Sogang University, Sinsu-dong, Mapo-gu, Seoul 121-741, Republic of Korea*
Sang-Jin Sin[^2] [[email protected]]{}
*Department of Physics, Hanyang University, Seoul 133-791, Republic of Korea*
.3in
**Abstract**
We consider the model of holographic QCD with asymptotic freedom and gluon condensation in its vacuum. It consists of the color D4-branes and D0-branes as a background and the flavor D8-branes as a probe. By taking a specific field theory limit, the effective coupling decreases to vanish in UV region. We then introduce the uniformly distributed baryons in terms of the baryon vertices and study the density dependence of chiral condensate, which is evaluated using the worldsheet instanton method. In the confined phase, the chiral condensate as a function of density monotonically decreases in high baryon density. Such behavior is in agreement with the expectation, while in extremely low density it increases. We attribute this anomaly to the incorrect approximation of uniformity in very low density. In the deconfined phase the chiral condensate monotonically decreases in the whole region of density.
Chiral symmetry is one of the most important properties which control the behavior of the strongly interacting nucleon system and therefore studying its order parameter, the chiral condensate, especially in the dense matter is the most urgent and important to understand the properties of hadrons in nuclei, neutron stars and so on. So far many speculations have been made about the behavior of the chiral condensate in dense medium without definite proof. One of the intuitive and acceptable ones is that the chiral condensate has a finite value at zero density and vanishes at a certain high density where the chiral symmetry is restored, and that it interpolates between the two by monotonically decreasing behavior . However, no one succeeded in proving this from a first principle in a theory with strong coupling since there has not been any reliable calculational tools for the strongly interacting system especially in the presence of the dense fermions.
Since the gauge/gravity correspondence is good for strongly interacting gauge theory, it is tantalizing to see what happen if we apply this method to QCD and such application is called holographic QCD. The flavor physics of QCD has been studied in many models of holographic QCD, for example, the D3/D7 model , the D4/D6 model and the D4/D8 model . One can incorporate baryons into those models by the use of the baryon vertices , the D5-branes wrapping $S^5$ in the D3-branes background or the D4-branes wrapping $S^4$ in the D4-branes background. Such models with baryonic medium have been attracting many interests . However we have not found a model which has the behavior of chiral condensate in complete agreement with the speculation mentioned above.
Although the D4/D8 model is one of the most successful models of holographic QCD, it has several shortcomings: current quark mass can not be incorporated, and as a consequence the chiral condensate can not be naturally introduced. Various prescriptions for this problem have been suggested by . The method by allows us to calculate the quark mass and chiral condensate $\langle {\bar \psi}\psi \rangle$ from the open Wilson line, whose expectation value is given by the minimal surface of the open string worldsheet surrounded by the D8-branes, the worldsheet instanton. Therefore the chiral condensate can be evaluated by the area of that minimal surface in the analogy of the computation of Wilson loop by . Following this method, we have studied the chiral condensate in both confined and deconfined phases . There, we found that the chiral condensate decreases in low density but increases in high density so that the low density behavior in this model agrees with the expectation, but the high density one does not, as in all other known models.
The D4/D6 and D4/D8 models use the common background yielded by the D4-branes, in which the gluon sector of four-dimensional dual gauge theory becomes non-supersymmetric due to an anti-periodic $S^1$ compactification of fermions . This is an advantage of this background, because the ordinary QCD does not have supersymmetry. However the effective coupling $e^\phi$ diverges as one goes to UV, namely, the theory has opposite behavior to asymptotic freedom. In this paper we improve this problem by introducing D0-branes smeared on the D4-branes. We take a specific field theory limit to achieve the asymptotic freedom. In general, the D$p$-branes smeared on the D$(p+4)$-branes play the role of instantons, so that the D$p$-branes provide a gluon condensate to the field theory on D$(p+4)$-branes .
The rest of the paper goes as follows. In Section 2 we suggest the new model whose background is given by taking the specific field theory limit of the supergravity solution of the D4-branes on which the D0-branes are smeared. In Section 3 we consider the D8-branes in the background of confined phase and then numerically calculate the chiral condensate depending on the density of baryons introduced as baryonic D4-branes (baryon vertices). In Section 4 we comment on the chiral condensate in deconfined phase which is analyzed in the similar way to the confined phase. Finally Section 5 is devoted to the conclusions.
In this section, we propose a new model of holographic QCD which consists of $N_4$ color D4-branes, $N_0$ D0-branes and $N_f$ flavor D8 and anti-D8-branes. The D0-branes are smeared on the color D4-branes and induce a gluon condensate. The theory has $U(N_4)$ color gauge symmetry and $U(N_f) \times U(N_f)$ chiral symmetry. The flavor D-branes are treated as a probe embedded into the background produced by the color D-branes.
We consider $N_4$ D4-branes on which $N_0$ (Euclidean) D0-branes are homogeneously smeared. The configuration of these D-branes are shown in Table 1.
In IIA supergravity the non-extremal black brane solution describing such D-branes has already been known . The solution has the following metric and dilaton: $$\eqalignno{
&ds^2 = \sqrt{H_0 \over H_4} dX_\mu dX^\mu
+{F \over \sqrt{H_0 H_4}} \bigl(dX^4\bigr)^2
+\sqrt{H_0 H_4}\biggl({d\rho^2 \over F} +\rho^2 d\Omega_4^2 \biggr) \,, &\bbsolution{a} \cr
&e^{\phi} = g_s \biggl({H_0^3 \over H_4}\biggr)^{1 \over 4} \,, \quad
g_s := e^{-\phi_0} \,, &\bbsolution{b}
}$$ where $\mu = 0,1,2,3$ which are contracted by $\eta_{\mu\nu} = {\rm diag}(-1,1,1,1)$. $H_0$, $H_4$ and $F$ are the harmonic functions: $\rho_0$ and $\rho_4$ are related with the horizon, $\rho_h$, and the D-brane charges, $N_0$ and $N_4$, by where $V_4$ denotes the volume of the space on which the D0-branes are smeared, [*i.e.*]{}, $V_4 := \int dX^0 dX^1 dX^2 dX^3$.
There are two Ramond-Ramond (RR) fluxes, $G_{(4)}$ and $G_{(8)}$, which are associated with the D4-branes and D0-branes respectively. Since the fluxes obey $(2\pi\ell_s)^{-3}\int G_{(4)} = N_4$ and $(2\pi\ell_s)^{-7}\int G_{(8)} = N_0$ by the definition of the D-brane charges, one can write down the fluxes as where $\Omega_4$ is the volume of unit four-sphere, [*i.e.*]{}, $\Omega_4 = 8\pi^2/3$.
Let us consider the Maldacena limit , Since $U$ has the dimension of energy, the system under the limit lies in the finite energy configuration. The leading term of $H_4$ as $\ell_s \to 0$ is where $g_{\rm YM5}$ is the five-dimensional Yang-Mills coupling defined by $g_{\rm YM5}^2 = 2\pi g_s \ell_s$, while $H_0$ is where $U_h := \rho_h \ell_s^{-2}$. We fix the density of D0-branes, to be large but finite, and introduce the UV cutoff, $U_\infty$, by so that we concentrate on the IR region defined by . Actually the analyses in the rest of this paper basically depend on the configuration of IR. Thus the harmonic function $H_0$ is reduced to Substituting and into , we obtain The remarkable point is that, within this region, the coupling $e^\phi$ decreases as a function of $U$, which is interpreted as an energy scale. The RR fluxes does not change by our field theory limit. Note that taking $n_0 \to 0$ limit is not consistent with .
For later convenience, we introduce the coordinates rescaled by $R$ and $\ell_s$ as so that $x^\mu$, $y$ and $r$ are dimensionless coordinates. We defined the radius $R$ by following the notation in . Then the metric and dilaton become $$\eqalignno{
&R^{-2} ds^2 =
\sqrt{\nu} dx_\mu dx^\mu
+{r^3 f_M(r) \over \sqrt{\nu}} dy^2
+{\sqrt{\nu} \over r^3} \biggl({dr^2 \over f_M(r)} +r^2 d\Omega_4^2 \biggr) \,, &\CONFmetricdl{a} \cr
&e^{\phi} = g_s {\nu^{3/4} \over r^{3/2}} \,, \quad
f_M(r) = 1 -\biggl({r_M \over r}\biggr)^3 \,, \quad
\nu := {(2\pi\ell_s)^4 n_0 \over N_4} \,, \quad
v_4 := \int d^4x = {V_4 \over R^4} \,, \cr
&&\CONFmetricdl{b} \cr
}$$ where $r_M := \ell_s^2 U_h /R$. The factor, $\ell_s^4$, in $\nu$ is an artifact caused by , and we recall that, in taking the field theory limit, $n_0$ is finite but $\nu$ is not. The RR fluxes are rewritten as
We compactify the $y$ direction by a circle with a period $\beta_y$ in the same manner as , so that we obtain the four-dimensional dual gauge theory. In order for the smoothness at $r = r_M$, the period $\beta_y$ has to be $$\beta_y = {4\pi \sqrt{\nu} \over 3r_M^2} \,.$$ Then the Kaluza-Klein mass and the Yang-Mills coupling in the four-dimensional dual gauge theory are described as
The background given by and corresponds to the confined phase of the dual gauge theory. If we take the T-dual on the $y$ direction, this D4+D0 becomes the D3+D($-1$) system which was studied by . The D$(-1)$-brane plays a role of axion/gluon condensation in the dual gauge theory, because a D$p$-brane can be realized as an instanton on D$(p+4)$-branes. Therefore the D0-branes smeared on the D4-branes behave as the instantons and provide the “axion condensation”.
We need to clarify the parameter region in which the supergravity approximation is valid. First, in order to suppress the loop correction, the effective string coupling must be small, namely, $e^\phi \ll 1$. Since $e^\phi$ decreases as $r$, this constraint is realized as $g_s \nu^{3/4}r_M^{-3/2} \ll 1$, that is, where $\lambda$ is the ’t Hooft coupling, $\lambda := g_{\rm YM}^2 N_4$, in four dimensions. Second, the curvature must be smaller than $\ell_s^{-2}$ for neglecting a higher curvature correction. We calculate the scalar curvature from the metric , Since this is the increasing function for large $r$, the UV cutoff $r_\infty (= \ell_s^2 U_\infty/R)$ is manifestly necessary so that ${\cal R}(r_\infty) \ll \ell_s^{-2}$, namely, $r_\infty \ll \sqrt{\nu}R^2 \ell_s^{-2}$. This constraint with $r_M < r_\infty$ leads to
Here we recall , which provides the condition, $r_M \ (\leq r_\infty) \ll \nu^{1/3}$. This is equivalent in terms of the field theory parameters to The three conditions, , and , are consistent with the ’t Hooft limit, namely, the $N_4 \to \infty$ limit with fixing $\lambda$ to a large finite number as well as fixing the ratio, $n_0/N_4$, to a very large finite number.
We can also write down the background geometry corresponding to the deconfined phase. By exchanging the coordinates and double Wick-rotation, namely, $(x^0, y) \to (iy, ix^0)$, in and , we obtain the metric, $$\eqalignno{
&R^{-2} ds^2 = -{r^3 f_T(r) \over \sqrt{\nu}} \bigl(dx^0\bigr)^2
+\sqrt{\nu} \bigl(d{\vec x}^2 +dy^2 \bigr)
+{\sqrt{\nu} \over r^3} \biggl({dr^2 \over f_T(r)} +r^2 d\Omega_4^2 \biggr) \,, &\DECmetricdl{a} \cr
&e^{\phi} = g_s {\nu^{3/4} \over r^{3/2}} \,, \quad
f_T(r) = 1 -\biggl({r_T \over r}\biggr)^3 \,, \quad
\nu = {(2\pi\ell_s)^4 N_0 \over R^4 v_4}{1 \over N_4} \,, \quad
v_4 := \int dx^1 dx^2 dx^3 dy \,, \cr
&&\DECmetricdl{b} \cr
}$$ and the RR fluxes, The transition between the background geometry for the confined phase and the one for the deconfined phase is known as the Hawking-Page transition. Since we exchanged the time direction with the radial one, the configuration of D-branes is slightly modified as Table 2.
Since the geometry is a black hole, we can introduce a temperature by compactifying the Euclidean time direction $\tau\, (:=ix^0)$ with a period $\beta_\tau$ which is determined to be $(4\pi/3)\sqrt{\nu} r_T^{-2}$. Then it leads to the (Hawking) temperature,
It is remarkable that the power of effective couplings $e^\phi$, and , has a negative value, $-3/2$. This implies the UV asymptotic freedom. On the other hand, the dilaton in the D4-branes background without D0-branes behaves like $e^\phi \sim r^{+3/4}$ , that is, asymptotically non-free. Therefore we expect that the UV behavior in our holographic QCD model is improved by virtue of the D0-branes.
In order to consider the flavor dynamics in the confined phase, we embed $N_f$ D8-branes and $N_f$ anti-D8-branes into the background , in which the D8-branes and anti-D8-branes are connected with each other, so that they become $N_f$ U-shape D8-branes (see fig. 1). This implies the chiral symmetry breaking from $U(N_f) \times U(N_f)$ to $U(N_f)$. The D8-branes are stretched on $(x^0,x^1,x^2,x^3,r,S^4)$ and the collective coordinate $y$ depends only on $r$ (see Table 3).
Then the action of D8-branes without a worldvolume gauge field is where $\CT = N_f\mu_8 R^9 \Omega_4/g_s$, $\mu_8 = (2\pi)^{-8}\ell_s^{-9}$ and $'$ denotes the derivative with respect to $r$. We calculate the equation of motion from the action , Integrating this equation with respect to $r$ and imposing the boundary condition, we obtain Therefore we can describe the separation between the D8-branes at the UV cutoff, $r = r_\infty$, as
Now let us turn on the baryons which are homogeneously distributed in $\Bbb{R}^3(x^1,x^2,x^3)$. These baryons are realized by baryon vertices, that is, the (baryonic) D4-branes wrapping $S^4$ . We assume that the baryon vertices are located on the D8-branes at $r = r_c$. From the viewpoint of the worldvolume theory on the D8-branes, the number of baryons corresponds to $U(1)$ charge, because the Chern-Simons action includes the term, $A^{U(1)}_0 \tr (F^{SU(N_f)})^2$.
More concretely we study the action of D8-branes. It consists of the DBI action and the Chern-Simons action, $$S_8 = S_{\rm DBI} +S_{\rm CS} \,.$$ The $U(N_f)$ worldvolume gauge field, $\CA$, on the D8-branes can be decomposed as $$\CA = A^{SU(N_f)} + {1 \over \sqrt{2N_f}}A^{U(1)} \,.$$ Since we are interested in the U(1) baryon charge, we focus on the time component of $U(1)$ gauge field, $A^{U(1)}_0$, which depends only on $r$, and we assume that the other components are equal to zero. For later convenience we define the dimensionless field by Due to this gauge field, the DBI action is modified as
We introduce the baryon vertices at $r=r_c$. We then have to take into account the Chern-Simons term coupling with the four-form flux $G_{(4)}$, because this includes the term inducing the baryon charge, Since the baryon vertices are at $r=r_c$ and are homogeneously distributed in $\Bbb{R}^3$, $\tr F^2$ is identified with the baryon number density as Therefore becomes We note that, although there is the other RR flux, $G_{(8)}$, in , it does not couple with $a(r)$, the time component of $U(1)$ gauge field.
Since the baryonic D4-branes are attached on the D8-branes at $r=r_c$, the force from D4-branes has to balance with that from the D8-branes .
We remind the total action of D8-branes given by and , [*i.e.*]{}, $$\eqalignno{
&S_8 = \CT \int d^4x dr
\bigl[-\CL_{\rm DBI} +3 N_B a(r) \delta(r-r_c) \bigr] \,, &\conACT{a} \cr
&\CL_{\rm DBI} = \nu \sqrt{r^2f_M(r) \bigl(y'(r)\bigr)^2+{\nu \over r^4 f_M(r)} -{1 \over r} \bigl(a'(r)\bigr)^2} \,. &\conACT{b}
}$$ The equations of motion for $y(r)$ and $a(r)$ are respectively $$\eqalignno{
&{d \over dr}\Biggl[{r^2f_M(r)y'(r) \over \CL_{\rm DBI}}\Biggr] = 0 \,, &\conEOM{a} \cr
&d'(r) = 3 N_B \delta(r-r_c) \,, &\conEOM{b}
}$$ where we introduced the electric displacement field, From we obtain $$d(r) = 3 N_B \,,$$ that is, the electric displacement field becomes the constant proportional to the baryon density. With the boundary condition, $y'(r_0) = \infty$, we can solve as
We are now considering the baryonic D4-branes which are attached on the D8-branes at $r=r_c$. Then we can imagine two cases about the shape of D8-branes. These depend on the direction of force from the baryonic D4-branes. One is the case as a where the V-shape D8-branes are pulled down by the falling D4-branes. We remember that this is the case for the generalized Sakai-Sugimoto model, [*i.e.*]{}, the D4-branes background without the D0-branes . The other is as b where the W-shape D8-branes are hung up by the floating D4-branes.
In order to clarify the shape of D8-branes in our D4+D0 model, we calculate the on-shell action of baryonic D4-branes. The baryon vertices with density $n_B$ are identified with the $n_B v_3$ D4-branes wrapping $S^4$. Note that $v_3\, (:= \int d^3x)$ is the volume of $\Bbb{R}^3$. Therefore the on-shell action of baryonic D4-branes is Since the energy, $\CE_4$, decreases as $r_c$, the D4-branes yield an upward force (to the positive direction of $r$). Therefore our D4+D0 model prefers the W-shape D8-brane than the V-shape one. We also need to take into account the possibility as c, in which the W-shape D8-branes and the baryonic D4-branes are connected by fundamental strings and the position of D4-branes is $r_b\, (\geq r_c)$ . The action of the baryonic D4-branes and fundamental strings are described as $$\eqalignno{
S_{4+{\rm st}} &= -(n_B v_3) \biggl[\mu_4 \int dx^0 d\Omega_4\,e^{-\phi}\sqrt{-\det g_{\rm D4}}
+{1 \over 2\pi\ell_s^2}\int dx^0 \int_{r_c}^{r_b} dr\sqrt{-\det g_{\rm string}}\biggr]_{\hbox{\sevenrm on-shell}} \cr
&= -\CT v_3 d\sqrt{\nu} \int dx^0 \biggl[{1 \over 3\sqrt{r_b}}
+\int_{r_c}^{r_b}{dr \over \sqrt{r^3-r_M^3}}\biggr]
=: -\int dx^0\, \CE_{4+{\rm st}} \,.
}$$ Since the derivative of the energy $\CE_{4+{\rm st}}$ with respect to $r_b$ is always positive, $\CE_{4+{\rm st}}$ has a minimum at $r_b = r_c$. In other words, the configuration like b is favorable than that like c. Finally the W-shape D8-branes on which the baryonic D4-branes directly attached (b) is the (at least classically) stable configuration .
For the W-shape D8-branes we calculate the separation at the UV cutoff $r = r_\infty$ in terms of , $$\eqalignno{
l &= 2\biggl(-\int_{r_c}^{r_0}+\int_{r_0}^{r_\infty}\biggr) dr\, |y'(r)| \cr
&= 2\biggl(\int_{r_0}^{r_c}+\int_{r_0}^{r_\infty}\biggr) dr {\sqrt{\nu} \over r^3 f_M(r)}\biggl({r^2 f_M(r)(\nu^2 +rd^2) \over r_0^2 f_M(r_0)(\nu^2 +r_0 d^2)} -1\biggr)^{-{1 \over 2}} \,. &\conSep
}$$ Now let us calculate the force balance condition between the D8-branes and the baryonic D4-branes (see ). The Legendre transform of the DBI action of D8-branes is $$\eqalignno{
{\tilde S}_8 &:= \CT \int d^4x dr \Bigl(a'(r) d+\CL_{\rm DBI}\Bigr) \cr
&= \CT \int d^4x dr \sqrt{(\nu^2 +rd^2)\biggl(r^2f_M(r)\bigl(y'(r)\bigr)^2 +{1 \over r^4 f_M(r)}\biggr)} \,. &\conLegeight
}$$ From this equation, we can read that the effective tension at $r=r_c$ is The angle between the D8-branes and the $r$ axis at $r=r_c$ is computed as The D8-branes provide the force, $f_8\cos \theta$, along the $r$ direction. On the other hand, the force from the D4-branes is evaluated in terms of , As a result, from , and , we obtain the force balance condition, $f_8\cos\theta = f_4$, namely,
In order to consider the chiral condensate, we adopt the method suggested by , in which the open Wilson line operator, $$\CO^j_i(x^\mu) = \psi_L^{\dagger j}(x^\mu,y=-l/2)
{\cal P}e^{\int_{-l/2}^{l/2} dy (iA_y +\Phi)} \psi_{Ri}(x^\mu,y=l/2) \,,$$ was introduced. $i,j$ are indices of the fundamental representation of $U(N_f)$. The AdS/CFT correspondence allows us to evaluate the vacuum expectation value of the open Wilson line operator by the use of the on-shell action of fundamental string whose boundary is the flavor D8-branes (see ), that is to say, $$\langle \CO_i^j \rangle \simeq \delta_i^j \langle \CO \rangle \,, \quad
\langle \CO \rangle = e^{-S_\CO} \,, \quad
S_\CO ={1 \over 2\pi\ell_s^2}\int_\Sigma d^2\sigma \sqrt{\det g} \,,$$ to leading order of $\alpha'\, (=\ell_s^2)$. One can realize $\langle \CO \rangle$ as the chiral condensate $\langle{\bar \psi}\psi\rangle$. $S_\CO$ is, in other words, the area of the worldsheet instanton. By the use of , $S_\CO$ can be written down as We note that $S_\CO$ does not have UV divergence, because we have introduced the UV cutoff, $r_\infty$, in our model as we have discussed in Section 2. However $S_\CO$ still has a large value depending on the cutoff. So we consider $\langle \CO \rangle$ divided by the large constant,
Let us evaluate the chiral condensate numerically. In order to sweep out $r_M$ from numerical computations, we rescale the variables as $$w:={r \over r_M} \,, \quad {\hat d}:= \sqrt{r_M} d \,, \quad
{\hat l} := r_M^2 l \,.$$ Then the force balance condition is rewritten as $$w_0^2 (1-w_0^{-3})(\nu^2+w_0{\hat d}^2)
= w_c^2 (1-w_c^{-3})\biggl(\nu^2 +w_c {\hat d}^2{35+w_c^{-3} \over 36}\biggr)
=: Q(\nu,{\hat d},w_c) \,.$$ This equation determines $w_0$ as a function of $\nu$, ${\hat d}$ and $w_c$. From , and , the separation and chiral condensate are described as $$\eqalignno{
&{\hat l} = 2\biggl(\int_{w_0}^{w_c}+\int_{w_0}^{w_\infty}\biggr) dw {\sqrt{\nu} \over w^3-1}\biggl({w^2 (1-w^{-3})(\nu^2 +w{\hat d}^2) \over Q(\nu,{\hat d},w_c)} -1\biggr)^{-{1 \over 2}} \,, &\conSepK \cr
&{\langle \CO \rangle \over \langle \CO \rangle_\infty}
= \exp \biggl({R^2 \over \pi \ell_s^2 r_M}I\biggr) \,, \cr
&\quad I = \biggl(\int_{w_0}^{w_c}+\int_{w_0}^{w_\infty}\biggr) dw {\sqrt{\nu} \over w^2+w+1}\biggl({w^2 (1-w^{-3})(\nu^2 +w{\hat d}^2) \over Q(\nu,{\hat d},w_c)} -1\biggr)^{-{1 \over 2}} \,. &\conregCCK
}$$ By fixing $\nu$, ${\hat l}$ and $w_\infty$, we calculate the density ${\hat d}$ dependence of $\exp(I)$.
For instance, when we set $w_\infty=10^4$, ${\hat l} = 0.1$ and $\nu = 0.1,1,10$, we obtain . Note that the contribution from large $w$ to the integrations in and is very small, therefore $\exp(I)$ is not sensitive to the cutoff $w_\infty$. We can see in that the chiral condensate, $\langle \CO \rangle\, (=\langle \CO \rangle_\infty(e^I)^{R^2/(\pi\ell_s^2r_M)})$, increases in very low density. This is different from our intuition from ordinary QCD. On the other hand, in high density, the chiral condensate is a monotonically decreasing function of ${\hat d}$ and converges to a finite value as ${\hat d} \to \infty$. In all other known models of holographic QCD , the chiral condensate increases in high density and it is completely opposite to the expectation from ordinary QCD. Therefore one can say that our model drastically improves the high density behavior of chiral condensate.
The background corresponding to the deconfined phase has been given by and . We embed the $N_f$ D8-branes (and $N_f$ anti-D8-branes) into it in the same manner as we did in the confined phase, that is to say, we identify the worldvolume coordinates of the D8-branes with $(x^0,x^1,x^2,x^3,r,\Omega_4)$ and assume that the collective coordinate, $y$, depends only on $r$. Since we are interested in the $U(1)$ baryon charge, we take into account the $U(1)$ gauge field denoted by . Then the DBI action is described as
There are two Chern-Simons terms coupling with the $U(1)$ gauge field, $a(r)$. One is the coupling with the four-form flux , [*i.e.*]{}, $\int G_{(4)} \wedge \omega_{(5)}$, originated from the background D4-branes. As we explained in Section 3.1, $\omega_{(5)}$ includes $A^{U(1)}\tr (F^{SU(N_f)})^2$. Therefore it provides us the source term of the $U(1)$ baryon charge, in the same way as . This also implies that the $N_B$ baryonic D4-branes wrapping $S^4$ are located at $r = r_c$.
In the deconfined phase there is the other Chern-Simons term induced by the eight-form flux , $G_{(8)}$, of the D0-branes, namely, $\int G_{(8)} \wedge \CA_0 dx^0$. This term includes We note that this Chern-Simons term yields the $U(1)$ charge even if there is not a baryon ($N_B=0$).
In this subsection we concentrate on the case of no baryon, that is, $N_B = 0$. From and , the action of D8-branes is We calculate the equations of motion from this action, $$\eqalignno{
&{d \over dr}\biggl[{\nu r^2 f_T(r) y'(r) \over \CL_{\rm DBI}^{\rm (dec)}} -3a(r)\biggr] = 0 \,, &\decEOMNoB{a} \cr
&{d \over dr}\biggl[{\nu a'(r) \over r \CL_{\rm DBI}^{\rm (dec)}} -3y(r) \biggr] = 0 \,. &\decEOMNoB{b}
}$$ Since these equations are symmetric under the constant shifts of the fields, we redefined the fields by Then, integrating with respect to $r$, we obtain $$\eqalignno{
&\psi'(r) =
{3\alpha(r) \over \sqrt{r^3 f_T(r) \Bigl\{ 9r^3f_T(r)\bigl(\psi(r)\bigr)^2 -9\bigl(\alpha(r)\bigr)^2 +r^2f_T(r) \Bigr\}}} \,, &\decEOMNoBre{a} \cr
&\alpha'(r) =
{3 \sqrt{r^3 f_T(r)} \psi(r) \over \sqrt{9r^3f_T(r)\bigl(\psi(r)\bigr)^2 -9\bigl(\alpha(r)\bigr)^2 +r^2f_T(r)} } \,, &\decEOMNoBre{b}
}$$ where the integration constants are absorbed by $C_1$ and $C_2$. Note that, when $(\psi(r), \alpha(r))$ is a solution of , $(-\psi(r), -\alpha(r))$ is also a solution.
The trivial solution is $\alpha(r) = \psi(r) = 0$. When we put a pair of boundary conditions, $y(r_\infty)=\pm l/2$, this solution leads to a pair of solutions, $y(r) =\pm l/2$, which represents the parallel D8-branes and anti-D8-branes separated from each other by $l$. This configuration preserves the chiral symmetry, $U(N_f) \times U(N_f)$.
Let us look for a non-trivial solution. In order to find a U-shape solution whose tip is located at $r=r_0$, we impose the boundary condition: From , one can interpret $\psi'(r_0) =\infty$ to $r_0^2 f(r_0) -9\bigl(\alpha(r_0)\bigr)^2 = 0$. Therefore the boundary condition can be replaced with We are now considering the $N_B = 0$ case, [*i.e.*]{}, there is no source induced by a baryon. Therefore there does not exit a cusp singularity on the D8-branes. In terms of the solution of and , such a smooth U-shape D8-brane is described as where $C_1$ and $C_2$ in are determined so that $y(r)$ and $a(r(y))$ are smooth at $y(r_0) = 0$. Since it is difficult to solve analytically, we perform numerical computation. For instance, in fixing $r_T=1$, we obtain the solutions depicted in . The D8-branes and anti-D8-branes are smoothly connected at $r=r_0$, so that the U-shape D8-branes appear. This implies the chiral symmetry breaking from $U(N_f) \times U(N_f)$ to $U(N_f)$. Although the gauge field diverges at $r \to \infty$, we neglect the large $r$ region. Because, as we discussed in Section 2.2, the UV cutoff $r_\infty$ is necessary due to the validity of supergravity approximation.
We shall turn on the baryons which are the $n_B v_3$ baryonic D4-branes located at $r=r_c$. The total action of D8-branes is given by , and because of $N_B \neq 0$. Although the equation of motion for $y(r)$ is same as , the one for $a(r)$ is slightly different from due to the Chern-Simons term . By redefining the equation of motion for $a(r)$ has a simple expression, $d'(r) = 3N_B \delta(r-r_c)$. Therefore $d(r)$ is a constant equal to $3N_B$.
In terms of we write down the on-shell action of the baryonic D4-branes wrapping $S^4$ as Since the energy $\CE_4^{\rm (dec)}$ monotonically increases as $r_c$, the baryonic D4-branes obtain a force toward the negative direction of $r$. Therefore this force from the D4-branes balances with the force from the V-shape D8-branes in a manner as a. By using the solution of $\psi(r)$ and $\alpha(r)$ that we obtained in the previous subsection, we can describe the V-shape D8-brane as Note that the gauge field $a(r)$ is not continuous at $r = r_c$ because there is the baryonic D4-branes, namely, the source of $U(1)$ charge.
The Legendre transformed action of D8-branes is up to total derivative. From and we calculate the force balance condition, The separation of V-shape D8-branes at $r=r_\infty$ is In the same way as Section 3.3, we shall study the density dependence of the chiral condensate $\langle \CO\rangle$ by $$\eqalignno{
{\langle \CO\rangle \over \langle \CO\rangle_\infty}
&= \exp \biggl[{R^2 \over \pi\ell_s^2}\int_{r_c}^{r_\infty}dr \int_0^{y(r)}dy {\sqrt{\nu} \over r^3 f_T(r)}
-{R^2 \over \pi\ell_s^2}\int_{r_c}^{r_\infty}dr \int_0^{l/2}dy {\sqrt{\nu} \over r^3 f_T(r)} \biggr] \cr
&= \exp \biggl({R^2\sqrt{\nu} \over 2\pi\ell_s^2} I_{(D)}\biggr) \,, \quad
I_{(D)} = \int_{r_c}^{r_\infty} dr {2y(r) -l \over \sqrt{r^3 -r_T^3}} \,. &\decCCK
}$$
For instance we numerically evaluate $\exp(I_{(D)})$ with fixed $l=0.5$ and $r_T=\nu = 1$, which is depicted in . It shows that the chiral condensate monotonically decreases as the density $d$. This agrees very well with our expectation from the ordinary QCD. The $\exp(I_{(D)})$ for $l=0.5$ and $r_T=\nu = 1$ is approximately equal to 0.966 at $d=0$, and converges to a finite value, $0.947$, at $d =\infty$ (see ).
In this paper we proposed a model of holographic QCD whose background has the asymptotic freedom. It is the specific field theory limit of the supergravity solution of the color D4-branes on which the D0-branes are smeared. In taking the limit, we fixed the ratio $\nu$ to be finite. Into this background we have introduced the flavor D8-branes and the baryonic D4-branes in order to study the chiral condensate in dense baryonic medium. The dilaton in the background of D4-branes without D0-branes behaves like $e^\phi \sim r^{3/4}$. On the other hand, the dilaton in our background behaves like $e^\phi\sim r^{-3/2}$, that is, the effective coupling is asymptotically free. Although it is still different from $(\log r)^{-1}$ in ordinary QCD, the D0-branes improve the asymptotic freedom in our model.
We have studied the chiral condensate by use of the open Wilson line, which we can evaluate by the area of open string worldsheet whose boundary is the D8-branes. In the confined phase the baryonic D4-branes balance with the D8-branes which have W-shape rather than V-shape. This is because the force from the D4-branes is toward the positive direction of the radial coordinate. This can be understood that the curvature increases for large radius so that the gravity is acting on baryon vertex upward.
Under this force balance condition we have calculated the chiral condensate concretely with the fixed ${\hat l}$, the separation between the D8-branes at the UV cutoff. The results for three cases of $\nu$ are shown in . All of the three have the common features as follows. In the region of very low density of baryons, the chiral condensate increases as the density. This behavior is not what one expects from the QCD. We attribute this anomaly to the approximation of the uniform distribution of baryons in the space $\Bbb{R}^3$, which is not a good one when the density of baryon is small. In high density we have obtained a remarkable result, that is, the chiral condensate monotonically decreases. This is in good agreement with our intuition from the ordinary QCD. When the density goes to infinity, the chiral condensate converges to a certain constant.
We have studied also the deconfined phase. We numerically solved the equations of motion without the baryons, and obtained the U-shape solution of D8-branes. It seems that the solution of the gauge field has UV divergence. However we do not need to seriously worry about the divergence, because basically the UV cutoff is necessary in our background due to the validity of supergravity approximation of bulk theory. Turning on baryons, we have calculated the chiral condensate. It is different from the confined phase that the D8-brane balancing with the baryonic D4-branes has V-shape in the deconfined phase. As a result, it monotonically decreases in whole range of the density (see ). This is in agreement with the expectation from ordinary QCD. In the deconfined phase there is the solution of parallel D8-branes, which preserves the chiral symmetry. Therefore we expect the chiral symmetry restoration occurs in certain high density . One can guess that this transition is clarified by comparing the free energies of two configurations, namely, the V-shape D8-branes with baryon vertices and the parallel D8-branes. However there is subtlety due to the UV cutoff. Therefore we need to develop a way beyond the UV cutoff.
[**Acknowledgement**]{}
This work was supported by the NRF grant through the SRC program CQUeST with grant number 2005-0049409 and partially supported by the WCU project of Korean Ministry of Education, Science and Technology (R33-2008-000-10087-0).
[^1]: $^\dagger$
[^2]: $^\ddagger$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the reinforcement learning problem of training multiple agents in order to maximize a shared reward. In this multi-agent system, each agent seeks to maximize the reward while interacting with other agents, and they may or may not be able to communicate. Typically the agents do not have access to other agent policies and thus each agent observes a non-stationary and partially-observable environment. In order to obtain multi-agents that act in a decentralized manner, we introduce a novel algorithm under the framework of centralized learning, but decentralized execution. This training framework first obtains solutions to a multi-agent problem with a single centralized joint-space learner. This centralized expert is then used to guide imitation learning for independent decentralized multi-agents. This framework has the flexibility to use any reinforcement learning algorithm to obtain the expert as well as any imitation learning algorithm to obtain the decentralized agents. This is in contrast to other multi-agent learning algorithms that, for example, can require more specific structures. We present some theoretical error bounds for our method, and we show that one can obtain decentralized solutions to a multi-agent problem through imitation learning.'
author:
- |
Alex Tong Lin\
Department of Mathematics\
UCLA\
`[email protected]`\
Mark J. Debord\
NAVAIR\
Katia Estabridis\
NAVAIR\
Gary Hewer\
NAVAIR\
Stanley Osher\
UCLA\
bibliography:
- 'references.bib'
title: 'CESMA: Centralized Expert Supervises Multi-Agents'
---
Introduction
============
Reinforcement Learning (RL) is the problem of finding an action policy that maximizes reward for an agent embedded in an environment [@SuttonB18]. It has recently has seen an explosion in popularity due to its many achievements in various fields such as, robotics [@levine2016end], industrial applications [@datacentercooling], game-playing [@mnih2015human; @silver2017mastering; @silver2016mastering], and the list continues. However, most of these achievements have taken place in the single-agent realm, where one does not have to consider the dynamic environment provided by interacting agents that learn and affect one another.
This is the problem of Multi-agent Reinforcement Learning (MARL) where we seek to find the best action policy for each agent in order to maximize their reward. The settings may be cooperative, and thus they might have a shared reward, or the setting may be competitive, where one agent’s gain is another’s loss. Some examples of a multi-agent reinforcement learning problem are: decentralized coordination of vehicles to their respective destinations while avoiding collision, or the game of pursuit and evasion where the pursuer seeks to minimize the distance between itself and the evader while the evader seeks the opposite. Other examples of multi-agent tasks can be found in [@Panait2005] and [@maddpg].
The key difference between MARL and single-agent RL (SARL) is that of interacting agents, which is why the achievements of SARL cannot be absentmindedly transferred to find success in MARL. Specifically, the state transition probabilities in a MARL setting are inherently non-stationary from the perspective of any individual agent. This is due to the fact that the other agents in the environment are also updating their policies, and so the Markov assumptions typically needed for SARL convergence are violated. This aspect of MARL gives rise to instability during training, where each agent is essentially trying to learn a moving target. In this work, we present a novel method for MARL in the cooperative setting (with shared reward). Our method first trains a centralized expert with full observability, and then uses this expert as a supervisor for independently learning agents. There are a myriad of imitation/supervised learning algorithms, and in this work we focus on adapting DAgger (Dataset Aggregation) [@DAgger] to the multi-agent setting. After the imitation learning stage, the agents are able to successfully act in a decentralized manner. We call this algorithm Centralized Expert Supervises Multi-Agents (CESMA). CESMA adopts the framework of centralized training, but decentralized execution [@KRAEMER201682], the end goal of which is to obtain multi-agents that can act in a decentralized manner.
Related works
=============
The most straight-forward way of adapting single-agent RL algorithms to the multi-agent setting is by having agents be independent learners. This was applied in [@Tan:1997:MRL:284860.284934], but this training method gives stability issues, as the environment is non-stationary from the perspective of each agent [@Matignon:2012:RIR:2349641.2349642; @Busoniu2010MultiagentRL; @claus1998dynamics]. This non-stationarity was examined in [@OmidshafieiPAHV17], and stabilizing experience replay was studied in [@foerster2017stabilising]
Another common approach to stabilizing the environment is to allow the multi-agents to communicate. In [@SukhbaatarSF16], they examine this using continuous communications so one may backpropagate to learn to communicate. And in [@foerster2016learning], they give an in-depth study of communicating multi-agents, and also provide training methods for discrete communication. In [@jpaulos_schen], they decentralize a policy by examining what to communicate and by utilizing supervised learning, although they mathematically solve for a centralized policy and their assumptions require homogeneous communicating agents.
Others approach the non-stationarity issue by having the agents take turns updating their weights while freezing others for a time, although non-stationarity is still present [@egorov2016multi]. Other attempts adapt $Q$-learning to the multi-agent setting: Distributed Q-Learning [@Lauer00analgorithm] updates Q-values only when they increase,and updates the policy only for actions that are not greedy with respect to the $Q$-values, and Hysteretic $Q$-Learning [@hysteretic] provides a modification. Other approaches examine the use of parameter sharing [@gupta2017cooperative] between agents, but this requires a degree of homogeneity of the agents. And in [@tesauro2004extending], their approach to non-stationarity was to input other agents’ parameters into the Q function. Other approaches to stabilize the training of multi-agents are in [@sukhbaatar2016learning], where the agents share information before selecting their actions.
From a more centralized view point, [@oliehoek2008optimal; @qmix; @sunehag2017value] derived a centralized $Q$-value function for MARL, and in [@DBLP:journals/corr/UsunierSLC16], they train a centralized controller and then sequentially select actions for each agent. The issue of an exploding action space was examined in [@tavakoli2018action].
A few works that follow the framework of centralized training, but decentralized execution are: RLar (Reinforcement Learning as Rehearsal) [@KRAEMER201682], COMA (Counterfactual Multi-Agent), and also [@ijcai2018-774; @DBLP:journals/corr/abs-1810-11702] – where the idea of knowledge-reuse is examined. In [@NIPS2017_6887], they examine decentralization of policies from an information-theoretic perspective. There is also MADDPG [@maddpg] where they train in a centralized-critics decentralized-actors framework; after training completes, the agents are seperated from the critics and can execute in a fully distributed manner.
For surveys of MARL, see articles in [@bu2008comprehensive; @panait2005cooperative].
Background
==========
In this section we briefly review the requisite material needed to define MARL problems. Additionally we summarize some of the standard approaches in general reinforcement learning and discuss their use in MARL. **Dec-POMDP:** A formal framework for multi-agent systems is called a decentralized partially-observable Markov decision process (Dec-POMDP) [@bernstein2005bounded]. A Dec-POMDP is a tuple $(I, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, P, R)$ where $I$ is the finite set of agents indexed $1$ to $M$, $S$ is the set of states, $\mathcal{A}_i$ is the set of actions for agent $i$, and thus $\prod_{i=1}^M \mathcal{A}_i$ is the joint action space, $\mathcal{O}_i$ is the observation space of agent $i$, and thus $\prod_{i=1}^M \mathcal{O}_i$ is the joint observation space, $P = P(\textbf{s}', \textbf{o} | \textbf{s}, \textbf{a})$ (where $\textbf{o} = (o_1,\ldots,o_M)$ and similarly for the others) is the state-transition probability for the whole system, and $R:\mathcal{S} \times \prod_{i=1}^M\mathcal{A} \rightarrow \mathcal{R}$ is the reward. In the case when the joint observations $\textbf{o}$ equals the world state of the system, then we call the system a decentralized Markov decision process (Dec-MDP).
**DAgger:** The Dataset Aggregation (DAgger) algorithm [@DAgger] is an iterative imitation learning algorithm that seeks to learn a policy from expert demonstration. The main idea is to allow the learning policy to navigate its way through the environment, and have it query the expert on states that the it sees. It does this by starting with a policy $\hat{\pi}_2$ which learns from the dataset of expert trajectories $\mathcal{D}_1$ through supervised learning. Using $\hat{\pi}_2$, a new dataset is generated by rolling out the policy and having the expert provide supervision on the decisions that the policy made. This new dataset is aggregated with the existing set into $\mathcal{D}_2 \supset \mathcal{D}_1$. This process is iterated, i.e. a new $\hat{\pi}_3$ is trained, another new dataset is obtained and aggregated into $\mathcal{D}_3 \supset \mathcal{D}_2$ and so on. Learning in this way has been shown to be more stable and have nicer convergence properties as learning utilizes trajectories seen from the learner’s state distribution, as opposed to only the expert’s state distribution.
**Policy Gradients (PG):** One approach to RL problems are policy gradient methods [@NIPS1999_1713]: instead of directly learning state-action values, the parameters $\theta$ of the policy $\pi_\theta$ are instead adjusted to maximize the objective, $$J(\theta) = \mathbb{E}_{s\sim p^\pi, a\sim \pi_\theta}\left[\nabla_\theta \log \pi_\theta(a|s) Q^\pi(s,a) \right]$$ where $p^\pi$ is the state distribution from following policy $\pi$. The gradient of the above expression can written as [@NIPS1999_1713; @SuttonB18]: $$\nabla_\theta J(\theta) = \mathbb{E}_{s\sim p^\pi, a\sim \pi_\theta} [\nabla_\theta \log \pi_\theta(s|a) Q^\pi(s,a)]$$ Many policy gradient methods seek to reduce the variance of the above gradient estimate, and thus many study how one estimates $Q^\pi(s,a)$ above [@SchulmanMLJA15]. For example, if we let $Q^\pi(s,a)$ be the sample return $R^t = \sum_{i=t}^T \gamma^{i-t} r_i$, then we get the REINFORCE algorithm [@REINFORCE]. Or one can choose to learn $Q^\pi(s,a)$ using temporal-difference learning [@Sutton1988LearningTP; @SuttonB18], and would obtain the Actor-Critic algorithms [@SuttonB18]. Other policy gradients algorithms are: DPG [@SilverDPG], DDPG [@LillicrapHPHETS15], A2C and A3C [@MnihBMGLHSK16], to name a few.
Policy Gradients have been applied to multi-agent problems; in particular the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) [@maddpg] uses an actor-critic approach to MARL, and this is the main baseline we test our method against. Another policy gradient method is by [@FoersterFANW17] called Counterfactual Multi-Agent (COMA), who also uses an actor-critic approach.
Methods
=======
In this section, we explain the motivation and method of our approach: Centralized Expert Supervises Multi-Agents (CESMA).
Treating a multi-agent problem as a single agent problem {#subsec:matosa}
--------------------------------------------------------
Intuitively, an optimal strategy of a multi-agent problem could be found by a centralized expert with full observability. This is because the centralized controller has the most information available about the environment, and therefore would not pay a high of cost of partial-observability that independent learners might. This is discussed more in \[sec:theory\].
To find this centralized expert, we treat a multi-agent problem as a single agent problem in the joint observation and action space of all agents. This is done by concatenating the observations of all agents into one observation vector for the centralized expert, and the expert learns outputs that represent the joint actions of the agents.
Our framework does not impose any other particular constraints on the expert. Any expert architecture that outputs an action that represents the joint-actions of all of the agents may be used. Due to that, we are free to use any standard RL algorithm for the expert such as DDPG, DQN, or potentially even analytically derived experts.
Curse of dimensionality and some reliefs
----------------------------------------
When training a centralized expert, both the observation space and action space can grow exponentially. For example, if we use a DQN for our centralized expert then the number of output nodes will typically grow exponentially. This is due to each output needing to correspond to an element in the joint action space $\prod_{i=1}^M \mathcal{A}_i$.
This problem has been studied by QMIX [@qmix] and VDNs (Value-Decomposition Networks) [@sunehag2017value], where exponential scaling of the output space is solved by having separate $Q$ values for each agent and then using the sum as a system $Q$. Other techniques such as action branching [@tavakoli2018action] have also been considered.
In our experiments, we use DDPG (with Gumbel-Softmax action selection if the environment is discrete, as MADDPG does also) to avoid the exploding input nodes of the observation space, as well as exploding output nodes of the action space. Under this paradigm, the input and output nodes only grow linearly with the number of agents.
CESMA for supervised learning on the multi-agents
-------------------------------------------------
In order to perform imitation learning to decentralize the centralized expert policy, we adapt DAgger to the multi-agent setting. There are many ways DAgger can be applied to multi-agents, but we implement a method that best allows the theoretical analysis from [@DAgger] to apply: Namely after training the expert, we do supervised learning on a single neural network with disconnected components, each of which corresponds to a multi-agent.
In more detail, after training a centralized expert $\pi^*$, we initialize the $M$ agents $\pi_{\theta_1}, \ldots, \pi_{\theta_M}$, and initialize the dataset $\mathcal{D}$. The agents then step through the environment, storing each observation the multi-agents encounter along with its action label: $(\textbf{o}, \textbf{a}_{\text{expert}})$, where $\textbf{a}_{\text{expert}}=\pi^*(\textbf{o})$. After $\mathcal{D}$ has reached a sufficient size, at every $k$th time step (chosen by the practitioner; we used $k=1$ in our experiments), we sample a batch from this dataset $\{(\textbf{o}^{(b)}, \textbf{a}^{(b)}_{\text{expert}})\}_{b=1}^B$, and then distribute observation $o^{(b)}_i$ and action-label $a^{(b)}_{\text{expert},i}$ to agent $i$. The observation and action label is then used to perform supervised learning on the agents. Having a shared dataset of trajectories in this way allows us to view $(\pi_{\theta_1}, \ldots, \pi_{\theta_M})$ as a single neural-network with disconnected components, and thus the error bounds from [@DAgger] directly apply, as discussed in Section \[sec:theory\].
See Figure \[fig:supervise\_multiagents\] for a diagram. The appendix contains pseudo-code for our method.
[r]{}[0.5]{} 
The aformentioned procedure is sufficient when the agents do not have to communicate in order to solve the environment (e.g. they all have full observability, or their observations are isomorphic to full observability, or the partial observability is not too detrimental). When communication is involved, then we have to modify the above method.
CESMA for multi-agents that communicate {#subsec:cesma_comm}
---------------------------------------
In order to train multi-agents that communicate, especially in an environment where communication is necessary (see Section \[subsec:coopcomm\] for two such cases), we have to slightly modify the above procedure. Training the centralized expert in a communication scenario follows the original procedure, but since the expert has full observability, it has no incentive to communicate with itself. Indeed, most of the time the reward function is independent of the communication (although if one desired to act covertly, then this may not be the case). Thus any observations and actions regarding communication are not utilized when training the expert. In the next phase, the agents learn to communicate amongst themselves.
We leave the full details in the appendix, and opt here to provide an overview: In the environments we test, the communication action by an agent at time step $t-1$ will appear to other agents at the next time step $t$. Then in our dataset of observations, we store together the observations from time step $t-1$ and $t$.
The idea is to separate the losses of the agent actions into an *action loss* and a *communication loss*. In order to obtain the action loss, we do supervised learning on the observation and actions at time $t$ where the input is the observation but with updated communications from the other agents, and the label is the centralized expert action label. The communication loss for agent $i$ is computed by first obtaining agent $i$’s output communication action at $t-1$. This is then combined with the observations and actions and of all other agents (at time $t$). We query the expert on the correct action for the other agents, and we use this supervised learning loss to backpropagate the errors through to agent $i$ weights. Indeed, our main motivation was the chain rule.
By using this procedure we allow the decentralized agents to learn communications that help reduce the penalty associated with operating in a partially-observable environment. Section \[ref:theorycomms\] discusses this in more detail.
In this way, we have alleviated a bit the issue of communication being a sparse reward, which is because even when an agent communicates to another agent, the receiving agent is still learning, so it is unclear if the broadcasting agent was correct, or the receiving agent was correct; we have a double ambiguity. With an expert supervisor, the correct action by the acting agent is clear.
Theoretical analysis {#sec:theory}
====================
Although the proposed framework could handle a myriad of imitation learning algorithms, such as Forward Training [@ross2010efficient], SMILe [@ross2010efficient], SEARN [@daume2009search], and more, we use DAgger in our experiments, and thus we follow its theoretical analysis, and provide some multi-agent extensions. And since our method can be viewed as using a single-agent learner with disconnected components so that we can trivially decompose it into multi-agents, then this gives us direct theoretical insights [@DAgger].
In our setting, the centralized expert observes the joint observations of all agents, and thus it is a function $\pi^*:\mathcal{O}_1 \times \cdots \times \mathcal{O}_M \rightarrow \mathcal{A}_1 \times \cdots \times \mathcal{A}_M$, and we can decompose $\pi^*$ into, $$\pi^*(\textbf{o}) = (\pi^*_1(\textbf{o}) , \ldots, \pi^*_M(\textbf{o}))$$ where $\pi^*_i:\mathcal{O}_1 \times \cdots \times \mathcal{O}_M \rightarrow \mathcal{A}_i$. In order to decentralize our policy, our goal is to find policies $\pi_1, \ldots, \pi_M$ such that, $$\begin{split}
\pi^*(\textbf{o}) &= (\pi^*_1(\textbf{o}) , \ldots, \pi^*_M(\textbf{o})) \\
&\overset{\text{want}}{=} (\pi_i(o_1), \ldots, \pi_M(o_M))
\end{split}$$ Note that $\pi^*_i$ is able to observe the joint observations while $\pi_i$ is only able to observe its own local observation $o_i$. This means we may encounter issues where $$\pi^*_i(o_1, \ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M) = a_i, \quad \text{but} \quad \pi^*_i(o_1', \ldots, o_{i-1}', o_i, o_{i+1}', \ldots, o_M') = a_i'$$ so we want $$\pi_i(o_i) = a_i \text{ or } a_i',$$ Thus the agent policy can act sub-optimally in certain situations, being unaware of the global observation. This unfortunate situation not only afflicts our algorithm, but any multi-agent training algorithm (and in general, any algorithm attempting to solve a POMDP).
We call this the *partial observability problem of decentralization* (note partial observability afflicts any algorithm trying to solve a POMDP (e.g. multi-agent systems), but here we examine from the viewpoint of decentralization).
More concretely, we can say there is is a partial observability problem of decentralization if there exists observations $(o_1, \ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M)$ and $(o_1', \ldots, o_{i-1}', o_i, o_{i+1}', \ldots, o_M')$ such that $\pi^*(o_1, \ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M) \neq \pi^*(o_1', \ldots, o_{i-1}', o_i, o_{i+1}', \ldots, o_M')$.
As it pertains to our algorithm, if we can bound the cost of this problem, then we can obtain an error bound. Then suppose $\ell$ is the $0$-$1$ loss of matching the expert policy $\pi^*$, and suppose the partial observability problem of decentralization is such that $$\mathbb{E}_{\textbf{o} \sim d_{(\pi^{(i)}_1, \ldots, \pi^{(i)}_M)}} \left[\ell(\textbf{o}, (\pi_1,\ldots,\pi_M) \right] \ge c_p$$ for all iterations $i$, observations $\textbf{o}$ and policies $(\pi_1,\ldots, \pi_M)$. Then we have the following theorem (whose proof is in the appendix),
\[thm:TCP\] If the number of iterations $N$ is $\tilde{O}(T)$, and $J(\pi)$ is the cost (negative of the reward) of executing policy $\pi$, then there exists a joint multi-agent policy $\hat{\pi} \in \{\hat{\pi}^{(i)}\}_{i=1}^N$ such that $$\begin{split}
J(\hat{\pi}) &\le J(\pi^*) + T\epsilon_N + O(1) \\
&\le J(\pi^*) + Tc_p + O(1)
\end{split}$$ where, $$\epsilon_N = \min_{(\pi_1, \ldots, \pi_M) \in \Pi_1\times \cdots \times \Pi_M} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\,\sim\, d_{ (\pi^{(i)}_1, \ldots, \pi^{(i)}_M) }} [\ell(\textbf{o}, (\pi_1, \ldots, \pi_M)]$$
Under these assumptions, this implies that decentralizing the policy always requires paying a cost of partial observability, independent of the number of iterations $N$. Although in our experiments, we sometimes find this cost is negligible.
The need for communication {#ref:theorycomms}
--------------------------
The most effective setting for our methods is when the multi-agents are all able to observe the full joint observation. From the perspective of each agent the only non-stationarity when learning comes from other agents’ policies, as opposed to from the non-agent portion of the environment.
In the situation when each agent only has local observations, then to avoid the partial observability problem of decentralization, there is an incentive to communicate. Namely, we want for the multi-agent policy $(\pi_1, \ldots, \pi_M)$, $$\begin{split}
\pi^*(\textbf{o}) &= (\pi^*_1(\textbf{o}) , \ldots, \pi^*_M(\textbf{o})) \\
&\overset{\text{want}}{=} (\pi_i(o_1, c_1), \ldots, \pi_M(o_M, c_M))
\end{split}$$ where $c_i$ is the communication from either all or only some of the other agents, to agent $i$. Note that we can view $c_i$ as a function $c_i:\mathcal{O}_1 \times \cdots \times \mathcal{O}_{i-1} \times \mathcal{O}_{i+1} \times \cdots \times \mathcal{O}_{M} \rightarrow \mathcal{C}_i$ (where $\mathcal{C}_i$ is some communication action space). Then under the following conditions, we have the error bound:
\[thm:comm\] If the multi-agent communication $c_i:\mathcal{O}_1 \times \cdots \times \mathcal{O}_{i-1} \times \mathcal{O}_{i+1} \times \cdots \times \mathcal{O}_{M} \rightarrow \mathcal{C}_i$ satisfies the following condition: $$\begin{split}
&\pi^*(o_1,\ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M) \neq \pi^*(o_1',\ldots, o_{i-1}', o_i, o_{i+1}', \ldots, o_M') \\
&\qquad \text{implies that } c_i(o_1,\ldots, o_{i-1}, o_{i+1}, \ldots, o_M) \neq c(o_1',\ldots, o_{i-1}', o_{i+1}', \ldots, o_M')
\end{split}$$ for all $i=1,\ldots, M$, then there is no partial observability cost of decentralization, and thus if the number of iterations $N$ is of $\tilde{O}(T)$, then there exists a policy $\hat{\pi}_{1,\ldots,M} \in \{ \hat{\pi}_{1,\ldots,M}^{(i)} \}_{i=1}^N$ such that $\mathbb{E}_{\textbf{o} \sim d_{\hat{\pi}_{1,\ldots,M}}} \left[ \ell(\textbf{o}, \hat{\pi}_{1,\ldots,M})\right] \le \epsilon_N + O(1/T)$. Then this implies, $J(\hat{\pi}_{1,\ldots,M}) \le J(\pi^*_{1,\ldots,M}) + \epsilon_N T + O(1)$
The proof is in the appendix. We remark that a naive communication protocol satisfying the above conditions is when $c_i$ is the identity operator.
Experiments
===========
Our experiments are conducted in the Multi-Agent Particle Environment [@mordatch2017emergence; @maddpg] provided by OpenAI, which has basic simulated physics (e.g. Newton’s law) and multiple multi-agent scenarios.
Preliminaries
-------------
In order to conduct comparisons to MADDPG, we also use the DDPG algorithm with the Gumbel-Softmax [@jang2016categorical; @maddison2016concrete] action selection for discrete environments, as they do. For the single-agent centralized expert neureal network, we always make sure the number of parameters matches (or is lower) than that of MADDPG’s. For the decentralized agents, we use the same number of parameters as the decentralized agents in MADDPG (i.e. the actor part). We always use the discount factor $\gamma=0.9$, as that seemed to work best both for our centralized expert, and also MADDPG. Following their experimental procedure, we average our experiments over three runs. And for the decentralization, we trained three separate centralized experts, and used each of them to obtain three decentralized policies. Full details of our hyperparameters is in the appendix. Rewards are also averaged over 1,000 episodes. And we always use two-hidden layer neural networks. Brief descriptions of each environment are provided, and fuller descriptions and some example pictures are placed in the appendix.
Cooperative Navigation – Homogeneous and Nonhomogenous Agents {#subsec:coopnav}
-------------------------------------------------------------
[r]{}[0.75]{}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
Here we examine the situation of $N$ agents occupying $N$ landmarks in a 2D plane, and the agents are either homogeneous or heterogeneous. The (continuous) observations of each agent are the relative positions of other agents, the relative positions of each landmark, and its own velocity. The agents do not have access to others’ velocities so we have partial observability. The reward is based on how close each landmark has an agent near it, and the actions of each agent are discrete: up, down, left, right, and do nothing.
In Figure \[fig:learning\_curves\], in all cases the fully centralized expert is able to achieve a lower reward than MADDPG and DDPG. We are also able to decentralize the expert policy (which was chosen to be the one with lowest reward) so as to reach this same reward. And we remark that our method seems to work better with more agents.
The six nonhomogeneous case works as a good experiment to see what happens when we stop the centralized expert before it truly converges. In this case, decentralization to achieve the same reward as the expert is quickest and occurs within the first 5,000 episodes. Intuitively, it makes sense that a suboptimal solution allows faster convergence.
We observe our method improves sample efficiency over MADDPG: we picked centralized experts that achieved a reward of -350 (averaged over its past 1,000 episodes) which took the centralized expert around 50,000 episodes. Decentralizing these polices so that we obtain a reward higher than -355 took around 7,000 episodes, so in total it took 57,000 episodes to obtain decentralized policies achieving a reward higher than -355. But it takes MADDPG around 80,000 to sometimes more than 100,000 episodes to obtain rewards higher than -355.
We also examined the use of DQNs, one with an exponential number of output nodes, and a QMIX/Centralized VDN that sums the individual agents $Q$-values so we get linear growth in action. And we also test decentralization performance when each agent has its own dataset of trajectories, in the 3 nonhomogeneous agents scenario. We hypothesized that the nonhomogeneity of the agents may have an effect on convergence, but this turned out not to be so. See Figure \[fig:learning\_curves\] for the reward curves.
Cooperative Communication {#subsec:coopcomm}
-------------------------
Here we we adapt CESMA to a task that involves communication. In this scenario, the communication action taken by each agent at time step $t-1$ will appear to the other agents at time step $t$. Although we require continuous communication to backprop, in practice we can use the softmax operator to provide the bridge between the discrete and continuous. And during decentralized execution, our agents are able to act with discrete communication inputs.
We examine three scenarios for CESMA that involve communication, and use the training scenario described in section \[subsec:cesma\_comm\]. The first scenario called the “speaker and listener" environment has a speaker who broadcasts the correct goal landmark (in a “language" it must learn) out of a possible 3 choices, and the listener, who is blind to the correct goal landmark, must use this information to move there. Communication is a necessity in this environment. The second scenario is cooperative navigation with communication and here we have two/three agents whose observation space includes the goal landmark of the *other* agent, and not their own, and there are three/five possible goal landmarks. We see in figure \[fig:learning\_curves\] that we achieve a lower reward and in a more sample efficient manner. For the speaker and listener environment, the centralized expert near-immediately converges, and same for the decentralization process. And MADDPG has a much higher variance in its convergence. In the cooperative navigation with communication scenarios, the story is similar, that the centralized expert quickly converges, and the decentralization process is near immediate.
Reward vs. loss, and slow and fast learners
-------------------------------------------
[r]{}[0.5]{}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
\[fig:loss\_v\_reward\]
In our experiments with cooperative navigation, when using the cross entropy loss, we did not find an illuminating correlation between the reward and the loss. We reran the experiments in a truer DDPG fashion by solving a continuous version of the environment, and used the mean-squared error for the supervised learning. We examined the loss in the cooperative navigation task with 3 agents, both homogeneous and nonhomogeneous agents. We plot the figures in \[fig:loss\_v\_reward\]. We found that in these cases, the reward and loss were negatively correlated as expected, namely that we achieved a higher reward as the loss decreased. In the nonhomogeneous case, we plot each individual agents’ reward vs its loss and found that the big and slow agent had the biggest loss, followed by the medium agent, and the small and fast agent being the quickest learner. This example demonstrates that in nonhomogeneous settings, some agents may be slower to imitate the expert than others.
Conclusion
==========
We propose a MARL algorithm, called Centralized Expert Supervises Multiagents (CESMA), which takes the training paradigm of centralized training, but decentralized execution. The algorithm first trains a centralized expert policy, and then adapts DAgger to obtain decentralized policies that execute in a decentralized fashion.
Experiment where each agent has its own dataset of trajectories
===============================================================
Here we describe in a little bit more detail our procedure for experimenting with individual dataset trajectories for each agent, vs a shared dataset. Namely, we plot the learning curves for decentralizing a policy in the two cases: (1) When each agent has its own dataset of trajectories, or (2) when there is a shared dataset of trajectories (which is the one we use in the experiments). We tested on the cooperative navigation environment with 3 nonhomogeneous agents. We hypothesized that the nonhomogeneity of the agents would have an effect on the shared reward, but this turned out not to be so. But it is interesting to note that in the main text, we found that the some agents had a bigger loss when doing supervised learning from the expert.
Experiment with DQNs
====================
Here we examined decentralizing DQNs. We briefly described it in the main text, and give a bit more info here. We used the cross entropy loss for the supervised learning portion, and used the cooperative navigation environment with 3 nonhomogenous agents. The DQNs we used are: the exponential actions DQN, which is just a naive implementation of DQNs for the multi-agents, and a Centralized VDN where the system $Q$ value is the sum of the individual agent $Q$ values. We used a neural network with 200 hidden units, batch size 64, and for the exponential DQN, we used a learning rate and $\tau$ of $5\times 10^{-4}$, and for the QMIX/Centralized VDN DQN we used a learning rate and $\tau$ of $10^{-3}$. We also used a noisy action selection for exploration.
Pseudo-algorithm of CESMA (without communication)
=================================================
.
Detailed description of CESMA with communicating agents
=======================================================
The main intuition for our idea is very simple: the chain rule. In order to minimize notational burden, we first leave out some details, and then provide them after: If the supervised learning loss function is $\ell$, the expert is $\pi^*$, and if we have agents: $\pi_{1}, \ldots, \pi_{M}$, then in order to update the communication of agent $i$ to agent $j$, then denoting $c_{-i}$ the communications from all other agents to agent $j$ excluding agent $i$, then we want to minimize: $$\min_{\pi_i} \ell(\pi^*(\textbf{o})_j, \;\; \pi_{j}(o_j, c_{-i} \cup \pi_{i}(o_i)_{\text{comm}})))$$ and so we should backpropagate through other agents’ actions (namely the physical actions which have the expert label, and not the communication action which does not) to update agent $i$’s communication. We note both $c_{-i}$ and $\pi_i(o_i)_{\text{comm}}$ are communication actions that are obtained from the current policy of the agents, i.e. we re-query the agents for their communication actions.
And to update the action loss of agent $i$, we want to minimize $$\min_{\pi_i} \ell(\pi^*(\textbf{o})_i, \;\; \pi_{i}(o_i, \textbf{c})_{\text{action}})$$ (note bold characters are vectors, e.g. $\textbf{o} = (o_1,\ldots, o_M)$). We note that $\textbf{c}$ is an up-to-date communication action, which we obtain by re-querying the other agents for their communication actions.
The above leaves out some notation for ease of understanding and to get the main idea. The technically and notationally correct version is:
If
- the supervised learning loss function is $\ell$,
- the expert is $\pi^*$,
- the agents at the current iteration are: $\pi_{1}, \ldots, \pi_{M}$,
- $o_i$ is the observation of agent $i$,
- $\hat{o}_i$ is the observation of agent $i$ at the next timestep (i.e. after observing $o_i$ and communications from other agents, and then taking a step in the environment, to get a new observation),
- $c_{i}$ is the communication action of other agents towards agent $i$ that accompanies observation $o_i$,
- $\hat{c}_i$ is the updated (i.e. using the most up-to-date policy) communication action of agents towards agent $i$, i.e. $\hat{c}_i = (\pi_j(o_j, c_j)_{\text{comm}, i})_{j\neq i}$
then we want to minimize: $$\tag{communication loss}
\min_{\pi_i} \frac{1}{M-1}\sum_{j=1, j\neq i}^M \ell(\pi^*(\hat{\textbf{o}})_j, \;\; \pi_{j}(\hat{o}_j, \; \hat{c}_j)_{\text{action}})$$ where we note $\hat{c}_j$ contains the communication action from agent $i$ to agent $j$, and thus we can backprop, and the subscript “action" means the physical action (and not the communication action).
And for the action loss, we want to minimize: $$\tag{action loss}
\min_{\pi_i} \ell(\pi^*(\hat{\textbf{o}})_i, \;\; \pi_{i}(\hat{o}_{i}, \;\hat{c}_{i} )_{\text{action}})$$ where $\hat{c}_i$ is the updated communications from all other agents, to agent $i$ (i.e. we use the current most-up-to-date policy to form $\hat{c}_i$. The need to use an updated version of communication $\hat{c}_i$ is because the language of the agents tends to change more drastically as training progresses. (One thing we did try was to set a low size limit on the trajectory dataset $\mathcal{D}$, but preliminary results showed it did not help much). But we do note that in order to construct $\hat{c}_i$, we need to use the observation/communications from time step $t-1$, so reliance on an older language is still there. One can imagine going all the way back to $t=0$ in order to update the communications of the whole trajectory, but for our experiments, going back one timestep sufficed.
We give a pseudocode in algorithm \[alg:supervising\_multiagents\].
A diagram of the action loss and communication loss is given in figure \[fig:action\_loss\] (action loss) and \[fig:comm\_loss\] (communication loss).
We remark that we also considered the case of a hybrid objective, where the actions are learned by supervised learning from the expert, and the communication is learned similar to a standard RL algorithm (e.g. the $Q$-values are communication actions). Preliminary results showed this did not work well.
![A diagram of the computation of the action loss for agent $i$. The $\pi_j$’s are the most up-to-date policies, and $\hat{c}_i$ is an updated communication from the other agents.[]{data-label="fig:action_loss"}](figures3/supervised_learning_multiagents_with_comm_action_loss_2_2.png)
![A diagram of the computation of the communication loss for agent $i$. The $\pi_j$’s are the most-up-to-date policies, and the $\hat{c}_j$’s are updated communications from the other agents. Note we calculate the action loss of agent $j$, $j\neq i$, and then backpropagate this loss to agent $i$’s weights.[]{data-label="fig:comm_loss"}](figures3/supervised_learning_multiagents_with_comm_comm_loss_2_2.png)
More detailed descriptions of the environments used in the experiments
======================================================================
Cooperative navigation
----------------------
The goal of this scenario is to have $N$ agents occupy $N$ landmarks in a 2D plane, and the agents are either homogeneous or heterogeneous. The environment consists of:
- Observations: The (continuous) observations of each agent are the relative positions of other agents, the relative positions of each landmark, and its own velocity. Agents do not have access to other’s velocities, and thus each agent only *partially observes* the environment (aside from not knowing other agents’ policies).
- Reward: At each timestep, if $A_i$ is the $i$th agent, and $L_j$ the $j$th landmark, then the reward $r_t$ at time $t$ is, $$\begin{aligned}
r_t = -\sum_{j=1}^N \min \left\{ \| A_i - L_j \| : i=1,\ldots, N \right\}
\end{aligned}$$ This is a sum over each landmark of the minimum agent distance to the landmark. Agents also receive a reward of $-1$ at each timestep that there is a collision.
- Actions: Each agents’ actions are discrete and consist of: up, down, left, right, and do nothing. These actions are acceleration vectors (except do nothing), which the environment will take and simulate the agents’ movements using basic physics (i.e. Newton’s law).
Speaker listener
----------------
In this scenario, the goal is for the listener agent to reach a goal landmark, but it does not know which is the goal landmark. Thus it is reliant on the speaker agent to provide the correct goal landmark. The observation of the speaker is just the color of the goal landmark, while the observation of the listener is the relative positions of the landmark. The reward is the distance from the landmark.
- Observations: The observation of the speaker is the goal landmark. The observation of the listener is the communication from the speaker, as well as the relative positions of each goal landmark.
- Reward: The reward is merely the negative (squared) distance from the listener to the goal landmark.
- Actions: The actions of the speaker is just a communication, a 3-dimensional vector. The actions of the listener are the five actions: up, down, left, right, and do nothing.
Cooperative navigation with communication
-----------------------------------------
In this particular scenario, we have one version with 2 agents and 3 landmarks, and another version with 3 agents and 5 landmarks. Each agent has a goal landmark that is only known by the other agents. Thus the each agent must communicate to the other agents its goal. The environment consists of:
- Observations: The observations of each agent consist of the agent’s personal velocity, the relative position of each landmark, the goal landmark for the other agent (an 3-dimensional RGB color value), and a communication observation from the other agent.
- Reward: At each timestep, the reward is the sum of the distances between and agent and its goal landmark.
- Actions: This time, agents have a movement action and a communication action. The movement action consists of either not doing anything, or outputting an acceleration vector of magnitude one in the direction of up, down, left, or right; so do nothing, up, down, left right. The communication action is a one-hot vector; here we choose the communication action to be a 10-dimensional vector.
Hyperparameters
===============
- For all environments, we chose the discount factor $\gamma$ to be $0.9$ for all experiments, as that seemed to benefit both the centralized expert as well as MADDPG (and as well as independently trained DDPG). And we always used a two-hidden-layer neural network for all of MADDPG’s actors and critics, as well as the centralized expert, and the decentralized agents. The training of MADDPG used the hyperparameters from the MADDPG paper [@maddpg], with the exception of having $\gamma=0.9$ (instead of 0.95), as that improved MADDPG’s performance.
- For the cooperative navigation environments with 3 agents, for both homogeneous and nonhomogeneous: Our centralized expert neural network was a two-hidden-layer neural network with 225 units each (as that matched the number of parameters for MADDPG when choosing 128 as their number of hidden units for each of their 3 agents), and we used a batch size of 64. The learning rate was 0.001, and $\tau=0.001$. We also clipped the gradient norms to $0.1$. When decentralizing, each agent was a two-hidden-layer neural network with 128 units (as in MADDPG), where we trained with a batch size of 32 and a learning rate of 0.001. In our experiment comparing with MADDPG, we use the cross entropy loss. The MADDPG and DDPG parameters were 128 hidden units, and we clipped gradients norms at 0.5, with a learning rate of 0.01.
- For the cooperative navigation with 6 agents, for both homogeneous and nonhomogeneous: Our centralized expert neural network was a two-hidden-layer neural network with 240 units each (as that matched the number of parameters for MADDPG when choosing 128 as their number of hidden units for each of their 3 agents’ actor and critic), and we used a batch size of 32. The learning rate was 0.0001, and $\tau=0.0001$. We also clipped the gradient norms to $0.1$. When decentralizing, each agent was a two-hidden-layer neural network with 128 units (as in MADDPG), where we trained with a batch size of 32 and a learning rate of 0.001. In our experiment comparing with MADDPG, we use the cross entropy loss. The MADDPG and DDPG parameters were 128 hidden units, and we clipped gradients norms at 0.5, with a learning rate of 0.01.
- For the speaker and listener environment: Our centralized expert neural network was a two-hidden-layer neural network with 64 units each (which gave a lower number of parameters than MADDPG when choosing 64 as their number of hidden units for each of their 2 agents’ actor and critic), and we used a batch size of 32. The learning rate was 0.0001, and $\tau=0.001$. When decentralizing, each agent was a two-hidden-layer neural network with 64 units (as in MADDPG), where we trained with a batch size of 32 and a learning rate of 0.001. In our experiment comparing with MADDPG, we use the cross entropy loss. The MADDPG and DDPG parameters were 64 hidden units, and we clipped gradients norms at 0.5, with a learning rate of 0.01.
- For the cooperative navigation with communication environment: Our centralized expert neural network was a two-hidden-layer neural network with 95 units each (which matched the number of parameters as MADDPG when choosing 64 as their number of hidden units for each of their 2 agents’ actor and critic), and we used a batch size of 32. The learning rate was 0.0001, and $\tau=0.0001$. When decentralizing, each agent was a two-hidden-layer neural network with 64 units (as in MADDPG), where we trained with a batch size of 32 and a learning rate of 0.001. In our experiment comparing with MADDPG, we use the cross entropy loss. The MADDPG and DDPG parameters were 64 hidden units, and we clipped gradients norms at 0.5, with a learning rate of 0.01.
- We also run all the environments for 25 time steps.
Proofs of theorems
==================
Mathmetical notation and preliminaries
--------------------------------------
For completeness, we provide here proofs of the theorems from the main text. The proofs heavily borrow from [@DAgger].
In order to reduce notational burden, we denote, $$\pi_{1,\ldots,M} = (\pi_1, \ldots, \pi_M)$$ so that $$\pi_{1,\ldots,M}(\textbf{o}) = (\pi_1(o_1), \ldots, \pi_M(o_M))$$ where $\textbf{o} = (o_1, \ldots, o_M)$ is the joint multi-agent observation. And we notate $\hat{\pi}_{1,\ldots,M}$ similarly.
We also denote $d^t_{\pi_{1,\ldots,M}}$ the distribution of states seen at time $t$, and $d_{\pi_{1,\ldots,M}} = \frac{1}{T} \sum_{t=1}^T d^t_{\pi_{1,\ldots,M}}$ the average distribution of states encountered if we follow $\pi_{1,\ldots,M}$ for $T$ time-steps.
And letting $C(\textbf{o}, \textbf{a})$ be the cost of executing action $\textbf{a}$ in state $\textbf{o}$, and letting $C_{\pi_{1,\ldots,M}}(\textbf{o}) = \mathbb{E}_{\textbf{a} \sim \pi_{1,\ldots, M}(\textbf{o})} \left[ C(\textbf{o}, \textbf{a}) \right]$, then we let $$J(\pi_{1,\ldots,M}) = \sum_{t=1}^T \mathbb{E}_{\textbf{o} \sim d^t_{\pi_{1,\ldots, M}}} \left[ C_{\pi_{1,\ldots,M}}(\textbf{o}) \right] = T \mathbb{E}_{\textbf{o}\sim d_{\pi_{1,\ldots, M}}} \left[ C_{\pi_{1,\ldots,M}}(\textbf{o}) \right]$$
We use the following lemma which gives a total variation bound:
If $d_{\hat{\pi}_{1,\ldots,M}^{(i)}}$ is the average distribution of states encountered by $\hat{\pi}_{1\ldots,M}^{(i)}$ (the policy that best mimics the expert on past trajectories at iteration $i$), and if $d_{\pi_{1\ldots,M}}$ is the average distribution of states encountered by the policy $\pi_{1\ldots,M}^{(i)}$ (which picks expert actions with probability $\beta_i$ and $\hat{\pi}_{1,\ldots,M}^{(i)}$ otherwise), then we have $\|d_{\pi^{(i)}_{1,\ldots,M}} - d_{\hat{\pi}_{1,\ldots,M}^{(i)}} \|_1 \le 2T \beta_i$.
Suppose $d$ is the distribution of states encountered over $T$ steps for a policy $\pi_{1,\ldots,M}^{(i)}$ that picks expert actions $\pi^*$ at least once. Then since $\pi_{1,\ldots,M}^{(i)}$ picks actions from $\hat{\pi}_{1,\ldots,M}^{(i)}$ over $T$ steps with probability $(1-\beta_i)^T$, then we have $$d_{\pi_{1,\ldots,M}^{(i)}} = (1-\beta_i)^T d_{\hat{\pi}_{1,\ldots,M}^{(i)}} + (1-(1-\beta_i)^T) d$$ Then, $$\begin{split}
\left\|d_{\hat{\pi}_{1,\ldots,M}^{(i)}} - d_{\pi_{1,\ldots,M}^{(i)}} \right\|_1 = (1-(1-\beta_i)^T) \left\|d - d_{\hat{\pi}_{1,\ldots,M}^{(i)}} \right\|_1 \le 2(1-(1-\beta_i)^T) \le 2T\beta_i
\end{split}$$
Now we can prove:
If the number of iterations $N$ is $\tilde{O}(T)$, then there exists a policy $\hat{\pi}_{1,\ldots,M} \in \{\hat{\pi}_{1,\ldots,M}^{(i)}\}_{i=1}^N$ such that $\mathbb{E}_{\textbf{o} \sim d_{\hat{\pi}_{1,\ldots,M}}}[\ell(s, \hat{\pi}_{1,\ldots,M})] \le \epsilon_N + O(1/T)$ where $$\epsilon_N = \min_{\pi_{1,\ldots,M}\in \Pi_{1,\ldots,M}} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o} \sim d_{\pi^{(i)}_{1,\ldots,M}}}\left[\ell(\textbf{o}, \pi_{1,\ldots,M}) \right]$$
Let $\ell_{\text{max}} = \sup_{\textbf{o}, \pi_{1,\ldots,M}} \ell(\textbf{o}, \pi_{1,\ldots, M})$, and let $\{\beta_i\}_{i=1}^N$ be a non-increasing sequence, and suppose $n_\beta$ is the largest $n\le N$ where $\beta_n > \frac{1}{T}$.
Now, the lemma above implies, $$\mathbb{E}_{\textbf{o}\sim d_{\hat{\pi}_{1,\ldots,M}^{(i)}}}\left[\ell(s, \hat{\pi}_{1,\ldots,M}^{(i)}) \right] \le \mathbb{E}_{\textbf{o} \sim d_{\pi_{1,\ldots,M}^{(i)}}}\left[\ell(s, \hat{\pi}_{1,\ldots,M}^{(i)})\right] + 2\ell_{\text{max}} \min(1, T\beta_i)$$ Then let $\gamma_N$ be a constant bounding (the no-regret bound), $$\frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\sim \pi_{1,\ldots,M}^{(i)}} \left[ \ell(\textbf{o}, \pi_{1,\ldots, M}^{(i)}) \right] - \min_{\pi_{1,\ldots,M} \in \Pi_{1,\ldots,M}} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\sim \pi_{1,\ldots,M}^{(i)}} \left[ \ell(\textbf{o}, \pi_{1,\ldots, M}) \right]$$ Then we have the following set of inequalities, $$\begin{split}
\min_{\hat{\pi}_{1,\ldots, M} \in \{\hat{\pi}_{1,\ldots, M}^{(i)}\}_{i=1}^N} \mathbb{E}_{\textbf{o} \in d_{\hat{\pi}_{1,\ldots, M}}} \left[ \ell(s, \hat{\pi}_{1,\ldots, M}) \right] &\le \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\sim \hat{\pi}_{1,\ldots,M}^{(i)}} \left[ \ell(\textbf{o}, \hat{\pi}_{1,\ldots, M}^{(i)}) \right] \\
&\le \frac{1}{N} \sum_{i=1}^N \left( \mathbb{E}_{\textbf{o}\sim \pi_{1,\ldots,M}^{(i)}} \left[ \ell(\textbf{o}, \hat{\pi}_{1,\ldots, M}^{(i)}) \right] + 2\ell_{\text{max}} \min(1, T\beta_i) \right) \\
&\le \gamma_N + \frac{2 \ell_{\text{max}}}{N} \left[ n_\beta + T \sum_{i=n_\beta + 1}^N \beta_i \right] \\
& \qquad + \min_{\pi_{1,\ldots,M} \in \Pi_{1,\ldots, M}} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\sim \pi_{1,\ldots,M}^{(i)}} \left[ \ell(\textbf{o}, \pi_{1,\ldots, M}) \right] \\
&= \gamma_N + \epsilon_N + \frac{2\ell_{\text{max}}}{N} [n_\beta + T\sum_{i=n_\beta + 1}^N \beta_i]
\end{split}$$
So we have,
\[corr:first\_thm\] If the number of iterations $N$ is $\tilde{O}(T)$, then there exists a joint multi-agent policy $\hat{\pi} \in \{\hat{\pi}^{(i)}\}_{i=1}^N$ such that $$J(\hat{\pi}) \le J(\pi^*) + T\epsilon_N + O(1)$$ where, $$\epsilon_N = \min_{\pi_{1,\ldots,M} \in \Pi_{1,\ldots,M}} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\,\sim\, d_{\pi_{1,\ldots,M}^{(i)}}} [\ell(\textbf{o}, \pi_{1,\ldots,M}]$$
And now we can prove the Theorem 1 from the main text:
[1]{} Let $\ell = \ell(\textbf{o}, \pi_{1,\ldots,M})$ be the $0$-$1$ loss of matching the expert policy. And suppose we have the cost of partial observability, $$\mathbb{E}_{\textbf{o} \sim d_{\pi^{(i)}_{1,\ldots,M}}} \left[\ell(\textbf{o}, \pi_{1,\ldots,M}) \right] \ge c_p$$ If the number of iterations $N$ is $\tilde{O}(T)$, then there exists a joint multi-agent policy $\hat{\pi} \in \{\hat{\pi}^{(i)}\}_{i=1}^N$ such that $$\begin{split}
J(\hat{\pi}) &\le J(\pi^*) + T\epsilon_N + O(1) \\
&\le J(\pi^*) + Tc_p + O(1)
\end{split}$$ where, $$\epsilon_N = \min_{(\pi_1, \ldots, \pi_M) \in \Pi_1\times \cdots \times \Pi_M} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\,\sim\, d_{ (\pi^{(i)}_1, \ldots, \pi^{(i)}_M) }} [\ell(\textbf{o}, (\pi_1, \ldots, \pi_M)]$$
The condition on $c_p$ gives us that, $$\epsilon_N = \min_{(\pi_1, \ldots, \pi_M) \in \Pi_1\times \cdots \times \Pi_M} \frac{1}{N} \sum_{i=1}^N \mathbb{E}_{\textbf{o}\,\sim\, d_{ (\pi^{(i)}_1, \ldots, \pi^{(i)}_M) }} [\ell(\textbf{o}, (\pi_1, \ldots, \pi_M)] \le \frac{1}{N} N c_p = c_p.$$ And thus we have, $$\begin{split}
J(\hat{\pi}) &\le J(\pi^*) + T\epsilon_N + O(1) \\
&\le J(\pi^*) + Tc_p + O(1)
\end{split}$$
And now we prove Theorem \[thm:comm\] from the main text:
[2]{} If the multi-agent communication $c_i:\mathcal{O}_1 \times \cdots \times \mathcal{O}_{i-1} \times \mathcal{O}_{i+1} \times \cdots \times \mathcal{O}_{M} \rightarrow \mathcal{C}_i$ satisfies the following condition: $$\begin{split}
&\pi^*(o_1,\ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M) \neq \pi^*(o_1',\ldots, o_{i-1}', o_i, o_{i+1}', \ldots, o_M') \\
&\qquad \text{implies that } c_i(o_1,\ldots, o_{i-1}, o_{i+1}, \ldots, o_M) \neq c(o_1',\ldots, o_{i-1}', o_{i+1}', \ldots, o_M')
\end{split}$$ for all $i=1,\ldots, M$, then there is no partial observability problem of decentralization, and thus if the number of iterations $N$ is of $\tilde{O}(T)$, then there exists a policy $\hat{\pi}_{1,\ldots,M} \in \{ \hat{\pi}_{1,\ldots,M}^{(i)} \}_{i=1}^N$ such that $\mathbb{E}_{\textbf{o} \sim d_{\hat{\pi}_{1,\ldots,M}}} \left[ \ell(\textbf{o}, \hat{\pi}_{1,\ldots,M})\right] \le \epsilon_N + O(1/T)$. Then this implies, $J(\hat{\pi}_{1,\ldots,M}) \le J(\pi^*_{1,\ldots,M}) + \epsilon_N T + O(1)$
We need only show that the there is no partial observability problem of decentralization when the above conditions hold. As a note: we now have that the observation of agent $i$ is not just $o_i$, but now $(o_i, c_i)$.
Suppose for contradiction that there is a partial observability problem of decentralization, and thus there exists an agent $i$, observations $\textbf{o} = (o_1,\ldots, o_{i-1}, o_i, o_{i+1}, \ldots, o_M)$ and $\textbf{o}' = (o_1',\ldots, o_{i-1}', o_i o_{i+1}', \ldots, o_M')$ such that $$\pi^*(\textbf{o})_i = a_i, \quad \pi^*(\textbf{o}')_i = a_i', \quad a_i \neq a_i'.$$ Denote $\textbf{o}_{-i}$ as the observation without $o_i$ and similarly for $\textbf{o}_{-i}'$. We then see we must have, $$\pi_i(o_i, c_i(\textbf{o}_{-i})) = \pi_i(o_i, c_i(\textbf{o}_{-i}'))$$ *for all possible policies* $\pi_i$. But of course by the above conditions, $c_i(\textbf{o}_{-i}) \neq c_i(\textbf{o}_{-i}')$, and thus we can certainly construct a policy where the above does not hold true, a contradiction. And thus we must have that there is no partial observability problem of decentralization.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Kerry A. Seitz, Jr.[^1]'
- 'Alex Kennedy[^2]'
- Owen Ransom
- 'Bassam A. Younis'
- 'John D. Owens'
bibliography:
- 'finalPaper.bib'
date: 'August 30, 2013'
title: 'A GPU Implementation for Two-Dimensional Shallow Water Modeling'
---
[^1]: Corresponding author: [email protected]
[^2]: Corresponding author: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $\Delta$ be the optimal packing density of ${\mathbb{R}}^n$ by unit balls. We show the optimal packing density using two sizes of balls approaches $\Delta + (1 - \Delta) \Delta$ as the ratio of the radii tends to infinity. More generally, if $B$ is a body and $D$ is a finite set of bodies, then the optimal density $\smash{\Delta_{\{rB\} \cup D}}$ of packings consisting of congruent copies of the bodies from $\{rB\} \cup D$ converges to $\Delta_D + (1 - \Delta_D) \Delta_{\{B\}}$ as $r$ tends to zero.'
address: 'D. de Laat, CWI, Science Park 123, 1098 XG Amsterdam, The Netherlands'
author:
- '[David de Laat](http://www.daviddelaat.nl)'
title: Optimal densities of packings consisting of highly unequal objects
---
[^1]
Introduction
============
There has been extensive research into the determination of optimal monodisperse packing densities. A well-known example is Hales’s proof of the Kepler conjecture on the optimal sphere packing density in ${\mathbb{R}}^3$ [@Hales2005a]. More recently, packings with polydispersity have been investigated: New lower and upper bounds for the density of packings of spheres using several sizes have been given in respectively [@HopkinsStillingerTorquato2012] and [@LaatOliveiraVallentin2012]. In applications, sphere packings can be used to model many-particle systems, and here it is important to also consider polydispersity as this can “dramatically affect the microstructure and the effective properties of the materials” [@UcheStillingerTorquato2004]. In this note we discuss the case of wide dispersity; that is, the case where the size ratio of the larger to the smaller objects grows large. One would expect boundary behavior to become negligible as the ratio of the radii tends to infinity, which intuitively means the density converges to $\Delta + (1 - \Delta) \Delta$; see for example [@Torquato2001]. To the best of our knowledge a proof of this has not yet been published. Here we provide such a proof which uses standard techniques albeit it is not trivial.
We prove the following theorem:
[4.2]{} Suppose $B$ is a body and $D$ is a finite set of bodies. Then, $$\lim_{r \downarrow 0} \Delta_{\{rB\} \cup D}
= \Delta_D + (1 - \Delta_D) \Delta_{\{B\}}.$$
Here, given a set of bodies $D$, we denote by $\Delta_D$ the optimal packing density using the bodies from $D$. If we take $B$ to be the unit disk in ${\mathbb{R}}^2$ and $D = \{B\}$, then the theorem says the optimal packing density using two sizes of disks converges to $0.9913\ldots$ as the ratio of the radii goes to infinity. By using that the optimal sphere packing density in ${\mathbb{R}}^3$ is known to be $\pi/(3\sqrt{2}) = 0.7404\ldots$, the theorem says the optimal packing density using two sizes of balls converges to $0.9326\ldots$ as the ratio of the radii goes to infinity.
Packings and density
====================
We define a *body* to be a bounded subset of ${\mathbb{R}}^n$ that has nonempty interior and whose boundary has Lebesgue measure zero. Such a set is Jordan measurable, which means its Lebesgue measure can be approximated arbitrarily well by the volume of inner and outer approximations by finite unions of $n$-dimensional rectangles [@Bogachev2007]. Moreover, since the interior of a body is nonempty it contains a ball with strictly positive radius and hence has strictly positive Lebesgue measure.
A *packing* using a set of bodies $D$ is a set of congruent copies of the elements in $D$ such that the interiors of the copies are pairwise disjoint. In other words, a packing is of the form $$P = \Big\{ R_i B_i + t_i : i \in {\mathbb{N}}, R_i \in O(n), \, t_i \in {\mathbb{R}}^n, \, B_i \in D\Big\},$$ where $(R_i B_i^\circ + t_i) \cap (R_j B_j^\circ + t_j) = \emptyset$ for all $i \neq j$. Here $B_i^\circ$ denotes the interior of $B_i$, and $O(n)$ is the orthogonal group. Define $\Sigma_D$ to be the set of packings that use bodies from $D$ and $\Lambda_D$ the set of packings $P \in \Sigma_D$ that have *rational box periodicity*; that is, for which there exists a $p \in \mathbb Q$ such that $|P|+pe_i = |P|$ for all $i \in [n]$. Here $|P| = \bigcup P$ denotes the *carrier* of $P$ and $e_i$ is the $i$th unit vector.
The *density* and *upper density* (provided these exist) of a set $S \subseteq {\mathbb{R}}^n$ are defined as $$\rho(S) = \lim_{r \to \infty} \frac{\lambda(S \cap r C)}{r^n} \quad \text{and} \quad \overline\rho(S) = \limsup_{r \to \infty} \frac{\lambda(S \cap r C)}{r^n}.$$ Here $\lambda$ is the Lebesgue measure on ${\mathbb{R}}^n$, and $C$ is the axis-aligned unit cube centered about the origin. The upper density $\bar\rho(|P|)$ is defined for every $P \in \Sigma_D$, because for each $r > 0$, the set $|P| \cap r C$ is Lebesgue measurable with measure at most $r^n$. The density $\rho(|P|)$ is defined for every $P \in \Lambda_D$: Let $p \in \mathbb Q$ be a period of $P$, then $c = \lambda(|P| \cap kpC)/(kp)^n$ does not depend on $k \in {\mathbb{N}}$, and for $r$ inbetween $kp$ and $(k+1)p$ we have $$\frac{\lambda(|P| \cap kpC)}{((k+1)p)^n} \leq \frac{\lambda(|P| \cap rC)}{r^n} \leq \frac{\lambda(|P| \cap kpC)}{((k+1)p)^n} + \frac{((k+1)p)^n - (kp)^n}{((k+1)p)^n},$$ where both the rightmost term and $|\frac{\lambda(|P| \cap kpC)}{((k+1)p)^n} - c|$ converge to $0$ as $k \to \infty$.
We define the *optimal packing density* for packings that use bodies from $D$ by $$\Delta_D = \sup_{P \in \Sigma_D} \overline\rho(|P|) = \sup_{P \in \Lambda_D} \rho(|P|).$$ The second equality follows because for each $P \in \Sigma_D$, we can construct a packing from $\Lambda_D$ whose density is arbitrarily close to $\overline\rho(|P|)$ by taking the subpacking contained in a sufficiently large cube and tiling space with this part of the packing. One might wonder whether the optimal density depends on $C$ being a cube, but it follows from [@Groemer1963] that the optimal density $\Delta_D$ is also equal to $$\sup_{P \in \Lambda_D} \lim_{r \to \infty} \lambda(|P| \cap (rB+t))/\lambda(rB),$$ where $t$ is any point in ${\mathbb{R}}^n$ and where $B$ is any compact set that is the closure of its interior and contains the origin in its interior.
Approximating the interstitial space of a packing
=================================================
We first show that a packing, and hence the interstitial space of a packing, can be approximated uniformly by grid cubes. Given $S \subseteq {\mathbb{R}}^n$ and $k \in {\mathbb{Z}}$, define the packings $$G_k(S) = \Big\{ C_{k,t} : t \in {\mathbb{Z}}^n, \, C_{k,t} \subseteq S \Big\} \quad \text{and} \quad G^k(S) = \Big\{ C_{k,t} : t \in {\mathbb{Z}}^n, \, C_{k,t} \cap S \neq \emptyset \Big\},$$ where $C_{k,t}$ is the cube $[\frac{t_1}{2^k}, \frac{t_1+1}{2^k}] \times \cdots \times [\frac{t_n}{2^k}, \frac{t_n+1}{2^k}]$ having side length $2^{-k}$. Given a set $P$ of subsets of ${\mathbb{R}}^n$, let $P^c = {\mathbb{R}}^n \setminus |P|$.
\[lem:cube approx\] Let $D$ be a finite set of bodies. Then $$\rho(|G_k(|P|)|) \uparrow \rho(|P|) \quad \text{and} \quad \rho(|G^k(|P|)|) \downarrow \rho(|P|)$$ and hence $$\rho(|G_k(P^c)|) \uparrow 1-\rho(|P|) \quad \text{and} \quad \rho(|G^k(P^c)|) \downarrow 1-\rho(|P|)$$ as $k \to \infty$ for $P \in \Lambda_D$ uniformly.
Let $\varepsilon > 0$ and $P \in \Lambda_D$. Since $P$ has rational box periodicity, the packings $G_k(|P|)$ and $G^k(|P|)$ have rational box periodicity, which means the densities $\rho(|G_k(|P|)|)$ and $\rho(|G^k(|P|)|)$ are defined. For each $k \in {\mathbb{Z}}$ we have $$|G_k(|P|)| \subseteq |P| \subseteq |G^k(|P|)|,$$ hence $$\rho(|G_k(|P|)|) \leq \rho(|P|) \leq \rho(|G^k(|P|)|).$$
We have $$\rho(|G_k(|P|)|) = \lim_{r \to \infty} \frac{\lambda(|G_k(|P|)| \cap r C)}{r^n} \geq \lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \subseteq rC} \lambda(|G_k(B)|)$$ and $$\rho(|G^k(|P|)|) = \lim_{r \to \infty} \frac{\lambda(|G^k(|P|)| \cap r C)}{r^n} \leq \lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \cap rC \neq \emptyset} \lambda(|G^k(B)|).$$ Every $B \in D$ is Jordan measurable, so there exists a number $K = K(B, \varepsilon)$ such that $$\lambda(|G_k(B)|) \geq \lambda(B) - \varepsilon \quad \text{and} \quad \lambda(|G^k(B)|) \leq \lambda(B) + \varepsilon$$ for all $k \geq K$. Since $D$ is a finite set, this implies $$\rho(|G_k(|P|)|) \geq \lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \subseteq rC} (\lambda(B) - \varepsilon)$$ and $$\rho(|G^k(|P|)|) \leq \lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \cap rC \neq \emptyset} (\lambda(B) + \varepsilon)$$ for all $k \geq \max_{B \in D} K(B,\varepsilon)$. Since each body $B$ in the finite set $D$ is bounded, there exists a number $r_0 = r_0(D) \geq 0$ such that $$\begin{aligned}
\lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \subseteq rC} \lambda(B)
\geq \lim_{r \to \infty} \frac{\lambda(|P| \cap (r-r_0)C)}{r^n} = \lim_{r \to \infty} \frac{\lambda(|P| \cap rC)}{(r+r_0)^n} = \rho(|P|)\end{aligned}$$ and $$\begin{aligned}
\lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \cap rC \neq \emptyset} \lambda(B)
\leq \lim_{r \to \infty} \frac{\lambda(|P| \cap (r+r_0)C)}{r^n} = \lim_{r \to \infty} \frac{\lambda(|P| \cap rC)}{(r-r_0)^n}
= \rho(|P|).\end{aligned}$$ Moreover, each body in the finite set $D$ has nonempty interior, so there exists a constant $c = c(D)$ such that the number of congruent copies of elements from $D$ that fit in a cube of radius $r+r_0$ is at most $c r^n$. Hence, $$\lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \subseteq rC} \varepsilon \leq c\varepsilon \quad \text{and} \quad \lim_{r \to \infty} \frac{1}{r^n} \sum_{B \in P : B \cap rC \neq \emptyset} \varepsilon \leq c\varepsilon.$$ Hence, for all $k \geq \max_{B \in D} K(B,\varepsilon)$ we have $$\rho(|G_k(|P|)|) \geq \rho(|P|) - c\varepsilon \quad \text{and} \quad \rho(|G^k(|P|)|) \leq \rho(|P|) + c\varepsilon,$$ which implies $$\rho(|G_k(|P|)|) \uparrow \rho(|P|) \quad \text{and} \quad \rho(|G^k(|P|)|) \downarrow \rho(|P|),$$ and hence $$\rho(|G_k(P^c)|) = 1 - \rho(|G^k(|P|)|) \uparrow 1 - \rho(|P|)$$ and $$\rho(|G^k(P^c)|) = 1 - \rho(|G_k(|P|)|) \downarrow 1 - \rho(|P|),$$ as $k \to \infty$ for $P \in \Lambda_D$ uniformly.
Polydisperse packings
=====================
Let $D$ and $D'$ be sets of bodies. Given $P \in \Lambda_D$, define $$\Lambda_{D'}(P) = \Big\{ Q \in \Lambda_{D \cup D'} : Q = P \cup R, \, R \in \Lambda_{D'}\Big\}.$$ The optimal density of such packings is given by $\Delta_{D'}(P) = \sup_{Q \in \Lambda_{D'}(P)} \rho(|Q|)$. In the following lemmas we give the optimal density given that a part of the packing is already fixed.
\[lem:smalldiameter\] Suppose $D$ is a finite set of bodies. For every $\varepsilon > 0$ there is a scalar $R = R(D, \varepsilon) > 0$ such that $$\rho(|P|) + (1-\rho(|P|)) \Delta_{\{B\}} - \varepsilon \leq \Delta_{\{B\}}(P) \leq \rho(|P|) + (1-\rho(|P|)) \Delta_{\{B\}} + \varepsilon$$ for all $P \in \Lambda_D$ and all bodies $B$ with $\mathrm{diam}(B) \leq R$.
Let $0 < \varepsilon \leq 1$ and $P \in \Lambda_D$.
By Lemma \[lem:cube approx\] there exists an integer $K_1 = K_1(D, \varepsilon)$ such that $$\label{eq:v1}
\rho(|G_k(P^c)|) \geq 1 - \rho(|P|) - \varepsilon/2 \quad \text{for all} \quad k \geq K_1.$$ By the definition of density there exists a scalar $R_1 = R_1(k, \varepsilon) > 0$ such that for each body $B$ with $\mathrm{diam}(B) \leq R_1$ we can pack each cube in $G_k(P^c)$ with congruent copies of $B$ with density at least $\Delta_{\{B\}} - \varepsilon/2$. By taking the union of the packings of the cubes together with $\{P\}$ we obtain a packing from $\Delta_{\{B\}}(P)$ which has density at least $$\begin{aligned}
\rho(|P|) + \rho(|G_k(P^c)|) (\Delta_{\{B\}} - \varepsilon/2) &\geq \rho(|P|) + (1 - \rho(|P|) - \varepsilon/2)(\Delta_{\{B\}} - \varepsilon/2)\\
&\geq \rho(|P|) + (1-\rho(|P|)) \Delta_{\{B\}} - \varepsilon.\end{aligned}$$ This implies $$\Delta_{\{B\}}(P) \geq \rho(|P|) + (1-\rho(|P|)) \Delta_{\{B\}} - \varepsilon$$ for all $P \in \Lambda_D$ and all bodies $B$ with $\mathrm{diam}(B) \leq R_1$.
By Lemma \[lem:cube approx\] there exists an integer $K_2 = K_2(D, \varepsilon)$ such that $$\rho(|G^k(P^c)|) \leq 1 - \rho(|P|) + \varepsilon/3 \quad \text{for all} \quad k \geq K_2.$$ Again, by the definition of density there exists a scalar $R_2 = R_2(k, \varepsilon)$ such that for each body $B$ with $\mathrm{diam}(B) \leq r$ and each $Q \in \Lambda_{\{B\}}(P)$, the intersection of $|Q \setminus P|$ with a cube from $G^k(P^c)$ has density at most $\Delta_{\{B\}} + \varepsilon/3$ in that cube. So, $$\begin{aligned}
\rho(|Q|) & = \rho(|P|) + \rho(|Q \setminus P|) \leq \rho(|P|) + \rho(|G^k(P^c)|)(\Delta_{\{B\}} + \varepsilon/3)\\
&\leq \rho(|P|) + (1 - \rho(|P|) + \varepsilon/3)(\Delta_{\{B\}} + \varepsilon/3)\\
&\leq \rho(|P|) + (1 - \rho(|P|))\Delta_{\{B\}} + \varepsilon,\end{aligned}$$ hence $$\Delta_{\{rB\}}(P) \leq \rho(|P|) + (1-\rho(|P|)) \Delta_{\{B\}} + \varepsilon$$ for all $P \in \Lambda_D$ and all bodies $B$ with $\mathrm{diam}(B) \leq R_2$.
The proof is then complete by setting $R = \min \{R_1, R_2\}$.
Using the above result the following lemma is immediate.
\[lem:addballs\] Suppose $B$ is a body and $D$ is a finite set of bodies. Then, $$\lim_{r \downarrow 0} \Delta_{\{B_k\}}(P) = \rho(|P|) + (1 - \rho(|P|)) \Delta_{\{B\}} \quad \text{for} \quad P \in \Lambda_D \quad \text{uniformly}.$$
We prove the main result by using the uniform convergence in the above lemma.
\[thm:polydisperse:main theorem\] Suppose $B$ is a body and $D$ is a finite set of bodies. Then, $$\lim_{r \downarrow 0} \Delta_{\{rB\} \cup D}
= \Delta_D + (1 - \Delta_D) \Delta_{\{B\}}.$$
We have $\Lambda_{\{rB\} \cup D} = \bigcup_{Q \in \Lambda_D} \Lambda_{\{rB\}}(Q)$, so $$\Delta_{\{rB\} \cup D} = \sup_{P \in \Lambda_{\{rB\} \cup D}} \rho(|P|)
= \sup_{Q \in \Lambda_D} \; \sup_{P \in \Lambda_{\{rB\}}(Q)} \rho(|P|)
= \sup_{Q \in \Lambda_D} \Delta_{\{rB\}}(Q),$$ and $$\lim_{r \downarrow 0} \Delta_{\{rB\} \cup D} = \lim_{r \downarrow 0} \sup_{Q \in \Lambda_D} \Delta_{\{rB\}}(Q).$$ By Lemma \[lem:addballs\], we have $$\lim_{r \downarrow 0} \Delta_{\{rB\}}(Q) = \rho(|Q|) + (1 - \rho(|Q|)) \Delta_{\{B\}},$$ and since convergence is uniform for $Q \in \Lambda_D$, we can interchange limit and supremum and obtain $$\begin{aligned}
\lim_{r \downarrow 0} \sup_{Q \in \Lambda_D} \Delta_{\{rB\}}(Q)
&= \sup_{Q \in \Lambda_D} \lim_{r \downarrow 0} \Delta_{\{rB\}}(Q)
= \sup_{Q \in \Lambda_D} (\rho(|Q|) + (1 - \rho(|Q|)) \Delta_{\{B\}})\\
&= \Delta_{\{B\}} + (1 - \Delta_{\{B\}}) \sup_{Q \in \Lambda_D} \rho(|Q|)
= \Delta_{\{B\}} + (1 - \Delta_{\{B\}}) \Delta_D,\end{aligned}$$ and since $\Delta_{\{B\}} + (1 - \Delta_{\{B\}}) \Delta_D = \Delta_D + (1 - \Delta_D) \Delta_{\{B\}}$ this completes the proof.
As a special case of the above theorem we obtain the result mentioned in the abstract of this note: We have $\lim_{r \downarrow 0} \Delta_{\{B, r B\}} = \Delta + (1 - \Delta) \Delta$, where $B$ is the closed unit ball and where $\Delta = \Delta_{\{B\}}$. By iteratively applying the above theorem and rewriting the resulting expression we see that if $B_1, \ldots, B_k$ are bodies, then $$\begin{aligned}
\lim_{r_k \downarrow 0} \cdots \lim_{r_1 \downarrow 0} \Delta_{\{r_1B_1, \ldots, r_kB_k\}}
&= \Delta_{\{B_1\}} + (1-\Delta_{\{B_1\}}) \lim_{r_k \downarrow 0} \cdots \lim_{r_2 \downarrow 0} \Delta_{\{r_2B_2, \ldots, r_kB_k\}} \\
&= 1 -(1 - \Delta_{\{B_1\}}) \cdots (1 - \Delta_{\{B_k\}}).\end{aligned}$$ Moreover, if $B$ is a body and $D = \{ r B: r > 0\}$, then for every $k$ there is an $r \in {\mathbb{R}}^k$ such that $$\Delta_D \geq 1 - (1-\Delta_{\{B\}})^k - \frac{1}{k},$$ so we get the intuitive result $\Delta_D = 1$. The following proposition gives a strengthening of this result. Note that the condition $\inf_{B \in D} \Delta_{\{B\}} > 0$ here is always satisfied if we restrict to convex bodies.
Suppose $D$ is a set containing bodies of arbitrarily small diameter such that $s = \inf_{B \in D} \Delta_{\{B\}} > 0$. Then there exist packings consisting of congruent copies of bodies from $D$ whose density is arbitrarily close to $1$; that is, $\Delta_D = 1$.
Let $0 < \varepsilon < s$. Select an arbitrary $B_1 \in D$ and choose $P_1 \in \Lambda_{\{B_1\}}$ such that $\rho(|P_1|) \geq \Delta_{\{B_1\}} - \varepsilon$. By Lemma \[lem:smalldiameter\] there exists a body $B_2 \in D$ and a packing $P_2 \in \Lambda_{\{B_2\}}(P_1)$ such that $$\begin{aligned}
\rho(|P_2|)
&\geq \rho(|P_1|) + (1-\rho(|P_1|)) \Delta_{\{B_2\}} - \varepsilon\\
&= \Delta_{\{B_2\}} + (1 - \Delta_{\{B_2\}}) \rho(|P_1|) -\varepsilon\\
&\geq \Delta_{\{B_1\}} + \Delta_{\{B_2\}} - \Delta_{\{B_1\}} \Delta_{\{B_2\}} - 2\varepsilon\\
&\geq 1 - (1 - \Delta_{\{B_1\}} + \varepsilon)(1 - \Delta_{\{B_2\}} + \varepsilon) \geq 1 - (1 - (s - \varepsilon))^2.\end{aligned}$$ By continuing like this we see that for each $k$ there exists a packing $P_k \in \Lambda_D$ such that $$\rho(|P_k|) \geq 1 - (1 - (s - \varepsilon))^k,$$ which completes the proof.
[^1]: The author was supported by Vidi grant 639.032.917 from the Netherlands Organization for Scientific Research (NWO)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Motivated by some results in classical differential geometry, we give a constructive procedure for building up a connection over a (twisted) tensor product of two algebras, starting from connections defined on the factors. The curvature for the product connection is explicitly calculated, and shown to be independent of the choice of the twisting map and the module twisting map used to define the product connection. As a consequence, we obtain that a product of two flat connections is again a flat connection. We show that our constructions also behaves well with respect to bimodule structures, namely being the product of two bimodule connections again a bimodule connection. As an application of our theory, all the product connections on the quantum plane are computed.'
address: |
Department of Algebra, University of Granada\
Avda. Fuentenueva s/n, E-18071, Granada, Spain
author:
- Javier López Peña
nocite:
- '[@Landi97a]'
- '[@Gracia01a]'
- '[@Mourad95a]'
- '[@BeggsUNa]'
- '[@Jara03a]'
- '[@DuboisViolette95a]'
- '[@Connes86a]'
- '[@LopezUNa]'
- '[@Beck69a]'
title: Connections over twisted tensor products of algebras
---
Introduction {#introduction .unnumbered}
============
One of the main tools in classical differential geometry is the use of the tangent bundle associated to a manifold. The role of the algebra of functions on the manifold is taken by the sections of the tangent bundle, namely, the vector fields. As a dual of the vector fields space, the algebra of differential forms (endowed with the exterior product) turns out to be an useful tool in the study of global properties of the manifold, giving rise to invariants such as the de Rham cohomology. A problem arises when trying to compare vector fields and differential forms at different points of the manifold, the solution to it being given by the concepts of [***(linear) connection***]{} and [***covariant derivative***]{}, that allow us to define the derivative of a curve on a point of orders higher than one, hence giving us a way to speak about accelerations on a path. The notion of connection also has another meanings in physics, like the existence of an electromagnetic potential, which is equivalent to the existence of a connection in a rank one trivial bundle with fixed trivialization.
Jean–Louis Koszul gave in [@Koszul60a] a powerful algebraic generalization of differential geometry, in particular giving a completely algebraic description of the notion of connection. These notions were extended to a noncommutative framework by Alain Connes in [@Connes86a], what meant the dawn of noncommutative differential geometry. Much research has been done about the theory of connections in this context. On the one hand, Joachim Cuntz and Daniel Quillen, in their seminal paper [@Cuntz95a] started the theory of [***quasi-free***]{} algebras (also named [***formally smooth***]{} by Maxim Kontsevich or [***qurves***]{} by Lieven Le Bruyn), opening the way to an approach to noncommutative (algebraic) geometry (also dubbed [***nongeometry***]{} to avoid confusions with Michael Artin and Michel Van den Bergh’s style of noncommutative algebraic geometry). These formally smooth algebras are characterized by the projectiveness (as a bimodule) of the first order universal differential calculus, or equivalently as those algebras that admit a universal linear connection. On the other hand, in Connes’ style of noncommutative geometry, the study of the general theory of connections leads to the definition of the Yang–Mills action, which turns out to be nothing but the usual gauge action when we specialize it to the commutative case (cf. [@Connes86a], [@Landi97a], [@Gracia01a] and references therein).
In this paper, we deal with the problem of building up products of those connection operators. Basically, there are two different notions of “*product connection*” that one might want to build. Firstly, one might want to consider two different bundles over a manifold, each of them endowed with a connection, and then try to build a product connection on the (fibre) product bundle. A noncommutative version of this construction was given by Michel Dubois–Violette and John Madore in [@DuboisViolette99a], [@Madore95a]. Further steps on this direction, including its relations with the realization of vector fields as Cartan pairs as proposed by Andrzej Borowiec in [@Borowiec96a], have been given by Edwin Beggs in [@BeggsUNa]. The other possible notion of product connection, and the one with which we want to deal, refers to the consideration of the cartesian product of two given manifolds, and the building of a connection of the bundle associated to this product manifold.
Traditionally, when taking the passage from classical geometry to (noncommutative) algebra, the product space is associated with some kind of tensor product (the algebraic tensor product in the case of algebraic varieties, the topological tensor one when dealing with topological manifolds). In [@Cap95a], Andreas Cap, Herman Schichl and Jiří Vanžura pointed out the limitations of these approach and proposed a definition of “noncommutative cartesian product” of spaces by means of the so–called [***twisted tensor product***]{} of the algebras. A twisted tensor product is a particular case of the notion of [***distributive law***]{} given by Jon Beck in [@Beck69a], and may be regarded as a sort of local version of a braiding in a braided monoidal category. Some further insights on the interpretation of this algebraic construction from a geometrical point of view has been given by Alfons Van Daele and S. Van Keer in [@VanDaele94a], Stanislaw Woronowicz in [@Woronowicz96a], and the author and some collaborators in [@JaraUNa] (cf. also [@LopezUNa] for an interpretation of twisted tensor products from a deformation theory point of view). Following these ideas, we will show how to build a product connection on a twisted tensor product of two algebras.
In Section \[preliminaries\] we recall the notion of a (right) connection on an algebra, given as an operator $\nabla:E\to E\ot_A \O^1 A$, where $E$ is a (right) $A$–module and $\O^1 A$ a first order differential calculus over $A$, motivating our choice of the differential calculus and the modules on the building of the product connection, for which an explicit formula is given in the simplest case of the usual tensor product. Then, we recall the main notions we need about twisted tensor products, defined by means of twisting maps $R:B\ot A\to A\ot B$.
In Section \[construction\] we give the basic definition of our object of study, and prove that our definition actually yields a connection in the product space, and that this connection boils down nicely to the classical product connection in the commutative case.
To a connection we can always associate its [***curvature operator***]{}, obtained by squaring the extension of the connection to the whole differential calculus. The curvature operator leads to the definition of [***flat connections***]{} as those having 0 curvature. Flat connections have been used by Philippe Nuss in [@Nuss97a] in relation with noncommutative descent theory, and also by Edwin Beggs and Tomasz Brzezinski in [@Beggs05a], where they are interpreted as the differential of a certain complex in order to build a noncommutative de Rham cohomology with coefficients. We deal with the problem of describing the curvature of our product connection in Section \[curvature\], stating our main theorem, that gives us an explicit formula to compute the curvature for the product connection in terms of the curvatures of the factors: $$\theta (e\ot b, a\ot f) = i_E(\theta^E(e))\cd b + a\cd i_F(\theta^F(f)).$$ The most striking consequence of this theorem is the fact that the curvature does not depend neither on the twisting map $R$ nor on the module twisting map that we use to get the module structure, suggesting that the curvature remains invariant under all the deformations obtained by means of a twisted tensor product. As an immediate corollary, we have that the product of two flat connections is again a flat connection.
In Section \[bimodules\] we consider bimodule connections (in the sense introduced by Jihad Mourad in [@Mourad95a]) instead of one sided connections, and we find necessary and sufficient conditions for the product of two bimodule connections to be a bimodule connection. We conclude, in Section \[examples\], by illustrating our theory giving a complete description of all the product connections on the quantum plane $k_q[x,y]$.
Preliminaries
=============
Connections on algebras
-----------------------
Let $A$ be an associative, unital algebra over a field $k$, and $\O A =\bigoplus_{p\geq 0}\O^p A$ a differential calculus over $A$, that is, a differential graded algebra generated, as a differential graded algebra, by $\O^0 A\cong
A$, with differential $d=d_A$. Let $E$ be a (right) $A$–module; a [***(right) connection***]{} on $E$ is a linear mapping $$\nabla: E \lto E\otimes_{A}\O^1 A$$ satisfying the (right) [***Leibniz rule***]{}: $$\nabla(s\cdot a)=(\nabla s)\cdot a + s\otimes da\quad \forall\,s\in
E, a\in A.$$ Under these conditions, the mapping $\nabla$ can be extended in a unique way to an operator $$\nabla: E\otimes_A \O A\lto E\otimes_A\O A$$ of degree 1, by setting $$\label{connectionextension}
\nabla(s\otimes \omega)=\nabla{s}\otimes \omega + (-1)^{p}s\otimes
d\omega\quad\forall\,s\in E,\omega\in\O^p A,$$ where we are using the identification $(E\ot_A \O^1 A)\ot_A \O^n
A\cong E\ot_A \O^{n+1}A$. Regarding $E\ot_A \O A$ as a right $\O A$–module, we find that the following graded Leibniz rule is satisfied: $$\nabla(\sigma\omega)=(\nabla\sigma)\omega + (-1)^p\sigma
d\omega\quad\forall\,\sigma\in E\ot_A \O^p A, \omega\in \O A.$$ There are analogous concepts for left modules.
Usually, we will be interested on working with the universal differential calculus over an algebra $A$. Connections over the universal differential calculus will be called [***universal connections***]{}. It is a well known fact (cf. [@Cuntz95a Corollary 8.2]) that a right $A$–module admits a universal connection if, and only if, it is projective over $A$.
Whenever $A$ is a commutative algebra, the tensor product $E\ot_A F$ of two $A$–modules $E$ and $F$ is again an $A$–module. If $E$ and $F$ carry respective connections $\nabla^E$ and $\nabla^F$, we may build the [***tensor product connection***]{} on $E\ot_A F$ by defining $$\nabla^{E\ot_A F}:=\nabla^E\ot F + E\ot \nabla^F.$$ A possible generalization of this construction was given by Dubois–Violette and Madore in [@DuboisViolette99a], [@Madore95a]. If $E$ and $F$ are $A$–bimodules equipped with right connections $\nabla^E$ and $\nabla^F$, and such that there exists a linear mapping $$\sigma: \O^1 A\ot_A F \lto F\ot_A \O^1 A$$ satisfying that $$\nabla^F(am)=a\nabla^F(m) + \sigma(da\ot_A m)\quad\forall\,a\in
A,m\in F,$$ then we may define $$\nabla^{E\ot_A F}:E\ot_A F \lto E\ot_A F \ot \O^1 A$$ by setting $$\nabla^{E\ot_A F}:= (E\ot \sigma)\circ(\nabla^E\ot F) + E\ot
\nabla^F,$$ and this $\nabla^{E\ot_A F}$ is a right connection on $E\ot_A F$.
Our aim is to define a different kind of “*product connection*” with a more geometrical flavour. Namely, consider that our algebras $A=C^\infty(M)$ and $B=C^\infty(N)$ represent the algebras of functions over certain manifolds $M$ and $N$, and that $E=\cam(M)$ and $F=\cam(N)$ are the modules of vector fields on the manifolds. The algebra associated to the cartesian product of the manifolds is $C^\infty(M\times N)\cong C^\infty(M)\ot C^\infty (N)$ (more precisely, a suitable completion of the latest). For the modules of vector fields and differential $1$–forms, we have that $$\begin{gathered}
\cam(M\times N)\cong \cam(M)\ot C^\infty(N) \oplus C^\infty(M)\ot
\cam(N),\\
\O^1 (C^\infty(M)\ot C^\infty(N))\cong \O^1(C^\infty(M))\ot
C^\infty(N) \oplus C^\infty(M)\ot \O^1(C^\infty(N)),\end{gathered}$$ hence, a “*product connection*” of two connections defined on $E$ and $F$ should be defined as a linear mapping $$\nabla : E\ot B \oplus A\ot F \lto (E\ot B \oplus A\ot
F)\ot_{A\ot B} (\O^1 A\ot B \oplus A \ot \O^1B)$$
Firstly, realize that if $E$ is a right (resp. left) $A$–module, and $F$ is a right (resp. left) $B$–module, then $E\ot B\oplus A\ot
F$ is a right $(A\ot B)$–module, with actions $$\begin{gathered}
(e\ot b,a\ot f)\cd(\alpha\ot \beta):=(e\alpha\ot b\beta,a\alpha\ot
f\beta)\\
\text{(resp.}\quad (\alpha\ot \beta)\cd (e\ot b,a\ot f):=(\alpha
e\ot \beta b,\alpha a\ot \beta f)\ \text{)}\end{gathered}$$
For simplicity, we will only work with right connections. Left connections admit a similar treatment.
Product Connection
------------------
Suppose then that $E$ is a right $A$–module endowed with a (right) connection $\nabla^E$, and that $F$ is a right $B$–module endowed with a (right) connection $\nabla^F$. Let us consider the mappings $$\begin{aligned}
\nabla_1:E\ot B & \lto & (E\ot B\op A\ot F)\ot_{A\ot B}(\O^1 A\ot
B\op A\ot \O^1 B), \\
\nabla_2:A\ot F & \lto & (E\ot B\op A\ot F)\ot_{A\ot B}(\O^1 A\ot
B\op A\ot \O^1 B)\end{aligned}$$ respectively given by $$\begin{aligned}
\nabla_1 &:=& (E\ot \tau \ot u_B)\circ (\nabla^E\ot B) + (E\ot
u_A\ot u_B\ot \O^1 B)\circ (E\ot d_B),\ \text{and}\\
\nabla_2 &:=& (A\ot F\ot u_A\ot \O^1B)\circ (A\ot \nabla^F) +
(u_A\ot \tau\ot u_B)\circ (d_A\ot F),\end{aligned}$$ where $\tau$ represent classical flips. If we use the shorthand notation $\nabla^E(e)=e_i\ot d_A a_i$, where the summation symbol is omitted, the Leibniz rule for $\nabla^E$ is written as $$\label{Leibiz:E}
\nabla^E(e\alpha)=e_i\ot (d_A a_i)\alpha + e\ot d\alpha,$$ and we have that $$\begin{aligned}
\nabla_1((e\ot b)\cd(\alpha\ot \beta)) &=& \nabla_1(e\alpha\ot
b\beta) = \\
&=& e_i\ot b\beta \ot_{A\ot B} (d_A a_i)\alpha \ot 1 + e\ot b\beta
\ot_{A\ot B} d\alpha \ot 1 + \\
& & +\ e\alpha \ot 1 \ot_{A\ot B} 1 \ot d_B(b\beta)= \\
&=& e_i\ot b \ot_{A\ot B} (d_A a_i)\alpha \ot \beta + e\ot b
\ot_{A\ot B} d\alpha \ot \beta + \\
& & +\ e \ot 1 \ot_{A\ot B} \alpha \ot d_B(b)\beta + e \ot 1 \ot_{A\ot B} \alpha \ot b d_B \beta = \\
&=& \left(e_i\ot b \ot_{A\ot B} da_i\ot 1 + e\ot 1\ot_{A\ot B} 1\ot
db)\cd (\alpha\ot \beta \right) + \\
&& +\ e\ot b \ot_{A\ot B} d_A\alpha\ot \beta + e\ot b \ot_{A\ot B} \alpha\ot
d_B\beta = \\
&=& \nabla_1(e\ot b)\cd (\alpha\ot \beta) + (e\ot b) \ot_{A\ot
B} d(\alpha\ot \beta).\end{aligned}$$ A similar computation shows that $$\begin{aligned}
\nabla_2((a\ot f)\cd(\alpha\ot \beta)) &=& \nabla_2(a\ot f)\cd (\alpha\ot \beta) + (a\ot f) \ot_{A\ot
B} d(\alpha\ot \beta).\end{aligned}$$ Adding up these two equalities, we conclude that the map $$\begin{aligned}
\nabla : E\ot B \oplus A\ot F & \lto & (E\ot B \oplus A\ot
F)\ot_{A\ot B} (\O^1 A\ot B \oplus A \ot \O^1B)\\
(e\ot b,a\ot f) & \lmto & \nabla_1(e\ot b) + \nabla_2(a\ot f)\end{aligned}$$ verifies that $$\nabla((e\ot b,a\ot f)\cd (\alpha\ot \beta))= \nabla(e\ot b,a\ot f)\cd (\alpha\ot
\beta) + (e\ot b,a\ot f) \ot_{A\ot B} d(\alpha \ot \beta),$$ and henceforth, $\nabla$ is a (right) connection on the module $E\ot
B\op A\ot F$. We shall call this map the [***(classical) product connection***]{} of $\nabla^E$ and $\nabla^F$.
Twisted tensor products
-----------------------
Let $k$ be a field, used as a base field throughout. We denote $\otimes _k$ by $\otimes $, the identity $id_V$ of an object $V$ simply by $V$, and by $\tau :V\otimes W \rightarrow W\otimes V$, $\tau (v\otimes w)=w\otimes v$, the usual flip. All algebras are assumed to be associative unital $k$-algebras; the multiplication and unit of an algebra $D$ are denoted by $\mu _D:D\otimes
D\rightarrow D$ and respectively $u_D:k\rightarrow D$ (or simply by $\mu $ and $u$ if there is no danger of confusion).
We recall the twisted tensor product of algebras from [@Tambara90a], [@VanDaele94a], [@Cap95a]. If $A$ and $B$ are two algebras, a linear map $R:B\ot A\rightarrow A\ot B$ is called a [***twisting map***]{} if it satisfies the conditions $$\begin{aligned}
&&R(b\ot 1)=1\ot b,\;\;\;R(1\ot a)=a\ot 1, \;\;\;\forall \; a\in A,
\;b\in B, \label{tw1}\\
&&R\circ (B\ot \mu _A)=(\mu _A\ot B)\circ (A\ot R)\circ (R\ot A),
\label{tw2}\\
&&R\circ (\mu _B\ot A)=(A\ot \mu _B)\circ (R\ot B)\circ (B\ot R).
\label{tw3}\end{aligned}$$ If we denote by $R(b\ot a)=a_R\ot b_R$, for $a\in A$, $b\in B$, then (\[tw2\]) and (\[tw3\]) may be written as: $$\begin{aligned}
&&(aa')_R\ot b_R=a_Ra'_r\ot (b_R)_r, \label{tw4} \\
&&a_R\ot (bb')_R=(a_R)_r\ot b_rb'_R, \label{tw5}\end{aligned}$$ for all $a, a'\in A$ and $b, b'\in B$, where $r$ is another copy of $R$. If we define a multiplication on $A\ot B$, by $\mu _R=(\mu _A\otimes
\mu _B)\circ (A\otimes R\otimes B)$, that is $$\begin{aligned}
&&(a\ot b)(a'\ot b')=aa'_R\ot b_Rb', \label{multtw}\end{aligned}$$ then this multiplication is associative and $1\ot 1$ is the unit. This algebra structure is denoted by $A\ot _RB$ and is called the [***twisted tensor product***]{} of $A$ and $B$. This construction works also if $A$ and $B$ are algebras in an arbitrary monoidal category.
If $A\otimes _{R_1} B$, $B\otimes _{R_2} C$ and $A\otimes
_{R_3} C$ are twisted tensor products of algebras, the twisting maps $R_1$, $R_2$, $R_3$ are called [*compatible*]{} if they satisfy $$\begin{aligned}
&&(A\otimes R_2)\circ (R_3\otimes B)\circ (C\otimes R_1)=
(R_1\otimes C)\circ (B\otimes R_3)\circ (R_2\otimes A),\end{aligned}$$ see [@JaraUNa]. If this is the case, the maps $T_1:C\otimes
(A\otimes _{R_1} B)\rightarrow (A\otimes _{R_1} B)\otimes C$ and $T_2:(B\otimes _{R_2} C)\otimes A\rightarrow A\otimes (B\otimes
_{R_2} C)$ given by $T_1:=(A\otimes R_2)\circ (R_3\otimes B)$ and $T_2:=(R_1\otimes C)\circ (B\otimes R_3)$ are also twisting maps and $A\otimes _{T_2} (B\otimes _{R_2} C)\equiv (A\otimes _{R_1} B)
\otimes _{T_1} C$; this algebra is denoted by $A\otimes _{R_1} B
\otimes _{R_2} C$. This construction may be iterated to an arbitrary number of factors, see [@JaraUNa] for complete detail.
When we have a left $A$–module $M$, a left $B$–module $N$, a twisting map $R:B\otimes A\to A\otimes B$ and a linear map $\tau_{M,B}:B\otimes M \to M\otimes B$ such that $$\begin{aligned}
\label{moduletwistcondition1}
\tau_{M,B}\circ(\mu_B\otimes M) & = & (M\otimes \mu_B)\circ (\tau_{M,B}\otimes B)\circ (B\otimes
\tau_{M,B}), \\ \label{moduletwistcondition2}
\tau_{M,B}\circ(B\otimes \l_M) & = & (\l_M \otimes B)\circ (A\otimes \tau_{M,B})\circ (R\otimes
M),\end{aligned}$$ then the map $\l_{\tau_{M,B}}:(A\otimes_R B)\otimes (M\otimes N)\to
M\otimes N$ defined by $\l_{\tau_{M,B}}:=(\l_M\otimes \l_N)\circ
(A\otimes \tau_{M,B}\otimes N)$ yields a left $(A\otimes_R
B)$–module structure on $M\otimes N$, which furthermore is compatible with the inclusion of $A$. In this case, we say that $\tau_{M,B}$ is a [***(left) module twisting map***]{}. Unlike what happens for algebra twisting maps, usually is not enough to have a left $(A\otimes_R B)$–module structure on $M\otimes N$ in order to recover a module twisting map. Some sufficient conditions for this to happen are, for instance, requiring that $M$ is projective and $N$ is faithful (cf. [@Cap95a Theorem 3.8]).
Similarly, if we have a twisting map $R:B\otimes A\to A\otimes B$, a right $A$–module $M$ and a right $B$–module $N$, a linear map $\tau_{N,A}:N\otimes A\to A\otimes N$ such that $$\begin{aligned}
\label{rightmoduletwist1}
\tau_{N,A}\circ (N\otimes \mu_A) &=&(\mu_A\otimes N)\circ (A\otimes
\tau_{N,A})\circ (\tau_{N,A}\otimes A)\\
\label{rightmoduletwist2}
\tau_{N,A}\circ (\rho_B\otimes A) &=& (A\otimes \rho_B)\circ
(\tau_{N,A}\otimes B)\circ (N\otimes R).\end{aligned}$$ then the map $\rho_{\tau_{N,A}}:=(\rho_A\otimes \rho_B)\circ (M\otimes \tau_{N,A}\otimes B)$, yields a right $(A\ot_R B$–module action on $M\ot N$. In this case, we call $\tau_{N,A}$ a [***(right) module twisting map***]{}
Twisting maps also have a nice behaviour with respect to (universal) differential calculi. More concretely, we have the following result (cf. [@Cap95a]):
\[twistdifferentialforms\] Let $A$, $B$ be two algebras. Then any twisting map $R:B\otimes A\to
A\otimes B$ extends to a unique twisting map $\tilde{R}:\O B\otimes
\O A \to \O A\otimes \O B$ which satisfies the conditions $$\begin{aligned}
\label{twisteddiff1}
\tilde{R}\circ(d_B\otimes \O A)&=&(\eps_A \otimes d_B)\circ
\tilde{R}, \\
\label{twisteddiff2} \tilde{R}\circ (\O B\otimes d_A)&=&(d_A \otimes
\eps_B)\circ \tilde{R},\end{aligned}$$ where $d_A$ and $d_B$ denote the differentials on the algebras of universal differential forms $\O A$ and $\O B$, and $\eps_A$, $\eps_B$ stand for the gradings on $\O A$ and $\O B$, respectively. Moreover, $\O A\otimes_{\tilde{R}}\O B$ is a graded differential algebra with differential $d(\vphi\otimes
\omega):=d_A\vphi\otimes \omega + (-1)^{{\left\lvert\vphi\right\rvert}}\vphi\otimes
d_B\omega$.
Twisted tensor product connection {#construction}
=================================
In the former section we introduced the definition of a connection within the formalism of differential calculus over algebras, and showed how to build the product connection for a tensor product of two algebras, extending the definition of the classical product connection in differential geometry. In [@JaraUNa], we advocated that in noncommutative geometry the cartesian product should not be replaced at the algebraic level by the usual tensor product of algebras, but by a deformation of it, known as the twisted tensor product. In this section, we will show how to extend the definition of the product connection to a twisted tensor product of two algebras under suitable conditions.
Let $A$ and $B$ be algebras, $R:B\ot A\to A\ot B$ a twisting map, $E$ a right $A$–module endowed with a right connection $\nabla^E$, and $F$ a right $B$–module endowed with a right connection $\nabla^F$. In [@Cap95a] it is shown that we can lift the twisting map $R$ to a twisting map $\widetilde{R}:\O B\ot \O A\to \O
A\ot \O B$ on the graded differential algebras of (universal) differential forms, and that the algebra $$\O A\ot_{\widetilde{R}} \O B = \bigoplus_{n\in \n}\left(\bigoplus_{p+q=n}\O^p A\ot \O^q B\right)$$ is a differential calculus over $A\ot_{R} B$. For this differential calculus, the module of $1$–forms can be identified as $\O^1 A \ot
B \oplus A\ot \O^1 B$, with the natural action induced by the twisting map. As the situation is pretty much the same as in the tensor product case, the natural way for defining a “*twisted product*” connection of $\nabla^E$ and $\nabla^F$ would be considering a linear map $$\nabla: E\ot B \op A\ot F \lto \left(E\ot B \op A\ot
F\right)\ot_{A\ot_{\widetilde{R}}B} \left(\O^1 A\ot B \op A\ot \O^1
B\right).$$ The first step on making this map becoming a connection is giving a right $(A\ot_R B)$–module action on $E\ot B \op A\ot F$, which means finding a right $(A\ot_R B)$–module structure on both $E\ot
B$ and $A\ot F$. For the first one we may just use the twisting map and define: $$\label{eq:twistactionEB}
(e\ot b)\cd (\alpha\ot \beta):= e\alpha_R \ot b_R\beta.$$ For the second one, a sufficient way of giving a module structure is finding a (right) module twisting map $\tau_{F,A}:F\ot A\to A\ot F$, and then taking $$\label{eq:twistactionAF}
(a\ot f)\cd(\alpha\ot \beta):=a\alpha_\tau\ot f_\tau\beta.$$ The fact that the former definitions are indeed module actions follows directly from the fact that both $R$ and $\tau_{F,A}$ are right module twisting maps (cf. [@Cap95a], 3.12).
Following the lines given by the definition of the classical tensor product connection, in order to build $\nabla$ we have to find suitable maps $\nabla_1$ and $\nabla_2$. For the first one, it suffices to define
$$\begin{gathered}
\nabla_1:E\ot B \lto (E\ot B\op A\ot F)\ot_{A\ot B}(\O^1 A\ot
B\op A\ot \O^1 B)\\
\nabla_1:=(E\ot u_B\ot \O^1 A\ot B)\circ (\nabla^E\ot B) + (E\ot
u_B\ot u_A\ot \O^1 B)\circ (E\ot d_B).\end{gathered}$$
With this definition, when $R$ is the classical flip is a definition trivially equivalent to the one given in the former section, and we have that
$$\begin{aligned}
\nabla_1((e\ot b)\cd(\alpha\ot \beta)) &=& \nabla_1(a\alpha_R\ot
b_R\beta) = \\
&=& (E\ot u_B\ot \O^1 A\ot B)(\nabla^E(e\alpha_R)\ot b_R\beta)+\\
&& +(E\ot u_B\ot u_A\ot \O^1 B)(e\alpha_R\ot d(b_R\beta))\overset{1}{=}\\
&\overset{1}{=}& e_i\ot 1\ot_{A\ot_R B} (d_Aa_i)\alpha_R\ot b_R\beta +\\
&& +\ e\ot 1 \ot_{A\ot_R B} d\alpha_R\ot b_R\beta
+ \\
&& +\ e\alpha_R\ot 1\ot_{A\ot_R B}1\ot (d_B b_R)\beta + \\
&& + e\alpha_R\ot 1\ot_{A\ot_R B}1\ot b_R
d_B\beta=\\
&=& e_i\ot 1\ot_{A\ot_R B} (d_Aa_i)\alpha_R\ot b_R\beta +\\
&& +\ e\ot 1 \ot_{A\ot_R B} d\alpha_R\ot b_R\beta
+ \\
&&+\ e\ot 1\ot_{A\ot_R B}\alpha_R\ot (d_B b_R)\beta + \\
&& + e\ot 1\ot_{A\ot_R B}\alpha_R\ot b_R
d_B\beta \overset{2}{=}\\
&\overset{2}{=}& (e_i\ot 1\ot_{A\ot_R B}d_Aa_i \ot b + \\
&& +\ e\ot 1\ot_{A\ot_R
B}1\ot
b)\cd (\alpha\ot \beta)+\\
&&+\ e\ot b \ot_{A\ot_R B} d_A\alpha\ot \beta +\\
&& +\ e\ot b \ot_{A\ot_R B}
\alpha\ot
d_B\beta =\\
&=& \nabla_1(e\ot b)\cd (\alpha\ot \beta) + e\ot b \ot_{A\ot_R
B} d(\alpha\ot\beta),\end{aligned}$$
where in 1 we are using Leibniz’s rules (for the connection $\nabla^E$ and the differential $d_B$), in $2$ the definition of the action and the compatibility of the twisting map with the differential, as mentioned in equations and .
The definition of $\nabla_2$ is more involved, and we are forced to assume some extra conditions on the maps $R$ and $\tau_{F,A}$. Namely, assume that $R$ is invertible, with inverse $S:A\ot B\to
B\ot A$, that $\tau_{F,A}$ is invertible with inverse $\sigma_{A,F}:A\ot F\to F\ot A$, and such that the following relation, ensuring the compatibility of the module twisting map with the connection $\nabla^F$, is satisfied: $$\begin{gathered}
\label{eq:nabla2cond1} (A\ot \nabla^F)\circ
\tau_{F,A}=(\tau_{F,A}\ot\O^1 B)\circ(F\ot
\widetilde{R})\circ(\nabla^F\ot A).\end{gathered}$$ From this condition, that in Sweedler’s like notation is written as $$a_\tau\ot (f_\tau)_j\ot_B (db_\tau)_j =
(a_{\widetilde{R}})_\tau\ot (f_j)_\tau\ot_B ((db_j)_{\widetilde{R}})_\tau,$$ the module twisting conditions and for $\tau_{F,A}$, and the twisting map conditions and for $R$, we may easily deduce the following equalities: $$\begin{gathered}
\label{eq:nabla2cond1bis} (\sigma_{A,F} \ot \O^1 B)\circ (A\ot\nabla^F) = (F\ot\widetilde{R})\circ(\nabla^F\ot A)\circ \sigma_{A,F},\\
\label{eq:nabla2cond2} (\mu_A\ot F)\circ(A\ot
\tau_{F,A})=\tau_{F,A}\circ(F\ot \mu_A)\circ (\sigma_{A,F}\ot A),\\
\label{eq:nabla2cond3} \sigma_{A,F}\circ(A\ot
\lambda_F)\circ(\tau_{F,A}\ot B) = (\lambda_F\ot A)\circ(F\ot S), \\
\label{eq:nabla2cond4} \sigma_{A,F}\circ (\mu_A\ot F) = (F\ot
\mu_A)\circ
(\sigma_{A,F}\ot A)\circ (A\ot \sigma_{A,F}),\\
\label{eq:nabla2cond5} \sigma_{A,F}\circ(B\ot \lambda_F)\circ(R\ot
F) = (\lambda_F\ot A)\circ (B\ot \sigma_{A,F})\end{gathered}$$ and define the map $$\begin{gathered}
\nabla_2: A\ot F \lto (E\ot B\op A\ot F)\ot_{A\ot B}(\O^1 A\ot
B\op A\ot \O^1 B) \\
\nabla_2:= (A\ot F\ot u_B\ot \O^1 B)\circ(A\ot \nabla^F) +
(u_A\ot F\ot d_A\ot u_B)\circ \sigma\end{gathered}$$ then we have that $$\begin{aligned}
\nabla_2((a\ot f)\cd(\alpha\ot\beta)) &=&
\nabla_2(a\alpha_\tau\ot f_\tau\beta)= \\
&=& (A\ot F\ot u_A\ot \O^1
B)(a\alpha_\tau\ot\nabla^F(f_\tau\beta)) + \\
&& + \ 1\ot (f_\tau\beta)_\sigma\ot
d_A((a\alpha_\tau)_\sigma)\ot 1 \overset{\eqref{eq:nabla2cond4}}{=}\\
&\overset{\eqref{eq:nabla2cond4}}{=}& a \alpha_\tau\ot
(f_\tau)_j\ot_{A\ot_R B} 1 \ot d_B(b_\tau)_j\beta + \\
&& +\ a\alpha_\tau\ot f_\tau\ot_{A\ot_R B} 1\ot d_B\beta + \\
&& +\ 1\ot (f_\tau\beta)_{\sigma\overline{\sigma}}\ot_{A\ot_R B}
(d_Aa_{\overline{\sigma}})\alpha_{\tau\sigma}\ot 1 + \\
&&+\ 1\ot (f_\tau\beta)_{\sigma\overline{\sigma}}\ot_{A\ot_R B}
a_{\overline{\sigma}}d_A(\alpha_{\tau\sigma})\ot 1
\overset{\eqref{eq:nabla2cond1}}{=}\\
&\overset{\eqref{eq:nabla2cond1}}{=}& a (\alpha_{\widetilde{R}})_{\tau}\ot
(f_j)_\tau \ot_{A\ot_R B} 1 \ot (d_Bb_j)_{\widetilde{R}}\beta + \\
&& +\ a \ot f\ot_{A\ot_R B} \alpha\ot d_B\beta + \\
&& +\ 1\ot (f_\tau\beta)_{\sigma\overline{\sigma}}\ot_{A\ot_R B}
(d_Aa_{\overline{\sigma}})\alpha_{\tau\sigma}\ot 1 + \\
&&+\ a \ot (f_\tau\beta)_{\sigma}\ot_{A\ot_R B}
d_A(\alpha_{\tau\sigma})\ot 1 \overset{\eqref{eq:nabla2cond3}}{=}\\
& \overset{\eqref{eq:nabla2cond3}}{=} & a \ot
f_j \ot_{A\ot_R B} \alpha_{\widetilde{R}} \ot (d_Bb_j)_{\widetilde{R}}\beta
+ \\
&& +\ a \ot f\ot_{A\ot_R B} \alpha\ot d_B\beta + \\
&& +\ 1\ot (f\beta_S)_{\sigma}\ot_{A\ot_R B}
(d_Aa_{\sigma})\alpha_{S}\ot 1 + \\
&&+\ a \ot f\beta_S \ot_{A\ot_R B}
d_A(\alpha_{S})\ot 1 \overset{\eqref{eq:nabla2cond2}}{=}\\
&\overset{\eqref{eq:nabla2cond2}}{=}& (a \ot
f_j \ot_{A\ot_R B} 1\ot (d_Bb_j))\cd(\alpha\ot\beta) + \\
&& +\ a \ot f\ot_{A\ot_R B} \alpha\ot d_B\beta + \\
&& +\ 1\ot f_\sigma\beta_{S\overline{S}}\ot_{A\ot_R B}
d_A(a_{\sigma\overline{S}})\alpha_{S}\ot 1 + \\
&& +\ a \ot f \ot_{A\ot_R B} d_A\alpha\ot \beta =\\
&=& (a \ot
f_j \ot_{A\ot_R B} 1\ot (d_Bb_j))\cd(\alpha\ot\beta) + \\
&& +\ 1\ot f_\sigma\ot_{A\ot_R B}
d_A(a_{\sigma})\alpha\ot \beta + \\
&& +\ a \ot f\ot_{A\ot_R B} \alpha\ot d_B\beta + \\
&& +\ a \ot f \ot_{A\ot_R B}
d_A\alpha\ot \beta =\\
&=& \nabla_2(a\ot f)\cd(\alpha\ot \beta) + a\ot f\ot_{A\ot_R
B}d(\alpha\ot \beta).\end{aligned}$$ Henceforth, the mapping $$\nabla: E\ot B \op A\ot F \lto \left(E\ot B \op A\ot F\right)\ot_{A\ot_{\widetilde{R}}B} \left(\O^1 A\ot B \op A\ot \O^1 B\right)$$ defined as $$\label{productconnection}
\nabla(e\ot b,a\ot f) := \nabla_1(e\ot b) + \nabla_2(a\ot f)$$ is a (right) connection on the module $E\ot
B\oplus A\ot F$. We will call this connection the [***(twisted) product connection of $\nabla^E$ and $\nabla^F$***]{}.
Curvature on product connections {#curvature}
================================
In this section our aim is to study the curvature for the formerly defined product connections. If we have a connection $\nabla:E \to E \ot_A \O^1 A$, we will also denote by $\nabla:E\ot_A \O A\to E\ot_A \O A$ the extension given by , ocassionally denoting by $\nabla^{[n]}:E\ot_A \O^n A\to E\ot_A \O^{n+1}A$ its restriction to $E$-valued $n$–form. The [***curvature***]{} of the connection $\nabla$ is defined to be the operator $\theta:=\nabla^{[1]}\circ \nabla^{[0]}:E\to E\ot_A\O^2 A$. It is well known (cf. for instance [@Landi97a Sect. 7.2]) that the map $\theta$ is right $A$–linear. A connection $\nabla$ is said to be a [***flat connection***]{} whenever the associated curvature map is equal to 0. As curvature map may be extended to a (right) $\O A$–linear map $\theta:E\ot_A \O A \to E\ot_A \O A$ of degree 2 given at degree $n$ by $\theta^{[n]}:=\nabla^{[n+1]}\circ \nabla^{[n]}$, and it is easily checked that $\theta^{[n]}= (E\ot\mu_{\O A}) \circ (\O^n A\ot \theta)$, (cf. [@Beggs05a Prop 2.3]), we have that a flat connection can be used for building a noncommutative de Rham cohomology with a nontrivial coefficient bundle.
Let then $A$ and $B$ algebras, $R:B\ot A\to A\ot B$ a twisting map, $E$ a right $A$–module endowed with a right connection $\nabla^E$, and $F$ a right $B$–module endowed with a right connection $\nabla^F$ such that we can build the product connection $\nabla$ as in the former section, let also $\nabla=(\nabla^{[n]})$ denote the extension of $\nabla$ to $(E\ot B\oplus A\ot F)\ot_{A\ot_R B}(\O
A\ot_{\widetilde{R}} \O B)$. For $e\in E$, let us denote $\nabla^E(e)=e_i\ot_A d_A a_i$, and $\nabla^E(e_i):=e_{ij}\ot_A
d_Aa_{ij}$, where summation symbols are omitted. In the same spirit, for $f\in F$, we will denote $\nabla^F(f)=f_k\ot_B d_B b_k$, and $\nabla^F(f_k):=f_{kl}\ot_B d_B b_{kl}$. With this notation, the respective curvatures are written as $\theta^E(e)=e_{ij}\ot_A
d_Aa_{ij}d_A{a_i}$, $\theta^F(f)=f_{kl}\ot_B d_B b_{kl} d_B{b_k}$. We will also denote by $i_E$ and $i_F$ the canonical inclusions (as vector spaces) of $E\ot_A \O^2 A$ and $F\ot_B \O^2 B$ into $(E\ot
B\oplus A\ot F)\ot_{A\ot_{R}B}(\O A\ot_{\widetilde{R}}\O B)^2$. For a generic element $(e\ot b,a\ot f)\in (E\ot B\oplus A\ot F)$, using the definition of the product connection we have that $$\begin{aligned}
\nabla(e\ot b,a\ot f) &=& e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot b + e\ot 1\ot_{A\ot_{R}B} 1\ot d_Bb + \\
&& + 1\ot f_\sigma \ot_{A\ot_{R}B} d_A(a_\sigma)\ot 1 + a\ot f_k\ot_{A\ot_{R}B} 1\ot d_Bb_k\end{aligned}$$ Applying $\nabla^{[1]}$ to each of these four term we obtain: $$\begin{aligned}
\nabla^{[1]}(e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot b) & = & \nabla(e_i\ot
1)\cd (d_Aa_i\ot b) +\\
&& +\ (e_i\ot 1)\ot_{A\ot_{R}B} d(da_i\ot b) \overset{1}{=}\\
& \overset{1}{=} &(e_ij\ot 1 \ot_{A\ot_{R}B} d_Aa_{ij}\ot 1)\cd
(d_Aa_i\ot b) \\
&& - e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot d_Bb = \\
&= & e_{ij}\ot 1 \ot_{A\ot_{R}B} d_Aa_{ij}d_A a_i\ot b \\
&& - e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot d_Bb = \\
& = & i_E(\theta^E(e))\cd b - e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot d_Bb,\end{aligned}$$ $$\begin{aligned}
\nabla^{[1]}(e\ot 1\ot_{A\ot_{R}B} 1\ot d_Bb) & = & \nabla(e\ot 1)\cd 1\ot d_Bb + (e\ot 1)\ot_{A\ot_{R}B} d(1\ot d_Bb) = \\
& = & e_i\ot 1\ot_{A\ot_{R}B} d_Aa_i\ot d_Bb,\end{aligned}$$ $$\begin{aligned}
\nabla^{[1]}(1\ot f_\sigma \ot_{A\ot_{R}B} d_A(a_\sigma)\ot 1) & = & \nabla(1\ot f_\sigma)\cd (d_A(a_\sigma)\ot 1) + \\
&& + (1\ot f_\sigma)\ot_{A\ot_{R}B} d(d_A(a_\sigma)\ot 1) = \\
&=& (1\ot(f_\sigma)_k\ot_{A\ot_{R}B} 1\ot d_B(b_\sigma)_k)\cd (d_A (a_\sigma)\ot 1) = \\
& = & 1\ot(f_\sigma)_k\ot_{A\ot_{R}B} (d_A(a_\sigma))_{\widetilde{R}}\ot (d_B(b_\sigma)_k)_{\widetilde{R}} \overset{2}{=}\\
& \overset{2}{=} & -1\ot(f_\sigma)_k\ot_{A\ot_{R}B} d_A(a_\sigma)_{\widetilde{R}}\ot (d_B(b_\sigma)_k)_{\widetilde{R}}\overset{\eqref{eq:nabla2cond1bis}}{=} \\
&\overset{\eqref{eq:nabla2cond1bis}}{=}& -1\ot(f_k)_\sigma \ot_{A\ot_{R}B} d_A(a_\sigma)\ot d_Bb_k,\end{aligned}$$ $$\begin{aligned}
\nabla^{[1]}(a\ot f_k\ot_{A\ot_{R}B} 1\ot d_Bb_k) & = & \nabla(a\ot
f_k)\cd(1\ot d_Bb_k) + \\
&& +\ a\ot f_k\ot_{A\ot_{R}B} d(1\ot d_Bb_k) = \\
&=& (a\ot f_{kl} \ot_{A\ot_{R}B} 1\ot d_B b_{kl})\cd(1\ot d_Bb_k) + \\
&& + (1\ot (f_k)_\sigma \ot_{A\ot_{R}B} d_A a_\sigma\ot 1)\cd(1\ot d_Bb_k) =\\
&=& a\ot f_{kl} \ot_{A\ot_{R}B} 1\ot d_B b_{kl}d_Bb_k + \\
&& + 1\ot (f_k)_\sigma \ot_{A\ot_{R}B} d_A a_\sigma \ot d_Bb_k = \\
&=& a\cd i_F(\theta^F(f)) + 1\ot (f_k)_\sigma \ot_{A\ot_{R}B} d_A a_\sigma \ot d_Bb_k.\end{aligned}$$ where in $1$ we are using the definitions of $\nabla$ and the differential $d$, in $2$ the compatibility of $\widetilde{R}$ with $d_A$. Adding up these four equalities we obtain the following result:
The curvature of the product connection is given by $$\label{curvatureformula}
\theta (e\ot b, a\ot f) = i_E(\theta^E(e))\cd b + a\cd i_F(\theta^F(f)).$$
An interesting remark at the sight of the former result is that the product curvature does not depend neither on the twisting map $R$ nor on the module twisting map $\tau_{F,A}$, but only on the curvatures of the factors. As an immediate consequence of Equation we obtain the following result:
The product connection of two flat connections is a flat connection.
Henceforth, one might ask the question of describing the de Rham cohomology with coefficients in the sense of Beggs and Brzezinski (ref. [@Beggs05a]) for the (twisted) product connection of two flat connections. We will leave this problem for future works. It is also worth noticing that formula drops down in the commutative case to the classical formula for the curvature on a product manifold.
Bimodule connections {#bimodules}
====================
For many purposes, only considering right (or left) modules is not enough. On the one hand, if we want to apply our theory to $\ast$–algebras, then sooner or later we will be bond to deal with $\ast$–modules and hermitian modules, but since the involution reverses the order of the products, these notions only make sense when we consider bimodules. On the other hand, there is a special kind of connections, known as [***linear connections***]{}, obtained when we take $E=\O^1 A$. Since $\O^1 A$ is a bimodule in a natural way, there is no reason to neglect one of its structures restraining ourselves to look at it just as a one-sided module. Reasons for extending the notion of connection to bimodules have been largely discussed at [@Mourad95a], [@DuboisViolette99a] and references therein.
Different approaches for dealing with this problem have been tried. The first one, described by Cuntz and Quillen in [@Cuntz95a], consists on considering a couple $(\nabla^l,\nabla^r)$ where $\nabla^l$ is a left connection which is also a right $A$–module morphism, and $\nabla^r$ a right connection which is also a left $A$–module morphism. As it was pointed out in [@Dabrowski96a], this approach, though rising a very interesting algebraic theory, is not well suited for our geometrical point of view, since it doesn’t behave as expected when restricted to the commutative case. A different approach was introduced by Mourad in [@Mourad95a] for the particular case of linear connections and later generalized to arbitrary bimodules by Dubois-Violette and Masson in [@DuboisViolette96a] (see also [@DuboisViolette99a Chapter 10]). Their approach goes as follow: let $E$ be an $A$–bimodule; a [***(right) bimodule connection***]{} on $E$ is a right connection $\nabla:E\to E\ot_A\O^1 A$ together with a bimodule homomorphism $\sigma:\O^1 A\ot_A E \to E\ot_A \O^1 A$ such that $$\nabla(ma)=a\nabla(m) + \sigma(d_A(a)\ot_A m)\quad \text{for any $a\in A$, $m\in E$}.$$ Giving a right bimodule connection in the above sense is equivalent to give a pair $(\nabla^L,\nabla^R)$ consisting in a left connection $\nabla^L$ and a right connection $\nabla^R$ that are $\sigma$–compatible, meaning that $$\label{sigmacompatible}
\nabla^R = \sigma\circ \nabla^L.$$
A weaker definition of $\sigma$–compatibility, namely requiring that equation holds only in the center $Z(E):=\{m\in E:\ am=ma\ \forall a\in A\}$ of $E$ rather than in the whole bimodule, has also been studied in [@Dabrowski96a].
So, assume that we have $E$ bimodule over $A$, $\nabla^E$ a bimodule connection on $E$ with respect to the morphism $\vphi:\O^1A\ot_A E\to E\ot_A\O^1 A$, and $F$ a bimodule over $B$ endowed with $\nabla^F$ a bimodule connection with respect to the bimodule morphism $\psi:\O^1 B\ot_B F\to F\ot_B\O^1 B$. As before, let $R:B\ot A\to A\ot B$ an invertible twisting map with inverse $S$, and assume also that we have a right module twisting maps $\tau_{F,A}:F\ot A\to A\ot F$ satisfying condition and a left module twisting map $\tau_{B,E}:B\ot E\to E\ot B$ satisfying condition $$\label{eq:nabla1cond1}
(\nabla^E\ot B)\circ \tau_{B,E} = (E\ot \widetilde{R})\circ (\tau_{B,E}\ot \O^1 A)\circ (B\ot \nabla^E),$$ which is the analogous of condition , and such that $(E\ot B)\oplus(A\ot F)$ becomes an $A\ot_R B$ bimodule with left action $$(\alpha\ot \beta)\cd (e\ot b,a\ot f):= (\alpha e_\tau\ot \beta_\tau b, \alpha a_R\ot \beta_R f),$$ then we have that $$\begin{aligned}
\nabla((\alpha\ot\beta)(e\ot b)) &=& \nabla_1(\alpha e_\tau \ot \beta_\tau b) = \\
&=& (\alpha e_\tau)_i\ot 1\ot_{A\ot_R B} d_A(a'_i)\ot \beta_\tau b+\\
&& +\ \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot d_B(\beta_\tau b) = \\
&=& \alpha(e_\tau)_i\ot 1 \ot_{A\ot_R B} d_A(a_\tau)_i\ot\beta_\tau b+\\
&& +\ (e_\tau)_\vphi\ot 1\ot_{A\ot_R B} (d_A\alpha)_\vphi\ot \beta_\tau
b + \\
&& +\ \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot d_B(\beta_\tau)b +\\
&& +\ \alpha
e_\tau\ot 1\ot_{A\ot_R B} 1\ot \beta_\tau d_B b
\overset{\eqref{eq:nabla1cond1}}{=} \\
&\overset{\eqref{eq:nabla1cond1}}{=}& \alpha(e_i)_\tau \ot 1
\ot_{A\ot_R B} (d_Aa_i)_{\widetilde{R}}\ot(\beta_\tau)_{\widetilde{R}} b+\\
&& \ \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot \beta_\tau d_B b + \\
&&+\ (e_\tau)_\vphi\ot 1\ot_{A\ot_R B} (d_A\alpha)_\vphi\ot \beta_\tau
b +\\
&& +\ \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot d_B(\beta_\tau)b = \\
&=& (\alpha\ot\beta)\nabla_1(e\ot b) +\\
&& +\ (e_\tau)_\vphi\ot 1\ot_{A\ot_R
B} (d_A\alpha)_\vphi\ot \beta_\tau b + \\
&&+\ \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot d_B(\beta_\tau)b.
\end{aligned}$$ On the other hand, $$\begin{aligned}
\nabla((\alpha\ot\beta)(a\ot f)) &=& \nabla_2(\alpha a_R \ot \beta_R f) = \\
&=& 1\ot (\beta_R f)_\sigma \ot_{A\ot_R B} d_A((\alpha a_R)_\sigma)\ot 1
+\\
&& +\ \alpha a_R\ot (\beta_R f)_k\ot_{A\ot_R B}1\ot d_Bb'_k)
\overset{\eqref{eq:nabla2cond4}}{=} \\
&\overset{\eqref{eq:nabla2cond4}}{=}& 1\ot (\beta_R f)_{\sigma\bar{\sigma}} \ot_{A\ot_R B} d_A (\alpha_{\bar{\sigma}}(a_R)_\sigma)\ot 1+ \\
&&+\ \alpha a_R\ot (\beta_R f)_k\ot_{A\ot_R B} 1\ot d_B(b'_k)
\overset{\eqref{eq:nabla2cond5}}{=}\\
&\overset{\eqref{eq:nabla2cond5}}{=}& 1\ot (\beta
f_\sigma)_{\bar{\sigma}} \ot_{A\ot_R B} d_A (\alpha_{\bar{\sigma}}a_\sigma)\ot 1
+\\
&& +\ \alpha a_R\ot (\beta_R f)_k\ot_{A\ot_R B} 1\ot d_B(b'_k) = \\
&=& 1\ot (\beta f_\sigma)_{\bar{\sigma}} \ot_{A\ot_R B} d_A
(\alpha_{\bar{\sigma}})a_\sigma \ot 1 +\\
&& +\ 1\ot (\beta f_\sigma)_{\bar{\sigma}}
\ot_{A\ot_R B} \alpha_{\bar{\sigma}}d_A (a_\sigma)\ot 1 + \\
&& +\ \alpha a_R\ot \beta_R f_k\ot_{A\ot_R B} 1\ot d_Bb_k +\\
&& +\ \alpha a_R\ot f_\psi \ot_{A\ot_R B} 1\ot (d_B(\beta_R))_\psi =\\
&=& (\alpha\ot \beta)\nabla_2(a\ot f) + \\
&& +\ 1\ot (\beta f_\sigma)_{\bar{\sigma}} \ot_{A\ot_R B} d_A
(\alpha_{\bar{\sigma}})a_\sigma \ot 1 + \\
&&+\, \alpha a_R\ot f_\psi \ot_{A\ot_R B} 1\ot (d_B(\beta_R))_\psi.
\end{aligned}$$ Adding up these two equalities we obtain $$\begin{aligned}
\nabla((\alpha\ot\beta)(e\ot b, a\ot f)) &=& (\alpha\ot \beta)\nabla (e\ot
b, a\ot f) + \\
&&+\ \xi(d(\alpha\ot\beta)\ot_{A\ot_R B} (e\ot b, a\ot f)),\end{aligned}$$ where the map $\xi: (\O^1A\ot B\oplus A\ot \O^1B)\ot_{A\ot_R B}
(E\ot B\oplus A\ot F) \to (E\ot B\oplus A\ot F)\ot_{A\ot_R
B}(\O^1A\ot B\oplus A\ot \O^1B)$ is defined by $\xi:=\xi_{11}+\xi_{12}+\xi_{21}+\xi_{22}$, being $$\begin{gathered}
\xi_{11}(d_A\alpha\ot\beta\ot_{A\ot_R B} e\ot b) := (e_\tau)_\vphi\ot 1\ot_{A\ot_R B} (d_A\alpha)_\vphi\ot \beta_\tau
b,\\
\xi_{12}(\alpha\ot d_B\beta \ot_{A\ot_R B} e\ot b) := \alpha e_\tau\ot 1\ot_{A\ot_R B} 1\ot
d_B(\beta_\tau)b,\\
\xi_{21}(d_A\alpha\ot\beta\ot_{A\ot_R B} a\ot f) := 1\ot (\beta f_\sigma)_{\bar{\sigma}} \ot_{A\ot_R B} d_A (\alpha_{\bar{\sigma}})a_\sigma \ot
1,\\
\xi_{22}(\alpha\ot d_B \beta \ot_{A\ot_R B} a\ot f) := \alpha a_R\ot f_\psi \ot_{A\ot_R B} 1\ot (d_B(\beta_R))_\psi.\end{gathered}$$ Hence, in order to show that the product connection $\nabla$ is a bimodule connection we only have to show that $\xi$ is a bimodule morphism, which is equivalent to prove that all the $\xi_{ij}$ are bimodule morphisms.
\[lemmaxi11\] The map $\xi_{11}$ is a left $(A\ot_R B)$–module morphism, if, and only if, the equality $$\label{xi11compatibility}
(\vphi\ot B)\circ (\O^1 A\ot\tau_{B,E})\circ (\widetilde{R}\ot E)=(E\ot \widetilde{R})\circ (\tau_{B,E}\ot \O^1 A)\circ (B\ot \vphi)$$ is satisfied in $B\ot \O^1 A\ot E$.
In order to check that the compatibility condition is necessary, just apply the compatibility with the module action to an element of the form $1\ot b\ot \omega\ot 1\ot e\ot 1$.
Conversely, assuming condition , we have that $$\begin{gathered}
\xi_{11}((x\ot y)\cd (d\alpha\ot\beta\ot_{A\ot_R B} e\ot b))
= \\
= \xi_{11}(x(d\alpha)_{\widetilde{R}}\ot y_{\widetilde{R}}\beta \ot_{A\ot_R
B} e\ot b) = \\
= (e_\tau)_\vphi\ot 1\ot_{A\ot_R B} (x(d\alpha)_{\widetilde{R}})_\vphi
\ot (y_{\widetilde{R}}\beta)_\tau b \overset{[1]}{=}\\
\overset{[1]}{=} x(e_\tau)_\vphi\ot 1\ot_{A\ot_R B}
((d\alpha)_{\widetilde{R}})_\vphi \ot (y_{\widetilde{R}}\beta)_{\tau} b
\overset{[2]}{=}\\
\overset{[2]}{=} x ((e_{\tau})_{\bar{\tau}})_\vphi\ot 1\ot_{A\ot_R B}
((d\alpha)_{\widetilde{R}})_{\vphi}\ot
(y_{\widetilde{R}})_{\bar{\tau}}\beta_\tau b
\overset{\eqref{xi11compatibility}}{=}\\
\overset{\eqref{xi11compatibility}}{=}
x((e_\tau)_\vphi)_{\bar{\tau}}\ot 1\ot_{A\ot_R B}
((d\alpha)_{\vphi})_{\widetilde{R}}\ot
(y_{\bar{\tau}})_{\widetilde{R}}\beta_\tau b = \\
= x((e_{\tau})_\vphi)_{\bar{\tau}}\ot 1\ot_{A\ot_R B} ((1\ot
y_{\bar{\tau}})\cd((d\alpha)_{\vphi}\ot \beta_\tau b)) = \\
= x((e_{\tau})_\vphi)_{\bar{\tau}}\ot y_{\bar{\tau}}\ot_{A\ot_R B}
(d\alpha)_{\vphi}\ot \beta_\tau b =\\
= (x\ot y)\xi_{11}(d\alpha\ot\beta\ot_{A\ot_R B} e\ot b),
\end{gathered}$$ where in \[1\] we are using that $\vphi$ is a left module map, in \[2\] that $\tau$ is a module twisting map.
It is straightforward checking that $\xi_{11}$ is a right module map, and thus left to the reader. In a completely analogous way, it is straightforward to check that $\xi_{22}$ is a left module map, whilst for the right module condition we need a compatibility relation similar to . More concretely, we have the following result, whose proof is analogous to the one of Lemma \[lemmaxi11\]:
The map $\xi_{22}$ is a right $A\ot_R B$–module morphism if, and only if, the equality $$\label{xi22compatibility}
(A\ot \psi)\circ(\widetilde{R}\ot F)\circ(\O^1 B\ot \tau_{F,A})=(\tau_{F,A}\ot \O^1 B)\circ(F\ot \widetilde{R})\circ (\psi\ot A)$$ is satisfied in $\O^1 B\ot F\ot A$.
For $\xi_{12}$ and $\xi_{21}$, the right (resp. left) module map conditions are also straightforward. We will show now that $\xi_{12}$ is a left module map, the proof that $\xi_{21}$ is a right module map being analogous.
$$\begin{aligned}
\xi_{12}((x\ot y)\cd(\alpha\ot d\beta\ot_{A\ot_R B})e\ot b) &=& \xi_{12}(x\alpha_R\ot yd\beta \ot_{A\ot_R B} e\ot b) =\\
&=& \xi_{12}(x\alpha_R\ot d(y_R\beta) \ot_{A\ot_R B} e\ot b) - \\
&& - \xi_{12}(x\alpha_R\ot d(y_R) \ot_{A\ot_R B} e_\tau \ot \beta_\tau b)=\\
&=& x\alpha_R e_\tau \ot 1\ot_{A\ot_R B} 1\ot (d(y_R\beta)_\tau)b - \\
&& - x\alpha_R (e_\tau)_{\bar{\tau}}\ot 1\ot_{A\ot_R B} 1\ot d((y_R)_{\bar{\tau}}) \beta_{\tau} b\overset{[1]}{=}\\
&\overset{[1]}{=}& x\alpha_R (e_\tau)_{\bar{\tau}} \ot 1\ot_{A\ot_R B} 1\ot (d((y_R)_{\bar{\tau}}\beta_\tau))b - \\
&& - x\alpha_R (e_\tau)_{\bar{\tau}}\ot 1\ot_{A\ot_R B} 1\ot d((y_R)_{\bar{\tau}}) \beta_{\tau} b =\\
&=& x\alpha_R (e_\tau)_{\bar{\tau}} \ot 1\ot_{A\ot_R B} 1\ot d((y_R)_{\bar{\tau}})\beta_\tau b + \\
&& +x\alpha_R (e_\tau)_{\bar{\tau}} \ot 1\ot_{A\ot_R B} 1\ot (y_R)_{\bar{\tau}}(d(\beta_\tau))b-\\
&& - x\alpha_R (e_\tau)_{\bar{\tau}}\ot 1\ot_{A\ot_R B} 1\ot d((y_R)_{\bar{\tau}}) \beta_{\tau} b =\\
&=& x\alpha_R (e_\tau)_{\bar{\tau}} \ot 1\ot_{A\ot_R B} 1\ot (y_R)_{\bar{\tau}}(d(\beta_\tau))b \overset{[2]}{=} \\
&\overset{[2]}{=}& x(\alpha e_\tau)_{\bar{\tau}}\ot 1 \ot_{A\ot_R B} 1\ot y_{\bar{\tau}}(d\beta_\tau)b = \\
&=& x(\alpha e_\tau)_{\bar{\tau}}\ot y_{\bar{\tau}} \ot_{A\ot_R B} 1\ot (d\beta_\tau)b = \\
&=& (x\ot y)\cd \xi_{12}(\alpha\ot d\beta\ot_{A\ot_R B})e\ot b),
\end{aligned}$$
where in \[1\] and \[2\] we use that $\tau_{F,A}$ is a module twisting map.
Summarizing, we have proved the following result:
Let $E$ be a bimodule over $A$, $(\nabla^E,\vphi)$ a bimodule connection on $E$, $F$ a bimodule over $B$, $(\nabla^F,\psi)$, $R:B\ot A\to A\ot B$ an invertible twisting map; $\tau_{F,A}:F\ot A\to A\ot F$ a right module twisting map satisfying condition and $\tau_{B,E}:B\ot E\to E\ot B$ a left module twisting map satisfying condition . Assume also that conditions and are satisfied, then the product connection of $\nabla^E$ and $\nabla^F$ is a bimodule connection with respect to the morphism $\xi$.
Examples
========
Let us start by recalling some facts from [@Cuntz95a]. For any projective (right) module $E$ over an algebra $A$, there exists a module $E'$ such that $E\oplus E' = A^n$, and we have two canonical mappings $$p:A^n=E\oplus E' \lto E\quad \text{and } \lambda: E
\hookrightarrow E\oplus E',$$ we can then define the map $\nabla_0:=(p\ot\operatorname{Id})\circ (A^n\ot d)\circ
(\lambda\ot \operatorname{Id})$ as the composition given by $$\xymatrix{
E\ot_A\O^p A \ar[rr]^{\lambda\ot \operatorname{Id}}&&
A^n\ot_A \O^p A \ar[rr]^{A^n\ot d} &&
\O^{p+1} A \ar[rr]^{p\ot \operatorname{Id}} && E\ot_A\O^{p+1}A
}$$ The operator $\nabla_0$ is a (flat) connection on $E$, called the [***Grassmann connection***]{} on $E$.
Physicists sometimes use the shorthand notation $\nabla_0=pd$ to denote the Grassmann connection.
It is also well known (cf. for instance [@Cuntz95a]) that the space of all linear connections over a projective module $E$ is an affine space modeled on the space of $A$–module morphisms $\operatorname{End}_A(E)\ot_A \O^1 A$, and henceforth we can write any linear connection $\nabla$ on $E$ as $\nabla=\nabla_0 + \alpha$, being $\alpha\in \operatorname{End}_A(E)\ot_A \O^1 A$, where the “*matrix*” $\alpha$ is called the [***gauge potential***]{} of the connection $\nabla$.
Product connections on the quantum plane $k_q[x,y]$
---------------------------------------------------
Consider now $A:=k[x]$ the polynomial algebra in one variable. Since any projective module over $A$ is free (actually, by Quillen-Suslin Theorem, any projective module over any polynomial ring is free) it is enough to consider connections for modules of the form $E=A^m$. If we denote by $\{e_i\}_{i=1,\dotsc,
m}$ the canonical generator set for $E$, we may write the Grassmann connection on $E$ as $$\nabla_0^E(a_1,\dotsc,a_m)=e_1\ot_A da_1 + \dotsb + e_m\ot_A
da_m\in E\ot_A \O^1 A.$$ Analogously, let $B:=k[y]$, $F:=B^n$ with canonical generating system $\{f_j\}_{j=1,\dotsc,n}$ and Grassmann connection $$\nabla_0^F(b_1,\dotsc, b_n)=f_1\ot_B db_1 + \dotsb +
f_n\ot_Bdb_n.$$
Recall that the quantum plane $k_q[x,y]$ may be seen as the twisted tensor product $k[x]\ot_R k[y]$ with respect to the twisting map obtained by extension of $R(y\ot x):=q x\ot y$. This is an invertible twisting map which extends to an invertible module twisting map $\tau_{F,A}:F\ot A\to A\ot F$ in a natural way. For elements $e\ot b \in E\ot B$, where $e=(a_1,\dotsc,a_m)$, and a generator $x\ot f$ with $f=(y^{i_1},\dotsc, y^{i_n})$ of $A\ot F$, using the definition of our product connection given by Equation , we have that the product of the Grassmann connections is $$\begin{aligned}
\nabla^{gr}(e\ot b,x\ot f) & = & \left( \sum e_i\ot 1\ot da_i\right)\ot b + e\ot 1\ot 1\ot db + \\
& & + x\ot \left( \sum f_k\ot 1\ot dy^{i_k}\right) + 1\ot (q^{-i_1} y^{i_1},\dotsc,q^{-i_n} y^{i_n})\ot dx\ot 1
\end{aligned}$$
If we introduce the notation $\lambda_q (p(y)):=p(qy)$, we can give the former expression for an element $a\ot f$ of the form $a=x^j$, $f=(b_1,\dotsc,b_n)\in F$ as $$\begin{aligned}
\nabla^{gr}(e\ot b,a\ot f) & = & \sum_i e_i\ot 1\ot da_i\ot b + e\ot 1\ot 1\ot db + \\
& & + \sum_k a\ot f_k\ot 1\ot db_k + \sum_k 1\ot \lambda_{q^{-j}}(b_k)\ot d(x^j)\ot 1
\end{aligned}$$
Now, for a generic connection $\nabla^E$ over the module $E$, there must exist a potential $\alpha^E=\vphi_i\ot \omega_i\in
\operatorname{End}{E}\ot_A\O^1 A$ given by $\alpha^E(a_1,\dotsc,a_m)=\sum_{i,j}
\vphi_i(a_j)\ot \omega_i$ such that $\nabla^E=\nabla^E_0+\alpha^E$. In the same way, for a generic connection $\nabla^F$ on $F$ there must exist a potential $\alpha^F=\sum_k \psi_k\ot \eta_k$, given by $\alpha^F(b_1,\dotsc,
b_n)=\sum_{k,l}\psi_k{b_l}\ot\eta_k$, and such that $\nabla^F=\nabla^F_0+\alpha^F$. Applying the formula for the product connection to $\nabla^E$ and $\nabla^F$ we easily observe that $$\nabla(e\ot b,a\ot f) = \nabla^{gr}(e\ot b,a\ot f) + \sum_{i,j}\vphi_i(a_j)\ot 1\ot \omega_i\ot b + \sum_{k,l}a\ot\psi_k(b_l)\ot 1\ot\eta_k,$$ expression that tells us the formula for all possible product connections on the quantum plane.
\[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{}
[DVMMM95]{}
E. J. Beggs and T. Brzezinski, *The [S]{}erre spectral sequence of a noncommutative fibration for de [R]{}ham cohomology*, To appear in Acta Math. (Preprint 2005), arxiv: math.QA/0508194.
J. Beck, *Distributive laws*, Lecture Notes in Mathematics **80** (1969), 119–140.
E. Beggs, *Braiding and exponentiating noncommutative vector fields*, math.QA/0306094.
A. Borowiec, *Cartan pairs*, Czech J. Phys. **46** (1996), 1197–1202.
A. Connes, *Non-commutative differential geometry*, Publ, Math. IHES **62** (1986), 44–144.
J. Cuntz and D. Quillen, *Algebra extensions and nonsingularity*, J. Amer. Math. Soc. **8** (1995), 251–289.
A. Cap, H. Schichl, and J. Vanžura, *On twisted tensor products of algebras*, Comm. Algebra **23** (1995), 4701–4735.
L. Dabrowski, P. M. Hajac, G. Landi, and Siniscalco P., *Metrics and pairs of left and right connections on bimodules*, J. Math. Phys. **37** (1996), no. 9, 4635–4646.
M. Dubois-Violette, *Lectures on graded differential algebras and noncommutative geometry*, Proceedings of the Workshop on Noncommutative Differential Geometry and its Application to Physics, Shonan-Kokusaimura., 1999.
Michel Dubois-Violette and Thierry Masson, *[On the first-order operators in bimodules.]{}*, Lett. Math. Phys. **37** (1996), no. 4, 467–474 (English).
M. Dubois-Violette, J. Madore, T. Masson, and J. Mourad, *Linear connections on the quantum plane*, Lett. Math Phys. **35** (1995), 351–359.
Jose M. Gracia-Bond[í]{}a, Joseph C. Varilly, and H[é]{}ctor Figueroa, *Elements of noncommutative geometry*, first ed., Advanced Texts, Birkhäuser, Berlin, 2001.
P. Jara and D. Llena, *Lie bracket of vector fields on noncommutative geometry*, Czech J. Physics **53 (9)** (2003), 743–758.
P. Jara Martínez, J. López Peña, F. Panaite, and F. Van Oystaeyen, *On iterated twisted tensor products of algebras*, math.QA/0511280.
J. L. Koszul, *Fibre bundles and differential geometry*, Tata Institue of Fundamental Research (Bombay), 1960.
Giovanni Landi, *Noncommutative spaces and their geometry*, Springer-Verlag, Berlin, 1997.
J. Lopez, F. Panaite, and F. Van Oystaeyen, *General twisting of algebras*, Adv. Math. (to appear).
J. Madore, *Noncommutative differential geometry and its physical applications*, Cambridge University Press, 1995.
J. Mourad, *Linear connections in non–commutative geometry*, Class. Quantum Grav. **12** (1995), 965–974.
P. Nuss, *Noncommutative descent and nonabelian cohomology*, K-Theory . **12** (1997), 23–74.
D. Tambara, *The coendomorphism bialgebra of an algebra*, J. Fac. Sci. Univ. Tokyo Sect. IA Math. **37** (1990), 425–456.
A. Van Daele and S. Van Keer, *The [Y]{}ang–[B]{}axter and [P]{}entagon equation*, Compositio Math. **91** (1994), 201–221.
S.L. Woronowicz, *An exmaple of a braided locally compact group*, Proceedings of the [IX]{}th [M]{}ax [B]{}orn [S]{}ymposium, Polish Scientific Publishers, 1996.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
-
-
- '[^1]'
- '[^2]'
-
bibliography:
- 'ref.bib'
title: |
Improvement of heavy-heavy current for calculation of\
$\bar{B}\to D^{(*)}\ell\bar{\nu}$ form factors using Oktay-Kronfeld heavy quarks
---
Acknowledgements {#acknowledgements .unnumbered}
================
The research of W. Lee is supported by the Creative Research Initiatives Program (No. 2017013332) of the NRF grant funded by the Korean government (MEST). W. Lee would like to acknowledge the support from the KISTI supercomputing center through the strategic support program for the supercomputing application research (No. KSC-2015-G2-002). Computations were carried out on the DAVID clusters at Seoul National University. J.A.B. is supported by the Basic Science Research Program of the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. 2015024974).
[^1]:
[^2]: Speaker
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we study power series having a fixed point of multiplier $1$. First, we give a closed formula for the residue fixed point index, in terms of the first coefficients of the power series. Then, we use this formula to study wildly ramified power series in positive characteristic. Among power series having a multiple fixed point of small multiplicity, we characterize those having the smallest possible lower ramification numbers in terms of the residue fixed point index. Furthermore, we show that these power series form a generic set, and, in the case of convergent power series, we also give an optimal lower bound for the distance to other periodic points.'
address:
- 'Department of Mathematics, Linnaeus University, V[ä]{}xj[ö]{}, Sweden'
- 'Department of Mathematics, University of Rochester. Hylan Building, Rochester, NY 14627, U.S.A.'
author:
- Jonas Nordqvist
- 'Juan Rivera-Letelier'
title: Residue fixed point index and wildly ramified power series
---
Introduction
============
Consider an open subset $U$ of $\operatorname{\mathbb{C}}$ and a holomorphic map $f \colon U \to \operatorname{\mathbb{C}}$. For a fixed point $z_0$ of $f$, the derivative $f'(z_0)$ is invariant under coordinate changes. In the case $z_0$ is isolated as a fixed point of $f$, a related invariant is defined by the countour integral $$\label{eq:19}
\operatorname{index}(f, z_0)
{\coloneqq}\frac{1}{2\pi i} \oint \frac{{\hspace{1pt}\operatorname{d}\hspace{-1pt}}z}{z - f(z)},$$ where we integrate on a sufficiently small simple closed curve around $z_0$ that is positively oriented. The complex number is invariant under coordinate changes and is called the *residue fixed point index of $f$ at $z_0$*. Together with the related holomorphic fixed point formula, it is one of the basic tools in complex dynamics, see, *e.g.*, [@Mil06c §12] for background, and [@BuffEpstein2002; @Buff03; @BuffXavEcalle2013] for some results where the residue fixed point index plays an important r[ô]{}le. See also [@Silverman2007 Exercise 5.10] for an extension to an arbitrary ground field.
In the case $f'(z_0) \neq 1$, a direct computation shows that is equal to $\frac{1}{1 - f'(z_0)}$. We give a closed formula for in the case $f'(z_0) = 1$, in terms of the first coefficients of the power series expansion of $f$ about $z_0$ (Theorem \[thm:closed-formula\] in §\[sec:closed-formula\]). This formula holds for an arbitrary ground field. We also show that the residue fixed point index is invariant under coordinate changes, and use it to study normal forms. We also study the behavior of the residue fixed point under iteration.
In our succeeding results, we restrict to ground fields of positive characteristic and power series having the origin as a fixed point of multiplier $1$. Such power series are called *wildly ramified*.[^1] See, *e.g.*, [@Sen1969; @Keating1992; @LaubieSaine1998; @Win04] for background on wildly ramified power series, [@KallalKirkpatrick2019; @LaubieMovahhediSalinier2002; @LindahlNordqvist2018; @LindahlRiveraLetelier2013; @LindahlRiveraLetelier2015; @Fransson2017; @RiveraLetelier2003] for results related to this paper, and [@HermanYoccoz1983; @Lindahl2004; @LindahlZieve2010; @Ruggiero2015] and references therein for local dynamics of analytic germs in positive characteristic. See also, *e.g.*, [@johnson1988; @Camina2000] and references therein, for the myriad of group-theoretic results about the “Nottingham group”, which is the group under composition formed by the wildly ramified power series.
Every wildly ramified power series has associated a sequence of “lower ramification” numbers. It encodes the multiplicity of the origin for the iterates of the power series. We study the lower ramification numbers of power series for which the multiplicity at the origin is small. First, we characterize those power series having the smallest possible lower ramification numbers. They are characterized by the nonvanishing of [É]{}calle’s “iterative residue”, which is a dynamical version of the residue fixed point index (Theorem \[thm:q-ramification\] in §\[sec:q-ramification\]). As a consequence, we obtain that these power series form a generic set. In the case of convergent power series, we also give an optimal lower bound for the distance to other periodic points (Theorem \[thm:lower-bound\] in §\[sec:lower-bound\]). This gives an affirmative solution to [@LindahlRiveraLetelier2013 Conjecture 1.2], for generic multiple fixed points of a fixed and small multiplicity, and to [@KallalKirkpatrick2019 Conjecture 4.3].
We proceed to describe our results more precisely.
Closed formula for the residue fixed point index {#sec:closed-formula}
------------------------------------------------
Our first result is a closed formula for the residue fixed point index of a fixed point of multiplier $1$. We allow an arbitrary ground field, and an arbitrary power series about a fixed point. In particular, we allow non-convergent power series. To simplify the notation, throughout the rest of the paper we restrict to the case of a power series $f$ fixing the origin, and denote $\operatorname{index}(f, 0)$ by $\operatorname{index}(f)$.
Let $\operatorname{\mathbb{K}}$ be a field and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ satisfying $f(0)=0$ and $f(z)\neq z$. The *residue fixed point index of $f$ at $0$*, denoted by $\operatorname{index}(f)$, is the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $$\frac{1}{z-f(z)}.$$
Clearly, this definition agrees with in the case where $\operatorname{\mathbb{K}}= \operatorname{\mathbb{C}}$, $z_0 = 0$, and $f$ is holomorphic on a neighborhood of $0$.
To state our first result, denote by $\mathbb{N}$ the set of nonnegative integers and for an integer $q \ge 1$ and $(\operatorname{\iota}_0, \ldots, \operatorname{\iota}_q)$ in $\operatorname{\mathbb{N}}^{q + 1}$, define $$|(\operatorname{\iota}_0, \ldots, \operatorname{\iota}_q)| {\coloneqq}\sum_{j = 0}^{q} \operatorname{\iota}_j
\text{ and }
\| (\operatorname{\iota}_0, \ldots, \operatorname{\iota}_q) \| {\coloneqq}\sum_{j = 1}^q j \operatorname{\iota}_j.$$
\[thm:closed-formula\] Let $\operatorname{\mathbb{K}}$ be a field, $q\geq 1$ an integer, and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ of the form $$\label{psform}f(z)
=
z\left(1 + \sum_{j=q}^{+\infty} a_jz^j\right), \text{ with } a_q \neq 0.$$ Then we have $$\label{eq:closed-formula}
\operatorname{index}(f)
=
- \frac{1}{a_q^{q+1}}\sum_{\substack{\operatorname{\iota}\in \operatorname{\mathbb{N}}^{q+1} \\ |\operatorname{\iota}| = q, \| \operatorname{\iota}\| = q }} (-1)^{q-\operatorname{\iota}_0}\binom{q-\operatorname{\iota}_0}{\operatorname{\iota}_{1},\ldots,\operatorname{\iota}_{q}}\prod_{j = 0}^q a_{q + j}^{\iota_j}.$$
We also show that the residue fixed point index is invariant under coordinate changes (Proposition \[conj\] in §\[invarresidue\]) and use the residue fixed point index to study normal forms (Proposition \[nf\] in §\[sec:nf\]). Both of these results, together with Theorem \[thm:closed-formula\], are used to prove our results below. In Appendix \[app:resit\] we use Theorem \[thm:closed-formula\] to study the behavior under iterations of the residue fixed point index, and of the closely related “iterative residue” defined below.
Wildly ramified power series {#sec:q-ramification}
----------------------------
Let $\operatorname{\mathbb{K}}$ be a field, and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $f(0) = 0$ and $f(z) \neq z$. The *multiplicity of $0$ as a fixed point of $f$* is the lowest degree of a nonzero term in $f(z) - z$. We denote it by $\operatorname{mult}(f)$.
From now on we assume the characteristic $p$ of $\operatorname{\mathbb{K}}$ is positive. The power series $f$ is *wildly ramified* if $\operatorname{mult}(f) \ge 2$, or equivalently, if $0$ is a multiple fixed point of $f$. Note that $f$ is wildly ramified if and only if $f'(0) = 1$. For a wildly ramified power series $f$, the *lower ramification numbers $\{ i_n(f) \}_{n = 0}^{+\infty}$ of $f$* are defined by $$i_n(f) {\coloneqq}\operatorname{mult}(f^{p^n})-1.$$ See, *e.g.*, [@Sen1969; @Keating1992; @LaubieSaine1998; @Win04] and references therein for background on wildly ramified power series and their lower ramification numbers. Due to their relation to ultrametric dynamics, they have been studied in, *e.g.*, [@RiveraLetelier2003 §3.2], [@LindahlRiveraLetelier2015; @LindahlRiveraLetelier2013; @LindahlNordqvist2018]. Note that the lower ramification numbers are invariant under coordinate changes.
If we put $$q {\coloneqq}\operatorname{mult}(f) - 1 \ge 1,$$ then the results of Sen in [@Sen1969] imply that, in the case $q \le p - 1$, for every integer $n \ge 0$ we have $$\label{eq:q-minimal}
i_n(f) \ge q (1+p+ \cdots +p^n),$$ see Proposition \[prop:qramifisminramif\] in §\[sec:lowramif\]. Following [@Fransson2017], for an integer $q \ge 1$ that is not divisible by $p$, we say that $f$ is *$q$-ramified* if equality holds in for every $n$. In the case $q = 1$, $1$-ramified power series are also known as “minimally ramified” [@LaubieMovahhediSalinier2002; @LindahlRiveraLetelier2013; @LindahlRiveraLetelier2015]. $q$-Ramified power series appear naturally as reductions of invertible elements of formal groups, see for example [@LaubieMovahhediSalinier2002 *Proposition* 4.2] for the case $q = 1$, and [@LaubieMovahhediSalinier2002 *Corollaire* 3.12] for general $q$ not divisible by $p$. Note that when $q$ is divisible by $p$, for every $n \ge 1$ we have $i_n(f) = i_0(f)p^n$ [@Sen1969], so we cannot have equality in .
Our next result characterizes $q$-ramified power series when $q \le p - 1$, and shows that $q$-ramified power series are generic among power series having the origin as a fixed point of multiplicity $q + 1$. We restrict to odd $p$, as the case $p = 2$ is treated in [@LindahlRiveraLetelier2015; @LindahlRiveraLetelier2013]. As in [@LindahlRiveraLetelier2013 Theorem E], our characterization is best stated in terms of the “iterative residue”, which is a dynamical variant of the residue fixed point index introduced by [É]{}calle in the complex setting. For a power series $f$ satisfying ${f(0) = 0}$ and ${f(z) \neq z}$, the *iterative residue of $f$* is defined by[^2] $$\label{eq:17}
\operatorname{r\acute{e}sit}(f)
{\coloneqq}\frac{1}{2}\operatorname{mult}(f) - \operatorname{index}(f).$$ See, *e.g.*, [@Ecalle1975 §[I]{}], or [@Mil06c §12] for background on the iterative residue.
\[thm:q-ramification\] Let $p$ be an odd prime number and $\operatorname{\mathbb{K}}$ a field of characteristic $p$. Furthermore, let $q$ be in $\{1, \ldots, p - 1 \}$, and let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ satisfying $\operatorname{mult}(f) = q+1$. Then $f$ is $q$-ramified if and only if $\operatorname{r\acute{e}sit}(f) \neq 0$.
Let $q \ge 1$ be an integer, $x_q$, $x_{q + 1}$, …indeterminates over $\operatorname{\mathbb{K}}$, and consider the generic power series $$f(\zeta) {\coloneqq}\zeta \left( 1 + \sum_{j = q}^{+ \infty} x_j \zeta^j \right).$$ Then by Theorem \[thm:closed-formula\], $x_q^{q + 1} \operatorname{r\acute{e}sit}(f)$ is equal to
$$\label{eq:genericity polynomial}
\left( \frac{q + 1}{2} \right) x_q^{q + 1} + \sum_{\substack{\operatorname{\iota}\in \operatorname{\mathbb{N}}^{q+1} \\ |\operatorname{\iota}| = q, \| \operatorname{\iota}\| = q }} (-1)^{q-\operatorname{\iota}_0}\binom{q-\operatorname{\iota}_0}{\operatorname{\iota}_{1},\ldots,\operatorname{\iota}_{q}}\prod_{j = 0}^q x_{q + j}^{\iota_j},$$
which is a polynomial in $x_q$, $x_{q + 1}$, …, $x_{2q}$ with coefficients in $\operatorname{\mathbb{F}}_p$.[^3] Thus, the following corollary is a direct consequence of Theorem \[thm:q-ramification\].
\[c:genericity\] Let $p$ be an odd prime number, $\operatorname{\mathbb{K}}$ a field of characteristic $p$, and $q$ in $\{1, \ldots, p - 1 \}$. Then, among power series with coefficients in $\operatorname{\mathbb{K}}$ for which the origin is a fixed point of multiplicity $q + 1$, those that are $q$-ramified are generic.
The following corollary is essentially a reformulation of the previous corollary in terms of the *Nottingham group $\mathcal{N}(\operatorname{\mathbb{K}})$*, which is the group under composition formed by all wildly ramified power series with coefficients in $\operatorname{\mathbb{K}}$. Since the work of Johnson [@johnson1988], this group has been extensively studied for its interesting group-theoretic properties. See for instance the survey article [@Camina2000].
Given an integer $q \ge 1$, consider the subgroup of $\mathcal{N}(\operatorname{\mathbb{K}})$, $$\mathcal{N}_q(\operatorname{\mathbb{K}})
{\coloneqq}\{ f \text{ power series with coefficients in~$\operatorname{\mathbb{K}}$ satisfying~$\operatorname{mult}(f) \ge q + 1$} \}.$$ Note that in the case $q = 1$, we have $\mathcal{N}_1(\operatorname{\mathbb{K}}) = \mathcal{N}(\operatorname{\mathbb{K}})$.
Let $p$ be an odd prime number, $\operatorname{\mathbb{K}}$ a field of characteristic $p$, and $q$ in $\{1, \ldots, p - 1 \}$. Then, an element $f$ of $\mathcal{N}_q(\operatorname{\mathbb{K}})$ is $q$-ramified if and only if $\operatorname{r\acute{e}sit}(f) \neq 0$. In particular, $q$-ramified power series are generic in $\mathcal{N}_q(\operatorname{\mathbb{K}})$.
This answers [@KallalKirkpatrick2019 Question 1.4] for $q$ in $\{1, \ldots, p - 1\}$.
In the case $q=1$, Theorem \[thm:q-ramification\] was shown by Lindahl and the second named author [@LindahlRiveraLetelier2013 Theorem E]. This last result also applies to the case $p=2$, and asserts that a power series of the form with $q = 1$ is $1$-ramified if and only if $$\operatorname{r\acute{e}sit}(f) \neq 0
\text{ and }
\operatorname{r\acute{e}sit}(f) \neq 1.$$ In the case $q=2$, Theorem \[thm:q-ramification\] was shown by the first named author [@Fransson2017 Theorem 1], with $\operatorname{r\acute{e}sit}(f)$ replaced by . In the case $q = 3$ and $\operatorname{\mathbb{K}}= \operatorname{\mathbb{F}}_p$, Theorem \[thm:q-ramification\] was shown by Kallal and Kirkpatrick in the first version of [@KallalKirkpatrick2019], with $\operatorname{r\acute{e}sit}(f)$ replaced by . After a preliminary version of this paper was completed, we received a new version of [@KallalKirkpatrick2019] proving Theorem \[thm:q-ramification\] when restricted to those $q$ satisfying $q^2 < p$, and with $\operatorname{r\acute{e}sit}(f)$ replaced by .
Theorem \[thm:q-ramification\] and its corollaries are not expected to extend to the case $q \ge p + 1$ not divisible by $p$. In fact, we give examples showing that the conclusion of Theorem \[thm:q-ramification\] is false for $q = p + 1$, see Example \[ex:p + 1\] in §\[sec:furth-results-exampl\]. About genericity, if $q \ge p + 1$ is not divisible by $p$, then the results of Laubie and Sa[ï]{}ne in [@LaubieSaine1998] imply that the inequality fails in general, even for $n = 1$. Thus, for $q \ge p + 1$ the $q$-ramified power series are not expected to be generic among power series having $0$ as a fixed point of multiplicity $q + 1$. So, the following question arises naturally.
Let $p$ be a prime number, $\operatorname{\mathbb{K}}$ a field of characteristic $p$, and $q \ge p + 1$ an integer that is not divisible by $p$. How are the lower ramification numbers of a generic power series in $\mathcal{N}_q(\operatorname{\mathbb{K}})$? [^4]
In the case $q = p + 1$, it seems that for a generic power series satisfying $\operatorname{mult}(f) = q + 1$, we have for every $n \ge 0$ $$i_n(f) = 1 + p + \cdots + p^{n + 1}.$$ See also Example \[ex:p + 1\] in §\[sec:furth-results-exampl\], and the discussion following it.
Periodic points of wildly ramified power series {#sec:lower-bound}
-----------------------------------------------
Our next result is about the distribution of periodic points of a convergent $q$-ramified power series. To state it, we introduce some notation. Given an ultrametric field $(\operatorname{\mathbb{K}},|\cdot|)$, denote by $$\mathcal{O}_{\operatorname{\mathbb{K}}}
{\coloneqq}\{ \zeta \in \operatorname{\mathbb{K}}: |\zeta| \le 1 \},
\text{ and }
\mathfrak{m}_{\operatorname{\mathbb{K}}}
{\coloneqq}\{ \zeta \in \operatorname{\mathbb{K}}: |\zeta| < 1 \},$$ the ring of integers of $\operatorname{\mathbb{K}}$ and the maximal ideal of $\mathcal{O}_{\operatorname{\mathbb{K}}}$, respectively.
\[thm:lower-bound\] Let $p$ be an odd prime number, let $q$ be in $\{1, \ldots, p - 1 \}$, and let $(\operatorname{\mathbb{K}}, |\cdot|)$ be an ultrametric field of characteristic $p$. Furthermore, let $f$ be a power series with coefficients in $\mathcal{O}_{\operatorname{\mathbb{K}}}$ of the form $$f(\zeta)
\equiv
\zeta(1 + a\zeta^q) \mod \langle \zeta^{q+2} \rangle, \text{ with } a\neq0.$$ Then, for every fixed point $\zeta_0$ of $f$ in $\mathcal{O}_{\operatorname{\mathbb{K}}}$ that is different from $0$ we have $|\zeta_0|\geq |a|$, and for every periodic point $\zeta_0$ of $f$ in $\mathcal{O}_{\operatorname{\mathbb{K}}}$ that is not a fixed point, we have $$\label{normbound}
|\zeta_0|
\ge
|a| \cdot |\operatorname{r\acute{e}sit}(f)|^\frac{1}{p}.$$
We give explicit examples for which equality holds in for every periodic point that is not fixed, when $q \le p - 3$ (Example \[e:optimality\] in §\[sec:furth-results-exampl\]). We recall that by Theorem \[thm:closed-formula\] we can explicitly compute $\operatorname{r\acute{e}sit}(f)$, see also , so the lower bound in Theorem \[thm:lower-bound\] is effective. Note also that the lower bound given by Theorem \[thm:lower-bound\] is trivial in the case that $f$ is not $q$-ramified, because by Theorem \[thm:q-ramification\] we have $\operatorname{r\acute{e}sit}(f) = 0$ in this case.
Note that every convergent power series about $0$ without constant term is conjugated to a power series with coefficients in $\mathcal{O}_{\operatorname{\mathbb{K}}}$ by a scale change. So, the following corollary is a direct consequence of Theorem \[thm:lower-bound\].
Let $\operatorname{\mathbb{K}}$ be an ultrametric field of positive characteristic, and let $q \ge 1$ be an integer that is strictly smaller than the characteristic of $\operatorname{\mathbb{K}}$. Moreover, let $f$ be a $q$-ramified power series with coefficients in $\operatorname{\mathbb{K}}$ that converges on a neighborhood of the origin. Then the origin is isolated as a periodic point of $f$.
Combined with Corollary \[c:genericity\] and [@LindahlRiveraLetelier2013 Theorem E with $p = 2$], the previous corollary implies the following result as a direct consequence.
\[c:generic isolation\] Let $p$ be a prime number and fix $m$ in $\{2, \ldots, p \}$. Then, over a field of characteristic $p$, a generic fixed point of multiplicity $m$ is isolated as a periodic point.
This corollary solves [@LindahlRiveraLetelier2013 Conjecture 1.2] in the affirmative, for generic multiple fixed points of a fixed and small multiplicity, as well as [@KallalKirkpatrick2019 Conjecture 4.3]. In the case $m = 2$, Corollary \[c:generic isolation\] is [@LindahlRiveraLetelier2015 Main Theorem].
In the case $q = 1$, Theorem \[thm:lower-bound\] was shown by Lindahl and the second named author [@LindahlRiveraLetelier2015 Theorem B]. This last result also applies to $p = 2$. In the case ${q = 2}$, and for power series with integer coefficients, Theorem \[thm:lower-bound\] was shown by Lindahl and the first named author [@LindahlNordqvist2018 Theorem A].
Organization {#sec:organziation}
------------
In §\[sec:invariance\] and in Appendix \[app:resit\], we study the residue fixed point index over a field of arbitrary characteristic. Theorem \[thm:closed-formula\] is shown in §\[sec:proof-of-closed-formula\], the invariance of the residue fixed point index under coordinate changes is shown in §\[invarresidue\], and in §\[sec:nf\] we study normal forms. All these results are used in the in the proof of Theorems \[thm:q-ramification\] and \[thm:lower-bound\]. In Appendix \[app:resit\], we study the behavior under iterations of the iterative residue.
In §\[shortproof\] we give a short proof of Theorem \[thm:q-ramification\] that relies on a result of Laubie and Sa[ï]{}ne in [@LaubieSaine1998]. After some preliminaries on lower ramification numbers in §\[sec:lowramif\], this proof is given in §\[sec:proof of ramification\].
In §\[s:periodic points\] we give a self-contained proof of Theorem \[thm:q-ramification\], and the proof of Theorem \[thm:lower-bound\]. We obtain both of these from our main technical result that we state as the “Main Lemma” at the beginning of §\[s:periodic points\]. The proof of this result occupies §\[s:proof of Main Lemma\]. In §\[sec:self-contained-proof\], we use the Main Lemma and the results in §\[sec:invariance\] to obtain more information about the coefficients of the iterates of a wildly ramified power series as in Theorem \[thm:q-ramification\]. This is stated as Proposition \[prop:deltaiterates\], and it implies Theorem \[thm:q-ramification\] as a direct consequence. It is also the main new ingredient in the proof of Theorem \[thm:lower-bound\], which is given in §\[sec:lowerbound\].
In §\[sec:furth-results-exampl\], we gather several examples illustrating our results.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We would like to thank the referees for their valuable comments and corrections that helped improve the exposition of the paper.
The first named author acknowledges support from *Kungliga Vetenskapsakademien*, grant MG2018-0011, for his visit to the second named author at University of Rochester. He would also like to thank the second named author for his hospitality and for providing an excellent working environment during said visit. Finally, the first named author would also like to thank his supervisor Karl-Olof Lindahl for fruitful discussions in the early stages of this project.
The second named author acknowledges partial support from NSF grant DMS-1700291.
The residue fixed point index {#sec:invariance}
=============================
In this section we prove the closed formula (Theorem \[thm:closed-formula\]) and the invariance under coordinate changes of the residue fixed point index. The former is proved in §\[sec:proof-of-closed-formula\], and the latter is stated and proved in §\[invarresidue\]. In §\[sec:nf\] we also use the residue fixed point index to study normal forms of wildly ramified power series.
Given a ring $R$ and elements $a_1, \ldots, a_n$ of $R$, denote by $\langle a_1,\ldots,a_n \rangle$ the ideal generated by $a_1, \ldots, a_n$. Furthermore, denote by $R[[z]]$ the ring of power series with coefficients in $R$ in the variable $z$, and denote by $\operatorname{ord}_z$ the $z$-adic valuation on $R[[z]]$, *i.e.*, for a nonzero $f$ in $R[[z]]$ the valuation $\operatorname{ord}_z(f)$ is the unique integer $j$ such that $f$ is in $z^jR[[z]] \setminus z^{j+1}R[[z]]$, and for $f = 0$ we have $\operatorname{ord}_z(0) {\coloneqq}+\infty$.
Closed formula for the residue fixed point index {#sec:proof-of-closed-formula}
------------------------------------------------
In this section we prove Theorem \[thm:closed-formula\], after the following lemma.
\[sumlittlea\] Let $\operatorname{\mathbb{K}}$ be a field, $q\geq 1$ an integer, and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ of the form . Then $- a_q^{q + 1} \operatorname{index}(f)$ is equal to the coefficient of $z^{q}$ in $$\label{eq:sumlittlea}
\sum_{r=0}^qa_q^r(-1)^{q-r}(a_{q+1}z+\cdots+a_{2q}z^{q})^{q-r}.$$
From the definition, $\operatorname{index}(f)$ is equal to the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $$\label{eq:laurentindf}
\begin{split}
\frac{1}{z-f(z)}
& =
-\frac{1}{a_qz^{q+1}+a_{q+1}z^{q+2}+\cdots+a_{2q}z^{2q+1}+\cdots}
\\ & =
-\frac{1}{a_qz^{q+1}}\cdot\frac{1}{1 + \frac{a_{q+1}}{a_q}z +\frac{a_{q+2}}{a_q}z^2+\cdots}
\\ & =
-\frac{1}{a_q^{q+1}z^{q+1}}\sum_{j=0}^{+ \infty}a_q^{q-j}(-1)^j\left(a_{q+1}z +a_{q+2}z^2+\cdots\right)^j.
\end{split}$$ Thus, $\operatorname{index}(f)$ is equal to the coefficient of $z^q$ in the sum in . Note that for $k\geq 2q+1$, the coefficient $a_k$ does not contribute to the coefficient of $z^q$ in the sum in . Also for $j>q$, the corresponding term in the sum in has no term in $z^q$. Hence, $\operatorname{index}(f)$ is equal to the coefficient of $z^q$ in , as claimed.
In view of Lemma \[sumlittlea\], it is sufficient to compute the coefficient of $z^q$ in . Using the multinomial theorem and regrouping, is equal to $$\begin{gathered}
\sum_{r=0}^qa_q^{r}(-1)^{q-r}\sum_{\substack{(\operatorname{\iota}_1, \ldots, \operatorname{\iota}_q) \in \operatorname{\mathbb{N}}^q \\ \operatorname{\iota}_{1}+\ldots+\operatorname{\iota}_{q} =q-r}}\binom{q-r}{\operatorname{\iota}_{1},\ldots, \operatorname{\iota}_{q}}\prod_{j=1}^{q} (a_{q+j} z^j)^{i_{j}}
\\ =
\sum_{\substack{\operatorname{\iota}\in \operatorname{\mathbb{N}}^{q + 1} \\ |\operatorname{\iota}| = q}}(-1)^{q-\operatorname{\iota}_0}\binom{q-\operatorname{\iota}_0}{\operatorname{\iota}_{1},\ldots,\operatorname{\iota}_{q}} \left(\prod_{j = 0}^q a_{q + j}^{\iota_j}\right)z^{\|\iota\|}.\end{gathered}$$ In the last expression, the term in $z^{q}$ is given by restricting the sum to those multi-indices $\operatorname{\iota}$ satisfying $\| \operatorname{\iota}\| = q$. This proves the theorem.
The residue fixed point index is invariant {#invarresidue}
------------------------------------------
This section is devoted to prove the following proposition.
\[conj\] Let $\operatorname{\mathbb{K}}$ be a field. Then, among power series $f$ with coefficients in $\operatorname{\mathbb{K}}$ and satisfying $f(0)=0$ and $f(z) \neq z$, the residue fixed point index is invariant under coordinate changes. That is, for every power series $\varphi$ with coefficients in $\operatorname{\mathbb{K}}$ such that $\varphi(0) = 0$ and $\varphi'(0) \neq 0$, the power series $\widehat{f} {\coloneqq}\varphi \circ f \circ \varphi^{-1}$ satisfies $$\operatorname{index}(\widehat{f}) = \operatorname{index}(f).$$
The proof of this proposition is given after the following lemma.
\[powersarezero\] Let $\operatorname{\mathbb{K}}$ be a field and $\varphi$ a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $\varphi(0) = 0$ and $\varphi'(0) \neq 0$. Then for every integer $N\geq 1$, the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $$\frac{\varphi'(z)}{\varphi(z)^{N+1}}$$ is zero.
Put $\varphi(z) = {\displaystyle \sum_{j=0}^{+\infty} a_jz^j}$ and for a field automorphism $\sigma$ of $\operatorname{\mathbb{K}}$ put $$\varphi^\sigma(z) {\coloneqq}\sum_{j=0}^{+\infty} \sigma(a_j)z^j.$$
If the characteristic of $\operatorname{\mathbb{K}}$ is zero or if the characteristic of $\operatorname{\mathbb{K}}$ is positive and it does not divide $N$, then the lemma is clear as $$\frac{\varphi'(z)}{\varphi(z)^{N+1}} = \left(-\frac{1}{N}\cdot \frac{1}{\varphi(z)^N}\right)'.$$ So we assume $\operatorname{\mathbb{K}}$ is of characteristic $p>0$ and that $N$ is divisible by $p$. Let $\ell\geq 1$ be the largest integer such that $p^\ell \mid N$, and put $n{\coloneqq}p^{-\ell}N$. Moreover, denote by $\operatorname{Frob}\colon \operatorname{\mathbb{K}}\to\operatorname{\mathbb{K}}$ the Frobenius automorphism, given by $\operatorname{Frob}(z) {\coloneqq}z^p$, and put $\sigma{\coloneqq}\operatorname{Frob}^\ell$. Then we have $$\label{eq:frob}
\frac{\varphi'(z)}{\varphi(z)^{N +1}}
=
\frac{(\varphi^\sigma)'(z^{p^\ell})}{\varphi^\sigma(z^{p^\ell})^{n+1}}
\cdot \left( \frac{\varphi^\sigma(z^{p^\ell})}{(\varphi^\sigma)'(z^{p^\ell})} \cdot \frac{\varphi'(z)}{\varphi(z)} \right).$$ Since $n$ is not divisible by $p$, the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $\frac{(\varphi^{\sigma})'(z)}{(\varphi^{\sigma}(z))^{n+1}}$ is zero. So the coefficient of $\frac{1}{z^{p^\ell}}$ in the Laurent series expansion about $0$ of $\frac{(\varphi^\sigma)'(z^{p^\ell})}{\varphi^\sigma(z^{p^\ell})^{n+1}}$ is zero. Together with $$\operatorname{ord}_z\left(\frac{\varphi^{\sigma}(z^{p^\ell})}{(\varphi^\sigma)'(z^{p^\ell})}\cdot \frac{\varphi'(z)}{\varphi(z)}\right) = p^\ell-1,$$ this implies that the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $\frac{\varphi'(z)}{\varphi(z)^{N +1}}$ is zero, which is the desired assertion.
If $f'(0) \neq 1$, then $\operatorname{index}(f)$ is equal to $\frac{1}{1-f'(0)}$, which is easily seen to be invariant under coordinate changes. Assume $f'(0) = 1$, and put $$\Delta(z) {\coloneqq}f(z)-z
\text{ and }
q{\coloneqq}\operatorname{ord}_z(\Delta(z))-1.$$ Our hypothesis $f(z)\neq z$ implies that $q$ is finite and our assumption $f'(0)=1$ implies that $q\geq 1$.
Let $\varphi$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $\varphi(0)=0$ and $\varphi'(0) \neq 0$, and put $$\widehat{f}{\coloneqq}\varphi^{-1}\circ f \circ \varphi
\text{ and }
\widehat{\Delta}(z) {\coloneqq}\widehat{f}(z)-z.$$ Clearly $\widehat{f}'(0) = 1$, so $\operatorname{ord}_z(\widehat{\Delta}(z))\geq 2$. Moreover, $$\label{DeltaH}
\begin{split}
\Delta \circ \varphi(z)
& =
\varphi(\widehat{f}(z)) - \varphi(z)
\\ & =
\varphi(z + \widehat{\Delta}(z)) - \varphi(z)
\\ & \equiv
\varphi'(z)\widehat{\Delta}(z) \mod \langle \widehat{\Delta}(z)^2\rangle.
\end{split}$$ Since $\operatorname{ord}_z(\Delta) = q+1$ and $\operatorname{ord}_z(\varphi') = 0$, we conclude that $$\operatorname{ord}_z(\Delta\circ \varphi) = q+1
\text{ and }
{\operatorname{ord}_z(\varphi'\cdot \widehat{\Delta})} = \operatorname{ord}_z(\widehat{\Delta}).$$ On the other hand, by (\[DeltaH\]) we have $\operatorname{ord}_z(\Delta\circ\varphi-\varphi'\cdot\widehat{\Delta})\geq 2\operatorname{ord}_z(\widehat{\Delta})$ and therefore $$\operatorname{ord}_z(\widehat{\Delta}) = \operatorname{ord}_z(\Delta \circ \varphi) = q+1.$$ Using (\[DeltaH\]) again we obtain $$\Delta \circ \varphi \equiv \varphi'\cdot \widehat{\Delta} + \langle z^{2q+2} \rangle,$$ and conclude that $\operatorname{index}(\widehat{f})$ is equal to the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of $$\frac{\varphi'}{\Delta\circ \varphi}.$$
Putting $$\left(\frac{1}{\Delta}\right)(z) {\coloneqq}\sum_{i = -(q+1)}^{+\infty} a_i z^i,$$ we have $$\left(\frac{\varphi'}{\Delta\circ \varphi}\right)(z)
=
\sum_{N = 0}^{q} a_{- (N + 1)} \frac{\varphi'(z)}{\varphi(z)^{N + 1}}
+
\sum_{i = 0}^{+\infty} a_i \varphi(z)^i \varphi'(z).$$ By Lemma \[powersarezero\], the coefficient of $\frac{1}{z}$ in the Laurent series expansion about $0$ of the right-hand side is equal to that of $a_{-1} \frac{\varphi'(z)}{\varphi(z)}$, which is clearly equal to $a_{-1}$. This completes the proof of the proposition.
Normal forms in positive characteristic {#sec:nf}
---------------------------------------
Let $\operatorname{\mathbb{K}}$ be a field and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $q {\coloneqq}\operatorname{mult}(f) - 1$ is finite and satisfies $q \ge 1$. In the case of $\operatorname{\mathbb{K}}=\operatorname{\mathbb{C}}$, or more generally if $\operatorname{\mathbb{K}}$ is of characteristic zero, there exists a (formal) power series conjugating $f$ to the polynomial $$\label{eq:22}
z(1+z^q + \operatorname{index}(f)z^{2q}).$$ When $\operatorname{\mathbb{K}}$ is of characteristic zero, this polynomial is called the *normal form of $f$*.
This statement is false if $\operatorname{\mathbb{K}}$ is of positive characteristic. Our goal in this section is to prove the following proposition giving a sufficient condition for $f$ to have the same normal form up to a high order.
\[nf\] Let $p$ be a prime number and $\operatorname{\mathbb{K}}$ a field of characteristic $p$. Moreover, let $q$ be in $\{1, \ldots, p - 1 \}$, and let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ satisfying $\operatorname{mult}(f) = q + 1$. Then, $f$ is conjugated to a power series with coefficients in a finite extension of $\operatorname{\mathbb{K}}$, of the form $$\label{eq:nff}
z(1 + z^q + \operatorname{index}(f)z^{2q})\mod \langle z^{2q + p + 1} \rangle.$$
The proof of this proposition is given after the following lemma.
\[removeterms\] Let $\operatorname{\mathbb{K}}$ be a field, $q \ge 1$ an integer, and $f$ a power series with coefficients in $\operatorname{\mathbb{K}}$ of the form $$f(z)
=
z\left(1 + \sum_{j=q}^{+\infty} a_jz^j\right), \text{ with } a_q \neq 0.$$ Then, for every integer $k \ge 1$ such that $a_{q + k} \neq 0$ and $k\neq q$ in $\operatorname{\mathbb{K}}$, there is $c$ in $\operatorname{\mathbb{K}}$ such that for the polynomial $\varphi(z) {\coloneqq}z(1 + c z^k)$, we have $$\varphi \circ f \circ \varphi^{-1}(z) \equiv z(1 + a_qz^q + \cdots + a_{q+k-1}z^{q+k-1}) \mod \langle z^{q+k+2}\rangle.$$
Let $c$ be a constant in $\operatorname{\mathbb{K}}$ to be chosen later, and put $$\varphi(z) {\coloneqq}z(1 + cz^k)
\text{ and }
\widehat{f}(z)
{\coloneqq}\varphi \circ f \circ \varphi^{-1}(z)
= z \left( 1 + \sum_{j = q}^{+ \infty} \widehat{a}_jz^j \right).$$ Then we find $$\begin{split}
\varphi \circ f(z)
& \equiv
z (1 + a_qz^q + \cdots + a_{q+k}z^{q+k})(1 + cz^k(1 + a_qz^q)^k)
\mod \langle z^{q+k+2} \rangle
\\ & \equiv
z(1 + cz^k + a_qz^q + \cdots + a_{q+k-1}z^{q+k-1}
\\&\qquad
+ ((k+1)ca_q + a_{q+k})z^{q+k}) \mod \langle z^{q+k+2} \rangle,
\end{split}$$ and $$\begin{split}
\widehat{f} \circ \varphi(z)
& \equiv
z (1 + cz^k)(1 + \widehat{a}_qz^q (1 + cz^k)^q + \widehat{a}_{q + 1} z^{q + 1} + \cdots + \widehat{a}_{q+k}z^{q+k})
\\ &
\quad \mod \langle z^{q+k+2} \rangle
\\ & \equiv
z(1 + cz^k + \widehat{a}_qz^q +\cdots + \widehat{a}_{q+k-1}z^{q+k-1}
\\&\qquad
+ ((q+1)c\widehat{a}_q + \widehat{a}_{q + k})z^{q+k}) \mod \langle z^{q+k+2} \rangle.
\end{split}$$ Equating both expression yields $$a_q = \widehat{a}_q,\ldots, a_{q+k-1} = \widehat{a}_{q+k-1},$$ and $$\widehat{a}_{q + k} = (k - q)c a_q + a_{q+k}.$$ By our assumption $k\neq q$ in $\operatorname{\mathbb{K}}$, we can take $c = - \frac{a_{q+k}}{a_q(k-q)}$ to obtain $\widehat{a}_{q + k} = 0$.
Denote by $a \neq 0$ the coefficient of $z^{q+1}$ in $f$, and let $\gamma$ in a finite extension of $\operatorname{\mathbb{K}}$ be such that $\gamma^q = a^{-1}$. Note that the power series $\widehat{f}(z) {\coloneqq}\gamma^{-1} f(\gamma z)$ satisfies $\operatorname{mult}(\widehat{f}) = q + 1$ and that the coefficient of $z^{q + 1}$ in $\widehat{f}$ is equal to $1$.
Since by assumption $q$ is in $\{1,\ldots, p - 1\}$, we can apply Lemma \[removeterms\] successively with $k = 1, \ldots, q - 1$, to obtain that there is a polynomial $\varphi$ with coefficients in $\operatorname{\mathbb{K}}[\gamma]$, such that $\varphi(0) = 0$, $\varphi'(0) = 1$, and $$g(z)
{\coloneqq}\varphi \circ \widehat{f} \circ \varphi^{-1}(z)
\equiv
z (1 + z^q) \mod \langle z^{2q + 1} \rangle.$$ Note that by Theorem \[thm:closed-formula\] the coefficient of $z^{2q + 1}$ in $g$ is equal to $\operatorname{index}(g)$ and by Proposition \[conj\] we have $\operatorname{index}(g) = \operatorname{index}(\widehat{f}) = \operatorname{index}(f)$. Thus, $$g(z)
\equiv
z (1 + z^q + \operatorname{index}(f) z^{2q}) \mod \langle z^{2q + 2} \rangle.$$ Finally, we apply Lemma \[removeterms\] successively with $k = q + 1, \ldots, q + p - 1$, to obtain that there is a polynomial $\phi$ with coefficients in $\operatorname{\mathbb{K}}[\gamma]$, such that $\phi(0) = 0$, $\phi'(0) = 1$, and $$\phi \circ g \circ \phi^{-1}(z)
\equiv
z (1 + z^q + \operatorname{index}(f) z^{2q}) \mod \langle z^{2q + p + 1} \rangle.
\qedhere$$
$q$-Ramified power series {#shortproof}
=========================
After some preliminaries on lower ramification numbers in §\[sec:lowramif\], in §\[sec:proof of ramification\] we give a short proof of Theorem \[thm:q-ramification\] that relies on a result of Laubie and Sa[ï]{}ne in [@LaubieSaine1998]. See §\[sec:self-contained-proof\] for a self-contained proof of Theorem \[thm:q-ramification\].
Lower ramification numbers {#sec:lowramif}
--------------------------
In this section we fix a prime number $p$ and a field $\operatorname{\mathbb{K}}$ of characteristic $p$. Recall that for a power series $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ and an integer $n \ge 1$, the lower ramification number $i_n(f)$ of $f$ is $$i_n(f) = \operatorname{mult}(f^{p^n})-1.$$ Lower ramification numbers have been studied by several authors, *e.g.*, [@Sen1969; @Keating1992; @LaubieSaine1998; @LaubieMovahhediSalinier2002]. A central theorem of Sen [@Sen1969 Theorem 1] states that if for some $n\geq0$ we have $i_n(f) < +\infty$, then $$i_n(f) \equiv i_{n-1}(f) \pmod{p^n}.$$ The following consequence of Sen’s theorem shows that for $q$ in $\{1, \ldots, p - 1 \}$, a $q$-ramified power series can be thought of as minimal in the sense that for every integer $n$ the lower ramification number $i_n(f)$ is least possible.
\[prop:qramifisminramif\] Let $p$ be a prime number and $\operatorname{\mathbb{K}}$ a field of characteristic $p$. Then for every $q$ in $\{1,\ldots,p-1\}$, and every power series $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ satisfying $\operatorname{mult}(f) = q+1$, we have for every integer $n \ge 1$ $$\label{ineq}
i_n(f) \geq q(1 + p + \cdots + p^n).$$
The proof of this proposition is given after the following lemma. To state this lemma, we introduce some notation. Let $R$ be a ring, and $f$ a power series in $R[[z]]$ of the form $f(z) \equiv z \mod \langle z^2 \rangle$. Following [@RiveraLetelier2003 *Exemple* 3.19] and [@LindahlRiveraLetelier2013], define recursively for every integer $m\geq 0$ the power series $\Delta_m$ by $$\label{eq:27}
\Delta_0(z) {\coloneqq}z,$$ and for $m \ge 1$ by $$\label{eq:23}
\Delta_m(z)
{\coloneqq}\Delta_{m-1}(f(z)) - \Delta_{m-1}(z).$$ If $R$ is of characteristic zero, then for every prime number $p$ a direct computation shows that we have $$\label{eq:7}
\Delta_p(z) \equiv f^p(z)-z \mod \langle p \rangle.$$ In the case $R$ is of characteristic $p$, we have $\Delta_p(z) = f^p(z)-z$.
\[lemma:delta\] Let $p$ be a prime number and $\operatorname{\mathbb{K}}$ a field of characteristic $p$. Given a wildly ramified power series $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$, let $( \Delta_m )_{m = 0}^{+\infty}$ be as above. Then for every integer $m \ge 1$ we have $$\label{eq:24}
\operatorname{ord}_\zeta(\Delta_m)-\operatorname{ord}_\zeta(\Delta_{m-1})
\geq
\operatorname{ord}_\zeta(\Delta_1)-1.$$
Put $q {\coloneqq}{\operatorname{ord}_\zeta(\Delta_1)-1}$, $f(\zeta) = \zeta\left(1 + \sum_{i = q}^{+ \infty} b_i\zeta^i\right)$, $r {\coloneqq}\operatorname{ord}_{\zeta}(\Delta_m)$, and $\Delta_m(\zeta)=\sum_{i = r}^{+ \infty} a_i\zeta^i$. Then $$\Delta_{m+1}(\zeta)
=
\sum_{i =r}^{+ \infty} a_i\zeta^i\left[(1+b_q\zeta^q+\cdots)^i - 1\right],$$ and therefore $\operatorname{ord}_\zeta(\Delta_{m+1}) \ge r+q$.
We prove by induction in $n$. To prove for $n = 1$, let $(\Delta_m)_{m = 0}^{+ \infty}$ be as in and . Then for every integer $m \ge 1$ we have $\operatorname{ord}_\zeta(\Delta_m) - \operatorname{ord}_\zeta(\Delta_{m-1}) \geq q$ by Lemma \[lemma:delta\]. An induction argument combined with gives $$i_1(f) = \operatorname{ord}_\zeta(\Delta_p) - 1 \geq qp = p i_0(f).$$ But by Sen’s theorem we have $i_1(f) \equiv i_0(f) \pmod{p}$, so $$\label{firstinduc}
i_1(f) \geq qp + q.$$ This proves for $n = 1$.
Let $n \ge 1$ be an integer for which holds, and put $g(\zeta) {\coloneqq}f^{p^n}(\zeta)$. Let $(\widehat{\Delta}_m)_{m = 0}^{+ \infty}$ be the sequence $(\Delta_m)_{m = 0}^{+ \infty}$ given by and with $f$ replaced by $g$. Then by Lemma \[lemma:delta\] for every integer $m \ge 1$ we have $$\operatorname{ord}_\zeta(\widehat{\Delta}_m)-\operatorname{ord}_\zeta(\widehat{\Delta}_{m-1})
\geq
\operatorname{ord}_{\zeta}(\widehat{\Delta}_1)
=
i_0(g).$$ An induction argument together with , implies $$\label{eq:26}
i_{n + 1}(f)
=
i_1(g)
=
\operatorname{ord}_\zeta(\widehat{\Delta}_p) - 1
\geq
p i_0(g)
=
p i_n(f).$$ If the inequality in our induction assumption is strict, then we have $$i_{n + 1}(f)
\ge
p + p q (1 + p + \cdots + p^n)
>
q(1 + p + \cdots + p^{n + 1}).$$ If equality holds in , then by Sen’s theorem we have $$i_{n + 1}(f) \equiv q(1 + p + \cdots + p^n) \pmod{p^{n+1}}.$$ Combined with , this implies $$i_{n+1}(f)
\geq
q + p q (1 + p + \cdots + p^n)
=
q(1 + p + \cdots + p^{n + 1}).$$ In all the cases we obtain with $n$ replaced by $n + 1$. This completes the proof of the induction step, and of the the proposition.
Proof of Theorem \[thm:q-ramification\] {#sec:proof of ramification}
---------------------------------------
In the proof of Theorem \[thm:q-ramification\] we use the following result of Laubie and Sa[ï]{}ne.
\[laubiecorr\] Let $p$ be a prime number, $\operatorname{\mathbb{K}}$ a field of characteristic $p$, and $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ such that $f(0)=0$ and $f'(0)=1$. If $$p\nmid i_0(f)
\text{ and }
i_1(f) < (p^2-p+1)i_0(f),$$ then for every integer $n\geq 1$ we have $$i_n(f) = i_0(f) + (1 + p + \cdots + p^n)(i_1(f)-i_0(f)).$$
In view of this result, the proof of Theorem \[thm:q-ramification\] reduces to show that for $q$ in $\{1, \ldots, p - 1 \}$ and $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ satisfying $i_0(f) = q$, the conditions $$i_1(f) = q (p + 1)
\text{ and }
\operatorname{r\acute{e}sit}(f) \neq 0$$ are equivalent. The following is the key ingredient, together with Proposition \[nf\] and the invariance of the residue fixed point index under coordinate changes shown in §\[sec:invariance\].
\[prop:deltaiteratesshort\] Let $p$ be an odd prime number and consider the rings $$\operatorname{\mathbb{Z}}_{(p)} {\coloneqq}\left\{\frac{m}{n} \in \operatorname{\mathbb{Q}}: m, n \in \operatorname{\mathbb{Z}}, p\nmid n \right\},$$ $$F_1 {\coloneqq}\operatorname{\mathbb{Z}}_{(p)}[x_0, x_1],
\text { and }
F_\infty {\coloneqq}\operatorname{\mathbb{Z}}_{(p)}[x_0, x_1,x_2,\ldots].$$ Then for each integer $q \ge 1$ not divisible by $p$, the power series $\widehat{f}$ in $F_\infty[[\zeta]]$ defined by $$\widehat{f}(\zeta)
{\coloneqq}\zeta\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + \zeta^{2q}\sum_{i=1}^{+ \infty} x_{i+1}\zeta^i\right),$$ satisfies $$\widehat{f}^p(\zeta)
\equiv
\zeta \left(1 + x_0^{p - 1} \left( x_0^2 \frac{q + 1}{2} - x_1 \right) \zeta^{q(p + 1)}\right) \mod \langle p, \zeta^{q(p + 1) + 2} \rangle.$$
The proof of Theorem \[thm:q-ramification\] is given at the end of this section, after the proof of this proposition. To prove this proposition we use the strategy introduced in [@RiveraLetelier2003 *Exemple* 3.19] and [@LindahlRiveraLetelier2013], using and . We also use the following elementary lemma.
\[wilson\] Let $p$ be an odd prime number, $a$ and $b$ in $\operatorname{\mathbb{F}}_p$ such that $a\neq 0$, and let $w \colon \operatorname{\mathbb{F}}_p \to \operatorname{\mathbb{F}}_p$ be defined by $w(n) {\coloneqq}an +b$. Denoting $s' {\coloneqq}-a^{-1}b$, we have $$\prod_{s \in \operatorname{\mathbb{F}}_p \setminus \{s'\}} w(s) = -1
\text{ and }
\sum_{s \in \operatorname{\mathbb{F}}_p \setminus\{s'\}} \frac{1}{w(s)} = 0.$$
We use the fact that the nonconstant affine map $w$ is a bijection of $\operatorname{\mathbb{F}}_p$. Together with Wilson’s theorem this implies the first assertion. The second assertion follows from the fact that, since $p$ is odd, the sum of all nonzero elements in $\operatorname{\mathbb{F}}_p$ is $0$.
Let $(\Delta_m)_{m = 0}^{+ \infty}$ be given by and . For each integer $m \ge 1$ define $\alpha_m$, and $\beta_m$ in the ring $F_1 {\coloneqq}\operatorname{\mathbb{Z}}_{(p)}[x_0, x_1]$ by the recursive relations $$\begin{aligned}
\alpha_{m+1} & {\coloneqq}x_0 (qm+1)\alpha_m, \label{receq1short}
\\
\beta_{m+1} & {\coloneqq}\left[x_0^2 \binom{qm+1}{2} + x_1(qm+1)\right]\alpha_m + x_0 (q(m+1)+1)\beta_m,
\label{receq2short}\end{aligned}$$ with initial conditions $\alpha_1 {\coloneqq}x_0$ and $\beta_1 {\coloneqq}x_1$. We prove by induction that for every integer $m \ge 1$ we have $$\label{deltaclaimshort}
\Delta_m(\zeta)
\equiv
\alpha_m\zeta^{qm+1} + \beta_m\zeta^{q(m+1)+1} \mod \langle \zeta^{q(m + 1) + 2} \rangle.$$ For $m=1$ this holds by definition. Assume further that it is valid for some $m\geq1$. Then $$\begin{split}
\Delta_{m+1}(\zeta)
& = \Delta_m(\widehat{f}(\zeta)) - \Delta_m(\zeta)
\\ &\equiv \alpha_m\zeta^{qm+1}\left[\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + \cdots\right)^{qm+1} - 1\right]\\
&\qquad + \beta_m\zeta^{q(m+1)+1}\left[\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + \cdots\right)^{q(m+1)+1}-1\right]
\\ & \qquad
\mod \langle \zeta^{q(m+2)+2} \rangle
\\
&\equiv \alpha_m\left[ \zeta^{q(m+1)+1}x_0 (qm+1) + \zeta^{q(m+2)+1}\left( x_0^2 \binom{qm+1}{2} + x_1 (qm+1)\right)\right] \\
&\qquad+ \beta_m\zeta^{q(m+2)+1}x_0 (q(m+1)+1) \mod \langle \zeta^{q(m+2)+2} \rangle.
\end{split}$$ In view of and , this proves the induction step and .
By and , to prove the proposition it is sufficient to prove $$\label{eq:5short}
\alpha_p \equiv 0 \mod p F_1
\text{ and }
\beta_p \equiv x_0^{p - 1} \left( x_0^2 \frac{q + 1}{2} - x_1 \right) \mod p F_1.$$ We do this by solving explicitly the linear recurrences described in , and . By telescoping , we obtain for every $m \ge 1$ the solution $$\label{alpham}
\alpha_m = x_0^m \prod_{j=1}^{m-1}(qj+1).$$ Taking $m = p$ we obtain the first congruence in .
On the other hand, inserting in yields $$\beta_{m+1} = \left(x_0^2 \frac{qm}{2} + x_1 \right) x_0^m \prod_{j=1}^m (qj+1) + x_0 (q(m+1)+1)\beta_m.$$ Noting that for every $j \ge 0$ we have $qj + 1 > 0$, we utilize the substitution $$\beta^*_m {\coloneqq}\beta_m \bigg/ \left( x_0^{m - 1} \prod_{j=1}^m (qj+1) \right),$$ which yields $$\beta^*_{m+1}
=
\beta^*_{m} + \left(x_0^2 \frac{qm}{2} + x_1 \right)\frac{1}{q(m+1)+1}.$$ Using $\beta^*_1 = \frac{x_1}{q + 1}$, we obtain inductively for every $m \ge 1$ $$\beta^*_m
=
\sum_{r=1}^{m} \left(x_0^2 \frac{q(r - 1)}{2} + x_1 \right)\frac{1}{qr + 1}.$$ Equivalently, $$\label{betam}
\beta_m
=
x_0^{m - 1} \sum_{r=1}^{m}\left[ \left( x_0^2 \frac{q(r - 1)}{2} + x_1 \right) \prod_{j \in \{1, \ldots, m\} \setminus \{ r \}} (qj+1) \right].$$ When $m = p$ every term in the sum above contains a factor $p$, except for the unique $r$ in $\{1, \ldots, p \}$ such that $qr \equiv -1 \pmod{p}$. Denote by $r_0$ this value of $r$. Then by Lemma \[wilson\], we have $$\begin{split}
\beta_p
&\equiv
x_0^{p - 1} \left(\frac{ x_0^2 q(r_0 - 1)}{2} + x_1\right) \prod_{j \in \{ 1, \ldots, p \} \setminus \{ r_0 \}} (qj+1) \mod p F_1
\\ &\equiv
x_0^{p - 1} \left(x_0^2 \frac{q+1}{2} - x_1\right) \mod p F_1.
\end{split}$$ This proves the second congruence in and thus the proposition.
By Proposition \[nf\] and our hypothesis that $q$ is in $\{1, \ldots, p-1\}$, we have that $f$ is conjugated to a power series $g$ in $\operatorname{\mathbb{K}}[[\zeta]]$ of the form $$g(\zeta) \equiv \zeta(1 + \zeta^q + \operatorname{index}(f)\zeta^{2q}) \mod \langle \zeta^{3q+2} \rangle.$$ Since $$i_0(g) = i_0(f) = q
\text{ and }
i_1(g) = i_1(f),$$ by Proposition \[laubiecorr\] the series $f$ is $q$-ramified if and only if $i_1(g) = q (p + 1)$.
Let $\operatorname{\mathbb{Z}}_{(p)}$ and $F_\infty$ be as in Proposition \[prop:deltaiteratesshort\]. Moreover, let $h \colon F_\infty \to \operatorname{\mathbb{K}}$ be the unique ring homomorphism extending the reduction map $\operatorname{\mathbb{Z}}_{(p)} \to \operatorname{\mathbb{F}}_p$, such that $h(x_1) = \operatorname{index}(f)$ and such that for every $i \ge 2$ the element $h(x_i)$ of $\operatorname{\mathbb{K}}$ is the coefficient of $\zeta^{2q + i}$ in $g$. Then $h$ extends to a ring homomorphism $F_{\infty}[[ \zeta ]] \to \operatorname{\mathbb{K}}[[ \zeta ]]$ that maps $\widehat{f}$ to $g$. So, Proposition \[prop:deltaiteratesshort\] implies $$g^{p}(\zeta) - \zeta
\equiv
\operatorname{r\acute{e}sit}(f) \zeta^{q(p+1)+1} \mod \langle \zeta^{q(p+1)+2} \rangle.$$ This proves that $i_1(g) = q(p + 1)$ if and only if $\operatorname{r\acute{e}sit}(f) \neq 0$ and completes the proof of the theorem.
Periodic points of $q$-ramified power series {#s:periodic points}
============================================
In this section we give a self-contained proof of Theorem \[thm:q-ramification\], and the proof of Theorem \[thm:lower-bound\]. In doing so, we obtain more information about the coefficients of the iterates of a wildly ramified power series as in Theorem \[thm:q-ramification\] (Proposition \[prop:deltaiterates\] in §\[sec:self-contained-proof\]). This extra information is used to prove Theorem \[thm:lower-bound\] in §\[sec:lowerbound\].
The main ingredients in the proofs of Theorems \[thm:q-ramification\] and \[thm:lower-bound\] are the results on the residue fixed point index in §\[sec:invariance\], and the following result that is proved in §\[s:proof of Main Lemma\].
Let $p$ be an odd prime number, and let $\operatorname{\mathbb{Z}}_{(p)}$, $F_1$ and $F_{\infty}$ be the rings defined in Proposition \[prop:deltaiteratesshort\]. Moreover, let $q \ge 1$ be an integer that is not divisible by $p$, and $\ell \ge 1$ an integer satisfying $$\ell \equiv q \pmod{p},
\text{ and }
\ell \le p - 1
\text{ or }
2\ell + 1 \le q.$$ Then the power series $\widehat{f}$ in $F_\infty[[\zeta]]$ defined by $$\widehat{f}(\zeta) {\coloneqq}\zeta\left(1 + x_0\zeta^{q} + x_1\zeta^{q+\ell} + \zeta^{q+2\ell}\sum_{i=1}^\infty x_{i+1}\zeta^{i}\right),$$ satisfies the following property: There are $\beta$ and $\gamma$ in $F_1$ such that $$\begin{aligned}
\label{e:main beta}
\beta
& \equiv
\begin{cases}
x_0^{p - 1} \left( x_0^2 \frac{q + 1}{2} - x_1 \right) \mod p F_1
& \text{if $q \le p - 1$};
\\
- x_0^{p - 1} x_1 \mod p F_1
& \text{if $q \ge p + 1$},
\end{cases}
\\
\label{e:main gamma}
\gamma
& \equiv
\begin{cases}
- x_0^{p - 2} \left(x_0^2 \frac{q + 1}{2} - x_1 \right)^2 \mod p F_1
& \text{if $q \le p - 1$};
\\
- x_0^{p - 2} x_1^2 \mod p F_1
& \text{if $q \ge p + 1$},
\end{cases}
\intertext{ and }
\widehat{f}^p(\zeta)
& \equiv
\zeta \left(1 + \beta \zeta^{qp + \ell} + \gamma \zeta^{qp + 2\ell} \right) \mod \langle p, \zeta^{qp + 2\ell + 2} \rangle.\end{aligned}$$
Self-contained proof of Theorem \[thm:q-ramification\] {#sec:self-contained-proof}
------------------------------------------------------
The goal of this section is to deduce the following proposition from the Main Lemma, which is a more precise version of Theorem \[thm:q-ramification\]. It is also one of the main ingredients of the proof of Theorem \[thm:lower-bound\], which is given in §\[sec:lowerbound\].
\[prop:deltaiterates\] Let $p$ be an odd prime number and $\operatorname{\mathbb{K}}$ a field of characteristic $p$. Furthermore, let $q$ be in $\{1, \ldots, p-1 \}$, let $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ be of the form $$f(\zeta)
\equiv
\zeta(1 + a_0\zeta^q + a_1\zeta^{2q}) \mod \langle \zeta^{3q+2}\rangle, \text{ with } a_0\neq 0,$$ and for each integer $n\geq1$, put $$\begin{aligned}
\chi_n
& {\coloneqq}a_0^{\frac{p^{n+1}-1}{p-1}}\left(\frac{q+1}{2}-\frac{a_1}{a_0^2}\right)^{\frac{p^n-1}{p-1}},
\intertext{ and }
\psi_n
& {\coloneqq}-a_0^{\frac{p^{n+1}-1}{p-1}+1}\left(\frac{q+1}{2}-\frac{a_1}{a_0^2}\right)^{\frac{p^n-1}{p-1}+1}.\end{aligned}$$ Then we have $$f^{p^n}(\zeta) - \zeta
\equiv
\chi_n\zeta^{q\frac{p^{n+1}-1}{p-1}+1} + \psi_n\zeta^{q\frac{p^{n+1}-1}{p-1}+q+1} \mod \langle \zeta^{q\frac{p^{n+1}-1}{p-1}+q+2} \rangle.$$ In particular, $f$ is $q$-ramified if and only if $$\operatorname{r\acute{e}sit}(f) = \frac{q + 1}{2} - \frac{a_1}{a_0^2} \neq 0.$$
The proof of Proposition \[prop:deltaiterates\] is given after the following lemma.
\[elimination\] Let $p$ be an odd prime number, $q$ in $\{1,\ldots,p-1\}$, and $d \ge 1$ an integer satisfying $d\equiv 1 \pmod{p}$. Furthermore, let $\operatorname{\mathbb{K}}$ be a field of characteristic $p$ and let $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$ be of the form $$f(\zeta)
\equiv
\zeta\left(1+ a_0\zeta^{qd} + a_1\zeta^{q(d+1)}\right) \mod \langle \zeta^{q(d+1)+2} \rangle, \text{ with } a_0\neq 0.$$ Then there is a polynomial $\varphi$ with coefficients in $\operatorname{\mathbb{K}}$ such that $\operatorname{mult}(\varphi) \ge {q + 2}$, and such that $\varphi$ conjugates $f$ to a power series $g$ satisfying $$\begin{aligned}
\label{eq:12}
g(\zeta)
& \equiv
\zeta\left(1 + a_0\zeta^{qd} + a_1\zeta^{q(d+1)}\right) \mod \langle \zeta^{q(d+1) + p + 1} \rangle,
\intertext{ and }
\notag
g^p(\zeta)
& \equiv f^p(\zeta) \mod \langle \zeta^{i_1(f)+ q + 2} \rangle.
\end{aligned}$$
Noting that $qd \equiv q \pmod{p}$, we can apply Lemma \[removeterms\] successively with $q$ replaced by $qd$, and with $$k
=
q + 1, \ldots, q + p - 1,$$ to obtain a polynomial $\varphi$ satisfying $\operatorname{mult}(\varphi) \ge q + 2$, such that $g {\coloneqq}\varphi \circ f \circ \varphi^{-1}$ satisfies .
To prove the second assertion, note that $\varphi$ also conjugates $f^p$ to $g^p$, so by Lemma \[removeterms\] $$i_1(f) = i_1(g)
\text{ and }
f^p(\zeta) \equiv g^p(\zeta) \mod \langle \zeta^{i_1(f) + \operatorname{mult}(\varphi)} \rangle.$$ The desired assertion follows from the inequality $\operatorname{mult}(\varphi) \ge q + 2$. This completes the proof of the lemma.
The last assertion is a direct consequence of the first and of .
To prove the first assertion, for each integer $n\geq 0$ put $d_n {\coloneqq}1 + p + \cdots + p^n$, and note that $$d_n \equiv 1 \pmod{p},
\text{ and }
d_np+1 = d_{n + 1}.$$ We first prove by induction that for every integer $n \ge 0$ there are $\chi_n$ and $\psi_n$ in $\operatorname{\mathbb{K}}$, such that $$\label{eq:13}
f^{p^n}(\zeta)
\equiv
\zeta\left(1 + \chi_{n}\zeta^{qd_n} + \psi_{n}\zeta^{q(d_n+1)}\right) \mod \langle \zeta^{q(d_n+1)+2} \rangle.$$ The case $n = 0$ is trivial, with $$\label{eq:10}
\chi_0 = a_0
\text{ and }
\psi_0 = a_1.$$ Let $n \ge 0$ be a given integer, and assume the desired assertion is true for $n$. By Lemma \[elimination\] there is a power series $g$ with coefficients in $\operatorname{\mathbb{K}}$ such that $$\begin{aligned}
\notag
g(\zeta)
& \equiv
\zeta\left(1 + \chi_{n}\zeta^{qd_n} + \psi_{n}\zeta^{q(d_n+1)}\right) \mod \langle \zeta^{q(d_n+2)+2} \rangle,
\intertext{ and }
\label{eq:14}
g^p(\zeta)
& \equiv
f^{p^{n + 1}}(\zeta) \mod \langle \zeta^{i_{n + 1}(f) + q + 2} \rangle.\end{aligned}$$ Define $\operatorname{\mathbb{Z}}_{(p)}, F_1$ and $F_\infty$ as in Proposition \[prop:deltaiteratesshort\]. Moreover, let $\widehat{g}$ in $F_\infty[[\zeta]]$ be of the form $$\widehat{g}(\zeta)
{\coloneqq}\zeta\left(1+x_0 \zeta^{qd_n} + x_1 \zeta^{q(d_n + 1)} + \zeta^{q(d_n + 2)}\sum_{j=1}^{+ \infty}x_{j+1} \zeta^j\right),$$ let $h\colon F_\infty \to \operatorname{\mathbb{K}}$ be the unique ring homomorphism extending the reduction map $\operatorname{\mathbb{Z}}_{(p)} \to \operatorname{\mathbb{F}}_p$, such that $h(x_0) = \chi_n$, $h(x_1) = \psi_n$, and such that for every $i\geq 2$ the element $h(x_i)$ of $\operatorname{\mathbb{K}}$ is the coefficient of $\zeta^{q(d_n+2)+i}$ in $\widehat{g}$. Then $h$ extends to a ring homomorphism $F_\infty[[\zeta]] \to \operatorname{\mathbb{K}}[[\zeta]]$ that maps $\widehat{g}$ to $g$. In the case $n = 0$, note that $\widehat{f}$ in the Main Lemma is equal to $\widehat{g}$, so $$\begin{gathered}
g^p(\zeta)
\equiv
\zeta \left( 1 + \chi_0^{p+1}\left(\frac{q+1}{2} - \frac{\psi_0}{\chi_0^2}\right)\zeta^{q(p+1)}
\right. \\ \left.
- \chi_0^{p+2} \left(\frac{q+1}{2}-\frac{\psi_0}{\chi_0^2}\right)^2\zeta^{q(p+2)} \right) \mod \langle \zeta^{q(p+2)+2} \rangle.
\end{gathered}$$ Together with with $n = 0$, this implies $$i_1(f) = i_1(g) \ge q(p + 1) = qd_1,$$ and with $n = 1$, $$\label{eq:11}
\chi_1
{\coloneqq}\chi_0^{p+1}\left(\frac{q+1}{2} - \frac{\psi_0}{\chi_0^2} \right)
\text{ and }
\psi_1
{\coloneqq}- \chi_0^{p+2} \left(\frac{q+1}{2}-\frac{\psi_0}{\chi_0^2}\right)^2.$$ In the case $n \ge 1$, the Main Lemma with $q$ replaced by $qd_n$ and $\ell$ replaced by $q$, implies $$g^p(\zeta)
\equiv
\zeta \left( 1 - \chi_n^{p - 1} \psi_n \zeta^{q(d_np+1)} - \chi_n^{p - 2} \psi_n^2 \zeta^{q(d_n p + 2)}\right)
\mod \langle \zeta^{q(d_n p+2)+2} \rangle.$$ Together with this implies $$i_{n + 1}(f) = i_1(g) \ge q(d_n p + 1) = q d_{n + 1}$$ and with $$\label{eq:15}
\chi_{n + 1}
=
- \chi_n^{p - 1} \psi_n
\text{ and }
\psi_{n + 1}
=
- \chi_n^{p - 2} \psi_n^2.$$ This completes the proof of the induction step and of for every integer $n \ge 0$. Then the proposition follows from a direct computation using the recursion , together with and .
Lower bound of the norm of periodic points {#sec:lowerbound}
------------------------------------------
The goal of this section is to prove Theorem \[thm:lower-bound\]. We first introduce some notation and recall a result from [@LindahlRiveraLetelier2015].
Let $(\operatorname{\mathbb{K}}, | \cdot |)$ be an ultrametric field, and recall that $\mathcal{O}_{\operatorname{\mathbb{K}}}$ denotes the ring of integers of $\operatorname{\mathbb{K}}$, and $\mathfrak{m}_{\operatorname{\mathbb{K}}}$ the maximal ideal of $\mathcal{O}_{\operatorname{\mathbb{K}}}$. Denote the residue field of $\operatorname{\mathbb{K}}$ by $\widetilde{\operatorname{\mathbb{K}}} {\coloneqq}\mathcal{O}_{\operatorname{\mathbb{K}}} / \mathfrak{m}_{\operatorname{\mathbb{K}}}$, and for an element $a$ of $\mathcal{O}_{\operatorname{\mathbb{K}}}$, denote by the $\widetilde{a}$ its reduction in $\widetilde{\operatorname{\mathbb{K}}}$. The reduction of a power series $f$ in $\mathcal{O}_{\operatorname{\mathbb{K}}}[[\zeta]]$, is the power series $\widetilde{f}$ in $\widetilde{\operatorname{\mathbb{K}}}[[\zeta]]$ whose coefficients are the reductions of the corresponding coefficients of $f$. For a power series $f$ in $\mathcal{O}_{\operatorname{\mathbb{K}}}[[\zeta]]$, the *Weierstrass degree* $\operatorname{wideg}(f)$ of $f$ is the order in $\widetilde{\operatorname{\mathbb{K}}}[[\zeta]]$ of the reduction $\widetilde{f}$ of $f$. Note that if $\operatorname{wideg}(f)$ is finite, then the number of zeros of $f$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$, counted with multiplicity, is less than or equal to $\operatorname{wideg}(f)$, see, *e.g.*, [@Lang2002 §[VI]{}, Theorem 9.2].
In the case the characteristic $p$ of $\widetilde{\operatorname{\mathbb{K}}}$ is positive, and $f$ is a wildly ramified power series in $\mathcal{O}_{\operatorname{\mathbb{K}}}[[\zeta]]$, it is well-known that the minimal period of every periodic point of $f$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$ is a power of $p$.
Let $p$ be a prime number and $\operatorname{\mathbb{K}}$ field of characteristic $p$. For a wildly ramified power series $f$ in $\operatorname{\mathbb{K}}[[\zeta]]$, define for each integer $n \geq 0$ the element $\delta_n(f)$ of $\operatorname{\mathbb{K}}$ as follows: Put $\delta_n(f) {\coloneqq}0$ if $i_n(f) = +\infty$, and otherwise let $\delta_n(f)$ be the coefficient of $\zeta^{i_n(f)+1}$ in $f^{p^n}(\zeta)$.
\[lem24\] Let $p$ be a prime number and $(\operatorname{\mathbb{K}},|\cdot|)$ an ultrametric field of characteristic $p$. Then, for every wildly ramified power series $f$ in $\mathcal{O}_{\operatorname{\mathbb{K}}}[[\zeta]]$, the following properties hold.
1. Let $w_0$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$ be a fixed point of $f$ different from $0$. Then we have $$|w_0|\geq |\delta_0(f)|$$ with equality if and only if $$\operatorname{wideg}(f(\zeta)-\zeta)= i_0(f)+2.$$
2. Let $n\geq1$ be an integer and $\zeta_0$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$ a periodic point of $f$ of minimal period $p^n$. If in addition $i_n(f)<+\infty$, then we have $$|\zeta_0|
\geq
\left|\frac{\delta_n(f)}{\delta_{n-1}(f)}\right|^{\frac{1}{p^n}},$$ with equality if and only if $$\label{widegzeta}
\operatorname{wideg}\left(\frac{f^{p^n}(\zeta)-\zeta}{f^{p^{n-1}}(\zeta)-\zeta}\right) = i_n(f)-i_{n-1}(f)+p^n.$$ Moreover, if (\[widegzeta\]) holds, then the cycle containing $\zeta_0$ is the only cycle of minimal period $p^n$ of $f$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$, and for every point $\zeta_0'$ in this cycle $|\zeta_0'|=\left|\frac{\delta_n(f)}{\delta_{n-1}(f)}\right|^{\frac{1}{p^n}}$.
The assertion about fixed points is a direct consequence of ${\delta_0(f) = a}$ and Lemma \[lem24\](1).
To prove the statement about periodic points that are not fixed, note first that this statement holds trivially in the case $\operatorname{r\acute{e}sit}(f) = 0$. Thus, we assume that $\operatorname{r\acute{e}sit}(f) \neq 0$, and therefore $f$ is $q$-ramified by Theorem \[thm:q-ramification\]. In particular, for every integer $n \ge 1$ we have $i_n(f) < + \infty$. On the other hand, by Proposition \[prop:deltaiterates\] we have for every integer $n \ge 1$ $$\delta_n(f) = a^{\frac{p^{n + 1} - 1}{p - 1}} \operatorname{r\acute{e}sit}(f)^{\frac{p^n-1}{p-1}}.$$ Hence, by Lemma \[lem24\](2) we have for every periodic point $\zeta_0$ in $\mathfrak{m}_k$ of minimal period $p^n$, $$\label{normbound2}
|\zeta_0|
\geq
\left|\frac{\delta_n(f)}{\delta_{n-1}(f)}\right|^{\frac{1}{p^n}}
=
\left| a^{p^n} \operatorname{r\acute{e}sit}(f)^{p^{n-1}} \right|^{\frac{1}{p^n}}
=
|a| \cdot | \operatorname{r\acute{e}sit}(f)|^{\frac{1}{p}}.$$ This completes the proof of Theorem \[thm:lower-bound\].
Equality in is, as seen in Lemma \[lem24\], given by a condition on the reduction of $f$. In the case of equality, for $q$-ramified power series *all* periodic points in the open unit disk, which are not fixed by $f$, in fact lie on the sphere about the origin of radius $|\delta_0(f)| \cdot |\operatorname{r\acute{e}sit}(f)|^{\frac{1}{p}}$, see Example \[e:optimality\] in §\[sec:furth-results-exampl\].
Proof of the Main Lemma {#s:proof of Main Lemma}
=======================
The goal of this section is to prove the Main Lemma. We use the strategy introduced in [@RiveraLetelier2003 §3.2] and [@LindahlRiveraLetelier2013], using the power series $(\Delta_m)_{m = 0}^{+ \infty}$ defined by and . The proof is naturally divided into the cases $q \le p - 1$ and $q \ge p + 1$.
[[**Case 1, $q \le p - 1$.**]{}]{} Note that in this case we have $\ell = q$. For each integer $m \ge 1$ define $\alpha_m$, $\beta_m$ and $\gamma_m$ in $F_1$ by the recursive relations $$\begin{aligned}
\alpha_{m+1} & {\coloneqq}x_0 (qm+1)\alpha_m \label{receq1}\\
\beta_{m+1} & {\coloneqq}\left[ x_0^2 \binom{qm+1}{2} + x_1(qm+1)\right]\alpha_m + x_0 (q(m+1)+1)\beta_m \label{receq2}\\
\gamma_{m+1} & {\coloneqq}\left[ x_0^3 \binom{qm+1}{3} + x_0x_1 qm (qm+1)\right]\alpha_m \label{receq4}\\
& \quad + \left[ x_0^2 \binom{q(m+1)+1}{2}+ x_1(q(m+1)+1)\right]\beta_m \notag
\\ & \quad + x_0 (q(m+2)+1)\gamma_m, \notag\end{aligned}$$ with initial conditions $\alpha_1 {\coloneqq}x_0$, $\beta_1 {\coloneqq}x_1$, and $\gamma_1 {\coloneqq}0$. We claim that for every integer $m \ge 1$ we have $$\label{deltaclaim}
\Delta_m(\zeta)
\equiv
\alpha_m\zeta^{qm+1} + \beta_m\zeta^{q(m+1)+1} + \gamma_m\zeta^{q(m+2)+1}\mod \langle \zeta^{q(m + 2) + 2} \rangle.$$ For $m=1$ this holds by definition. Assume this is valid for some $m\geq1$. Then $$\begin{split}
\Delta_{m+1}(\zeta)
& = \Delta_m(\widehat{f}(\zeta)) - \Delta_m(\zeta)\\
&\equiv \alpha_m\zeta^{qm+1}\left[\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + x_2\zeta^{3q+1} + \cdots\right)^{qm+1} - 1\right]\\
&\qquad + \beta_m\zeta^{q(m+1)+1}\left[\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + x_2\zeta^{3q+1} + \cdots\right)^{q(m+1)+1}-1\right]\\
&\qquad + \gamma_m\zeta^{q(m+2)+1}\left[\left(1 + x_0 \zeta^q + x_1\zeta^{2q} + x_2\zeta^{3q+1} + \cdots\right)^{q(m+2) +1}-1\right]\\
&\qquad \mod \langle \zeta^{q(m+3)+2} \rangle\\
&\equiv
\alpha_m\left[ \zeta^{q(m+1)+1}x_0 (qm+1) + \zeta^{q(m+2)+1}\left(x_0^2 \binom{qm+1}{2} + x_1(qm+1)\right)
\right. \\ & \qquad \left.
+ \zeta^{q(m+3)+1}\left( x_0^3 \binom{qm+1}{3} + x_0 x_1 qm (qm+1)\right) \right]
\\
&\qquad+ \beta_m\Bigg[ \zeta^{q(m+2)+1} x_0 (q(m+1)+1)
\\ & \qquad
+ \zeta^{q(m+3)+1}\left( x_0^2 \binom{q(m+1)+1}{2} + x_1(q(m+1)+1)\right) \Bigg]
\\
&\qquad + \gamma_m\zeta^{q(m+3)+1} x_0 (q(m+2)+1) \mod \langle \zeta^{q(m+3)+2} \rangle,
\end{split}$$ which proves the induction step and .
In view of and , to prove the Main Lemma with $q \le p - 1$, it is sufficient to prove $$\label{e:main alpha}
\alpha_p \equiv 0 \mod p F_1,$$ with $\beta = \beta_p$, and $\gamma = \gamma_p$. The first 2 are given by Proposition \[prop:deltaiteratesshort\], so we only need to prove the latter. To do this, we solve explicitly, utilizing the explicit solutions of and given in the proof of Proposition \[prop:deltaiteratesshort\]. Assume first $q \equiv - 1 \pmod{p}$. By and with $m = p - 1$, we have $$\alpha_{p - 1} \equiv 0 \mod p F_1
\text{ and }
\beta_{p - 1} \equiv - x_0^{p - 2} x_1 \mod p F_1.$$ Combined with with $m = p - 1$, this implies $$\gamma_p
\equiv
- x_0^{p - 2} x_1^2 \mod p F_1.$$ This proves with $\gamma = \gamma_p$, when $q \equiv - 1 \pmod{p}$.
It remains to prove with $\gamma = \gamma_p$, when $q \not\equiv - 1 \pmod{p}$. Denote by $r_0$ the unique $r$ in $\{1, \ldots, p - 1 \}$ such that $qr \equiv - 1 \pmod{p}$. By our assumption $q \not\equiv - 1 \pmod{p}$, we have $r_0 \neq 1$ and therefore $$\label{eq:4}
r_0 \in \{2, \ldots, p - 1 \}.$$ Noting that for every $j \ge 0$ we have $qj + 1 > 0$, we use the substitution $$\gamma^*_m {\coloneqq}\frac{\gamma_mx_0^2}{(q(m+1)+1)(qm+1)\alpha_m}.$$ Note that by we have $$\gamma^*_m =
\gamma_m\bigg/ \left( x_0^{m - 2} \prod_{j=1}^{m + 1} (qj+1) \right).$$ On the other hand, by and we get $$\frac{\beta_m}{\alpha_m} = \frac{1}{x_0} \sum_{r=1}^m \left(x_0^2\frac{q(r-1)}{2} + x_1\right)\frac{qm+1}{qr+1}.$$ By plugging these equations into , we obtain $$\begin{split}
\gamma^*_{m+1}
& =
\gamma^*_m + \frac{qm}{(q(m+1)+1)(q(m+2)+1)} x_0^2 \left( x_0^2 \frac{qm-1}{6} + x_1 \right)
\\
& \quad +\frac{1}{q(m+2)+1} \left( x_0^2 \frac{q(m+1)}{2} + x_1 \right)
\sum_{r=1}^{m} \left(x_0^2 \frac{q(r - 1)}{2} + x_1 \right) \frac{1}{qr + 1}.
\end{split}$$ Using $\gamma^*_{1} = 0$ and defining for every integer $s$ $$H(s) {\coloneqq}x_0^2 \frac{qs}{2}+x_1,$$ we obtain inductively for each $m\geq 1$ $$\begin{gathered}
\gamma^*_{m}
=
\sum_{s=1}^{m-1} \left[ \frac{qs}{(q(s+1)+1)(q(s+2)+1)} x_0^2 \left(x_0^2 \frac{qs - 1}{6} + x_1 \right)
\right. \\ \left.
+ \frac{H(s+1)}{q(s+2)+1}\sum_{r=1}^{s}\frac{H(r - 1)}{qr + 1} \right].\end{gathered}$$ Equivalently, $$\begin{gathered}
\label{dmexplicit}
\gamma_{m}
=
x_0^{m - 2} \sum_{s=1}^{m-1}\left[ x_0^2 qs \left(x_0^2 \frac{qs - 1}{6} + x_1 \right) \prod_{j\in \{1,\ldots,m+1\}\setminus\{s+1,s+2\}}(qj+1)
\right. \\ \left.
+ H(s+1)\sum_{r=1}^{s}H(r - 1) \prod_{j\in \{1,\ldots,m+1\}\setminus \{r, s+2\}}(qj+1)\right].\end{gathered}$$ Setting $m = p$, for every $s$ in $\{1, \ldots, p - 1\}$ we have by Lemma \[wilson\] $$\prod_{\substack{j\in \{1,\ldots,p+1\} \\ j \not\in \{s+1,s+2\}}}(qj+1)
\equiv
\begin{cases}
- \frac{q(p+1)+1}{q(r_0 + 1)+1} \equiv -\frac{q+1}{q} \mod p\operatorname{\mathbb{Z}}_{(p)}
& \text{if~$s = r_0 - 1$};
\\
- \frac{q(p+1)+1}{q(r_0 - 1)+1} \equiv \frac{q+1}{q} \mod p\operatorname{\mathbb{Z}}_{(p)}
& \text{if $s = r_0 - 2$};
\\
0
& \text{otherwise}.
\end{cases}$$ Analogously, for every $s$ in $\{1, \ldots, p - 1 \}$ and $r$ in $\{1, \ldots, s\}$, we have $$\prod_{\substack{j\in \{1,\ldots,p+1\} \\ j \not\in \{r, s+2\}}} (qj+1)
\equiv
\begin{cases}
- \frac{q+1}{qr + 1} \mod p\operatorname{\mathbb{Z}}_{(p)}
& \text{if $s = r_0 - 2$};
\\
- \frac{q+1}{q(s + 2) + 1} \mod p\operatorname{\mathbb{Z}}_{(p)}
& \text{if~$s \ge r_0$ and~$r = r_0$};
\\
0
& \text{otherwise}.
\end{cases}$$ Combined with with $m = p$ and $$\label{eq:2}
H(r_0 - 1) \equiv - x_0^2 \frac{q + 1}{2} + x_1 \mod p \operatorname{\mathbb{Z}}_{(p)},$$ these congruences imply $$\label{eq:3}
\begin{split}
\gamma_p
& \equiv
- x_0^{p} q(r_0 - 1) \left( x_0^2 \frac{q(r_0 - 1) - 1}{6} + x_1 \right) \frac{q + 1}{q}
\\ & \quad
+ x_0^{p} q (r_0 - 2) \left( x_0^2 \frac{q (r_0 - 2) - 1}{6} + x_1 \right) \frac{q + 1}{q}
\\ & \quad
- x_0^{p - 2} H(r_0 - 1) \sum_{r = 1}^{r_0 - 2} H(r - 1) \frac{q + 1}{qr + 1}
\\ & \quad
- x_0^{p - 2} \sum_{s = r_0}^{p - 1} H(s + 1)H(r_0 - 1) \frac{q + 1}{q(s + 2) + 1} \mod p F_1
\\ & \equiv
- x_0^{p} (q + 1) H(r_0 - 1)
\\ & \quad
- x_0^{p - 2} (q + 1) H(r_0 - 1)
\sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \frac{H(r - 1)}{q r + 1}
\mod p F_1.
\end{split}$$ By , we have $$\begin{gathered}
\label{eq:5}
\sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \frac{H(r - 1)}{q r + 1}
\\
\begin{aligned}
& \equiv
\sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \left( \frac{x_0^2}{2} + \frac{H(r_0 - 1)}{q r + 1} \right)
\mod p F_1
\\ & \equiv
- x_0^2 + H(r_0 - 1) \sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \frac{1}{q r + 1}
\mod p F_1.
\end{aligned}\end{gathered}$$ On the other hand, by the second assertion of Lemma \[wilson\], we have $$\begin{gathered}
\label{eq:25}
\sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \frac{1}{q r + 1}
\\
\begin{aligned}
& \equiv
\frac{1}{q(p + 1) + 1} - \frac{1}{q(r_0 - 1) + 1} - \frac{1}{q(r_0 + 1) + 1}
\mod p \operatorname{\mathbb{Z}}_{(p)}
\\ & \equiv
\frac{1}{q + 1}
\mod p \operatorname{\mathbb{Z}}_{(p)}.
\end{aligned}\end{gathered}$$ Together with , , and , this implies with $\gamma = \gamma_p$ and completes the proof of the Main Lemma in the case $q \le p - 1$.
[[**Case 2, $q \ge p + 1$.**]{}]{} Note that in this case our hypotheses on $\ell$ imply in all the cases that $q \ge 2 \ell + 1$. For each integer $m \ge 1$ define ${\widehat{\alpha}}_m$, ${\widehat{\beta}}_m$ and ${\widehat{\gamma}}_m$ in $F_1$ by the recursive relations $$\begin{aligned}
& {\widehat{\alpha}}_{m+1} {\coloneqq}x_0(qm+1){\widehat{\alpha}}_m \label{receq1b}\\
&{\widehat{\beta}}_{m+1} {\coloneqq}x_1(qm+1){\widehat{\alpha}}_m + x_0(qm+\ell+1){\widehat{\beta}}_m \label{receq2b}\\
& {\widehat{\gamma}}_{m+1} {\coloneqq}x_1(qm +\ell+1){\widehat{\beta}}_m + x_0(qm+2\ell+1){\widehat{\gamma}}_m\label{receq4b},\end{aligned}$$ with initial conditions ${\widehat{\alpha}}_1 {\coloneqq}x_0$ ${\widehat{\beta}}_1 {\coloneqq}x_1$, and ${\widehat{\gamma}}_1 {\coloneqq}0$. We claim that for every integer $m \ge 1$ we have $$\label{deltaclaimhat}
\Delta_m(\zeta)
\equiv
{\widehat{\alpha}}_m\zeta^{qm+1} + {\widehat{\beta}}_m\zeta^{qm+\ell+1} + {\widehat{\gamma}}_m\zeta^{qm+2\ell+1}\mod \langle \zeta^{qm+2\ell+2} \rangle.$$ For $m=1$ this holds by definition. Assume further this is valid for some $m\geq1$. Then, using $q \ge 2 \ell + 1$, we have $$\begin{split}
\Delta_{m+1}(\zeta)
& = \Delta_{m}(\widehat{f}(\zeta)) - \Delta_{m}(\zeta)\\
& \equiv {\widehat{\alpha}}_m\zeta^{qm+1}\left[\left(1 + x_0\zeta^{q} + x_1\zeta^{q+\ell} + x_2\zeta^{q+2\ell + 1} + \cdots\right)^{qm+1} - 1\right]\\
&\quad + {\widehat{\beta}}_m\zeta^{qm+\ell+1}\left[\left(1 +x_0\zeta^{q} + x_1\zeta^{q+\ell} + x_2\zeta^{q+2\ell + 1} + \cdots\right)^{qm + \ell + 1}-1\right]\\
&\quad + {\widehat{\gamma}}_m\zeta^{qm + 2\ell + 1}\left[\left(1 + x_0\zeta^{q} + x_1\zeta^{q+\ell} + x_2\zeta^{q+2\ell + 1} + \cdots\right)^{qm + 2\ell + 1}-1\right]\\
&\quad \mod \langle \zeta^{q(m + 1) + 2\ell + 2} \rangle
\\
&\equiv{\widehat{\alpha}}_m\left(\zeta^{q(m+1) + 1}x_0(qm+1) + \zeta^{q(m+1)+\ell+1}x_1(qm+1)\right) \\
&\quad+ {\widehat{\beta}}_m\left(\zeta^{q(m+1)+\ell+1}x_0(qm+\ell+1) + \zeta^{q(m+1)+2\ell+1}x_1(qm+\ell+1)\right)
\\
& \quad+ {\widehat{\gamma}}_m\zeta^{q(m+1)+2\ell+1}x_0(qm+2\ell+1) \mod \langle \zeta^{q(m+1) + 2\ell +2} \rangle,
\end{split}$$ which proves the induction step and the claim .
In view of and , to complete the proof of the Main Lemma in the case $q \ge p + 1$, it is sufficient to prove $$\label{e:main alpha bis}
{\widehat{\alpha}}_p \equiv 0 \mod p F_1,$$ with $\beta = {\widehat{\beta}}_p$, and with $\gamma = {\widehat{\gamma}}_p$. The linear recursion described in , and can be solved explicitly. By telescoping , we obtain for every $m \ge 1$ the solution $$\label{alphamhat}
{\widehat{\alpha}}_m
=
x_0^{m}\prod_{j=1}^{m-1}(qj+1).$$ Taking $m = p$, this implies .
On the other hand, inserting in yields $${\widehat{\beta}}_{m+1} = x_0^{m}x_1\prod_{j=1}^m (qj+1) + x_0(qm+\ell+1){\widehat{\beta}}_m.$$ Then, an induction argument shows that for every $m \ge 1$ we have $$\label{betamhat}
{\widehat{\beta}}_m \equiv x_0^{m-1}x_1\sum_{r=1}^{m}\prod_{j \in \{1, \ldots, m\} \setminus \{ r \}} (q j+1) \mod p F_1.$$ When $m = p$ every term in the sum above contains a factor $p$, except for the unique $r_0$ in $\{1, \ldots, p - 1 \}$ satisfying $qr_0 \equiv -1 \pmod{p}$. Then by Lemma \[wilson\], we have $$\begin{split}
{\widehat{\beta}}_p
&\equiv
x_0^{p-1}x_1\prod_{j \in \{ 1, \ldots, p \} \setminus \{ r_0 \}} (q j+1) \mod p F_1
\\ &\equiv
-x_0^{p-1}x_1 \mod p F_1.
\end{split}$$ This proves with $\beta = {\widehat{\beta}}_p$.
To prove with $\gamma = {\widehat{\gamma}}_p$, assume first $q \equiv - 1 \pmod{p}$. Then by with $m = p - 1$, , and Lemma \[wilson\] we have $$\begin{split}
{\widehat{\gamma}}_p
& \equiv
x_1 {\widehat{\beta}}_{p - 1} \mod p F_1
\\ & \equiv
x_0^{p - 2} x_1^2 \sum_{r = 1}^{p - 1} \prod_{j \in \{1, \ldots, p - 1\} \setminus \{ r \}} (q j + 1) \mod p F_1
\\ & \equiv
x_0^{p - 2} x_1^2 \prod_{j \in \{2, \ldots, p - 1\}} (1 - j) \mod p F_1
\\ & \equiv
- x_0^{p - 2} x_1^2 \mod p F_1.
\end{split}$$ It remains to prove with $\gamma = {\widehat{\gamma}}_p$ in the case $q \not\equiv - 1 \pmod{p}$. Note that in this case $r_0 \neq 1$. Inserting in , we obtain $$\begin{gathered}
\label{eq:9}
{\widehat{\gamma}}_{m+1} \equiv x_0^{m-1}x_1^2\sum_{r = 1}^{m}\prod_{j \in \{1, \ldots, m+1\} \setminus \{ r \}} (q j+1)
\\
+ x_0(q(m+2)+1){\widehat{\gamma}}_m \mod p F_1.
\end{gathered}$$ For every $m \ge 1$ define ${\check{\gamma}}_m$ in $F_1$ recursively, by ${\check{\gamma}}_1 {\coloneqq}0$ and for $m \ge 1$, by $$\label{eq:1}
{\check{\gamma}}_{m + 1}
{\coloneqq}x_0^{m-1}x_1^2\sum_{r=1}^{m}\prod_{j \in \{1, \ldots, m+1\} \setminus \{ r \}} (q j+1)
+ x_0(q(m+2)+1){\check{\gamma}}_m.$$ Note that by for every $m \ge 1$ we have ${\check{\gamma}}_m \equiv {\widehat{\gamma}}_m \mod p F_1$. Using that for every $j \ge 0$ we have $q j + 1 > 0$, and the substitution $${\check{\gamma}}^*_m
{\coloneqq}{\check{\gamma}}_m \bigg/ \left(x_0^{m-2}x_1^2\prod_{j=1}^{m+1} (q j+1) \right),$$ we obtain $${\check{\gamma}}_{m+1}^*
=
{\check{\gamma}}_{m}^* + \frac{1}{q(m+2)+1}\sum_{r=1}^{m}\frac{1}{q r + 1}.$$ Inductively we have $$\label{eq:8}
{\check{\gamma}}_m^*
=
\sum_{s=1}^{m-1}\frac{1}{q (s+2) + 1}\sum_{r=1}^{s}\frac{1}{q r + 1},$$ which is a rational number. Since $r_0 \neq 1$, for every $r$ in $\{1, \ldots, p + 1 \} \setminus \{ r_0 \}$ we have that $\frac{1}{q r + 1}$ is in $\operatorname{\mathbb{Z}}_{(p)}$. Thus, taking $m = p$ in , and using , we obtain $$\begin{split}
(q r_0 + 1) {\check{\gamma}}_p^*
& \equiv
\sum_{\substack{r \in \{1, \ldots, p + 1 \} \\ r \not\in \{ r_0 - 1, r_0, r_0 + 1 \}}} \frac{1}{q r + 1} \mod p \operatorname{\mathbb{Z}}_{(p)}
\\ & \equiv
\frac{1}{q + 1} \mod p \operatorname{\mathbb{Z}}_{(p)}.
\end{split}$$ Using Lemma \[wilson\], we obtain $$\begin{split}
{\check{\gamma}}_p
& \equiv
x_0^{p-2}x_1^2 \frac{1}{q + 1} \prod_{j \in \{1, \ldots, p + 1 \} \setminus \{ r_0 \}} (q j + 1) \mod p F_1
\\ & \equiv
- x_0^{p-2}x_1^2 \mod p F_1.
\end{split}$$ This completes the proof of with $\gamma = {\widehat{\gamma}}_p$ and of the Main Lemma.
Further results and examples {#sec:furth-results-exampl}
============================
In this section we gather several examples illustrating our results and state some further consequences of our main theorems.
\[ex:p + 1\] The following example shows that the conclusion of Theorem \[thm:q-ramification\] is false when $q = p + 1$ and $p$ is odd. Consider the polynomial with coefficients in $\operatorname{\mathbb{F}}_p$, $$P(\zeta)
{\coloneqq}\zeta (1 + \zeta^{p + 1} + \zeta^{p + 2} + \zeta^{2(p + 1)}).$$ A direct computation using shows that $\operatorname{r\acute{e}sit}(P) = 1$. On the other hand, using the Main Lemma with $q = p + 1$, $\ell = 1$, and $x_0=x_1=1$, we have $$i_1(P) = p^2 + p + 1
<
i_0(P)(p + 1),$$ so $P$ is not $(p + 1)$-ramified.
There is another natural source of power series $f$ that satisfy $i_0(f) = p + 1$ and that are not $(p + 1)$-ramified. Let $g$ in $\operatorname{\mathbb{K}}[[\zeta]]$ be a $1$-ramified power series, and put ${f {\coloneqq}g^p}$. Then $$i_0(f)
=
i_1(g)
=
p+1
\text{ and }
i_1(f) = i_2(g) = 1+p+p^2
<
i_0(f) (p+1),$$ so $f$ is not $(p+1)$-ramified. For concreteness, let $a$ in $\operatorname{\mathbb{K}}$ be different from $1$, and assume that $g$ is of the form $$g(\zeta)
\equiv
\zeta(1 + \zeta + a \zeta^2) \mod \langle \zeta^4 \rangle.$$ In view of , we have $\operatorname{r\acute{e}sit}(g) = 1 - a \neq 0$, so $g$ is $1$-ramified by Theorem \[thm:q-ramification\]. On the other hand, for $p = 3$, $5$ and $7$ a computation shows that $\operatorname{r\acute{e}sit}(f) = (1 - a)^p \neq 0$. Thus, in contrast with the situation for $q$ in $\{1, \ldots, p - 1 \}$ in Theorem \[thm:q-ramification\], for $q = p + 1$ and $p = 3$, $5$ and $7$ the nonvanishing of the iterative residue does not imply $(p + 1)$-ramification.[^5] So, the following question arises naturally.
For which $1$-ramified power series $g$ in $\operatorname{\mathbb{K}}[[\zeta]]$ do we have $\operatorname{r\acute{e}sit}(g^p) \neq 0$?
\[ex:p - 1\] The following example illustrates Theorem \[thm:q-ramification\] in the case $q = p - 1$. A direct computation shows that for the polynomial $P(\zeta) {\coloneqq}\zeta + \zeta^p$, we have for every integer $n \ge 1$ $$P^{p^n}(\zeta) = \zeta + \zeta^{p^{p^n}}.$$ In particular, $i_n(P) = p^{p^n} - 1$, and therefore $P$ is not $(p - 1)$-ramified. This is consistent with Theorem \[thm:q-ramification\], since by Theorem \[thm:closed-formula\] we have $\operatorname{r\acute{e}sit}(P) = \operatorname{index}(P) = 0$.
\[e:optimality\] This example shows that the lower bound in Theorem \[thm:lower-bound\] is optimal for $p \ge 5$ and $q \le p - 3$. Let $p \ge 3$ be a prime number, ($\operatorname{\mathbb{K}}$, $| \cdot |$) an ultrametric field of characteristic $p$, and $q$ in $\{1, \ldots, p - 1 \}$. Furthermore, let $a$ and $b$ in $\operatorname{\mathbb{K}}$ be such that $0 < |a| < 1$ and $|b|=1$, and let $f$ be a power series in $\operatorname{\mathbb{K}}[[z]]$ satisfying $$f(\zeta) \equiv \zeta(1 + a\zeta^q + b\zeta^{q+1}) \mod \langle \zeta^{2q + 4} \rangle.$$ A direct computation using shows that $$\operatorname{r\acute{e}sit}(f) = \frac{q + 1}{2} + (-1)^q \frac{b^q}{a^{q + 1}} \neq 0,$$ so by Theorem \[thm:q-ramification\] the series $f$ is $q$-ramified. In the case $q \le p - 2$, by the reduction $\widetilde{f}$ of $f$ satisfies $\operatorname{r\acute{e}sit}(\widetilde{f}) = \frac{q + 2}{2}$. Assuming further that $q \le p - 3$, we have $\operatorname{r\acute{e}sit}(\widetilde{f}) \neq 0$, and we obtain that $\widetilde{f}$ is ${(q+1)}$-ramified by Theorem \[thm:q-ramification\]. This implies that in Lemma \[lem24\] holds for every integer $n \ge 1$. It follows that for every periodic point $\zeta_0$ of $f$ in $\mathfrak{m}_{\operatorname{\mathbb{K}}}$ that is not fixed, we have $$|\zeta_0| = |a| \cdot |\operatorname{r\acute{e}sit}(f)|^{\frac{1}{p}},$$ see the proof of Theorem \[thm:lower-bound\].
The following result is a direct consequence of Theorems \[thm:q-ramification\] and \[thm:lower-bound\] for fixed points whose multiplier is a root of unity, compare with [@LindahlRiveraLetelier2015 Corollary C].
\[c:higher order\] Let $\operatorname{\mathbb{K}}$ be an ultrametric field of odd characteristic, let $\gamma$ in $\operatorname{\mathbb{K}}$ be a root of unity, and denote by $q \ge 1$ the order of $\gamma$. Moreover, let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ satisfying $f(0)=0$ and $f'(0)=\gamma$. If $$q' {\coloneqq}\operatorname{mult}(f^q) - 1 \le p - 1
\text{ and }
\operatorname{r\acute{e}sit}(f^q) \neq 0,$$ then $f^q$ is $q'$-ramified. In particular, if $f$ converges on a neighborhood of the origin, then the origin is isolated as a periodic point of $f$.
Let $\operatorname{\mathbb{K}}$ be an ultrametric field of characteristic $7$, and note that $2$ is a root of unity in $\operatorname{\mathbb{K}}$ of order $3$. Let $f$ be a power series with coefficients in $\mathcal{O}_{\operatorname{\mathbb{K}}}$ such that $$f(\zeta)
\equiv
2\zeta + \zeta^2 \mod \langle \zeta^{13} \rangle.$$ A direct computation shows that $$f^3(\zeta) \equiv \zeta(1 + \zeta^6 + \zeta^7) \mod \langle \zeta^{13} \rangle.$$ In particular, $\operatorname{mult}(f^3) - 1 = 6 > 3$, so $f$ is not minimally ramified in the sense of [@LindahlRiveraLetelier2015], and we cannot apply Corollary C of that paper to $f$. However, by we have $\operatorname{r\acute{e}sit}(f^3) \neq 0$, so Corollary \[c:higher order\] applies to $f^3$ and it implies that $f^3$ is $6$-ramified and that the origin is isolated as a periodic point of $f^3$, and hence of $f$.
Iterative residue in positive characteristic {#app:resit}
============================================
In this section we study the behavior of the iterative residue under iteration, which is defined for a power series $f$ with coefficients in a field of characteristic different from $2$, by . For a ground field of characteristic zero, this behavior can be understood from a relatively easy computation using the normal form .[^6] For a ground field of positive characteristic, not every power series $f$ is formally conjugated to , so we cannot apply this strategy. We use instead the closed formula for the residue fixed point index in Theorem \[thm:closed-formula\].
\[resitcharp\] Let $\operatorname{\mathbb{K}}$ a field of characteristic different from $2$, and let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $$f(0) = 0,
f'(0) = 1
\text{ and }
f(z) \neq z.$$ Then, for every integer $n \ge 1$ that is not divisible by the characteristic of $\operatorname{\mathbb{K}}$, we have $$\label{eq:16}
\operatorname{r\acute{e}sit}(f^n)
=
\frac{1}{n}\operatorname{r\acute{e}sit}(f).$$
For a field of characteristic $2$, the formula defining the iterative residue is meaningless. Instead, we study the behavior of the residue fixed point index under iteration.
\[resitchar2\] Let $\operatorname{\mathbb{K}}$ be a field of characteristic $2$, and let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $q {\coloneqq}\operatorname{mult}(f) - 1 \ge 1$. Then, for every odd integer $n \ge 1$ we have $$\operatorname{index}(f^n)
=
\begin{cases}
\operatorname{index}(f) +1
& \text{if $q$ is even and $n\equiv 3 \pmod{4}$};
\\
\operatorname{index}(f)
& \text{otherwise.}
\end{cases}$$
The proofs of Proposition \[resitcharp\] and \[resitchar2\] are given after the following lemma. For a field $\operatorname{\mathbb{K}}$ of positive characteristic, and an integer $n \ge 0$, we use $\binom{n}{2}$ to denote the reduction of this integer in the prime field of $\operatorname{\mathbb{K}}$.
\[fniter\] Let $\operatorname{\mathbb{K}}$ be a field, let $f$ be a power series with coefficients in $\operatorname{\mathbb{K}}$ such that $q {\coloneqq}\operatorname{mult}(f) - 1 \ge 1$, and denote by $a$ the coefficient of $z^{q + 1}$ in $f(z)$. Then, for every integer $n \ge 1$ we have $$\label{fncharp}
f^n(z) -z\equiv n (f(z)-z) + \binom{n}{2}(q+1)a^2z^{2q+1} \mod \langle z^{2q+2} \rangle.$$
We proceed by induction. The lemma holds trivially for $n=1$. Assume that holds for an integer $n\geq1$. Put $\Phi(z) {\coloneqq}\frac{f(z)-z}{z}$ and note that $$\Phi(z)
\equiv
a z^q \mod \langle z^{q + 1} \rangle,
\text{ and }
\Phi(f(z))
\equiv
\Phi(z) + q a^2 z^{2q} \mod \langle z^{2q + 1} \rangle.$$ Together with the induction hypothesis, this implies $$\begin{split}
f^n\circ f(z)
& \equiv
f(z) + n f(z) \Phi(f(z)) + \binom{n}{2}(q+1)a^2f(z)^{2q+1} \mod \langle z^{2q+2} \rangle\\
&\equiv
z + z \Phi(z) + n z (1 + \Phi(z))\left(\Phi(z) + q a^2z^{2q}\right)\\
&\qquad
+ \binom{n}{2}(q+1)a^2 z^{2q + 1} \mod \langle z^{2q+2} \rangle\\
&\equiv z + (n+1) z \Phi(z) + \left(n + \binom{n}{2}\right) (q+1) a^2z^{2q+1}\mod \langle z^{2q+2} \rangle\\
&\equiv z + (n+1) (f(z) - z) + \binom{n+1}{2} (q+1) a^2z^{2q+1} \mod \langle z^{2q+2} \rangle.
\qedhere
\end{split}$$
Given a field $\operatorname{\mathbb{K}}$, an integer $q \ge 1$, and $a_q$, …, $a_{2q}$ in $\operatorname{\mathbb{K}}$, denote by $P_q(a_q,\ldots,a_{2q})$ the right-hand side of . Note that for every $\lambda$ in $\operatorname{\mathbb{K}}$ we have $$\label{eq:21}
P_q( a_q, \ldots, a_{2q} + \lambda a_q^2)
=
P_q(a_q,\ldots, a_{2q}) + \lambda.$$ If in addition $\lambda$ is nonzero, then we also have $$\label{eq:18}
P_q( \lambda a_q, \ldots, \lambda a_{2q})
=
\frac{1}{\lambda} P_q(a_q,\ldots, a_{2q}).$$
Put $$f(z) = z(1 + a_q z^q + \cdots + a_{2q}z^{2q} + \cdots),$$ so that $a_q \neq 0$. A direct computation shows that for every integer $n \ge 1$, we have $$f^n(z)
\equiv
z(1 + n a_q z^q) \mod \langle \zeta^{q + 2} \rangle.$$ In particular, if $n$ is not divisible by the characteristic of $\operatorname{\mathbb{K}}$, then $\operatorname{mult}(f^n) = q + 1$. On the other hand, by Theorem \[thm:closed-formula\], Lemma \[fniter\], , and , we have $$\begin{split}
\label{eq:20}
\operatorname{index}(f^n)
& =
P_q\left(na_q, \ldots, na_{2q-1}, na_{2q} + \binom{n}{2} (q+1) a_q^2 \right)
\\ & =
\frac{1}{n} P_q\left(a_q, \ldots, a_{2q-1}, a_{2q} \right) + \frac{1}{n^2} \binom{n}{2} (q+1)
\\ & = \frac{1}{n} \left[ \operatorname{index}(f) + \frac{1}{n} \binom{n}{2} (q+1) \right].
\end{split}$$ If the characteristic of $\operatorname{\mathbb{K}}$ is different from $2$, then by the definition of the iterative residue we have $$\begin{gathered}
n \operatorname{r\acute{e}sit}(f^n)
=
n \frac{\operatorname{mult}(f^n)}{2} - n\operatorname{index}(f^n)
\\ =
n \frac{q + 1}{2} - \operatorname{index}(f) - \frac{n - 1}{2} (q + 1)
=
\operatorname{r\acute{e}sit}(f).\end{gathered}$$ This proves Proposition \[resitcharp\]. In the case the characteristic of $\operatorname{\mathbb{K}}$ is $2$, Proposition \[resitchar2\] follows from and from the fact that, in $\operatorname{\mathbb{K}}$, we have $n = 1$ and $$\binom{n}{2}(q + 1)
=
\begin{cases}
1
& \text{if $q$ is even and $n\equiv 3 \pmod{4}$;}
\\
0
& \text{otherwise.}
\end{cases}
\qedhere$$
[[LRL]{}16b]{}
Xavier Buff and Adam L. Epstein. A parabolic [P]{}ommerenke-[L]{}evin-[Y]{}occoz inequality. , 172(3):249–289, 2002.
Xavier Buff, Jean Écalle, and Adam Epstein. Limits of degenerate parabolic quadratic rational maps. , 23(1):42–95, 2013.
Xavier Buff. Virtually repelling fixed points. , 47(1):195–209, 2003.
Rachel Camina. The [N]{}ottingham group. In [*New horizons in pro-[$p$]{} groups*]{}, volume 184 of [*Progr. Math.*]{}, pages 205–221. Birkhäuser Boston, Boston, MA, 2000.
J. Écalle. Théorie itérative: introduction à la théorie des invariants holomorphes. , 54:183–258, 1975.
M. Herman and J.-C. Yoccoz. Generalizations of some theorems of small divisors to non-[A]{}rchimedean fields. In [*Geometric dynamics ([R]{}io de [J]{}aneiro, 1981)*]{}, volume 1007 of [*Lecture Notes in Math.*]{}, pages 408–447. Springer, Berlin, 1983.
D. L. Johnson. The group of formal power series under substitution. , 45(3):296?302, 1988.
Kevin Keating. , 41(3):314–321, 1992.
Kenz [Kallal]{} and Hudson [Kirkpatrick]{}. . , page arXiv:1611.01077, Nov 2016.
Serge Lang. . Springer, New York, rev. 3. ed. edition, 2002.
Karl-Olof Lindahl. On [S]{}iegel’s linearization theorem for fields of prime characteristic. , 17(3):745–763, 2004.
Fran[ç]{}ois Laubie, Abbas Movahhedi, and Alain Salinier. , 132(1):57–98, 2002.
Karl-Olof Lindahl and Jonas Nordqvist. Geometric location of periodic points of 2-ramified power series. , 465(2):762–794, 2018.
Karl-Olof Lindahl and Juan Rivera-Letelier. Generic parabolic points are isolated in positive characteristic. , 29(5):1596–1621, 2016.
Karl-Olof Lindahl and Juan Rivera-Letelier. Optimal cycles in ultrametric dynamics and minimally ramified power series. , 152(1):187–222, 2016.
F. Laubie and M. Saïne. Ramification of some automorphisms of local fields. , 72(2):174–182, 1998.
Karl-Olof Lindahl and Michael Zieve. On hyperbolic fixed points in ultrametric dynamics. , 2(3):232–240, 2010.
John Milnor. , volume 160 of [*Annals of Mathematics Studies*]{}. Princeton University Press, Princeton, NJ, third edition, 2006.
Jonas Nordqvist. Characterization of 2-ramified power series. , 174:258–273, 2017.
Jonas [Nordqvist]{}. . , page arXiv:1909.10782, Sep 2019.
Juan Rivera-Letelier. , 287(xv):147–230, 2003.
Matteo Ruggiero. Classification of one-dimensional superattracting germs in positive characteristic. , 35(7):2242–2268, 2015.
Shankar Sen. . , 90:33–46, 1969.
Joseph H. Silverman. . Springer, New York, N.Y., 2007.
Jean-Pierre Wintenberger. Automorphismes des corps locaux de caractéristique [$p$]{}. , 16(2):429–456, 2004.
[^1]: This terminology arises from the study of field automorphisms. Every power series $f$ with coefficients in a field $\operatorname{\mathbb{K}}$ that satisfies $f(0) = 0$ and $f'(0) = 1$, defines a field automorphism of $\operatorname{\mathbb{K}}[[t]]$ given by $g \mapsto g \circ f$. When $\operatorname{\mathbb{K}}$ is of positive characteristic, this type of field automorphism is traditionally known as *wildly ramified*, due to the behavior of its associated ramification numbers.
[^2]: We keep [É]{}calle’s notation “$\operatorname{r\acute{e}sit}$”, an abbreviation of the French “*r[é]{}sidue it[é]{}ratif*”.
[^3]: Note that this polynomial is isobaric of degree $q(q + 1)$.
[^4]: Recently, the first named author answered this question completely in [@Nor1909].
[^5]: The situation is now clear form the recent characterization of $(p + 1)$-ramification by the first named author in [@Nor1909].
[^6]: See also [@Mil06c Lemma 12.9] for a different approach for convergent power series.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The issue of spectrum scarcity in wireless networks is becoming prominent and critical with each passing year. Although several promising solutions have been proposed to provide a solution to spectrum scarcity, most of them have many associated tradeoffs. In this context, one of the emerging ideas relates to the utilization of cognitive radios (CR) for future heterogeneous networks (HetNets). This paper provides a marriage of two promising candidates (i.e., CR and HetNets) for beyond fifth generation (5G) wireless networks. More specifically, a joint power allocation and user assignment solution for the multi-user underlay CR-based HetNets has been proposed and evaluated. To counter the limiting factors in these networks, the individual power of transmitting nodes and interference temperature protection constraints of the primary networks have been considered. An efficient solution is designed from the dual decomposition approach, where the optimal user assignment is obtained for the optimized power allocation at each node. The simulation results validate the superiority of the proposed optimization scheme against conventional baseline techniques.'
author:
- 'Khush Bakht, Furqan Jameel, Zain Ali, Wali Ullah Khan, Imran Khan, Guftaar Ahmad Sardar Sidhu, and Jeong Woo Lee [^1][^2][^3] [^4] [^5] [^6] [^7]'
bibliography:
- 'References.bib'
title: Power Allocation and User Assignment Scheme for Beyond 5G Heterogeneous Networks
---
Beyond 5G, Cognitive Radio (CR), Dual Decomposition, User Fairness, Heterogeneous Networks (HetNets)
Introduction
============
In the last few years, the wireless systems have evolved to the point where a homogeneous cellular networks have achieved near optimal performance [@8340813]. These advancements in the homogeneous cellular networks, though significant, may not be enough to support beyond fifth generation (5G) wireless networks [@Gao]. To do so, dynamic and exhaustive improvements in spectral efficiency are needed. Typically, a HetNet consists of main macro base stations (MBSs), a few pico base stations (PBSs), and several femto base stations (FBSs). An MBS in HetNets has a high transmission power and greater coverage area which is then overlaid with low-powered PBS and FBS [@xu2018robust; @2a].
There are different purposes of PBS and FBS in the HetNets. In conventional homogeneous cellular network, the mobile terminal is associated to a BS on the basis of downlink signal-to-interference-and-noise ratio (SINR). In HetNets, the SINR-based association principle leads to having load balancing issue among MBS and PBS.
Related Works
-------------
By the end of 2020, it is anticipated that up to 50 billion devices will exist in the world including static and mobile platforms [@8581856]. Due to this reason, there has been an upsurge in the research on HetNets to provide efficient and long-term solutions [@sun2018energy]. Zhang *et al.* in [@3] investigated the problem of user association in HetNets. They considered an optimization problem with different traffic capacity limits, quality of service (QoS) requirements and power budget constraints. The proposed scheme of user association improved the performance of the network. The authors of [@4] performed joint optimization of resource allocation and user association in HetNets. They presented and compared three allocation strategies, i.e., orthogonal deployment, co-channel deployment, and partially shared deployment.
Liu *et al.* presented a fair user association scheme in [@9]. In [@11], the HetNets using orthogonal frequency division multiple access (OFDMA) network were presented by the authors. The aim was to manage radio resource by maximizing the throughput of user having minimum rate. Similarly, the authors of [@12] managed the radio resources by a scheduling algorithm implemented by a central global resource controller (GRC). The algorithm optimized the attributes such as fairness among users, spectral efficiency and battery lifetime. Relay-based HetNets were considered by the authors of [@qin2019cross], wherein, they proposed a method to suppress inter-cell and intra-cell interference. Their proposed scheme was shown to outperform existing baseline methods in terms of sum-rate performance.
Semov *et al.* proved that in HetNets by taking the geographical position into account the users throughput and fairness can be improved. [@13]. A similar concept was employed in the form of relays by the authors of [@14] to ensure proportional fairness among user equipment by taking into account backhaul links between the relays and BS. . In [@9f], the authors studied a single cell HetNet having one macro and one picocell for efficient resource allocation such that the energy efficiency is maximized by proposing an iterative resource allocation algorithm. To achieve the spectral efficiency of the system more advanced dynamic spectrum access techniques (DSA) should be employed. The cognitive radio (CR) is an efficient DSA technique[@15] that allows secondary (unlicensed) users (SU) to access the spectrum of primary (licensed) users (PU) in an opportunistic way[@16]. In CR, spectrum sharing can be classified as spectrum overlay and spectrum underlay.
Of late, the CR-based HetNets have been drawing a lot of research attention nowadays due to their dynamic resource allocation property. The power adjustment of BS and mobile users can be achieved by dynamic resource allocation [@7f]. In this regard, the author in [@18] maximized the energy efficiency subject to power and interference constraint in OFDMA based CR networks. The authors applied the convex optimization theory and proposed an iterative algorithm. In [@14f], the authors considered the resource allocation problem for rate maximization in multi-user cognitive heterogeneous networks. The authors considered the maximum transmit power of cognitive microcell base station and cross-tier interference constraints, simultaneously. They converted the non-convex optimization problem into a geometric programming problem and solved it in a distributed way using the Lagrange dual method.
**Acronym** **Definition**
------------- -----------------------------------------------
5G Fifth Generation
AWGN Additive White Gaussian Noise
BS Base Station
CR Cognitive Radio
D2D Device to Device
DAS Distributed Antenna System
DSA Dynamic Spectrum Access
FBS Femto Base Station
GRC Global Resource Controller
HetNet Heterogeneous Networks
KKT Karush Kuhn Tucker
MBS Macro Base Station
NBS Nash Bargaining Solution
OFDMA Orthogonal Frequency Division Multiple Access
PBS Pico Base Station
PSD Power Spectral Density
PU Primary User
QoS Quality of Service
RRM Radio Resource Management
SQP Sequential Quadratic Programming
SU Secondary User
SINR Signal to Interference Plus Noise Ratio
: List of acronyms.[]{data-label="tab1"}
Motivation and Contributions
----------------------------
Although the research works reported in recent years have considered the overall system’s performance maximization for CR-based HetNets, the problem of fairness among different users has received little attention. The potential problem arises when the schemes proposed for the sum-rate maximization assign very few or no resources to some of the users with higher fading conditions. This uneven distribution of resources results in degrading the achievable performance for different users. This problem may become more serious for the CR-based HetNets due to the limiting factor of increased interference. Thus, optimization of the transmission for user fairness under more practical constraints becomes essential. To the best of our knowledge, the resource optimization and user assignment techniques for fair rate allocation have not been jointly investigated in the literature due to the higher level of complexity involved in finding the optimal solution.
To fill this gap in the literature and provide a comprehensive solution to the user fairness problem in CR-based HetNets, we provide a joint strategy for power allocation and user fairness. In particular, we consider the joint power optimization and user association problem in HetNets for achieving fairness among different users. We first formulate a joint optimization problem subject to power and interference constraints. Then, to provide a less complex and an efficient solution, we design an algorithm from the dual decomposition strategy. The presented numerical results indicate the importance and utility of our scheme in comparison to the baseline methodologies.
Organization
------------
The remainder of this paper is organized as follows. The system model and problem formulation are presented in Section 2. In Section 3, the proposed solution has been described, while Section 4 discusses the simulation results. Finally, Section 5 presents some concluding remarks and future research directions. In addition, the list of acronyms used throughout this paper has been provided in Table \[tab1\].
{width="90.00000%"}
System Model and Problem Formulation
====================================
This section describes the considered system model and provides the steps related to problem formulation.
System Model
------------
A downlink CR transmission is considered, where the secondary HetNet system is reusing the spectrum of the primary network in an underlay mode, as shown in Fig. \[Fig.1\]. The HetNet consists of a single macrocell overlaid with multiple pico BSs, intended to transmit data to multiple users. The channel bandwidth is distributed among MBS and PBSs in such a way that MBS and PBSs are non-orthogonal to each other, while each PBS is orthogonal to other PBSs in the HetNet.
In this model, BSs are denoted as $BS_b$, such that $\mathbb{B}={BS_b|b=1,2,3,...,B}$, where $BS_1$ is modeled as macro and the remaining BSs act as pico BSs. Further, there are $A$ users and $C$ channels allocated to BS. The signal-to-interference-and-noise ratio (SINR) of $a$-th user from the $b$-th BS at the $c$-th channel is given as in [@4] $$\begin{aligned}
SINR_{a,c,b} = \frac{ P_{a,c,b} g_{a,c,b}}{P'_{a,c,b'} f_{a,c,b}+\sigma_a^2}, \end{aligned}$$ where $P_{a,c,b}$ is the transmit power for $a$-th user connected to $b$-th BS at $c$-th channel, $\alpha_{a,c,b}$ denotes the association variable which belongs to set ${ \{0,1} \}$, $\sigma_a^2$ is the power spectral density (PSD) of the noise while $g_{a,c,b}$ and $f_{a,c,b'}$ are the channel gains from intended BS to user and from interfering BS to the user, respectively. Furthermore, $P'_{a,c,b'}$ denotes the transmit power of interfering BS.
Problem Formulation
-------------------
One of our key objectives in this article is fair maximization of the data rate of each user in the network. To facilitate mathematical analysis, we introduce a binary variable $\alpha_{a,c,b}$, such that $$\begin{aligned}
\alpha_{a,c,b}=\!\begin{cases}
1,\,\,\, &\text{when $a$-th user is associated with $b$-th BS}\\ &\text{through $c$-th channel,}\\
0,\,\,&\text{otherwise,} \nonumber
\end{cases}.\end{aligned}$$ Based on the above expression, the channel allocation follows $$\begin{aligned}
&\sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b}=1\,\,\,\quad\quad\quad \, \forall\ a={1,2,3,...A}.\label{a3}\end{aligned}$$ To protect PUs from the interference and to improve the performance of the system, resource allocation at BSs need to be performed such that $$\begin{aligned}
\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b} P_{a,c,b} h_{a,c,b}\leq I_{th},\label{a4}\end{aligned}$$ where $P_{a,c,b}$ is the power allocated by $b$-th BS to $c$-th channel allocated to $a$-th user and $h_{a,c,b}$ represents the gain of interference channel $c$ from $b$-th BS to primary BS. Finally, to ensure that the power consumed by each BS is less than or equal to the power budget, power allocation at each BS needs to be ensured as $$\begin{aligned}
\sum\limits_{a=1}^A \sum\limits_{c=1}^C P_{a,c,b}\alpha_{a,c,b}\leq P_b\quad\ \forall\ b={1,2,3,...B}.\label{a5}\end{aligned}$$ To achieve fairness in rates of different users, a max-min-based optimization framework is adopted such that the problem is stated, mathematically as $$\begin{aligned}
\textbf{P1:}\quad&\max_{P_{a,c,b}, \alpha_{a,c,b}} \min\,\, \log_2 \left(1+SINR_{a,c,b} \right),\\
&\text{s.t.}\quad (\ref{a3}), (\ref{a4})\,\, \text{and}\,\,(\ref{a5})\nonumber,\end{aligned}$$ where $P_b$ denote total power available at $b$-th BS and $I_{th}$ represents the sum interference threshold. Here, the first constraint ensures that the total power allocated by $b$-th BS must be within the available power budget.
Proposed Solution
=================
To provide a viable solution to the aforementioned problem, we propose a solution based on duality theory. As is evident from the above expression, the allocation of channels and power loading are strongly coupled variables, thus, a joint optimization approach is needed. The problem **P1** is a mixed binary integer programming and requires an exhaustive search. Fortunately, it has been shown by [@zain] that for a sufficiently large number of subchannels the gap between the dual solution and the primal solution reduces to zero, regardless of the non-convexity of the original problem. Thus, we exploit the duality theory to decompose and optimally solve the coupled problem.
To convert the complex max min problem into standard optimization form, we introduce intermediate variable $t$ such that $$\begin{aligned}
t\ \leq \log_2 \left(1+SINR_{a,c,b} \right)\quad \forall a,c,b,\label{a10}\end{aligned}$$ The introduction of the intermediate variable transforms the problem as $$\begin{aligned}
\textbf{P2:}\quad&\max_{P_{a,c,b},t} \ t \label{obj1}\\
&\text{s.t.} \nonumber\quad (\ref{a3}), (\ref{a4}), (\ref{a5})\,\,\text{and}\,\,(\ref{a10}).\end{aligned}$$ To obtain an immediate solution of auxiliary variables, from the structure of objective in (\[obj1\]), we utilize the fact that for any $y \geq 0$, minimizing $y$ is equivalent to minimization of $y^{2}$. It is a known fact that maximizing $y$ is equal to minimizing $-y$. Hence, to transform the problem into a standard minimization problem we replace the objective in (\[obj1\]) with its negative. $$\begin{aligned}
\textbf{P3:}\quad&\min_{P_{a,c,b},t} -t^{2} \label{obj2}\\
&\text{s.t.} \nonumber\quad (\ref{a3}), (\ref{a4}), (\ref{a5})\,\,\text{and}\,\,(\ref{a10}).\end{aligned}$$ Now, substituting $x=-t$, we obtain The dual function associated with (\[obj3\]) is given by $$\begin{aligned}
&D \left(\lambda_{a}, \eta_{b},v \right)= \min_{x,P_{a,c,b},\alpha_{a,c,b}}\nonumber\\&\sum\limits_{a=1}^A\lambda_{a}\Big(-x- \sum\limits_{c=1}^C \sum\limits_{b=1}^B\alpha_{a,c,b} \log_2\! \left(1+SINR_{a,c,b}\right)\Big) \nonumber\\&+ x^{2} +\sum\limits_{b=1}^B \eta_b \Bigg(\sum\limits_{a=1}^A \sum\limits_{c=1}^C\alpha_{a,c,b} P{a,c,b}-P_{b} \Bigg) \nonumber\\& +v \left(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b}P_{a,c,b}h_{a,c,b} -I_{th}\right), \label{aa1}\\
&\text{s.t.}\quad\sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b}=1,\ \forall\ a={1,2,3,...A}.\nonumber\end{aligned}$$ The expression in (\[aa1\]) can be written as $$\begin{aligned}
&D \left(\lambda_{a}, \eta_{b},v \right)= \min_{x,P_{a,c,b},\alpha_{a,c,b}}x^{2} +\sum\limits_{a=1}^A\sum\limits_{c=1}^C \sum\limits_{b=1}^B\alpha_{a,c,b}\nonumber\\&\bigg(-\lambda_{a}\log_2 \left(1\!\!+\!SINR_{a,c,b} \right) + \eta_b P{a,c,b} +v P_{a,c,b}h_{a,c,b}\bigg) \nonumber\\&\!\! -\!\sum\limits_{a=1}^A\!\lambda_{a} x - \sum\limits_{b=1}^B\eta_{b} P_{b}-v I_{th}\label{aa2},\\
&\text{s.t.}\quad\sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b}=1,\ \forall a={1,2,3,...A}.\nonumber\end{aligned}$$ For any given channel allocation, dual decomposition guides to solve the following sub-problems $$\begin{aligned}
\textbf{P4:}\quad&\min_{x} \ \bigg(x^2- \ x \sum\limits_{a=1}^A \lambda_{a} \bigg), \label{22}\end{aligned}$$ $$\begin{aligned}
\textbf{P5:}\quad&\min_{P_{a,c,b}} \Big(-\lambda_{a} \log_2 (1+SINR_{a,c,b} \nonumber\\&+ \eta_b P_{a,c,b} + v P_{a,c,b} h_{a,c,b} \Big). \label{23a}\end{aligned}$$ Using the KKT conditions to find the solution of the problems given in (\[22\]) and (\[23a\]), we get $$\begin{aligned}
x^{*}= \frac{1}{2} \bigg( \sum\limits_{a=1}^A \lambda_{a} \bigg)^{+},\end{aligned}$$ $$\begin{aligned}
P_{a,c,b}^{*}= \left(\frac{\Phi_{a,c,b} - P_{a,c,b}' \alpha_{a,c,b} f_{a,c,b} }{g_{a,c,b} \left( \eta_b + v h_{a,c,b} \right)}\right)^{+}, \end{aligned}$$ where $\Phi_{a,c,b}=\lambda_{a} g_{a,c,b} + \sigma_a^2 \eta_b + v h_{a,c,b}$, for all $a,c,b$, when $(\Psi)^{+}=\max(0,\Psi)$. The detailed derivation steps have been provided in the Appendix.
Now, to find the optimum value of $\alpha_{a,c,b}$, following optimization problem is considered $$\begin{aligned}
&\min_{\alpha_{a,c,b}}\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b} \bigg(-\lambda_{a} \log_2 \Big(1+SINR_{a,c,b} \Big)\nonumber\\& + \eta_b P_{a,c,b}^{*}+ v P_{a,c,b}^{*} h_{a,c,b} \bigg), \label{23}\\
&\text{s.t.}\quad\sum\limits_{c=1}^C \sum\limits_{b=1}^B \alpha_{a,c,b}=1,\ \forall a={1,2,3,...A}.\nonumber \end{aligned}$$ Evidently, the optimal solution can be found, such that $$\begin{aligned}
\alpha_{a,c,b}^{*}=\begin{cases}
& \text{for}\, a=\text{arg} \min_{c}\!\Big(-\lambda_{a} \log_2 (1+SINR_{a,c,b})\nonumber\\& + \eta_b P_{a,c,b}^{*}+ v P_{a,c,b}^{*} h_{a,c,b} \Big),\\
&\text{otherwise}.
\end{cases}.\tag{18}\end{aligned}$$ The dual problem is convex, hence, sub-gradient method can be adopted to find the solution. The dual variables are updated at each iteration as $$\begin{aligned}
&\lambda_{a}^{ \left(itr \right)}=\bigg(\lambda_{a}^{ \left(itr-1 \right)}+\delta^{ \left(itr-1 \right)} \nonumber \\
&\times \Big(-\sum\limits_{b=1}^B \sum\limits_{c=1}^C\log_{2}(1\!\!+\!SINR_{a,c,b})-x \Big) \bigg),\tag{19}\\
&\eta_b^{ \left(itr \right)}= \bigg(\eta_b^{ \left(itr-1 \right)}+ \delta^{ \left(itr-1 \right)} \Big( \sum\limits_{a=1}^A \sum\limits_{c=1}^C P_{a,c,b}-P_b \Big) \bigg),\tag{20}\\
& v^{ \left(itr \right)}=\bigg(v^{ \left(itr-1 \right)}+\delta^{ \left(itr-1 \right)} \nonumber \\
&\times \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P_{a,c,b} h_{a,c,b}-I_{th}\Big) \bigg), \tag{21}\end{aligned}$$ where $\delta$ is the step size. To obtain the joint optimization solution, power loading and channel allocation are updated in each iteration.
Simulation Results
==================
In this section, we present the performance of the proposed scheme. The simulation environment for the proposed schemes is MATLAB. The noise PSD is $\sigma_a^2= 0.1$. For simulation we have considered the maximum number of users in the system is $A=30$, the total channels available in the system is $C=20$ and the total BS available in the system is $B=5$. For the sake of the evaluation, we have compared our proposed optimal allocation and optimal power scheme (OAOP) with baseline schemes, i.e., fixed channel allocation and optimal power loading (FAOP) and fixed channel allocation and fixed power loading (FAFP).
In Fig. \[Fig.2\], peak to average rate ratio (PR) has been plotted against different parameters. Different values of PR show fairness among the users. Moreover, the effect of changing picocell power $PR_p$ on PR has also been shown. Considering $P_p=0.25$, the percentage gap between FAFP and FAOP is $49.74\%$, between FAFP and OAOP is $22.59\%$ and between FAOP and OAOP is $45.42\%$. This is because optimization on channel allocation along with optimal power loading makes the problem more flexible compared to the cases when only power loading is optimized and if channel allocation and power loading are fixed. Also for $P_p=0.5$ the percentage gap between FAFP and FAOP is $49.83\%$, between FAFP and OAOP is $25.44\%$ and between FAOP and OAOP is $51.05\%$. Finally, increasing the value to $P_p=2$, the percentage gap between FAFP and FAOP is $56.99\%$, between FAFP and OAOP is $33.12\%$ and between FAOP and OAOP is $58.11\%$.
![Peak to average ratio with changing picocell power.[]{data-label="Fig.2"}](Peaktoavgratiowithchangingpicopower.eps)
Fig. \[Fig.3\] is showing the PR with changing interference constraint $PR_{I_{th}}$. The interference constraint has been varied from 20 to 40. Considering $I_{th}=20$, the percentage gap between FAFP and FAOP is $55.08\%$, between FAFP and OAOP is $31.83\%$ and between FAOP and OAOP is $57.79\%$. Also for $I_{th}=25$ the percentage gap between FAFP and FAOP is $53.39\%$, between FAFP and OAOP is $30.57\%$ and between FAOP and OAOP is $57.27\%$. Moreover, if $I_{th}=30$, then the percentage gap between FAFP and FAOP is $52.41\%$, between FAFP and OAOP is $29.95\%$ and between FAOP and OAOP is $57.14\%$. Further increasing the value of $I_{th}=35$, the percentage gap between FAFP and FAOP is $52.00\%$, between FAFP and OAOP is $29.62\%$ and between FAOP and OAOP is $57.11\%$. Finally, increasing the value to $I_{th}=40$, the percentage gap between FAFP and FAOP is $51.86\%$, between FAFP and OAOP is $28.91\%$ and between FAOP and OAOP is $56.31\%$. Comparison of schemes in terms of the percentage gap shows that as the interference constraint $I_{th}$ increases the percentage gap is decreasing. Moreover, the average gap between the graphs of FAFP with FAOP and FAOP with OAOP scheme is near to each other. Graphs also show that OAOP outperforms all other schemes in terms of fairness. The reason behind this is the additional flexibility provided by optimum channel allocation.
![Peak to average ratio with changing $I_{th}$.[]{data-label="Fig.3"}](PeaktoavgwithchangingIth.eps)
The effect of changing the power of macro base station $PR_m$ on the peak to average rate ratio is shown in Fig. \[Fig.4\]. The macro power has been varied from 10 to 30. Considering $PR_m=10$, the percentage gap between FAFP and FAOP is $54.64\%$, between FAFP and OAOP is $31.83\%$ and between FAOP and OAOP is $58.26\%$. Also for $PR_m=15$ the percentage gap between FAFP and FAOP is $56.05\%$, between FAFP and OAOP is $34.46\%$ and between FAOP and OAOP is $61.48\%$. Moreover, if $PR_m=20$, then the percentage gap between FAFP and FAOP is $56.65\%$, between FAFP and OAOP is $36.27\%$ and between FAOP and OAOP is $64.03\%$. Further increasing the value of $PR_m=25$, the percentage gap between FAFP and FAOP is $57.09\%$, between FAFP and OAOP is $37.59\%$ and between FAOP and OAOP is $65.84\%$. Finally, increasing the value to $PR_m=30$, the percentage gap between FAFP and FAOP is $57.49\%$, between FAFP and OAOP is $38.61\%$ and between FAOP and OAOP is $67.15\%$. Hence, comparing all the three schemes show that the OAOP is the superior scheme.
![Peak to average ratio with changing macro power.[]{data-label="Fig.4"}](Peaktoavgwithchangingmacropower.eps)
In the results presented here, the sum throughput has also been maximized along with achieving fairness among the users. Fig. \[Fig.5\] shows the sum throughput versus pico power $P_p$. The $P_p$ has been varied from 0.25 to 2. The percentage gap of OAOP scheme with FAOP is $38\%$, between OAOP with FAFP is $7\%$ and between FAOP with FAFP is $18.42\%$ where the value of $P_p=0.25$. Then, increasing the value of $P_p=0.5$ percentage gap of OAOP with FAOP is $42.56\%$, between OAOP with FAFP is $7.825\%$ and between FAOP with FAFP is $18.38\%$. Further increasing the value to $P_p=2$ OAOP with FAOP is $48.51\%$, between OAOP with FAOP is $10.12\%$ and between FAOP with FAFP is $20.85\%$.
![Sum throughput with changing pico power.[]{data-label="Fig.5"}](Sumthroughputwithchangingpicopower.eps)
Fig. \[Fig.6\] shows the sum throughput versus $I_{th}$. The percentage gap of OAOP scheme with FAOP is $45.98\%$, between OAOP with FAFP is $9.73\%$ and between FAOP with FAFP is $21.17\%$ where the value of $I_{th}=20$. Further increasing the value of $I_{th}=25$ percentage gap of OAOP with FAOP is $46.09\%$, between OAOP with FAFP is $10.22\%$ and between FAOP with FAFP is $22.17\%$. Moreover, if $I_{th}=30$, percentage gap between OAOP with FAOP is $46.21\%$, between OAOP with FAFP is $10.62\%$ and between FAOP with FAFP is $23\%$. Considering $I_{th}=35$, OAOP with FAOP is $46.32\%$, between OAOP with FAFP is $10.98\%$ and between FAOP with FAFP is $23.71\%$. Further increasing the value to $I_{th}=40$ OAOP with FAOP is $46.47\%$, between OAOP with FAOP is $11.01\%$ and between FAOP with FAFP is $23.71\%$. Hence, the sum throughput of OAOP scheme is maximum.
![Sum throughput with changing $I_{th}$.[]{data-label="Fig.6"}](SumthroughputwithchangingIth.eps)
The effect of varying macro power $P_m$ on the sum throughput is shown in Fig. \[Fig.7\]. The $P_m$ has been varied from 10 to 30. The percentage gap of OAOP scheme with FAOP is $45.49\%$, between OAOP with FAFP is $9.73\%$ and between FAOP with FAFP is $21.40\%$ where the value of $P_m=10$. Further increasing the value of $P_m=15$ percentage gap of OAOP with FAOP is $46.69\%$, between OAOP with FAFP is $9.65\%$ and between FAOP with FAFP is $20.67\%$. Moreover, if $P_m=20$, percentage gap between OAOP with FAOP is $47.27\%$, between OAOP with FAFP is $9.61\%$ and between FAOP with FAFP is $20.32\%$. Considering $P_m=25$, OAOP with FAOP is $47.74\%$, between OAOP with FAFP is $9.58\%$ and between FAOP with FAFP is $20.06\%$. Further increasing the value to $P_m=30$ OAOP with FAOP is $48.13\%$, between OAOP with FAOP is $9.56\%$ and between FAOP with FAFP is $19.86\%$. The results show that OAOP scheme is the most optimal.
![Sum throughput with changing macro power.[]{data-label="Fig.7"}](Sumthroughputwithchangingmacropower.eps)
Fig. \[Fig.8\] shows the values of sum throughput with a changing number of users $U$. The $U$ has been varied from 20 to 100. Considering $U=20$, the percentage gap of OAOP scheme with FAOP is $61.07\%$, between OAOP with FAFP is $14.82\%$ and between FAOP with FAFP is $24.27\%$. Further increasing $U=40$ percentage gap of OAOP with FAOP is $53.28\%$, between OAOP with FAFP is $10.99\%$ and between FAOP with FAFP is $20.63\%$. Moreover, if $U=60$, the percentage gap between OAOP with FAOP is $48.81\%$, between OAOP with FAFP is $10.79\%$ and between FAOP with FAFP is $20.10\%$.
![Sum throughput with increasing users[]{data-label="Fig.8"}](Sumthroughputwithincreasingusers.eps)
Fig. \[Fig.9\] shows the convergence of dual variable $\lambda_a$, $\eta$ and $v$ with initial value as 0.6, $\eta$ as 0.6 and v as 0.1. The smaller the value of step size more fine the convergence is, a bigger value leads to fast convergence.
![Convergence of $\eta,\lambda,v$ with increasing iterations.[]{data-label="Fig.9"}](CombinedFigure.eps)
Conclusion
==========
This paper optimized the power allocation and user association to achieving fairness among different users in underlay CR-based HetNets. Specifically, we adopted a max-min-based fairness framework under various practical constraints. Problem was first transformed into a standard maximization optimization and then dual decomposition is used to solve the integer programming problem. The dual problem was solved through the sub-gradient method. The power allocation at each node is obtained from the convex optimization techniques for the known power allocation at all other nodes. The results are compared with the sub-optimal scenarios when the fixed channel allocation is assumed for the optimized/non-optimized power allocation. The results showed that the proposed scheme outperforms all other candidates and the spectral efficiency has been improved due to the simultaneous transmission with the primary network.
Appendix {#appendix .unnumbered}
========
The Lagrangian L associated with the optimization problem **P5** is $$\begin{aligned}
&\min_{P_{a,c,b}} \left(-\lambda_{a} \log_2 \left(1+SINR_{a,c,b} \right) + \eta_b P_{a,c,b} + v P_{a,c,b} h_{a,c,b} \right). \nonumber\end{aligned}$$ In the above expression, we have converted our problem from a constrained problem to an unconstrained problem. $$\begin{aligned}
&
\min_{P_{a,c,b \geq 0}} \Bigg(-\lambda_{a} \log_2 \bigg(1+\frac{ P_{a,c,b} g_{a,c,b}}{ \sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2} \bigg) \nonumber\\&+ \eta_b P_{a,c,b} + v P_{a,c,b} h_{a,c,b} \Bigg). \nonumber\end{aligned}$$ By substituting the value of SINR, we get $$\begin{aligned}
&
= \frac{\partial}{\partial P_{a,c,b}} \Bigg(-\lambda_{a} \log_2 \bigg(1+\frac{P_{a,c,b} g_{a,c,b}}{\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2} \bigg)\nonumber\\& + \eta_b P_{a,c,b} + v P_{a,c,b} h_{a,c,b} \Bigg). \nonumber\end{aligned}$$
$$\begin{aligned}
= &-\lambda_a \Bigg(\frac{1}{1+\frac{P_{a,c,b}+g_{a,c,b}}{\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2} }\Bigg)\nonumber\\&\times \frac{\partial}{\partial P_{a,c,b}} \Bigg( 1+ \frac{P_{a,c,b} g_{a,c,b}}{ \sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2} \Bigg)+ \eta_b + v h_{a,c,b}. \nonumber\end{aligned}$$
Now, solving the partial derivative as $$\begin{aligned}
& = \eta_b + v h_{a,c,b}-\lambda_a \nonumber \\&\times\Bigg( \frac{\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2 }{\Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2 \Big)+ P_{a,c,b} g_{a,c,b}} \Bigg) \nonumber \\& \times\Bigg( \frac{ g_{a,c,b} \sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2} {\Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2 \Big)^2} \Bigg). \nonumber \end{aligned}$$
Taking L.C.M. of previous step and finding the partial derivative of internal function, we have $$\begin{aligned}
&= -\lambda_a \Bigg(\frac{g_{a,c,b} \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2 \Big)^2}{ \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2 \Big)+ P_{a,c,b} g_{a,c,b}} \Bigg) \nonumber \\& \times\Bigg( \frac{1}{\Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2\Big)^2} \Bigg)+ \eta_b + v h_{a,c,b}. \nonumber\end{aligned}$$
Further simplification results in $$\begin{aligned}
&
= -\lambda_a \Bigg(\frac{g_{a,c,b}}{ \sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b}+\sigma_a^2+ P_{a,c,b} g_{a,c,b}} \Bigg) \nonumber \\
& + \eta_b + v h_{a,c,b}.\end{aligned}$$
By canceling the common terms and simplifying, we have $$\begin{aligned}
&
-\eta_b -v h_{a,c,b}=-\lambda_{a} \nonumber \\
& \Bigg( \frac {g_{a,c,b}}{ \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b} + \sigma_a^2 \Big)+ P_{a,c,b} g_{a,c,b}}\Bigg) \nonumber
\end{aligned}$$ $$\begin{aligned}
&
\frac{1}{ \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b} + \sigma_a^2 \Big)+ P_{a,c,b} g_{a,c,b}} = \frac{ \eta_b +v h_{a,c,b}}{\lambda_{a,c,b} g_{a,c,b}}. \nonumber \end{aligned}$$
After some mathematical simplifications, we get $$\begin{aligned}
&
\bigg(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b} + \sigma_a^2 \bigg)+ P_{a,c,b} g_{a,c,b}= \frac{\lambda_{a,c,b} g_{a,c,b}}{\eta_b +v h_{a,c,b}}. \nonumber
\end{aligned}$$
Inverting both sides of the equation as $$\begin{aligned}
&
P_{a,c,b} g_{a,c,b}= \Big(\frac{\lambda_{a,c,b} g_{a,c,b}}{ \eta_b +v h_{a,c,b}} \Big)- \Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b} + \sigma_a^2 \Big). \nonumber \end{aligned}$$
After cross multiplying, we have $$\begin{aligned}
&
P_{a,c,b}= \frac{\lambda_{a,c,b} g_{a,c,b}-\bigg( \Xi \eta_b + v h_{a,c,b} \bigg)}{\eta_{b}+v h_{a,c,b}}. \nonumber\end{aligned}$$ where $\Xi=\Big(\sum\limits_{a=1}^A \sum\limits_{c=1}^C \sum\limits_{b=1}^B P'_{a,c,b} f_{a,c,b} + \sigma_a^2 \Big)$.
[^1]: Khush Bakht is with the Department of Electronic Engineering, Fatima Jinnah Women University Rawalpindi 46000, Pakistan (email: [email protected]).
[^2]: Furqan Jameel is with the Department of Communications and Networking, Aalto University, 02150 Espoo, Finland. (email: [email protected]).
[^3]: Zain Ali, Guftaar Ahmad Sardar Sidhu are with the Department of Electrical Engineering, COMSATS University, Islamabad 45550, Pakistan (email: [email protected], [email protected]).
[^4]: Wali Ullah Khan is with the School of Information Science and Engineering, Shandong University, Qingdao 266237, China (email: [email protected]).
[^5]: Imran Khan is with the University of Engineering and Technology, Peshawar Pakistan (email: [email protected]).
[^6]: Jeong Woo Lee is with the School of Electrical and Electronics Engineering, Chung-Ang University, Seoul, South Korea (email: [email protected]).
[^7]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We calculate numerically the quasiparticle effective mass ($m^*$) renormalization as a function of temperature and electron density in two- and three-dimensional electron systems with long-range Coulomb interaction. In two dimensions, the leading temperature correction is linear and positive with the slope being a universal density independent number in the high density limit. We predict an enhancement of the effective mass at low temperatures and a non-monotonic temperature dependence at higher temperatures ($T/T_F
\sim 0.1$) with the peak shifting toward higher temperatures as density decreases. In three dimensions, we find that the effective mass temperature dependence is nonlinear and non-universal, and depends on the electron density in a complicated way. At very high densities, the leading correction is positive, while at lower densities it changes sign and the effective mass decreases monotonically from its zero temperature value with increasing temperature.
author:
- Ying Zhang
- 'S. Das Sarma'
bibliography:
- 'massprb.bib'
title: Temperature dependent effective mass renormalization in a Coulomb Fermi liquid
---
Introduction
============
In the Fermi liquid theory the interacting electron system is composed of weakly interacting quasiparticles at low energies with long quasiparticle lifetimes. The effective mass of a quasiparticle, which can be viewed as the bare mass of a free electron being renormalized by electron-electron interactions, is an important and fundamental Fermi liquid parameter. For decades theorists have been exploring the effective mass renormalization in two- and three- dimensional interacting electron systems (2DES and 3DES). In spite of this great deal of theoretical activity concentrating almost entirely on the density dependence of the effective mass renormalization, the temperature dependence of the effective mass has not been studied until very recently. Besides the considerable difficulties involved in the finite temperature numerical and analytical many-body calculations in 2DES and 3DES, the reason that this issue has not been addressed before can also be explained by the fact that the Fermi energy in 3D metals is typically $10^4$K, and therefore any finite temperature effects are negligible. In the past decade, however, low density 2DES have been attracting attention, and several experiments have been performed to measure the 2D effective mass [@oldexp; @exp]. The temperature dependence of the quasiparticle effective mass in 2DES is of considerable experimental interest since the Fermi energy in realistic 2DES may be $1K$ or lower, which makes the issue of the temperature dependence of 2D Fermi liquid parameters extremely important. In addition, the temperature dependence of the Fermi liquid parameters such as the effective mass is obviously of considerable fundamental theoretical significance.
The $T=0$ quasiparticle effective mass renormalization in an electron system interacting through the long-range Coulomb interaction is one of the oldest many-body problems in theoretical condensed matter physics, and a number of theoretical calculations of 3D and 2D electron effective mass have been carried out [@rice; @gellmann; @ting; @vinter; @jalabert; @marmorkos; @book] in the literature. In fact, the Coulomb interaction induced electron effective mass renormalization at $T=0$ is standard text-book material [@book] in electronic many-body theory. Essentially all of these calculations, both analytical and numerical (and both 2D and 3D), are based on the leading-order dynamically screened interaction one-loop self-energy evaluation (the so-called RPA or ‘GW’ self-energy approximation) because this approximation is really the only meaningful nontrivial calculation that can actually be carried out, and (perhaps more importantly) because this RPA self-energy is asymptotically exact in the weakly interacting high density regime. There have been a few finite temperature RPA self-energy calculations over the years [@chaplik; @gq; @hu; @dassarma] mostly in the context of low dimensional systems, but none for the temperature dependence of the effective mass renormalization in interacting electron systems. Very recently, Chubukov and Maslov [@chubukov] considered the problem of temperature corrections to the 2D Fermi liquid theory for the case of a short-ranged interaction. In particular, they showed that the leading many-body temperature correction is linear in 2D similar to the results which we reported recently for the long-range Coulomb interaction in 2DES [@short].
In the current paper we present a calculation of the density and temperature dependent effective mass renormalization by the Coulomb interaction in 2DES and 3DES at arbitrary densities and temperatures (i.e. [*not*]{} necessarily restricted to high densities and low temperatures). We work within the random-phase approximation (RPA), or equivalently in the ring-diagram approximation for the self-energy, which gives exact results in the high-density limit ($r_s \ll 1$) but is known to be qualitatively reliable at relatively low densities as well. RPA is perhaps the only manageable way to perform any non-trivial quantitative calculations in electronic many-body systems, and the finite temperature RPA effective mass renormalization is certainly a problem of intrinsic interest. In two dimensions, our numerical results predict a non-monotonic effective mass temperature dependence. The leading temperature dependence is linear and positive, with the low-temperature slope being independent of the electron density in the high density limit. The temperature at which the effective mass is maximum at a particular density moves toward higher temperatures as density decreases. In three dimensions, we find that the effective mass temperature dependence is non-universal and depends on the electron density in a complicated way. At very high densities, the leading correction is positive, while at lower densities it changes sign and decreases monotonically from its zero temperature value. This is in contrast to the 2D results where the effective mass always increases (linearly) with temperature at low temperatures, and then decreases with temperature beyond a density-dependent characteristic temperature. We find the 3D temperature correction to the effective mass to be nonlinear in contrast to our 2D results.
We express the quasiparticle effective mass $m^*(n, T) \equiv m^*(r_s,
T/T_F)$ in units of the bare band mass $m$ (which is, by definition, a constant) and present our results as a function of the usual dimensional interaction parameter $r_s$ (the average inter-electron separation measured in the units of Bohr radius) and the dimensionless temperature $T/T_F$ where $T_F = E_F / k_B$ is the Fermi temperature. Note that $r_s \propto n^{-1/2}~(n^{-1/3})$ in 2D (3D) systems, and $T_F \propto n~(n^{2/3})$ in 2D (3D) systems, where $n$ is the appropriate 2D (per unit area) or 3D (per unit volume) electron density. Note that the dimensionless interaction and temperature parameters $r_s$ and $T/T_F$ are [*not*]{} independent parameters since they both depend on the electron density. We also note that $r_s
\ll 1$ (high-density) and $r_s \gg 1$ (low density) limits are respectively the weak- and the strong-interaction limits of the electron system (at $T=0$), and $T/T_F \ll 1$ and $T/T_F \gg 1$ are respectively the low-temperature (quantum) and the high-temperature (classical) limits. We consider the electron system to be a uniform jellium system with the noninteracting kinetic energy dispersion being the usual parabolic dispersion. We use $\hbar = k_B = 1$ throughout.
The structure of our paper is as follows: In section \[sec:form\] we provide the formalism which we will use in this paper. In section \[sec:method\] we explain in detail the numerical method we are using in the effective mass calculations. In section \[sec:results\] we present all our numerical results for 2D and 3D effective mass, comparing to analytical results in the high density limit. In section \[sec:PPA\] we discuss a special approximation method, the plasmon-pole approximation, and present our effective mass results using this method. In section \[sec:decay\], we calculate the imaginary self-energy of quasiparticles and discuss the validity of the quasiparticle approximation at finite temperatures. We provide a conclusion and discussion of our results in section \[sec:con\].
Formalism {#sec:form}
=========
In this section we give the theoretical formalism, the basic equations, and the notations which will be used throughout the paper.
Effective mass
--------------
In a system of interacting fermions the retarded Green’s function can be written as $$\label{Gr}
G_R ({\bf k}, \omega)
= {1 \over\omega -\epsilon_0({\bf k})+\mu + \Sigma({\bf k},\omega)},$$ where $\epsilon_0({\bf k}) = k^2 /2m$ is the spectrum of non-interacting fermions, $\mu$ is the chemical potential, and $\Sigma({\bf k},\omega)$ is the quasiparticle self-energy, the imaginary (real) part of which determines the lifetime (effective mass) of the quasiparticle. The quasiparticle energy can be obtained by solving the Dyson’s equation [@book] $$\label{eq:dyson}
\epsilon({\bf k}) = \epsilon_0({\bf k})+ {\rm
Re\,}\Sigma({\bf k},\epsilon({\bf k})).$$ The quasiparticle effective mass can be written by definition as $$\begin{aligned}
\label{eq:offshell}
{m^* \over m} &=& {m \over k} {d \over d k} \epsilon({\bf k})
\Big|_{k=k_F} \nonumber \\
&=& {1 - {\partial \over \partial \omega} {\rm
Re\,} \Sigma\left({\bf k}, \omega\right)
\over 1 + {m \over k} {\partial \over \partial p} {\rm Re\,}
\Sigma({\bf k}, \omega)} \Big|_{k = k_F, \omega = 0}. \end{aligned}$$ Note that in the above equation $\omega = 0$ is measured from the renormalized chemical potential $\mu^*$, which is given by $$\label{eq:mu}
\mu^* = \mu + {\rm Re} \Sigma(k_F, 0).$$
All the above equations are exact, while the RPA approximation for $\Sigma({\bf k}, \omega)$ that we are going to use is the first order perturbation theory in the dynamically screened interaction. There has been extensive discussion [@rice; @ting; @book; @diverge] on whether one should use exact Eq.(\[eq:offshell\]) for calculating the effective mass or it is more consistent to use the so-called on-shell approximation, keeping only the first order interaction terms in the expression for the effective mass (since $\Sigma$ is calculated only to first order in the dynamically screened interaction): $$\label{eq:onshellm}
{m^* \over m} = {1 \over
1 + {m \over k} {d \over d k}
\Sigma \left({\bf k}, \xi_{\bf k} \right)\big|_{k=k_F} },$$ where $\xi_{\bf k} = k^2 /(2m) - \mu$. Note that all the quantities on the right side of Eq. (\[eq:onshellm\]) are in the leading order in effective interaction. There are compelling arguments in favor of the latter choice: the on-shell approximation is believed to be more accurate as it effectively accounts for some higher order diagrams and satisfies the Ward identity. We have extensively discussed this issue elsewhere [@diverge].
Obviously, the two equations for calculating $m^*$ are identical in the high-density limit $r_s \ll 1$. However, in the region of $r_s >
1$, they give very different results. In what follows, we use Eq. (\[eq:onshellm\]) for all the numerical results shown in this paper because we believe the on-shell approximation to be the superior one in our case. Both formulae give similar temperature dependence for $m^*(T)$. The main qualitative results of the paper are insensitive to the choice of the on-shell or off-shell formula for the effective mass.
Self-energy in the RPA approximation.
-------------------------------------
![Feynman diagram for RPA self-energy calculation. Solid lines denote the free electron Green’s function and the dashed lines the bare Coulomb potential.[]{data-label="fig1"}](fig1.eps){width="2in"}
Within RPA, the finite temperature electron self-energy can be expressed in terms of the Feynman diagrams shown in Fig. \[fig1\], and can be written in the Matsubara formalism as [@book]: $$\label{eq:EMats}
\Sigma ({\bf k}, \nu_n) = - T \sum\limits_{\omega_m} {\cal
G}({\bf k}-{\bf q}, \nu_n - \omega_m) {\cal D}({\bf q},\omega_m),$$ where $ \nu_n = \pi (2 n +1) T$ is the fermion Matsubara frequency, $\omega_m = 2 \pi m T$ is the boson Matsubara frequency with $n$ and $m$ integers, and $T$ the temperature. The function ${\cal D}({\bf
q},\omega_m)$ denotes the coupling to a collective mode (phonon, plasmon, electron-hole excitation, etc.), [*i.e.*]{}, ${\cal D}$ is the bosonic propagator for the effective interaction. In our case, the function is the dynamically screened Coulomb interaction given by the sum of the ring or bubble diagrams: $$\label{eq:scr_Coulomb}
{\cal D}({\bf q},\omega_m) = {v_0({\bf q}) \over 1 + v_0({\bf q})
\mbox{\Large $\pi$} ({\bf q},\omega_m) },$$ where $v_0({\bf q})$ is the bare Coulomb interaction and $\mbox{\Large
$\pi$}({\bf q},\omega_m)$ is the (bare) polarization operator, which is defined as $$\begin{aligned}
\label{eq:PMats}
\mbox{\Large $\pi$} \left({\bf q},\omega_m\right) &=&
2 \sum\limits_{\nu_n} \int {d ^d {\bf p} \over \left(2
\pi \right)^d} {\cal G}^{(0)} ({\bf p}, \nu_n)
\nonumber \\
&&~~~~~~~~~~~\cdot {\cal G}^{(0)}({\bf p + q}, \nu_n + \omega_m), \end{aligned}$$ where $d$ is the dimension of the system and “$(0)$” denotes the non-interacting system. We mention that Eqs. (\[eq:scr\_Coulomb\]) and (\[eq:PMats\]) together form what is called the RPA for an electron gas, where the bare Coulomb interaction is dynamically screened by the electron dielectric function, which is formed from the infinite series of the polarization bubbles. The corresponding electron self-energy, obtained in the leading-order expansion in the dynamically screened interaction ${\cal D}$, is conventionally called the RPA self-energy approximation, although the “dynamical Hatree-Fock” approximation or the “Ring Diagram Approximation” may be a more appropriate terminology.
For calculations, it is more convenient to use the self-energy defined as a function of the real frequency $\omega$ rather than the Matsubara one. Using the standard procedure of analytic continuation, one obtains the following expression for the analytically continued self-energy: $$\begin{aligned}
\label{eq:Sig_gen}
&&\Sigma^R \left( {\bf k},\omega \right) =
-\int {d^d {\bf q} \over (2 \pi)^d}
\int\limits_{-\infty}^{+\infty} {d \nu \over 2 \pi}
\Big[ \nonumber \\
&&~~~~~~~{\rm Im\,} G_R^{(0)}
\left({\bf k} - {\bf q}, \nu + \omega \right)
D_R\left({\bf q}, -\nu \right) \tanh ( {\nu +
\omega \over 2 T} ) \nonumber \\
&&~~~+ G_R^{(0)}
\left({\bf k} - {\bf q}, \nu + \omega \right)
{\rm Im\,} D_R\left({\bf q}, \nu \right) \coth ( {\nu
\over 2 T} ) \Big], \end{aligned}$$ where functions labeled with index “R” are retarded functions, [ *i.e.*]{} functions analytical in the upper half-planes of the complex frequency. The corresponding effective interaction can be written as: $$\label{eq:scr_Coulomb1}
D_R({\bf q},\omega ) = {v({\bf q}) \over 1 + v({\bf q})
\Pi_R ({\bf q},\omega ) },$$ where the retarded polarizability can be obtained from Eq.(\[eq:PMats\]) using the following identities: $$\label{PRA}
\Pi_R ({\bf q},\omega)=\mbox{\Large $\pi$}
({\bf q},i \omega_n \to \omega + i \eta),$$ where $\eta$ is a real infinitesimal positive number.
Note that we will almost always use retarded quantities unless otherwise stated. Thus without causing any confusion, we can drop the superscript “R”.
Effective interaction {#sec:form:eff}
---------------------
The next step toward deriving the renormalization of mass is to obtain expressions for the effective coupling $D({\bf q},\omega)$. We use the long-range bare Coulomb interaction to get $$\begin{aligned}
\label{eq:bareV}
v_0^{(2D)}(q) &=& {2 \pi e^2 \over q}, \nonumber \\
v_0^{(3D)}(q) &=& {4 \pi e^2 \over q^2},\end{aligned}$$ and the effective interaction $$\label{eq:scr_Coulomb2}
D({\bf q},\omega) = {v_0({\bf q}) \over 1 + v_0({\bf q}) \Pi
({\bf q},\omega) } = {v_0({\bf q}) \over \epsilon({\bf q},
\omega)},$$ where $\epsilon({\bf q}, \omega) \equiv 1 + v_0 \Pi$ is the RPA dynamical dielectric function. In the RPA the full polarizability is approximated by the bare polarizability as in Eq. (\[eq:PMats\]), which is just the bare bubble diagram.
Analytical properties of the propagator $\Pi_0$ (where $\Pi_0$ denotes the $T=0$ form for the bare polarizability with $\Pi$ denotes the finite $T$ bare polarizability) are non-trivial even at zero temperature. The zero temperature polarization for 2DES and 3DES are well-known and shown below. For 2D $T=0$ case we have $$\begin{aligned}
\label{eq:2DP}
\Pi^{\mbox{2D}}_0 (q, \omega, \mu)
= -{m \over \pi} + {m^2 \over \pi q^2}
\Bigg[ \sqrt{ (\omega + {q^2 \over 2m} )^2 - {2 \mu q^2 \over 2m} }
\nonumber \\
- \sqrt{ (\omega - {q^2 \over 2m} )^2
- {2 \mu q^2 \over 2m} } \Bigg],\end{aligned}$$ where $\mu$ is the chemical potential, the frequency $\omega$ can be any complex number, and the branch cut of the square roots are taken so that the imaginary part is positive. For 3D ($T=0$) case we have $$\begin{aligned}
\label{eq:2DPT}
\Pi^{\mbox{3D}}_0 (q, \omega, \mu)
&=& {k_\mu m \over 2 \pi^2 q^2} \Bigg\{ 1 \nonumber \\
&& \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
+ {m^2 \over 2 k_\mu q^3}
\left[4 \mu \epsilon_q - (\epsilon_q + \omega)^2 \right]
\ln \left({\epsilon_q + q v_\mu + \omega
\over \epsilon_q + q v_\mu + \omega} \right) \nonumber \\
&& \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
+ {m^2 \over 2 k_\mu q^3}
\left[4 \mu \epsilon_q - (\epsilon_q - \omega)^2 \right]
\ln \left({\epsilon_q + q v_\mu - \omega
\over \epsilon_q + q v_\mu - \omega} \right) \Bigg\},\end{aligned}$$ where $\epsilon_q = q^2 /2m$, $\mu$ is the chemical potential and $\mu
= k_\mu^2 /2m = m v_\mu^2 /2$, and the frequency $\omega$ can be any complex number.
Finite temperature polarizability can be easily obtained from those at zero temperature using the following identity: $$\label{eq:finiteTP}
\Pi (q, \omega, \mu; T) = \int\limits_0^\infty d \mu'
{\Pi_0(q, \omega, \mu') \over 4 T \cosh^2 ({\mu' - \mu \over 2T}) }.$$ We find Eq. (\[eq:finiteTP\]) to be the most convenient numerical method for obtaining the finite-$T$ polarizability.
Dimensionless parameters {#sec:form:para}
------------------------
Our 2D and 3D electron system can be characterized by two parameters, namely density ($n$) and temperature ($T$). This immediately leads to two dimensionless parameters $r_s$ and $T/T_F$ characterizing the system with $r_s$ being the effective zero-temperature interaction strength and $T/T_F$ being the effective temperature (note that they are [*not*]{} independent). The definition of $r_s$ is the following. In 2DES $r_s$ is defined such that $$\begin{aligned}
\label{eq:2Drs}
\pi r_s^2 a_B^{2} n &=& 1,\\
k_F r_s a_B &=& \sqrt{2},\end{aligned}$$ where $n$ is the 2D electron density, $k_F$ is the Fermi momentum, and $a_B = (m e^2)^{-1}$ is the Bohr radius. In 3DES $r_s$ is defined such that $$\begin{aligned}
\label{eq:3Drs}
4 \pi n a_B^3 r_s^3 /3 &=& 1, \\
k_F r_s a_B &=& (9 \pi /4)^{1/3}.\end{aligned}$$ The Fermi temperature $T_F \equiv E_F \equiv k_F^2 /(2m)$, which goes as $T_F \propto r_s^{-2}$ in both 2D and 3D.
Numerical methods in $m^*$ calculations in RPA {#sec:method}
==============================================
In this section we explain in detail our numerical approach for the effective mass calculation within RPA. In carrying out the integrations of self-energy in Eq. (\[eq:Sig\_gen\]) in order to obtain the effective mass, we use three different techniques, namely frequency sum, frequency integration, and plasmon-pole approximation (PPA). The first two techniques are equivalent, and we explain them in detail in this section. PPA is a further approximation of RPA, which has been extensively used in the literature [@lundqvist; @dassarma; @vinter]. We discuss the PPA in section \[sec:PPA\]. Since there is no existing literature on the finite temperature effective mass or self-energy calculation to check our numerical results, it is crucial for us to use these different techniques to ensure the correctness of our numerical calculations. We mention here that our frequency sum results and frequency integration results agree well with each other. The frequency integration result is numerically relatively more noisy and therefore in this paper we will only show the frequency sum results. We also check our numerical results against the already known results at $T=0$ and against analytical calculations in the $T/T_F, r_s \to 0 $ limit.
Frequency integration technique {#sec:method:int}
-------------------------------
Eq. (\[eq:Sig\_gen\]) gives the general formula for the RPA self-energy at real frequencies. It can also be written in a more succinct way as
$$\begin{aligned}
\label{eq:Eint}
\Sigma ({\bf k}, \omega) &=&
- \int \frac{d^d q}{(2 \pi)^d}
v_0(q) n_F ( \xi_{{\bf q} - {\bf k}} ) \nonumber \\
&& -\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \varepsilon}{2 \pi}
\frac{ 2 v_0(q) \mbox{Im} \epsilon^{-1} (q, \varepsilon )} {
\varepsilon + \omega + i
\eta - \xi_{{\bf q} - {\bf k}} } \nonumber \\
&&~~~~\cdot \left[ n_F ( \xi_{{\bf q} - {\bf k}} ) + n_B ( \varepsilon )
\right], \end{aligned}$$
where $n_F(x) = 1/(\exp(x/T) + 1)$ is the Fermi function and $n_B(x) =
1/(\exp(x/T) - 1)$ the Bose function. This method of calculating the self-energy involves integration over real frequencies, and therefore we call it the frequency integration method. It is also known as the spectral or the Lehmann representation of the self-energy. The derivation of Eq. (\[eq:Eint\]) from Eq. (\[eq:Sig\_gen\]) is given in the appendix.
The self-energy of Eq. (\[eq:Eint\]) is composed of two parts: the exchange part and the correlation part. The (frequency independent) exchange part is also known as the Hartree-Fock self-energy, and its contribution to the effective mass at $T=0$ is singular in both 2D and 3D. Not surprisingly, this singularity is cancelled out by contributions from the correlation part of the self-energy. Effective mass is derived from the self-energy through Eq. (\[eq:onshellm\]), and we therefore need to obtain the real part of Eq. (\[eq:Eint\]) by putting $i \eta$ to be $0$ and regarding the frequency integration as a principal value integration. It is easy to derive from Eq. (\[eq:Eint\]) that the imaginary part of the self-energy can be written as $$\begin{aligned}
\label{eq:Eimag}
\mbox{Im} \Sigma({\bf k}, \omega) &=& - \int \frac{d^d q}{(2 \pi)^d} v_0(q)
\mbox{Im}\epsilon^{-1}({\bf q}, \xi_{{\bf q} - {\bf k}} - \omega))
\nonumber \\
&&~~~~~\cdot \left[n_B(\xi_{{\bf q} - {\bf k}} - \omega)
+n_F(\xi_{{\bf q} - {\bf k}}) \right].\end{aligned}$$ The ${\rm Im} \Sigma$ is not needed in the effective mass calculation since $m^*$ is a Fermi surface property. But it is important to have some idea of the magnitude of ${\rm Im} \Sigma$ in order to ensure that quasi-particles are well defined at finite $T$.
Numerically carrying out the integration in Eq. (\[eq:Eint\]) is non-trivial: for each momentum $\bf q$ and frequency $\omega$, a three dimensional integration is required to obtain $\Sigma({\bf q},
\omega)$, and what makes the problem even more difficult is that the $\mbox{Im} \epsilon^{-1}({\bf q}, \omega) $ term in the integrand is highly non-monotonic. A careful examination of the dynamical dielectric function tells us that at $T=0$, $\mbox{Im}
\epsilon^{-1}({\bf q}, \omega)$ contains delta-functions at plasmon excitation frequencies, and at finite temperatures these delta-functions broaden into sharp peaks. Integration over these sharps peaks requires special care. For each ${\bf q}$, the position (i.e., frequency) of the sharp peaks can be determined by solving $\mbox{Re} [\epsilon({\bf q}, \omega)] = 0$, and their weight can be determined from $\mbox{Re} [\epsilon({\bf q}, 0)]$ using the Kramers-Krönig relations.
One advantage of the frequency integration method is that in Eq. (\[eq:Eint\]) we can directly put $T=0$ to obtain the zero temperature result, in contrast to the frequency sum method which we will describe in detail below.
Frequency sum technique {#sec:method:sum}
-----------------------
Due to the great numerical difficulty in carrying out the frequency integration method introduced above (because it involves integration over highly non-monotonic or singular functions), it is advisable to seek alternatives. At zero temperature, previous works in calculating self-energy and related quantities often transform the real frequency integration into integrations over imaginary frequencies using the analytic properties of the dielectric function. The purpose of this contour distortion is to avoid singularities along the real axis. At finite temperature, a similar approach can be adopted. At finite temperature, what is different from the zero temperature case is that we transform the integration into an imaginary frequency summation (or Matsubara frequency summation). Hu [*at al*]{}. [@hu] showed in detail how to perform such a transformation from the real-frequency integration to an imaginary frequency summation. Following the technique of contour distortion introduced in Ref. [@hu] we can write the RPA self-energy as a sum of the Matsubara frequency along the imaginary axis: $$\begin{aligned}
\label{eq:Esum}
\Sigma({\bf k}, \omega) =&-& \int \frac{d^d q}{(2 \pi)^d}
v_0(q) n_F(\xi_{{\bf q} - {\bf k}}) \nonumber \\
&-& \int \frac{d^d q}{(2 \pi)^d} v_0(q) \left[
\frac{1}{\epsilon(q, \xi_{{\bf q} - {\bf k}} - \omega)} - 1 \right]
\nonumber \\
&&~~~~~\cdot \left[n_B(\xi_{{\bf q} - {\bf k}} - \omega)
+n_F(\xi_{{\bf q} - {\bf k}}) \right] \nonumber \\
&-& \int \frac{d^d q}{(2 \pi)^d} T \sum_{\omega_n}
v_0(q) \left[ \frac{1}{\epsilon(q, i \omega_n)} - 1 \right]
\nonumber \\
&& ~~~~ \cdot \frac{1}{i \omega_n - (\xi_{{\bf q} - {\bf k}} -
\omega)},\end{aligned}$$ where the frequency sum is over even Matsubara frequencies $i \omega_n
= i 2 n \pi T$ with $n$ integers. The above expression contains three parts, namely the exchange part, the residue part and the line part from top to bottom in Eq. (\[eq:Esum\]). The proof of the equivalence between Eq. (\[eq:Esum\]) and Eq. (\[eq:Eint\]) is provided below.
Since the exchange part exists in both Eq. (\[eq:Esum\]) and Eq. (\[eq:Eint\]), we only need to consider the correlation part of the self-energy $$\begin{aligned}
\label{eq:Esum1}
\Sigma^{\mbox{cor}} ({\bf k}, \omega) &=&
-\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi}
\frac{ 2 v_0(q) \mbox{Im} \epsilon^{-1} (q, \nu )} { \nu + \omega + i
\eta - \xi_{{\bf q} - {\bf k}} } \nonumber \\
&&~~~~\cdot \left[ n_F ( \xi_{{\bf q} - {\bf k}} ) + n_B ( \nu )
\right].\end{aligned}$$ We choose the contour as in Fig. \[fig2\]. It is easy to see that the integration over real axis can be transformed into integration over contour $\cal C$, so that we have $$\begin{aligned}
\label{eq:Esum2}
\Sigma^{\mbox{cor}} ({\bf k}, \omega) &=&
-\int \frac{d^d q}{(2 \pi)^d} \oint_{\cal C} \frac{d \nu}{2 \pi i}
\frac{ v_0(q) (\epsilon^{-1} (q, \nu )-1)} { \nu + \omega + i
\eta - \xi_{{\bf q} - {\bf k}} } \nonumber \\
&&~~~~\cdot \left[ n_F ( \xi_{{\bf q} - {\bf k}} ) + n_B ( \nu )
\right].\end{aligned}$$ This is because $\epsilon({\bf q}, \omega -i \eta)= \epsilon^*({\bf
q}, \omega+i \eta)$, and therefore the integration of the real part of the integrand right above the real axis in the positive direction and right below the real axis in the negative direction cancel each other, and the corresponding integration of the imaginary part on these two lines are equal to each other. The $-1$ after $\epsilon^{-1}({\bf q}, \omega)$ is inserted to make the integration on the arc part of contour $\cal C$ vanish as the radius of the contour approaches infinity. Now we are left to evaluate the residues within contour $\cal C$, the positions of which are denoted by crosses in Fig. \[fig2\]. Note that the analytic property of the dielectric function $\epsilon^{-1}({\bf q}, \omega)$ is very important in this approach. The transformation requires that $\epsilon^{-1}({\bf q},
\omega)$ is analytic in the upper and the lower half of the complex plane, which is true for electron gas systems. The single residue at $\xi_{{\bf q} - {\bf k}} - \omega - i \eta$ right below the real axis produced by the denominator of the integrand produces the residue part of the self-energy. This part can be easily derived as $$\begin{aligned}
\label{eq:Eres}
\Sigma^{\mbox{res}} &=& -
\int \frac{d^d q}{(2 \pi)^d} v_0(q) \left[
\frac{1}{\epsilon(q, \xi_{{\bf q} - {\bf k}} - \omega)} - 1
\right] \nonumber \\
&&~~~~~\cdot \left[ n_B(\xi_{{\bf q} - {\bf k}} - \omega)
+n_F(\xi_{{\bf q} - {\bf k}}) \right].\end{aligned}$$ The residues at $\omega_n = 2 n \pi T$ on the imaginary axis (the third term in Eq. (\[eq:Esum\])), which are produced by the Bose function $n_B(\nu)$, lead to the line part of the self-energy. This part can be written as $$\begin{aligned}
\label{eq:Eline}
\Sigma^{\mbox{line}} &=& -
\int \frac{d^d q}{(2 \pi)^d} T \sum_{\omega_n}
v_0(q) \left[ \frac{1}{\epsilon(q, i \omega_n)} - 1 \right]
\nonumber \\
&& ~~~~ \cdot \frac{1}{i \omega_n - (\xi_{{\bf q} - {\bf k}} -
\omega)}.\end{aligned}$$ From Eq. (\[eq:Eres\]) and Eq. (\[eq:Eline\]) we have $\Sigma^{\mbox{cor}} = \Sigma^{\mbox{res}} + \Sigma^{\mbox{line}}$, and we thus obtain Eq. (\[eq:Esum\]).
![Contour of integration for the derivation of self-energy formula for the frequency sum method. The thick lines on real axis denotes the branch cut for $\epsilon^{-1}({\bf q}, \omega)$. The crosses mark the poles due to the integrand; the ones on the imaginary axis are due to $n_B(\omega)$, and the isolated pole is due to the denominator.[]{data-label="fig2"}](fig2.eps){width="2in"}
The frequency sum method proves to be a far more efficient numerical technique for calculating the self-energy than the frequency integration method due to the absence of the strong non-monotonicity and singularity in the real frequency dependence of the integrand. One thing to notice is that at high temperatures, higher Matsubara frequency terms can be neglected because $(\epsilon^{-1}({\bf q}, i
\omega_n) - 1) \to 0$ when $\omega_n \to \infty$, while at low temperatures a large number of Matsubara terms have to be kept in the sum in order to ensure accuracy. At zero temperature, the frequency sum turns into an integration over imaginary frequencies, and we have $$\begin{aligned}
\label{eq:Esum0}
\Sigma ({\bf k}, \omega) &=& - \int_{{\cal R}_1}
\frac{d^d q}{(2 \pi)^d} v_0(q) \nonumber \\
&&+ \int_{{\cal R}_2} \frac{d^d q}{(2 \pi)^d} \frac{v_0(q)}
{\epsilon({\bf q}, \xi_{{\bf q} - {\bf k}} - \omega)} \nonumber \\
&&-\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi}
\left[ \frac{1}{\epsilon({\bf q}, i \nu)} - 1 \right] \nonumber \\
&&~~~~~~\cdot \frac{\omega - \xi_{{\bf q} - {\bf k}}}
{\nu^2 + (\omega -\xi_{{\bf q} - {\bf k}})^2 },\end{aligned}$$ where the integration region ${\cal R}_1$ denotes the region where $|{\bf k} - {\bf q}| < k $, and ${\cal R}_2$ denotes the integration region where$|{\bf k} - {\bf q}|$ is in between $k$ and $k_F$. This explicit formula for self-energy is exactly what previous works (see, e.g. Ref. [@ting]) used to calculate the zero temperature RPA self-energy.
It is obvious that the frequency independent exchange part of the self-energy is real. By noticing that $\epsilon({\bf q}, -\omega_n) =
\epsilon^*({\bf q}, \omega_n)$, we can see that the line part of the self-energy is real as well. Thus the only contribution to the imaginary part of the self-energy comes from the residue part, which gives the same result as Eq. (\[eq:Eimag\]) in the frequency integration method.
RPA Results for $m^*(r_s, T/T_F)$ {#sec:results}
=================================
In this section we present our numerical results for effective mass in 2D and 3D electron systems within RPA. We first present in section \[sec:result:zeroT\] results for the zero temperature effective mass to compare with earlier works. Our finite temperature results for 2DES are presented in section \[sec:result:2D\] and those for 3DES in section \[sec:result:3D\]. In section \[sec:result:beyond\] we present results for a model bare potential where the Coulomb interaction is cut-off by a finite length so that the bare interaction is short-ranged. We do this in order to investigate the model dependence of our results.
Zero temperature effective mass {#sec:result:zeroT}
-------------------------------
We first present our extreme low temperature results ($T/T_F \approx
10^{-4}$) in Fig. \[fig3\] and Fig. \[fig4\], to be compared with the existing $T=0$ results [@rice; @gellmann; @ting; @vinter; @jalabert; @marmorkos; @book]. We calculate $m^*(r_s)$ in the $r_s = 0 - 10$ range, showing that the effective mass renormalization could be as large as 4.5 for dilute ($r_s \sim
10$) 2DES and 3 for ($r_s \sim 10$) 3DES. We emphasize that the results presented in Fig. \[fig3\] and Fig. \[fig4\] are entirely based on the $T \to 0$ limit of our finite temperature theory. They are in [*quantitative*]{} agreement with the existing $T = 0$ 2D RPA effective mass calculations [@ting] (which are restricted to the $r_s < 5$ regime) and are consistent with the existing $T = 0$ 3D effective mass calculations at low $r_s$ [@rice]. This serves as a stringent check on our numerical approaches.
![Calculated $T \sim 0$ effective mass as a function of $r_s$ in a 2DES. Inset: the result in low $r_s$ region.[]{data-label="fig3"}](fig3.eps){width="2.5in"}
![Calculated $T \sim 0$ effective mass as a function of $r_s$ in a 3DES. Inset: the result in low $r_s$ region.[]{data-label="fig4"}](fig4.eps){width="2.5in"}
It is clear from Figs. \[fig3\] and \[fig4\] that both our 2D and 3D results show the non-monotonic dependence of $m^*(r_s)$ on $r_s$ in the high-density regime (i.e. in the $r_s \ll 1$ regime). This nonmonitonic low-$r_s$ behavior for $m^*(r_s)$ at $T=0$ has been reported in the earlier literature [@rice; @ting].
We emphasize that the numerical results given in Figs. \[fig3\] and \[fig4\] are obtained by putting $T/T_F \approx 10^{-4}$ in our finite-temperature formalism.
Finite temperature effective mass in 2DES {#sec:result:2D}
-----------------------------------------
In Fig. \[fig5\] and Fig. \[fig6\] we show our calculated 2D $m^*(T)$ as a function of $T/T_F$ for different values of the 2D interaction parameter $r_s$ ($=0.1 - 10$). In the low temperature region the effective mass first rises to some maximum, and then decreases as temperature increases. This nonmonotonic trend is systematic, and the value of $T/T_F$ where the effective mass reaches the maximum increases with increasing $r_s$. The initial increase of $m^*(T)$ is almost linear in $T/T_F$ as $T \to 0$, and the slope $\frac{d(m^*/m)}{d(T/T_F)}$ is almost independent of $r_s$ for very small $r_s$ ($<1$) (which is shown in Fig. \[fig6\]), but increases with $r_s$ for larger $r_s$ values. It is important to notice that this non-monotonic temperature dependence of $m^*(T)$ with a maximum around $T/T_F \lesssim 1$ persists all the way to $r_s \to 0$, which suggests that it is not an artifact of our approximation scheme since RPA become [*exact*]{} as $r_s \to 0$.
![Calculated 2D effective mass as a function of $T/T_F$ for different $r_s$: $r_s = 10 \to 1$ from top to bottom; Inset: $r_s
= 5 - 1 $ from top to bottom. Note that $T_F \propto r_s^{-2}$, making the absolute temperature scale lower for higher $r_s$ values.[]{data-label="fig5"}](fig5.eps){width="2.5in"}
![Calculated 2D effective mass as a function of $T/T_F$ for low $r_s$ values: $r_s = 1.0 \to 0.1$ from top to bottom.[]{data-label="fig6"}](fig6.eps){width="2.5in"}
In Fig. \[fig7\] and Fig. \[fig8\] we show the dependence of the effective mass renormalization as a function of the interaction parameter $r_s$ for a few values of [*fixed*]{} temperature (rather than fixed $T/T_F$, remembering that $T_F \propto r_s^{-2}$ since $T_F
\propto n$ and $r_s \propto n^{-1/2}$). Fig. \[fig7\] shows the effective mass for high $T$ and large $r_s$ values while Fig. \[fig8\] concentrates on the low $T$ region. The calculated $m^*(r_s)$ for fixed $T$ values are quite striking: For low fixed values of $T$, $m^*/m$ initially increases with $r_s$ even faster than the corresponding $T=0$ result, eventually decreasing with $r_s$ at large enough values (where the corresponding $T/T_F$ values become large enough). This nonmonotonic behavior of $m^*(r_s)$ as a function of $r_s$ for fixed temperatures showing a temperature-dependent maximum (with the value of $r_s$ at which the $m^*$ peak occurs decreasing with increasing $T$ as in Fig. \[fig7\]) is complementary to the nonmonotonicity of $m^*(T)$ in Fig. \[fig5\] as a function of $T/T_F$ (at fixed $r_s$) and arises from the relationship between the dimensionless variables $T/T_F$ ($\propto r_s^{-2}$) and $r_s$ ($\propto T_F^{-1/2}$) due to their dependence on the carrier density (i.e. $T_F \propto n$ and $r_s \propto n^{-1/2}$). At large $r_s$ and high temperature, Fig. \[fig7\] shows that the effective mass increases from below unity with increasing $r_s$. This is the region where the exchange part of the self-energy dominates, and it can be easily shown that the exchange self-energy produces this peculiar effect on the $r_s$ dependence of $m^*(r_s)$ at fixed high $T$ values. Since this region is hardly accessible by experiments, and moreover the quasiparticles may not even be well-defined at such high $T/T_F$ values, we do not further discuss the physics related to this region.
![Calculated 2D effective mass as a function of $r_s$ at fixed value of temperatures. T is in the unit of $T_F$ at $r_s =
1$.[]{data-label="fig7"}](fig7.eps){width="2.5in"}
![Calculated 2D effective mass as a function of $r_s$ at fixed value of temperatures. T is in the unit of $T_F$ at $r_s =
1$. This plot is similar to Fig. \[fig7\] but concentrating on the low temperature region.[]{data-label="fig8"}](fig8.eps){width="2.5in"}
One immediate consequence of our results shown in Figs. \[fig5\] and \[fig7\] is that $m^* (T/T_F, r_s) \equiv m^*(T, n)$ in 2DES could show a strong enhancement at low (but finite) temperatures and low electron densities (large $r_s$). Comparing with the actual system parameters for 2D electrons in Si inversion layers and GaAs heterostructures (and taking into account the quasi-2D form factor effects [@ando] neglected in our strictly 2D calculation) we find that, consistent with recent experimental findings [@exp], our theoretical calculations predict (according to Figs. \[fig5\] and \[fig7\] as modified by subband form factors) $m^*/m$ to be enhanced by a factor of $2-4$ for the experimental densities and temperatures used in recent measurements [@exp]. Due to the approximate (i.e. RPA) nature of our theory we do not further pursue the comparison with experimental data in this paper since the main goal of this paper is to discuss the temperature dependence of $m^*(r_s, T/T_F)$ which has not yet been reported in the literature. A direct experimental observation of an increasing $m^*(T)$ at low temperatures in 2DES will be a striking confirmation of our theory.
Finite temperature effective mass in 3DES {#sec:result:3D}
-----------------------------------------
In Fig. \[fig9\] and Fig. \[fig10\] we show our calculated 3D $m^*(T)$ as a function of $T/T_F$ for different $r_s$ values. In Fig. \[fig9\] $r_s$ varies from $1$ to $10$ while in Fig. \[fig10\] $r_s$ is from $0.1$ to $1$. The 3D temperature dependence of the effective mass shows very different characteristics from that of 2D. Fig. \[fig9\] shows that for $r_s > 1$ the effective mass decreases monotonically with increasing $T$ at low temperatures. However for $r_s << 1$, as shown in Fig. \[fig10\], the effective mass increases with increasing $T$ in the temperature region we are considering. We therefore conclude that in 3DES the sign of the slope $\frac{d(m^*/m)}{d(T/T_F)|_{T=0}}$ is non-universal, which differs from that of 2DES where the above mentioned slope is always positive for all $r_s$. Another interesting feature is that the sign of $\frac{d(m^*/m)}{d(T/T_F)}|_{T=0}$ matches the sign of $-\frac{d(m^*/m)}{d(r_s)}|_{T=0}$ very well. In particular, $m^*(T)$ decreases with increasing $T$ (at low $T$) in the “larger” $r_s$-regime where the corresponding $T=0$ $m^*(r_s)$ shows an increasing mass with increasing $r_s$. Similarly, $m^*(T)$ increases (at low $T$) with increasing $T$ in the $r_s$-regime where the corresponding $m^*(r_s; T=0)$ shows decreasing $m^*$ with increasing $r_s$.
![Calculated 3D effective mass as a function of $T/T_F$ for different $r_s$: $r_s = 10 \to 1$ from top to bottom.[]{data-label="fig9"}](fig9.eps){width="2.5in"}
![Calculated 3D effective mass as a function of $T/T_F$ for low $r_s$ values: $r_s = 1.0 \to 0.1$.[]{data-label="fig10"}](fig10.eps){width="2.5in"}
Model short-range bare interaction {#sec:result:beyond}
----------------------------------
So far in all of our calculations we have been using the realistic long-ranged Coulomb interaction for the bare potential as in Eq. (\[eq:bareV\]). A question naturally arises: how is the temperature dependence that we find in our calculations related to the long range nature of the interaction between electrons? Therefore we also calculate the effective mass in 2DES and 3DES using simple (parameterized) finite-range interaction model $$\begin{aligned}
\label{eq:newV}
v^{\mbox{2D}}(q) &=& {2 \pi e^2 \over q + a}, \nonumber \\
v^{\mbox{3D}}(q) &=& {4 \pi e^2 \over q^2 + a^2}.\end{aligned}$$ where $a$ is the cut-off wavevector which eliminates the long wavelength Coulomb divergence.
Our numerical calculation shows that as $a/k_F \to 0$, we recover the $m^*(T)$ behavior of the bare Coulomb interaction results in both 2D and 3D. As $a/k_F$ increases, the mass renormalization in both 2D and 3D is suppressed, but all the qualitative features of the temperature dependence persist. In 2DES, as $a/k_F$ increases, the temperature where the effective mass reaches the maximum decreases, and the effective mass enhancement (from the $T=0$ value to the maximum) decreases, but the linear-$T$ dependence at low $T$ and the non-monotonic trend remain unchanged. In 3DES, as $a/k_F$ increases, the $r_s$ region where $\frac{d(m^*/m)}{d(r_s)}|_{T=0} < 0$ shrinks, but the consistency between the sign of $-\frac{d(m^*/m)}{d(r_s)}|_{T=0}$ and the sign of $\frac{d(m^*/m)}{d(T/T_F)}|_{T=0}$ remains.
From these result we conclude that the qualitative features of the temperature dependence are model independent and [*not*]{} peculiar to the bare interaction being Coulombic. This conclusion is further reinforced by the recent report of a linearly $T$ dependent electronic specific heat in a short-range interaction model [@chubukov]. It may be worth while, however, to note that RPA is specific to the long-range Coulomb interaction in giving an exact result in the high-density $r_s \to 0$ limit, and there is nothing special about RPA in the case of short-range interaction.
Plasmon-Pole approximation {#sec:PPA}
==========================
We now apply a simple-to-use dynamical approximation to calculate $m^*(T)$. The plasmon-pole approximation (PPA) has often been used [@vinter; @dassarma; @lundqvist] to obtain the electron self-energy in the literature. It is a simple technique for carrying out the frequency sum or integration in the RPA self-energy calculation by using a spectral pole (i.e. a delta function) ansatz for the dynamical dielectric function $\epsilon({\bf k}, \omega)$. In other words, it is an approximation to the RPA. The PPA ansatz assumes that $$\label{eq:ansatz}
- 2 \mbox{Im}\frac{1}{\epsilon({\bf k}, {\omega})} =
C_k (\delta(\omega - \bar{\omega}_k) - \delta(\omega + \bar{\omega}_k)),$$ where the pole $\bar{\omega}_k$ and the spectral weight $C_k$ of the PPA propagator in Eq. (\[eq:ansatz\]) are determined by using the the Kramers-Krönig relation (i.e. causality) $$\label{eq:KK}
\mbox{Re} \frac{1}{\epsilon(k, 0)} = 1 + \frac{2}{\pi} \int_0^{\infty}
\frac{1}{\omega} d \omega \mbox{Im} \frac{1}{\epsilon(k, \omega)}$$ and the $f$-sum rule (i.e. current conservation) $$\label{eq:sumrule}
\int_0^{\infty} \omega d \omega \mbox{Im}
\frac{1}{\epsilon(k, \omega)}
= - \frac{\pi}{2} \omega_P^2(k).$$ Putting Eq. (\[eq:ansatz\]) in Eq. (\[eq:KK\]) and Eq. (\[eq:sumrule\]) we have $$\begin{aligned}
\label{eq:Ck}
C_k &=& \pi \omega_P(k) \sqrt{1 - \mbox{Re} \epsilon^{-1} (k, 0)}, \\
\label{eq:omegabar}
\bar{\omega}_k &=& \frac{\omega_P(k)}{\sqrt{1 - \mbox{Re}
\epsilon^{-1} (k, 0) }},\end{aligned}$$ where $\omega_P(k)$ in Eqs. (\[eq:sumrule\]), (\[eq:Ck\]) and (\[eq:omegabar\]) is the long wavelength plasmon frequency which is defined as $$\label{eq:omegaP}
\lim_{\omega \to \infty} \mbox{Re} [\epsilon(k, \omega)]
= 1 - \frac{\omega_P^2(k)}{\omega^2}.$$ It is well-known that in 2DES $$\label{eq:omegaP2D}
\omega_P^2(k) = \frac{2 \pi n e^2}{m} k,$$ and in 3DES $$\label{eq:omegaP3D}
\omega_P^2(k) = \frac{4 \pi n e^2}{m}.$$
We mention that $\bar{\omega}_k$ in Eq. (\[eq:ansatz\]) does [ *not*]{} correspond to the real plasmon dispersion in the electron liquid, but simulates the whole excitation spectra of the system behaving as an effective plasmon at low momentum and as the single-particle electron-hole excitation at large momentum, as constrained by the Kramers-Krönig relation and the $f$-sum rule. Details on the PPA are available in literature [@vinter; @dassarma; @lundqvist], including its finite-temperature generalization [@dassarma]. The PPA, which is known to give results close to the full RPA calculation of self-energy, allows a trivial carrying out of the frequency sum in the retarded self-energy function leading to: $$\begin{aligned}
\label{eq:EPPA}
\mbox{Re}\Sigma({\bf k}, \omega) =&-& \int \frac{d^2 q}{(2 \pi)^2}
v_0(q) n_F(\xi_{{\bf q} - {\bf k}}) \nonumber \\
&+& \int \frac{d^2 q}{(2 \pi)^2} v_0(q) C_q \Big[
\frac{ n_B( \bar{\omega}_q ) + n_F ( \xi_{{\bf q} - {\bf k}} ) }
{\bar{\omega}_q - ( \xi_{{\bf q} - {\bf k}} - \omega ) } \nonumber \\
&&~~~~~~~+
\frac{ n_B( -\bar{\omega}_q ) + n_F ( \xi_{{\bf q} - {\bf k}} ) }
{\bar{\omega}_q + ( \xi_{{\bf q} - {\bf k}} - \omega) } \Big],\end{aligned}$$ where $C_q$ and $\bar{\omega}_k$ only depend on $\epsilon(k, 0)$ at finite temperatures, and are determined by Eq. (\[eq:Ck\]) and Eq. (\[eq:omegabar\]). Obviously the PPA provides a great simplification of the problem since the most numerically demanding part of the calculation (the frequency sum or integration) is trivially done. It should be noted, however, that although the PPA is known to produce a reliable approximation to ${\rm Re} \Sigma$, it, by definition, fails completely for ${\rm Im} \Sigma$.
We present our PPA results for the 2D effective mass as a function of $T/T_F$ at fixed $r_s$ values in Fig. \[fig11\]. One immediate observation by comparing Fig. \[fig5\] and Fig. \[fig11\] is that even though PPA provides a very good approximation for the self-energy (indeed, our numerical results for PPA self-energy and RPA self-energy match very well), it fails to provide accurate result for the effective mass. The zero temperature effective mass generated by PPA is almost half of that from RPA, and the temperatures where $m^*$ maximizes shift to higher $T$ values in the PPA compared with RPA. But the qualitative behavior of $m^*(r_s, T/T_F)$ is similar in the PPA and RPA for the 2DES as is clear by comparing Figs. \[fig11\] and \[fig6\].
![Calculated 2D PPA effective mass as a function of $T/T_F$ at fixed value of $r_s$.[]{data-label="fig11"}](fig11.eps){width="2.5in"}
From our results of 3D PPA effective mass calculation presented in Fig. \[fig12\] we can see that they are different from RPA results even qualitatively. In fact, our RPA results for $m^*(r_s, T/T_F)$ are similar in both 2D and 3D.
![Calculated 3D PPA effective mass as a function of $T/T_F$ at fixed value of $r_s$.[]{data-label="fig12"}](fig12.eps){width="2.5in"}
Quasiparticle decay {#sec:decay}
===================
The quasiparticle decay rate (or the inverse lifetime) is given [@chaplik; @gq] by the imaginary part of the self-energy. As we have discussed in section \[sec:method:sum\], the imaginary part of the quasiparticle self-energy can be calculated from Eq. (\[eq:Esum\]). It is also obvious that only the second term in Eq. (\[eq:Esum\]) contributes to the imaginary self-energy: the first term is obviously real, and the last term is also real because $\epsilon({\bf q}, -\omega_n) = \epsilon^*({\bf q}, \omega_n)$. Thus we have $$\begin{aligned}
\label{eq:ImE}
\mbox{Im} \Sigma({\bf k}, \omega) =
&-& \int \frac{d^d q}{(2 \pi)^d} v_0(q) \mbox{Im}
\frac{1}{\epsilon(q, \xi_{{\bf q} - {\bf k}})}
\nonumber \\
&&\cdot \left[n_B(\xi_{{\bf q} - {\bf k}} - \omega)
+n_F(\xi_{{\bf q} - {\bf k}}) \right]. \end{aligned}$$
Fig. \[fig13\] and Fig. \[fig14\] show the calculated imaginary self-energy on the Fermi surface in 2D and 3D respectively. The quasiparticle decay (i.e. finite ${\rm Im} \Sigma$) here arises entirely from having a finite temperature. The results show that the magnitude of the imaginary self energy increases with increasing $r_s$ and $T/T_F$. It is obvious from Eq. (\[eq:ImE\]) that the imaginary self-energy vanishes on the Fermi surface at $T = 0$ as it must since the quasiparticles are perfectly well-defined at $T=0$ for $k=k_F$. As $T$ increases, the magnitude of imaginary self-energy remains small compared to the Fermi energy up to a certain temperature, and the quasiparticles on the Fermi surface remain well-defined up to that temperature. The important question is whether the finite temperature quasiparticles are sufficiently well-defined for the interesting behavior of $m^*(T)$ we discussed in section \[sec:results\] to be experimentally observable. If the quasiparticles are ill-defined (i.e. ${\rm Im} \Sigma(k_F) > E_F$ in the temperature regime of interest) then obviously all the interesting temperature dependence of $m^*(T)$ predicted by us is only of academic interest since the large broadening will make it impossible to define quasiparticles, let alone their effective mass. Examining the results of Figs. \[fig13\] and \[fig14\] compared with those presented in section \[sec:results\] it is clear that there is a well-defined regime of $(r_s, T/T_F)$ values where $m^*(T/T_F)$ shows non-trivial temperature dependence with the condition $E_F \gg \left| {\rm Im} \Sigma(k_F) \right|$ well-satisfied so that quasiparticles are well-defined. Although this is not unexpected since $\left|{\rm Im} \Sigma(T) \right| \sim T^2$ for $T/T_F \ll 1$ whereas $m^*(T)/m \approx 1 + {\cal O}(T)$ in 2D, it is nevertheless important to see that ${\rm Im} \Sigma$ remains small in magnitude in the $(r_s, T/T_F)$ regime of interest.
![Calculated magnitude of the 2D RPA imaginary self-energy of quasiparticles on Fermi surface as a function of $r_s$ at different values of $T/T_F$.[]{data-label="fig13"}](fig13.eps){width="2.5in"}
![Calculated 3D RPA imaginary self-energy of quasiparticles on Fermi surface as a function of $r_s$ at different values of $T/T_F$.[]{data-label="fig14"}](fig14.eps){width="2.5in"}
Earlier theoretical work on the quasiparticle damping of 2D interacting electron systems can be found in Refs. [@jalabert; @marmorkos; @chaplik; @gq].
Discussion and conclusion {#sec:con}
=========================
In this work we have obtained detailed results for the temperature dependence of the quasiparticle effective mass, $m^*(r_s, T/T_F)$, at arbitrary values of temperature and density in 2D and 3D electron systems interacting via the long range Coulomb interaction. Our central approximation is the RPA (i.e. the dynamically screened Hatree-Fock self-energy approximation), which is the leading-order one-loop self-energy calculation in a dynamically screened effective interaction expansion. RPA is exact in the high-density ($r_s \to 0$) limit at $T=0$, and is therefore a controlled nontrivial approximation which is empirically known to work well for $r_s > 1$ (e.g. metals with $r_s \sim 3 - 6$ and 2D semiconductors with $r_s \sim 1 - 10$). We also calculate the finite-temperature imaginary self-energy (i.e. the quasiparticle decay rate or broadening) to ensure that the broadening remains small in the $(r_s, T/T_F)$ parameter regime of our interest where $m^*(T)$ shows interesting temperature dependence.
As mentioned earlier in the paper, it is well-known that at $T=0$, $m^*(r_s)$ can be exactly calculated (in both 2D and 3D) in the asymptotic $r_s \to 0$ limit by systematically expanding the RPA self-energy since ring diagrams (included in the RPA) are the most divergent diagrams in the $r_s \to 0$ limit. Such a zero temperature $r_s$-expansion of RPA gives the following formula for $m^*(r_s)$ in both 2D and 3D: $$\label{eq:m0T}
\left. {m^*(r_s) \over m} \right|_{r_s \to 0} = 1
+ a r_s (b + \ln r_s) + {\cal O} (r_s^2),$$ where $a$, $b$ are constants of order unity, What we find numerically is that the leading temperature correction to this effective mass formula is linear in $T/T_F$ in 2D and nonlinear in $T/T_F$ in 3D. In this paper we have calculated $m^*(r_s, T/T_F)$ numerically for the one-loop dynamically screened Hatree-Fock RPA self-energy theory for arbitrary $r_s$ and $T/T_F$ finding nontrivial temperature dependence of the effective mass at all densities.
Our most important result is the unexpected discovery of a strong temperature-dependent quasiparticle effective mass $m^*(T)$ at low temperatures in 2DES. Since the temperature scale for the temperature dependence of $m^*(T)$ is the Fermi temperature which tends to be high ($\sim 10^4 K)$ in the 3D electron liquids (i.e. metals), our temperature-dependent effective mass results for 3D systems are mostly of theoretical interest since any actual $T$-dependence of $m^*(T)$ in the $T/T_F \lesssim 10^{-4}$ regime will be miniscule. Our numerical results for the calculated $m^*(T)$ in 2D systems are consistent with a [*linear*]{} leading-order temperature correction for the 2D quasiparticle effective mass: Results in Figs. \[fig5\] and \[fig6\] can be well fitted to the formula $m^*(T) \approx 1 +
A^{\rm 2D} (r_s) + B^{\rm 2D} (r_s) (T/T_F) + \cdots$ for small $T/T_F$ where the slope $B^{\rm 2D} (r_s)$ seems to be a constant independent of $r_s$ (i.e. density) at least in the high-density ($r_s
\ll 1$) limit; for $r_s > 1$ the slope $B^{\rm 2D} (r_s)$ has a weak density dependence increasing somewhat with increasing $r_s$ (but our approximation scheme, RPA, becomes less quantitatively reliable at large $r_s$, therefore it is possible that the slope $d(m^*/m)/d(T/T_F)$ is indeed independent of $r_s$ in the $T \to 0$ limit). In addition to this interesting (and unexpected) linear leading-order temperature correction to the quasiparticle effective mass, we also find $B^{\rm 2D} (>0)$ to be positive for all $r_s$, indicting that in 2DES, the leading-order temperature correction to the effective mass is positive. Thus, $m^*(T)$ increases with increasing $T$ at first before eventually decreasing as $T/T_F$ increases substantially, leading to a maximum in $m^*(T)$ at some intermediate temperature $T^*(r_s) \sim 0.5 T_F$ which is only weakly density dependent (except of course through $T_F$ itself). All three of these 2D findings (i.e. linear leading-order $T/T_F$ dependence of $m^*$, increasing $m^*$ with $T/T_F$ at low temperatures, and the nonmonotonic behavior with a maximum in $m^*(T/T_F)$ occurring at $T^*
\sim 0.5 T_F$) are surprising and unexpected. In principle, these predictions can be experimentally tested since our calculations presented in section \[sec:decay\] show that the quasiparticles remain reasonably well-defined (i.e. the broadening, ${\rm Im}
\Sigma$, remains small) all the way to $T^*$ and perhaps even above $T^*$. This is reasonable since the many-body correction to $m^*$ is linear in $T/T_F$ whereas the broadening ${\rm Im} \Sigma \sim
(T/T_F)^2$, ensuring that for $T/T_F <1$, the quasiparticle effective mass is a well-defined quantity. In contrast to the linear (with positive slope) leading-order $T$-dependence we find for all $r_s$ in our calculated 2D $m^*(T)$, our 3D results show non-universal $m^*(r_s, T/T_F)$ behavior. In 3D, $m^*(T/T_F)$ increases with increasing $T/T_F$ at low temperatures only for very high densities (small $r_s$) – for larger $r_s$ values $m^*(T)$ decreases monotonically with increasing temperature (in sharp contrast to the striking non-monotonicity in $m^*(T)$ in 2D) and this decrease is more consistent with a nonlinear leading-order temperature dependence (rather than a linear one as in 2D). Our best guess for our numerical results shown in Fig. \[fig9\] and \[fig10\] is the following equation: $m^*/m \approx 1 + A^{\rm 3D} (r_s) + C^{\rm 3D} (T/T_F)^l
\ln (T_F/T) + \cdots$, where $l$ is a number of the order one (note that numerically fixing the number $l$ is difficult and need much more work), $C^{\rm 3D} > 0$ for $r_s < r_s^*$ and $C^{\rm 3D} < 0$ for $r_s > r_s^*$ with $r_s^*$ being approximately the $r_s$ value where $A^{\rm 3D}$ changes from being negative to positive.
We comment that our numerical results for $m^*(T)$ are consistent (actually agree very well) with the very recent analytical work [@short; @vd] on the temperature corrections to the effective mass renormalization in 2D and 3D Fermi liquid. The analytical work is necessarily restricted to the $r_s \to 0$ and $T/T_F \to 0$ limit where the infinite series of ring diagrams to the electron self-energy (depicted in Fig. \[fig1\]) provides an [*exact*]{} leading-order asymptotic answer to the problem with the following result $$\begin{aligned}
\label{eq:ana}
{m^*(r_s, T/T_F) \over m} = 1 &+& A(r_s)
+ B(r_s) \left({T \over T_F} \right) \nonumber \\
&+& C(r_s) \left({T \over T_F} \right)^2 \ln \left({T \over T_F} \right)
+ \cdots,\end{aligned}$$ with $B(r_s) \equiv B^{\rm 2D}$, a constant, in 2D, and $B(r_s) \equiv
0$ in 3D. Our numerical results are consistent with this exact result, but our numerical results apply also in the non-asymptotic region where $T/T_F$ and $r_s$ are not necessarily small. In this non-asymptotic regime (where $r_s$ is [*not*]{} small, actually $r_s$ may be large in 2D semiconductor systems) RPA is by no means an exact theory, but we have recently argued [@diverge] that RPA remains extremely well-valid (if somewhat uncontrolled) even for $r_s \gg 1$. We also emphasize a point in this context that seems not to have been widely appreciated in the literature. The point is that RPA becomes a progressively better approximation as $T/T_F$ increases at a fixed $r_s$ (for any $r_s$), because the system is becoming more classical in the $T/T_F \gg 1$ regime where RPA is again exact. Thus in the $(r_s, T/T_F)$ parameter space (see Fig. \[fig15\]) RPA is exact as $r_s \to 0$ (the high density limit) and as $T_F \to 0$ (the high-temperature or equivalently the low density limit) or as $T \to
\infty$. Thus the regime of validity of RPA is greatly enhanced at finite temperature, and in fact even at very large $r_s$ (i.e. very low density) RPA becomes exact as $T$ is raised (because $T/T_F \gg 1$ limit is more easily achieved at low densities).
![Schematic validity of RPA. The shaded area denotes the region where RPA is considered to be valid. Line A denotes a certain density value above which RPA is valid at $T=0$ (e.g. the vertical line A may correspond to the $r_s = 1$ condition so that for higher density, i.e. to the right to line A, RPA is valid even at $T=0$). Line B denotes the line of $T_F \propto n$. In the region above line B, RPA is again valid. Therefore for any fixed value of density $n$ (or equivalently fixed $r_s$), RPA is a better approximation with increasing $T$, whereas for fixed value of temperature, low density values (or large $r_s$) counter-intuitively makes RPA to valid again since RPA is valid for $T>T_F$.[]{data-label="fig15"}](fig15.eps){width="3in"}
Finally, we comment on the anomalous (often referred to as “nonanalytic” [@chubukov]) nature of the temperature corrections to the quasiparticle effective mass in 2D systems (but [*not*]{} in 3D) as manifested in the [*linear*]{} leading-order temperature correction we find in interacting 2D electron systems. This particular feature is apparently generic in 2D and [*not*]{} due merely to our using the long-range bare Coulomb interaction, because in Ref. [@chubukov] the same linear-$T$ correction is found in calculations using a zero-range bare interaction although the sign of the slope is negative in the zero-range interaction case. This kind of (leading-order) linear temperature correction is quite common in 2D electron systems due to the peculiar form of the 2D polarizability with a $T=0$ cusp at $2k_F$. This leading-order linear-$T$ correction is interesting because the naive expectation in a Fermi system (based on the usual Sommerfeld expansion of the Fermi functions) is that the leading-order correction in a “normal” situation should always be ${\cal O} (T/T_F)^2$ for all electronic properties. In 2D electron systems it seems that the generic situation is “anomalous”, i.e. the leading-order temperature correction is ${\cal O} (T/T_F)$ rather than the “normal” quadratic Fermi behavior expected on the basis of analytic Sommerfeld expansion of Fermi functions. In 2D interacting electron systems, therefore, all leading-order thermal corrections to electronic properties are much stronger (by a factor of $T_F/T$, which is a large number as $T \to 0$) than the quadratic Fermi gas behavior. This anomalous nonanalyticity, which may have important consequences for fermionic quantum critical phenomena, obviously has important experimental implications since it is much easier to observe a linear temperature correction than a quadratic one at low temperatures.
This work is supported by the US-ONR, the NSF, and the LPS.
Appendix {#appendix .unnumbered}
========
Here we provide a proof of for equivalence between Eq. (\[eq:Esum\]) and Eq. (\[eq:Eint\]).
$$\begin{aligned}
\label{eq:Eint1}
&&\!\!\!\!\!\!\!\!\!\!
\Sigma ({\bf k}, \omega)=
-\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi}
\Big\{ \nonumber \\
&&~\mbox{Im} G_R({\bf k} - {\bf q}, \nu + \omega)
D_R({\bf q}, \nu)
\tanh (\frac{\nu+\omega}{2 T} )
\nonumber \\
&&+G_R({\bf k} - {\bf q}, \nu + \omega)
\mbox{Im}D_R({\bf q}, \nu)
\coth (\frac{\nu}{2 T}) \Big\} \nonumber\end{aligned}$$
$$\begin{aligned}
&=&-\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi} v_0(q)
\Big\{ \nonumber \\
&&- \pi \delta(\nu + \omega - \xi_{{\bf q} - {\bf k}})
\epsilon^{-1} ({\bf q}, \nu)
\tanh (\frac{\nu+\omega}{2 T} ) \nonumber \\
&&+ \frac{1}{\nu + \omega + i \eta - \xi_{{\bf q} - {\bf k}}}
\mbox{Im} \epsilon^{-1} ({\bf q}, \nu)
\coth ( \frac{\nu}{2 T} ) \Big\}.\end{aligned}$$
Using Kramers-Krönig relations for $\epsilon^{-1}(q, \nu)$ in the above equation, we have $$\begin{aligned}
\label{eq:Eint2}
&&\!\!\!\!\!\!\!\!\!\!
\Sigma ({\bf k}, \omega) =
-\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi} v_0(q)
\Big\{ \nonumber \\
&&- \pi \delta(\nu + \omega - \xi_{{\bf q} - {\bf k}})
\nonumber \\
&& ~~~~\cdot \left[1 + \int \frac{d \nu'}{\pi}
\frac{ \mbox{Im} \epsilon^{-1} (q,\nu') }{\nu' - \nu - i \eta}
\right] \tanh (\frac{\nu' +\omega}{2 T} ) \nonumber \\
&& + \frac{1}{\nu + \omega + i \eta - \xi_{{\bf q} - {\bf
k}}}
\mbox{Im} \epsilon^{-1} ({\bf q}, \nu)
\coth ( \frac{\nu}{2 T} ) \Big\} \nonumber \end{aligned}$$ $$\begin{aligned}
&=& -\int \frac{d^d q}{(2 \pi)^d} v_0(q) \Big\{ \nonumber \\
&& -\frac{1}{2} \left[1 + \int \frac{d \nu}{\pi} \frac{ \mbox{Im}
\epsilon^{-1} (q, \nu) }{\nu + \omega + i \eta
- \xi_{{\bf q} - {\bf k}}} \right]
\tanh (\frac{\xi_{{\bf q} - {\bf k}}}{2 T} ) \nonumber \\
&& + \int \frac{d \nu}{2 \pi}
\frac{1}{\nu + \omega + i \eta - \xi_{{\bf q} - {\bf k}}}
\mbox{Im} \epsilon^{-1} ({\bf q}, \nu)
\coth ( \frac{\nu}{2 T} ) \Big\} \nonumber \end{aligned}$$ $$\begin{aligned}
&=& \int \frac{d^d q}{(2 \pi)^d} v_0(q) \frac{1}{2}
\tanh ( \frac{\xi_{{\bf q} - {\bf k}}}{2 T} ) \nonumber \\
&&+\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi}
\frac{ v_0(q) \mbox{Im} \epsilon^{-1} (q, \nu )} { \nu + \omega +
i \eta - \xi_{{\bf q} - {\bf k}} } \nonumber \\
&&~~~~~~~\cdot \left[ \tanh ( \frac{\xi_{{\bf q} - {\bf k}}}{2 T} )
- \coth ( \frac{\nu}{2 T} ) \right] \nonumber \end{aligned}$$ $$\begin{aligned}
&=& \mbox{Const} - \int \frac{d^d q}{(2 \pi)^d}
v_0(q) n_F ( \xi_{{\bf q} - {\bf k}} ) \nonumber \\
&&~~~~~~~ -\int \frac{d^d q}{(2 \pi)^d} \int \frac{d \nu}{2 \pi}
\frac{ 2 v_0(q) \mbox{Im} \epsilon^{-1} (q, \nu )} { \nu + \omega
+ i \eta - \xi_{{\bf q} - {\bf k}} } \nonumber \\
&&~~~~~~~~~~~~~~~~~~~~~\cdot \left[ n_F ( \xi_{{\bf q} - {\bf k}} )
+ n_B ( \nu ) \right].\end{aligned}$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In view of the numerous experimental results recently released, we provide in this letter an update on the performance of our simple Regge model for strangeness electroproduction on the nucleon. Without refitting any parameters, a decent description of all measured observables and channels is achieved. We also give predictions for spin transfer observables, recently measured at Jefferson Lab which have high sensitivity to discriminate between different theoretical approaches.'
address:
- '$^{a}$ IPN Orsay, F-91406 Orsay, France'
- '$^{b}$ CEA/Saclay, DAPNIA/SPhN, F-91191 Gif-sur-Yvette Cedex, France'
- '$^{c}$ University Mainz, D-55099 Mainz, Germany'
author:
- 'M. Guidal $^{a}$, J.-M. Laget $^{b}$ and M. Vanderhaeghen$^{c}$'
title: 'Exclusive electromagnetic production of strangeness on the nucleon : review of recent data in a Regge approach.'
---
Regge theory provides a simple and elegant framework to describe exclusive hadronic reactions above the resonance region [@regge1; @regge2; @storrow]. We have presented in references [@ourpapers1; @ourpapers2; @ourpapers3; @ourpapers4]) a Regge model for meson photo- and electroproduction reactions which is based on the exchange of one or two meson Regge trajectories in the $t$-channel. The very few free parameters in this approach are the coupling constants of the first particle materialization of the Regge trajectory at the hadronic vertices along with the mass scales of monopole form factors at the electromagnetic vertices in the case of electroproduction. Basically, all available observables are decently and “economically" described by this approach for a plethora of elementary channels : photo- and electroproduction on the nucleon of $\pi^{0,\pm}$ and $K^+$ as well as $\rho^0, \omega, \phi,\gamma$ [@laget1; @laget2] and $\eta, \eta^\prime$ [@taiwan]. Surprisingly, in some cases, data down to W$<2$ GeV center of mass energies, therefore supposedly in the resonance region, can be succesfully described : this could be interpreted as a manifestation of the reggeon-resonance duality hypothesis [@dolen]. Why this duality seems to work for some channels and not for some others still remains an open question.
Recently, numerous experimental data have been released by the Jefferson Lab, ELSA, GRAAL and SPRING8 facilities in the kaon production sector. This motivates this study which compares the Regge model to these new observables and channels, without any change of the parameters. Our model is fully described in Refs. [@ourpapers1; @ourpapers2; @ourpapers3; @ourpapers4]. In those works, we found that, for strangeness electromagnetic production, it allows to describe the $\gamma^{(*)} + p \rightarrow K^+
+ \Lambda$ and $\gamma^{(*)} + p \rightarrow K^+ + \Sigma$ reactions through the exchange of only two trajectories in the $t$-channel : $K$ and $K^*$. The coupling constants at the ($K,(\Lambda,\Sigma),N$) and ($K^*,(\Lambda,\Sigma),N$) vertices were derived and fitted from the photoproduction study [@ourpapers1; @ourpapers2] where all existing high energy data could be satisfactorily described : $$\begin{aligned}
&&g_{K\Lambda N}=-11.54 \; , \; g_{K^*\Lambda N}=-23.0 \; ,
\; \kappa_{K^*\Lambda N}=2.5, \nonumber\\
&&g_{K\Sigma N}=4.48 \; \; \; , \; g_{K^*\Sigma N}=-25.0 \; ,
\; \kappa_{K^*\Sigma N}=-1.0
\label{ctes}\end{aligned}$$
In the electroproduction case, the other parameters of the model are the two (squared) mass scales of the monopole form factors at the $\gamma,K,(K,K^*)$ vertices, which were taken both equal to 1.5 GeV$^2$ [@ourpapers4] in order to fit the high $Q^2$ behavior of the data.
Finally, let’s recall that one essential feature of the model is the way gauge invariance is restored for the $t$-channel $K$ exchange by proper “reggeization" of the $s$-channel nucleon pole contribution. As detailed further below, this is the key element to describe the slow decrease with $Q^2$ of the $R=\sigma_L /\sigma_T$ ratio for the $K^+ \Lambda$ channel, a feature which was found to be difficult to accomodate in all other approaches.
These data, first published in [@baker], have actually been recently re-analysed [@mohring] by the JLab collaboration E93018. We compare on Fig. \[fig:mohring\] the Regge model predictions [@ourpapers4] with these new $Q^2$ dependence for the experimental transverse ($\sigma_T$) and longitudinal ($\sigma_L$) cross sections, along with their ratio ($R$), for both the $\Lambda$ and $\Sigma$ channels. The corrections due to the new analysis are non-negligible for the absolute values of the longitudinal and transverse cross sections and affect significantly the slopes of the $Q^2$ dependences. Our (unchanged) model gives now a much better description of $\sigma_L$ but it significantly underestimates $\sigma_T$ at large $Q^2$, for both channels. However, the ratio $R$ is still very well reproduced and displays a slow decrease as $Q^2$ increases.
The curves of the lower panels of Fig. \[fig:mohring\] illustrate that the origin of the decrease with $Q^2$ of the ratio $R$ is actually not so much the result of the Regge treatment of the $K$ and $K^*$ $t$-channel propagators but, rather, as mentionned before, of the particular way gauge invariance is restored (i.e. by reggeizing the nucleon $s$-channel and kaon $t$-channel diagrams and assigning to them the [*same*]{} electromagnetic form factor). The reggeization is nevertheless necessary to ensure a correct normalization of ($\sigma_T$) and ($\sigma_L$) as standard Feynman propagators would produce cross sections higher by factors of 2 to 5 at these energies.
=9.5 cm =10. cm
A better agreement with these new data could certainly be achieved by changing the values of the $K$ and $K^*$ form factor mass scales but this would destroy the nice agreement with the other kaon electroproduction data [@bebek77b; @brauel79; @feller; @bebek77; @azemoon; @brown; @bebek74], for which $W>$ 2.1 GeV and which were presented in Ref. [@ourpapers4]. We prefer to interpret this discrepancy as room for potential additionnal processes at these lower energies (W=1.84 GeV), such as $s$-channel resonances. It has indeed already been observed that the well known nucleon resonances have larger transverse photo-couplings than longitudinal ones [@pdg; @burkert].
Fig. \[fig:mohring2\] shows the $Q^2$ dependence of the $\Sigma \over
\Lambda$ ratio for both transverse and longitudinal cross sections. Again, a good agreement with the data is found without any additional refitting of the parameters of the model. In our $t$-channel approach, only the kaon exchange contributes to $\sigma_L$ (like for the pion form factor in pion electroproduction, the best way to access the kaon form factor at intermediate and large $Q^2$ is to isolate $\sigma_L$, as it is mostly insensitive to $K^*$ exchange). This is why the $\Sigma\over\Lambda$ ratio for $\sigma_L$ is basically constant in our model. However, for the transverse part of the cross section, the $K$ and $K^*$ contribute with different weights for the $\Lambda$ and $\Sigma^0$ channels and therefore, the $\Sigma\over\Lambda$ ratio is no longer constant for $\sigma_T$.
=9.5 cm =10. cm
Fig. \[fig:schuh\] displays the photoproduction data with, in particular, the latest results from the JLab/CLAS collaboration [@schuh] which are about 20-25% above the previous Bonn/Saphir data [@bonn]. For the $\Lambda$ channel, there are structures in the experimental data (“bumps" around W$\approx$ 1.75 GeV and 1.95 GeV) which hint to $s$-channel resonances excitations and which our model can obviously not reproduce as it is purely a $t$-channel model ; however, the JLab experimental data (for the forward angles) lies now, in average, very close to the Regge model, unchanged from Ref. [@ourpapers1; @ourpapers2]. This supports, at the 10-15% level and for this channel, the concept of a global duality between the sum of all $t$-channel exchanges and all $s$-channels excitations.
=9.5 cm =10. cm
As to the $\Sigma^0$ channel, the Bonn/Saphir data [@bonn] largely exceeds the calculation of our model. The data point to a prominent $s$-channel resonance structure around W $\approx$ 1.9 GeV. It would be interesting to explore the kinematical region $W>$ 2.1 GeV to see if our model overestimates the experimental data which would therefore make up for the underestimation of our model in the $W<$ 2.1 GeV and which would restore, in average, the duality idea. This high energy domain can be explored by the JLab CLAS collaboration up to $E_\gamma \approx$ 6 GeV (W$\approx$ 3.5 GeV).
The $K^0 \Sigma^+$ channel can be calculated in a straightforward way, without any additional parameter by taking into account only the $K^{0*}$ $t$-channel exchange (as a real photon cannot elastically couple to the neutral spinless $K^0$) and using $g^2_{\gamma K^0 K^{*0}}=
g^2_{\gamma K^+ K^{*+}} \Gamma_{K^{*0}\to\gamma K^0} / {\Gamma_{K^{*+}\to\gamma K^+}}
\approx$ 2.3 $g^2_{\gamma K^+ K^{*+}}$ with the radiative widths taken from Ref. [@pdg]. Figure \[fig:schuh\] shows a very nice agreement between our calculation and the Bonn/Saphir data of Ref. [@bonn2] which can be interpreted as an absence of $s$-channel excitation in this channel (which brings constraint on the isospin of the hinted resonance around W$\approx$1.9 GeV decaying into the $\Sigma^0$ channel). If the $\Sigma^+$ channel is indeed produced by pure $t$-channel exchange, these data provide us with a strong constraint (and in our case, a nice confirmation) of the strength of the $K^*$ exchange, which is the only allowed leading Regge trajectory and whose coupling constants have been derived independently [@ourpapers2].
=9.5 cm =10. cm
Turning now to polarization observables, Fig. \[fig:spring8\] shows the photon asymmetry recently measured by the Spring-8 collaboration at the LEPS facility [@spring8] in photoproduction for both the $\Lambda$ and $\Sigma$ channels. The general trend of the data is reproduced by the model, in particular the sharp rise of the beam asymmetry at very forward angles (indication of the dominance of the natural parity K\* exchange). At smaller center of mass energies and larger angles, the discrepancies between theory and data become more pronounced. This should not come as a surprise as Regge theory is essentially valid at high energies and forward angles. It is nevertheless certainly of great interest to probe the limits of this model through such data. Let’s also remind that the Regge model describes well at forward angles the magnitude and the sign of the $\Lambda$ and $\Sigma$ recoil polarizations [@ourpapers4].
An electroproduction experiment at JLab [@carman] has measured for the first time, the transfered polarisation from a longitudinally polarized beam to the $\Lambda$ recoil hyperon in the exclusive reaction $e + p \rightarrow e^\prime + K^+ + \Lambda$. We show on Fig. \[fig:carman1\] the only two transferred polarization components $P^\prime_{x^\prime}$ and $P^\prime_{z^\prime}$[^1] which are non-zero when integrated over $\Phi$, the azimuthal angle between the leptonic and the hadronic planes. The longitudinal spin transfer component ($P^\prime_{z^\prime}$) is well reproduced at the very forward angles. The sideways component spin transfer component ($P^\prime_{x^\prime}$) is well reproduced over the full angular range.
=9.5 cm =10. cm
We have plotted, along with the standard $K$+$K^*$ reggeized exchanges which constitutes our full model, the individual contributions of the reggeized $K$ and $K^*$ exchanges. We also show the calculation carried out with standard Feynman propagators of the type $1/(t-m^2_{(K,K^*)})$, instead of Regge propagators of the type $s^{\alpha_{(K,K^*)}(t)}$. It can be seen that these observables are actually barely sensitive to these variations and that basically any model based on $K$ and/or $K^*$ $t$-channel exchange, whatever is the chosen prescription (Feynman or Regge type pole), gives a decent description of the data. However, it should be noted that, when introducing $s$-channel resonances processes (see, for instance the isobaric models cited in Ref. [@carman]), such double polarization observables are very sensitive to resonance properties and allow to discriminate rather precisely between various sets of nucleon resonances participating in the reaction.
In summary, the latest experimental results released in the domain of open strangeness electromagnetic production on the nucleon confirm that our “simple" Regge model surprisingly reproduces the gross features of the data, even for $W<$ 2 GeV. It thus provides an economical description and a simple explanation of the data, hinting that a sort of reggeon-resonance duality is at work here. Where it fails, it gives a useful hint that other mechanisms than simple $t$-channel mechanisms are necessary. In the difficult task of putting into evidence new nucleon resonances in the strangeness channel which are overlapping or hidden into large backgrounds, this sort of clue, i.e. the contribution of non-resonant $t$-channel and Born mechanisms, is certainly much needed.
This work was supported in part by the French CNRS/IN2P3, the French Commissariat à l’Energie Atomique, the Deutsche Forschungsgemeinschaft (SFB443) and the European Commission IHP program (contract HPRN-CT-2000-00130).
[99]{} T. Regge, Nuovo Cimento [**14**]{}, 951 (1959). T. Regge, Nuovo Cimento [**18**]{}, 947 (1960). J. K. Storrow, Phys. Rep. [**103**]{}, 317 (1984). M. Guidal, J.-M. Laget and M. Vanderhaeghen, Phys. Lett. B [**400**]{}, 6. (1997). M. Guidal, J.-M. Laget and M. Vanderhaeghen, Nucl. Phys. A [**627**]{}, 645 (1997). M. Guidal, J.-M. Laget and M. Vanderhaeghen, Phys. Rev. C [**57**]{}, 1454 (1998). M. Guidal, J.-M. Laget and M. Vanderhaeghen, Phys. Rev. C [**61**]{}, 025204 (2000). J.-M. Laget, Phys. Lett. B [**489**]{}, 313 (2000). F. Cano and J.-M. Laget, Phys. Lett. B [**551**]{}, 317 (2003). W.-T. Chiang, S. N. Yang, L. Tiator, M. Vanderhaeghen, D. Drechsel, nucl-th/0212106. R. Dolen, D. Horn and C. Schmidt, Phys. Rev. [**166**]{}, 1768 (1968). G. Niculescu et al., Phys. Rev. Lett. [**81**]{}, 1805 (1998). R. M. Mohring et al., Phys. Rev. C [**67**]{}, 055205 (2003). C.J. Bebek, C.N. Brown, R.V. Kline, F.M. Pipkin, S.W. Raither, L.K. Sisterson, A. Browman, K.M. Hanson, D. Larson and A. Silverman, Phys. Rev. D [**15**]{}, 3082 (1977). P. Brauel, T. Canzler, D. Cords, R. Felst, G. Grindhammer, M. Helm, W.-D. Kollmann, H. Krehbiel and M. Schädlich, Z. Phys. C [**3**]{}, 101 (1979). P. Feller, D. Menze, U. Opara, W. Schulz and W.J. Schwille, Nucl. Phys. B [**39**]{}, 413 (1972). C.J. Bebek, C.N. Brown, P. Bucksbaum, M. Herzlinger, S.D. Holmes, C.A. Lichtenstein, F.M. Pipkin, S.W. Raither and L.K. Sisterson, Phys. Rev. D [**15**]{}, 594 (1977). T. Azemoon, I. Dammann, C. Driver, D. Lüke, G. Specht, K. Heinloth, H. Ackermann, E. Ganssauge, F. Janata and D. Schmidt, Nucl. Phys. B [**95**]{}, 77 (1975). C. N. Brown, C.R. Canizares, W.E. Cooper, A.M. Eisner, G.J. Feldman, C.A. Lichtenstein, L.Litt, W. Lockeretz, V.B. Montana and F.M. Pipkin, Phys. Rev. Lett. [**28**]{}, 1086 (1972). C. J. Bebek et al., Phys. Rev. Lett. [**32**]{}, 21 (1974). Particle Data Group, K. Hagiwara et al., Phys. Rev. D [**66**]{}, 010001 (2002). V. Burkert et al., Phys. Rev. C [**67**]{}, 035204 (2003). J.W.C. McNabb et al., nucl-ex/0305028. M.Q. Tran et al., Phys. Lett. B [**445**]{}, 20 (1998). S. Goers et al., Phys. Lett. B [**464**]{}, 331 (1999). R. G. T. Zegers et al., nucl-ex/0302005. D. Carman et al., Phys. Rev. Lett. [**90**]{}, 131804 (2003).
[^1]: where $x^\prime$ and $y^\prime$ refer to the cartesian coordinates in the $\gamma^*-p$ center of mass frame with the $z$-axis along the direction of the produced kaon
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The low-energy properties of one-dimensional quantum liquids are commonly described in terms of the Tomonaga-Luttinger liquid theory, in which the elementary excitations are free bosons. To this approximation the theory can be alternatively recast in terms of free fermions. In both approaches, small perturbations give rise to finite lifetimes of excitations. We evaluate the decay rate of fermionic excitations and show that it scales as the eighth power of energy, in contrast to the much faster decay of bosonic excitations. Our results can be tested experimentally by measuring the broadening of power-law features in the density structure factor or spectral functions.'
author:
- 'K. A. Matveev'
- 'A. Furusaki'
date: 'December 19, 2013'
title: 'Decay of fermionic quasiparticles in one-dimensional quantum liquids'
---
According to Landau’s Fermi liquid theory [@landau1956Fermiliquid; @landauStat_Mech_II] the low-energy properties of systems of interacting fermions can be described in terms of a gas of weakly interacting quasiparticles. The latter retain fermionic statistics, and their energy depends linearly on the momentum $p$ measured from the Fermi surface, $\varepsilon=v(p-p_F)$. A quasiparticle can decay by exciting a particle-hole pair, but the resulting decay rate $\tau^{-1}\propto\varepsilon^2$ is small compared to the energy.
For a one-dimensional system with linear spectrum, conservation of momentum automatically implies conservation of energy, leading to a divergent decay rate. This results in a failure of the Fermi liquid theory in one dimension. Instead, one-dimensional systems are commonly treated in the framework of the Tomonaga-Luttinger liquid theory [@haldane1981luttinger; @giamarchi2004quantum], where the elementary excitations are bosons. In terms of original fermions, the bosons correspond to the particle-hole pairs. At small momentum $q$, excitation energy is a linear function of $q$, and the Hamiltonian of the system is given by $$\label{eq:Luttinger_Hamiltonian}
H_{TL}=\sum_q v|q|b_q^\dagger b_q^{},$$ where $b_q$ is the boson annihilation operator. In the case of a system of interacting one-dimensional fermions with linear spectrum, the Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) can be derived using the standard bosonization procedure [@haldane1981luttinger; @giamarchi2004quantum].
In general, however, the spectrum of the original particles is not linear, and in addition to $H_{TL}$ the full Hamiltonian contains corrections that account for the effect of the curvature of the spectrum. At low energies, these corrections are small and can often be neglected. For example, the simplest correction contains terms cubic in bosons, $b_{q_1+q_2}^\dagger b_{q_1}b_{q_2}$, and represents an irrelevant perturbation to the Hamiltonian (\[eq:Luttinger\_Hamiltonian\]). On the other hand, this and other irrelevant perturbations enable the decay of bosonic excitations. Thus, similarly to the quasiparticles in a Fermi liquid, bosonic excitations in a Luttinger liquid have a finite decay rate. The evaluation of the lifetimes of bosonic excitations is a challenging problem. However, the basic result for the boson decay rate $\tau^{-1}\propto\varepsilon^2$ [@samokhin_lifetime_1998] can be understood simply as an uncertainty $q^2/m$ of the energy of the particle-hole pair with momentum $q$ caused by the curvature of the spectrum near the Fermi point.
Instead of describing the properties of a Luttinger liquid in terms of bosonic excitations, one can formulate an alternative theory based on quasiparticles with Fermi statistics. This is accomplished by noticing that the Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) gives an exact description of the excitations of a system of noninteracting fermions with linear dispersion [@mattis_exact_1965; @rozhkov_fermionic_2005]. The new fermionic excitations coincide with the original particles if the latter do not interact. It is important, however, that in systems with arbitrarily strong interactions, the quasiparticles are only weakly interacting, in analogy with the Fermi liquid theory in higher dimensions. The interactions result in scattering of the fermionic excitations and give rise to a finite lifetime.
In this paper we evaluate the decay rate of the fermionic excitations in a one-dimensional quantum liquid and show that it scales with energy as $\tau^{-1}\propto\varepsilon^8$. At low energies the respective lifetimes are much longer than those of bosonic excitations. We therefore show that the fermionic picture has a significant advantage over the conventional bosonic one when the curvature of the spectrum is important. Experimentally, the decay of fermionic excitations should manifest itself as broadening of sharp features at the quasiparticle mass shell in the density structure factor and spectral functions.
The various phenomena caused by spectral curvature in one-dimensional systems have been the subject of active study in the last few years [@imambekov_one-dimensional_2012]. Notably, the decay rate of excitations in a weakly interacting Fermi gas was studied in Ref. . In a system with quadratic correction to the spectrum of the fermions, the decay of hole excitations at zero temperature is forbidden by the conservation laws, whereas for the particle excitations the result $\tau^{-1}\propto\varepsilon^8$ was obtained. Unlike Ref. , we are interested in a system with arbitrary interaction strength. It is important to stress that in this case the fermionic excitations are not the original particles studied in Ref. , but the true quasiparticle excitations of the Luttinger liquid defined via the refermionization procedure [@rozhkov_fermionic_2005].
The simplest way to introduce fermionic quasiparticles in the Tomonaga-Luttinger liquid is by noticing that its Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) has the same basic form for both interacting and noninteracting fermions, although the formal transformation from fermions to bosons does depend on interactions. Let us consider a system of free fermions described by the Hamiltonian $$\label{eq:H_0}
H_0=\sum_k\varepsilon_k a_k^\dagger a_k^{}
+\sum_p\tilde\varepsilon_{p} \tilde a_p^\dagger \tilde a_p^{},$$ Here $a_k$ and $\tilde a_p$ are the fermion operators for particles on the right- and left-moving branches, with momenta $k$ and $p$ measured from the respective Fermi points. If the spectra of the fermions are linear, $\varepsilon_k=vk$ and $\tilde\varepsilon_p=-vp$, the Hamiltonian (\[eq:H\_0\]) can be brought to the form (\[eq:Luttinger\_Hamiltonian\]) using the standard bosonization prescription [@haldane1981luttinger; @giamarchi2004quantum]
\[eq:bosonization\] $$\begin{aligned}
b_q&=&i\left(\frac{2\pi\hbar}{qL}\right)^{1/2}
\sum_k a_k^\dagger a_{k+q}^{},
\quad
q>0,
\\
b_q&=&-i\left(\frac{2\pi\hbar}{|q|L}\right)^{1/2}
\sum_p \tilde a_{p}^\dagger \tilde a_{p+q}^{},
\quad
q<0,\end{aligned}$$
where $L$ is the system size. For nonvanishing interactions, the bosonization transformation is somewhat more complicated. It is controlled by the so-called Luttinger liquid parameter $K$, which takes values $K<1$ for repulsive interactions and $K>1$ for the attractive ones [@giamarchi2004quantum].
Since the bosonic Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) is equivalent to the fermionic Hamiltonian (\[eq:H\_0\]) with linear spectrum, instead of the standard Tomonaga-Luttinger liquid theory with bosonic excitations $b_q$ one can construct an equivalent theory based on the free fermion quasiparticles $a_k$ and $\tilde a_p$. The latter will coincide with the original fermions constituting the Luttinger liquid only in the absence of interactions, i.e., at $K=1$.
In most physical systems the spectrum is not strictly linear. To account for the curvature, one has to add to the Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) various perturbations that are irrelevant in the renormalization group sense. Such perturbations are given by operators with scaling dimensions higher than 2. Because each bosonic operator is mapped to a pair of fermion operators on the same branch \[see Eq. (\[eq:bosonization\])\], the irrelevant perturbations have even numbers of fermion operators on each branch. In the simplest case, the perturbation consists of two fermion operators, e.g., $k^la_k^\dagger a_k^{}$ with $l=2,3,\ldots$. These perturbations can be included into the Hamiltonian (\[eq:H\_0\]) by allowing for the nonlinear quasiparticle spectrum $$\label{eq:varepsilon}
\varepsilon_k=vk+\frac{k^2}{2m^*}+\frac{\lambda}{6}\,k^3+\ldots$$ and $\tilde\varepsilon_p=\varepsilon_{-p}$.
More complicated perturbations contain four, six, or more quasiparticle operators. For instance, the only such operator of scaling dimension 3 is [@rozhkov_fermionic_2005] $$\label{eq:V_3}
V_3=\frac{\gamma_3}{L}\sum_{kk'pp'}
(k+k'-p-p')\,\delta_{k+p,k'+p'}
\tilde a_{p'}^\dagger \tilde a_p^{} a_{k'}^\dagger a_k^{}.$$ Here $\gamma_3$ is a constant, and the Kronecker delta reflects the conservation of momentum. The perturbation (\[eq:V\_3\]) describes an effective interaction of two quasiparticles. Perturbations of higher dimensions include interactions of any number of quasiparticles. Because the interactions between the fermionic quasiparticles are described by irrelevant perturbations, they are weak at low energies.
In the presence of interactions, the fermionic quasiparticles must have a finite decay rate, which is the main subject of this paper. Specifically, we consider a state with a filled Fermi surface of states with $k<0$ and $p>0$ on the right- and left-moving branches, respectively, and one additional quasiparticle in state $k_1>0$ on the right-moving branch. The two-particle scattering processes do not simultaneously conserve the momentum and energy of the system, so the simplest allowed process is three-particle scattering. Furthermore, at zero temperature two of the three quasiparticles must be on the same branch, while the third is on the other branch [@khodas2007fermi], Fig. \[fig:decay-process\].
![A three-particle process leading to the decay of a quasiparticle in the state $k_1$ above the Fermi level involves two additional quasiparticles below the Fermi level, one, $k_2$, on the same branch and the other, $p$, on the opposite branch.[]{data-label="fig:decay-process"}](decay-process.eps){width=".45\textwidth"}
The decay rate is then found from Fermi’s golden rule $$\begin{aligned}
\label{eq:golden_rule}
\frac1\tau&=&\frac{\pi}{\hbar}
\sum_{p,k_1',k_2'>0\atop p',k_2<0}
\Big|{\mathcal A}_{k_1,k_2,p}^{k_1',k_2',p'}\Big|^2
\nonumber\\
&&\times
\delta(\varepsilon_{k_1}+\varepsilon_{k_2}+\tilde\varepsilon_{p}
-\varepsilon_{k_1'}-\varepsilon_{k_2'}-\tilde\varepsilon_{p'}),\end{aligned}$$ where the transition matrix element is defined in terms of the $T$-matrix as $$\label{eq:matrix_element}
{\mathcal A}_{k_1,k_2,p}^{k_1',k_2',p'}=
\langle
a_{k_1'} a_{k_2'} \tilde a_{p'}|T|\tilde a_{p}^\dagger a_{k_2}^\dagger a_{k_1}^\dagger
\rangle.$$ A three-particle scattering event can be accomplished in the second order in perturbations coupling two particles, such as the term (\[eq:V\_3\]). In addition, a contribution to the amplitude (\[eq:matrix\_element\]) can be obtained in the first order in three-particle coupling, which appears in perturbations with scaling dimensions higher than 4.
Because the conservation of momentum imposes restrictions on the final states of the three particles, the matrix element (\[eq:matrix\_element\]) must have the form $\mathcal
A\delta_{k_1+k_2+p,k_1'+k_2'+p'}$. Furthermore, since we are interested in the decay rate of a quasiparticle with small momentum $k_1$, and all the other quasiparticles have momenta smaller than $k_1$ in absolute value, one may expect to be able to take the limit $k_1, k_1', \ldots\to0$ and replace $\mathcal A$ with the resulting constant. This would be incorrect because the quasiparticles on the same branch are identical fermions, and thus the matrix element (\[eq:matrix\_element\]) is antisymmetric with respect to permutations $k_1\leftrightarrow k_2$ and $k_1'\leftrightarrow k_2'$. We therefore conclude that $$\label{eq:matrix_element_ansatz}
{\mathcal A}_{k_1,k_2,p}^{k_1',k_2',p'}=
\frac{\Lambda}{L^2}(k_1-k_2)(k_1'-k_2')\delta_{k_1+k_2+p,k_1'+k_2'+p'},$$ where $\Lambda$ is symmetric with respect to the above permutation and takes a constant value at $k_1\to0$. We note that in the limit of weak interactions the three-particle scattering amplitude does have the form (\[eq:matrix\_element\_ansatz\]) at $k_1\to0$, cf. Eq. (47) in Ref. .
Substituting the scattering amplitude (\[eq:matrix\_element\_ansatz\]) into Eq. (\[eq:golden\_rule\]) we obtain the expression $$\label{eq:decay_rate_result}
\frac1\tau=\frac{3}{5120\pi^3}
\frac{\Lambda^2\varepsilon^8}{\hbar^5 m^* v^{10}}$$ for the decay rate of a fermionic quasiparticle with energy $\varepsilon=vk_1$. The strong suppression of the quasiparticle decay at low energies, $\tau^{-1}\propto \varepsilon^8$, is our main result.
Similar to the decay rate for weakly interacting fermions, the eighth power of energy is a combined effect of the weak scattering amplitude, $|\mathcal A|^2\propto \varepsilon^4$, and small phase space volume for three-particle scattering, $\nu\propto \varepsilon^4$ [@khodas2007fermi; @pereira_spectral_2009; @ristivojevic_relaxation_2013]. The former result is a direct consequence of the Fermi statistics of the quasiparticles, whereas the latter is limited to systems where the quadratic correction to the spectrum (\[eq:varepsilon\]) is not forbidden by symmetry. For example, the result (\[eq:decay\_rate\_result\]) does not apply to interacting fermions on a tight-binding chain at half filling. Apart from this limitation, our result (\[eq:decay\_rate\_result\]) is rather generic. Most importantly, it is not limited to systems of weakly interacting fermions. In particular, it applies at strong interactions, the so-called Wigner crystal regime, when the Luttinger-liquid parameter $K\ll1$. One should keep in mind, however, that in this case the description of the system in terms of weakly interacting fermionic excitations is expected to be valid only at particularly small energies $\varepsilon\ll \sqrt K\,vp_F$ [@lin_thermalization_2013].
The prefactor in our result (\[eq:decay\_rate\_result\]) for the decay rate is expressed in terms of the parameter $\Lambda$ introduced in Eq. (\[eq:matrix\_element\_ansatz\]). A microscopic theory for it can be developed in the limit of weak interactions [@khodas2007fermi; @ristivojevic_relaxation_2013]. At arbitrary interaction strength, an analytic microscopic treatment is possible only for integrable models, in which excitations are expected to have infinite lifetimes, i.e., $\Lambda=0$. On the other hand, it is possible to obtain an expression for $\Lambda$ in terms of the parameters $v$, $m^*$, and $\lambda$ of the excitation spectrum (\[eq:varepsilon\]) and their dependences on the density $n$ and momentum per particle $\kappa$ of the liquid.
The most straightforward approach involves identifying the possible perturbations to the Hamiltonian (\[eq:H\_0\]) up to the scaling dimension 5 and performing the evaluation of the $T$-matrix up to second order in such perturbations. This is a laborious procedure that will be discussed elsewhere. Here we pursue an alternative approach based on the idea that a quasiparticle can often be treated as a mobile impurity in the Luttinger liquid [@ogawa1992fermi; @neto1996dynamics; @khodas2007fermi; @imambekov2008exact; @pereira_spectral_2009; @schecter_dynamics_2011; @matveev_scattering_2012].
Luttinger liquid theory applies only to low-energy properties of the system. Thus its Hamiltonian (\[eq:Luttinger\_Hamiltonian\]) accounts only for the excitations with energies below certain bandwidth $D$. The exact value of $D$ is usually not important, as long as it is small compared to the characteristic energy scales of the problem, such as the Fermi energy. In our discussion so far the quasiparticles were constructed out of bosonic excitations of the Luttinger liquid, i.e., the quasiparticle energy $\varepsilon \ll D$. Alternatively, one can choose $D\ll\varepsilon$, in which case the quasiparticle is not part of the Luttinger liquid and should be treated as a mobile impurity. Let us consider a special case of the three-particle scattering process (\[eq:matrix\_element\]) for which $k_2'=-Q$ and $k_2=-Q-\delta Q$, where $Q\gg D/v$, and all the remaining momenta are such that $|k_1|, |k_1'|, |p|, |p'|\ll D/v$. According to Eq. (\[eq:matrix\_element\_ansatz\]), to leading order in small $\delta Q$, the scattering matrix element for this process is given by $$\label{eq:low_momentum_scattering}
\mathcal A = \frac{\Lambda}{L^2}Q^2\,
\delta_{q-\tilde q, \delta Q},$$ where $q=k_1-k_1'$ and $\tilde q=p'-p$ are the momenta of the particle-hole pairs collapsed near the right Fermi point and created near the left one, respectively.
Since the particle-hole pairs correspond to bosons of the standard Luttinger liquid theory (\[eq:Luttinger\_Hamiltonian\]), one can think of this process as scattering of a hole from state $Q$ to $Q+\delta Q$ while absorbing a boson $q$ and emitting a boson $\tilde
q$. Such processes were studied in Ref. . The expression for the respective scattering amplitude can be brought to the form [@matveev_equilibration_2013] $$\label{eq:t-total}
t_{q,\tilde q} = -\frac{\sqrt{|q\tilde q|}}{2\pi\hbar L}\,Y_{Q},$$ with $$\begin{aligned}
\label{eq:Y}
&&Y_Q=
\partial^2_{LR}\varepsilon_Q
-\frac{1}{m_Q^*}
\frac{\partial_L\varepsilon_{Q}}{v+v_Q}
\frac{\partial_R\varepsilon_{Q}}{v-v_Q}
+\partial_L v_Q
\frac{\partial_R\varepsilon_{Q}}{v-v_Q}
\nonumber\\
&&\hspace{0.5em} -\partial_R v_Q
\frac{\partial_L\varepsilon_{Q}}{v+v_Q}
+\frac{v\partial_n K}{\sqrt K}
\bigg(
\frac{\partial_R\varepsilon_Q}{v-v_Q}
+\frac{\partial_L\varepsilon_Q}{v+v_Q}
\bigg).\end{aligned}$$ The quasiparticle velocity and mass here depend on momentum, $v_Q=\varepsilon_Q'$ and $1/m_Q^*=\varepsilon_Q''$, and the following shorthand notation is used:
$$\begin{aligned}
\label{eq:derivatives}\hspace{-2em}
\partial_R &=& \sqrt{K}\partial_n
+\frac{\pi\hbar}{\sqrt K}\partial_\kappa,
\\
\partial_L &=& \sqrt{K}\partial_n
-\frac{\pi\hbar}{\sqrt K}\partial_\kappa,
\\
\partial^2_{LR} &=& K\partial_n^2-\frac{(\pi\hbar)^2}{K}\partial_\kappa^2.\end{aligned}$$
Because scattering of the hole by two bosons in the Luttinger liquid is a special case of the three-fermion scattering event, we can use the above result to determine $\Lambda$. To this end, we substitute the $T$-matrix $t_{q,\tilde q}\,a_{-Q}^\dagger a_{-Q-\delta Q}b_{\tilde
q}^\dagger b_{q}^{}$ into Eq. (\[eq:matrix\_element\]) and use Eq. (\[eq:bosonization\]) for the boson operators. The resulting scattering amplitude has the form (\[eq:low\_momentum\_scattering\]) with $\Lambda=Y_Q/Q^2$. As expected, at $Q\to0$ the latter expression has a finite limit $$\begin{aligned}
\label{eq:Lambda}
\Lambda&=&
\frac12\left(\partial^2_{LR}\frac{1}{m^*}-2\pi\partial_L\lambda\right)
-\frac{\partial_L v}{4v}\partial_L\frac{1}{m^*}
+\frac{(\partial_L v)^2}{4m^*v^2}
\nonumber\\ &&
-\left(\frac{\partial_L v}{4v}
+\frac{m^*}{2}\partial_L\frac{1}{m^*}\right)
\left(\partial_{R}\frac{1}{m^*}-2\pi\lambda\right).\end{aligned}$$ It is worth noting that the absence of the inversion symmetry in the above expression is caused by our choice of the quasiparticle on the right-moving branch.
Expression (\[eq:Lambda\]) relates $\Lambda$ to the parameters $v$, $m^*$, and $\lambda$ of the quasiparticle spectrum (\[eq:varepsilon\]) and their dependences on the particle density $n$ and momentum per particle $\kappa$ of the liquid. In combination with Eq. (\[eq:decay\_rate\_result\]) it gives a complete expression for the decay rate of fermionic quasiparticles in a spinless quantum liquid. Our result is valid at any strength of the interactions between the physical particles constituting the liquid. Although our discussion focused on liquids of spinless fermions, the results are also applicable to one-dimensional systems of interacting bosons.
Relaxation of excitations in a one-dimensional system with spins was recently observed in experiment with quantum wires [@barak_interacting_2010]. To test our results, the spins may be polarized by a sufficiently strong in-plane magnetic field. More generally, the decay of quasiparticles may be observed as the broadening of sharp features in the density structure factor or spectral functions. Both types of measurements can, in principle, be performed in experiments with two parallel quantum wires. The density structure factor controls the Coulomb drag in such devices [@pustilnik_coulomb_2003], whereas the spectral functions can be measured in experiments with momentum-resolved tunneling between the wires [@auslaender_spin-charge_2005]. The same experiments would also measure the excitation spectrum, enabling a quantitative test of our results (\[eq:decay\_rate\_result\]) and (\[eq:Lambda\]).
The authors are grateful to L. I. Glazman, M. Pustilnik, and Z. Ristivojevic for helpful discussions. This work was supported by UChicago Argonne, LLC, under contract No. DE-AC02-06CH11357, by JSPS KAKENHI Grant No. 24540338, and by the RIKEN iTHES Project. This work was supported in part by the National Science Foundation under Grant No. PHYS-1066293 and the hospitality of the Aspen Center for Physics.
[21]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ** (, ).
, ****, ().
, ** (, ).
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, , , , , , , , ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Neutrino oscillations are experimentally observable only as a result of interference between neutrino states with different masses and THE SAME ENERGY. All interference effects between neutrino states having different energies are destroyed by the intereaction between the incident neutrino and the neutrino detector. Erroneous results are frequently obtained by neglecting the neutrino-detector interactions.'
address: |
Department of Particle Physics Weizmann Institute of Science, Rehovot 76100, Israel\
School of Physics and Astronomy, Raymond and Beverly Sackler Faculty of Exact Sciences, Tel Aviv University, Tel Aviv, Israel\
High Energy Physics Division, Argonne National Laboratory, Argonne, IL 60439-4815, USA\
\
[email protected]\
\
author:
- 'Harry J. Lipkin[^1]'
title: 'Stodolsky’s Theorem and Neutrino Oscillation Phases – for pedestrians'
---
\#1[$$\label{#1}}
\def\eeq{$$]{}
\#1[\#1]{} \#1[\#1]{}
Stodolsky[@Leo] has given a very simple answer to the confusion that still arises in discussions of the phase of neutrino oscillations. The relevant literature producing this confusion has recently been summarized and clarified[@Okun]. The purpose of this note is to support this excellent analysis[@Okun] without engaging in a direct debate against the confusing articles and also to present a “pedestrian" version of Stodolsky’s work which is hopefully understandable to students and experimentalists.
The detection of a neutrino always involves its interaction with a detector that is part of an environment described by a density matrix in which the energy is diagonal. Unless this interaction with the environment is turned off, and no experiment can do this, all relative phase information between neutrino states with different energies is destroyed.
The question is not whether states of the same momentum and different energies are coherent, states of the same energy and different momentum or states of the same velocity. There have been many irrelevant arguments about these issues. But states with different ENERGIES ARE NEVER COHERENT in any realistic experiment. States of the same energy and different momenta can be coherent, but may not be. This depends upon the way the measurement is made. But states with different energies can not be coherent.
This discussion refers only to neutrino detectors, The usual detector is a nucleon, which changes its state after absorbing a neutrino and emitting a charged lepton, and is initially either in an energy eigenstate or in a statistical mixture in thermal equilibriam with its surroundings. No neutrino detector has ever been prepared in a coherent mixture of energy eigenstates and no such detector has been proposed for future experiments .
All arguments about Lorentz invariance are irrelevant. The detector chooses a particular Lorentz frame where the detector is at rest and described by a density matrix in which the energy is diagonal and no interference between states of different energies can be observed.
Most treatments do not consider at all the quantum mechanics of the detector. Since the detector is a quantum system (e.g. a nucleon) which undergoes a transition together with the neutrino-to-charged-lepton transition, and the initial and final states of the detector are not measured, the transition probability is the square of the transition matrix element for the whole system, averaged over detector initial states and summed over detector final states. This immediately kills all interference between neutrino states with different energies as they are accompanied by different detector states which must have different energies because energy in conserved in the process. The detector states with different energies are orthogonal to one another and all interference terms between them vanish because of this detector orthogonality.
In this context the “factor-of-two" arguments are seen to be missing an essential point in the actual neutrino oscillation experiments; namely the role of the detector as a quantum-mechanical system entangled with the neutrino.
The standard textbook neutrino-oscillation wave function, a coherent linear combination of states with different energies, is not found in real experiments. Thus considerable confusion remains even though coherence, interference and dephasing have been extensively discussed and clarified [@Leo; @Dost; @QM; @NeutHJL; @Kayser; @GoldS; @Pnonexp; @MMNIETO; @GrossLip; @okun1; @pnow98]. Elementary quantum mechanics and quantum statistical mechanics tell us that the components of the density matrix describing a neutrino detector and having different energies are never coherent[@Leo], while neutrino components with different masses and different momenta must be coherent to cancel components with the wrong flavor just outside the neutrino source.. This coherence between source states having the same energy and different momenta can produce coherence between neutrino states with the same energy and different masses.
This physics is illustrated in detail in a toy model[@pwhichfin] for the detection of a neutrino as a transition between an initial state of a neutrino and a detector and a final state of a muon and the same detector. The wave function for the initial state of neutrino and detector is \_i(,D) = \_[k=1]{}\^[N\_]{} where $N_\nu$ is the number of neutrino mass states, $E_\nu$, $m_k$ and $\vec P_k$ denote the neutrino energy, mass and momentum and $D_i(E_i)$ is the initial state of the detector with energy $E_i$. If the detector is a muon detector the final detector state after neutrino absorption is \_f(\^,D) = \_[k=1]{}\^[N\_]{} where $E_\mu$ and $\vec P_\mu$ denote the muon energy and momentum, $D^\mp_{kf}$ is the final detector state produced in the “path $k$"; i.e. after the absorption of a neutrino with mass $m_k$ and emission of a $\mu^\pm$, and $E=E_\nu + E_i$ is the total energy which is conserved in the transition.
The transition in the detector occurs on a nucleon, whose co-ordinate is denoted by by $\vec X$, and involves a charge exchange denoted by the isospin operator $I_{\mp}$ and a momentum transfer $\vec P_k -\vec P_\mu$. The detector transition matrix element is therefore given by T\^ = I\_e\^[i(P\_k -P\_) X]{}
The overlap between the final detector wave functions after the transitions absorbing neutrinos with masses $m_k$ and $m_j$ is then = e\^[i(P\_j-P\_k) X]{}
If the quantum fluctuations in the position of the active nucleon in the initial state of the detector are small in comparison with the oscillation wave length, $\hbar /(\vec P_j-\vec P_k)$, |P\_j -P\_k|\^2 |X\^2| 1 1 - (1/2)|P\_j -P\_k|\^2 |X\^2| 1
There is thus effectively a full overlap between the final detector states after absorption of different mass neutrinos, and a full coherence between the neutrino states with the same energy and different momenta.
The total energies of the final muon and detector produced after absorption of neutrinos with different energies are different. These muon-detector states are thus orthogonal to one another and there is no coherence between detector states produced by the absorption of neutrinos with different energies.
There have been suggestions for bypassing Stodolsky’s theorem by exploiting some kind of energy-time uncertainty to detect interference between components having different energies in the neutrino wave function. The time of flight of the neutrino from source to detector might be measured by detecting the muon emitted together with the neutrino in a pion decay in the source and measuring precisely the times of emission in the source and of absorbtion in the detector.
However, if the quantum fluctuations in the position of the active nucleon in the initial state of the detector are small in comparison with the oscillation wave length, eqs. (\[WP26\]) and (\[WP27\]) apply and the coherence and relative phase of the components in the neutrino wave function having the same energy and different momenta are preserved . This relative phase completely determines the flavor output of the detector; i.e. the relative probabilities of producing a muon or an electron. These probabilities in all realistic cases are essentially independent of energy over the relevant energy range. Thus the relative phases and coherence between components in the neutrino wave function with different energies is irrelevant. All energies give the same muon/electron ratio whether they add coherently or incoherently and time measurements cannot change the muon/electron ratio observed at the detector.
It is a pleasure to thank Maury Goodman, Yuval Grossman, Boris Kayser, Lev Okun,and Leo Stodolsky for helpful discussions and comments.
[^1]: Supported in part by grant from US-Israel Bi-National Science Foundation and by the U.S. Department of Energy, Division of High Energy Physics, Contract W-31-109-ENG-38.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The problem of determining cyclotomic numbers in terms of the solutions of certain Diophantine systems has been treated by many authors since the age of Gauss. In this paper we obtain an explicit expression for cyclotomic numbers of order $2\ell^{2}$ in terms of the coefficients of the Jacobi sums of lower orders. At the end, we illustrate the nature of two matrices corresponding to two types of cyclotomic numbers.'
address:
- 'Md Helal Ahmed @ Centre for Applied Mathematics, Central University of Jharkhand, Ranchi-835205, India'
- 'Jagmohan Tanti @ Centre for Applied Mathematics, Central University of Jharkhand, Ranchi-835205, India'
- 'Azizul Hoque @Azizul Hoque, Harish-Chandra Research Institute, Chhatnag Road, Jhunsi, Allahabad 211 019, India.'
author:
- 'Md Helal Ahmed, Jagmohan Tanti and Azizul Hoque'
---
Introduction
============
Let $p$ be an odd prime integer and $q=p^r$ with $r\geq 1$ an integer. Let $e$ be a non-trivial divisor of $q-1$. Then $q=ek+1$ for some positive integer $k$. Suppose $\gamma$ is a generator of the cyclic group $\mathbb{F}^{\times}_{q}$. For a primitive $e$-th root $\zeta_e$ of unity, define a multiplicative character $\chi_e$ on $\mathbb{F}^{\times}_{q}$ by $\chi_e(\gamma)=\zeta_e$. For $0\leq i, j\leq e-1$, the cyclotomic numbers $(i, j)_e$ of order $e$ are defined as follows: $$(i,j)_e:=\#\{v\in\mathbb{F}_q\setminus \{0,-1\}\mid {\rm ind}_{\gamma}v\equiv a \pmod e , {\ \rm ind}_{\gamma}(v+1)\equiv b \pmod e\}.$$ We now extent $\chi_e$ to a map from $\mathbb{F}_q$ to $\mathbb{Q}({\zeta_e})$ by taking $\chi_e(0)=0$. The Jacobi sums of order $e$ is defined by $$J_e(i,j)= \sum_{v\in \mathbb{F}_q} \chi_e^i(v) \chi_e^j(v+1).$$ The cyclotomic numbers $(i,j)_e$ and the Jacobi sums $J_e(i,j)$ are well connected by the following relations: $$\label{01}
\sum_i\sum_j(i,j)_e\zeta_e^{ai+bj}=J_e(a,b),$$ and $$\label{00}
\sum_i\sum_j\zeta_e^{-(ai+bj)}J_e(i,j)=e^{2}(a,b)_e.$$ and show that if we want to calculate all the cyclotomic numbers $(i,j)_e$ of order $e$, it is sufficient to calculate all the Jacobi sums $J_e(i,j)$ of the same order, and vice-versa.
Cyclotomic numbers are one of the most important objects in number theory and in other branches of mathematics. These number have been extensively used in coding theory, cryptography and in other branches of information theory. One of the central problems in the study of these numbers is the determination of all cyclotomic numbers of a specific order for a given field in terms of solutions of certain Diophantine system. This problem has been treated by many mathematicians including C. F. Gauss who had determined all the cyclotomic numbers of order $3$ in the field $\mathbb{F}_q$ with prime $q\equiv 1\pmod 3$. Cyclotomic numbers of order $e$ over the field $\mathbb{F}_q$ with characteristic $p$, in general, can not be determined only in terms of $p$ and $e$, but that one requires a quadratic partition of $q$ too.
Complete solutions to this cyclotomic number problem have been computed for some specific orders. For instance, the cyclotomic numbers of prime order $e$ in the finite field $\mathbb{F}_q$ with $q=p^r$ and $p\equiv 1 \pmod e$ have been investigated by many authors (see, [@KR1985] and the references therein). Cyclotomic numbers of small composite order have been investigated by many authors, but most of the results involve the classical ambiguity. The problem of removal of this ambiguity may led to composite moduli. The first result in composite case is due to L. E. Dickson [@DI1935] who had computed cyclotomic numbers of orders $4, 6, 8, 10$ and $12$ over the field $\mathbb{F}_q$ for $r=1$. In [@ST1967], the author showed that the cyclotomic numbers of orders $3, 4, 6$ over the field $\mathbb{F}_q$ are well determined by a unique representation of $p^r$ in terms of a particular binary quadratic form $x^2+dy^2$, and that of order $8$ are determined by two such forms. L. E. Dickson also determined the cyclotomic numbers of orders $ 14, 22$ in [@DI1935-2], and that of orders $9, 15, 16, 18, 20, 24$ in [@DI1935-3] over the field $\mathbb{F}_q$ for $r=1$. Following the techniques of [@DI1935], A. L. Whiteman [@WH1960] computed the cyclotomic numbers of order $12$ over the field $\mathbb{F}_q$ for $r=1$. He also computed the cyclotomic numbers of order $16$ over the same field in [@WH1957]. These cyclotomic numbers of order $16$ were also determined independently by E. Lehmer in [@LE1954] and R. J. Evans and J. R. Hill in [@EH1979]. J. B. Muskat and A. L. Whiteman gave the complete theory for the cyclotomic numbers of order $20$ in [@MW1970]. J. B. Muskat analysed completely the cyclotomic numbers of order $14$ in [@MU1966] and that of orders $24, 30$ in [@MU1968]. In [@FMSW1986], the authors discussed the cyclotomic numbers of order $15$ over the field $\mathbb{F}_{q}$ for $q=p^2$ when $p\equiv 4, 11\pmod{15}$. For $\ell$ an odd prime, Katre and Rajwade [@KR1985] solved the cyclotomy problem of order $\ell$ whereas Acharya and Katre [@AK1995] determined the cyclotomic numbers of order $2\ell$ over the field $\mathbb{F}_q$ for $q=p^r$ with the prime $p\equiv 1\pmod \ell$ in terms of the solutions of certain Diophantine systems. Recently Shirolkar and Katre [@SK2011] determined the cyclotomic numbers of order $\ell^2$ with prime $\ell$ in terms of the coefficients of Jacobi sums of orders $\ell$ and $\ell^2$.
In this paper, We obtain a formula for the cyclotomic numbers of order $2\ell^2$ over the finite field $\mathbb{F}_q$ for $q=p^r\equiv 1\pmod {2\ell^2}$ in terms of the coefficients of Jacobi sums of orders $\ell$, $2\ell$, $\ell^2$ and $2\ell^2$. These coefficients can be evaluated in terms of the Dickson-Hurwitz sums of of orders $\ell$, $2\ell$, $\ell^2$ and $2\ell^2$. We also discuss the nature of two types of cyclotomic numbers in terms of their associated matrix representations.
Preliminaries
=============
For an odd prime $\ell$, let $p$ be a prime and $q=p^r$ with an integer $r\geq 1$ satisfying $q\equiv 1\pmod {2\ell^2}$. We write $q=2\ell^2k+1$ for some positive integer $k$. In next two subsequent subsection we state some well known properties [@SK2011] of cyclotomic numbers and Dickson-Hurwitz sums which are needed to give an explicit expression for cyclotomic numbers of order $2\ell^{2}$
Properties of Cyclotomic Numbers
--------------------------------
1. $(a,b)_{2\ell^2}=(a' ,b')_{2\ell^2}$ whenever $a\equiv a'\pmod{2\ell^2}$ and $b\equiv b'\pmod{2\ell^2}$.\
2. $(a,b)_{2\ell^2}= (2\ell^2-a,b-a)_{2\ell^{2}}$ along with the following: $$\label{2.1}
(a,b)_{2\ell^{2}}=\begin{cases}
(b,a)_{2\ell^{2}}\hspace*{1.751cm} \text{ if } k \text{ is even or } q=2^r,\\
(b+\ell^{2},a+\ell^{2})_{2\ell^{2}}\hspace*{2mm} \text{ if } k \text{ is odd}.
\end{cases}.$$
3. $$\label{2.2}
\sum_{a=0}^{2\ell^{2}-1}\sum_{b=0}^{2\ell^{2}-1}(a,b)_{2\ell^{2}}=q-2.$$
4. $$\label{2.3}
\sum_{b=0}^{2\ell^{2}-1}(a,b)_{2\ell^{2}}=k-n_{a},$$
where $n_{a}$ is given by $$n_{a}=\begin{cases}
1 \quad \text{ if } a=0, 2\mid k \text{ or if } a=\ell^2, 2\nmid k;\\
0 \quad \text{otherwise}.
\end{cases}$$
5. $$\label{2.4}
\sum_{a=0}^{2\ell^{2}-1}(a,b)_{2\ell^{2}}=\begin{cases}
k-1 \hspace*{3mm}\text{ if } b =0; \\
k \hspace*{1cm}\text{ if } 1\leq b \leq 2\ell^{2}-1.
\end{cases}$$
Properties of Dickson-Hurwitz sums
----------------------------------
Dickson-Hurwitz sums, $B_{2\ell^2}(a,b)$ [@ZE1971] of order $2\ell^2$ are defined for two positive integers $a$ and $ b$ modulo $2\ell^2$ by $$\label{2.5}
B_{2l^{2}}(a,b)=\sum_{h=0}^{2\ell^{2}-1}(h,a-bh)_{2\ell^{2}}.$$
1. $$\label{2.6}
\sum_{a=0}^{2l^{2}-1}B_{2l^{2}}(a,b)=q-2.$$
2. $$\label{2.7}
B_{2l^{2}}(a,0)=\begin{cases}
k-1 \hspace*{3mm}\text{ if } a =0; \\
k \hspace*{1cm}\text{ if } 1\leq a \leq 2\ell^{2}-1.
\end{cases}$$
3. $$\begin{aligned}
\label{xx}
B_{2\ell^{2}}(\ell^{2}+k,n) &\equiv \sum_{a=0}^{2\ell^{2}-1}B_{2\ell^{2}}(a-\ell,n)\pmod{2\ell^{2}}\nonumber \\
&\equiv \sum_{a=0}^{2\ell^{2}-1}B_{2\ell^{2}}(a,n)\pmod{2\ell^{2}}\nonumber\\
&\equiv q-2\pmod {2\ell^2},\end{aligned}$$
where $0\leq k \leq \phi(2\ell^{2})-1$.
Relations between Jacobi Sums and Dickson-Hurwitz Sums
------------------------------------------------------
For any positive integers $m$ and $n$ modulo $2\ell^2$, we recall the following relation (see, relation (8) in [[@DI1935-2]]{}), $$\label{2.8}
J_{2l^{2}}(m,n)=\sum_{a=0}^{2\ell^{2}-1}\zeta_{2\ell^2}^{na}\sum_{b=0}^{2\ell^{2}-1}\zeta_{2\ell^2}^{-(m+n)b}(a,b)_{2\ell^{2}},$$ where $\zeta_{2\ell^2}$ is a primitive $2\ell^{2}$-th root of unity.
We also recall (see, relation (9) in [[@DI1935-2]]{}), $$\label{2.9}
J_{2\ell^{2}}(m,n)=(-1)^{nk}J_{2\ell^{2}}(-m-n,n).$$ Puttig $m=2$ in and $m=-2-n$ in , we obtain respectively: $$J_{2\ell^{2}}(2,n)=(-1)^{nk}J_{2\ell^{2}}(-2-n,n)$$ and $$J_{2\ell^{2}}(-2-n,n)=\sum_{a=0}^{2\ell^{2}-1}\sum_{b=0}^{2\ell^{2}-1}\zeta_{2\ell^2}^{na+2b}(a,b)_{2\ell^{2}}.$$
Eliminating $b$ by the use of $na+2b\equiv i \pmod {\ell^{2})}$, we obtain $$J_{2\ell^{2}}(-2-n,n)=\sum_{i=0}^{\ell^{2}-1}\sum_{a=0}^{\ell^{2}-1}\zeta_{\ell^2}^{i}\big(a,\frac{i-na}{2}\big)=(-1)^{nk} J_{2\ell^{2}}(2,n),$$ Writing $B_{\ell^{2}}(i,n)=\sum_{a=0}^{\ell^{2}-1}(a,(i-na)/2)$, we get
$$\label{2.10}
(-1)^{nk}J_{2\ell^{2}}(2,n)=\sum_{i=0}^{\ell^{2}-1}B_{\ell^{2}}(i,n)\zeta_{\ell^2}^{i}.$$
where $\zeta_{\ell^2}$ is a $\ell^{2}$-th root of unity.\
Also from [@DI1935-2], we have $$\label{2.11}
(-1)^{nk}J_{2\ell^{2}}(1,n)=\sum_{i=0}^{2\ell^{2}-1}B_{2\ell^{2}}(i,n)\zeta_{2\ell^2}^{i}.$$
Some results on Dickson-Hurwitz sums
====================================
In this section, we determine the trace of $J_{2\ell^2}(1,n)\zeta_{2\ell^2}^{-t}$, $J_{2\ell^2}(2, n)\zeta_{2\ell^2}^{-t}$ and also obtain a result that shows a relation between the coefficients $d_{i,n}$ and Dickson-Hurwitz sums of order $2\ell^{2}$ which are needed in the subsequent sections. Let $\zeta_\ell$ be a primitive $\ell$-th root of unity. Then $Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(\zeta_\ell)=-\ell$, $ Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(\zeta_{\ell^2})=\ell$ and $Tr_{\mathbb{Q}(_{2\ell^2})/\mathbb{Q}}(\zeta_{2\ell^2})=0$. Then using we compute for a positive integer $t$, $$\begin{aligned}
\label{2.12}
&Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(1,n)\zeta_{2\ell^2}^{-t})=\sum_{i=0}^{2\ell^{2}-1}B_{2\ell^2}(i,n)Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(\zeta_{2\ell^2}^{i-t})\nonumber\\
&= \ell(\ell-1)B_{2\ell^2}(t,n)+\sum_{\substack{i=0\\ i\neq t}}^{2\ell^{2}-1}B_{2\ell^2}(i,n)Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(\zeta_{2\ell^2}^{i-t})\nonumber\\
& = \ell(\ell-1)B_{2\ell^2}(t,n)-\ell\sum_{u=1}^{2l-1}B_{2\ell^2}(\ell u+t,n)+\ell\sum_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+t,n).\end{aligned}$$ Similary using , we obtain $$\begin{aligned}
\label{2.13}
Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(2,n)\zeta_{\ell^2}^{-t})&=\sum_{i=0}^{\ell^{2}-1}B_{\ell^2}(i,n)Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(\zeta_{\ell^2}^{i-t})\nonumber\\
&=\ell(\ell-1)B_{\ell^2}(t,n)-l\sum_{u=1}^{l-1}B_{\ell^2}(lu+t,n).\end{aligned}$$ Recently, we [@AT2018] determined the coefficients of Jacobi sums of order $2\ell^{2}$ in terms of Dickson-Hurwitz sums of the same order. More precisely, we proved:
\[AT1\] Let $\ell$ be an odd prime. Let $p$ be a prime such that $q=p^{r}\equiv 1 \pmod {2\ell^{2}}$ with an integer $r\geq 1$. Then for any positive integer $n$, $$J_{2\ell^{2}}(1,n)=\sum_{i=0}^{\ell(\ell-1)-1} d_{i,n}\zeta_{2\ell^2}^{i},$$ where the coefficents $d_{i, n}$ are given by $$d_{i,n}=B_{2\ell^{2}}(i,n)\mp B_{2\ell^{2}}(\ell(\ell-1)+j,n)-B_{2\ell^{2}}(\ell^{2}+k,n)\pm B_{2\ell^{2}}((2\phi(\ell^{2})+\ell)+j,n),$$ satisfying $ 0\leq j \leq \ell-1$, $j\equiv i \pmod { 2\ell^{2}}$, $k\equiv i \pmod{2\ell^{2}}$, $0\leq i, k \leq \phi(2\ell^{2})-1$.
The next result gives a relation between the coeffients $d_{i,n}$ and Dickson-Hurwitz sums of order $2\ell^2$. This result is useful to determine the cyclotomic numbers of order $2\ell^2$. Here, we prove:
\[thm1\] Let $\ell$ be an odd prime, and $t$ and $n$ be two positive integers modulo $2\ell^{2}$. Then $$\begin{aligned}
& \ell(\ell-1)B_{2\ell^2}(t,n)-\ell\sum\limits_{u=1}^{2\ell-1}B_{2\ell^2}(\ell u+t,n)+\ell\sum\limits_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+t,n)\\
&= \ell(\ell-1)d_{t,n}-\ell\sum\limits_{u=\ell-1}^{2\ell-3}d_{u\ell+t,n}-\ell(q-2),\end{aligned}$$ where $t=j\ell+s$ with $0 \leq j \leq 2\ell-1$ and $0 \leq s \leq \ell-1$.
Let $$D(n)=\ell(\ell-1)B_{2\ell^2}(t,n)-\ell\sum\limits_{u=1}^{2\ell-1}B_{2\ell^2}(\ell u+t,n)+\ell\sum\limits_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+t,n).$$ Then putting $t=j\ell+s$ with $0 \leq j \leq 2\ell-1$ and $0 \leq s \leq \ell-1$, we obtain $$D(n)=\ell(\ell-1)B_{2\ell^2}(j\ell+s,n)-\ell\sum\limits_{u=1}^{2\ell-1}B_{2\ell^2}(\ell(u+j)+s,n)+\ell\sum\limits_{u=1}^{\ell-1}B_{2\ell^2}(\ell(2u+j)+s,n).$$ This can be written as $$\begin{aligned}
D(n)&=\ell(\ell -1)B_{2\ell^2}(j\ell +s,n)-\ell \sum\limits_{u=1}^{2\ell -j-3}B_{2\ell^2}(\ell (u+j)+s,n) \\
&- \ell \sum\limits_{u=2\ell -j-1}^{2\ell -1}B_{2\ell^2}(\ell (u+j)+s,n)-\ell B_{2\ell^2}((\ell (2\ell-j-2+j))+s,n)\\
& +\ell \sum\limits_{u=1}^{\ell-1}B_{2\ell^2}(\ell (2u+j)+s,n).
\end{aligned}$$ Since $\ell B_{2\ell^2}((\ell (2\ell-j-2+j))+s,n)=B_{2\ell^2}(\ell^{2}+k,n))$, so that $$\begin{aligned}
D(n)&=\ell(\ell-1)(B_{2\ell^2}(j\ell+s,n)-B_{2\ell^2}(\ell^{2}+k,n))+\ell\sum\limits_{u=1}^{\ell-1}B_{2\ell^2}(\ell(2u+j)+s,n)\\
& -\ell \bigg(\sum\limits_{u=1}^{2\ell-j-3}B_{2\ell^2}(\ell(u+j)+s,n)-B_{2\ell^2}(\ell^{2}+k,n)\bigg)-\ell \bigg(\sum\limits_{u=2\ell-j-1}^{2\ell-1}B_{2\ell^2}(\ell(u+j)\\
& +s,n)-B_{2\ell^2}(\ell^{2}+k,n)\bigg)-\ell^{2}B_{2\ell^2}(\ell^{2}+k,n).
\end{aligned}$$ Applying , we obtain $$\begin{aligned}
D(n)&=\ell(\ell-1)(B_{2\ell^2}(j\ell+s,n)\mp B_{2\ell^2}(\ell(\ell-1)+s,n)\pm B_{2\ell^2}(2\ell^{2}-\ell+s,n)\\
& -B_{2\ell^2}(\ell^{2}+k,n)) +l\sum\limits_{u=1}^{\ell-1}(B_{2\ell^2}(\ell(2u+j)+s,n)\mp B_{2\ell^2}(\ell(\ell-1)+s,n)\\
& \pm B_{2\ell^2}(2\ell^{2}-\ell+s,n)-B_{2\ell^2}(\ell^{2}+k,n)) -\ell \bigg(\sum\limits_{u=1}^{2\ell-j-3}B_{2\ell^2}(\ell(u+j)+s,n)\\
& \mp B_{2\ell^2}(\ell(\ell-1)+s,n)\pm B_{2\ell^2}(2\ell^{2}-\ell+s,n)-B_{2\ell^2}(\ell^{2}+k,n)\bigg)\\
&-\ell \bigg(\sum\limits_{u=2\ell-j-1}^{2\ell-1}B_{2\ell^2}(\ell(u+j)+s,n)\mp B_{2\ell^2}(\ell(\ell-1)+s,n)\pm B_{2\ell^2}(2\ell^{2}-\ell+s,n)\\
&-B_{2\ell^2}(\ell^{2}+k,n)\bigg)-\ell(q-2).\end{aligned}$$ We now applying Theorem \[AT1\] and to get, $$\begin{aligned}
D(n)&=\ell(\ell-1)d_{j\ell+s,n}+\ell\sum_{u=1}^{\ell-1}d_{\ell(2u+j)+s,n}-\ell
\sum_{u=1}^{2\ell-j-3}d_{\ell(u+j)+s,n}\\
& -\ell\sum_{u=2\ell-j-1}^{2\ell-1}d_{\ell (u+j)+s,n}-\ell(q-2). \end{aligned}$$
Puting $u+j=x$ in second sum and $u+j\equiv x \ \pmod{ (2\ell-1)}$ in third sum, we have $$D(n)=\ell(\ell-1)d_{j\ell+s,n}+\ell\sum_{u=1}^{\ell-1}d_{\ell(2u+j)+s,n}-\ell
\sum_{x=0}^{2\ell-3}d_{\ell x+s,n}-\ell(q-2).$$
Consider $u\ell+t\equiv \ell x+s \pmod{ \ell(\ell-1)}$. Then $$D(n)=\ell(\ell-1)d_{t,n}+\ell\sum_{u=1}^{\ell-1}d_{\ell(2u+j)+s,n}-\ell\sum_{u=0}^{2l-3}d_{ul+t,n}-\ell(q-2).$$ Puting $2u+j\equiv m+j \pmod{(2\ell-1)}$ to obtain $$D(n)=\ell(\ell-1)d_{t,n}+\ell\sum_{u=0}^{\ell-2}d_{u\ell+t,n}-\ell\sum_{u=0}^{2\ell-3}d_{ul+t,n}-\ell(q-2).$$ Since for every $m+j, u\ell+t\equiv (m+j)\ell+s \pmod{\ell(\ell-1)}$ with $u\in \{0,1,2,3,\cdots, \ell-2\}$, we obtain $$D(n)=\ell (\ell -1)d_{t,n}+\ell \sum_{u=0}^{\ell-2}d_{u\ell +t,n}-\ell \sum_{u=0}^{2\ell -3}d_{u\ell +t,n}-
\ell (q-2).$$ This further implies $$D(n)=\ell(\ell-1)d_{t,n}-\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+t,n}-\ell(q-2).$$
Main Results
============
In this section, we obtain an explicit expression for cyclotomic numbers of order $2\ell^{2}$ in terms of the coefficients of Jacobi sums of orders $\ell$, $2\ell$, $\ell^{2}$ and $2\ell^{2}$. These coefficients of Jacobi sums can be evaluated in terms of Dickson-Hurwitz sums of orders $\ell$, $2\ell$, $\ell^{2}$ and $2\ell^{2}$. Here, we mainly prove:
\[thm2\] Let $\ell$ be an odd prime. Let $p$ be a prime such that $q=p^r\equiv 1 \pmod {2\ell^{2}}$ with an integer $r\geq 1$. Then $$\begin{aligned}
&4\ell^4 (a,b)_{2l^{2}}= \ell^{4}(a,b)_{\ell^{2}}+4\ell^{2}(a,b)_{2\ell}-\ell^{2}(a,b)_{\ell}-\ell(q-2)(4\ell^{2}-3)+\sum_{i=2}^{2\ell^{2}-1}\{\ell(\ell-1)\\
&\times d_{ia+b,n}-\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+ia+b,n}\} +\sum_{j=1}^{2\ell^{2}-1}\{\ell(\ell-1)d_{a+jb,n}-\ell\sum_{u=\ell-1}^{2\ell-3}
d_{u\ell+a+jb,n}\}\\
&-\sum_{i\text{ odd}}^{2\ell-1}\{\varepsilon((\ell^{2}+1)/2)(\ell ia+2b))b_{((\ell^{2}+1)/2)(\ell ia+2b),n} -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(\ell ia+2b),n}\}\\
& -\sum_{j\text{ odd}}^{2\ell-1}\{\varepsilon((\ell^{2}+1)/2)(2a+\ell jb))b_{((\ell^{2}+1)/2)(2a+\ell jb),n} -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(2a+\ell jb),n}\},\end{aligned}$$ where $\varepsilon(t)$ is given by $$\label{3.1}
\varepsilon(t)=
\begin{cases}
\ell^2\hspace*{4mm} \text{if } 0\leq j\leq \ell-2, \text{ i.e. } 0\leq t\leq \ell^2-\ell,\\
-\ell \hspace*{3mm} \text{if } j=\ell-1, \text{i.e. } \ell^2-\ell\leq t\leq \ell^2-1.
\end{cases}$$
To prove this theorem, we need an important result of D. Shirolkar and S. A. Katre [@SK2011 Lemma 6.1]. We recall that results for convenience. We consider the Jacobi sum $J_{\ell^2}(1,n)$ of order $\ell^2$, $$J_{\ell^2}(1,n)=\sum_{i=0}^{\ell(\ell-1)}b_{i,n}\zeta_{\ell^2}^i.$$
\[lm1\] For two positive integers $t$ and $n$ modulo $\ell^2$, define $$C(t,n):= \ell(\ell-1)B_{\ell^2}(t,n)-\ell B_{\ell^2}(u\ell+t, n).$$ Let $0\leq t\leq \ell^2-1$. Write $t=j\ell+s$, where $0\leq j\leq \ell^2-1$ and $0\leq s\leq \ell^2-1$. Then $$C(t,n)=\varepsilon(t)b_{t,n}-\ell\sum_{\ell=0}^{\ell-2}b_{u\ell+t, n},$$ where $\varepsilon(t)$ is given by .
Proof of Theorem \[thm2\] {#proof-of-theorem-thm2 .unnumbered}
-------------------------
Applying , we obtain $$4\ell^{4}(a,b)_{2\ell^{2}} = \sum_{i}\sum_{j}J_{2\ell^{2}}(i,j) \zeta_{2\ell^2}^{-(ai+bj)}.$$ Let us consider $G=\text{Gal}(\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q})$ and $G'=\text{Gal}(\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q})$. Then the above equation can be written as
$$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & =(a,b)_{\ell^{2}}+\sum_{i \text{ odd }}^{2\ell^{2}-1}J_{2\ell^{2}}(i,0)\zeta_{2\ell^2}^{-ia}+\sum_{j \text{ odd}}^{2\ell^{2}-1}J_{2\ell^{2}}(0,j)\zeta_{2\ell^2}^{-jb} \\ &+\sum_{i=1}^{2\ell^{2}-1}\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(i,1)\zeta_{2\ell^2}^{-(ia+b)})+\sum_{j=1}^{2\ell^{2}-1}\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(1,j)\zeta_{2\ell^2}^{-(a+jb)}) \\ &-\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(1,1)\zeta_{2\ell^2}^{-(a+b)})+\sum_{i\text{ odd}}^{2\ell-1}\sum_{\sigma \in G'}\sigma (J_{2\ell^{2}}(\ell i,2)\zeta_{2\ell^2}^{-(\ell ia+2b)})\\ &+\sum_{j\text{ odd}}^{2\ell-1}\sum_{\sigma \in G'}\sigma (J_{2\ell^{2}}(2,\ell j)\zeta_{2\ell^2}^{-(2a+\ell jb)})+\sum_{i=1}^{2\ell-1}\sum_{j=1}^{2\ell-1}J_{2\ell^{2}}(\ell i,\ell j)\zeta_{2\ell^2}^{-(\ell ia+\ell jb)}\\ &-\sum_{i=1}^{\ell-1}\sum_{j=1}^{\ell-1}J_{2\ell^{2}}(2\ell i,2\ell j)\zeta_{2\ell^2}^{-(2\ell ia+2\ell jb)}.\end{aligned}$$
Since the second and third terms in R H S are zero, so that $$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & = \ell^{4}(a,b)_{\ell^{2}}+\sum_{i=1}^{2\ell^{2}-1}\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(i,1)\zeta_{2\ell^2}^{-(ia+b)})+\sum_{j=1}^{2\ell^{2}-1}\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(1,j)\zeta_{2\ell^2}^{-(a+jb)})\\
&-\sum_{\sigma \in G}\sigma (J_{2\ell^{2}}(1,1)\zeta_{2\ell^2}^{-(a+b)})-\sum_{i\text{ odd}}^{2\ell-1}\sum_{\sigma \in G'}\sigma (J_{2\ell^{2}}(\ell i,2)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(\ell ia+2b)})\\ & -\sum_{j\text{ odd}}^{2\ell-1}\sum_{\sigma \in G'}\sigma (J_{2\ell^{2}}(2,\ell j)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(2a+\ell jb)}) +\sum_{i=1}^{2\ell-1}\sum_{j=1}^{2\ell-1}J_{2\ell^{2}}(\ell i,\ell j)\zeta_{2\ell^2}^{-(\ell ia+\ell jb)}\\ & -\sum_{i=1}^{\ell-1}\sum_{j=1}^{\ell-1}J_{2\ell^{2}}(2\ell i,2\ell j)\zeta_{2\ell^2}^{-(2\ell ia+2\ell jb)}.\end{aligned}$$ This implies, $$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & = \ell^{4}(a,b)_{\ell^{2}}+\sum_{i=1}^{2\ell^{2}-1}Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(i,1)\zeta_{2\ell^2}^{-(ia+b)})
\\ & +\sum_{j=1}^{2\ell^{2}-1}Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(1,j)\zeta_{2\ell^2}^{-(a+jb)})-Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(1,1)\zeta_{2\ell^2}^{-(a+b)}) \\ & -\sum_{i\text{ odd}}^{2\ell-1}Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(\ell i,2)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(\ell ia+2b)})\\ &- \sum_{j\text{ odd}}^{2\ell-1}Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(2,\ell j)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(2a+\ell jb)}) \\ & +\sum_{i=1}^{2\ell-1}\sum_{j=1}^{2\ell-1}J_{2\ell^{2}}(\ell i,\ell j)\zeta_{2\ell^2}^{-(\ell ia+\ell jb)} -\sum_{i=1}^{\ell-1}\sum_{j=1}^{\ell-1}J_{2\ell^{2}}(2\ell i,2\ell j)\zeta_{2\ell^2}^{-(2\ell ia+2\ell jb)}.\end{aligned}$$ Applying , we get $$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & = \ell^{4}(a,b)_{\ell^{2}}+\sum_{i=1}^{2\ell^{2}-1}Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(i,1)\zeta_{2\ell^2}^{-(ia+b)})\nonumber
\\ & +\sum_{j=1}^{2\ell^{2}-1}Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(1,j)\zeta_{2\ell^2}^{-(a+jb)})-Tr_{\mathbb{Q}(\zeta_{2\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(1,1)\zeta_{2\ell^2}^{-(a+b)})\nonumber \\ & -\sum_{i\text{ odd}}^{2\ell-1}Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(\ell i,2)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(\ell ia+2b)})\nonumber \\ &- \sum_{j\text{ odd}}^{2\ell-1}Tr_{\mathbb{Q}(\zeta_{\ell^2})/\mathbb{Q}}(J_{2\ell^{2}}(2,\ell j)\zeta_{\ell^2}^{-((\ell^{2}+1)/2)(2a+\ell jb)})\nonumber \\ & +
4\ell^2(a, b)_{2\ell}-\ell^2(a,b)_{\ell}.\end{aligned}$$ By the use of and , we get $$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & = \ell^{4}(a,b)_{\ell^{2}}+\sum_{i=1}^{2\ell^{2}-1}\{\ell(\ell-1)B_{2\ell^2}(ia+b,n)-\ell\sum_{u=1}^{2\ell-1}B_{2\ell^2}(\ell u+ia+b,n)\\ & +\ell\sum_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+ia+b,n)\} +\sum_{j=1}^{2\ell^{2}-1}\{\ell(\ell-1)B_{2\ell^2}(a+jb,n)\\ & -\ell\sum_{u=1}^{2\ell-1}B_{2\ell^2}(\ell u+a+jb,n)+\ell\sum_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+a+jb,n)\} \\ & - \ell(\ell-1)B_{2\ell^2}(a+b,n)+\ell\sum_{u=1}^{2\ell-1}B_{2\ell^2}(\ell u+a+b,n)\\ & -\ell\sum_{u=1}^{\ell-1}B_{2\ell^2}(2\ell u+a+b,n) -\sum_{i\text{ odd}}^{2\ell-1}\{\ell(\ell-1)B_{\ell^2}(((\ell^{2}+1)/2)(\ell ia+2b),n)\\ & -\ell\sum_{u=1}^{\ell-1}B_{\ell^2}(u\ell+((\ell^{2}+1)/2)(\ell ia+2b),n)\} -\sum_{j\text{ odd}}^{2\ell-1}\{\ell(\ell-1)
B_{\ell^2}(((\ell^{2}+1)/2)\\
&\times (2a+\ell jb),n)-\ell\sum_{u=1}^{\ell-1}B_{\ell^2}(u\ell+((\ell^{2}+1)/2)(2a+\ell jb),n)\} +4\ell^{2}(a,b)_{2\ell}\\ &
-\ell^{2}(a,b)_{\ell}. \end{aligned}$$ Employing Theorem \[thm1\] and Lemma \[lm1\], we obtain $$\begin{aligned}
& 4\ell^{4}(a,b)_{2\ell^{2}} = \ell^{4}(a,b)_{\ell^{2}}+4\ell^{2}(a,b)_{2\ell}-\ell^{2}(a,b)_{\ell}+\sum_{i=2}^{2\ell^{2}-1}\{\ell(\ell-1)d_{ia+b,n}-\ell(q-2)\\
&
-\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+ia+b,n}\} +\sum_{j=1}^{2\ell^{2}-1}\{\ell(\ell-1)d_{a+jb,n} -\ell(q-2)-\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+a+jb,n}\}\\
& -\sum_{i\text{ odd}}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(\ell ia+2b))b_{((\ell^{2}+1)/2)(\ell ia+2b),n} -\ell \sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(\ell ia+2b),n}\} \\ &
-\sum_{j\text{ odd}}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(2a+\ell jb))b_{((\ell^{2}+1)/2)(2a+\ell jb),n} -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(2a+\ell jb),n}\}.\end{aligned}$$ This can be further simplified as $$\begin{aligned}
4\ell^{4}(a,b)_{2\ell^{2}} & = \ell^{4}(a,b)_{\ell^{2}}+4\ell^{2}(a,b)_{2\ell}-\ell^{2}(a,b)_{\ell}-\ell(q-2)(4\ell^{2}-3)\\
& +\sum_{i=2}^{2\ell^{2}-1}\{\ell(\ell-1)d_{ia+b,n}-\ell\sum_{u=\ell-1}^{2\ell-3}
d_{u\ell+ia+b,n}\} +\sum_{j=1}^{2\ell^{2}-1}\{\ell(\ell-1)d_{a+jb,n}\\ & -\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+a+jb,n}\}-\sum_{i\text{ odd}}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(\ell ia+2b))b_{((\ell^{2}+1)/2)(\ell ia+2b),n}\\ & -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(\ell ia+2b),n}\}-\sum_{j\text{ odd}}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(2a+\ell jb))\\ & \times b_{((\ell^{2}+1)/2)(2a+\ell jb),n} -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(2a+\ell jb),n}\}.\end{aligned}$$ This completes the proof.
Substituting the well known formulae for cyclotomic numbers of order $\ell$ and $2\ell$ in Theorem \[thm2\], we obtain the following:
Let $p, q$ and $\ell$ as in Theorem \[thm2\] with $r=1$. Then $$\begin{aligned}
&4\ell^{4}(a,b)_{2\ell^2} = \ell^{4}(a,b)_{\ell^{2}}-\ell(q-2)(4\ell^{2}-3)-\{(-1)^{b}+(-1)^{a+k}+(-1)^{a+b}\}\\
& \times \{\ell+\sum_{m=1}^{\ell-1}b_{m}(\ell) +\sum_{u=0}^{\ell-2}\sum_{m=1}^{\ell-1}b_{m}(2u+1)\}+(-1)^{b}\ell\{b_{\nu(-a)}(\ell)\\
& +\sum_{u=0}^{\ell-1}b_{\nu(b-2au-2a)}(2u+1)\}+(-1)^{a+b}\ell\{b_{\nu(b)}(\ell) +\sum_{u=0}^{\ell-1}b_{\nu(a+2bu+b)}(2u+1)\}\\
& +(-1)^{a+k}\ell\{b_{\nu(-b)}(\ell) +\sum_{u=0}^{\ell-1}b_{\nu(a-2bu-2b)}(2u+1)\}+\sum_{i=2}^{2\ell^{2}-1}\{\ell(\ell-1)d_{ia+b,n}\\ & -\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+ia+b,n}\} +\sum_{j=1}^{2\ell^{2}-1}\{\ell(\ell-1)d_{a+jb,n}-\ell\sum_{u=\ell-1}^{2\ell-3}d_{u\ell+a+jb,n}\}\\ & -\sum_{i\text{ odd }}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(\ell ia+2b))b_{((\ell^{2}+1)/2)(\ell ia+2b),n} -\ell \sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(\ell ia+2b),n}\}\\ & -\sum_{j\text{ odd}}^{2\ell-1}\{\varepsilon(((\ell^{2}+1)/2)(2a+\ell jb))b_{((\ell^{2}+1)/2)(2a+\ell jb),n} -\ell\sum_{u=0}^{\ell-2}b_{u\ell+((\ell^{2}+1)/2)(2a+\ell jb),n}\},\end{aligned}$$ where $b_{0}(n)=0$,\
$\nu(b)= \begin{cases}
\varLambda(b)/2 \hspace*{8mm} \text{if } b \text{ is even};\\
\varLambda(b+\ell)/2 \hspace*{2mm} \text{if } b \text{ is odd},
\end{cases}$\
and $\varLambda(r)$ is defined as the least non-negative residue of $r$ module $2\ell$.
For any odd prime $p\geq 5$, $p^2\equiv 1 \pmod 3$ holds. Thus to calculate cyclotomic numbers of order $2\ell^{2}$ with prime $\ell\geq 5$, it is sufficient to calculate $2\ell^2+(2\ell^2-3)+(2\ell^2-6)+(2\ell^2-9)+\cdots+2$ distinct cyclotomic numbers of order $2\ell^2$. However $\ell=3$, it is enough to calculate $2\ell^2+(2\ell^2-3)+(2\ell^2-6)+(2\ell^2-9)+\cdots +1$ distinct cyclotomic numbers of order $2\ell^2$.
Matrix assocated to cyclotomic numbers
======================================
In this section, we illustrate two types of cyclotomic numbers. We first consider $q=p=19\equiv 1 \pmod{18}$ and $q=p=37\equiv 1 \pmod {18}$. Then using property of cyclotomic numbers the $324$ pairs of two paremeters numbers $(a,b)_{18}$ can be reduced to $64$ distinct pairs (see Table \[table1\] and Table \[table2\]). On evaluating cyclotomic numbers corresponding to $64$ distinct pair, we obtain the complete tables in a form of matrices $A$ and $B$. Primary interest is to know about the determinant, eigen values, characteristic polynomial and minimal polynomial of $A$ and $B$. It is also important to know the nature of the matrices obtained by changing the generator of $\mathbb{F}_q^{\times}$. Since the entries of exactly one row of $A$ are zero, thus $det(A)=0$. We observe that the characteristic as well as minimal polynomial of $A$ is $m_A(x)=x^{18}$. We also see that all the eigenvalues of $A$ are equal, and in fact they all are zero. Using GP/PARI (version 2.9.2), we obtain $det(B)=-1$ and the characteristic as well as minimal polynomial is $$\begin{aligned}
m_B(x)&=x^{18}-x^{17}-17x^{16}+16x^{15}+120x^{14}-105x^{13}-455x^{12}+364x^{11}\\ & +1001x^{10}-715x^{9}-1287x^{8}+792x^{7}+924x^{6}-462x^{5}-330x^{4}
+120x^{3}+45x^{2}-9x-1.\end{aligned}$$ The eigenvalues of $B$ are $\lambda_{1}=-1.9712,\ \lambda_{2}=1.9928,\ \lambda_{3}=1.9355,\ \lambda_{4}=-1.8858,\ \lambda_{5}=1.8225,\ \lambda_{6}=-1.7460,\ \lambda_{7}=1.6570,\ \lambda_{8}=-1.5561,\ \lambda_{9}=1.4439,\ \lambda_{10}=-1.3213,\ \lambda_{11}=1.1893, \lambda_{12}=-1.0486,\ \lambda_{13}=0.9004,\ \lambda_{14}=-0.7457, \lambda_{15}=0.5856,\ \lambda_{16}=-0.4214,\ \lambda_{17}=0.2540,\ \lambda_{18}=-0.0849$, and all of them are distinct. It is noted that if we change the generator of $\mathbb{F}_{q}^{\times}$, then entries of $A$ (resp. $B$) get interchange among themselves but their nature remain as the same.
The matrix corresponding of Table \[table1\] is given by
$A$ =$\left[
\begin{array}{rrrrrrrrrrrrrrrrrr}
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 & 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 & 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 & 1 & 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 & 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 & 1 & 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 & 1 & 0 & 0
\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0
\\
0 & 0 & 0 & 0 & 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 & 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 \\
\end{array}\right]
$
It is noted if $k=1$, then the corresponding matrix is always singular. The Jordon-canonical form of $B$ is given by the diogonal matrix $D$ with diagonal entries\
$\displaystyle{ \lambda_{1},\lambda_{2},\lambda_{3},\lambda_{4},\lambda_{5},\lambda_{6},\lambda_{7},\lambda_{8}, \lambda_{9},\lambda_{10}, \lambda_{11},
\lambda_{12}, \lambda_{13},\lambda_{14},\lambda_{15},\lambda_{16},\lambda_{17},\lambda_{18}}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
M H Ahmed and J Tanti acknowledge Central University of Jharkhand, Ranchi, Jharkhand for providing necessary and excellent facilities to carry out this research. A Hoque acknowledges SERB-NPDF (PDF/2017/001958), Govt. of India for financial support.
[25]{}
V. V. Acharya and S. A. Katre, [Cyclotomic numbers of orders $2\ell, \ell$ an odd prime]{}, Acta Arith. **69** (1995), no. 1, 51–74.
M. H. Ahmed and J. Tanti, [*Jacobi sums of order $2\ell^2$*]{}, 2018, <arXiv:1807.06218>.
L. E. Dickson, [*Cyclotomy, higher congruences, and Waring’s problem*]{}, Amer. J. Math. **57** (1935), 391–424.
L. E. Dickson, [*Cyclotomy and trinomial congruences*]{}, Trans. Amer. Soc. **37** (1935), 363–380.
L. E. Dickson, [*Cyclotomy when $e$ is composite*]{}, Trans. Amer. Math. Soc. **38** (1935), 187–200.
R. J. Evans and J. R. Hill, [*The cyclotomic numbers of order sixteen*]{}, Math. Comp. **33** (1979), 827–835.
C. Friesen, J. B. Muskat, B. K. Spearman and K. S. Williams, [*Cyclotomy of order $15$ over $GF(p^{2})$, $p\equiv 4, 11 \pmod {15}$*]{}, Int. J. Math. Math. Sci. **9** (1986), 665–704.
S. A. Katre and A. R. Rajwade, [*Complete solution of the cyclotomic problem in $\mathbb{F}^{*}_q$ for any prime modulus $\ell$*]{}, $q=p^{\alpha}$, $p\equiv 1 \pmod {\ell}$, Acta Arith. **45** (1985), 183–199.
E. Lehmer, [*On the cyclotomic numbers of order sixteen*]{}, Canad. J. Math. **6** (1954), 449–454.
J. B. Muskat and A. L. Whiteman, [*The cyclotomic numbers of order twenty*]{}, Acta Arith. **17** (1970), 185–216.
J. B. Muskat, [*On Jacobi sums of certain composite orders*]{}, Trans. Amer. Math. Soc. **134** (1968), 483–502.
J. B. Muskat, [*The cyclotomic numbers of order fourteen*]{}, Acta Arith. **11** (1966), 263–279.
D. Shirolkar and S. A. Katre, [*Jacobi sums and cyclotomic numbers of order $\ell^2$*]{}, Acta Arith. **147** (2011), 33–49.
T. Storer, [*On the unique determination of the cyclotomic numbers for Galois fields and Galois domains*]{}, J. Combinatorial Theory **2** (1967), 296–300.
A. L. Whiteman, [*The cyclotomic numbers of order twelve*]{}, Acta Arith. **6** (1960), 53–76.
A. L. Whiteman, [*The cyclotomic numbers of order sixteen*]{}, Trans. Amer. Math. Soc. **86** (1957), 401–413.
Y. C. Zee, [*Jacobi sums of order $22$*]{}, Proc. Amer. Math. Soc. **28** (1971), 25–31.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We observed the near infrared emission in the wavelength range 2.28–2.5$\mu$m from the peculiar B\[e\]-star MWC349. The spectra contain besides the strong IR continuum the first overtone CO bands and most of the hydrogen recombination lines of the Pfund series, both in emission. We also modeled the spectra. The Pfund lines have a gaussian profile with a FWHM of $\sim
100$km/s, and it turned out that their emission is in LTE and optically thin. To explain the CO bands, several scenarios were investigated. We found that the CO band heads are formed under LTE and that the gas must have a temperature of 3500 to 4000K. The width of the $2\rightarrow 0$ band head indicates kinematical broadening of 50 to 60km/s. We can obtain fits to the measured spectra assuming that the CO gas has a column density of $5\cdot
10^{20}$cm$^{-2}$ and is located either at the inner edge of the rotating circumstellar disk. In this case, the disk must have a bulge which partly blocks the radiation so that the observer sees only a sector on the far side where the radial velocities are small. Or the CO emission originates in a wind with gaussian line profiles. Both fits are of equal quality and satisfactory. In a third alternative where the fit is less convincing, the CO emission is optically thin and comes from an extended Keplerian disk.
author:
- 'M. Kraus'
- 'E. Krügel'
- 'C. Thum'
- 'T. Geballe'
date: 'Received; accepted '
subtitle: 'I. First overtone bands from a disk or from a wind?'
title: CO band emission from MWC349
---
Introduction
============
After the discovery of CO band head emission in the BN object by Scoville et al. ([@scoville]) a large number of further detections in other young stellar objects (YSO) followed (see e.g. Geballe & Persson [@geballe]; Carr [@carr]; Chandler et al. [@chandler]; Greene & Lada [@green]; Najita et al. [@najita]).
Several scenarios were discussed to explain the origin of the hot (2500–5000 K) and dense ($n > 10^{11}$cm$^{-3}$) CO component (for some examples see Calvet et al. [@calvet]; Martin [@martin]), but the most likely location is a neutral disk or wind (Carr [@carr]; Chandler, Carlstrom, & Scoville [@chandler95]). The disk model is especially supported by high resolution spectroscopic observations of the CO $2\rightarrow 0$ band head. The shape of this band head shows for several YSO the kinematic signature of Keplerian rotation (Carr et al. [@carr93]; Carr [@carr95]; Najita et al. [@najita]) and is a powerful tracer for the existence of a circumstellar disk around young stellar objects.
Also the peculiar B\[e\]-star shows the first overtone CO bands in emission. These bands were observed first by Geballe & Persson ([@geballe]) with a velocity resolution of about 460km/s.
MWC349 is a binary system consisting of the main component MWC349A, classified as a B\[e\]-star, and the B0 III star companion, MWC349B, localized $2\arcsec$ west of MWC349A. We are only interested in the main component, in the following refered to as MWC349. Its evolutionary state is still unclear. It shows some characteristics of a pre-main sequence B\[e\]-type star as well as characteristics of a B\[e\] supergiant (e.g. Lamers et al. [@lamers]).
Cohen et al. ([@cohen]) determined its distance (1.2kpc), bolometric luminosity ($\sim 3\cdot 10^{4}$L$_{\sun}$) and the visual extinction towards it ($A_{\rm V}^{\rm ISM}\simeq 10$mag of which 2mag might be circumstellar).
The existence of a bulge of circumstellar dust around MWC349 has been known since long ago (e.g. Geisel [@geisel]). The proposition that MWC349 also has a disk was supported by observations of double-peaked emission lines (Hamann & Simon [@hamann], [@hamann88]) and by IR speckle interferometry (Leinert [@leinert]; Mariotti et al. [@mariotti]) which revealed a disk-like structure of the dust emission in the east-west direction, seen nearly edge on. An accumulation of neutral gas and dust in the equatorial plane of the star is also expected to be responsible for the bipolar structure of the optically thick wind zone seen in the VLA-map of White & Becker ([@white]). The mass loss rate found for a 50km/s wind velocity is $\sim 1.2\cdot 10^{-5}$M$_{\sun}$yr$^{-1}$ (Cohen et al. [@cohen]).
Another indicator for the circumstellar disk are the strong hydrogen recombination maser lines in the mm and submm range which also show the characteristic double-peaked profiles (e.g. Martín-Pintado et al. [@martin-pintado]). Recombination lines at different wavelengths sample different regions. With decreasing quantum number $n$, i.e. increasing frequency, one sees ionized gas closer to the star. The fact that the rotational velocity of recombination lines with decreasing $n$ displays a systematical increase led to the assumption that the double-peaked maser emission comes from the ionized atmosphere of a Keplerian rotating disk around a 25–30M$_{\sun}$ star (Thum, Martín-Pintado & Bachiller [@thum]; Thum et al. [@thum94]) whose free-free and free-bound emission was modelled by Kraus et al. ([@kraus]). In addition, Rodríguez & Bastian ([@rodriguez]) determined the inclination angle of the disk towards the line of sight to $15^{\circ}\pm 5^{\circ}$.
In this paper, we present new low resolution observations of the total first overtone CO band emission as well as high resolution observations of the $2\rightarrow 0$ and $3\rightarrow 1$ band heads. We discuss the probable location of the hot CO gas by modeling the emission for several scenarios.
Observations {#obs}
============
In 1997 we observed MWC349 with the UKIRT telescope in the NIR range ($1.85 - 2.5\,\mu$m) with a spectral resolution of $\sim 330$km/s (Fig.\[observ\]a). The spectrum contains several prominent hydrogen and helium recombination lines as well as the first overtone CO bands, all of them in emission.
-- ---- -- --------- ----------------
1 1.86904
2 1.87561 Pa$\alpha$
3 1.89282 ,
4 1.90867
5 1.94508 Br$\delta$,
6 1.95445
7 1.97470 , ,
8 1.98702 ,
9 1.99318 , ,
10 2.00130
11 2.05917
12 2.08819 (?),
13 2.11194
14 2.13831 , ,
15 2.14359 , , \[\]
16 2.16611 Br$\gamma$
17 2.20687 doublet, , (?)
18 2.21741
19 2.24113 (?)
-- ---- -- --------- ----------------
: Identified lines in the NIR spectrum of MWC349 (Figs. a and b).
\[identi\]
In two sessions in 1998 we reobserved MWC349 in the spectral range 2.285–2.341$\mu$m with higher spectral resolution (10–15km/s). The observations consist of four seperate subspectra with a small overlap at their edges. They are slightly offset relative to each other and at their low wavelength end, the fluxes are systematically overestimated. We calibrated the subspectra by smoothing them to the resolution of the spectrum in Fig.\[observ\]c, then scaled each spectrum with a factor near $\sim 1$ and combined them to the final plot of Figure \[hires\].
Unfortunately, the $3\rightarrow 1$ band head falls into the overlap region of two subspectra, so its strength is rather uncertain (marked by the dashed line within the figure). This fact turned out during the modeling to be a disadvantage.
Results
=======
Hydrogen recombination lines – the Pfund series {#pfund}
-----------------------------------------------
As can be seen in Figs.\[observ\]c and \[hires\], the spectrum of MWC349 in the wavelength range from 2.25 to 2.5$\mu$m contains beside the first overtone CO bands also the Pfund series (transitions of the form $n \rightarrow n' = 5$, denoted as Pf($n$)) of the hydrogen atom. Hamann & Simon ([@hamann]) already identified some of them in their velocity-resolved infrared spectroscopy of MWC349, but their resolution of $\sim$20km/s was lower than ours, so they could only resolve lines up to Pf(34). Their FWHM values of several unsmeared lines were about 100km/s and agree with ours.
Many of the other emission lines they observed are double-peaked which is a hint for a rotating medium. At first glance we cannot exclude the Pfund lines to take part of the rotation because they are superimposed on the CO bands, but their profile looks more than a gaussian than a double-peaked one. A gaussian velocity component of about 50km/s, however, is not a surprising feature for MWC349. Also the cm and mid IR hydrogen recombination lines as well as the ’pedestal‘ feature of the double-peaked mm recombination maser lines show such a velocity component (Smith et al. [@smith]; Thum, Martín-Pintado & Bachiller [@thum]). It is often ascribed to the wind of MWC349.
To model the hydrogen emission, we assume the Pfund lines to be optically thin. Their intensity is given by $$\label{intens}
I_{\nu} = N_{n} A_{n,5} h\nu~\Phi_{\rm H}(\nu)~.$$ To evaluate it, we need the number density $N_{n}$ of the atoms in level $n$, the Einstein coefficients, $A_{n,5}$, and the profile function of the hydrogen gas, $\Phi_{\rm H}(\nu)$.
In view of the measured line widths, we choose a gaussian profile function defined as $$\label{gaussprofil}
\Phi_{\rm H} (\nu) = \frac{1}{\sqrt{\pi}\frac{\nu_{0}}{c}v}\,\exp\left[-\left(\frac{\nu -
\nu_{0}}{\frac{\nu_{0}}{c}v}\right)^{2}\right]$$ $v$ is the most probable velocity which is in the case of the thermal velocity given by $v=\sqrt{2kT/m}$ for particles having a Maxwellian velocity distribution. In our case, $v\simeq
$50km/s. We also take into account the local standard of rest velocity of $\sim$8km/s (Thum, Martín-Pintado & Bachiller [@thum]; Thum et al. [@thum95]).
The Einstein coefficients, $A_{nn'}$, can be calculated, following Menzel & Pekeris ([@menzel]). If we restrict ourselves to Pfund lines with $n\geq 18$ (see Fig.\[observ\]c), they can be approximated by
$$\label{a_n5}
A_{n,5} \simeq 3.023\cdot 10^{9}~n^{-5.005}~.$$
The number density, $N_{n}$, of the level $n$ follows from Saha’s equation $$\label{bes_n}
N_{n} = b_{n} N_{\rm i} N_{\rm e} g_{n}~\frac{h^{3}}{2\,(2\pi
m_{\rm e}kT_{\rm e})^{3/2}}~{\rm e}^{\frac{h\,Ry}{n^{2}kT_{\rm e}}}$$ where $N_{\rm i}$ and $N_{\rm e}$ are the number densities of the protons and electrons, $T_{\rm e}$ and $m_{\rm e}$ are the electron temperature and electron mass, $g_{n}=2n^{2}$ is the statistical weight of the state $n$, and $Ry$ is the Rydberg constant. The $b_{n}$-factors which describe the deviation from LTE and which depend on the electron temperature and density were taken from the electronic Tables of Storey & Hummer ([@storey]) for Menzel’s case B recombination theory.
Detailed modeling of the Pfund line spectrum performed by Kraus ([@kraus00]) showed that the emission arises within the innermost parts of the ionized wind where the electron densities are highest. Within these regions it is expected that $N_{\rm e}\leq 10^{9}$cm$^{-3}$. Electron densities as high as $10^{8}$cm$^{-3}$ have already been found by Strelnitski et al.([@strelnitski96]) and Thum et al.([@thum98]). For such high electron densities the emission is clearly in LTE, i.e. the $b_{n}$-factors are $\sim\,1$. Therefore, non LTE effects are negligible.
In our case of optically thin Pfund lines in LTE and with a constant electron temperature the intensity ratios of two neighbouring Pfund lines is according to Eq.(\[intens\]) independent of the electron density $$\frac{I_{{\rm Pf}(n+1)}}{I_{{\rm Pf}(n)}}\sim \left( \frac{n}{n+1}\right)^{5.005}
\cdot \exp\left(-\frac{hRy}{kT_{\rm e}}\frac{2n+1}{(n^{2}+n)^{2}}\right)~.$$ and always smaller than 1, which means that the intensity of the Pfund lines decreases with increasing quantum number. This behaviour can qualitatively be seen in Fig.\[pf\_syn1\] where we plotted the synthetic Pfund lines which we smoothed to the spectral resolution of 10–15km/s (solid line) and 330km/s (dashed line), respectively. The wavelength difference between the neighbouring Pfund lines decreases with increasing $n$ leading to a blend of the individual lines. The onset of this blend depends on the width of the lines and on the spectral resolution and leads to a hydrogen ’continuum‘.
$T_{\rm e}$ \[K\] $v_{\rm gauss}$ \[km/s\] $A^{\rm ISM}_{\rm v}$ \[mag\] $\int N_{\rm e}^{2}dV$ \[cm$^{-3}$\]
------------------- -------------------------- ------------------------------- --------------------------------------
$\leq 10^{4}$ $\sim 50$ 10 $6.35\cdot 10^{60}$
: Parameters for hydrogen used for the calculations of the Pfund lines shown in Fig.[\[pf\_subtract\]]{}(middle panel).
\[pf\_param\]
The interstellar visual extinction towards MWC349 is about $A_{\rm v}^{\rm ISM} \simeq
10$mag and has to be taken into account. Fitting the Pfund lines with the parameters displayed in Table \[pf\_param\] leads to the quantity $\int N_{\rm e}^{2}dV$, where $V$ is the volume of the emitting region.
The middle panel of Figure \[pf\_subtract\] contains the modelled Pfund series added to the continuum emission of MWC349 (dashed line) taken from Kraus et al. ([@kraus]). These two components were subtracted from the observations (upper panel) and the resulting spectrum (lower panel) can mainly be ascribed to the CO first overtone band emission. To fit this observed CO spectrum is the aim of the following sections.
Theory of the CO bands
----------------------
The energy of a diatomic molecule in rotational level $J$ and vibrational level $v$ can be expanded in the following way (Dunham [@dunham32a], [@dunham32b]) $$E(v,J) = hc~\sum_{k,l} Y_{k,l}\,\left( v+\frac{1}{2}\right)^{k}(J^{2}+J)^{l}~.$$ The parameters $Y_{k,l}$ are for the CO molecule taken from Farrenq et al. ([@farrenq]). The first overtone bands result from coupled vib-rot transitions in the ground electronic state and obey the selection rules $\Delta v = 2$ and $\Delta J =
\pm 1$.
In the following, we assume the CO gas to be in LTE. Then, the levels are populated following a Boltzmann distribution, $$\label{n_vib_rot}
N_{vJ} = \frac{N}{Z} (2J + 1)\,e^{\frac{-E(v,J)}{kT}}$$ where $N$ and $T$ are the total number density and temperature of CO molecules and $Z$ is the total partition function, given as the product of the vibrational and the rotational partition function
$$Z = Z_{v}\cdot Z_{J} = \sum_{v} e^{\frac{-E_{v}}{kT}} \cdot \sum_{J}
(2J + 1)\,e^{\frac{-E_{J}}{kT}}~.$$
To account for optical depth effects, we calculate the line intensities from the transfer equation $$\label{line_intens}
I_{\nu} = B_{\nu}(T)\,(1-e^{-\tau_{\nu}})\,.$$ The optical depth is $\tau_{\nu} = \int\kappa_{\nu}\,ds$ with the absorption coefficient per cm $$\label{kappa}
\kappa_{\nu} = \frac{c^{2}N_{vJ} A_{vJ;v'J'}}{8\pi \nu^{2}}\,\left( \frac{2j + 1}
{2j' + 1}\cdot\frac{N_{v'J'}}{N_{vJ}} - 1\right)\,\Phi_{\rm CO}(\nu)$$ where $\Phi_{\rm CO}(\nu)$ is the profile function of the CO gas. The Einstein coefficients, $A_{vJ;v'J'}$, are from Chandra, Maheshwari & Sharma ([@chandra]).
Let $\lambda_{\rm min}$ denote the onset of the ($2\rightarrow 0$) bandhead. We define it to be the wavelength where the intensity of the $(2,51)\rightarrow (0,50)$ line (see Fig.\[line\]) has dropped by a factor $x \sim 5$ from its maximum at $\lambda_0$; the exact number is not important. We determine $\lambda_{\rm min}$ from Figure \[mini\] to $(2.29285\pm 0.00005)\,\mu$m.
The value of $\lambda_{\rm min}$ depends of course on the various mechanisms of line broadening. Most processes lead to a gaussian or nearly gaussian profile, like thermal (turbulent) motion of mean velocity $v_{\rm th}$ ($v_{\rm turb}$), detector resolution $v_{\rm res}$, spherical wind $v_{\rm wind}$ (see e.g. Hamann & Simon [@hamann]). Their superposition results again in a gaussian line of width $v_{\rm gauss} = \sqrt{v_{\rm th}^{2} + v_{\rm turb}^{2} + v_{\rm wind}^{2} +
v_{\rm res}^{2}}$.
Disk rotation, on the other hand, produces a double-peaked profile. In a gas ring rotating at velocity $v_{\rm rot}$ a line with laboratory frequency $\nu_0$ is shifted up to a maximum frequency $\nu_0\cdot(1+v_{\rm rot}/c)$. In a mixture of disk rotation and gaussian broadening one finds the following expression for the minimum wavelength, $$\label{l_min}
\frac{1}{\lambda_{\rm min}} = \frac{1}{\lambda_{0}}\left[ 1+\frac{v_{\rm lsr} +
v_{\rm rot}}{c}\right] \left[\frac{v_{\rm gauss}}{c}\,\sqrt{\ln x} + 1\right]\,,$$ where $v_{\rm lsr} = 8$km/s is the velocity shift of MWC349 with respect to the local standard of rest. This relation only holds if the emission in the rise of the spectrum is optically thin. The thermal velocity $v_{\rm th}$ of CO molecules at 3500K is less than $\sim$1km/s and negligible in comparison with the spectral resolution $v_{\rm res} = 10\ldots 15$km/s. The same can be said about the turbulent velocity which is not likely to exceed the speed of sound.
Eq.(\[l\_min\]) can now be used to relate $v_{\rm wind}$ to the rotational velocity $v_{\rm rot}$. The result is shown in Figure \[v\_turb\]. The dashed lines above and below the solid line are for the lower and upper limit of $\lambda_{\rm min}$. If the CO gas performs no rotation, the gas must emanate from the disk in a wind with about $\sim$50km/s, the same velocity as for the Pfund lines. Without a wind, one derives a rotational velocity of $\sim 58\pm
7$km/s. This value is larger than other rotational velocities around MWC349 observed up to now, the highest being 42.5km/s for (Hamann & Simon [@hamann]).
Modeling the CO band emission {#model}
=============================
A gaussian line profile for the CO gas {#turb_mot}
--------------------------------------
First, we neglect rotation and assume a pure gaussian line profile. The velocity of the CO gas is taken from Figure\[v\_turb\] as $v_{\rm gauss} \approx 50$km/s.
For the calculation of the optical depth we assume the absorption coefficient to be constant along the line of sight. Then $\tau_{\nu}$ is simply given as the product of the absorption coefficient per CO molecule, $\kappa_{\nu}$, times the CO column density, $\tau_{\nu} = \kappa_{\nu} \cdot N_{\rm CO}$.
In Fig.\[kappa\_turb\] we plotted $\kappa_{\nu}$ for temperatures between 3000K and 5000K, and a gaussian velocity of the gas of 50km/s. In the wavelength range around the $3\rightarrow 1$ band head (2.322–2.325$\mu$m) the absorption coefficient is higher than in the region around the $2\rightarrow 0$ band head (2.293–2.295$\mu$m). This effect becomes stronger with increasing temperature. Thus, the turnover from optically thin to optically thick emission depends not only on the column density but also on the temperature: the $3\rightarrow 1$ band head becomes optically thick at lower temperature and lower column density than the $2
\rightarrow 0$ band head.
Next, we vary the CO column density and fix $T_{\rm CO}$ at 4000K. The synthetic spectra in Figure \[turb\_syn\] are calculated for the case of extreme optical depth (upper line) and for CO column densities (from top to bottom) of $2\cdot
10^{21}, 10^{21}, 5\cdot 10^{20}$, and $10^{20}$cm$^{-2}$. With increasing column density the optical depth enlarges and the intensity tends towards its limiting blackbody value. In addition, the spectra show two characteristic features: a varying intensity ratio of the two band heads and a broadening of the band head structure.
The first property can better be seen in Figure \[verhaelt\] where we plotted the flux density ratio of the $3\rightarrow 1$ and $2\rightarrow 0$ band head with temperature for different column densities. This ratio decreases with increasing column density, i.e. increasing $\tau$, and approaches 0.96 for $\tau\rightarrow\infty$.
The determination of the column density with only the help of this ratio is, however, not definite because the flux density ratio also depends on temperature. From our observations we get no information at all on the column density in this way because the observed strength of the $3\rightarrow 1$ band head is subject to some uncertainty as discussed in Section \[obs\].
The second characteristic feature is the broadening of the band head structure with increasing column density, especially at the long wavelength side. In Figure \[turb\_tief\] the observed $2\rightarrow 0$ band head region (histogram) is overlaid on the theoretical spectra with column densities of $2\cdot 10^{21}$ (dashed line), $5\cdot 10^{20}$ (solid line), and $10^{20}$cm$^{-2}$ (optically thin case, dotted line).
At first glance, it seems that the observations might also be fitted by optically thin emission, but the upper panel of Figure \[thick\_thin\] which includes a somewhat larger wavelength interval is clearly inconsistent with the optically thin case whereas the observations can be reconciled with a synthetic spectrum of column density $N_{\rm CO}
\simeq 5\cdot 10^{20}$cm$^{-2}$ (lower panel of Figure \[thick\_thin\]). An additional disagreement is that the flux density ratio of the bandheads is highest for the optical thin case which can not be confirmed by the observations.
So we take $N_{\rm CO} = 5\cdot 10^{20}$cm$^{-2}$ as the best fit CO column density. The synthetic spectra of the $2\rightarrow 0$ band head look very similar for temperatures in the range of $3000 - 5000$K. For a more sensitive limitation of the temperature range we must take a look at the $3\rightarrow 1$ band head. From there we conclude that $T_{\rm CO} =$3500–4000K. Unfortunately, an exact determination of the CO temperature fails because the higher bandheads are not available within our high resolution spectrum.
-------------------- -------------- ----------------- ----------------------- --------------------- --------------------
$N_{\rm CO}$ $T_{\rm CO}$ $v_{\rm gauss}$ $A^{\rm ISM}_{\rm v}$ $A_{\rm CO}$ $M_{\rm CO}$
$[{\rm cm}^{-2}$\] \[K\] \[km/s\] \[mag\] \[cm$^{2}$\] \[g\]
$5\cdot 10^{20}$ 3500 $\sim 50$ 10 $1.36\cdot 10^{26}$ $3.2\cdot 10^{24}$
-------------------- -------------- ----------------- ----------------------- --------------------- --------------------
: Parameters for CO used for the calculations shown in Fig.[\[turb\]]{}. $A_{\rm v}^{\rm ISM}$ denotes the foreground extinction, $A_{\rm CO}$ and $M_{\rm CO}$ are the emitting CO area projected to the sky and the mass of CO derived from the fit, respectively.
\[param\_turb\]
With the derived values for column density and temperature we model the synthetic CO band spectrum taking into account the foreground extinction, $A^{\rm ISM}_{\rm v}$, and fitted it to the observations (Figure \[turb\]; Table \[param\_turb\] summarizes the parameters used). The lower panel of Fig.\[turb\] shows the residue of observed minus calculated spectrum. Some remaining features may result from observational uncertainties (indicated as dashed parts), for example the dip around 2.3225$\mu$m and the region around 2.314$\mu$m, whereas the features at 2.3253 and 2.3084$\mu$m seem to be yet unidentified emission lines.
For a distance to MWC349 of 1.2kpc we find a CO emission area projected on the sky of $A_{\rm CO}\simeq 1.36\cdot 10^{26}$cm$^{2}$ and a mass of $M_{\rm CO}\simeq 3.2
\cdot 10^{24}$g. Our $A_{\rm CO}$ is about 5 times larger than the value of $2.5\cdot
10^{25}$cm$^{2}$ found by Geballe & Persson ([@geballe]) who could only derive a lower limit.
CO bands from a Keplerian rotating disk {#disk}
---------------------------------------
We investigate whether the CO bands in MWC349 can alternatively be explained by emission from the circumstellar disk. For demonstration, we first calculate the optically thin emission from an infinitesimally narrow ring. For Kepler rotation, the orbital velocity is $$\label{kepler}
v_{\rm rot} = \sqrt{GM_*/r}\,.$$ and the line of sight velocity, $v_{\rm los}$, of a gas element in the ring $$v_{\rm los} = v_{\rm rot}\cdot\cos \theta\cdot\cos i~.$$ Here $G$ denotes the gravitational constant, $r$ the ring radius, $M_{*}$ the stellar mass, $i$ the inclination angle of the ring, and $\theta$ the azimuthal angle. The resulting profile of the CO band head smoothed to the observational resolution of 15km/s shows a characteristic shoulder at short and a maximum at long wavelengths relative to an unbroadened band head (see Fig.\[rot\]). The distance between shoulder and maximum decreases with decreasing orbital velocity. The CO profile of a rotating ring is evidently very different from what is observed in MWC349.
We next calculate the emission from a whole disk viewed nearly edge-on, using the radiative transfer of Eq.(\[line\_intens\]). We assume power laws for the temperature and surface density, $$\label{temp}
T(r) = T_{0} \cdot \left(\frac{r}{r_{i}}\right)^{-p} = T_{0} \cdot \left(\frac{v}
{v_{i}}\right)^{2p}$$ and $$\label{sigma}
\Sigma(r) = \Sigma_{0} \cdot \left(\frac{r}{r_{i}}\right)^{-q} = \Sigma_{0} \cdot
\left(\frac{v}{v_{i}}\right)^{2q}~.$$ $r_i$ is the inner radius of the Kepler disk where the rotational velocity has its maximum value $v_{i}$. We fix the inner radius by putting $v_{i} =
60$km/s (see Figure \[v\_turb\]). Further, we assume $T_0 = 5000$K, which is the dissociation limit of CO molecules, and an inclination angle of $10\degr$. The radius $r_{\rm out}$, out to which the CO (2$\to$1) band is thermally excited, is a free parameter; the velocity there equals $v_{\rm
out}$. For these disk calculations we surmize that the CO bands at 2 micron are for radii greater than $r_{\rm out}$ only subthermally excited due to lower densities. Therefore their contribution to the overall spectrum is probably small. The CO disk is geometrically flat so that along each line of sight through the inclined disk the density and temperature are constant. The total flux from the disk is obtained from a straight forward numerical integration.
--------------- --------------- ------------------ -------------------- --------------------
$v_{\rm out}$ $r_{\rm out}$ $T(r_{\rm out})$ $N_{0}$ $N(r_{\rm out})$
$[$km/s\] \[AU\] \[K\] \[cm$^{-2}$\] \[cm$^{-2}$\]
40 14.4 3333 $8.5\cdot 10^{16}$ $2.5\cdot 10^{16}$
30 25.6 2500 $5.8\cdot 10^{16}$ $7.3\cdot 10^{15}$
--------------- --------------- ------------------ -------------------- --------------------
: Parameters for the fit of the CO disk calculations shown in Fig.. Fixed parameters are the mass of 26$M_{\sun}$, $T_{0}=
5000$K, $p = 0.5$, $q = 1.5$, and $v_i = 60$km/s which leads to $r_i\simeq
6.4$AU.
\[disk\_param\]
By a systematic variation of the free parameters ($\Sigma_0, v_{\rm out}, p,q$), we find that the shape of the spectrum is strongly influenced by the column density at the inner edge and by the size of the disk, whereas the exponents $p$ and $q$ play a minor role; we therefore fix them following Hayashi ([@hayashi]) to $p=0.5$ and $q=1.5$. Similar numbers are derived for the dust disk (Kraus et al. [@kraus]).
The disk models then contain only two free parameters: the outer radius, $r_{\rm
out}$, and the column density at the inner edge, $N_0$. To explain the observed band head profile with a rotating disk, one needs a large spread in the rotational velocity because otherwise one cannot surpress the shoulders seen in Fig. \[rot\]. The large velocity spread implies a large emitting area, and therefore a low column density, so that the emission becomes very optically thin ($\tau\simeq 10^{-4}$), even at the inner edge of the CO disk. The outer velocity $v_{\rm out}$ must lie in the range between 30 and 40km/s. If $v_{\rm out}$ were higher, the shoulder of the band head would become too broad, if it were smaller, the profile would get too narrow. Our best models of a rotating disk are presented in Fig. \[disk\_fit\]. Considering the quality of the data, they are marginally acceptable.
We add that if one assumes dust and gas to be thermally decoupled, it is in these models nevertheless possible that the strength of the CO band head amounts to $f \sim 5$% of the continuum, as observed, although the CO emission is very optically thin. As the dust emits like a blackbody, the equation $f\,B_{\nu}(T_{d}) =
\tau_{\nu}\,B_{\nu}(T_{\rm CO})$ is approximately fulfilled for $T_{d}\sim
1000$K, $T_{\rm CO} \sim 4000$K and $\tau\sim 10^{-4}$.
Discussion
==========
Mathematically, there is also the possibility to combine the wind model of Fig. 11 (solid line), which is based on a gaussian profile of $\sim 50$km/s half width, with a Kepler rotation as long as $v_{\rm rot} \le 40$km/s. This is shown in Figure \[fit\], where a rotational component with $v_{\rm rot} =
40$km/s (black curve) is added to the gaussian velocity profile. Otherwise the basic parameters are the same as for the wind (LTE; CO column density $\sim
5\cdot 10^{20}$cm$^{-2}$; $T_{\rm CO} = 3500\ldots 4000$K).
Although the fit is again very good, this scenario meets the following difficulties: If the rotational velocity $v_{\rm rot} \le 40$km/s, the stellar mass of 26M$_\odot$ implies a distance greater than 14AU. If the CO gas is coupled to the dust, it should have the same temperature of $\sim 950$K as the dust grains at that distance (Kraus et al. [@kraus]), which is much too low. If CO is located above the disk and thermally decoupled from the dust, it might achieve the required temperature of $\sim 3500$K by radiative heating. But the CO column density of $5\cdot 10^{20}$cm$^{-2}$ then implies a gas surface density which is at least an order of magnitude greater than what follows from the dust disk (adopting standard conversion factors $\Sigma_{\rm dust} :
\Sigma_{\rm H} \simeq 10^{-2}$, $N_{\rm CO} : N_{\rm H} \simeq 10^{-4}$).
The scenario of a pure wind is very vague in its details, but gives a remarkably good fit. We might think of dense molecular blobs emanating from the disk. The total emission area of the blobs is $\sim 0.6$AU$^{2}$.
$\theta [\degr]$ $v_{\rm rot,los}$\[km/s\]
------------- ------------------ ---------------------------
dashed line 180 $-96.5 \dots 96.5$
solid line 75 $-59.0 \dots 59.0$
dotted line 33 $-27.6 \dots 27.6$
: Angles $\theta$ and velocity ranges $v_{\rm rot,los}$ for the calculated sectors shown in Figure . The inclination angle was taken as $10\degr$ and the rotational velocity is 98km/s.
\[seg\_param\]
But there is still another alternative with $N_{\rm CO} \sim 5 \cdot
10^{20}$cm$^{-2}$, which we discuss now. If the CO band emission arises in the disk, it must come from inside the evaporation radius of the dust, $r_{\rm
evap}$, because of its high temperature of $\sim 3500$K. According to the Kraus et al. model of the dust disk, the distance is then smaller than 2.88AU according to a rotational velocity greater than 90km/s. Let us assume $v_{\rm rot} \simeq 100$km/s. To avoid the problems with the line profile in case of such a high $v_{\rm rot}$, as discussed in subsection \[disk\] and demonstrated in Figure \[rot\], we propose a configuration as depicted in Figure \[skizze4\]. The disk has a bulge and because of its small inclination angle $i\sim 10\degr$, this bulge blocks the light from the inner edge of the disk on the near side to the observer. We therefore receive emission only from sector 1 in Figure \[skizze4\] where the [*radial*]{} velocity is low. In Figure \[segment\] we show computations of the flux received from sector 1 for different opening angles $\theta$. The ranges of radial velocities in the sectors are summarized in Table \[seg\_param\].
Conclusions
===========
We present low and high resolution spectra in the near infrared of the B\[e\]-star MWC349. The wavelength interval 2.285–2.342$\mu$m contains mainly emission of the first overtone bands of the CO molecule as well as the Pfund series of atomic hydrogen. From modeling the Pfund lines under the assumption that they are optically thin, we find that they come from the inner part of the region around MWC349 and are in LTE. The hydrogen line profiles are approximately gaussian and their width indicates a wind velocity of $\sim
50$km/s.
The CO molecules are at a temperature of 3500–4000K, the population of the vibrational levels is close to LTE. The width of the $2\rightarrow 0$ band head indicates a velocity broadening of the order of 50–60km/s, depending on the broadening mechanism, wind or rotation. The location of the hot CO gas producing the band emission could not be clarified, but we were able to propose several scenarios.
Because the CO as well as the Pfund lines can be fitted with a gaussian velocity component of $\sim 50$km/s, the first possibility is that the CO bands arise in the windy transition layer between the disk and the HII region, presumably in dense clumps (see Fig.13). The band head has in this case an optical depth of order unity ($N_{\rm CO}\simeq 5\cdot 10^{20}$cm$^{-2}$).
Alternatively, the CO might form in a thin layer above the dust disk at radii between 6.5 and 15AU. In this scenario, the CO emission is optically very thin ($N_{\rm CO}\simeq 6\cdot 10^{16}$cm$^{-2}$; $\tau\simeq 10^{-4}$), and the gas is thermally decoupled from the dust. As we could only produce fits of mediocre quality (see Fig.15), due to the characteristic line profiles under Keplerian rotation (see Figure \[rot\]), this is not our favorite configuration.
Finally, we suggest that the CO bands come from the inner edge of the circumstellar disk. Because of the required high temperature and column density, the CO gas must be located inside the evaporation radius of the dust, i.e. at distances less than $\sim 3$AU from the star. Although this would imply a high rotational velocity component which would usually lead to a characteristic shoulder in the band profile, which is not observed, the disk may well have a bulge, as depicted in Figure \[skizze4\]. This bulge absorbs the CO emission from the near part of the rotating inner disk edge, so that the observer sees only a sector on the far side where the radial velocities are all smaller than $\sim 60$km/s. With such a geometrical configuration, a satisfactory fit is also possible (see Figure \[segment\]).
We would like to thank Frank Shu, the referee, for his helpful comments
Calvet, N., Patiño, A., Magris, C.G., D‘Alessio, P., 1991, ApJ 380, 617 Carr, J.S., 1989, ApJ 345, 522 Carr, J.S., 1995, Ap&SS 224, 25 Carr, J.S., Tokunaga, A.T., Najita, J., Shu, F.H., Glassgold, A.E., 1993, ApJ 411, L37 Chandler, C.J., Carlstrom, J.E., Scoville, N.Z., Dent, W.R., Geballe, T.R., 1993, ApJ 412, L71 Chandler, C.J., Carlstrom, J.E., Scoville, N.Z., 1995, ApJ 446, 793 Chandra, S., Maheshwari, V.U., Sharma, A.K., 1996, A&AS 117, 557 Cohen, M., Bieging, J.H., Dreher, J.W., Welch, W.J., 1985, ApJ 292, 249 Dunham, J. L., 1932a, Phys.Rev. 41, 713 Dunham, J. L., 1932b, Phys.Rev. 41, 721 Farrenq, R., Guelachvili, G., Sauval, A. J., Grevesse, N., Farmer, C. B., 1991, J.Mol.Spectrosc. 149, 375 Geballe, T.R., Persson, S.E., 1987, ApJ 312, 29 Geisel, S.L., 1970, ApJ 161, L105 Green, T.P., Lada, C.J., 1996, ApJ 461, 345 Hamann, F., Simon, M., 1986, ApJ 311, 909 Hamann, F., Simon, M., 1988, ApJ 327, 876 Hayashi, C., 1981, Suppl.Prog.Theor.Phys. 70, 35 Kraus, M., 2000, PhD-Thesis, University of Bonn Kraus, M., Krügel, E., Hengel, C., Thum, C., 2000 [*in preparation*]{} Lamers, H.J.G.L.M., Zickgraf, F.-J., de Winter, D., Houziaux, L., Zorec, J., 1998, A&A 340, 117 Leinert, C., 1986, A&A 155, L6 Mariotti, J.M., Chelli, A., Foy, R., Léna, P., Sibille, F., Tchountonov, G., 1983, A&A 120, 237 Martin, S.C., 1997, ApJ 478, L33 Martín-Pintado, J., Bachiller, R., Thum, C., Walmsley, M., 1989, A&A 215, L13 Menzel, D.H., Pekeris, C.L., 1935, MNRAS 96, 77 Najita, J., Carr, J.S., Glassgold, A.E., Shu, F.H., Tokunaga, A.T., 1996, ApJ, 462, 919 Rodríguez, L.F., Bastian, T.S., 1994, ApJ 428, 324 Scoville, N., Hall, D.N.B., Kleinmann, S.G., Ridgeway, S.T., 1979, ApJ 232, L121 Smith, H.A., Strelnitski, V., Miles, J.W., Kelly, D.M., Lacy, J.H., 1997, AJ 114, 2658 Storey, P.J., Hummer, D.G., 1995, MNRAS 272, 41 Strelnitski, V.S., Haas, M.R., Smith, H.A., Erickson, E.F., Colgan, S.W.J., Hollenbach, D.J., 1996a, Sci 272, 1459 Thum, C., Martín-Pintado, J., Bachiller, R., 1992, A&A 256, 507 Thum, C., Martín-Pintado, J., Quirrenbach, A., Matthews, H.E., 1998, A&A 333, L63 Thum, C., Matthews, H.E., Martín-Pintado, J., Serabyn, E., Planesas, P., Bachiller, R., 1994b, A&A 283, 582 Thum, C., Strelnitski, V.S., Martín-Pintado, J., Matthews, H.E., Smith, H.A., 1995, A&A 300, 843 White, R.L., Becker, R.H., 1985, ApJ 297, 677
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this letter we investigate the information provided by the “compass rose" (Crack, T.F. and Ledoit, O. (1996), Journal of Finance, 51(2), pg. 751-762) patterns revealed in phase portraits of daily stock returns. It has been initially suggested that the compass rose is just a manifestation of price clustering and discreteness and the tick size, factors that can affect the unbiasedness of an array of statistical tests based on stock returns. We show that this may not entirely be the case.'
address: 'Durham Business School, University of Durham, Mill Hill Lane, Durham, DH13LB, UK.'
author:
- 'Constantinos E. Vorlow'
title: 'Stock Price Clustering and Discreteness: The “Compass Rose" and Predictability'
---
Price Clustering and Discreteness ,Microstructure ,Compass rose ,Nonlinear and Complex Dynamics ,Surrogate Data Analysis.\
(JEL G10; G12; G22; Z00). 02.50.-r ,02.50.Tt ,05.45.Tp ,05.45.Ac .
Introduction
============
Price clustering and discreteness is considered an important chapter of the “market microstructure" literature with serious implications for tests of market efficiency, risk evaluation techniques and optimal design of securities procedures. The “*compass rose*", introduced dy @Crack96, is simply the manifestation of price clustering and discreteness in two or three dimensional phase portraits. More precisely, a pattern is usually revealed in scatter diagrams of daily percentage returns against their lagged values: rays emanating from the center of the portrait (the origin of the Cartesian axis system), generating a compass rose like formation of clusters of points. @Crack96 identify three conditions necessary for the appearance of the compass rose pattern:
1. The daily price changes of the stock should be small relative to the price level;
2. daily price changes should also be effected in discrete jumps of a small number of ticks and
3. the price of the stock should vary over a relatively wide range.
If any of the above points is violated, compass rose patters will be very weak or fail to realize altogether.
Several papers have recently appeared on the compass rose theme. Some mainly confirm the compass rose as a result of the tick size, price clustering and discreteness and investigate or relax the above conditions under which it appears (see [@Szpiro98], [@Wang00], [@Chen97], [@Lee99], [@Gleason00], [@Wang02] and [@McKenzie:03]) . Others (see [@Fang02], [@Kramer97], [@Koppl:01] and [@Amilon:03]) also concentrate on how the above factors affect the validity of statistical tests based on stock return sequences that reveal this “nanostructure" [@Szpiro98]. The compass rose should have appeared in several other works (such as [@EnrightPhD], [@Chen93], [@kar95], [@BM:96], [@Franses98:book], [@Franses00:book] and [@kar00]). However this was not achieved due to an unfortunate choice of graphical representation style or resolution and of the length (history) of the sequences examined.
@Crack96 suggested originally that the compass rose may not offer any help in predicting returns series. This is due to an apparent absence of any strong temporal continuity pattern in the phase portraits (see Fig. \[fig:fig1\]a). In this letter we show that from a small manipulation of the information that appears in the compass roses, we can gain additional information on the dynamics of stock returns processes in a very costless algorithmically way. An initial analysis of this information suggests the presence of strong nonlinear and possibly deterministic dynamics.
A different view of the compass rose
====================================
Price clustering and discreteness, as manifested in the compass rose, appears in the form of rays that emanate from the center of the phase portraits (see Fig. \[fig:fig1\](a) were we have a detail of the compass rose for the returns of the TESCO stock, chosen randomly from stocks forming to the FTSE100 index). It has been suggested that there is no predictability in such phase portraits [@Crack96]. This is due to temporal structure and information being concealed in the pase portraits of return sequences. One could initially adopt the view that the distribution of the points in the compass rose in Fig. \[fig:fig1\] could have been generated by a suitably discretized random Gaussian process. However [@Vorlow:03] showed that there may be a more delicate temporal pattern hidden under noise in the compass rose.
\[ Insert Fig. \[fig:fig1\] about here. \]
In this letter we follow the approach of @Koppl:01 and instead of examining directly the distribution of points in the phase portraits, we choose to model the value of the arcs $\Phi$ formed by the line joining the point with the origin (0,0) and the horizontal axis (Fig. \[fig:fig1\]a). We measure this in degrees for simplicity. An advantage of such “recoding" of the compass rose information is that all prices will range between 0 and 2$\pi$ or 0 and 360 degrees. It will also allow us to observe more efficiently the clustering and obtain a non-subjective view of the compass rose pattern (refer to [@Crack96] and [@Koppl:01] for discussions). Indeed, the investigation of the distribution of the compass rose arc values for TESCO (Fig. \[fig:fig1\]b), shows that there is concentration of points across the horizontal and vertical axis (0, 90, 180 and 270 (-90) degrees) and the diagonal rays (close to 45, 45+90, 45+180 and -45 degrees i.e., the NE-SW and NW-SE directions of the compass rose). This is referred to as *X-skewness* in @Koppl:01 and attributed to the presence of big-players and herding (see [@Koppl:96a] and [@Koppl:96b]). They also suggested that X-skewing may be inconsistent with ARCH effects and demonstrate this with simulations.[^1] [@Mand:99; @Mandsciam:99] also revealed a similar result through a different approach. Judging from the above literature, it seems that several differently fabricated nonlinear and non-random sequences can pass as (G)ARCH processes. There is also the case that stock returns may be characterized by more complex dynamics, not excluding deterministic or chaotic structures (refer to [@Kyrtsou2; @Kyrtsou1] and [@Vorlow:00; @Vorlow:03; @Vorlow:04] for more details).
If we plot the arc values $\Phi$ as a time series we also obtain a pattern that confirms this clustering of points (see Fig. \[fig:fig1\]c). Plotting a sorted version of the sequence in Fig. \[fig:fig1\](c), provides us with the pattern observed in \[fig:fig1\](d), where the plateaus indicate more clearly where the clustering occurs. The intensity of the compass rose patterns may differ between stocks, however the clustering along the main directions as discussed above, usually prevails. Searching for some type of temporal dependency, we looked into the phase portraits of the arcs. In Fig. \[fig:fig2\](a) we have the first lag phase portrait and in Fig. \[fig:fig2\](b) the second lag one. We can see very curious patterns arising in both diagrams. There are grids (Fig. \[fig:fig2\]b) and rays (Fig. \[fig:fig2\]a) which correspond to the main rays of the compass rose as seen in Fig. \[fig:fig1\](c) and (d). These patterns need more analysis.
\[ Insert Fig. \[fig:fig2\] about here. \]
In order to provide evidence of some kind of dependency in the above patterns we calculated the BDS test (see [@Brock87] and [@BrockBaek91]) for the TESCO sequence of arc values $\Phi$. In table \[tab:tab1\] we see clearly that for a range of dimensions $d_E$ (2 to 10), the BDS test reports non IID dynamics for neighborhood area ranges between 0.5 and 1.5 times the standard deviation of arcs. When the size becomes twice the standard deviation (which is regarded as a large radius), we can accept independence only up to dimension 3. This is an initial indication that the dynamics as observed from the sequence of arcs, may contain some dependency that could be used for forecasting purposes.
\[ Insert table \[tab:tab1\] about here. \]
To back up this result, we calculated the BDS test within a Surrogate Data Analysis framework (SDA: [@Theiler92:surr]). SDA is a permutation test framework, similar to bootstrapping, and is used to test specific nulls that exclude certain dynamics. Every null comes with its own tailor made simulation procedures for creating surrogate data sets from the original sequences (refer to [@Kaplan95], [@Kantz], [@Schreiber96a; @Schreiber00] and [@Kugium01] for more details). In nutshell, via SDA one searches for large discrepancies between the statistic values on the original and surrogate sequences. When this occurs, one can safely reject the hypothesis that the original sequence stems from a process that is in compliance with the null. SDA is appropriate here as it allows us to bypass the limitations of chaotic invariance measures (such as Lyapunov exponents and dimension statistics) due to small data sets and noisy information. Although the BDS is a test for independence and not nonlinear determinism and complexity, under the SDA framework it can be used to test for absence of stochastic randomness via the exclusion of the null hypothesis.
For our case we used the null of the $\Phi$ sequences being *a monotonic nonlinear transformation of linearly filtered noise*, which is also regarded as the “most interesting". This implies strongly, the absence of stochastic (random) dynamics. We analyzed data from 53 FTSE 100 index stocks, spanning the period 01/01/1970 to 5/30/2003 (a maximum of 8717 observations). In tables \[tab:tab2\] and \[tab:tab3\] we present the results for 5% and 2.5% levels of statistical significance, for neighborhood sizes $\epsilon$ ranging from 0.5 to 2 times the standard deviation of $\Phi$. We clearly see that for up to 1.5 times the standard deviation, the null can be safely rejected (observe the large biases). However, for $\epsilon$ equalling twice the standard deviation, we can not always reject the null (but this is only for a few cases). Our SDA results do not reject the case of forecastability on the basis of arc values. They also suggest the presence of complex dynamics and the possibility of some level of determinism.
\[ Insert table \[tab:tab2\] about here. \]
\[ Insert table \[tab:tab3\] about here. \]
Conclusions
===========
By adopting an approach similar to [@Koppl:01] we showed that the compass rose [@Crack96] can provide useful information for understanding further the dynamics of stock return sequences. Moreover, we do not reject the case of these dynamics being complex and forecastable. We also do not exclude the possibility of nonlinear determinism. However more research is needed on this controversial area (see [@Lo:88], [@MM92], and [@Hsieh:91]). An interesting area for future research would be to experiment with differently simulated processes (such as purely chaotic, stochastic, (G)ARCH and mixtures of these) and examine similarities or dissimilarities with the results for the framework we follow in this letter. It would also be interesting to see how inhomogeneous sampling of such processes could alter the results we report here. That could provide useful information on how our view of the stock return dynamics changes as we move from high-frequency data to lower frequencies, which is also an issue that has been concerning the general compass rose literature.
[39]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix
Amilon, H., 2003. [GARCH]{} estimation and discrete stock prices: an application to low-priced [A]{}ustralian stocks. Economics Letters 81 (2), 215–222.
Andreou, A. S., Pavlides, G., Karytinos, A., 2000. Nonlinear time-series analysis of the Greek exchange-rate market. International Journal of Bifurcation and Chaos 10 (7), 1729–1759.
Andrew, L. W., MacKinlay, C. A., 1988. Stock [M]{}arket [P]{}rices do not [F]{}ollow [R]{}andom [W]{}alks: [E]{}vidence from a [S]{}imple [S]{}pecification [T]{}est. Review of Financial Studies 1 (1), 41–66.
Antoniou, A., Vorlow, C. E., 2000. Recurrence plots and financial time series analysis. Neural Network World 10 (1-2), 131–146.
Antoniou, A., Vorlow, C. E., 2004. [P]{}rice [C]{}lustering and [D]{}iscreteness: [I]{}s there [C]{}haos behind the [N]{}oise? Under Review in Physica A: Statistical Mechanics and Its Applications. <http://arxiv.org/abs/cond-mat/0407471>
Antoniou, A., Vorlow, C. E., 2004. Recurrence quantification analysis of wavelet pre-filtered index returns. Physica A: Statistical Mechanics and Its Apllications Forthcoming.
Brealy, R. A., Meyers, S. C., 1996. Principles of [C]{}orporate [F]{}inance, 4th Edition. McGraw-Hill, New York.
Brock, W., Dechert, W., Scheinkman, J., 1987. A test for independence based upon the correlation dimension. Working paper, University of Winsconsin.
Brock, W. A., Baek, E. G., 1991. Some [T]{}heory of [S]{}tatistical [I]{}nference for [N]{}onlinear [S]{}cience. Review of Economic Studies 58 (4), 697–716.
Broussard, J., Koppl, R., 1996. Big [P]{}layers and and the [R]{}ussian [R]{}uble: [E]{}xplaining [V]{}olatility [D]{}ynamics. Managerial Finance 25 (1), 49–63.
Chen, A., 1997. The square compass rose: the evidence from [T]{}aiwan. Journal of Multinational Financial Management 7 (2), 127–144.
Chen, P., 1993. Searching for economic chaos: A challenge to econometric practice and nonlinear tests. In: Day, R. H., Chen, P. (Eds.), Nonlinear dynamics and evolutionary economics. Oxford University Press, Oxford; New York; Toronto and Melbourne, pp. 217–53.
Crack, T. F., Ledoit, O., 1996. Robust structure without predictability: The “compass rose" pattern of the stock market. Journal of Finance 51 (2), 751–762.
Enright, A. J., 1992. Searching for chaotic components in financial time-series. Ph.d. thesis, Pace University.
Fang, Y., 2002. The compass rose and random walk tests. Computational Statistics & Data Analysis 39 (3), 299–310.
Franses, P. H., 1998. Time series anslysis models for business and economic forecasting. Cambridge University Press.
Franses, P. H., Dijk, D. v., 2000. Nonlinear time series models in empirical finance. Cambridge University Press.
Gleason, K. C., Lee, C. I., Mathur, I., 2000. An explanation for the compass rose pattern. Economics Letters 68 (2), 127–133.
Hsieh, D. A., 1991. Chaos and nonlinear dynamics: Application to financial markets. Journal of Finance 46 (5), 1839–77.
Kantz, H., Schreiber, T., 1997. Nonlinear [T]{}ime [S]{}eries [A]{}nalysis. No. 7 in Cambridge Nonlinear Science series. Cambridge University Press, UK.
Kaplan, D. T., Glass, L., 1995. Understanding nonlinear dynamics. Textbooks in mathematical sciences. Springer-Verlag, New York.
Koppl, R., Nardone, C., 2001. The [A]{}ngular [D]{}istribution of [A]{}sset [R]{}eturns in [D]{}elay [S]{}pace. Discrete Dynamics in Nature and Society 6, 101–120.
Koppl, R., Yeager, L., July 1996. Big [P]{}layers and [H]{}erding in [A]{}sset [M]{}arkets: [T]{}he [C]{}ase of the [R]{}ussian [R]{}uble. Explorations in Economic History 33 (3), 367–383.
Kramer, W., Runde, R., 1997. Chaos and the compass rose. Economics Letters 54 (2), 113–118.
Kugiumtzis, D., 2001. On the [R]{}eliability of the [S]{}urrogate [D]{}ata [T]{}est for [N]{}onlinearity in the [A]{}nalysis of [N]{}oisy [T]{}ime [S]{}eries. International Journal of Bifurcation and Chaos 7 (11), 1881–1896.
Kyrtsou, C., Terraza, M., 2002. Stochastic [C]{}haos or [ARCH]{} [E]{}ffects in [S]{}tock [S]{}eries? a [C]{}omparative [S]{}tudy. International Review of Financial Analysis 11 (4), 407–431.
Kyrtsou, C., Terraza, M., 2003. Is it [P]{}ossible to [S]{}tudy [C]{}haotic and [ARCH]{} [B]{}ehaviour [J]{}ointly? [A]{}pplication of a [N]{}oisy [M]{}ackey-[G]{}lass [E]{}quation with [H]{}eteroskedastic [E]{}rrors to the [P]{}aris [S]{}tock [E]{}xchange [R]{}eturns [S]{}eries. Computational Economics 21 (3), 257–276.
Lee, C. I., Gleason, K. C., Mathur, I., 1999. A comprehensive examination of the compass rose pattern in futures markets. The Journal of Futures Markets 19 (5), 541–564.
Mandelbrot, B. B., 1999. A multifractal [W]{}alk down [W]{}all [S]{}treet. Scientific American 280 (2), 70–74.
Mandelbrot, B. B., 1999. Renormalization and fixed points in finance, since 1992 263 (1).
Mayfield, E. S., Mizrach, B., 1992. On determining the dimension of real-time stock-price data. Journal of Business and Economic Statistics 10 (3), 367–74.
McKenzie, M. D., Frino, A., 2003. The tick/volatility ratio as a determinant of the compass rose: empirical evidence from decimalisation on the [NYSE]{}. Accounting & Finance 43 (3), 331–331.
Papaioannou, G., Karytinos, A., 1995. Nonlinear time series analysis of the stock exchange: The case of an emerging market. International Journal of Bifurcation and Chaos 5 (6), 1557–1585.
Schreiber, T., Schmitz, A., 1996. Improved surrogate data for nonlinearity tests. Phys. Rev. Lett. 77 (4), 635–638.
Schreiber, T., Schmitz, A., 2000. Surrogate time series. Physica D 142 (3-4), 346–382.
Szpiro, G. G., 1998. Tick size, the compass rose and market nanostructure. Journal of Banking & Finance 22 (12), 1559–1569.
Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., Farmer, J. D., 1992. Testing for nonlinearity in time series: the method of surrogate data. Physica D: Nonlinear Phenomena 58 (1-4), 77–94.
Wang, E., Hudson, R., Keasey, K., 2000. Tick size and the compass rose: further insights. Economics Letters 68 (2), 119–125.
Wang, H., Wang, C., 2002. Visibility of the compass rose in financial asset returns: A quantitative study. Journal of Banking & Finance 26 (6), 1099–1111.
0.0in 0.0in
[rrrrr]{}\
$\epsilon$: & $0.5 \times s$ & $ 1.0 \times s$ & $ 1.5
\times s$ & $ 2.0 \times s$\
$d_E$ & 0.897 & 1.794 & 2.691 & 3.588\
2 & 745.31 & 241.38 & 65.39 & 3.45\
3 & 846.67 & 223.27 & 56.12 & 2.84\
4 & 1010.50 & 220.27 & 49.64 & 1.68\
5 & 1243.98 & 222.22 & 44.97 & 0.95\
6 & 1577.48 & 229.07 & 41.54 & 0.39\
7 & 2053.94 & 239.21 & 38.59 & $-$0.18\
8 & 2743.93 & 252.22 & 36.10 & $-$0.72\
9 & 3741.75 & 269.69 & 34.16 & $-$1.08\
10 & 5201.71 & 291.13 & 32.85 & $-$1.16\
\
2 & 0.00 & 0.00 & 0.00 & 0.00\
3 & 0.00 & 0.00 & 0.00 & 0.00\
4 & 0.00 & 0.00 & 0.00 & 0.09\
5 & 0.00 & 0.00 & 0.00 & 0.34\
6 & 0.00 & 0.00 & 0.00 & 0.69\
7 & 0.00 & 0.00 & 0.00 & 0.86\
8 & 0.00 & 0.00 & 0.00 & 0.47\
9 & 0.00 & 0.00 & 0.00 & 0.28\
10 & 0.00 & 0.00 & 0.00 & 0.25\
------------------------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- --------------
Neighborhood size $\epsilon$: $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$ $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$ $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$
FTSE ALL SHARE - PRICE INDEX 1249.61 339.45 48.02 $-$21.75 $-$1246.52 $-$338.09 $-$47.34 22.31 2.53 1.54 1.19 1.16
FTSE 100 - PRICE INDEX 152.69 97.72 35.55 20.52 $-$152.91 $-$97.70 $-$35.56 $-$20.58 0.87 1.00 0.94 1.21
ALLIED DOMECQ 823.28 259.24 68.38 1.29 $-$821.57 $-$258.47 $-$67.80 $-$0.93 1.81 1.00 0.99 1.06
AMVESCAP 159.36 130.18 60.63 28.46 $-$158.57 $-$129.63 $-$59.80 $-$27.64 1.20 1.24 1.26 1.25
ASSD.BRIT.FOODS 290.34 163.95 62.85 21.22 $-$289.08 $-$162.77 $-$61.89 $-$20.08 1.37 1.13 1.32 1.26
AVIVA 928.47 270.09 66.89 0.16 $-$927.36 $-$269.49 $-$66.61 0.11 1.38 0.96 1.17 1.13
BARCLAYS 1076.24 281.21 62.10 $-$5.50 $-$1076.56 $-$281.26 $-$62.33 5.20 1.21 1.16 1.09 1.13
BOC GROUP 829.83 257.29 67.76 3.58 $-$828.10 $-$256.45 $-$67.26 $-$3.16 1.70 1.05 0.91 0.92
BOOTS GROUP 1076.14 283.31 64.05 $-$5.67 $-$1076.77 $-$283.69 $-$64.38 5.31 1.01 0.92 0.79 0.81
BP 1115.66 295.08 63.04 $-$7.99 $-$1114.58 $-$294.74 $-$62.90 8.04 1.32 1.04 1.24 1.26
BRIT.AMERICAN TOBACCO 961.19 273.01 66.07 $-$3.63 $-$959.84 $-$272.50 $-$65.74 3.80 1.23 0.82 0.75 0.76
BRITISH LAND 568.83 223.68 66.58 4.10 $-$567.36 $-$223.00 $-$66.02 $-$3.65 1.02 0.94 0.72 0.74
BUNZL 192.12 138.33 59.76 23.61 $-$191.62 $-$137.83 $-$59.07 $-$23.05 1.01 0.97 1.18 1.10
CADBURY SCHWEPPES 653.11 233.87 70.17 6.34 $-$651.55 $-$232.74 $-$69.52 $-$5.83 1.25 0.79 0.70 0.68
DAILY MAIL ’A’ 89.23 96.50 47.69 35.62 $-$88.25 $-$95.23 $-$46.43 $-$33.71 0.78 0.95 0.85 0.77
DIAGEO 619.63 225.74 67.63 5.33 $-$618.52 $-$224.98 $-$67.00 $-$4.83 1.01 0.88 0.84 0.71
DIXONS GP. 341.08 172.89 60.51 13.26 $-$340.48 $-$172.19 $-$60.22 $-$12.95 1.19 1.12 1.11 1.22
EMAP 72.70 69.19 35.03 20.02 $-$72.73 $-$68.86 $-$35.12 $-$20.23 1.26 1.27 1.26 1.07
EXEL 433.53 203.06 69.26 9.61 $-$433.38 $-$203.00 $-$69.29 $-$9.73 1.01 0.82 0.90 0.87
FOREIGN & COLONIAL 711.36 237.89 63.39 6.31 $-$709.53 $-$236.99 $-$62.93 $-$5.96 1.00 1.04 1.19 1.07
GKN 979.42 279.06 65.76 $-$2.93 $-$978.52 $-$278.57 $-$65.37 3.30 1.39 0.88 0.81 0.80
GLAXOSMITHKLINE 1069.02 282.23 62.10 $-$6.11 $-$1068.62 $-$282.08 $-$61.98 6.14 1.79 1.12 1.04 1.02
GRANADA 459.67 200.60 65.15 9.49 $-$459.69 $-$200.59 $-$65.08 $-$9.41 1.27 1.19 1.11 1.16
GUS 128.72 111.26 50.83 35.48 $-$128.49 $-$110.76 $-$50.36 $-$34.92 0.89 0.92 0.93 0.83
HANSON 804.72 251.67 61.16 $-$2.19 $-$804.64 $-$251.61 $-$61.11 2.20 1.06 1.09 0.90 0.98
HILTON GROUP 777.65 248.02 70.54 4.30 $-$775.18 $-$246.83 $-$69.75 $-$3.75 1.92 1.20 1.21 1.18
IMP.CHM.INDS. 1007.77 283.71 65.69 $-$5.80 $-$1007.17 $-$283.64 $-$65.66 5.77 1.64 0.92 1.22 1.21
JOHNSON MATTHEY 378.71 178.51 63.97 14.39 $-$378.47 $-$178.21 $-$63.81 $-$14.15 0.91 0.99 0.89 0.89
LAND SECURITIES 1089.61 282.28 62.79 $-$6.22 $-$1088.38 $-$281.71 $-$62.41 6.55 1.83 1.30 1.22 1.20
LEGAL & GENERAL 970.82 269.13 65.65 0.04 $-$969.14 $-$268.62 $-$65.34 0.22 1.09 1.10 0.88 0.92
MARKS & SPENCER GROUP 912.70 272.35 64.73 $-$2.03 $-$909.93 $-$270.77 $-$63.62 3.04 1.67 1.36 1.32 1.26
MORRISON (WM) SPMKTS. 90.75 90.64 45.66 44.19 $-$89.83 $-$89.01 $-$44.06 $-$41.71 0.93 0.97 0.76 1.07
NEXT 253.11 151.42 55.97 18.84 $-$252.81 $-$150.97 $-$55.62 $-$18.50 1.02 1.16 0.94 1.13
PEARSON 444.27 196.80 64.87 8.11 $-$444.02 $-$196.49 $-$64.80 $-$8.04 1.08 0.84 0.94 0.84
PROVIDENT FINL. 266.69 155.41 58.69 16.83 $-$266.57 $-$155.36 $-$58.50 $-$16.76 0.77 0.98 1.11 0.77
PRUDENTIAL 910.07 267.31 65.78 0.13 $-$909.09 $-$266.63 $-$65.29 0.37 1.08 0.98 0.79 0.79
RECKITT BENCKISER 836.20 258.41 67.44 2.10 $-$836.20 $-$258.47 $-$67.68 $-$2.30 1.33 1.29 1.19 1.24
REED ELSEVIER 898.32 261.80 63.61 $-$1.99 $-$898.15 $-$261.72 $-$63.60 1.93 1.37 1.14 1.04 1.01
RENTOKIL INITIAL 197.81 131.19 56.03 28.14 $-$196.12 $-$129.70 $-$54.49 $-$26.38 0.97 0.94 0.96 0.82
REXAM 697.01 237.79 72.49 6.51 $-$695.08 $-$236.83 $-$71.66 $-$5.71 1.38 0.90 0.90 0.88
RIO TINTO 849.08 260.87 66.48 $-$3.08 $-$848.51 $-$260.71 $-$66.40 3.19 1.59 1.37 1.33 1.31
ROYAL BANK OF SCOTLAND 433.58 198.72 65.90 9.33 $-$432.41 $-$198.07 $-$65.28 $-$8.83 0.88 1.06 0.93 0.86
SAINSBURY (J) 698.76 227.30 59.38 0.21 $-$698.30 $-$227.22 $-$59.18 $-$0.21 1.15 1.29 1.29 1.21
SCHRODERS 72.28 84.11 40.53 41.80 $-$71.93 $-$83.50 $-$39.86 $-$40.50 1.14 1.17 1.16 1.20
SCOT. & NEWCASTLE 750.36 245.66 67.65 0.87 $-$749.30 $-$244.78 $-$67.08 $-$0.40 1.09 0.72 0.94 0.95
SHELL TRANSPORT & TRDG. 1199.79 294.59 65.20 $-$6.32 $-$1198.76 $-$293.78 $-$64.65 6.72 1.46 1.07 1.04 1.04
SMITH & NEPHEW 500.14 208.90 71.11 14.25 $-$497.28 $-$206.89 $-$69.41 $-$12.80 1.64 1.67 1.60 1.60
SMITHS GROUP 587.34 222.17 69.99 10.48 $-$586.92 $-$222.11 $-$70.07 $-$10.66 1.36 0.92 1.27 1.29
STD.CHARTERED 623.16 223.03 67.43 5.56 $-$622.41 $-$222.38 $-$66.93 $-$5.16 1.24 1.05 0.84 0.93
TESCO 745.31 241.38 65.24 3.24 $-$744.56 $-$240.87 $-$64.96 $-$2.97 1.55 1.15 1.15 1.14
TOMKINS 77.66 73.89 37.40 25.49 $-$77.65 $-$73.70 $-$37.12 $-$25.38 1.17 1.29 0.94 0.93
UNILEVER (UK) 1017.26 278.21 59.93 $-$7.77 $-$1017.15 $-$278.20 $-$59.90 7.76 1.39 1.04 1.09 0.98
WHITBREAD 851.02 262.03 66.73 3.23 $-$849.13 $-$260.92 $-$65.95 $-$2.46 1.79 1.20 1.11 1.11
WOLSELEY 172.73 129.30 52.87 27.14 $-$172.66 $-$129.15 $-$52.69 $-$27.06 0.82 0.95 1.03 0.89
WPP GROUP 71.43 60.94 38.30 21.64 $-$71.01 $-$60.48 $-$37.74 $-$21.25 1.11 0.81 0.82 0.82
------------------------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- --------------
: Surrogate Data Analysis results on arc values $\Phi$ for 53 companies in the FTSE100. Discriminating statistic: BDS test (embedding dimension 2). Neighborhood size $\epsilon_1 = 0.5
\times s$, $\epsilon_2 = 1.0 \times s$, $\epsilon_3 = 1.5 \times
s$ and $\epsilon_4 = 2.0 \times s$, where $s =$ standard deviation of $\Phi$ sequence. Biases and standard errors reported for significance level $\alpha = 5\%$.[]{data-label="tab:tab2"}
------------------------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- --------------
Neighborhood size $\epsilon$: $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$ $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$ $\epsilon_1$ $\epsilon_2$ $\epsilon_3$ $\epsilon_4$
FTSE ALL SHARE - PRICE INDEX 1249.61 339.45 48.02 $-$21.75 $-$1245.98 $-$338.20 $-$47.50 22.15 1.97 0.98 0.98 1.01
FTSE 100 - PRICE INDEX 152.69 97.72 35.55 20.52 $-$152.69 $-$97.82 $-$35.55 $-$20.54 0.98 0.93 0.86 1.12
ALLIED DOMECQ 823.28 259.24 68.38 1.29 $-$821.92 $-$258.54 $-$68.02 $-$1.07 1.35 0.96 0.97 0.92
AMVESCAP 159.36 130.18 60.63 28.46 $-$158.41 $-$129.22 $-$59.67 $-$27.18 0.90 0.83 0.85 0.89
ASSD.BRIT.FOODS 290.34 163.95 62.85 21.22 $-$288.64 $-$162.40 $-$61.51 $-$19.84 1.35 1.33 1.37 1.30
AVIVA 928.47 270.09 66.89 0.16 $-$926.88 $-$269.44 $-$66.56 0.05 1.36 1.18 1.23 1.17
BARCLAYS 1076.24 281.21 62.10 $-$5.50 $-$1075.93 $-$281.31 $-$62.19 5.38 1.36 1.14 1.16 1.11
BOC GROUP 829.83 257.29 67.76 3.58 $-$828.19 $-$256.50 $-$67.12 $-$3.08 1.62 1.03 0.89 0.84
BOOTS GROUP 1076.14 283.31 64.05 $-$5.67 $-$1075.91 $-$283.18 $-$63.98 5.69 1.26 1.01 0.95 0.92
BP 1115.66 295.08 63.04 $-$7.99 $-$1114.62 $-$294.54 $-$62.78 8.18 1.70 1.34 1.19 1.20
BRIT.AMERICAN TOBACCO 961.19 273.01 66.07 $-$3.63 $-$959.60 $-$272.40 $-$65.67 3.99 1.37 1.08 1.00 0.94
BRITISH LAND 568.83 223.68 66.58 4.10 $-$567.30 $-$222.67 $-$65.59 $-$3.19 1.17 0.95 1.03 1.00
BUNZL 192.12 138.33 59.76 23.61 $-$191.61 $-$137.79 $-$59.05 $-$22.97 0.99 0.98 0.90 0.84
CADBURY SCHWEPPES 653.11 233.87 70.17 6.34 $-$651.52 $-$232.88 $-$69.47 $-$5.61 1.35 1.04 1.02 1.00
DAILY MAIL ’A’ 89.23 96.50 47.69 35.62 $-$88.45 $-$95.38 $-$46.60 $-$33.87 1.07 0.99 1.01 1.14
DIAGEO 619.63 225.74 67.63 5.33 $-$618.36 $-$225.24 $-$67.24 $-$5.02 1.26 1.15 1.19 1.22
DIXONS GP. 341.08 172.89 60.51 13.26 $-$340.07 $-$172.18 $-$59.87 $-$12.58 0.97 0.99 0.88 0.99
EMAP 72.70 69.19 35.03 20.02 $-$72.78 $-$68.80 $-$34.77 $-$20.02 1.04 1.08 0.87 1.04
EXEL 433.53 203.06 69.26 9.61 $-$433.28 $-$202.75 $-$69.08 $-$9.38 1.04 1.04 0.99 0.97
FOREIGN & COLONIAL 711.36 237.89 63.39 6.31 $-$709.87 $-$236.98 $-$62.74 $-$5.78 1.23 1.21 1.03 1.08
GKN 979.42 279.06 65.76 $-$2.93 $-$978.64 $-$278.57 $-$65.46 3.20 1.33 1.10 1.04 1.08
GLAXOSMITHKLINE 1069.02 282.23 62.10 $-$6.11 $-$1068.51 $-$281.94 $-$61.86 6.30 1.29 1.03 1.03 1.04
GRANADA 459.67 200.60 65.15 9.49 $-$459.69 $-$200.39 $-$65.16 $-$9.38 1.15 1.29 1.14 1.22
GUS 128.72 111.26 50.83 35.48 $-$128.32 $-$110.95 $-$50.35 $-$34.85 1.15 1.12 1.14 1.09
HANSON 804.72 251.67 61.16 $-$2.19 $-$804.75 $-$251.79 $-$61.30 2.07 1.09 1.03 0.99 1.02
HILTON GROUP 777.65 248.02 70.54 4.30 $-$776.09 $-$247.12 $-$69.98 $-$3.90 1.31 1.17 1.10 1.14
IMP.CHM.INDS. 1007.77 283.71 65.69 $-$5.80 $-$1007.48 $-$283.85 $-$65.79 5.67 1.33 1.03 0.88 0.88
JOHNSON MATTHEY 378.71 178.51 63.97 14.39 $-$378.02 $-$177.88 $-$63.54 $-$13.91 1.03 1.01 0.89 0.88
LAND SECURITIES 1089.61 282.28 62.79 $-$6.22 $-$1088.71 $-$281.87 $-$62.38 6.55 1.26 1.10 0.99 0.95
LEGAL & GENERAL 970.82 269.13 65.65 0.04 $-$969.44 $-$268.38 $-$65.24 0.33 1.36 1.08 0.96 0.96
MARKS & SPENCER GROUP 912.70 272.35 64.73 $-$2.03 $-$909.91 $-$270.94 $-$63.82 2.77 1.33 0.95 0.78 0.85
MORRISON (WM) SPMKTS. 90.75 90.64 45.66 44.19 $-$90.24 $-$89.74 $-$44.68 $-$42.54 0.93 0.90 0.92 0.93
NEXT 253.11 151.42 55.97 18.84 $-$252.66 $-$151.03 $-$55.45 $-$18.50 0.82 0.96 0.82 0.73
PEARSON 444.27 196.80 64.87 8.11 $-$443.83 $-$196.62 $-$64.64 $-$7.87 1.09 1.02 1.00 0.94
PROVIDENT FINL. 266.69 155.41 58.69 16.83 $-$266.55 $-$155.49 $-$58.70 $-$16.89 1.04 1.29 1.20 1.09
PRUDENTIAL 910.07 267.31 65.78 0.13 $-$909.18 $-$267.05 $-$65.59 $-$0.00 1.22 1.01 1.06 1.11
RECKITT BENCKISER 836.20 258.41 67.44 2.10 $-$835.56 $-$257.93 $-$67.04 $-$1.72 1.20 1.12 1.00 0.98
REED ELSEVIER 898.32 261.80 63.61 $-$1.99 $-$897.92 $-$261.77 $-$63.47 2.14 1.37 1.13 1.12 1.13
RENTOKIL INITIAL 197.81 131.19 56.03 28.14 $-$196.05 $-$129.48 $-$54.33 $-$26.04 0.91 1.19 1.09 1.17
REXAM 697.01 237.79 72.49 6.51 $-$695.26 $-$236.76 $-$71.64 $-$5.70 1.53 1.21 1.17 1.21
RIO TINTO 849.08 260.87 66.48 $-$3.08 $-$848.18 $-$260.53 $-$66.23 3.30 1.03 0.98 0.85 0.86
ROYAL BANK OF SCOTLAND 433.58 198.72 65.90 9.33 $-$432.32 $-$198.01 $-$65.34 $-$8.82 1.15 0.96 0.81 0.83
SAINSBURY (J) 698.76 227.30 59.38 0.21 $-$698.48 $-$227.33 $-$59.46 $-$0.22 0.93 0.95 0.83 0.82
SCHRODERS 72.28 84.11 40.53 41.80 $-$71.77 $-$83.11 $-$39.59 $-$40.11 1.03 1.13 1.10 1.26
SCOT. & NEWCASTLE 750.36 245.66 67.65 0.87 $-$749.08 $-$245.00 $-$67.06 $-$0.26 1.03 0.96 0.91 0.89
SHELL TRANSPORT & TRDG. 1199.79 294.59 65.20 $-$6.32 $-$1198.60 $-$294.14 $-$65.03 6.45 1.58 1.17 1.10 1.07
SMITH & NEPHEW 500.14 208.90 71.11 14.25 $-$497.73 $-$207.15 $-$69.72 $-$12.99 1.17 1.20 1.06 1.11
SMITHS GROUP 587.34 222.17 69.99 10.48 $-$586.43 $-$221.51 $-$69.40 $-$9.90 1.28 0.97 1.12 1.14
STD.CHARTERED 623.16 223.03 67.43 5.56 $-$623.18 $-$222.82 $-$67.38 $-$5.56 1.18 1.10 1.13 1.21
TESCO 745.31 241.38 65.24 3.24 $-$745.08 $-$241.37 $-$65.23 $-$3.33 1.12 1.16 1.10 1.08
TOMKINS 77.66 73.89 37.40 25.49 $-$77.72 $-$73.55 $-$37.02 $-$25.33 0.87 0.98 1.10 0.90
UNILEVER (UK) 1017.26 278.21 59.93 $-$7.77 $-$1017.42 $-$278.30 $-$60.03 7.68 1.28 0.98 0.98 0.96
WHITBREAD 851.02 262.03 66.73 3.23 $-$849.28 $-$261.35 $-$66.25 $-$2.81 1.44 1.10 1.12 1.16
WOLSELEY 172.73 129.30 52.87 27.14 $-$172.58 $-$129.19 $-$52.74 $-$26.87 1.06 1.08 1.10 0.96
WPP GROUP 71.43 60.94 38.30 21.64 $-$71.32 $-$60.89 $-$38.02 $-$21.43 1.00 1.01 0.73 0.94
------------------------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- --------------
: Surrogate Data Analysis results on arc values $\Phi$ for 53 companies in the FTSE100. Discriminating statistic: BDS test (embedding dimension 2). Neighborhood size $\epsilon_1 = 0.5
\times s$, $\epsilon_2 = 1.0 \times s$, $\epsilon_3 = 1.5 \times
s$ and $\epsilon_4 = 2.0 \times s$, where $s =$ standard deviation of $\Phi$ sequence. Biases and standard errors reported for significance level $\alpha = 2.5\%$.[]{data-label="tab:tab3"}
\
\
\
[^1]: The multimodality of the distribution in the histogram of Fig. \[fig:fig1\](b) may also be an indication of more complex dynamics. See in [@Kantz] and [@Kaplan95] for discussions.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A notion of a coring extension is defined and it is related to the existence of an additive functor between comodule categories that factorises through forgetful functors. This correspondence between coring extensions and factorisable functors is illustrated by functors between categories of descent data. A category in which objects are corings and morphisms are coring extensions is also introduced.'
address: ' Department of Mathematics, University of Wales Swansea, Singleton Park, Swansea SA2 8PP, U.K.'
author:
- Tomasz Brzeziński
title: A note on coring extensions
---
Introduction
============
Given two algebras $A$, $B$ over a commutative ring $k$, an algebra extension or an algebra map $B\to A$ can be equivalently characterised as a $k$-additive functor $F: \cM_A\to\cM_B$ with the factorisation property $$\xymatrix{\cM_A\ar[rd]_{U_A}\ar[rr]^F& & \cM_B \ar[ld]^{U_B}\\
& \cM_k, &}$$ where $U_A$, $U_B$ are forgetful functors (cf. [@Par:Ver]). Through this correspondence, morphisms of $k$-algebras can be defined as functors having such a factorisation property. This point of view is taken up in a recent paper by Pareigis [@Par:ten], in which functors between categories of entwined modules are studied, conditions for the factorisation property are derived and these are then suggested as the definition of morphisms between entwining structures. The resulting notion of morphisms of entwining structures is different from the one introduced earlier in [@Brz:mod]. Since any entwining structure gives rise to a coring such that the entwined modules can be identified with its comodules (cf. [@Brz:str]), it is natural to look at the results of [@Par:ten] from the coring point of view. This is the aim of the present note in which, rather than changing the established notion of a morphism of corings (cf. [@Gom:sep], [@BrzWis:cor Section 24]), we introduce the notion of an [*extension of corings*]{} or a [*coring extension*]{} and show that such extensions arise from and – provided they satisfy a suitable purity condition (for example in the case of corings associated to entwining structures) – give rise to $k$-additive functors with a (suitable) factorisation property.
We work over a commutative associative ring $k$ with a unit. All algebras are over $k$, associative and with a unit. The symbol $\otimes$ between $k$-modules and $k$-module maps means tensor product over $k$. As a rule we do not decorate $\otimes$ between elements, unless there is a danger of confusion. For a $k$-algebra $A$, the category of right $A$-modules and right $A$-linear maps is denoted by $\cM_A$. The product map in $A$ is denoted by $\mu_A:A\ot A\to A$ and the unit (either as an element of $A$ or as a $k$-linear map $k\to A$) is denoted by $1_A$. Given a $k$-algebra $A$, coproduct in an $A$-coring ${{\mathcal C}}$ is denoted by $\DC :{{\mathcal C}}\to {{\mathcal C}}\ot_A{{\mathcal C}}$, and the counit is denoted by $\eC:{{\mathcal C}}\to A$. We use the Sweedler sigma notation, i.e., for all $c\in {{\mathcal C}}$, $$\DC(c) = \sum c\sw1\ot c\sw 2, \qquad (\DC\ot_A{{\mathcal C}})\circ\DC(c) =
({{\mathcal C}}\ot_A\DC)\circ\DC(c) = \sum c\sw1\ot c\sw 2\ot c\sw 3,$$ etc. For an $A$-coring ${{\mathcal C}}$, the [*left dual ring*]{} is defined as a $k$-module ${}^*{{\mathcal C}}= \lhom A {{\mathcal C}}A$ with the unit $\eC$ and the product, for all $f,g\in {}^*{{\mathcal C}}$, $c\in {{\mathcal C}}$, $f*g(c) = \sum g(c\sw 1f(c\sw 2))$.
The category of right ${{\mathcal C}}$-comodules and right ${{\mathcal C}}$-colinear maps is denoted by $\cM^{{\mathcal C}}$. For a right ${{\mathcal C}}$-comodule $M$, $\varrho^M:M\to M\ot_A{{\mathcal C}}$ denotes a coaction. Recall that $\cM^{{\mathcal C}}$ is built upon the category of right $A$-modules, in the sense that every right ${{\mathcal C}}$-comodule is a right $A$-module, coactions and morphisms are right $A$-linear maps (with additional compatibility conditions). On elements, $\varrho^M$ is denoted by the Sweedler notation $\varrho^M(m) = \sum m\sw 0\ot m\sw 1$ (but see an exception in the proof of Theorem \[thm.main\]). Similar notational conventions apply to coalgebras and their comodules. A detailed account of the theory of corings and comodules can be found in [@BrzWis:cor].
Extensions of corings and factorisable functors {#sec.main}
===============================================
Recall that, given an $A$-coring ${{\mathcal C}}$ and a $B$-coring $\cD$, a $({{\mathcal C}},\cD)$-bicomodule is a left ${{\mathcal C}}$-comodule that is at the same time a right $\cD$-comodule with ${{\mathcal C}}$-colinear $\cD$-coaction. The ${{\mathcal C}}$-colinearity of $\cD$-coaction is equivalent to $\cD$-colinearity of ${{\mathcal C}}$-coaction.
\[def.exten\] Let $A$ and $B$ be $k$-algebras. A $B$-coring $\cD$ is called a [*right extension*]{} of an $A$-coring ${{\mathcal C}}$ provided ${{\mathcal C}}$ is a $({{\mathcal C}},\cD)$-bicomodule with the left regular coaction $\DC$.
For example, if ${{\mathcal C}}$ and $\cD$ are $A$-corings and $\gamma:{{\mathcal C}}\to\cD$ is an $A$-coring morphism, then ${{\mathcal C}}$ is a $({{\mathcal C}},\cD)$-bicomodule with the left regular coaction $\DC$ and the right coaction $\varrho^{{\mathcal C}}=({{\mathcal C}}\ot_A\gamma)\circ\DC$. Thus any $A$-coring morphism gives rise to a coring extension.
Definition \[def.exten\] implies in particular that if $\cD$ is a right extension of ${{\mathcal C}}$, then necessarily ${{\mathcal C}}$ is a right $B$-module and $\DC$ is a right $B$-linear map. This leads to the following
\[def.mea\] Let $A$ and $B$ be $k$-algebras. An $A$-coring ${{\mathcal C}}$ is said to [*measure $B$ to $A$*]{} if there exists a left $A$-linear map $\nu:{{\mathcal C}}\ot B\to A$ rendering commutative the following diagrams:
$ \xymatrix{{{\mathcal C}}\ar[rd]_{\eC}\ar[rr]^{{{\mathcal C}}\ot 1_B}& & {{\mathcal C}}\ot B \ar[ld]^{\nu}\\
& A, &}$
$ \xymatrix{{{\mathcal C}}\ot B\ot B\ar[d]_{\DC\ot B\ot B}\ar[rr]^{{{\mathcal C}}\ot \mu_B}&& {{\mathcal C}}\ot B\ar[rr]^\nu&& A \\
{{\mathcal C}}\ot_A{{\mathcal C}}\ot B\ot B\ar[rr]^{{{\mathcal C}}\ot_A\nu\ot B}&& {{\mathcal C}}\ot_AA\ot B\ar[rr]^\simeq &&{{\mathcal C}}\ot B\ar[u]_{\nu}.}$
The map $\nu$ is called a [*${{\mathcal C}}$-measuring*]{} of $B$ to $A$.
\[prop.equiv\] Let ${{\mathcal C}}$ be an $A$-coring and $B$ an algebra. ${{\mathcal C}}$-measurings of $B$ to $A$ are in bijective correspondence with algebra maps $B\to {}^*{{\mathcal C}}$.
There is a bijective correspondence between left $A$-module maps $\nu: {{\mathcal C}}\ot B\to A$ and $k$-linear maps $\chi:B\to {}^*{{\mathcal C}}$ provided by the hom-tensor isomorphism $$\lhom A {{{\mathcal C}}\ot B} A\simeq \hom k B{ \lhom A{{\mathcal C}}A} = \hom k B {{}^*{{\mathcal C}}}.$$ Explicitly, for all $b\in B$, $c\in {{\mathcal C}}$, $\nu(c\ot b)= \chi(b)(c)$. Since the counit $\eC$ of ${{\mathcal C}}$ is the unit in ${}^*{{\mathcal C}}$, the map $\chi$ is unital if and only if $\nu(c\ot 1_B) = \chi(1_B)(c) = \eC(c)$. Thus the unitality of $\chi$ is equivalent to the commutativity of the diagram (a) in Definition \[def.mea\] for $\nu$. Second, the multiplicativity of $\chi$ means that, for all $b,b'\in B$, $\chi(bb') = \chi(b)*\chi(b')$, i.e., for all $c\in {{\mathcal C}}$, $$\chi(bb')(c) = \sum \chi(b')(c\sw 1\chi(b)(c\sw 2)).$$ Therefore, $\chi$ is a multiplicative map if and only if $$\begin{aligned}
\nu (c\ot bb') &=& \chi(bb')(c) = \sum \chi(b')(c\sw 1\chi(b)(c\sw 2))\\
& =& \sum \nu(c\sw 1\chi(b)(c\sw 2)\ot b') =
\sum \nu (c\sw 1 \nu(c\sw 2 \ot b)\ot b'),\end{aligned}$$ i.e., the diagram (b) in Definition \[def.mea\] is commutative.
In view of Proposition \[prop.equiv\], any [*$B^{op}|A$-coring*]{} in the sense of [@Tak:mor Definition 3.5], any [*right rational pairing*]{} of corings in the sense of [@ElKGom:sem] or a [*measuring left $A$-pairing*]{} of [@Abu:rat] are examples of a ${{\mathcal C}}$-measuring. We illustrate the notion of a ${{\mathcal C}}$-measuring with a number of additional examples.
\[ex.measure\] \
An algebra $A$, viewed as a trivial $A$-coring, measures $B$ to $A$ if and only if there is an algebra map $B\to A$.
Given algebras $A$ and $B$, let $\Sigma$ be a $(B,A)$-bimodule that is finitely generated and projective as a right $A$-module and let ${{\mathcal C}}=\Sigma^*\ot_B\Sigma$ be the corresponding comatrix $A$-coring (cf. [@KaoGom:com]). Then ${{\mathcal C}}$-measurings of $B$ to $A$ are in bijective correspondence with right $B$-module structures on $\Sigma$ that make $\Sigma$ a $(B,B)$-bimodule.
Given an algebra map $\iota: B\to A$, take ${{\mathcal C}}=A\ot_B A$ the canonical Sweedler coring. Fix a left $B$-module structure on $A$ provided by the map $\iota$, i.e., $ba := \iota(b)a$. Then ${{\mathcal C}}$-measurings of $B$ to $A$ are in bijective correspondence with right $B$-module structures on $A$ that make $A$ a $(B,B)$-bimodule.
Let $A$ be a $k$-algebra and $C$ be a $k$-coalgebra with coproduct $\Delta_C$ and counit $\eps_C$. If ${{\mathcal C}}=A\ot C$ is the coring associated to an entwining structure $(A,C,\psi)$ then ${{\mathcal C}}$-measurings of $B$ to $A$ are in bijective correspondence with [*entwined measurings*]{} in the sense of [@Par:ten Remark 2.2], i.e., with $k$-linear maps $f: C\ot B\to A$ making the following diagrams: $$\xymatrix{C\ar[d]_{\eps_C}\ar[rr]^{C\ot 1_B}& & C\ot B \ar[d]^{f}\\
k\ar[rr]^{1_A}& &A}$$ and $$\xymatrix{C\ot B\ot B\ar[d]_{\Delta_C\ot B\ot B}\ar[rrr]^{C\ot \mu_B}&&& C\ot B\ar[rrr]^f&&& A \\
C\ot C\ot B\ot B\ar[rr]^{C\ot f\ot B}&& C\ot A\ot B\ar[rr]^{\psi\ot B} &&A\ot C\ot B\ar[rr]^{A\ot f}&&A\ot A \ar[u]_{\mu_A}}$$ commute.
[*Check.*]{}(1) This follows immediately from Proposition \[prop.equiv\], since ${}^*A \simeq A$ as $k$-algebras.
\(2) Recall from [@KaoGom:com] that ${{\mathcal C}}= \Sigma^*\ot_B\Sigma$ is an $A$-coring with coproduct and counit, for all $s\in \Sigma$, $s^*\in \Sigma^*$, $$\DC(s^*\ot_B s) = \sum_{i\in I} s^*\ot_B
e_i\ot_A e_i^*\ot_B s, \qquad \eC(s^*\ot_B s) = s^*(s),$$ where $\{e_i\in \Sigma, e^*_i\in \Sigma^*\}_{i\in I}$ is a finite dual basis of $\Sigma_A$. Recall also that ${}^*{{\mathcal C}}\simeq \lend B\Sigma^{op}$, where the product in $\lend B\Sigma$ is given by $fg(s) = f(g(s))$. Therefore, by Proposition \[prop.equiv\], ${{\mathcal C}}$-measurings are in bijective correspondence with anti-algebra maps $B\to \lend B\Sigma$, i.e., with right $B$-module structures on $\Sigma$ such that $\Sigma$ is a $(B,B)$-bimodule.
\(3) This is a special case of (2), simply take $\Sigma = A$ and view $A$ as a left $B$-module via the map $\iota$.
\(4) Recall that an entwining structure consists of an algebra $(A,\mu_A, 1_A)$, a coalgebra $(C,\Delta_C,\eps_C)$ and a $k$-linear map $\psi: C\ot A\to A\ot C$ satisfying a number of conditions (cf., e.g., [@BrzWis:cor Section 32]). In this case, ${{\mathcal C}}= A\ot C$ is an $A$-bimodule via $a(a'\ot c)a'' = aa'\psi(c\ot a'')$ and it has a coproduct and counit $\DC(a\ot c) = a\ot \Delta_C(c)$, $\eC(a\ot c) =a\eps_C(c)$. In the view of the isomorphism $\lhom A {A\ot C\ot B} A \simeq \hom k{C\ot B} A $ any ${{\mathcal C}}$-measuring $\nu: {{\mathcal C}}\ot B\to A$ corresponds to a $k$-linear map $f:C\ot B\to A$, via $\nu(a\ot c\ot b) = af(c\ot b)$. With the help of this identification one immediately checks that the diagram (a) in Definition \[def.mea\] for $\nu$ is equivalent to the first of diagrams in (3) for $f$. As to the second pair of diagrams, introduce the explicit notation $\psi(c\ot a) = \sum_\alpha a_\alpha\ot c^\alpha$, take any $c\in C$, $b,b'\in B$, use the diagram (b) in Definition \[def.mea\] and the definition of the right $A$-multiplication on ${{\mathcal C}}$ to compute $$\begin{aligned}
f(c\ot bb') &=& \nu(1_A\ot c\ot bb') = \sum \nu((1_A\ot c\sw 1)\nu(1_A\ot c\sw 2\ot b)\ot b')\\
&=& \sum_\alpha \nu (f(c\sw 2\ot b)_\alpha\ot c\sw 1^\alpha\ot b') = \sum_\alpha f(c\sw 2\ot b)_\alpha f(c\sw 1^\alpha\ot b').\end{aligned}$$ This is exactly the contents of the second of the diagrams in (3). Similarly one proves that if $f$ makes this diagram commutative, then also $\nu$ renders commutative the diagram (b) in Definition \[def.mea\]. In particular, if in Example \[ex.measure\] (4) the trivial entwining $\psi: C\ot A\to A\ot C$, $c\ot a\mapsto a\ot c$ is taken, then ${{\mathcal C}}= A\ot C$ measures $B$ to $A$ if and only if $C$ measures $B$ to $A$ in the sense of Sweedler [@Swe:Hop p. 138] (this justifies the choice of the name). Also, the combination of Example \[ex.measure\] (4) and Proposition \[prop.equiv\] leads to an equivalent description of entwined measurings as algebra maps $B\to \#_\psi(C,A)$, where $\#_\psi(C,A)$ is a $\psi$-twisted convolution algebra defined as a $k$-module $\hom k C A$ with the unit $\eps_C$ and with the product, for all $f,g\in \hom k C A$ and $c\in C$, $$(f \#_\psi g)(c) = \sum_\alpha f(c\sw 2)_\alpha g(c\sw 1^\alpha).$$ This follows immediately from the fact that $\#_\psi(C,A)$ is isomorphic to the left dual ring of the coring ${{\mathcal C}}= A\ot C$ associated to an entwining structure $(A,C,\psi)$.
Examples \[ex.measure\] indicate that the notion of a coring measuring can be understood as one that unifies the notions of an algebra map, a bimodule structure and entwined measuring. The relationship between a measuring and a coring extension is revealed in the following
\[lemma.mea\] Let $A$, $B$ be $k$-algebras and let ${{\mathcal C}}$ be an $A$-coring. Then the following statements are equivalent:
there exists a right $B$-module structure on ${{\mathcal C}}$ such that ${{\mathcal C}}$ is an $(A,B)$-bimodule and the coproduct $\DC$ is a $B$-linear map;
${{\mathcal C}}$ measures $B$ to $A$.
\(1) ${\Rightarrow}$ (2) Suppose that ${{\mathcal C}}$ is an $(A,B)$-bimodule with a $B$-linear coproduct $\DC$, and let $\varrho_{{\mathcal C}}: {{\mathcal C}}\ot B\to {{\mathcal C}}$ be the right $B$-multiplication. Define $\nu = \eC\circ\varrho_{{\mathcal C}}: {{\mathcal C}}\ot B\to A$. The condition (a) in Definition \[def.mea\] for $\nu$ follows then from the commutative diagram $$\xymatrix{{{\mathcal C}}\ar[d]_\eC\ar@{=}[rrd]\ar[rr]^{{{\mathcal C}}\ot 1_B}& & {{\mathcal C}}\ot B \ar[d]^{\varrho_{{\mathcal C}}}\\
A& & {{\mathcal C}}\ar[ll]^\eC ,}$$ in which the right upper triangle is commutative by the unitality of the multiplication $\varrho_{{\mathcal C}}$. Since $\DC$ is a right $B$-module morphism, $$\xymatrix{{{\mathcal C}}\ot B\ar[d]_{\DC\ot B}\ar[rr]^{\varrho_{{\mathcal C}}} & &{{\mathcal C}}\ar[d]^{\DC}\ar[rrd]^{\simeq} &&\\
{{\mathcal C}}\ot_A{{\mathcal C}}\ot B\ar[rr]^{{{\mathcal C}}\ot_A\varrho_{{\mathcal C}}}&&{{\mathcal C}}\ot_A{{\mathcal C}}\ar[rr]^{{{\mathcal C}}\ot_A\eC} && {{\mathcal C}}\ot_A A}$$ is a commutative diagram. The right-hand triangle is simply the counit axiom. Note that the composition of maps in the bottow row equals ${{\mathcal C}}\ot_A\nu$. Since $\varrho_{{\mathcal C}}$ is an associative multiplication and the above diagram commutes, we obtain the following commutative diagram $$\xymatrix{{{\mathcal C}}\ot_A{{\mathcal C}}\ot B\ot B\ar[d]_{{{\mathcal C}}\ot_A\nu\ot B} && {{\mathcal C}}\ot B\ot B\ar[d]_{\varrho_{{\mathcal C}}\ot B}\ar[ll]_{\DC\ot B\ot B}\ar[rr]^{{{\mathcal C}}\ot \mu_B}&& {{\mathcal C}}\ot B\ar[rr]^{\varrho_{{\mathcal C}}}&& {{\mathcal C}}\ar@{=}[dll]\ar[d]^{\eC} \\
{{\mathcal C}}\ot_AA\ot B\ar[rr]^\simeq &&{{\mathcal C}}\ot B\ar[rr]^{\varrho_{{\mathcal C}}}&&{{\mathcal C}}\ar[rr]^{\eC}&& A.}$$ The outer rectangle in the above diagram is equivalent to the condition (b) in Definition \[def.mea\] for $\nu= \eC\circ\varrho_{{\mathcal C}}$. Thus we conclude that $\nu$ is a ${{\mathcal C}}$-measuring of $B$ to $A$ as required.
\(2) ${\Rightarrow}$ (1) Given a ${{\mathcal C}}$-measuring $\nu$, define $\varrho_{{\mathcal C}}: {{\mathcal C}}\ot B\to {{\mathcal C}}$ by $c\ot b\mapsto \sum c\sw 1\nu(c\sw 2\ot b)$. Then condition (a) in Definition \[def.mea\] for $\nu$ implies that, for all $c\in {{\mathcal C}}$, $$\varrho_{{\mathcal C}}(c\ot 1_B) = \sum c\sw 1\nu(c\sw 2\ot 1_B) = \sum c\sw 1\eC(c\sw 2) = c.$$ Furthermore, the use of condition (b) (in the second equality below), the definition of $\varrho_{{\mathcal C}}$ in terms of $\nu$, and the right $A$-linearity of $\DC$ give, for all $c\in{{\mathcal C}}$, $b,b'\in B$, $$\begin{aligned}
\varrho_{{\mathcal C}}(c\ot bb') &=& \sum c\sw 1\nu(c\sw 2\ot bb') = \sum c\sw 1\nu(c\sw 2\nu(c\sw 3\ot b)\ot b')\\
& = & \sum \varrho_{{\mathcal C}}(c\sw 1\nu(c\sw 2\ot b)\ot b') = \varrho_{{\mathcal C}}(\varrho_{{\mathcal C}}(c\ot b)\ot b'). \end{aligned}$$ Thus ${{\mathcal C}}$ is a right $B$-module with the multiplication $\varrho_{{\mathcal C}}$. The map $\varrho_{{\mathcal C}}$ is a composition of left $A$-linear maps, hence a left $A$-linear map, i.e., ${{\mathcal C}}$ is an $(A,B)$-bimodule. Finally, since the coproduct is a coassociative right $A$-linear map, for all $b\in B$ and $c\in{{\mathcal C}}$, $$\DC(\varrho_{{\mathcal C}}(c\ot b)) = \sum\DC(c\sw 1\nu(c\sw 2\ot b)) = \sum c\sw 1\ot c\sw 2\nu(c\sw 3\ot b) = \sum c\sw 1\ot \varrho_{{\mathcal C}}(c\sw 2\ot b).$$ This means that the coproduct $\DC$ is a right $B$-linear map as required.
The main result of this note is contained in the following
\[thm.main\] Let ${{\mathcal C}}$ be an $A$-coring and $\cD$ be a $B$-coring.
If there exists a $k$-additive functor $F:\cM^{{\mathcal C}}\to \cM^\cD$ with a factorisation property $$\xymatrix{\cM^{{\mathcal C}}\ar[rd]_{U^{{\mathcal C}}}\ar[rr]^F& & \cM^\cD \ar[ld]^{U^\cD}\\
& \cM_k, &}$$ where $U^{{\mathcal C}}$, $U^\cD$ are forgetful functors, then $\cD$ is a right extension of ${{\mathcal C}}$.
Let $\cD$ be a right extension of ${{\mathcal C}}$ such that, for all ${{\mathcal C}}$-comodules $(N,\varrho^N)$, the right $B$-module map $\varrho^N \ot_A {{\mathcal C}}- N\ot_A\DC$ is $\cD\ot_B \cD$-pure. Then there exists a $k$-additive functor $F:\cM^{{\mathcal C}}\to \cM^\cD$ with a factorisation property as in (1).
\(1) Let $F:\cM^{{\mathcal C}}\to \cM^\cD$ be a $k$-additive functor that factorises through forgetful functors $U^{{\mathcal C}}$ and $U^\cD$. The factorisation property means that for any $M\in \cM^{{\mathcal C}}$, $F(M) = M$ as $k$-modules. Similarly, for any morphism $f$ in $\cM^{{\mathcal C}}$, $F(f) =f$ as $k$-linear maps. This implies that for any right ${{\mathcal C}}$-comodule $M$ there exist an action $\varrho_M :M\ot B\to M$ and a coaction $\varrho^M: M\to M\ot_B\cD$, and any $k$-linear map $f$ that is a morphism in $\cM^{{\mathcal C}}$ is also a morphism in $\cM^\cD$ (the functoriality of action and coaction). In particular, ${{\mathcal C}}$ is a right ${{\mathcal C}}$-comodule with the regular coaction $\DC$, hence ${{\mathcal C}}$ is a right $B$-module and there exists a right $\cD$-coaction on ${{\mathcal C}}$, $\varrho^{{\mathcal C}}:{{\mathcal C}}\to {{\mathcal C}}\ot_B\cD$. In addition, $\DC$ is a left $A$-module map. Equivalently, for any $a\in A$, the right ${{\mathcal C}}$-colinear map $\ell_a:{{\mathcal C}}\to {{\mathcal C}}$, $c\mapsto ac$ is a morphism in $\cM^{{\mathcal C}}$. Thus the $k$-linear map $\ell_a$ is a morphism in $\cM^\cD$, i.e., $\varrho^{{\mathcal C}}$ is a left $A$-module map. Furthermore, ${{\mathcal C}}\ot_A{{\mathcal C}}$ is a right ${{\mathcal C}}$-comodule with the coaction ${{\mathcal C}}\ot_A\DC$, hence it is a right $B$-module and there exists a right $\cD$-coaction, $\varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}:{{\mathcal C}}\ot_A{{\mathcal C}}\to {{\mathcal C}}\ot_A{{\mathcal C}}\ot_B\cD$. For any $c\in {{\mathcal C}}$, consider a right ${{\mathcal C}}$-comodule map $\ell^c: {{\mathcal C}}\to {{\mathcal C}}\ot_A{{\mathcal C}}$, $c'\mapsto c\ot c'$. Since $\DC$ is a right ${{\mathcal C}}$-comodule map too, the functoriality of $\cD$-coactions implies that, for all $c\in {{\mathcal C}}$, $$(\ell^c\ot_B\cD)\circ\varrho^{{\mathcal C}}= \varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}\circ\ell^c, \qquad \varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}\circ\DC = (\DC\ot_B\cD)\circ\varrho^{{\mathcal C}},$$ in $\cM_k$. Putting these two together we obtain, $$\begin{aligned}
\sum c\sw 1\ot \varrho^{{\mathcal C}}(c\sw 2) &=& \sum (\ell^{c\sw 1}\ot_B\cD)\circ \varrho^{{\mathcal C}}(c\sw 2) = \sum \varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}\circ\ell^{c\sw 1}(c\sw 2)\\
&=& \sum \varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}(c\sw 1 \ot c\sw 2) = \varrho^{{{\mathcal C}}\ot_A{{\mathcal C}}}\circ\DC(c)\\
&=& (\DC\ot_B\cD)\circ\varrho^{{\mathcal C}}(c).\end{aligned}$$ This means that the coaction $\varrho^{{\mathcal C}}$ is a left ${{\mathcal C}}$-comodule map, hence ${{\mathcal C}}$ is a $({{\mathcal C}},\cD)$-bicomodule, i.e., $\cD$ is a coring extension of ${{\mathcal C}}$.
\(2) This is contained in [@BrzWis:cor 22.3, Erratum]. In detail, suppose that $\cD$ is a coring extension of ${{\mathcal C}}$ and let $\nu : {{\mathcal C}}\ot B\to A$ be the measuring corresponding to the right $B$-multiplication on ${{\mathcal C}}$ as in Lemma \[lemma.mea\]. Write $\sigma: {{\mathcal C}}\to {{\mathcal C}}\ot_B\cD$ for the right $\cD$-coaction on ${{\mathcal C}}$. Define a $k$-linear functor $F:\cM^{{\mathcal C}}\to\cM^\cD$ as follows.
Take any right ${{\mathcal C}}$-comodule $M$ and define a map $\varrho_M: M\ot B\to M$, $m\ot b\mapsto \sum m\sw 0\nu(m\sw 1\ot b)$. Following the same steps as in the proof (2) ${\Rightarrow}$ (1) of Lemma \[lemma.mea\] one easily verifies that $M$ is a right $B$-module with multiplication $\varrho_M$. We write $m. b := \varrho_M(m\ot b)$. Furthermore, if $f:M\to N$ is a morphism in $\cM^{{\mathcal C}}$, then for all $m\in M$, $b\in B$, $$\begin{aligned}
f(m.b) &=& \sum f(m\sw 0\nu(m\sw 1\ot b)) = \sum f(m\sw 0)\nu(m\sw 1\ot b)\\
&=& \sum f(m)\sw 0\nu(f(m)\sw 1\ot b) = f(m).b.\end{aligned}$$ The second equality follows from the $A$-linearity of $f$, while the third one is a consequence of the fact that $f$ is a ${{\mathcal C}}$-comodule map. The first and last equalities follow from the definition of the $B$-multiplication $\varrho_M$ on $M$. Therefore, $f$ is a right $B$-linear map, and thus we have constructed a functor $\cM^{{\mathcal C}}\to \cM_B$. Now we need to define a $\cD$-coaction on any ${{\mathcal C}}$-comodule.
Start with the right ${{\mathcal C}}$-comodule isomorphism $M\simeq M\Box_{{\mathcal C}}{{\mathcal C}}$ (cf. [@BrzWis:cor 22.4]) provided by the right ${{\mathcal C}}$-coaction on $M$. Here $\Box_{{\mathcal C}}$ denotes the cotensor product over ${{\mathcal C}}$. By applying the above functor $\cM^{{\mathcal C}}\to \cM_B$ we obtain a right $B$-module isomorphism and we can construct a map $$\varrho^M : \xymatrix@R=2pt{M\ar[r]^(.4){\simeq}& M\Box_{{\mathcal C}}{{\mathcal C}}\ar[rr]^(.35){M\Box_{{\mathcal C}}\sigma} &&M\Box_{{\mathcal C}}({{\mathcal C}}\ot_B\cD)\simeq M\ot_B\cD\\
m\ar@{|->}[rrr]&&&\sum m\sw 0\eC(m\sw 1\su 0)\ot m\sw 1\su 1,}$$ where $\sum m\sw 0\ot m\sw 1\in M\ot_A{{\mathcal C}}$ denotes the ${{\mathcal C}}$-coaction on $M$, while $\sigma(c) = \sum c\su 0\ot c\su 1\in {{\mathcal C}}\ot_B\cD$ denotes the $\cD$-coaction on ${{\mathcal C}}$. Note that $\varrho^M$ is well-defined by the purity assumption. We claim that $\varrho^M$ is a right $\cD$-coaction. First, $\varrho^M$ is a right $B$-module map as a composition of $B$-module maps. Note that, for all $m\in M$, $$\sum m\sw 0\eC(m\sw 1\su 0)\eD(m\sw 1\su 1) = \sum m\sw 0\eC(m\sw 1) = m,$$ so that $\varrho^M$ is a counital map. It remains to check the coassociativity of $\varrho^M$. This is done by a rather lengthy but straightforward calculation, the details of which are displayed below. Take any $m\in M$ and compute $$\begin{aligned}
(\varrho^M&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ot_B\cD)\circ\varrho^M(m) = \sum \varrho^M (m\sw 0\eC(m\sw 1\su 0))\ot m\sw 1\su 1 \\
&=& \sum m\sw 0\sw 0\eC((m\sw 0\sw 1\eC(m\sw 1\su 0))\su 0)\ot (m\sw 0\sw 1\eC(m\sw 1\su 0))\su 1\ot m\sw 1\su 1\\
&=& \sum m\sw 0\eC((m\sw 1\sw 1\eC(m\sw 1\sw 2\su 0))\su 0)\ot (m\sw 1\sw 1\eC(m\sw 1\sw 2\su 0))\su 1\ot m\sw 1\sw 2\su 1\\
&=& \sum m\sw 0\eC((m\sw 1\su 0\sw 1\eC(m\sw 1\su 0\sw 2))\su 0)\ot (m\sw 1\su 0\sw 1\eC(m\sw 1\su 0\sw 2))\su 1\ot m\sw 1\su 1\\
&=& \sum m\sw 0\eC(m\sw 1\su 0\su 0)\ot m\sw 1\su 0\su 1\ot m\sw 1\su 1\\
&=& \sum m\sw 0\eC(m\sw 1\su 0)\ot m\sw 1\su 1\sw 1\ot m\sw 1\su 1\sw 2\\
&=& ({{\mathcal C}}\ot_B\DD)\circ\varrho^M(m).\end{aligned}$$ The fourth equality is a consequence of the fact that ${{\mathcal C}}$ is a $({{\mathcal C}},\cD)$-bicomodule and the purity assumption is used to derive the penultimate equality. Thus $\varrho^M$ is a right $\cD$-coaction.
We already know that if $f:M\to N$ is a morphism in $\cM^{{\mathcal C}}$, then $f$ is a right $B$-module map. Take any $m\in M$ and compute $$\begin{aligned}
\varrho^N(f(m)) &=& \sum f(m)\sw 0\eC(f(m)\sw 1\su 0)\ot f(m)\sw 1\su 1 = \sum f (m\sw 0)\eC(m\sw 1\su 0)\ot m\sw 1\su 1\\
&=& \sum f (m\sw 0\eC(m\sw 1\su 0))\ot m\sw 1\su 1 = (f\ot_B\cD)\circ\varrho^M(m),\end{aligned}$$ where the second equality follows from the ${{\mathcal C}}$-colinearity and the third one from the $A$-linearity of $f$. Thus $f$ is a morphism of right $\cD$-comodules. Put together, all this means that we have constructed a $k$-additive functor $F:\cM^{{\mathcal C}}\to \cM^\cD$. It is obvious from this construction that $F$ factorises through the forgetful functors $U^{{\mathcal C}}$ and $U^\cD$ as required.
In view of Lemma \[lemma.mea\], a trivial $B$-coring $B$ is a right extension of ${{\mathcal C}}$ if and only if ${{\mathcal C}}$ measures $B$ to $A$. Thus Theorem \[thm.main\] leads immediately to the following characterisation of measurings in terms of functors with a factorisation property.
Let $A$ and $B$ be $k$-algebras and let ${{\mathcal C}}$ be an $A$-coring. The following statements are equivalent:
${{\mathcal C}}$ measures $B$ to $A$;
there exists a $k$-additive functor $F:\cM^{{\mathcal C}}\to \cM_B$ with a factorisation property $$\xymatrix{\cM^{{\mathcal C}}\ar[rd]_{U^{{\mathcal C}}}\ar[rr]^F& & \cM_B \ar[ld]^{U_B}\\
& \cM_k, &}$$ where $U^{{\mathcal C}}$, $U_B$ are forgetful functors.
If ${{\mathcal C}}$ and $\cD$ are corings corresponding to entwining structures $(A,C,\psi)$, $(B,D,\Psi)$ then Theorem \[thm.main\] implies [@Par:ten Theorem 2.3]. As many general results about corings, Theorem \[thm.main\] finds an application in non-commutative descent theory.
Given an algebra extension $\iota: B\to A$, the category of right comodules of the associated Sweedler coring ${{\mathcal C}}= A\ot_B A$ is isomorphic to the category of [*(right) descent data*]{} $\Desc (A| B)$ (cf. [@BrzWis:cor 25.4]). The objects in $\Desc (A|B)$ are pairs $(M,f)$, where $M$ is a right $A$-module and $f: M\to M\ot_B A$ is a right $A$-module map rendering commutative the following diagrams $$\xymatrix{M\ar[r]^{f}\ar@{=}[dr] & M\ot_B A \ar[d]^{\varrho_{M|B}}\\
&M,} \quad
\xymatrix{M\ar[r]^{f}\ar[dd]_f& M\ot_B A\ar[rd]^{f\ot_B A} &\\
&& M\ot_B A\ot_B A.\\
M\ot_B A\ar[r]^\simeq& M\ot_BB\ot_B A \ar[ru]_{M\ot_B\iota\ot_B A}&}$$ Here $\varrho_{M|B}: M\ot_B A\to M$ is the factorised (through $M\ot A\to M\ot_B A$) $A$-multiplication on $M$. Using this identification of right comodules of the Sweedler coring $A\ot_BA$ with (right) descent data of the ring extension $B\to A$ we can thus derive the following corollary of Theorem \[thm.main\].
\[cor.descent\] Let $A$, $B$, $D$ be $k$-algebras and let $\iota_B :D\to B$ and $\iota_A: B\to A$ be algebra maps. Then the following statements are equivalent.
There exists a $k$-additive functor $F:\Desc (A|B)\to\Desc (B|D)$ with a factorisation property $$\xymatrix{\Desc (A|B)\ar[rd]_{U^{A|B}}\ar[rr]^F& & \Desc (B|D) \ar[ld]^{U^{B|D}}\\
& \cM_k, &}$$ where $U^{A|B}$, $U^{B|D}$ are forgetful functors.
View $A$ as a left $B$-module via the map $\iota_A$. There exist:
a left $B$-linear right $B$-multiplication $\varrho_A:A\ot B\to A$;
a $(B,B)$-bimodule map $\varphi:A\to A\ot_B A\ot_D B$ ($A$ is a right $B$-module via $\varrho_A$) rendering commutative the following three diagrams:\
(a) $
\xymatrix{A\ar[rr]^{\simeq}\ar[d]_\varphi & & B\ot_B A \ar[d]^{\iota_A\ot_B A}\\
A\ot_B A\ot_D B\ar[rr]^{A\ot_B\varrho_{A|D}} && A\ot_B A,}$\
(b) $
\xymatrix{A\ar[r]^{\varphi}\ar[d]_\varphi & A\ot_B A\ot_D B\ar[rr]^{A\ot_B\varphi\ot_DB}& & A\ot_B A\ot_B A\ot_D B\ot_D B \ar[d]_{\mu_{A|B}\ot_B A\ot_D B\ot_D B}\\
A\ot_B A\ot_D B\ar[r]^{\simeq} & A\ot_B A\ot_D D\ot_D B\ar[rr]^{A\ot_B A\ot_D\iota_B\ot_D B} && A\ot_B A\ot_D B\ot_D B,}$\
(c) $
\xymatrix{A\ar[r]^{\simeq}\ar[d]_\varphi & B\ot_B A \ar[rr]^{\iota_A\ot_B \varphi} && A\ot_BA\ot_B A\ot_D B\\
A\ot_B A\ot_D B\ar[rrr]^{\simeq} &&& A\ot_B B\ot_B A\ot_D B\ar[u]_{A\ot_B\iota_A\ot_BA\ot_D B},}$\
\
where $\varrho_{A|D}: A\ot_D B\to A$ is the factorised (through $A\ot B\to A\ot_D B$) right $B$-multiplication $\varrho_A$ and $\mu_{A|B}: A\ot_B A\to A$ is the factorised (through $A\ot A\to A\ot_BA$) product $\mu_A$ in $A$.
The purity condition in Theorem \[thm.main\] (2) always holds for Sweedler corings, hence part (1) is equivalent to the statement that $A\ot_B A$ is a right coring extension of $B\ot_D B$. The contents of this statement is contained in part (2). We only explain the origin of maps and diagrams in part (2), the details of the proof are left to the reader. By Example \[ex.measure\](3), there must exist a right $B$-multiplication $\varrho_A$ as in (i). Furthermore, $A\ot_BA$ must be a right $B\ot_DB$-comodule, thus there exists a descent datum $(A\ot_BA,f)\in \Desc (B|D)$. The map $f: A\ot_B A\to A\ot_BA\ot_DB$ is a left $A$-module map as it corresponds to a coaction that is left ${{\mathcal C}}$-colinear. As a part of a descent datum, $f$ is right $B$-linear. In view of the isomorphism $\lrhom AB{A\ot_BA}{A\ot_BA\ot_DB}\simeq \lrhom BB{A}{A\ot_BA\ot_DB}$, $f$ can be equivalently given as a $(B,B)$-bimodule map $\varphi: A\to A\ot_BA\ot_DB$, $a'\varphi(a) = f(a'\ot a)$. The diagrams (a) and (b) are the defining diagrams for $f$ as a part of a descent datum written in terms of $\varphi$. The diagram (c) expresses the fact that $f$ is a left $A\ot_B A$-comodule map, as it corresponds to a coaction that makes $A\ot_B A$ into an $(A\ot_B A,B\ot_DB)$-comodule.
A functor obtained as a composition of any two functors between comodule categories that factorise through the forgetful functors also factorises through forgetful functors. The correspondence between such functors and coring extensions leads therefore to a category $\cex_k^r$ in which objects are corings understood as pairs $({{\mathcal C}}\!:\! A)$. Morphisms $({{\mathcal C}}\!:\! A)\to (\cD\!:\! B)$ are [*pure coring extensions*]{}, i.e. pairs $(\varrho_{{{\mathcal C}}}, \varrho^{{{\mathcal C}}})$ where $\varrho_{{{\mathcal C}}}: {{\mathcal C}}\ot B\to {{\mathcal C}}$ is a left ${{\mathcal C}}$-colinear $B$-action and $\varrho^{{{\mathcal C}}} :{{\mathcal C}}\to{{\mathcal C}}\ot_B\cD$ is a left ${{\mathcal C}}$-colinear $\cD$-coaction, such that, for all ${{\mathcal C}}$-comodules $(N,\varrho^N)$, the right $B$-module map $\varrho^N \ot_A {{\mathcal C}}- N\ot_A\DC$ is $\cD\ot_B \cD$-pure. A composition of morphisms $({{\mathcal C}}\!:\! A)\to (\cD\!:\! B)$ and $(\cD\!:\! B)\to (\cE\!:\! D)$ is derived from the composition of corresponding functors and comes out as $$\varrho_{\cD}\bullet\varrho_{{{\mathcal C}}} : \xymatrix@R=2pt{{{\mathcal C}}\ot D\ar[rr]^(.4){\varrho^{{{\mathcal C}}}\ot D}& & {{\mathcal C}}\ot_B\cD\ot D \ar[rr]^(.55){{{\mathcal C}}\ot_B \varrho_{\cD}} &&{{\mathcal C}}\ot_B\cD\ar[rr]^{{{\mathcal C}}\ot_B\eD}&& {{\mathcal C}}\ot_B B \simeq {{\mathcal C}}}$$ and $$\varrho^{\cD}\bullet\varrho^{{{\mathcal C}}} : \xymatrix@R=2pt{{{\mathcal C}}\ar[r]^(.4){\simeq}& {{\mathcal C}}\Box_\cD\cD \ar[rr]^(.35){{{\mathcal C}}\Box_\cD\varrho^{\cD}} &&{{\mathcal C}}\Box_\cD(\cD\ot_D\cE)\simeq {{\mathcal C}}\ot_D\cE,}$$ where the first isomorphism is provided by the coaction $\varrho^{{{\mathcal C}}|\cD}$, while the second one is obtained with the help of the counit in $\cD$ (compare the construction of coaction $\varrho^M$ in the proof of Theorem \[thm.main\] (2)).
Finally, we would like to point out that the results of this note can also be presented for left comodules of a coring thus leading to the notions of a left ${{\mathcal C}}$-measuring and a left coring extension. This is achieved by using the obvious left-right correspondence. Note, however, that a right coring extension is not necessarily a left coring extension, thus the left-right symmetry that exists for characterisation of algebra (or coalgebra) extensions does not exist in the general coring case.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank Laiachi El Kaoutit for pointing out to me the relationship between measurings and pairings, and the Engineering and Physical Sciences Research Council for an Advanced Fellowship. I am also grateful to Gabriella Böhm for spotting a mistake in the original version of this note.
[Bibliography]{}
J.Y. Abuhlail, Rational modules for corings. [*Comm. Algebra*]{} 31: 5793–5840, 2003. T. Brzeziński, On modules associated to coalgebra Galois extensions. [*J. Algebra*]{}, 215: 290–317, 1999.
T. Brzeziński, The structure of corings. Induction functors, Maschke-type theorem, and Frobenius and Galois-type properties. [*Alg. Rep. Theory*]{}, 5: 389–410, 2002.
T. Brzeziński and R. Wisbauer, [*Corings and Comodules*]{}. Cambridge University Press, Cambridge, 2003. Erratum: http://www-maths.swan.ac.uk/staff/tb/Corings.htm
L. El Kaoutit and J. Gómez-Torrecillas, Comatrix corings: Galois corings, descent theory, and a structure theorem for cosemisimple corings. [*Math. Z.*]{}, 244:887–906, 2003.
L. El Kaoutit, J. Gómez-Torrecillas and F. J. Lobillo, Semisimple corings. [*Algebra Colloq.*]{} 11, 427–442 (2004)..
J. Gómez-Torrecillas, Separable functors in corings. [*Int. J. Math. Math. Sci.*]{} 30: 203–225, 2002.
B. Pareigis, Vergessende Funktoren und Ringhomomorphismen. [*Math. Z.*]{} 93: 265–275, 1966.
B. Pareigis, Tensor products and forgetful functors of entwined modules. [*The Pumplün 70 Festschrift*]{}, RWTH Aachen University, Eds A. Krieg, S. Walcher, pp. 1-12, 2003.
M. E. Sweedler, W.A. Benjamin, Inc., New York, 1969
M. Takeuchi, $\sqrt{\rm Morita}$ theory: Formal ring laws and monoidal equivalences of categories of bimodules. [*J. Math. Soc. Japan*]{} 39: 301–336, 1987.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show, that the 2D XY-model with random phase shifts exhibits for low temperature and small disorder a phase with quasi-long-range order, and that the transition to the disordered phase is [*not*]{} reentrant. These results are obtained by heuristic arguments, an analytical renormalization group calculation, and a numerical Migdal-Kadanoff renormalization group treatment. Previous predictions of reentrance are found to fail due to an overestimation of the vortex pair density as a consequence of independent dipole approximations. At positions, where vortex pairs are energetically favored by disorder, their statistics becomes effectively fermionic. The results may have implications for a large number of related models.'
address: |
$^{(a)}$ Institut für Theoretische Physik, Universität zu Köln, D-50937 Köln, Germany\
$^{(b)}$ Landau Institute for Theoretical Physics, Kosygina 2, 117940 Moscow, Russia\
$^{(c)}$ Institute of Physics, 02-668 Warsaw, Poland
author:
- 'Thomas Nattermann$^{(a)}$, Stefan Scheidl$^{(a)}$, Sergey E. Korshunov$^{(b)}$, and Mai Suan Li$^{(c)}$'
date: February 1995
title: 'Absence of Reentrance in the Two-Dimensional XY-Model with Random Phase Shift'
---
= 10000
[2]{}
We reconsider in this paper the 2-dimensional XY-model $${\cal H} = - J\sum_{<i,j>} \cos (\phi_i - \phi_j - A_{ij})$$ with quenched random phase shifts $A_{ij}$ on the bonds, where $i,j$ run over the sites of a square lattice. For simplicity we assume, that the $A_{ij}$ on different bonds are uncorrelated and gaussian distributed with mean zero and variance ${\sigma}$.
Model (1) describes e.g. 2-dimensional XY-magnets with random Dzyaloshinskii-Moriya interaction [@RSN]. Other realizations are given by Josephson-junction arrays with positional disorder [@jja] and model vortex glasses [@mvg]. In particular, in the case of the so-called gauge glass model, one assumes $A_{ij}$ to be uniformly distributed between $0$ and $2\pi$. We expect, that our model with gaussian disorder is equivalent to the gauge glass model when $\sigma \to \infty$.
For vanishing $A_{ij}$ model (1) undergoes a Kosterlitz-Thouless (KT) transition, at which the spin-spin correlation exponent $\eta$ jumps from 1/4 to zero [@KT].
Weak disorder, $\sigma \ll 1$, should not change much this picture. In the spin wave approximation one obtains $\eta$ = $\frac {1}{2\pi} (T/J
+\sigma)$, which remains now finite even at $T = 0$. The features of the KT-transition are essentially preserved, but the transition is shifted to lower temperatures and the jump of $\eta $ at the transition is diminished [@RSN]. The actual transition temperature $T_c(\sigma) \le T_+(\sigma)$ depends on the bare value for the vortex core energy $E_c$, here $T_{\pm }=\frac{\pi}{4}J[1 {\pm
}{(1-8\sigma/\pi)}^{1/2}]$. In the limit $E_c \to \infty $, $
T_c=T_+$.
Strong disorder will suppress the quasi-long-range order of the KT phase [@mvg]. In particular, if $Q = \frac{1}{2\pi}\sum_{<plaq>}A_{ij}$ is of the order one, vortices are generated even at zero temperature. Here $\sum_{<plaq>}$ denotes the sum over the four bonds of an elementary plaquette.
Rubinstein, Shraiman and Nelson (RSN) [@RSN] extended the Coulomb gas description of the KT-transition [@KT] to the presence of randomly frozen dipoles arising from the random phase shifts. Surprisingly, they found a second (reentrant) transition at $T_{re}(\sigma)$ $(\le
T_-(\sigma))$ to a disordered phase at low temperatures (see Fig. 1). $T_{re}(\sigma)$ bends towards higher temperatures for increasing disorder. The two lines $T_{\pm} $ merge at $\sigma _c =
\pi /8$. For $\sigma>\sigma _c$ there is no ordered phase. The precise value of $T_{re}(\sigma)$ depends again on $E_c$. Similar results were obtained in Ref. \[2\].
Korshunov \[5\] has argued, that the intermediate phase in the range $T_{re}(\sigma) < T < T_c(\sigma)$ with quasi-long-range order is probably [*not stable*]{}, if in addition to the screening of Coulomb charges by neutral pairs of charges, considered in \[1\], screening by larger complexes of charges in different replicas are taken into account.
Experiments \[6\] as well as Monte Carlo studies \[7\] indicate no reentrance. Also, Ozeki and Nishimori \[8\] have shown for a general class of random spin systems, which include (1), that the phase boundary between the KT- and the paramagnetic phase is parallel to the temperature axis for [*low T*]{}. Thus they exclude a reentrant transition, provided the intermediate KT phase exists. However, they cannot rule out the possibility, that the KT-phase [*disappears*]{} completely, as suggested in \[5\].
We will argue below, that the reentrant transition is indeed an artefact of the calculation scheme used in \[1\], \[2\] and that the KT-phase is stable at low temperatures with $T_c(\sigma) \to 0$ for $\sigma \to \pi /8$ (see Fig.1). Since the renormalization group (RG) flow equations (4) (see below) derived in \[1\], which give rise to the reentrant behavior, appear as a subset of the more general RG equations for XY-systems with additonal symmetry-breaking \[9\] or random fields \[10\], as well as for solid films with quenched random impurities \[11\], also these systems have to be reconsidered, which we will postpone to forthcoming publications.
For the further discussion it is useful to decompose the Hamiltonian (1) into a spin-wave part ${\cal H}_{sw}$ and a vortex part ${\cal
H}_v$. Since the phase transition is governed by ${\cal H}_v$, we will omit ${\cal H}_{sw}$ completely. In the continuum description, the vortex part can be rewritten in the form \[1\] (for simplicity, we set the lattice constant equal to unity) $$\begin{aligned}
{\cal H}_v &=& - J \pi \sum _{i} m_i \{\sum_{j \not= i}m_j \ln |{\bf r}_i -
{\bf r}_j | +\nonumber \\
&+& 2\int d^2r Q({\bf r}) \ln |{\bf r}-{\bf r}_i|- \frac{E_c}{J\pi}m_i\} .\end{aligned}$$
The integer vortex charges $m_i$ satisfy $\sum _i m_i =0$. $Q({\bf
r})$ is a quenched random charge field, which is related to the phase shift ${\bf A} ({\bf r})$ by $2\pi Q(r) = -\partial_xA_y+
\partial _yA_x$. Here we made the replacement $A_{ij} \rightarrow {\bf A}
({\bf r})$ by going over to the continuum description. Since $$[{\bf A}]_d = 0, \qquad [A_\alpha({\bf r})A_\beta({\bf r'})]_d = \sigma
\delta _{\alpha \beta}\ ({\bf r}-{\bf r'}),$$ where $[...]_d$ denotes the disorder average, the random charges are [*anticorrelated*]{}.
The main result of the work of RSN \[1\] are the RG-flow equations (4a-c) (see also \[2\], \[9\], \[10\]), which describe the change of $J,
\sigma $ and the vortex number density $y$ after eliminating vortex degrees of freedom up to a length scale $e^l$
$$\begin{aligned}
\frac{dJ}{dl}&=& -4{\pi}^3 \frac{J^2}{T} y^2\label{4a}\\
\frac{dy}{dl}&=&(2-\pi \frac{J}{T} +\pi\frac{J^2}{T^2} \sigma)y\label{4b}\\
\frac{d\sigma}{dl}&=&0.\label{4c}\end{aligned}$$
Here we use the convention, that only the exchange constant $J$ is renormalized and the temperature plays merely the role of an unrenormalized parameter. For $\sigma \equiv 0$ the equations (4a), (4b) behave well defined for $T\rightarrow 0$. This becomes more clear, if we rewrite the vortex fugacity as $y=e^{-F_c/T}$, where $F_c$ is the (core) free energy of a single vortex on the scale $e^l$. Then (4b) takes the form ${dF_c}/{dl}=(\pi J - 2T - \pi
\frac{J^2}{T}\sigma )=2(T_+-T)(T-T_-)/T$.
For $\sigma > 0$, the last term on the r.h.s. of (4b) blows up at low $T$, leading to the reentrance transition mentioned above. Whereas for high temperatures the $1/T$ coefficient of the $\sigma$ term is plausible, since thermal fluctuations wipe out the random potential, we do not see a reason that this effect could lead to an [*unlimited*]{} growth of the effective disorder strength at very low temperatures. Clearly, (4b) cannot be valid at zero temperature. Contrary to RSN \[1\], we argue, that the equations (4a), (4b) are valid only for sufficiently high temperatures $T\ge
T^*(\sigma)>T_-(\sigma)$.
An indication for $T^*$ follows from the flow of the vortex entropy $S_c = -{\partial F_c}/{\partial T}$, $ {\partial S_c}/{\partial l}
=2-\pi\frac{J^2}{T^2}\sigma+ \pi(-\frac {\partial J}{\partial
T})(1-2\sigma\frac {J}{T})$. Since ${\partial J}/{\partial T} \le 0$, the entropy is [*reduced*]{} for $T<T^*=2J\sigma$, $\sigma\le \pi/8$, if one goes over to larger length scales. This leads finally to a negative entropy, which we consider as an artefact of the calculation \[1\] (see also \[2\], \[9\], \[10\], \[11\]). The vanishing of the entropy in disorderd systems usually signals a freezing of the system by approaching $T^*$ from high temperatures \[12\]. Similarly, the flow of the vortex energy $E_c=F_c+T S_c$, $
{\partial E_c}/{\partial l} =(1-2\sigma\frac {J}{T}) \pi (
J-T\frac{\partial J}{\partial T})$ leads for $T<T^*$ eventually to negative values of the core energy. Inevitably, this favours multiple occupancy of vortex positions. However, the resulting vortices of higher vorticity $|m|>1$ appear even in the presence of disorder much less likely than those with $|m|=1$: since their energy cost scales as $m^2$ whereas their energy gain scales only as $m$. This effective repulsion of vortices leads for $T<T^*$ to a much smaller vortex density than in the RSN-theory \[1\], which neglects completely the interaction between vortex dipoles.
For $T<T^*$ we expect the physics to be different from that described by Eqs. (4). Since $T^*(\sigma)$ intersects the RNS phase boundary at $\sigma = \pi /8$ where $T_+=T_-=J\pi/4$, the whole $(T,\sigma)$-range in which reentrance was observed belongs to the freezing region, which has to be reconsidered.
To find the correct behaviour at low temperatures, we consider first the system at $T=0$. A simple estimate shows, that then vortices will not be relevant if the disorder is weak. Indeed, the elastic energy of an isolated vortex of charge $ \pm m$ in a system of radius $R$ is $m^2 \pi J \ln R$, which has to be compared with the possible energy gain $E_r$ from the interaction of the vortex with the disorder. If we rewrite the second term in (2) as $\sum_i m_iV({\bf r}_i)$, we find $[V^2({\bf r}_i)] _d \simeq 2\pi J^2 \sigma \ln R$. Hence the [*typical*]{} energy gain is $-J (m^2 2\pi \sigma \ln R)^{1/2}$.
In order to find the [*maximal*]{} energy gain, we have to estimate the number $n(R)$ of vortex positions ${\bf r}_i$ in which the energies $V({\bf r}_i)$ are essentially uncorrelated. Two vortex positions ${\bf r}_i, {\bf r}_j$ have independent energies if $[V^2({\bf
r}_i)]_d \gg [V({\bf r}_i)V({\bf r}_j)]_d$ , a condition which can be rewritten with $$[(V({\bf r}_i)-V({\bf r}_j))^2]_d = 4\pi \sigma J^2\ln |{\bf r}_i-{\bf r}_j|
\equiv \Delta ^2({\bf r}_i - {\bf r}_j)$$ as $\ln|{\bf r}_i-{\bf r}_j| \approx (1-\epsilon)\ln R$ with $\epsilon
\ll 1$. Thus $n(R) \approx R^{2\epsilon}$ and the maximal energy gain from exploiting the tail of the gaussian distribution for $V({\bf
r}_i)$ is $E_r\approx -2J(m^2 \pi \epsilon \sigma)^{1/2}\ln R$. The total vortex free energy at $T=0$ is therefore $$F_c \approx J\pi (m^2-2(m^2 \epsilon\sigma/\pi)^{1/2})\ln R$$ and hence vortices should be irrelevant for weak disorder $\sigma \ll 1$.
In studying the behavior for $T=0$ but larger $\sigma$ we have to take into account the screening of the vortex and quenched random charges by other vortex pairs. This can be done most easily by using the dielectric formalism. Here we follow the treatment of Halperin \[13\] who showed, that screening by vortex pairs with separation between $R$ and $R+dR$ changes the coupling constant $J(R)$ (which corresponds to the inverse dielectric constant) as $$J(R+dR) = J(R)-4\pi^2 \sum_{m>0} \alpha_m(R) J^2(R) 2 \pi R p_m(R) dR.$$ Here $\alpha_m(R)$ is the polarizability and $p_m(R)$ is the probability density of a pair with charge $+m$ at ${\bf r}_1$ and charge $-m$ at ${\bf r}_2$, $R=|{\bf r}_1- {\bf r}_2|$.
For the calculation of $p_m{(R)}$ and $\alpha_m(R)$ we use the fact, that the interaction energy between a vortex pair and the disorder is gaussian distributed with a width $\Delta (R)$. If the density of pairs is sufficiently small, we may neglect the interaction between pairs and write $$\begin{aligned}
p_m{(R)}&=&\int^{-2|m|(\pi J\ln R+E_c)}_{-\infty}
\frac{dV}{\sqrt{2\pi} \Delta (R)}
e^{-\frac{V^2}{2\Delta ^2(R)}}\approx \nonumber \\
&\approx &\sqrt{\frac{\sigma}{2\pi^2 m^2 \ln R}}
R^{-\frac{m^2 \pi}{2\sigma}},\end{aligned}$$ where the $r.h.s.$ of (8) is valid only for $\sigma$, ${E_c}/{J} \ll
\ln R$. Since $p_{|m|>1}(R) \ll p_1(R)=: p(R)$, we will neglect double occupancy of vortex positions.
The polarizability $\alpha(R)=\alpha_1(R)$ can be calculated in a similar way and is found to be $\alpha (R) \approx R^2/T^*$ at large $R$. With $y^2=R^4 p(R)$ and $l=\ln R$ we get from (7) and (8) for $T=0$
$$\begin{aligned}
\frac{dJ}{dl}&=&- 4 \pi^3 \frac{J}{\sigma}y^2 \label{9a}\\
\frac{dy}{dl}&=&(2-\frac{\pi}{4\sigma})y \label{9b}\\
\frac{d\sigma}{dl}&=&0, \label{9c}\end{aligned}$$
where we again neglected terms of the order $\sigma/l$. These are the flow equations, which replace (4) at [*zero temperature*]{}. Within this approximation, the system undergoes a phase transition at $\sigma _c=\pi/8$ from a KT to a disordered phase, which is in qualitative agreement with our estimate (6). At $\sigma _c$ the exponent $\eta$ shows a universal jump from $1/16$ to zero. For $\sigma > \sigma_c$ the $y$ reaches a value of order magnitude unity on the scale $R\approx \xi$ with $$\xi\propto e^{1/b(1-\pi/{8\sigma})}.$$ $b$ is a constant, which depends on the details of the system. For $R>\xi$ our flow equations are no longer valid, since $y$ is no longer small. We identify $\xi$ with the correlation length in the disordered phase.
We discuss now the properties of the system at low but [*finite*]{} temperatures. The $T$-correction to our free energy estimate (6) are of the order $-2T\ln R$ (or smaller) and hence will not allow a reentrance transition. A more efficient way for thermal fluctuations to influence the low-$T$ behaviour would be the generation of [*uncorrelated*]{} frozen charges $Q(\bf r)$. However unlike to random field systems, where [*uncorrelated*]{} random fields are indeed generated from [*anticorrelated*]{} random fields \[14\], which destroy the ordered phase in 2 dimension at all non-zero $T$, we do not see such a mechanism here. The main difference consists in the existence of a double degenerated ground state in the random field system at $T=0$.
The physics at finite temperature can also be captured within the dielectric formalism. Neglecting again the interaction between vortex pairs at different positions, we calculate the [*normalized*]{} probability for a pair with charges $\pm m$ as $$p_m(R)=p_{-m}(R)=
\left[ \frac {e^{-E_m(R)/T}}{\sum_{m}e^{-E_m(R)/T}} \right]_d,
\label{def.p}$$ where $E_m(R)=2 m^2 (E_c + \pi J \ln R) + m (V({\bf r}_1)- V({\bf
r}_2))$ denotes the pair energy. At large $R$ holds $p_{|m|>1}(R)
\approx 0$, since the elastic energy cost $\propto m^2$ will be compensated with decreasing probability by an energy gain $\propto m$ due to disorder. We therefore drop occupancies $|m|>1$. Furthermore, for a given configuration of disorder, one of the two energies $E_{\pm
1}(R)$ is always so large, that the corresponding weight factor $e^{E_{\pm 1}(R)/T}$ can be neglected. After this approximation, the probability for a single pair reads $$p(R)= p_1(R)= \left[ \frac 1{1+e^{E_1(R)/T}} \right]_d.
\label{def.p1}$$ Eq. (\[def.p\]) thus effectively reduces to the disorder average of the [*Fermi distribution function*]{}. In other words: vortex pairs of vorticity one can be treated as [*non-interacting fermions*]{}. In the limit $T=0$, where this distribution function becomes step-like, Eq. (\[def.p1\]) immediately reduces to the previous expression (8). At finite temperature, the disorder average in Eq. (\[def.p1\]) is performed by splitting the integral over the disorder distribution into two contributions corresponding to $E_1(R) {\mbox{\hbox{ \lower-.6ex\hbox{$>$}\kern-.8em
\lower.5ex\hbox{$<$}\kern+.35em}}}0$. To leading order in $R$, we find $p(R) \sim R^{-\pi/2 \sigma}$ for $0 \leq T \leq
T^*=2J \sigma$, whereas $p(R) \sim R^{-2\pi J/T(1- \sigma J/T)}$ for $T \geq T^*$. Plugging these results into the definition of $y$, we obtain the flow equation (4b) in the whole range $T \geq T^*$, whereas Eq. (9b) is valid in the whole range $0\leq T \leq T^*$. Both equations coincide at the boundary $T =T^*$.
We add a few remarks: As long as $E_1(R) \gg T$, the Fermi distribution can be replaced by the Boltzmann distribution, as is usually done in the treatment of the KT transition \[4\]. The disorder average of the latter yields $p(R) \sim R^{-2\pi J/T(1- \sigma J/T)}$ and hence Eq. (4b) for [*all*]{} temperatures. However, for $T<T^*$ the condition $E_1(R) \gg T$ is no longer fulfilled for most of the vortex positions (see also our remarks below Eqs. (4)) and hence this approximation breaks down. Indeed, use of the Boltzmann distribution at low temperatures would lead to $p(R) \gg 1$, and the interaction between vortex pairs could no longer be neglected. It is therefore important to calculate $p(R)$ from (\[def.p1\]). An attempt to improve upon the Boltzmann-approximation consists in expanding (\[def.p1\]) into a power series in $e^{-E_1(R)/T}$. The $n$-th order term in the expansion yields a contribution $R^{-2\pi J/T(n-n^2
\sigma J/T)}$ to $p(R)$. The series is divergent, i.e. for large $R$ higher order terms are more important than lower order terms, irrespective of temperature. These higher order terms generate contributions to the flow equation $dy/dl$, which tend to blow up $y$ even faster. One might hence expect an instability of the ordered phase, similarly to the observation of Korshunov \[5\]. In fact, the above expansion and in particular the replacement of the Fermi- by the Boltzmann distribution are disqualified [*a posteriori*]{}.
We conclude, that $dy(l)/dl <0$ for all $T<T_+$. The polarizability at finite temperatures is given by $\alpha=R^2/(T+T^*)$ for $T<T^*$ and by $\alpha=R^2/(2T)$ for $T>T^*$. Thus $dJ/dl<0$ holds for all $T<T_+$ which is sufficient to guarantee the absence of reentrant phase topology. In the special case of $E_c \to \infty$ the phase boundary is given by $T_+ (\sigma)$ for $T\ge J\pi/4$ and a horizontal line $\sigma_c = \pi/8$ for smaller $T$, as shown bye bold lines in Fig. 1. This is consistent with the prediction of Ozeki and Nishimori \[8\] about the existence of a horizontal phase boundary. We expect the critical behavior at $T_+(\sigma)$ as discussed in \[1\] to be unchanged. At finite core energies, the actual transition temperature will be renormalized to $T_c(\sigma)< T_+(\sigma)$. Its value $T_c(0)$ is given by the KT flow equations without disorder and lies only slightly below $T_+(0)$ for large $E_c$. For small $\sigma$, the critical RG trajectory flows completely in the domain of equation (4b), where weak disorder induces weak additional screening. Therefore $T_c(\sigma)$ will smoothly decrease with increasing $\sigma$. We expect this function to end up in $T_c(\pi/8)=0$ monotonously, since flow equations vary monotonously in parameter space.
Our conclusions about the absence of reentrance are confirmed also by a discretized Migdal-Kadanoff renormalization group (MKRG) scheme \[15\] for model (1), which we consider in the last part of this paper. Our technique has been shown to be similar to that of José et al. \[16\]. Their approach is based on studying Migdal-Kadanoff recursion relations for the Fourier components of the (spatially uniform) potential.
In the discretized scheme \[15\] instead of allowing $\phi$ to be a continuous variable, we constrain it to take one of many discrete values which are uniformly distributed between $0$ and $2\pi$. Hamiltonian (1) is now defined for values of $\phi$ restricted to $2\pi k/q$, where $k = 0, 1, 2, \ldots , (q-1)$ and $q$ is a number of clock states. We define $$J_{ij} (q,k) \; = \; J \cos ( 2\pi k/q - A_{ij} ).
\label{eqn1}$$ The recursion relations for $J_{ij}(q,k)$ may be found in \[15\]. For the random 2D system, the numerical procedure is based on creating first a pool of $N_p$ bonds, each decomposed into $q$ components according to Eq. (\[eqn1\]). One then picks $N_p$ random batches of 4 such bonds (the corresponding rescaling factor is equal to 2) from the pool to generate a new pool of the coupling variables and the whole procedure is iterated. We consider typically $N_p$=2000 and $q$=100. The results depend on these parameters rather weakly.
It should be noted that Gingras and Sorensen \[16\] have tried to construct the phase diagram of the 2D random Dzyaloshinskii-Moriya model (this model is believed to be equivalent to (1) by the same discretized MKRG approach. In order to locate the Kosterlitz-Thouless phase, they study the scaling behavior of the absolute average height of the potential, $\bar{h}$, which is defined as follows $$\bar{h} \; = \; \langle | J_{ij}(q,0) - J_{ij}(q,q/4) | \rangle \; \; .
\label{eqn5}$$ Due to errratic behavior of $\bar{h}$ they could not draw the phase diagram. The reason here is that $\bar{h}$ representing only two clock states cannot correctly describe the system with many clock states.
To obtain the phase diagram one can consider the scaling of the maximal and minimal couplings for each effective bond or the scaling of the average of absolute values of all $q$ couplings. The scaling properties of these three quantities has been found to be essentially the same, so it is sufficient to focus on the maximum coupling $J_{max}(q,k)$. The details of this approach can be found in Ref. \[15\].
It should be noted that the discretized MKRG approach cannot rigorously reproduce the quasi-long-range XY order in 2$D$ \[15\]. The scale invariance of $J_{max}(q,k)$ in the KT-phase is merely approximate in this approach. In practice, the scale invariance of $J_{max}(q,k)$ persists for about 20 iterations. Further iterations lead to an eventual decrease of $J_{max}(q,k)$ at any nonzero $T$. Having this caveat in mind, we can locate the boundary between the paramagnetic and KT- phase (see Fig. 2). Thus the MKRG gives us additional evidence that the reentrance is absent in model (1).
To conclude, in the present paper we have shown by a combination of simple analytical arguments, a renormalization group calculation and a Migdal-Kadanoff RG scheme at finite $T$, that the 2-dimensional XY-model with random phase shifts does not exhibit a reentrant transition.
[**Acknowledgments**]{} T.N. acknowledges a stay at the ITP at Santa Barbara where this work was begun. It is a pleasure to thank M. Kardar, H. Orland, L.H. Tang for discussion on this problem. S.E.K. and M.S.L. acknowledge support from the SFB 341.
[99]{}
M. Rubinstein, B. Shraiman and D.R. Nelson, Phys. Rev. B [**27**]{}, 1800 (1983) E. Granato and J.M. Kosterlitz, Phys. Rev. B [**33**]{}, 6533 (1986), Phys. Rev. Lett. [**62**]{}, 823 (1989) C. Ebner and D. Stroud, Phys. Rev. B [**31**]{} 165 (1985); M.P.A. Fisher, T.A. Tokuyasu and A.P. Young, Phys. Rev. Lett. [**66**]{}, 2931 (1991) J.M. Kosterlitz and D.J. Thouless, J. Phys. C [**6**]{}, 1181 (1973) S.E. Korshunov, Phys. Rev. B [**48**]{}, 1124 (1993) M.G. Forrester, H. Jong-Lee, M. Tinkham and C.J. Lobb, Phys. Rev. B [**37**]{} 5966(1988); S.P. Beuz, M.G. Forrester, M. Tinkham and C.J. Lobb, Phys. Rev. B [**38**]{}, 2869 (1988) A. Chakrabarti and C. Dasgupta, Phys. Rev. B[**37**]{}, 7557 (1988); M.G. Forrester, S.P. Beuz and C.J. Lobb, Phys. Rev. B[**41**]{}, 8749 (1990) Y. Ozeki and H. Nishimori, J. Phys. A [**26**]{}, 3399 (1993) M. Paczuski and M. Kardar, Phys. Rev. B [**43**]{}, 8331 (1991) J.L. Cardy and S. Ostlund, Phys. Rev. B [**25**]{}, 68991 (1982); Y.Y. Goldschmidt and A. Houghton, Nucl. Phys. B [**210**]{}, 155 (1982); Y.Y. Goldschmidt and B. Schaub, Nucl. Phys. B [**251**]{}, 77 (1985) D.R. Nelson, Phys. Rev. B [**27**]{}, 2902 (1983) The situation resembles somewhat to that in certain spin glasses, where the vanishing of the entropy goes along with replica symmetry breaking, a mechanism which may also apply here. See e.g. D.J. Gross and M. Mézard, Nucl. Phys. B [**240**]{}, 431 (1984) B.L. Halperin, in Proceedings of the Kyoto Summer Institute on Low Dimensional Systems (1979) M. Schwartz, J. Villain, Y. Shapir and T. Nattermann, Phys. Rev. B [**48**]{}, 3095 (1993) M. Cieplak, J.R. Banavar, M.S. Li and A. Khurana, Phys. Rev. B [**45**]{}, 786 (1992); M. Cieplak, M.S. Li, and J.R. Banavar, Phys. Rev. B [**47**]{}, 5022 (1993) J.V. José, L.P. Kadanoff, S. Kirkpatrick and D. Nelson, Phys. Rev. B [**16**]{}, 1217 (1977) M.J.P. Gingras and E.S. Sorensen, Phys. Rev. B [**46**]{}, 3441 (1992)
=
=
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We calculate the asymptotic high-energy amplitude for electrons scattering at one ion as well as at two colliding ions, respectively, by means of perturbation theory. We show that the interaction with one ion [*eikonalizes*]{} and that the interaction with two ions [*causally decouples*]{}. We are able to put previous results on perturbative grounds and propose further applications for the obtained rules for interactions on the light cone. The formalism will be of use for the calculation of Coulomb corrections to electron-positron pair creation in heavy ion collisions. Finally we discuss the results and inherent dangers of the employed approximations.'
author:
- 'U. Eichmann${}^a$, J. Reinhardt${}^a$, S. Schramm${}^b$, W. Greiner${}^a$'
date: |
${}^a$Institut für Theoretische Physik,\
Johann Wolfgang Goethe-Universität, Frankfurt am Main, Germany\
${}^b$Gesellschaft für Schwerionenforschung mbH, Darmstadt, Germany
title: Electron Propagation in the Field of Colliding Nuclei at Ultrarelativistic Energies
---
Introduction
============
At ultrarelativistic energies, the theoretical treatment of scattering processes is extremely facilitated. On the one hand, the relevant equations themselves simplify, when terms of order ${\cal O}(1/\gamma^2)$ become negligible, on the other hand, the interactions simplify due to causality. In that way, high energy scattering becomes analytically accessible.
Eikonal approximations or optical models usually are formulated for the scattering of a highly energetic particle at a slow or even static center [@Abarbanel][@Torgerson]. We present a simple transformation of the covariant derivatives that is used to easily solve the opposite case. The transformation of the equations of motion for particles scattered by fast moving charge centers immediately generates the scattered wave describing the particle. Our results coincide with previous calculations performed in this reference frame [@Jackiw][@tHooft].
The summation of ladder graphs is shown to [*eikonalize*]{} as well [@Chang-Ma]. This was elegantly derived within the method of kinematically decoupling the components of the scattering process, and Lorentz transforming into the respective rest frames [@Chang-Fishbane] which inherently contains the advantages of a fast external potential.
Following a different approach we will exploit the same advantages. We perform a perturbative approach and directly approximate the external potential by its asymptotic high-energy limit which amounts to saying, that the longitudinal components of the exchanged photons can be discarded.
In doing so, one can directly rederive the amplitude for the scattering at one center and even put the recent result of Segev and Wells [@Segev] for the scattering amplitude for an electron moving in the field of two ultrarelativistic colliding ions on perturbative grounds. Moreover, one is allowed to go beyond their calculations and is provided deeper insight.
The derivations in this paper are formulated for electron scattering, but they can be immediately extended to cover the physically more relevant process of electron-positron pair production. The search for exact analytic expressions describing electron-positron pair production in heavy-ion collisions is motivated by the question whether Coulomb effects only play an inferior role at high energies. Such a conclusion might be drawn from a comparison between second-order perturbation theory results [@Bottcher] and calculations employing Furry-Sommerfeld-Maue wave functions [@Ionescu]. It should be mentioned, however, that the Coulomb distortions considered in these calculations only account for one ion, whereas the second ion enters as a perturbation.
Scattering of an electron off a fast moving source
==================================================
Transformation of the Dirac equation {#1iontrafo}
------------------------------------
We are searching for the asymptotic scattering solution of a Dirac particle from a fast moving Coulomb potential in the limit of very large collision energy. In the Lorentz gauge the Li$\acute{\rm e}$nard-Wiechert potentials for a point charge moving moving with uniform velocity $\beta$ in $+z$ direction read $$\begin{aligned}
\label{lwpot}
A_0&=&-\frac{Z\alpha \gamma}{\sqrt{\gamma^2(z-\beta t)^2 +
\vec{x}_\perp^2}}\\
\label{a0a3}
A_3&=&\beta A_0\end{aligned}$$ The equation of motion for the scattered particle becomes $$\label{dg}
\left[ \hat{\gamma_0}(i\partial_t -A_0)+\hat{\gamma_3}(i\partial_z +A_3)+
\hat{\vec{\gamma}}_\perp \cdot i\vec{\nabla}_\perp -
m\right] \psi =0$$ We set $c=\hbar=1$. The charge $e$ of the electron was absorbed into the definition of the potential. We make use of the external field approximation, i.e. we assume that the source is not influenced by the scattered particle and moves on a straight line. This treatment will be justified if the mass of the source particle is very large. To simplify the Dirac equation (\[dg\]) we use the operator identity [@Miura] $$(i\partial_x \mp
i\partial_x\ln \phi)^n=\phi^{\pm 1}(i\partial_x)^n\phi^{\mp 1}
\label{kovabl}$$ to rewrite the covariant derivatives. We must introduce two fields $\phi'$ and $\phi$ for the space and the time component of the vector-potential $A^\mu$ $$\begin{aligned}
A_0&=&i\partial_t \ln \phi\nonumber\\
A_3&=&i\partial_z \ln \phi'\end{aligned}$$ The field $\phi$ is determined to be $$\label{phidef}
\phi=e^{-i\int_{-\infty}^t dt' A_0}$$ The thus transformed Dirac equation reads (see Appendix \[adiractrans\]) $$\label{dtrans}
\left[\hat{\gamma}_0i\partial_t +\hat{\gamma}_3
(i\partial_z-\frac{1}{\beta^2\gamma^2}A_3)+
\hat{\vec{\gamma}}_\perp \cdot (i\vec{\nabla}_\perp +
i{\rm grad}_\perp
\ln \phi )- m\right]\tilde{\psi}=0$$ where $\tilde{\psi}=\psi/\phi = e^{i\int_{-\infty}^t dt' A_0}\psi$. The operator identity, together with the field $\phi$ defined in (\[phidef\]) has led to the elimination of the scalar part of the vector potential, i.e. to the temporal gauge, $A'_0=0$. For very large $\gamma$ one nearly has a purely transverse vector potential $\vec{A'}_\perp = i{\rm grad}_\perp
\ln \phi$ which is the negative time integral of the transverse electric field. From classical electrodynamics one knows, that the time integral of the transverse electric field is given by $$\label{etransint}
\int_{-\infty}^{\infty}\vec{E}_\perp=-2Z\alpha\vec{x}_\perp /
(\beta x_\perp^2)$$ This implies that $$\int_{-\infty}^\infty dt' A_0 = +\frac{1}{\beta}
Z\alpha \ln x_\perp^2 + {\cal C}$$ which reproduces (\[etransint\]) if the transverse gradient operator is applied. ${\cal C}$ is an infinite quantity which expresses the divergence of the phases in Coulomb scattering. Furthermore using (\[phidef\]) and (\[lwpot\]) it is easy to show, that the transverse vector potential exhibits a Heaviside step function dependence $\sim \theta (t -z)\vec{x}_\perp/x^2_\perp$ in the limit of very large $\gamma$. Now, since $t$ and $\gamma$ enter symmetrically in the integral, the limit $\gamma \to \infty$ corresponds to sending the upper bound of the integral to infinity. Therefore, all of the above is applicable and we find $$\label{potlimc}
\lim_{\gamma \to \infty} A_0 = +\delta (z-t) Z \alpha \ln x_\perp^2 + {\cal
C}'$$ The Coulomb phase ${\cal C}'$ in general will depend on $z$ and $t$. It can be removed by a gauge transformation, as is easily seen $$\label{cgtrafo}
\tilde{\psi}'=e^{-i\int_{-\infty}^{t}dt'{\cal C}'}\tilde{\psi}=
e^{+iZ\alpha \theta (t-z)
\ln x_\perp^2}\psi$$ This gauge transformation was first applied in [@Aichelburg]. The removal of the Coulomb phase yields a short range potential allowing for asymptotic plane wave solutions (see Appendix \[apotlim\]).
For $t\neq z$ the $t$ and $z$ dependence in both the transverse vector potential and the transformed spinor $\tilde{\psi}$ vanish in the limit $\gamma\to \infty$. By inverse transformation we find that $\psi$ solves a free Dirac equation on either side of the light front $t=z$ and can only differ by a phase.
The transformed wave function $\tilde{\psi}$ has the advantage of being continuous on the surface defined by $t=z$. In contrast, the wave function $\psi$ exhibits a discontinuous behaviour at the light front. There is a jump in that component of $\psi$ which couples to $\hat{\gamma}_-
=\hat{\gamma}_0-\hat{\gamma}_3$, the matrix structure of the interaction in the limit $\gamma \to \infty$. Using this property one directly finds for $\hat{\gamma}_-\psi$ at the discontinuity $$\label{psiout}
\hat{\gamma}_-\psi(t-z=0^+)=e^{-iZ\alpha \ln x_\perp ^2 }\hat{\gamma}_-\psi
(t-z=0^-)$$ where we ignored the irrelevant quantity ${\cal C}$.[^1] The complement $\hat{\gamma}_+\psi$ of these spinor components, where $\hat{\gamma}_+ = 2\hat{\gamma}_0
-\hat{\gamma}_-= \hat{\gamma}_0+\hat{\gamma}_+$ is continuous at $t=z$. Both parts of the spinor are coupled via the free Dirac equation on either side of the discontinuity.
By application of the LSZ-reduction formula one finds in general, that at very large scattering energies the $S$ matrix is determined by $\phi$, in which we recognize the well known eikonal form [@Fried]. Because of the identity (\[kovabl\]) this result holds independently of the power of the momentum entering in the respective wave equation. For that reason the expressions for the $S$ matrices for e.g. spinor or scalar particles only differ by an overall factor.
We first consider the properties of the previous result. The LHS of (\[psiout\]) can be expanded in plane waves. Since we consider scattering at the negative light front, we must substitute $d^3x \to dx_+d^2x_\perp$ [@Meggiolaro] and accordingly $d^3p \to dp_-d^2p_\perp$. The expansion coefficient corresponds to the $S$ matrix in momentum space, which is easily found to be $$\begin{aligned}
\label{smom}
S(p',p)&=&2\pi\delta
(p'_--p_-)
\Bigg[\left(\frac{4}{(\vec{p'}_\perp-\vec{p}_\perp)^2}
\right)^{1-iZ\alpha}\Gamma^2(1-iZ\alpha)\sin(\pi
iZ\alpha)\nonumber\\
&&+(2\pi)^2\delta(\vec{p'}_\perp-\vec{p}_\perp)\Bigg]
\overline{u}(p')\hat{\gamma}_-u(p)\end{aligned}$$ Here $u$ denotes the electron unit spinor. If the trajectory of the ion is shifted by the impact parameter $\vec{b}$, this result is simply multiplied by the factor $e^{i(\vec{p'}_\perp-\vec{p}_\perp)\cdot \vec{b}}$. $p$ and $p'$ are the incoming and outgoing momenta. We note that the negative light cone momentum $p_-=p_0-p_3$ is conserved in the scattering. The positive light cone momentum is fixed by the mass shell condition. The first term in the square brackets in (\[smom\]) corresponds to the $T$ matrix. Eq. (\[smom\]) represents a well known result which was previously derived in e.g. [@Abarbanel; @tHooft; @Jackiw; @Segev].
Perturbative Approach {#pertapp}
---------------------
In this section we want to derive the eikonal form of the $T$ matrix via perturbation theory. Several approximations are necessary to obtain the eikonal form, namely the neglect of the longitudinal components of the photon momentum, the conservation of the photon light cone momentum, as well as the simplification of the matrix structure of the interaction [@Chang-Ma]. The calculation shows that these approximations are the counterparts of the requirement of a vanishing longitudinal vector potential and the step function dependence of the transverse vector potential. Having this in mind we directly use the asymptotic high-energy expression of the potential. We then evaluate the terms of the perturbation series for the external-field scattering problem depicted by the Feynman graphs of Fig. \[fig01\].
The potential entering into our calculations is of the form $$\label{pot}
V_0(x)=V_3(x)=\delta (z-t)V_\perp(\vec{x}_\perp)$$ In the following calculations it will not be necessary to specify the explicit form of $V_\perp(\vec{x}_\perp)$. Problems related to the logarithmic potential obtained in the last section will be discussed in section \[discussion\]. We use the light-cone variables $$\left(\begin{array}{c}p_-\\ p_+\end{array}\right)=\left(\begin{array}{rr}
1&-1\\ 1&1\end{array}\right)\left(\begin{array}{c}p_0\\ p_3\end{array}\right)$$ The Feynman propagator describing the internal electron lines reads $$S_F(p)=\frac{1}{\hat{\gamma}_0p_0 -\hat{\vec{\gamma}}\cdot\vec{p}-
m+i\epsilon}=
\frac{\frac{1}{2}(\hat{\gamma}_-p_++\hat{\gamma}_+p_-)-
\hat{\vec{\gamma}}_\perp
\cdot \vec{p}_\perp +m}{p_+p_--p_\perp^2-m^2+i\epsilon}$$ The following products of gamma matrices in the light-cone representation are needed: $$\label{gamprod}
\begin{array}{ll}
\displaystyle
\hat{\gamma}_\pm \hat{\gamma}_\mp \hat{\gamma}_\pm &
\displaystyle
=4\hat{\gamma}_\pm\\
\displaystyle
\hat{\gamma}_\pm \hat{\vec{\gamma}}_\perp \hat{\gamma}_\mp&\displaystyle
= 2\hat{\vec{\alpha}}_\perp \hat{\gamma}_\mp\\
\displaystyle
\hat{\gamma}_\pm \hat{\vec{\gamma}}_\perp \hat{\gamma}_\pm
&\displaystyle =0\\
\displaystyle
\hat{\gamma}_\pm \hat{\gamma}_\mp&\displaystyle
=2\hat{\gamma}_0\hat{\gamma}_\mp\\
\displaystyle
\hat{\gamma}_\pm \hat{\gamma}_\pm&\displaystyle
=0\\
\end{array}$$ The amplitude for electron scattering in first order perturbation theory is $$\label{1order}
A^{(1)}_{p'p}=(2\pi) (-i) \delta (p'_--p_-)
F_{p'p}(V_\perp)\overline{u}(p')\hat{\gamma}_-u(p)$$ $F_{p'p}(\;\;)$ denotes the Fourier transform with respect to the transverse coordinates taken at the momentum $(\vec{p'}_\perp-\vec{p}_\perp)$ $$F_{p'p}(V_\perp)=\int d^2 x_\perp e^{-i\vec{x}_\perp \cdot
(\vec{p'}_\perp-\vec{p}_\perp)}V_\perp (\vec{x}_\perp)$$ In second order the amplitude reads $$\begin{aligned}
A^{(2)}_{p'p}&=&\int \frac{dk_+dk_-d^2k_\perp}{(2\pi)^4} (2\pi)^{2}
(-i)^2 i \delta (k_--p_-)\delta
(p'_--k_-)\frac{k_-}{k_-k_+-k_\perp^2 -m^2
+i\epsilon}\nonumber \\
\label{k+int}
&&F_{kp}(V_\perp)F_{p'k}(V_\perp)\overline{u}(p')\hat{\gamma}_-u(p)\\
&=&(2\pi) (-i)^2 \delta (p'_--p_-)
\frac{1}{2}F_{p'p}(V_\perp^2)\overline{u}(p')\hat{\gamma}_-u(p)
\nonumber\end{aligned}$$ The $k_+$ integral in (\[k+int\]) drops out using the symbolic substitution $$\label{Feynsubst}
1/(x+i\epsilon)\to P(1/x)-i\pi\delta (x)$$ since the principal value integral $P$ vanishes. It is interesting to note that the simple structure of the results (\[1order\]) and (\[k+int\]) is retained if one goes to higher orders of perturbation theory. The $n^{th}$ order amplitude factorizes into $n-1$ integrals of the form (\[k+int\]) which leads to $$\label{norder}
A^{(n)}_{p'p}=(2\pi) (-i)^n \delta (p'_--p_-)
\frac{1}{n!}F_{p'p}(V_\perp^n)\overline{u}(p')
\hat{\gamma}_-u(p)$$
This result is obtained by symmetrizing the $n-1$ integrals over the positive light cone momenta in (\[Feynsubst\]) yielding the expression $(-i2\pi)^{n-1}/n!\prod_i \delta (k_+^i)$ [@Chang-Ma]. This corresponds to reconsidering the different time orderings and finally dividing by $n!$ to prevent double counting. This symmetrization procedure directly shows that the principal value terms in (\[Feynsubst\]) do not contribute.
Clearly, with (\[norder\]) the perturbation series can be summed up to yield the result $$\label{tmatr1}
A_{p'p}=2\pi\delta (p'_--p_-){\cal
T}(\vec{p'}_\perp-\vec{p}_\perp)\overline{u}(p')\hat{\gamma}_-u(p)$$ Here we defined the momentum transfer function $${\cal
T}(\vec{p'}_\perp-\vec{p}_\perp)=F_{p'p}(e^{-iV_\perp(\vec{x}_\perp)}-1)$$ with $$V_\perp(\vec{x}_\perp)=\int_{-\infty}^{+\infty} dt V_0(x)$$ This result reproduces the eikonal form.
Solution in the field of several ions
=====================================
The case of two colliding ions
------------------------------
In the c.m. frame, the field of two ultrarelativistic colliding ions $A$ and $B$, cf. Figure \[fig02\], reads $$\label{pot2ion}
V_{0/3}(x)=\delta (z-t) V^A_\perp(\vec{x}_\perp)
\pm \delta (z+t)V^B_\perp(\vec{x}_\perp)$$
The identity (\[kovabl\]) can also be applied to potentials given by a superposition as is easily verified $$\label{mtrafomi}
(i\partial_x \mp i\sum \partial_x\ln \phi_i)^n=\left(\prod \phi_i\right)^
{\pm 1}
(i\partial_x)^n
\left(\prod \phi_i\right)^{\mp 1}$$ Since in the case of (\[pot2ion\]) we have two discontinuities, the asymptotic solution is not obtained as easily as in section \[1iontrafo\]. The explicit calculation is shown in Appendix \[asolu2ion\].
It is found, that the two ions couple to distinct components of the electron spinor. We show in this section, how this behaviour follows from perturbation theory and how it can be interpreted consistently.
We have to consider several new diagrams describing the alternate interaction of the electron with both ions. Using (\[gamprod\]) we find, that the contribution to the $T$ matrix of an arbitrary number of interactions with one ion that are sandwiched between interactions with the other ion (see Fig. \[fig03\]) , vanishes. The reason is, that we end up with an integral of the form $$A\sim\int \frac{dk_\pm}{(k_\pm p_\mp -k_\perp^2 -m^2 +i\epsilon)
(k_\pm p'_\mp -{k'}^2_\perp -m^2 +i\epsilon)}=0$$ The vanishing of this integral is immediately seen from Cauchy’s theorem since the contour can be closed in the upper half plane, where the integrand is analytic.
In the ultrarelativistic limit the electron will therefore interact with the ions separately, see Figure \[fig04\]. The separate interactions of the electron with the two ions $A$ and $B$ are linked in the following way $$\begin{aligned}
A^{tot}_{p'p}&=&\int \frac{d^2k_\perp}{(2\pi)^2}
{\cal T}_A(-\vec{p}_\perp+\vec{k}_\perp){\cal T}_B(\vec{p'}_\perp -
\vec{k}_\perp)\nonumber \\
&&\overline{u}(p')\frac{-\hat{\vec{\alpha}}_\perp
\cdot \vec{k}_\perp +
\gamma_0
m}{p'_+ p_- - {k}_\perp^2 -m^2 +i\epsilon}
\hat{\gamma}_+u(p)
\nonumber\\
&&+\int \frac{d^2k_\perp}{(2\pi)^2}
{\cal T}_A(-\vec{p}_\perp+\vec{k}_\perp){\cal T}_B(\vec{p'}_\perp -
\vec{k}_\perp)\nonumber \\
\label{Ttot}
&&\overline{u}(p')\frac{-\hat{\vec{\alpha}}_\perp
\cdot (\vec{p}_\perp +\vec{p'}_\perp -\vec{k}_\perp) +
\gamma_0
m}{p'_- p_+ -
(\vec{p}_\perp +\vec{p'}_\perp -{k}_\perp)^2 -m^2 +i\epsilon}
\hat{\gamma}_-u(p)\end{aligned}$$ Here we have already added both possible time orderings. ${\cal T}_A$ and ${\cal T}_B$ are the momentum transfer functions ${\cal T}$, defined in (\[tmatr1\]) for the interactions with ion $A$ and $B$, respectively. This result is equivalently obtained by using the discontinuous behaviour at the light fronts (see Appendix \[asolu2ion\]) and corresponds to the result of Segev and Wells [@Segev].
To understand the decoupling property, one has to consider the matrix structure of the potential. To this end we write down the Dirac equation in the following form $$\left[ i\partial_t+
\hat{\vec{\alpha}} \cdot i \vec{\nabla} - \hat{\gamma}_0m
-(1\pm \beta\hat{\alpha_z})A_0 \right] \psi =0$$ where the sign depends upon the direction of motion and $A_0$ is given by (\[lwpot\]). In the limit $\beta\to 1$ the operators $1/2(1\pm \beta\hat{\alpha_z})$ become orthogonal projection operators [@Segev]. The action of these operators can be understood if one recalls the standard form of Lorentz transformations [@Itzykson] in spinor space $$\label{ltrafo}
\psi'(x')=e^{-(i/4)\sigma_{\alpha \beta}\omega^{\alpha\beta}}\psi (x)$$ Here $\sigma_{\alpha \beta}=i/2[\gamma_\alpha,\gamma_\beta]$ and the exponent represents the product of the rapidity vector $\vec{\omega}$ times the generators of the Lorentz transformation. For a boost in $+z$ direction (\[ltrafo\]) simplifies to $$\begin{aligned}
\psi'(x')&=&e^{-\frac{\omega}{2}\hat{\alpha}_z}\psi (x)\nonumber \\
&=&{\rm cosh}\left(\frac{\omega}{2}\right)\left(1-{\rm
tanh}\left(\frac{\omega}{2}\right)\hat{\alpha}_z\right)\psi (x)\end{aligned}$$ Therefore, (see (\[lwpot\])), a Lorentz-transformed vector acting in spinor space $$(1\pm \beta\hat{\alpha_z})A_0=\gamma(1\pm
\beta\hat{\alpha_z})\frac{-Z\alpha}{r'}$$ can directly be obtained by a Lorentz transformation (\[ltrafo\]) accounting for the vectorial nature of the transformed object with a factor 2 in the exponent. The operators $(1\pm\hat{\alpha}_z)$ are $1/\gamma$ times a Lorentz transformation with effectively infinite rapidity. These operators project the Dirac spinors onto causally disconnected subspaces of the Hilbert space. Therefore it is simply causally impossible for the Dirac spinor to communicate alternately with both ions.
Therefore, even the exact expressions for the interaction of an electron with two colliding ultrarelativistic ions maintains the structure of the two-photon graph. We can interpret (\[Ttot\]) as the interaction of an electron in lowest order with a ”dressed“ potential of the form [@Baltz2] $$\label{modpot}
\tilde{V_0}(x)=\tilde{V_3}(x)
\sim\delta (z-t)\left(\left(\frac{1}{x_\perp}\right)^{2iZ\alpha}-1\right)$$ An inspection of (\[Ttot\]) reveals that the scattering amplitude is represented by a divergent integral. There are infrared divergencies caused by the poles of the momentum transfer functions ${\cal T}_A$ and ${\cal
T}_B$, see Appendix \[approp\]. It is interesting to note, that an explicit introduction of a photon mass describing a screened Coulomb potential does [*not*]{} yield a regularized expression for the functions ${\cal T}$. On the other hand, if the modified potential (\[modpot\]) is screened with a damping factor $e^{-\epsilon x_\perp}$, this leads in momentum space to $$\label{tscreened}
\tilde{V}(k)\sim
\frac{1}{(\epsilon^2+k_\perp^2)^{(1-i\alpha Z)}}\Gamma(2(1-i\alpha
Z))P_{1-2i\alpha Z}(\epsilon(\epsilon^2+ k_\perp^2)^{-\frac{1}{2}})$$ which resembles the propagator of a photon with mass $\epsilon$. $P_{1-2i\alpha Z}(\;\;)$ denotes a Legendre function.
We should stress however, that such artificial regularization procedures are not needed. There is a natural cut off since the condition for the applicability of the used approximations requires (see Appendix \[apotlim\]) $$\gamma \gg \frac{x_\perp}{|z-\beta t|}$$ which in momentum space translates into the condition $$k_\perp \gg \frac{\omega}{\gamma}$$ This lower bound for the transverse momentum corresponds to the cut off inherent in the Fourier transform of the potential, cf eq. (\[ftpot\]) The introduction of this cut off has the important property to restore the energy dependence of the amplitude which was lost when taking the limit $\gamma \to \infty$.
Solution in the field of channeled ions {#crystal}
---------------------------------------
Here we want to sketch briefly an extension of the formalism discussed so far to the case of more that two colliding charges. The causal decoupling of interactions with sources moving on the positive and negative light-cone, respectively, and the above interpretation of the interaction can be used to calculate the scattering amplitude of electrons (or more realistically electron-positron pair production) for a field configuration which corresponds to the channeling of an ion in a crystal.
We use the equal speed system, the crystal is moving in $-z$ direction. The crystal layers have a spatial distance $a\vec{e}_z$. In the ultrarelativistic case, the electron again interacts with the ion and the crystal layers separately and we get simple time orderings of the interaction. For $n$ crystal layers we have $n+1$ possibilities. For the sake of simplicity we formulate the perturbative description of the successive interactions of the electron with both the ion and the crystal layers directly with modified potentials of the form (\[modpot\]). One then obtains for the interaction with two neighbouring crystal layers the integral $$\begin{aligned}
A&=&-\delta (p'_+-p_+)e^{ip'_-\frac{a}{2}}\nonumber \\
&&\int\frac{dk_-d^2k_\perp}{(2\pi)^2}
\frac{\overline{u}(p')\hat{\gamma}_+u(p)}{-ik_-+
i\frac{k_\perp^2+m^2}{k_+}+\frac{\epsilon}{k_+}}e^{-ik_-\frac{a}{2}}{\cal
T}_{C_i}(\vec{k}_\perp-\vec{p}_\perp){\cal
T}_{C_{i+1}}(\vec{p'}_\perp-\vec{k}_\perp)\nonumber \\
&=&2\pi \delta (p'_+-p_+)
\int\frac{d^2k_\perp}{(2\pi)^2}e^{i\left(-\frac{k_\perp^2+m^2}{2p'_+}+
\frac{p'_-}{2}\right)a}\nonumber \\
&&{\cal T}_{C_i}(\vec{k}_\perp-\vec{p}_\perp)
{\cal T}_{C_{i+1}}(\vec{p'}_\perp-\vec{k}_\perp)
\overline{u}(p')\hat{\gamma}_+u(p)
\;\;\;\;a>0,\;\epsilon \to 0\end{aligned}$$ The subscripts $C_i$ and $C_{i+1}$ denote the scattering amplitudes from the interaction of the electron with the $i$th and the $(i+1)$th crystal layer. For $a<0$ (reverse direction of electron motion) the integral vanishes, which expresses, that the electron can not interact alternately with neighbouring crystal layers, due to causality. The derivation of this functional connection using (\[mtrafomi\]) is shown in Appendix \[asolu2ion\].
If the electron interacts with the ion between interacting with two distinct crystal layers, we get $$\begin{aligned}
A&=& \int \frac{d^2k_\perp d^2k'_\perp}{(2\pi)^4}
\frac{e^{ip'_-\frac{a}{2}}
\left(e^{-i ({k}_\perp^2
+m^2-i\epsilon )\frac{a}{2p_+}} -e^{-i ({k'}_\perp^2
+m^2-i\epsilon )\frac{a}{2p'_+}}\right)}{(p'_+({k}_\perp^2+m^2)-
p_+({k'}_\perp^2+m^2)-i \epsilon(p'_+-p_+))}
\nonumber \\
&&{\cal T}_{C_i}(\vec{k}_\perp-\vec{p}_\perp)
{\cal T}_A(\vec{k'}_\perp-\vec{k}_\perp)
{\cal T}_{C_{i+1}}(\vec{p'}_\perp-\vec{k'}_\perp) \nonumber \\
&&\overline{u}(p')
(-\hat{\vec{\alpha}}_\perp \cdot \vec{k}_\perp +\hat{\gamma}_0m)
(-\hat{\vec{\alpha}}_\perp \cdot
\vec{k'}_\perp +\hat{\gamma}_0m) \hat{\gamma}_+u(p)\end{aligned}$$ Successive interactions with different crystal layers factorize and any scattering process including intermediate interaction with the channeled ion gives the same amplitude. Further studies will have to show how these considerations can be put to use for the calculation of pair creation in channeling.
Discussion
==========
In the previous sections the potential of a fast moving charge has been substituted by its asymptotic high-energy expression. From a mathematical (and also a physical) point of view this is a problematic limit, since the required transformation is not an element of the Lorentz group. Furthermore, the potential (\[lwpot\]), a bounded operator in Hilbert space, gets transformed into an unbounded operator, and finally the number of spatial dimensions gets reduced from three to two.
The ansatz directly reflects the approximations made by Chang and Ma [@Chang-Ma] who neglected the longitudinal components of the photon momentum, giving the $\delta$-functions for the respective conserved light cone momenta. The above mentioned problems emerge here in the fact that the longitudinal components of the photon momentum never really vanish.
All approximations allow the well known conclusion, that the eikonal expression can be regarded as the contribution of all ladder diagrams in the high-energy limit and that it is completely compatible with a perturbative calculation.
In the case of two ions we mainly profited from the causal decoupling of the interactions implied by the presence of the factors $(1\pm \alpha_3)$. The matrix structure of the true interaction is given by $(1\pm \beta\hat{\alpha}_3)\approx
(1\pm \hat{\alpha}_3) \mp
\hat{\alpha}_3/2\gamma^2$, so that the leading corrections to this behaviour are suppressed with $1/\gamma^2$.
However, the considered calculations have inherent dangers. Fortunately we had to specify neither the transverse part of the potential nor its Fourier transform throughout our perturbative calculation. The first point may serve to generalize the validity of the result to any function $V_\perp(\vec{x}_\perp)$. However, a naive comparison even with the first order Born approximation would have failed due to the difficulties concerning the Fourier transform of the logarithm, whereas the logarithm in the argument of the exponential function is meaningful and correct. On the one hand the integration of the potential eliminates one dimension, which is finally recovered in the overall $\delta$ function for the light cone momenta. On the other hand the calculation of the scattering matrix between asymptotic states $(t\to \pm \infty)$ corresponds to the (unphysical) limit $\gamma \to \infty$. The detour via the Fourier transform of the ungauged potential gives a logarithm as well (see Appendix \[approp\]), depending, however, strongly upon a regularization mass $\mu$. This has its root in the fact, that $1/(k_\perp^2 + \mu^2)$ is not the correct two-dimensional photon propagator [@Grignani]. The divergent term $\lim_{\mu\to 0}\ln \mu^2$ is the term ${\cal C}$ in section \[1iontrafo\].
Now it is well known, that two-dimensional fields in the limit of vanishing mass are rather ill-defined objects, whereas the exponential of these fields is not. The Fourier transform of this exponential expression is elementary (see (\[smom\])). It can further be expanded into a Taylor series. Although it is not justified to identify the different terms with the Fourier transforms of the powers of the logarithm, the first term corresponds to the high energy limit of the Fourier transform of the retarded potential, which is rather accidental.
Nevertheless, the correct Fourier transform of the logarithm in two dimensions is obtained by Taylor expansion of ${\cal T}(k)$, but the limit $Z\alpha \to 0$ has to be taken [*after*]{} having integrated the expression with a test function [@Grignani]. $$\label{lnft}
\int d^2x_\perp e^{-i\vec{k}_\perp\cdot \vec{x}_\perp}\ln x_\perp^2
=\lim_{iZ\alpha\to 0}\frac{d}{d(iZ\alpha)}\left(\pi \frac{\Gamma(1-i\alpha
Z)}{\Gamma(i\alpha
Z)}\left(\frac{4}{k_\perp^2}\right)^{1-i\alpha Z}
\right)$$ This peculiarity is related to the fact, that the linearity of the Fourier transform is not strictly defined for the action on infinite series, resulting in the non-commutability of limiting procedures as in (\[lnft\]). Another way is to rewrite the logarithm by the $t$ integral of the gauged potential and expressing the integrand by means of its Fourier transform. One then finds for the Fourier transform of the logarithm in two dimensions [@Ferrari]: $$-\int d^2x_\perp e^{-i\vec{k}_\perp\cdot \vec{x}_\perp}
\ln x_\perp ^2 = \lim_{\lambda\to 0} 4 \pi \left(
\frac{1}{k_\perp^2 +\lambda^2} + \pi \delta^2(k_\perp) \ln \left(\frac{
\lambda^2}{\mu^2}\right)\right)$$ with $\lambda =\omega/\gamma$, $\mu = 2/e^C$. The condition $\lambda \to 0$ coincides with the limit $\gamma\to \infty$ (see Appendix \[approp\]). The correct treatment of this result again requires the limit to be taken after integrating the expression with a test function. Inverse transformation shows the required independence of the result of the regularzation parameter $\lambda$.
In view of the previous discussion one may infer that it is not justified to identify the eikonal expression using a gauged potential with first-order perturbation theory using the original potential, since this ignores the gauge transformation applied to the potential.
In any case the above considerations show, that the limit $\gamma \to \infty$ is pathological and that its implications should be studied with care.
Conclusion
==========
The transformation presented in section \[1iontrafo\] directly yields the scattering amplitude for (arbitrary) particles scattered at fast charge centers. Due to Lorentz invariance the expression for the amplitude holds even for the case of static scattering centers, which gives the classical form of the eikonal approximation. The essential ingredients in the static case – the vanishing of the spin current and the assumption that $\phi$ was a slowly varying function – have been replaced by the discontinuous behaviour at the light fronts in the presented case of fast scattering centers.
The gauge transformed high energy limit of the potential directly contains the necessary approximations mentioned in section \[pertapp\], and perturbative calculations using this potential can be done without further assumptions.
The results obtained by eikonal approximations or by application of the transformation (\[mtrafomi\]) are shown to be equivalent to the sum of all ladder graphs. It should therefore not be surprizing that it is possible to regain perturbative results from the eikonal expression. We showed however, that the obtained results must be studied with care, and that a wrong treatment only accidentally leads to correct results.
Finally, the interaction of the electron with several ions moving along the light cones was shown to decouple due to causality.
The high-energy scattering amplitude of electrons in the field of two colliding ions has the same structure as the second-order perturbative result [@Bottcher]. It can therefore be considered as a two-photon process with a modified potential of the form (\[modpot\]). The restrictions for the exchanged momenta imposed by the considered approximations are compatible with those encountered in the Weizsäcker–Williams method of virtual quanta. This method corresponds to a first-order Born approximation in the temporal gauge, considering only the transvere part of the interaction (the longitudinal part is suppressed by $1/\gamma^2$). The analogy of the obtained scattering amplitude to the two-photon process allows to adopt the approximations made in [@Bottcher], aiming to express the result in the Weizsäcker-Williams form. The cross section of the scattering process can therefore be obtained from the Klein–Nishina formula for Compton-scattering and photon distributions obtained from (\[modpot\]) in the temporal gauge [@Eichmann]. The simplified structure of the scattering amplitude allows for a study of the high-energy behaviour of electron-positron pair production, that accounts correctly for the Coulomb effects of both ions.
Acknowledgements {#acknowledgements .unnumbered}
================
U.E. would like to thank S.J. Chang for helpful comments, and D. Schwarz and F. Constantinescu for stimulating discussions. This work was supported by [*Deutsche Forschungsgemeinschaft*]{} DFG within the project Gr-243/44-2.
Transformation of the Dirac equation {#adiractrans}
====================================
According to (\[kovabl\]) we set $$\begin{aligned}
A_0&=& i\partial_t \ln \phi\\
A_3&=& i\partial_z \ln \phi'\end{aligned}$$ Since $A_3=\beta A_0$ and $\partial_t =-\beta \partial_z$ due to the fact, that the $t$ and $z$ dependence enters only via the combination ($z-\beta t$), we find $$\begin{aligned}
&&\hphantom{\Rightarrow\;\;}\frac{\partial_z\phi'}{\phi'}=-\beta^2
\frac{\partial_z\phi}{\phi}\\
&&\Rightarrow\;\; \phi'=\phi^{-\beta^2}\end{aligned}$$ i.e. $\phi=1/\phi'$ for $\beta \to 1$ as expected (see (\[kovabl\]),(\[a0a3\])). Inserting this into the Dirac equation, we find $$\begin{aligned}
&&\left[ \phi \hat{\gamma}_0i\partial_t \frac{1}{\phi} + \phi
\phi^{-\frac{1}{\gamma^2}}\hat{\gamma}_3i\partial_z
\frac{1}{\phi\phi^{-\frac{1}{\gamma^2}}} +\hat{\vec{\gamma}}_\perp
i\vec{\nabla}_\perp -m\right]\psi\\
&&=\left[\phi\left( \hat{\gamma}_0i\partial_t + \phi^{-\frac{1}{\gamma^2}}
\hat{\gamma}_3
i\partial_z \frac{1}{\phi^{-\frac{1}{\gamma^2}}}\right)\frac{1}{\phi}
+\hat{\vec{\gamma}}_\perp
i\vec{\nabla}_\perp -m\right]\psi\\
&&=\phi\left[ \hat{\gamma}_0i\partial_t + \hat{\gamma}_3\left(
i\partial_z -\frac{1}{\beta^2
\gamma^2}A_3\right) + \hat{\vec{\gamma}}_\perp (i \vec{\nabla}_\perp + i {\rm
grad}_\perp \ln \phi )-m \right]\tilde{\psi}\\
&&=0\end{aligned}$$ In the last step we introduced $\tilde{\psi}=\psi/\phi$.
In the ultrarelativistic limit terms of the order $1/\gamma^2$ are neglected. We end up with a Dirac equation coupled to a purely transverse vector potential exhibiting a Heaviside step function dependence $\sim \theta
(t-z)$. These properties are essential for our considerations.
Ultrarelativistic limit of the potential {#apotlim}
========================================
In this section we want to discuss the limit $\beta \to 1$ of the potential (\[lwpot\]). From section (\[1iontrafo\]) we expect the asymptotic form of the potential to be $$\lim_{\beta \to 1}\frac{-1}{\sqrt{(z-t)^2+\frac{x_\perp^2}{\gamma^2}}} =
\delta (z-t) \ln (x_\perp^2) + {\cal C}'$$ $t$ integration of (\[lwpot\]) determines ${\cal C}'$ $$\label{cprime}
{\cal C}'=\frac{-1}{|z-t|}- \delta (z-t)\ln (\gamma^2)$$ where we had to require $\gamma \gg x_\perp/|z-\beta t|$.\
An attempt to derive the limit by means of Fourier transformation of (\[lwpot\]) with respect to $z$ was presented in [@Jackiw]. The Fourier transform reads $$\begin{aligned}
\int dz e^{i\omega z} \frac{1}{\sqrt{(z-\beta t)^2+\frac{x_\perp^2}{\gamma^2}}}
&=& e^{i\omega \beta t}\int dz e^{i\omega z}
\frac{1}{\sqrt{z^2+\frac{x_\perp^2}{\gamma^2}}}\nonumber \\
&=& 2e^{i\omega \beta t} K_0\left(\frac{\omega
x_\perp}{\gamma}\right)\nonumber \\
\label{FTlim}
&\stackrel{\gamma\to
\infty}{\longrightarrow}& -2e^{i\omega t}\ln
\left(\frac{\omega x_\perp e^C}{2 \gamma}\right)\end{aligned}$$ The quantity $C$ here denotes Euler’s constant. The inverse Fourier transformation of this expression yields $$\label{arbilim}
\lim_{\beta \to 1}\frac{1}{\sqrt{(z-t)^2+\frac{x_\perp^2}{\gamma^2}}}
=\frac{1}{|z-t|} + g(x_\perp)\delta (z-t)$$ The coefficient $g(x_\perp)$ of the delta distribution in this result is not uniquely specified. Naive application of the textbook formula [@Lighthill] $$\label{lighthillf}
\int \frac{dk}{2\pi} \ln |k| e^{ikx} = -\frac{1}{2|x|}$$ would give $g(x_\perp)=2[\ln(x_\perp/2\gamma)+C]$, but (\[lighthillf\]) is valid only up to arbitrary multiples of $\delta (x)$. The validity of (\[FTlim\]), however, as well demands the condition $\gamma \gg \omega x_\perp$.
It is possible to find a gauge transformation that removes both, the long-range potential $1/|z-t|$ as well as $\delta (z-t)\ln (\gamma^2)$ (\[cprime\]). This is fulfilled with the gauge transformation [@Aichelburg] $$\label{gaugeob2}
\psi'=e^{iZ\alpha \ln\left(\gamma (z-t) +
\sqrt{1+\gamma^2(z-t)^2}\right)}\psi$$ The gauge-transformed potential reads $$\label{gpot}
A_0'=-\frac{Z\alpha\gamma}{\sqrt{\gamma^2(z-\beta t)^2 + x_\perp^2}} +
\frac{Z\alpha\gamma}{\sqrt{\gamma^2(z-\beta t)^2 + 1}}$$ and has the ultrarelativistic limit $$\label{potlim}
\lim_{\beta\to 1}A_0'=+\delta (z-t) \ln (x_\perp ^2)$$
The appearance of the logarithm follows immediately from the inhomogeneous Maxwell equations in the Lorentz gauge that reduces to a two dimensional Poisson equation in the limit $\beta\to 1$.
This gauge transformation has the advantage to yield a short-range potential that allows for asymptotic plane wave solutions. For this reason it was used to obtain a faster convergence in coupled channel calculations [@Eichler].
Solution of the Dirac equation with two discontinuities {#asolu2ion}
=======================================================
The Dirac equation for an electron moving in the field of two ultrarelativistic colliding ions $A$ and $B$ reads $$\label{de2ion}
\left[\frac{1}{2}\left(\hat{\gamma}_-i\partial_{\tau_-}+
\hat{\gamma}_+i\partial_{\tau_+}\right) +i\hat{\vec{\gamma}}_\perp
\cdot \vec{\nabla}_\perp -m -\frac{1}{2}\hat{\gamma}_-\delta (\tau_-)
V^A_\perp -\frac{1}{2}\hat{\gamma}_+\delta (\tau_+) V^B_\perp \right] \psi=0$$ where we used light cone variables $\tau_\pm=(t\pm z)/2$. One directly finds that $\hat{\gamma}_- \psi$ is discontinuous at $\tau_-$ through the action of ion $A$ and $\hat{\gamma}_+\psi$ is discontinuous at $\tau_+=0$ through the action of ion $B$, respectively.
We introduce $\psi_\pm=(1\pm \hat{\alpha}_z)\psi$ and use $2\psi=\psi_-+\psi_+$ to formulate the problem as follows $$\left(i\partial_{\tau_+} +i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m -\delta (\tau_+) V^B_\perp
\right)\psi_+ +\left(i\partial_{\tau_-}
+i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m -\delta (\tau_-) V^A_\perp\right)\psi_-=0$$ where (\[de2ion\]) has been multiplied by $2\hat{\gamma}_0$. This we rewrite as $$\begin{aligned}
&&\left(i\partial_{\tau_+} -\delta (\tau_+) V^B_\perp
\right)\psi_+ +\left(i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m\right)\psi_-\nonumber \\
\label{delambda}
&&=-\left(i\partial_{\tau_-}-\delta (\tau_-)
V^A_\perp\right)\psi_- -\left(i\hat{\vec{\alpha}}_\perp\cdot
\vec{\nabla}_\perp
-\gamma_0m\right)\psi_+\end{aligned}$$ By using the standard representation of Dirac matrices and simply rearranging the four equations (\[delambda\]) one obtains $$\begin{aligned}
\nonumber
&&\left[\left(\begin{array}{rr}1\!{\rm l}&0\\0&0\end{array}\right)
\left(i\partial_{\tau_+}
-\delta (\tau_+) V^B_\perp
\right)
+\left(\begin{array}{rr}0&0\\0&1\!{\rm l}\end{array}\right)
\left(i\partial_{\tau_-}-\delta (\tau_-)
V^A_\perp\right)-
m\left(\begin{array}{rr}0&1\!{\rm l}\\1\!{\rm l}&0\end{array}\right)\right.\\
&&\left.+i\partial_x
\left(\begin{array}{rr}0&-\sigma_y\\ \sigma_y&0\end{array}\right)
+i\partial_y\left(\begin{array}{rr}0&-i\sigma_x\\i\sigma_x&0\end{array}\right)
\right]\tilde{\psi}=0\\
\nonumber
{\rm where}
&&\tilde{\psi}=\left(\begin{array}{c}\psi_1+\psi_3\\ \psi_2-\psi_4\\
\psi_1-\psi_3\\ \psi_2+\psi_4\end{array}\right)\end{aligned}$$ corresponding to an isomorphic linear transformation [@Segev] with the matrix $$\label{segevtrafo}
\Lambda=\left(\begin{array}{rr}
1\!{\rm l}&\sigma_z\\1\!{\rm l}&-\sigma_z\end{array}\right)$$ Since $\Lambda$ is a bijection, each side of (\[delambda\]) has to be zero. Off the light fronts we therefore have the two equations $$\begin{aligned}
\label{psi+}
i\partial_{\tau_+}\psi_+&=&(i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m)\psi_-\\
i\partial_{\tau_-}\psi_-&=&(i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m)\psi_+\end{aligned}$$ According to (\[psiout\]) the discontinuities at the light fronts are described by $$\psi_-(\tau_-=0^+)=\phi^A(x_\perp)\psi_-(\tau_-=0^-)\;\;,\;\;\;
\psi_-(\tau_+=0^+)=\phi^B(x_\perp)\psi_+(\tau_-=0^-)$$ $\phi^A$ and $\phi^B$ are defined by (\[phidef\]) using the scalar parts of the potentials of the ions $A$ and $B$. Let us study the spinor $\psi_+$, evaluated at the surface $\tau_+=0^+$: $$\label{taupgr0}
\psi_+(\tau_+=0^+)=\phi^B \frac{i\hat{\vec{\alpha}}_\perp\cdot
\vec{\nabla}_\perp
-\gamma_0m}{p_-}\psi_-(\tau_+=0^-)$$ In the region $\tau_->0$ the electron already has interacted with ion $A$ and we can write $$\label{taupmgr0}
\psi_+(\tau_+=0^+,\tau_->0)
=\phi^B \frac{i\hat{\vec{\alpha}}_\perp\cdot
\vec{\nabla}_\perp
-\gamma_0m}{p_-}\phi^A (1-\alpha_z)\psi_p$$ where $\psi_p$ is the incoming plane wave at momentum $p$. This relation also can be obtained imeediately from (\[mtrafomi\]) and (\[psi+\]) for $\tau_+$ and $\tau_->0$. The operator $i\partial_+$ in (\[psi+\]) has been replaced by its eigenvalue $p_-$, the incoming negative light cone momentum. This is possible since $p_-$ is conserved in the interaction with ion $A$. The expansion of $\psi_+(\tau_+=0^+)$ in the plane-waves basis reads $$\psi_+(\tau_+=0^+,\tau_->0)
=\int\frac{dp'_+d^2p'_\perp}{(2\pi)^3}B(p',p)e^{-ip'_+\tau_-
+i\vec{p'}_\perp\cdot \vec{x}_\perp}u(p')$$ where we substituted $d^3p' \to dp'_+d^2p'_\perp$ [@Meggiolaro]. According to (\[taupgr0\]) the expansion coefficients are $$B(p',p)=\int_0^\infty d\tau_-\int d^2x_\perp e^{ip'_+\tau_-
-i\vec{p'}_\perp\cdot \vec{x}_\perp}\phi^B
\overline{u}(p')\frac{i\hat{\vec{\alpha}}_\perp\cdot \vec{\nabla}_\perp
-\gamma_0m}{p_-}\psi_-(\tau_+=0^-,\tau_->0)$$ In the region $\tau_->0$, $\tau_+<0$ the wave function $\psi_-$ is a freely propagating wave packet with a fixed light cone momentum $p_-$ and a superposition of transverse momenta $p_\perp$. The mass shell condition requires $p_-p_+=\vec{q}^2_\perp +m^2$. In this way $\psi_-(\tau_+<0,\tau_->0)$ can be obtained from $\psi_-(\tau_+<0,\tau_-=0^+)$. We have $$\psi_-(\tau_->0)=
\int\frac{d^2q_\perp}{(2\pi)^2}
e^{-i\frac{q_\perp^2+m^2}{p_-}\tau_--ip_-\tau_++
i\vec{q}_\perp\cdot \vec{x}_\perp}
\int d^2x'_\perp e^{i\vec{x'}_\perp\cdot
(\vec{p}_\perp-\vec{q}_\perp)}\phi^A
(1-\hat{\alpha}_z)u(p)$$ which leads to $$\begin{aligned}
B(p',p)&=&i\int \frac{d^2q_\perp}{(2\pi)^2} \int d^2x'_\perp
e^{i\vec{x'}_\perp\cdot (\vec{q}_\perp-\vec{p'}_\perp)}\phi^B\int d^2x_\perp
e^{i\vec{x}_\perp\cdot (\vec{p}_\perp-\vec{q}_\perp)}\phi^A\nonumber \\
\label{bexp}
&&\overline{u}(p')
\frac{\hat{\vec{\alpha}}_\perp\vec{q}_\perp -\gamma_0m}
{p'_+p_--q_\perp^2-m^2+i\epsilon}(1-\hat{\alpha}_z)u(p)\end{aligned}$$ Note, that the lower bound of the $\tau_-$ integration is 0, since we inserted the expression of $\psi_-$ for $\tau_->0$.
Together with the corresponding term for the reverse order of interactions with the two ions, (\[bexp\]) is the $S$ matrix for an electron scattered at the light fronts, first derived by Segev and Wells [@Segev] in an elegant way using the transformation (\[segevtrafo\]).
If both ions $A$ and $B$ move on positive light cones separated by the spatial distance $a\vec{e}_z$ (see section \[crystal\]), we obtain with (\[mtrafomi\]) for the interacting part of the spinor $\psi$ $$\begin{aligned}
\psi_+(\tau_+-a/2=0^+)&=&\phi^A\psi_+(\tau_+-a/2=0^-)\nonumber \\
&=&\phi^A\int\frac{d^2q_\perp}{(2\pi)^2}
e^{-i\frac{q_\perp^2+m^2}{p_+}\frac{a}{2}-ip_+\tau_-
+i\vec{q}_\perp\cdot \vec{x}_\perp}\nonumber \\
&&\int d^2x'_\perp e^{i\vec{x'}_\perp\cdot
(\vec{p}_\perp-\vec{q}_\perp)}\phi^B
(1+\hat{\alpha}_z)u(p)\end{aligned}$$ The expansion of $\psi_+$ in plane waves at the point $\tau_+=a/2+0^+$ yields the $S$ matrix of this process in momentum space $$\begin{aligned}
S(p',p)&=&2\pi \delta(p'_+-p_+)i\int \frac{d^2q_\perp}{(2\pi)^2}
e^{i\left(-\frac{q_\perp^2+m^2}{2p'_+}+\frac{p'_-}{2}\right)a}\nonumber \\
&& \int d^2x_\perp e^{i\vec{x}_\perp\cdot (\vec{q}_\perp-\vec{p'}_\perp)}
\phi^A \int d^2x_\perp
e^{i\vec{x}_\perp\cdot (\vec{p}_\perp-\vec{q}_\perp)}\phi^B \overline{u}(p')
(1+\hat{\alpha}_z)u(p)\end{aligned}$$ in accordance with section \[crystal\].
The photon propagator at high collision energies {#approp}
================================================
The four-dimensional Fourier transform of the potential (\[lwpot\]) reads $$\label{ftpot}
\int d^4x e^{ikx}\frac{-Z\alpha \gamma}{\sqrt{\gamma^2(z-\beta t)^2 +
\vec{x}^2_\perp}}=-(2\pi)^2 Z \alpha \delta (k_0-\beta k_3)
\frac{2}{\left(\frac{k_3}{\gamma}\right)^2 + k_\perp^2}$$ which has the following low and high-velocity limits $$\begin{aligned}
\lim_{\beta \to 0}&=&-(2\pi)^2 Z \alpha \delta (k_0)\frac{2}{|\vec{k}|^2}\\
\label{helimft}
\lim_{\beta \to 1}&=&-(2\pi)^2 Z \alpha \delta (k_0-k_3)\frac{2}{k^2_\perp}\end{aligned}$$
The last expression reflects the observation, that in the high-energy limit the longitudinal components $k_-$ and $k_+$ of the photon momentum can be dropped.
After having performed the gauge transformation (\[gaugeob2\]) and taken the limit $\gamma \to \infty$, the potential to be transformed is expression (\[potlim\]). Grignani and Mintchev [@Grignani] have shown, that it is wrong to identify the Fourier transform of (\[potlim\]) with (\[helimft\]) or with the regulated expression $1/(k_\perp^2 + \mu^2)$ with a regulating mass inserted by hand.
Calculating the time integral of $A_0$ in the eikonal expression and using (\[ftpot\]) one finds $$\int^\infty_{-\infty}dt A_0 =Z\alpha \ln (x_\perp^2) +Z\alpha
\lim_{\mu\to 0} \ln \mu^2$$ with $1/2e^C$ absorbed in $\mu$ as in [@Jackiw]. The term $\lim_{\mu\to 0} Z\alpha\ln \mu^2$ is the term ${\cal C}$ in section \[1iontrafo\] and is completely different from $\mu$ in eq (7) of [@Jackiw]!
One may attempt to calculate the two-dimensional Fourier transform of the logarithm from a Taylor expansion in powers of $iZ\alpha$ of the Fourier transform of the $T$ matrix (i.e. its transverse part) which is given by the following closed expression $$\begin{aligned}
{\cal T}(k_\perp)&=&
\left(\frac{4}{k^2_\perp}\right)^{1-i\alpha Z}\Gamma^2(1-i\alpha Z)
\sin(\pi i\alpha Z)\\
&=&\pi \frac{\Gamma(1-i\alpha Z)}{\Gamma(i\alpha
Z)}\left(\frac{4}{k^2_\perp}\right)^{1-i\alpha Z} \end{aligned}$$ The first terms of the Taylor expansion read $$\begin{aligned}
{\cal T}(k_\perp)&\approx& +4\pi i\alpha Z
\frac{1}{k^2_\perp} + 4\pi (i\alpha Z)^2 \frac{\ln
(k^2_\perp/4)+C}{k^2_\perp}\nonumber \\
\label{ttaylor}
&&+ 2\pi
(i\alpha Z)^3 \frac{\ln^2(k^2_\perp/4)+ 4C\ln(k^2_\perp/4)+4C^2}
{k^2_\perp} + \dots\end{aligned}$$ The second term would then correspond to the desired Fourier transform (times $(iZ\alpha)$), the third term correspondingly to $(iZ\alpha)^2$ times the Fourier transform of the square of the transverse part of the potential (\[potlim\]) that has to be compared with the result of Torgerson [@Torgerson].
This is, however, not justified, since the linearity of the Fourier transform is only guaranteed for finite sums and causes problems when applied to infinite series like the Taylor expansion of the exponential function. To get the correct result for the exact two dimensional euclidean photon propagator, the limit $iZ\alpha \to 0$ in $$\label{2dimpott}
\int d^2x_\perp e^{-i\vec{k}_\perp\vec{x}_\perp}\ln x_\perp^2
=\lim_{iZ\alpha\to 0}\frac{d}{d(iZ\alpha)}\left(\pi \frac{\Gamma(1-i\alpha
Z)}{\Gamma(i\alpha
Z)}\left(\frac{4}{k_\perp^2}\right)^{1-i\alpha Z}
\right)$$ has to be taken after having integrated the result with a test function. Performing the limit without this precaution gives the wrong result (\[helimft\]).
Another form of the correct Fourier transform was derived in [@Ferrari]. We obtain the equivalent form from the gauged potential $A'_0$ in (\[gpot\]). Since $$\label{lnpotdef}
-\ln x_\perp^2=\lim_{\gamma\to \infty}\int_{-\epsilon}^\epsilon dt \left(
\frac{\gamma}{\sqrt{\gamma^2t^2+x_\perp^2}}-
\frac{\gamma}{\sqrt{\gamma^2t^2+1}} \right)$$ ($\epsilon$ is arbitrary but finite) and $$\int dtd^2x_\perp e^{i\omega t-i\vec{k}_\perp\cdot\vec{x}_\perp} \left(
\frac{\gamma}{\sqrt{\gamma^2t^2+x_\perp^2}}-
\frac{\gamma}{\sqrt{\gamma^2t^2+1}} \right)=4\pi\left(
\frac{1}{\left(\frac{\omega}{\gamma}\right)^2+k_\perp^2}-2\pi
K_0\left(\frac{\omega}{\gamma}\right)\right)$$ we find by direct substitution $$-\int d^2x_\perp e^{-i\vec{k}_\perp\vec{x}_\perp}\ln x_\perp^2
=\lim_{\lambda \to 0}4\pi\left(\frac{1}{k_\perp^2+ \lambda^2}+\pi\delta^2
(k_\perp)\ln\left(\frac{\lambda^2}{\mu^2}\right)\right)$$ with $\lambda =
\omega/\gamma$, $\mu=2/e^C$. The limit has to be treated in the same way as in (\[2dimpott\]).
When naively taking the limit $iZ\alpha \to 0$ immediately, by chance one obtains the high energy limit of the ungauged potential.
[99]{} H.D.I. Abarbanel, C. Itzykson, Phys. Rev. Lett. 23 (1969) 53 R. Torgerson, Phys. Rev. 143 (1966) 1194 R. Jackiw, D. Kabat, M. Ortiz, Phys. Lett. B 277 (1992) 148 G. ’t Hooft, Phys. Lett. B 198 (1987) 61 S.J. Chang, S.K. Ma, Phys. Rev. 188 (1969) 2385; S.J. Chang, S.K. Ma, Phys. Rev. Lett. 22 (1969) 1336 S.J. Chang, P.M. Fishbane, Phys. Rev. D 2 (1970) 1104 B. Segev, J.C. Wells, Phys. Rev. A 57 (1998) 1849 C. Bottcher, M.R. Strayer, Phys. Rev. D 39 (1989) 1330 D.C. Ionescu, J. Eichler, Phys. Rev. A 48 (1993) 1176 R.M. Miura, J. Math. Phys. 9 (1968) 1202 P.C. Aichelburg, R.U. Sexl, Gen. Rel. Grav. 2 (1971) 303 H.M. Fried, Functional Methods and Models in Quantum Field Theory, MIT Press (1972) E. Meggiolaro, Phys. Rev. D 53 (1996) 3835 C. Itzykson, J.B. Zuber, Quantum Field Theory, McGraw-Hill (1985) A.J. Baltz, Phys. Rev. Lett. 78 (1997) 1231 G. Grignani, M. Mintchev, Phys. Rev. D 38 (1988) 3163 R. Ferrari, Nuovo Cimento A 19 (1974) 204 U. Eichmann, J. Reinhardt, W. Greiner, manuscript in preparation M.J. Lighthill, Introduction to Fourier analysis and generalized functions, Cambridge Univ. Press (1959) N. Toshima, J. Eichler, Phys. Rev. A 42 (1990) 3896
[^1]: The effect of the potential (\[potlimc\]) also can be described within the Aichelburg-Sexl metric. Two field-free regions of space-time meet at $z=t$, such that (the superscripts $<$ and $>$ denote $t>z$ and $t<z$, respectively) $$\begin{aligned}
\label{xtrgr}x_\perp^>&=&x_\perp^<\\
\label{zgr}z^>&=& z^<-Z\alpha \ln x_\perp^2\\
\label{tgr}t^>&=& t^<-Z\alpha \ln x_\perp^2\end{aligned}$$ The result (\[psiout\]) is then easily obtained by simply substituting (\[xtrgr\])-(\[tgr\]) into the plane wave at $t>z$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider a nonlocal aggregation equation with nonlinear diffusion which arises from the study of biological aggregation dynamics. As a degenerate parabolic problem, we prove the well-posedness, continuation criteria and smoothness of local solutions. For compactly supported nonnegative smooth initial data we prove that the gradient of the solution develops $L_x^\infty$-norm blowup in finite time.'
address:
- 'School of Mathematics, Institute For Advanced Study, Princeton NJ 08544'
- 'School of Mathematics, Institute For Advanced Study, Princeton NJ 08544'
author:
- Dong Li
- Xiaoyi Zhang
date: submitted Jun 2008
title: On a nonlocal aggregation model with nonlinear diffusion
---
Introduction and main results
=============================
In this paper we consider the following evolution equation with nonlinear diffusion: $$\begin{aligned}
\label{eq1}
\begin{cases}
\partial_t u + \partial_x (u \partial_x K * u) = r\partial_x (u^2 \partial_x u), \quad (t,x)
\in (0,\infty)\times{{\mathbb{R}}},\\
u(0,x)=u_0(x),
\end{cases}\end{aligned}$$ where $K$ is an even function of $x$ and has a Lipschitz point at the origin, e.g. $K(x) =e^{-|x|}$. Here $*$ denotes the usual spatial convolution on ${{\mathbb{R}}}$. The parameter $r> 0$ measures the strength of the nonlinear diffusion term. The function $u=u(t,x)$ represents population density in biology or particle density in material science. (see, for example, [@TBL06] [@EkWG98] [@HP05] [@HP06] [@Mek99] [@ToBe04] [@LRC01] [@TT98] [@FGLO99] [@Mu02] [@O80]). This equation was used in the study of biological aggregation such as insect swarms, fish schools and bacterial colonies. It is first derived by Bertozzi, Lewis and Topaz [@TBL06] as a modification (more precisely, the addition of the density-dependent term on the RHS of ) of an earlier classical model of Kawasaki [@K78]. According to [@TBL06], this modification gives rise to biologically meaningful clumping solutions (i.e. densities with compact support and sharp edges). For other similar one-dimensional models and their biological applications we refer the readers to [@Mek99] [@EkWG98] [@VCFH99] [@PD84] [@HM89] [@I84] [@I85] [@IN87] [@K78] [@MY82] [@NM83] and the references therein for more extensive background and reviews.
To understand the biological meaning of each term, one can rewrite as the classical continuity equation $$\begin{aligned}
\partial_t u + \partial_x ( u v) =0,\end{aligned}$$ where the velocity $v$ is related to the density $u$ by $$\begin{aligned}
v &= \partial_x (K* u) - r u\partial_x u \\
&=: v_a +v_d.\end{aligned}$$ Here $v_a$ is called the attractive velocity since as explained in [@TBL06] individuals aggregate by climbing gradients of the sensing function $s=K*u$. Due to the spatial convolution $v_a$ is a nonlocal transformation of the density $u$. The second term $v_d$ is called the dispersal(repulsive) velocity and is a spatially local function of both the population and the population gradient. Biologically $v_d$ represents the anti-crowding mechanism which operates in the opposite direction of population gradient and decreases as the population density drops. These constitutive relations of $v_a$, $v_d$, $u$ are natural in view of the basic biological assumption that aggregation occurs on a longer spatial scale than dispersal.
In the mathematics literature, the aggregation equations which have similar forms to have been studied extensively [@BeLa07] [@BoVe05] [@BoVe06] [@BuDiF06] [@BuCaMo07] [@La07] [@ToBe04]. The case of with $r=0$ and general choices of the kernel $K$ was considered by Bodnar and Velazquez [@BoVe06]. There by an ODE argument the authors proved the local well-posedness of *without the density-dependent term* for $C^1$ initial data. For a generic class of choices of the kernel $K$ and initial data, they proved by comparing with a Burgers-like dynamics, the finite time blowup of the $L_x^\infty$-norm of the solution. Burger and Di Francesco [@BuDiF06] studied a class of one-dimensional aggregation equations of the form $$\begin{aligned}
\partial_t u = \partial_x \left( u \partial_x (a(u) -K*u +V ) \right), \quad\text{ in }\,
{(0,\infty)\times{{\mathbb{R}}}},\end{aligned}$$ where $V:\, {{\mathbb{R}}}\to {{\mathbb{R}}}$ is a given external potential and the nonlinear diffusion term $a(\rho)$ is assumed to be either $0$ or a strictly increasing function of $\rho$. In the case of no diffusion ($a\equiv 0$) they proved the existence of stationary solutions and investigated the weak convergence of solutions toward the steady state. In the case of sufficiently small diffusion ($a(\rho)=\epsilon \rho^2$) they proved the existence of stationary solutions with small support. Burger, Capasso and Morale [@BuCaMo07] studied the well-posedness of an equation similar to but with a different diffusion term: $$\begin{aligned}
\partial_t u +\nabla \cdot (u \nabla K* u) = div( u \nabla u),\quad \text{in}\;
{(0,T)\times{{\mathbb{R}}}^d}.
\end{aligned}$$ For initial data $u_0 \in L^1_x({{\mathbb{R}}}^d) \cap L^\infty_x({{\mathbb{R}}}^d)$ with $u_0^2 \in H^1_x({{\mathbb{R}}}^d)$, they proved the existence of a weak solution by using the standard Schauder’s method. Moreover the uniqueness of entropy solutions was also proved there. In connection with the problem we study here, Laurent [@La07] has studied in detail the case of *without density-dependent diffusion* (i.e. $r=0$ ) and proved local and global existence results for a class of kernels $K$ with $H_x^s(s\ge 1)$ initial data. More recently Bertozzi and Laurent [@BeLa07] have obtained finite-time blowup of solutions for the case of without diffusion (i.e. $r=0$) in ${{\mathbb{R}}}^d (d\ge 2)$ assuming compactly supported radial initial data with highly localized support. Li and Rodrigo [@LR08] [@LiRod08prep] [@LiRod09] studied the case of with fractional dissipation and proved finite-time blowup or global wellposedness in various situations. We refer to [@LiRodZhang09] [@LiZhang09] for the cases with singular kernels and further detailed studies concerning sharp asymptotics and regularity of solutions. We also mention that Bertozzi and Brandman [@BB08] recently constructed $L_x^1\cap L_x^\infty$ weak solutions to in $R^d$ ($d\ge 2$) and with no dissipation ($r=0$) by following Yudovich’s work on incompressible Euler equations [@Yo63]. We refer the interested readers to [@PD84] [@HM89] [@I84] [@I85] [@IN87] [@K78] [@MY82] [@NM83] and the references therein for some further rigorous studies.
From the analysis point of view, equation is also connected with a general class of degenerate parabolic equations known as porous medium equations, which takes the form $$\begin{aligned}
\label{eq_poro}
\partial_t u= \partial_x ( u^m \partial_x u),\quad (t,x)\in (0,\infty)\times{{\mathbb{R}}},\end{aligned}$$ where $m$ is a real number. These equations describe the ideal gas flow through a homogeneous porous medium and other physical phenomena in gas dynamics and plasma physics [@M37] [@ZR66] [@Ar86]. Oleinik, Kalashnikov and Chzhou [@OKC58] proved the existence and uniqueness of a global weak solution of . Hölder continuity of the weak solution is established in [@Ar69] [@Kr69] [@Gi76] [@CF79]. As for higher regularities, the Lipschitz continuity of the pressure variable $v=\tfrac{m+1}m u$ is shown in [@Ar69] [@Di79] [@Be83] [@AC86] [@CVW87]. Aronson and Vazquez [@AV87] showed that $v$ becomes $C^\infty$ outside of the free boundary after some waiting time. For derivative estimates of the solution $u$, a local solution of in $W^{1,\infty}_x({{\mathbb{R}}})$ is constructed by Otani and Sugiyama [@OS96]. In the case of with $m$ being an even natural number, Otani and Sugiyama [@OS01] proved the existence of smooth solutions. We refer the interested readers to [@AABL01] [@AG93] [@Be72] [@BCW00] [@BV05] [@CC95] [@CF80] [@CV99] [@CVW87] [@CCV06] [@CEL84] [@CL71] [@Ko99] [@Va07] and references therein for more detailed studies and expositions.
Our starting point of the analysis of equation is to treat it as a degenerate parabolic problem with a nonlocal flux term. We focus on constructing and analyzing classical solutions of . The bulk of this paper is devoted to proving the existence and uniqueness of classical solutions to , which is done in section 2 and part of section 3. In the final part of section 3, we prove the continuation and blowup criteria of solutions. In the last section we prove that any smooth initial data with compact support will lead to blowup of the gradient in finite time. The analysis developed in this work can be extended to treat the $d$-dimensional ($d\ge 2$) case of which we will address in a future publication.
We now state more precisely our main results. The first theorem establishes the existence of smooth local solutions for initial data which is not necessarily nonnegative.
\[thm1\] Assume $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R)$ ($u_0$ is not necessarily nonnegative). Then there exists a positive $T_0 =T_0(\|u_0\|_{2}+\|\partial_x u_0 \|_{\infty})$ such that has a unique solution $u \in C^\infty([0,T_0]\times
{{\mathbb{R}}})$. In particular $u \in C([0,T_0]\cap H^k_x)$ for any $k\ge 0$. If $u_0\ge 0$, then $u(t)\ge 0$ for all $0\le t\le T_0$.
The assumptions on the initial data $u_0$ in Theorem \[thm1\] can be weakened significantly (see for example Theorem \[thm\_local\]). However in order to simplify the presentation, we do not state our theorems here in its most general form.
Our second theorem gives the blowup or continuation criteria of solutions. Roughly speaking, it says that all the $L^p_x$-norm of the solution cannot blow up and we can continue the solution as long as we have a control of the gradient of the solution.
\[thm2\] Assume $u \in C^\infty([0,T)\times {{\mathbb{R}}})$ is a maximal-lifespan solution obtained in Theorem \[thm1\]. Then either $T=+\infty$ or $T<+\infty$ and $$\begin{aligned}
\lim_{t\to T} \| \partial_x u(t) \|_{\infty} =+\infty.\end{aligned}$$ For any $2\le p<\infty$, there exists a generic constant $C$ such that $$\begin{aligned}
\| u(t) \|_{p} \le \|u_0 \|_{p} e^{Cpt}, \quad \forall\, t\in\ [0,T).\end{aligned}$$ If in addition $u_0\ge 0$, then $u(t) \ge 0$ for all $t\in [0,T)$, and we also have the $p$-independent estimate for all $2\le p\le
+\infty$: $$\begin{aligned}
\| u(t) \|_{p} \le \|u_0\|_{p} \exp \left ( Ct \|u_0\|_{2} e^{Ct} \right), \quad\forall\,
t\in[ 0,T).\end{aligned}$$ In particular if $p=+\infty$, then $$\begin{aligned}
\| u(t) \|_{\infty} \le \|u_0\|_{\infty} \exp\left( Ct \|u_0\|_{2} e^{Ct} \right),
\quad\forall\, t\in [0,T).\end{aligned}$$
The next theorem states that if we assume the initial data has a little bit more integrability, then the local solution will inherit this property. Note in particular that the $L^1_x$-norm of the solution is preserved for all time if the initial data is nonnegative and in $L^1_x({{\mathbb{R}}})$.
\[thm3\] Assume $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R)$ and $u_0 \in
L_x^p({{\mathbb{R}}})$ for some $1\le p<2$. Then the local solution obtained in Theorem \[thm1\] also satisfies $u\in C([0,T], L_x^p)$. If in addition $u_0 \ge 0$, then $u(t)\ge 0$ for any $t\in [0,T]$. If also $p=1$, then $\|u(t)\|_{1} = \|u_0\|_{1}$, i.e. $L_x^1$-norm of the solution is preserved.
The last theorem states that any solution with smooth nonnegative initial data will blow up in finite time.
\[thm4\] Let $u_0 \in C_c^\infty({{\mathbb{R}}})$ and $u_0 \ge 0$. Assume $u_0$ is not identically $0$. Then there exists time $T<\infty$ such that the corresponding solution with $u_0$ as initial data blows up at time $T$ in the sense that $$\begin{aligned}
\label{eq_thm41}
\lim_{t\to T} \| \partial_x u(t) \|_{\infty} =+\infty.\end{aligned}$$ However all the $L^p_x$-norm of $u$ remain finite at the time of blowup, i.e.: $$\begin{aligned}
\label{eq_thm42}
\sup_{0\le t < T} \| u(t) \|_{p} \le C(T)<\infty, \forall\; 1\le p \le \infty.\end{aligned}$$
In the case of *without diffusion* (i.e. $r=0$ in ), if we take the kernel $K(x)=e^{-|x|}$, then the result of Bodnar and Velazquez [@BoVe06] says that any solution with initial data $u_0$ satisfying a slope condition will blow up in finite time in the sense that $\|u(t)\|_{\infty}$ blows up. This is highly in contrast with our result here when the diffusion term does not vanish. In this case the solution will blow up at the level of the gradient, i.e. we have $\|\nabla u(t)\|_{\infty}$ tends to infinity while all the other $L^p_x$-norm remain finite as $t\to T$, where $T$ is the blowup time.
**Notations.** Throughout the paper we denote $L_x^p=L_x^p({{\mathbb{R}}})$ for $1\le p \le \infty$ as the usual Lebesgue space on ${{\mathbb{R}}}$. We also write $\|\cdot\|_p = \|\cdot\|_{L_x^p}$. For $s>0$, $s$ being an integer and $1\le p\le \infty$, $W^{s,p}_x = W^{s,p}_x({{\mathbb{R}}})$ denotes the usual Sobolev space $$\begin{aligned}
W^{s,p}_x = \bigl\{ f \in S^\prime({{\mathbb{R}}}): \, \| f\|_{W^{s,p}}
= \sum_{0\le j\le s} \| \partial_x^j f \|_{p} <\infty \bigr\}.\end{aligned}$$ When $p=2$, we denote $H^m_x=H^m_x({{\mathbb{R}}}) = W^{2,p}_x({{\mathbb{R}}})$ and $\|\cdot\|_{H^m}$ as its norm. Occasionally we shall use the Sobolev space of fractional power $H_x^s({{\mathbb{R}}})$ whose norm can be defined via Fourier transform: $$\|f\|_{ H^s}=\| (1+|\xi|)^s \hat f(\xi)\|_{L_{\xi}^2}.$$ For any two quantities $X$ and $Y$, we use $X \lesssim Y$ or $Y
\gtrsim X$ whenever $X \leq CY$ for some constant $C>0$. A constant $C$ with subscripts implies the dependence on these parameters. We write $A=A(B_1,\cdots,B_k)$ when we want to stress that a quantity $A$ depends on the quantities $B_1,\cdots,B_k$.
From now on we assume $r=1$ in without loss of generality. Same results hold for any $r>0$.
The regularized equation and its wellposedness
==============================================
Since is a degenerate parabolic equation, in order to construct a local solution, we have to regularize the equation. To this end, we consider the following regularized version of $$\begin{aligned}
\label{eq_WSR_1}
\begin{cases}
\partial_t u = (u^2+\epsilon) \partial_{xx} u + 2 u (\partial_x u)^2 - \partial_x(u\, K*\partial_x u ), \\
u(0,x)=u_0(x).
\end{cases}\end{aligned}$$ Here $\epsilon>0$ is a parameter. We are going to prove the following
\[prop1\]
Assume $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R)$ ($u_0$ is not necessarily nonnegative). Then there exists a positive $T_0 =T_0(\|\partial_x u_0 \|_{\infty}
+\|u_0\|_{2})$ ($T_0$ is independent of $\epsilon$) such that has a unique solution $u^\epsilon \in
C^\infty([0,T_0]\times {{\mathbb{R}}})$ for any $\epsilon>0$. In particular $u^\epsilon \in C([0,T_0];\, H^k_x)$ for any $k\ge 0$.
\[cor1\] Assume $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R)$ and $u_0 \in
L_x^p({{\mathbb{R}}})$ for some $1\le p<2$. Then the local solution obtained in Proposition \[prop1\] also satisfies $u^{\epsilon} \in
C([0,T_0],L_x^p)$. And $T_0$ can be chosen to be $T_0=T_0(\|u_0\|_{p}+\|\partial_x u_0\|_{\infty})$.
In the case $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R) $ with $u_0 \ge 0$. The regularized equation admits a global solution.
\[prop2\] Assume $u_0 \in \bigcap_{m=0}^\infty H^m_x(\mathbb R)$ with $u_0 \ge 0$. Then for any $T>0$ has a unique solution $u^\epsilon \in C^\infty([0,T]\times {{\mathbb{R}}}) $. In particular $u^\epsilon \in C([0,T]; \, H^k_x)$ for any $k\ge 0$. Also $u^\epsilon(t)\ge 0$ for any $t\ge 0$. If in addition $u_0 \in
L_x^p({{\mathbb{R}}})$ for some $1\le p<2$, then $u^\epsilon \in C([0,T];\,
L_x^p)$.
Proof of Proposition \[prop1\]
------------------------------
Our proof of Proposition \[prop1\] is reminiscent of the $L_x^\infty$ energy method used by Otani and Sugiyama [@OS01] where they dealt with the porous medium equation . Denote the set $$\begin{aligned}
B^k_{T} = \bigl\{ v\in C([0,T]; \, H^{2k+1}_x({{\mathbb{R}}})):\; \partial_{xx} v, \partial_t v \in L^2([0,T];\,H^{2k}_x({{\mathbb{R}}})) \bigr\}.\end{aligned}$$ As a very first step, we shall show the local existence of the solution to in $B^k_T$. At this point, we need the following lemma from [@OS01].
\[lem518\] Consider the initial value problem $$\begin{aligned}
\begin{cases}
\partial_t u = (u^2 +\epsilon) \partial_{xx} u + h(t,x), \\
u(0) =u_0,
\end{cases}\end{aligned}$$ where $h \in L^2(0,T; \, H^{2k}_x)$, $u_0 \in H^{2k+1}_x({{\mathbb{R}}})$, there exists a unique function $u=u(t,x) \in B^k_T$ which satisfies, for some constant $C_1=C_1(\| u_0\|_{H^{2k+1}_x}, \| h\|_{L_t^2 H_x^{2k}},
k,\epsilon)$, that $$\begin{aligned}
\| u\|_{L_t^\infty H^{2k+1}_x} + \epsilon^{\frac 12} \| u\|_{L_t^2 H^{2k+2}_x} \le C_1.\end{aligned}$$ For any $h_1$, $h_2 \in K_R^T=\bigl\{v:\, \| v\|_{L^2(0,T;\, H^{2k}_x)} \le R \bigr\}$, there exists a constant $C_2=C_2(R,k,\epsilon)$ such that $$\begin{aligned}
\| u_1 - u_2 \|_{L_t^\infty H^{2k+1}_x} + \epsilon^{\frac 12} \| u_1 -u_2 \|_{L_t^2 H^{2k+2}_x}
\le C_2 e^{C_2 T} \| h_1 -h_2 \|_{L_t^2 H^{2k}_x}.\end{aligned}$$
See [@OS01].
Based on this lemma, we establish the local existence of solutions of the regularized equation. Note however that the time of existence of the solution depends on the regularization parameter $\epsilon$. We have the following
\[lem\_local\_exist\] Let $u_0 \in H^{2k+1}_x({{\mathbb{R}}})$, $k\ge 2$. There exists $T_0=T_0(\|u_0\|_{H^{2k+1}},k, \epsilon)>0$, such that has a unique solution $u \in B^k_{T_0}$.
Introduce the auxiliary equation $$\begin{aligned}
\label{eq538}
\begin{cases}
\partial_t u = (u^2+\epsilon)\partial_{xx} u + h(t,x) \\
u(0)=u_0
\end{cases}\end{aligned}$$ By Lemma \[lem518\], for any $u_0 \in H^{2k+1}_x({{\mathbb{R}}})$, $h\in
L^2(0,T;H^{2k}_x)$, has a unique solution $u_h \in
B^k_T$. Next define the map $$\begin{aligned}
\phi(h) = 2 u_{h} (\partial_x u_h)^2 - \partial_x ( u_h K*\partial_x u_h).\end{aligned}$$ It suffices to show that for some suitable $R$ and $T_0$, $\phi$ is a contraction from the set $K^{T_0}_R = \bigl\{ v\in L^2(0,T_0;\, H^{2k}_x);\, \| v\|_{L^2(0,T_0;H^{2k}_x)} \le R \bigr\}$ into itself. First we show that $\phi$ maps $K_R^{T_0}$ into $K_R^{T_0}$. Take any $h\in K_R^T$, by the fact that $H^{s}_x({{\mathbb{R}}})$ is an algebra when $s>\frac 12$, we have $$\begin{aligned}
\| \phi(h) \|_{H^{2k}} & \le \|u_h \|_{H^{2k}} \| u_h \|_{H^{2k+1}}^2 + \| \partial_x K* \partial_x u_h u_h\|_{H_x^{2k}}
+ \| K* \partial_x u_h \partial_x u_h \|_{H^{2k}} \\
& \le C (\|u_h \|_{H^{2k+1}}^3+\|u_h\|_{H^{2k+1}}^2),\end{aligned}$$ where $C$ is a generic constant. Then Lemma \[lem518\] implies that $$\begin{aligned}
\| \phi(h) \|_{L^2(0,T;\, H^{2k}_x)} &\le CT^{\frac 12} (\|u_h \|_{L_t^{\infty}(0,T;H^{2k+1})}^3
+\|u_h \|_{L_t^{\infty}(0,T;H^{2k+1})}^2)\\
& \le T^{\frac 12} C_3(\|u_0\|_{H_x^{2k+1}}, R,k,\epsilon).\end{aligned}$$ This shows that if we take $T \le C_3^2/R^2$, then $\phi$ maps $K_R^T$ into $K_R^T$. Similarly we can show that $\phi$ is Lipschitz, i.e. for any $h_1$, $h_2 \in K_R^T$, we have $$\begin{aligned}
\| \phi(h_1) - \phi(h_2) \|_{L^2(0,T;\, H^{2k}_x)} \le C_4 T^{\frac 12} e^{C_4 T}
\| h_1- h_2\|_{L^2(0,T;\, H^{2k}_x)},\end{aligned}$$ where $C_4=C_4(\|u_0\|_{H^{2k+1}}, R,k,\epsilon)$. Therefore if $T$ is sufficiently small, then $\phi$ is a contraction on $K_R^T$. The lemma is proved.
Lemma \[lem\_local\_exist\] is not satisfactory since the time of existence of the solution depends on the regularization parameter $\epsilon$. The following lemma removes this dependence, and at the same time, weakens the dependence on the initial norm.
\[lem301\] Let $u_0 \in H^{2k+1}_x({{\mathbb{R}}})$, $k\ge 2$. Then there exists $T_0=T_0(\|\partial_x u_0\|_{\infty}
+\|u_0\|_2)$ such that for any $\epsilon>0$, has a unique solution $u^\epsilon \in B^k_{T_0}$. If $u_0\ge 0$, then $u^\epsilon(t)\ge 0$ for any $0\le t\le T_0$.
By Lemma \[lem\_local\_exist\], we can continue the solution as long as we can control the ${H_x^{2k+1}}$-norm of $u^{\varepsilon}$. In the following we give the a priori control of $u^\epsilon$. As we will see, the $H^{2k+1}_x$-norm of $u^{{\varepsilon}}$ will stay bounded on a time interval $[0,T_0]$ for a certain small $T_0$ depending only on the norms $\|\partial
_xu_0\|_{\infty}$, $\|u_0\|_2$. We are then left with the task of estimating the various Sobolev norms which we will do in several steps. For simplicity of notations we shall write $u^\epsilon$ as $u$ throughout this proof. `Step 1`: $L_x^p$-norm estimate for $2\le p \le +\infty$. Multiply both sides of by $|u|^{p-2} u$ and integrate over $x$, we have $$\begin{aligned}
\label{eq1144}
\frac 1 p \frac d {dt} \| u(t) \|_p^p & =-\epsilon(p-1) \int |u|^{p-2} u_x^2 dx -(p-1) \int u_x^2 |u|^p dx \notag \\
&\qquad+ (p-1) \int (\partial_x K* u) u |u|^{p-2} \partial_x u dx \\
&\le -(p-1) \int |u|^{p-2} \left( uu_x - \frac {\partial_x K*u} 2 \right)^2 dx \notag \\
&\qquad+ \frac {p-1} 4 \int |u|^{p-2} (\partial_x K*u )^2 dx \notag\\
& \le \frac {p-1} 4 \| \partial_x K * u \|_p^2 \| u\|_p^{p-2} \notag\\
& \le \frac {p-1}4 \| \partial_x K \|_1^2 \| u\|_p^p. \notag\end{aligned}$$ Gronwall’s inequality implies that $$\begin{aligned}
\label{eq948}
\| u(t) \|_p \le \|u_0\|_p \cdot \exp\left( \frac{p-1}4 \|\partial_x K\|_1^2 t \right), \quad\forall\, 2\le p<+\infty.\end{aligned}$$ In particular we have $$\begin{aligned}
\| u(t) \|_2 \le \|u_0\|_2 \exp \left( \frac14 \|\partial_x K\|_1^2 t \right).\end{aligned}$$ To control $L_x^\infty$-norm, we are going to choose $T_0<\frac 1
{\|\partial_x K\|_1}$. This implies that $\|u(t)\|_{2}$ has a uniform bound on any such $[0,T_0]$. Now use again the RHS of to get $$\begin{aligned}
\frac 1 p \frac d {dt} \| u(t) \|_p^p & \le (p-1) \int (\partial_x K* u) u |u|^{p-2} \partial_x u dx \\
& =\frac {p-1} p \int (\partial_x K* u) \partial_x (|u|^p) dx \\
&=-\frac {p-1} p \int (\partial_{xx}K*u) |u|^p dx.\end{aligned}$$ Now we use the fact that $K(x) =e^{-|x|}$ and therefore $\partial_{xx} K(x) = -2 \delta(x) +e^{-|x|}$, and this gives $$\begin{aligned}
\frac 1 p \frac d {dt} \| u(t) \|_p^p & \le -\frac {2(p-1)} p \int u |u|^p dx + \frac {p-1} p \int (e^{-|x|}*u) |u|^p dx \\
& \le \frac {2(p-1)} p \| u(t) \|_{p+1}^{p+1} + \frac {p-1} p \| u(t) \|_p^p \| e^{-|x|}\|_2 \| u\|_2 \\
&{\lesssim}\frac {2(p-1)} p \| u(t) \|_{p+1}^{p+1} + \frac {p-1} p \| u(t) \|_p^p \\
& {\lesssim}\frac {2(p-1)} p \| u\|_p^p \|u\|_\infty + \frac {p-1} p \| u\|_p^p.\end{aligned}$$ This implies $$\begin{aligned}
\frac d {dt} \|u(t)\|_p {\lesssim}\frac {2(p-1)} p \|u\|_p \|u\|_{\infty} + \frac {p-1} p \| u(t) \|_p.\end{aligned}$$ Integrating over $[0,t]$, we obtain $$\begin{aligned}
\| u(t) \|_p \le \|u_0\|_p + C\int_0^t (1+\|u(s)\|_{\infty}) \|u(s)\|_p ds.\end{aligned}$$ Letting $p\to \infty$, we obtain $$\begin{aligned}
\|u(t) \|_\infty \le \|u_0\|_\infty + C \int_0^t (1+\|u(s)\|_\infty) \|u(s)\|_\infty ds.\end{aligned}$$ Now it’s easy to see that if we choose $T_0$ sufficiently small depending on $\|u_0\|_2,\|u_0\|_\infty$, then we have $$\begin{aligned}
\label{ru}
\sup_{0\le t\le T_0} \|u(t)\|_{r} \le C \| u_0 \|_r, \quad \forall\, 2\le r\le \infty.\end{aligned}$$
`Step 2`: Control of $\|\partial_x u(t) \|_\infty$. At this point, one can appeal directly to the maximum principle (see for example Theorem 11.16 of [@Lie96]). We give a rather direct estimate here without using the maximum principle. Differentiating both sides of w.r.t $x$ and denoting $v=u_x$, we have $$\begin{aligned}
\label{eq1242}
\partial_t v = \epsilon v_{xx}+u^2 v_{xx} +6 u vv_x +2 v^3 +G,\end{aligned}$$ where $$\begin{aligned}
G= \partial_{xx} (uK*v ).\end{aligned}$$ Multiplying both sides of by $|v|^{p-2}v$ and integrating over $x$, we obtain $$\begin{aligned}
\label{eq1247}
\frac 1 p \frac d {dt} \| v(t) \|_p^p & = \int u^2 v_{xx} |v|^{p-2} v dx + \epsilon \int v_{xx} |v|^{p-2} vdx +6 \int u v_x |v|^p dx \notag\\
&\quad + 2 \int|v|^{p+2} dx + \int G(x) |v|^{p-2} vdx.\end{aligned}$$ Integrate by parts and we have $$\begin{aligned}
\int u^2 v_{xx} |v|^{p-2} v dx &= -2 \int u v_x |v|^p dx - (p-1) \int u^2 v_x^2 |v|^{p-2} dx,\\
\int u v_x |v|^p dx &= - \frac 1{1+p} \int |v|^{p+2} dx \le 0,\\
\epsilon \int v_{xx} |v|^{p-2} vdx &= -\epsilon(p-1) \int |v|^{p-2} v_x^2 dx \le 0.\end{aligned}$$ Plugging the above estimates into , we get $$\begin{aligned}
\frac 1p \frac d{dt} \|v(t)\|_p^p \le 2 \int |v|^{p+2} dx + \int G(x) |v|^{p-2} v dx.\end{aligned}$$ Now we compute $$\begin{aligned}
\int G |v|^{p-2} v dx =& \int \partial_{xx} (K*v u) |v|^{p-2} v dx \\
=& \int \partial_{xx}K* v \, u |v|^{p-2} v dx + 2 \int \partial_x K*v |v|^p dx + \int K*v v_x |v|^{p-2} v dx \\
{\lesssim}& \int |u| |v|^p dx + \int( e^{-|x|}* v) u |v|^{p-2} v dx + \| v\|_p^p \|v \|_\infty\\
{\lesssim}&\|v\|_p^p (\| u\|_\infty + \| v\|_\infty).\end{aligned}$$ Finally we obtain $$\begin{aligned}
\frac 1 p \frac d {dt} \| v(t) \|_p^ p & {\lesssim}\|v \|_{p+2}^{p+2} + \| v\|_p^p ( \| u\|_\infty + \|v\|_\infty) \\
& {\lesssim}\| v\|_p^p ( \|v\|_\infty^2 + \|v\|_\infty + \|u\|_\infty).\end{aligned}$$ Integrating over $t$, letting $p\to \infty$, and recalling $v=u_x$, we have $$\begin{aligned}
\label{rux}
\sup_{0\le t\le T_0} \|\partial_x u(t) \|_\infty & \le C \| \partial_x
u_0\|_\infty,\end{aligned}$$ for sufficiently small $T_0=T_0(\|u_0\|_\infty, \|u_0\|_2, \| \partial_x u_0 \|_\infty)$.
`Step 3:` Control of $\| \partial_{xx} u(t) \|_{p}$ for $2\le
p\le +\infty$. First take $ 2\le p <+\infty$ and compute $$\begin{aligned}
&\frac 1 p \frac d {dt} \int |\partial_{xx} u |^p dx \\
= & \frac 13 \int |\partial_{xx} u|^{p-2} \partial_{xx} u \cdot (u^3)_{xxxx} dx
+ \epsilon \int |\partial_{xx} u|^{p-2} \partial_{xx} u u_{xxxx} dx \\
&\quad -\int |\partial_{xx} u|^{p-2} \partial_{xx} u \partial_{xxx}((\partial_x K*u) u) dx.\end{aligned}$$ We further estimate by using integration by parts, $$\begin{aligned}
\label{eqMay222}
& \int |\partial_{xx} u|^{p-2} \partial_{xx} u \cdot (u^3)_{xxxx} dx \notag\\
= & 3 \int |\partial_{xx} u|^{p-2} \partial_{xx} u \cdot u^2 \partial_{xxxx} udx +24 \int |\partial_{xx} u|^{p-2}
\partial_{xx} u\cdot u\partial_x u \partial_{xxx} u dx \notag\\
&\quad +18\int |\partial_{xx} u|^p u \partial_{xx} u dx + 36\int |\partial_{xx} u|^p (\partial_x u)^2 dx \notag\\
= & - 3(p-1)\int (\partial_{xxx} u)^2 |\partial_{xx} u|^{p-2} u^2 dx +18(1-p)\int |\partial_{xx} u|^{p-2}
\partial_{xx} u\cdot u\partial_x u \partial_{xxx} u dx \notag\\
&\quad +18\int |\partial_{xx} u|^p (\partial_x u)^2 dx.\end{aligned}$$ Using Cauchy Schwartz inequality: $$ab\le \frac 16 a^2+\frac 32 b^2,$$ we bound the second term in as $$\begin{aligned}
|18(1-p)\int &|u_{xx}|^{p-2}\partial_{xx} u u\partial_x
u\partial_{xxx}u dx| \\
&\le 27(p-1)\int|u_{xx}|^p|\partial_x u|^2 dx+3(p-1)\int
|u_{xx}|^{p-2} u^2|\partial_{xxx} u|^2 dx.\end{aligned}$$ Hence, $$\begin{aligned}
\int |\partial_{xx} u|^{p-2} \partial_{xx} u \cdot (u^3)_{xxxx} dx
\le & (27p-9)\int |\partial_{xx} u|^p (\partial_x u)^2 dx.
$$ Also it is obvious that $$\begin{aligned}
&\epsilon \int |\partial_{xx} u|^{p-2} \partial_{xx} u u_{xxxx} dx \\
= & -{\varepsilon}p \int (\partial_{xxx} u)^2 | \partial_{xx} u|^{p-2} dx
\le 0.\end{aligned}$$ Similarly we have $$\begin{aligned}
& \int |\partial_{xx} u|^{p-2} \partial_{xx} u \cdot \partial_{xxx} ( (\partial_x K*u) u) dx \\
{\lesssim}& \int |\partial_{xx} u|^{p-2} \partial_{xx} u \partial_{xxx} u (\partial_x K* u) dx + \int| \partial_{xx} u|^{p} \partial_x K* \partial_x udx \\
&\quad + \int |\partial_{xx} u|^{p-2} \partial_{xx} u \partial_x u \partial_x K* \partial_{xx} u dx + \int |\partial_{xx} u|^{p-2} \partial_{xx} u
u \partial_{xx} K* \partial_{xx} u dx \\
{\lesssim}& \int |\partial_{xx} u|^p \partial_x K* \partial_x u dx + \int |\partial_{xx} u|^{p-2} \partial_{xx} u \partial_x u \partial_x K*
\partial_{xx} udx\\
&\quad + \int |\partial_{xx} u|^p u dx + \int |\partial_{xx} u|^{p-2} \partial_{xx} u u (e^{-|x|}* \partial_{xx} u) dx \\
{\lesssim}& \| \partial_{xx} u \|_p^p \cdot ( \| \partial_x u\|_\infty + \| u \|_\infty).\end{aligned}$$ Collecting all the estimates and we have $$\begin{aligned}
\frac 1 p \frac d {dt} \| \partial_{xx} u(t) \|_p^p & {\lesssim}\| \partial_{xx} u\|_{p}^{p} \| \partial_x u\|_\infty^2+ \| \partial_{xx} u \|_p^p ( \|\partial_x u\|_\infty^2
+\| \partial_x u\|_\infty+ \| u\|_\infty) \\
& {\lesssim}\| \partial_{xx} u\|_p^p ( \|\partial_x u\|_\infty^2
+\| \partial_x u\|_\infty+ \| u\|_\infty).\end{aligned}$$ Integrating over $t$ we have $$\begin{aligned}
\label{ruxx}
\sup_{0\le t\le T_0} \|\partial_{xx} u(t) \|_p \le e^{CpT_0} \| \partial_{xx} u_0\|_p.\end{aligned}$$ To estimate $\| u_{xx} \|_\infty$ we need to estimate $\| u_{xxx} \|_2$. We have $$\begin{aligned}
\frac d {dt} \| u_{xxx}(t) \|_2^2 & {\lesssim}- \int u_{xxxx} (u^3)_{xxxx} dx +\epsilon \int u_{xxxxx} u_{xxx} dx
- \int \partial_{xxxx}(\partial_x K*u u) u_{xxx} dx \\
& {\lesssim}-\int u_{xxxx}^2 u^2 dx + \int u_{xxxx} (uu_x u_{xxx} + u_{xx}u_x^2 + u_{xx}^2 u) dx \\
&\quad- \int \partial_{xxxx}(\partial_x K*u u) u_{xxx} dx \\
& {\lesssim}\int (u_{xxx}^2 u_x^2 + u_{xx}^4 + u_{xx}^2 u_x^2) dx- \int \partial_{xxxx}(\partial_x K*u u) u_{xxx} dx.\end{aligned}$$ Since $$\begin{aligned}
& \int \partial_{xxxx}( (\partial_x K*u) u) u_{xxx} dx \\
{\lesssim}& \int (\partial_x K*\partial_x u) u_{xxx}^2 dx + \int (\partial_x K*\partial_{xx} u) \partial_{xx} u \partial_{xxx} u dx \\
&\quad + \int (\partial_x K*\partial_{xxx}u) uu_{xxx} dx + \int u (\partial_{xx}K*\partial_{xxx}u) u_{xxx} dx \\
{\lesssim}& \| u_{xxx} \|_2^2 \|\partial_x u\|_\infty+ \| u_{xxx}\|_2 \| \partial_{xx} u\|_2^2 + \| \partial_{xxx} u\|_2^2 \| u\|_\infty.\end{aligned}$$ Finally we get $$\begin{aligned}
\frac d {dt} \| u_{xxx}(t) \|_2^2 &{\lesssim}\| u_{xxx}\|_2^2 ( \|u_x\|_2^2 + \| u_x\|_\infty +\|u\|_\infty) \\
&\quad + \|u_{xxx}\|_2 \| \partial_{xx} u\|_2^2 + \| u_{xx} \|_4^4 + \| u_{xx} \|_2^2 \| u_x\|_\infty^2.\end{aligned}$$ Gronwall then implies that $$\begin{aligned}
\sup_{0\le t\le T} \|\partial_{xxx} u(t) \|_2 \le C \| \partial_{xxx}
u_0\|_2.\label{three grad}\end{aligned}$$ Since $\|u_{xx}\|_\infty {\lesssim}\|u\|_2^{\frac 16} \|\partial_{xxx} u\|_2^{\frac 56}$, we conclude that $$\begin{aligned}
\sup_{0\le t \le T_0} \| u_{xx}(t)\|_{\infty} \le C.\end{aligned}$$
`Step 4`: Control of $\|\partial_x^{2k+1} u\|_2$. We compute $$\begin{aligned}
\frac d {dt} \| \partial_x^{2k+1} u\|_2^2 & = \frac 13 \int \partial_x^{2k+3} (u^3) \partial_x^{2k+1} udx
+ \epsilon \int \partial_x^{2k+3} u \partial_x^{2k+1} udx \\
&\quad -\int \partial_x^{2k+2} (K*\partial_x u u) \partial_x^{2k+1} u dx \\
&=: I+II+III\end{aligned}$$ For $I$, we integrate by parts and obtain $$\begin{aligned}
I&= -\frac 13 \int \partial_x^{2k+2}(u^3) \partial_x^{2k+2} u dx \notag\\
& =-\int u^2(\partial_x^{2k+2}u)^2 dx-C\int uu_x\partial_x^{2k+1}u
\partial_x^{2k+2}u dx\notag\\
&\qquad +\sum_{\substack{l+m+n= 2k+2\\l\le m\le n\le
2k}}C_{l,m,n}\int\partial_x^l u\partial_x^m u\partial_x^n
u\partial_x^{2k+2} udx\notag\\
&=-\int u^2(\partial_x^{2k+2}u)^2 dx+\frac C
2\int(u_x^2+uu_{xx})(\partial_x^{2k+1}u)^2 dx\label{049}\\
&\qquad+\sum_{\substack{l+m+n=2k+3\\l\le m\le n\le 2k+1}}C_{l,m,n}
\int\partial_x ^l u\partial_x^m u\partial x^n u\partial_x^{2k+1}u
dx\label{050}\end{aligned}$$ Discarding the negative term and using Hölder’s inequality, we bound as $$\eqref{049}{\lesssim}(\|u_x\|_{\infty}^2+\|u\|_{\infty}\|u_{xx}\|_{\infty})\|\partial_x^{2k+1}
u\|_2^2.$$ To estimate the summand in , we discuss several cases.
Case 1. $n=2k+1$. In view of the constraint, we have $(l,m)=(0,2)$ or $ (1,1)$. Hölder’s inequality then gives that $$\begin{aligned}
|\int \partial_x^l u\partial_x^m u\partial_x^n u\partial_x^{2k+1} u
dx|{\lesssim}(\|u_x\|_{\infty}^2+\|u\|_{\infty}\|u_{xx}\|_{\infty})\|\partial_x^{2k+1}
u\|_2^2.\end{aligned}$$
Case 2. $n\le 2k$, $l=0$. Since $l+m+n=2k+3$ we must have $m\ge 3$. We choose $2<p,q<\infty$ such that $\frac 1p+\frac 1q=\frac 12$ and use Hölder to get $$\begin{aligned}
|\int \partial_x^l u\partial_x^m u\partial_x^n u\partial_x^{2k+1} u
dx|{\lesssim}\|u\|_{\infty}\|\partial_x^m u\|_p\|\partial_x^n
u\|_q\|\partial_x^{2k+1} u\|_2.\end{aligned}$$ Using the interpolation inequality $$\|\partial_x^m u\|_p\le \|\partial_x^{2k+1}
u\|_2^{\theta}\|u\|_{\dot H^{\frac 52}}^{1-\theta},\
\theta=\frac{m-2-\frac 1p}{2k-\frac 32},$$ we have $$\begin{aligned}
|\int \partial_x^l u\partial_x^m u\partial_x^n u\partial_x^{2k+1} u
dx|&\le \|u\|_{\infty}\|\partial_x^{2k+1} u\|_2^2\|u\|_{\dot
H^{\frac
52}}\\
&\le \|u\|_{\infty}\|u\|_{H^3}\|\partial_x^{2k+1} u\|_2^2.\end{aligned}$$
Case 3. $n\le 2k$, $l=1$. In this case $m,n\ge 2$, we can choose $2<p,q<\infty $ such that $\frac 1p+\frac 1q=\frac 12$ and use the interpolation inequality $$\label{inter}
\|\partial_x^m u\|_p\le \|\partial_x^{2k+1} u\|_2^{\frac{m-1-\frac
1p}{2k-\frac 12}}\|u\|_{\dot H^{\frac 32}}^{1-\frac{m-1-\frac
1p}{2k-\frac 12}}$$ to get $$\begin{aligned}
|\int \partial_x^l u\partial_x^m u\partial_x^n u\partial_x^{2k+1} u
dx|&\le \|u_x\|_{\infty}\|\partial_x^m u\|_p\|\partial_x ^n
u\|_q\|\partial_x^{2k+1} u\|_2\\
&\le \|u_x\|_{\infty}\|\partial_x^{2k+1} u\|_2^2\|u\|_{\dot
H^{\frac 32}}\\
&\le \|u_x\|_{\infty}\|u\|_{H^2}\|\partial_x^{2k+1} u\|_2^2.\end{aligned}$$ Case 4. $n\le 2k$, $l,m,n\ge 2$. Choosing $2<p,q,r<\infty$ such that $\frac 1p+\frac 1q+\frac 1r=\frac 12$ and using the interpolation inequality , we get $$\begin{aligned}
|\int \partial_x^l u\partial_x^m u\partial_x^n u\partial_x^{2k+1} u
dx|&\le \|\partial_x^l u\|_p\|\partial_x^m u\|_q\|\partial_x^n
u\|_r\|\partial_x^{2k+1} u\|_2\\
&\le \|u\|_{\dot H^{\frac 32}}^2 \|\partial_x^{2k+1} u\|_2^2\\
&\le \|u\|_{H^2}^2\|\partial_x^{2k+1} u\|_2^2.\end{aligned}$$ Collecting all the estimates together, we conclude $$\begin{aligned}
I\le &(\|\partial_x u\|_{\infty}^2
+\|u\|_{\infty}\|u_{xx}\|_{\infty}
+\|u\|_{\infty}\|u\|_{H^3}\\
&+\|u_x\|_{\infty}\|u\|_{H^2}+\|u\|_{H^2}^2)\|\partial_x^{2k+1}
u\|_2^2.\end{aligned}$$
For term $II$, we simply have $$\begin{aligned}
II = -\epsilon \int (\partial_x^{2k+2} u)^2 dx \le 0.\end{aligned}$$ For III, we compute: $$\begin{aligned}
III & = - \int \partial_x^{2k+2} (K*\partial_x u) u \partial_x^{2k+1} u dx - \int K*\partial_x u \partial_x^{2k+2} u \partial_x^{2k+1} udx \\
&\quad + \sum_{\substack{l+m= 2k+2 \\l,m\le 2k+1}} C_{l,m}\int \partial_x^l(K*\partial_x u) \partial_x^m u \partial_x^{2k+1} u dx \\
& = - \int \partial_{xx} K* \partial_x^{2k+1} u u \partial_x^{2k+1} u dx + \frac 12 \int \partial_x K*\partial_x u (\partial_x^{2k+1} u)^2 dx \\
&\quad + \sum_{\substack{l+m= 2k+2 \\l,m\le 2k+1}}C_{l,m}
\int \partial_x K* \partial_x^l u \partial_x^m u \partial_x^{2k+1} u dx \\
& =2\int u(\partial_x^{2k+1} u)^2 dx -\int (e^{-|x|}*
\partial_x^{2k+1} u)
\partial_x^{2k+1} u u dx +\frac 12\int(\partial_x K*\partial_x u)(\partial_x^{2k+1} u)^2 dx\\
&\quad + \sum_{\substack{l+m\le 2k+2 \\l,m\le 2k+1}}
C_{l,m}\int \partial_x K* \partial_x^l u \partial_x^m u \partial_x^{2k+1} u dx \\
& {\lesssim}(\| u\|_\infty +\|\partial_x u\|_{2})\|
\partial_x^{2k+1} u\|_2^2 \\
&\qquad+ \sum_{\substack{l+m=2k+2 \\l,m\le 2k+1}}C_{l,m} \|
\partial_x^{2k+1} u\|_2 \| \partial_x^l u\|_2
\| \partial_x^m u\|_2 .\\
$$ Using the interpolation inequality $$\|\partial_x^l u\|_2\le \|\partial_x^{2k+1}
u\|_2^{\frac{l-1}{2k}}\|u_x\|_2^{1-\frac{l-1}{2k}},$$ we finally get $$III{\lesssim}(\|u\|_{\infty}+\|u_x\|_{2})\|\partial_x^{2k+1} u\|_2^2.$$ Summarizing the estimates of $I,II,III$ we obtain $$\begin{aligned}
\frac d{dt}\|\partial_x^{2k+1} u(t)\|_2^2& {\lesssim}\|\partial_x^{2k+1}u\|_2^2(\|u_x\|_{\infty}^2+\|u\|_{\infty}\|u_{xx}\|_{\infty}
\\
&\qquad+\|u\|_{\infty}\|u\|_{H^3}+\|u_x\|_{\infty}\|u\|_{H^2}+\|u\|_{\infty}+\|u_x\|_{2})\\
&{\lesssim}\|\partial_x^{2k+1} u\|_2^2(\|u\|_{H^3}+\|u\|_{H^3}^2).\end{aligned}$$ Using Gronwall and , , we get $$\begin{aligned}
\sup_{0\le t\le T_0} \| \partial^{2k+1}_x u(t) \|_2 {\lesssim}e^{CT_0} \| \partial^{2k+1}_x
u_0\|_2,\end{aligned}$$ for sufficiently small $T_0=T_0(\|u_0\|_2,\|\partial_x u_0\|_2)$.
This concludes the estimate of the $H_x^{2k+1}$-norm of $u^\epsilon$. Finally if $u_0\ge 0$, then by the weak maximum principle we have $u^\epsilon(t)
\ge 0$ for any $0\le t\le T_0$. The lemma is proved.
We are now ready to complete the
This follows directly from Lemma \[lem\_local\_exist\] and \[lem301\]. In particular note that in Lemma \[lem301\] the time of existence of the local solution does not depend on $\epsilon$.
Assume $u_0 \in \bigcap_{m=0}^\infty H_x^m({{\mathbb{R}}}) \bigcap L_x^p({{\mathbb{R}}})$ for some $1\le p<2$. Then using and Duhamel’s formula, we can write $u^\epsilon(t)$ as $$\begin{aligned}
u^\epsilon(t) &= e^{\epsilon t\partial_{xx}} u_0 + \int_0^t e^{\epsilon (t-s) \partial_{xx}}
\left( \frac 13 \partial_{xx} ((u^\epsilon)^3) \right)(s) \\
&\qquad - \int_0^t e^{\epsilon(t-s) \partial_{xx} } \partial_x ( u^\epsilon \,\partial_x K*u^\epsilon )(s) ds.\end{aligned}$$ Since $\|e^{\epsilon t\partial_{xx}} f\|_p {\lesssim}\|f \|_p$ for any $\epsilon>0$, we can estimate the $L^p_x$ norm of $u^\epsilon$ as $$\begin{aligned}
\| u^\epsilon(t) \|_p &{\lesssim}\| u_0\|_p +
\int_0^t \left \| \frac 13 \partial_{xx}((u^\epsilon)^3) - \partial_x( (\partial_x K*u^\epsilon) u^\epsilon) \right\|_p ds \\
& {\lesssim}\| u_0\|_p + \int_0^t \| u^\epsilon(s) \|_{W^{2,3p}}^3+ \| u^\epsilon(s) \|_{W^{1,2p}}^2 ds \\
& {\lesssim}\| u_0\|_p + \int_0^t \| u^\epsilon(s)\|_{H^3}^3 + \|u^\epsilon(s) \|_{H^2}^2 ds,\end{aligned}$$ where the last inequality follows from the Sobolev embedding. This estimate shows that $u^\epsilon(t) \in L_x^p$ for any $t$. The continuity (including right continuity at $t=0$) follows from similar estimates. We omit the details. Finally since for $1\le p<2$, $\|u_0\|_2 + \|\partial_x u_0\|_\infty {\lesssim}\|u_0\|_p + \| \partial_x u_0\|_\infty$, one can choose the time interval sufficiently small depending only on $\|u_0\|_p + \| \partial_x u_0\|_\infty$.
Proof of Proposition \[prop2\]
------------------------------
In the case $u_0 \ge 0$, we shall show the regularized equation has a global solution. The key point here is that by using the positivity, we can obtain the apriori boundedness of the $L_x^p$-norm ($2\le p\le \infty$) of the solution on any finite time interval. To control the $L_x^{\infty}$-norm of the gradient, we need the following lemma from [@Lie96].
\[lemM\] Let $\Omega = (0,T) \times {{\mathbb{R}}}$ and $u \in C^{2,1}(\Omega) \cap C( \bar \Omega) $ be a solution to the parabolic equation: $$\begin{aligned}
\begin{cases}
\partial_t u = b(t,x,u,u_x) u_{xx} +a(t,x,u,u_x), \quad (t,x) \in (0,T) \times {{\mathbb{R}}}\\
u|_{t=0} = u_0
\end{cases}\end{aligned}$$ with the following properties:
1. Uniform parabolicity: there exists constants $\beta_0$, $\beta_1>0$ such that $$\begin{aligned}
\beta_0 b(t,x,z,p) p^2 \ge |a(t,x,z,p)|,\quad\forall\, |p|\ge \beta_1.\end{aligned}$$
2. There is a constant $M>0$ such that $$\begin{aligned}
|u(t,x) - u(t,y) | \le M,\quad \forall\, (t,x), (t,y) \in \Omega.\end{aligned}$$
3. There exists $L_0>0$ such that $$\begin{aligned}
|u_0(x) -u_0(y) | \le L_0 |x-y|, \quad \forall\, x,y\in {{\mathbb{R}}}.\end{aligned}$$
Under all the above assumptions, we have the following bound: $$\begin{aligned}
\sup_{0\le t\le T} \| u_x(t) \|_\infty \le 2 (L_0+\beta_1) e^{\beta_0 M}.\end{aligned}$$
See Lemma 11.16 of [@Lie96].
Incorporating Lemma \[lemM\] with the idea of the proof of Lemma \[lem301\], we get the following global analogue of Lemma \[lem301\].
\[lem301B\] Let $u_0 \in H^{2k+1}_x({{\mathbb{R}}})$, $k\ge 2$ and $u_0 \ge 0$. Then for any $T>0$ and for any $\epsilon>0$, has a unique solution $u^\epsilon \in B^k_{T}$. Also $u^\epsilon(t)\ge 0$ for any $t\ge 0$.
The positivity of $u^\epsilon$ follows easily from the weak maximum principle. By Lemma \[lem\_local\_exist\] we only need to get an apriori control of the $H^{2k+1}_x$-norm of $u$ on an arbitrary time interval $[0,T]$.
`Step 1`: $L_x^p$-norm control for $2\le p\le +\infty$. By the same estimates as in Step 1 of the proof of Lemma \[lem301\] we have the following a priori $L_x^2$-norm estimate: $$\begin{aligned}
\|u(t) \|_2 \le \| u_0\|_2 \exp\left(\frac 14 \| \partial_x K\|_1^2 t \right).\end{aligned}$$ To obtain the $L_x^\infty$-norm estimate, we take $2<p<+\infty$ and compute $$\begin{aligned}
\frac 1 p \frac d {dt} \| u(t) \|_p^p &\le - \frac {2(p-1)} p \int u |u|^p dx + \frac {p-1} p \int (e^{-|x|}*u) |u|^p dx\end{aligned}$$ Since $u\ge 0$ we can drop the first term and continue to estimate $$\begin{aligned}
\frac 1 p \frac d {dt} \| u(t) \|_p^p \le C \|u_0\|_2 e^{Ct} \| u(t) \|_p^p,\end{aligned}$$ where $C$ is a generic constant independent of $p$. Gronwall then implies that $$\begin{aligned}
\label{eq1220A}
\| u(t) \|_p \le \|u_0\|_p \exp \left( Ct \|u_0\|_2 e^{Ct} \right).\end{aligned}$$ Letting $p\to \infty$, we obtain $$\begin{aligned}
\label{eq1220B}
\| u(t) \|_\infty \le \|u_0\|_\infty \exp \left( Ct \|u_0\|_2 e^{Ct} \right).\end{aligned}$$ This is the $L_x^\infty$-norm estimate we needed.
`Step 2`: Control of $\| \partial_x u(t) \|_\infty$. We shall apply Lemma \[lemM\] with $$\begin{aligned}
& b(t,x,z,p) =z^2 +\epsilon, \\
& a(t,x,z,p) = 2zp^2 + 2 z^2 -c(t,x) z -d(t,x)p,\end{aligned}$$ where $$\begin{aligned}
c(t,x) = (\partial_x K*u_x)(t,x),\\
d(t,x) = (\partial_x K*u)(t,x).\end{aligned}$$ By step 1 we have $$\begin{aligned}
|c(t,x)|+ |d(t,x)| &\le C \|u(t)\|_{\infty}\\
& \le C \|u_0\|_{\infty} exp(CT\|u_0\|_2 e^{CT}).\end{aligned}$$ Also $$\begin{aligned}
|u(t,x) -u(t,y)
| &\le 2\|u(t)\|_\infty \le 2 \|u_0\|_\infty \exp\left( CT \|u_0\|_2 e^{CT} \right). \\
& =: M\end{aligned}$$ and $$\begin{aligned}
|u_0(x) -u_0(y) | &\le \| \partial_x u_0\|_\infty |x-y| \\
&=: L_0|x-y|.\end{aligned}$$ Collecting all these estimates, we see that $$\begin{aligned}
\sup_{0\le t\le T} \| \partial_x u(t) \|_\infty \le C(\|u_0\|_2, \|\partial_x u_0\|_\infty, \epsilon, T)\end{aligned}$$ This concludes the gradient estimate.
`Step 3`: Control of the higher derivatives. This part of the estimates is exactly the same as the corresponding estimates in the proof of Lemma \[lem301\]. Note in particular that we do not need $T$ to be small once we obtain the a priori control of $\|u\|_p$ and $\|\partial_x u\|_\infty$ norms.
Now we are ready to complete
This follows directly from the local existence Lemma \[lem\_local\_exist\] and the a priori estimate Lemma \[lem301B\]. If $u_0 \in L_x^p({{\mathbb{R}}})$ for some $1\le p<2$, then by repeating the proof of Corollary \[cor1\] we obtain $u\in C([0,T], L_x^p)$ for any $T>0$. The proof is finished.
Proof of Theorem \[thm1\], \[thm2\] and \[thm3\]
================================================
In this section we prove our main theorems. Our solution is going to be the limit of the sequence of regularized solutions $u^{{\varepsilon}}$ which we constructed in the previous section. To obtain uniform control of the Sobolev norms of the regularized solutions, we have the following
\[lem\_apriori\] Assume $u_0 \in H_x^{2k+1}({{\mathbb{R}}})$, $k\ge 2$. Let $T_0=T_0(\| u_0\|_2+
\| \partial_x u_0 \|_\infty)$ and $u^{\epsilon} \in B^k_{T_0}$ be the corresponding unique solution to (see Lemma \[lem301\]). Then the set of functions $(u^{\epsilon})_{0<\epsilon<1}$ satisfies the following uniform estimates: $$\begin{aligned}
& \sup_{0\le t\le T_0}\| u^\epsilon \|_{H^{2k+1}}
+\sqrt{\epsilon} \| u^{\epsilon} \|_{L_t^2 H^{2k+2}([0,T_0]\times {{\mathbb{R}}})}
\le C( \| u_0\|_{H^{2k+1}}, k), \\
& \| u^\epsilon \partial_x^{2k+2} u^\epsilon \|_{L_t^2 L_x^2([0,T_0]\times {{\mathbb{R}}})} + \| \partial_t u^\epsilon\|_{L_t^2 H^{2k}([0,T_0]\times {{\mathbb{R}}})}
\le C( \| u_0\|_{H^{2k+1}}, k),\\
& \| \partial_x (\partial_x K*u^{\epsilon} u^\epsilon)
\|_{L_t^2 H^{2k}([0,T_0]\times {{\mathbb{R}}})} \le C( \| u_0\|_{H^{2k+1}}, k),\end{aligned}$$ where $C=C(\| u_0\|_{H^{2k+1}}, k)$ is a positive constant.
The estimates of $\|u^{\epsilon} \|_{L_t^\infty H^{2k+1}}$, $\sqrt{\epsilon} \| u^{\epsilon} \|_{L_t^2H^{2k+2}}$ and $ \|
u^\epsilon \partial_x^{2k+2} u^\epsilon \|_{L_t^2 L_x^2}$ can be recovered from step 4 of the proof of Lemma \[lem301\]. To bound $\| \partial_t u^\epsilon \|_{L_t^2 H^{2k}}$, we use to obtain (here we again drop the superscript ${\varepsilon}$ for simplicity): $$\begin{aligned}
\| \partial_t u \|_{H^{2k}} & \le \| \partial_t u
\|_2 + \| \partial_x^{2k} \partial_t u \|_2 \\
& \le \| (u^2+\epsilon) u_{xx} + 2 u u_x^2 - \partial_x ( K*\partial_x u u) \|_2 + \| \partial_x^{2k} (u^2 u_{xx}) \|_2 \\
&\quad + \epsilon \| \partial_x^{2k+2} u \|_2 + 2 \| \partial_x^{2k} (u u_x^2) \|_2 + \| \partial_x^{2k+1}( K*\partial_x u u) \|_2 \\
& \le \| u\|_{H^2}^3+ {\| u\|_{H^2}^2} + \| u^2 \partial_x^{2k+2} u\|_2 + \| u\|_{H^{2k+1}}^3 \\
& \quad + \epsilon \| \partial_x^{2k+2} u\|_2 + \| u\|_{H^{2k+1}}^3 + \| u\|_{H^{2k+1}}^2.\end{aligned}$$ Integrating over $[0,T_0]$, we have $$\begin{aligned}
\int_0^{T_0} \| \partial_t u(t)\|_{H^{2k}}^2 dt & {\lesssim}T_0 \sup_{0\le t\le T_0} \| u\|_{H^{2k+1}}^2 (1+ \| u\|_{H^{2k+1}} ) \\
& \quad + \int_0^{T_0} \| u^2 \partial_x^{2k+2} u(t) \|_2^2 dt + \epsilon \int_0^T \| \partial_x^{2k+2} u \|_2^2 dt \\
& \le C(T_0,k, \|u_0\|_{H^{2k+1}}).\end{aligned}$$ Since $T_0=T_0(\|u_0\|_2+\|\partial_x u_0\|_\infty)$, we obtain the desired bounds. The last estimate is a simple application of Hölder and Young’s inequality using the given estimates.
Now we define the set $$\begin{aligned}
D^k_{T_0} := &\Bigl\{ v \in C([0,T_0]; H_x^{2k}) \cap L^\infty(0,T_0; H_x^{2k+1})):\, \Bigr.\\
&\Bigl. \partial_t v \in L^2(0,T_0; H^{2k})),\, v^2\partial_{xx}v
\in L^2(0,T_0; H^{2k}) \Bigr\}.\end{aligned}$$ We shall prove the existence and the uniqueness of the solution of in this set. This is
\[thm\_local\] Assume $u_0 \in H^{2k+1}({{\mathbb{R}}})$ with $k\ge 2$. Then there exists $T_0=T_0(\|u_0\|_2 + \|\partial_x u_0 \|_\infty)$ such that has a unique solution $u \in D^k_{T_0}$.
`Step 1`. Contraction in $C([0,T],L_x^2)$. For any two $\epsilon_1, \epsilon_2>0$, let $u=u^{\epsilon_1}$, $v=u^{\epsilon_2}$ solve with the same initial data $u_0$. Denote $w=u-v$, then for $w$ we have the equation $$\begin{aligned}
\partial_t w &= \frac 13 (u^3-v^3)_{xx} + \epsilon_1 u_{xx} -\epsilon_2 v_{xx} - \partial_x (K*\partial_x w u) \\
&\qquad - \partial_x (K*\partial_x v w).\end{aligned}$$ Then $$\begin{aligned}
\frac 12 \frac d {dt} \| w(t) \|_2^2 & = \frac 13 \int (u^3-v^3)_{xx} w dx + \int (\epsilon_1 u_{xx} -\epsilon_2 v_{xx}) wdx \\
&\quad -\int \partial_x(K*\partial_x w u) w dx - \int \partial_x (K*\partial_x v w) wdx \\
& =: I +II+III+IV.\end{aligned}$$ Denote $F(u,v)=u^2+v^2+uv$. Clearly we always have $F(u,v)\ge 0$. We compute $$\begin{aligned}
I & = \frac 13 \int \partial_{xx}(w F(u,v)) w dx \\
& = \frac 13 \int \partial_{xx} w w F(u,v) dx \\
& = - \frac 13 \int (\partial_x w)^2 F(u,v) dx + \frac 16 \int w^2 \partial_{xx} F(u,v) dx \\
& \le \frac 16 \int w^2 \partial_{xx} F(u,v) dx \\
& {\lesssim}\| w\|_2^2 ( \| u\|_{H^3}^2 + \|v\|_{H^3}^2).\end{aligned}$$ For II we have the estimate $$\begin{aligned}
II \le (\epsilon_1 \|u\|_{H^2} + \epsilon_2 \|v \|_{H^2}) \| w\|_{2}.\end{aligned}$$ The term III can be bounded as: $$\begin{aligned}
III & = \int K*\partial_x w u w_x dx \\
&= -\int \partial_{xx}K*w uw dx -\int K*\partial_x w u_x w dx \\
& \le (\|u\|_\infty + \|\partial_x u\|_\infty) \| w\|_2^2.\end{aligned}$$ Similarly for IV we get $$\begin{aligned}
IV & {\lesssim}\| K*\partial_{xx} v\|_\infty \| w\|_2^2 \\
& {\lesssim}\|v \|_{H^1} \|w\|_2^2.\end{aligned}$$ Finally we have by Lemma \[lem\_apriori\] $$\begin{aligned}
\frac d {dt} \| w(t)\|_2^2 &{\lesssim}\| w(t) \|_2^2 ( \| u\|_{H^3}^2 + \| v\|_{H^3}^2 + \|u\|_{H^3} + \| v\|_{H^3} ) \\
&\quad + (\epsilon_1 \| u\|_{H^2} + \epsilon_2 \| v\|_{H^2} ) \|w(t)\|_{2} \\
& {\lesssim}C(\|u_0\|_{H^{2k+1}},k) \|w(t)\|_2^2 +
C(\|u_0\|_{H^{2k+1}},k) (\epsilon_1+\epsilon_2) \|w(t)\|_{2}\end{aligned}$$ Gronwall then gives $$\begin{aligned}
\| u^{\epsilon_1}(t)-u^{\epsilon_2}(t) \|_2 &= \|w(t)\|_2 \\
& \le C(\|u_0\|_{H^{2k+1}},k) (\epsilon_1+\epsilon_2) \exp \left( C(\|u_0\|_{H^{2k+1}},k) t \right).\end{aligned}$$ This implies that $(u^\epsilon)_{0<\epsilon<1}$ has a limit as $\epsilon\to 0$ in $C([0,T_0];\, L_x^2)$. Denote this limit as $u
\in C([0,T_0];\, L_x^2)$.
`Step 2`. We show that $u$ is our desired solution in $D^k_{T_0}$. By Lemma \[lem\_apriori\] and the interpolation inequality $$\begin{aligned}
\| f\|_{H^{2k}} {\lesssim}\|f\|_2^{\frac 1 {2k+1}} \| f\|_{H^{2k+1}}^{\frac {2k}{2k+1}},\end{aligned}$$ we obtain $$\begin{aligned}
u^{\epsilon} \to u \quad \text{in}\; C([0,T_0];\, H_x^{2k}), \quad \text{as $\epsilon\to 0$}.\end{aligned}$$ It then follows easily that $$\begin{aligned}
& (u^\epsilon)^2 \partial_{xx} u^{\epsilon} \to u^2 \partial_{xx} u, \quad
\text{in $C(0,T_0; L_x^2)$},\\
& u^\epsilon (\partial_x u^{\epsilon})^2 \to u (\partial_x u)^2, \quad \text{in $C(0,T_0;
L_x^2)$},\\
& \partial_x(K*\partial_x u^{\epsilon} u^\epsilon ) \to \partial_x
(K*\partial_x u u), \quad \text{in $C(0,T_0; L_x^2)$}.\end{aligned}$$ By Lemma \[lem\_apriori\], the functions $(u^\epsilon)^2 \partial_{xx}
u^\epsilon$, $u^\epsilon (\partial_x u^{\epsilon})^2$, $\partial_x
(K*\partial_x u^\epsilon u^\epsilon)$ are uniformly bounded in $L^2([0,T_0], H_x^{2k})$. We conclude that $u^2 \partial_{xx} u$, $u
(\partial_x u)^2$, $\partial_x (K*\partial_x u u) \in
L^2([0,T_0],H_x^{2k})$. Now since $u^\epsilon$ is also uniformly bounded in $L^\infty([0,T_0],H_x^{2k+1})$, we obtain $u \in
L^\infty([0,T_0],H_x^{2k+1})$. Similarly by Lemma \[lem\_apriori\], the set of functions $ (\partial_t u^{\epsilon})_{0<\epsilon<1}$ is uniformly bounded in $L^2([0,T_0],
H_x^{2k})$. By extracting a subsequence if necessary, we obtain for some $\epsilon_n \to 0$, $$\begin{aligned}
\partial_t u^{\epsilon_n} \rightharpoonup \partial_tu, \quad \text{in $L^2([0,T_0],H_x^{2k})$}.\end{aligned}$$ It follows that $\partial_t u \in L^2([0,T_0],H_x^{2k}) $. Finally by using Lemma \[lem\_apriori\] again we have $$\begin{aligned}
\epsilon_n \partial_{xx} u^{\epsilon_n} \to 0, \quad \text{in $L^2(0,T_0;H_x^{2k})$}.\end{aligned}$$ This shows that $u$ is our desired solution in $D^k_{T_0}$. The theorem is proved.
\[cor\_cont\] Let $u_0 \in H_x^{2k+1}({{\mathbb{R}}})$ with $k\ge 2$. Then there exists a unique solution $u$ of with maximal-lifespan $T^*$ such that only one of the following possibilities occur:
1. $T^*=\infty$ and $u \in D^k_T$ for any $T>0$.
2. $T^*<\infty$, $u\in D^k_T$ for any $T<T^*$, and $$\begin{aligned}
\lim_{t\to T^*} \| \partial_x u(t) \|_\infty =\infty.\end{aligned}$$
For any $2\le p<\infty$, there exists a generic constant $C$ such that $$\begin{aligned}
\| u(t) \|_p \le \|u_0 \|_p e^{Cpt}, \quad \forall\, t\ge 0.\end{aligned}$$ If in addition $u_0\ge 0$, then $u(t)\ge 0$ and we also have the $p$-independent estimate for all $2\le p\le +\infty$: $$\begin{aligned}
\label{eq1219}
\| u(t) \|_p \le \|u_0\|_p \exp \left ( Ct \|u_0\|_2 e^{Ct} \right), \quad\forall\, t\ge 0.\end{aligned}$$ In particular if $p=+\infty$, then $$\begin{aligned}
\| u(t) \|_{\infty} \le \|u_0\|_\infty \exp\left( Ct \|u_0\|_2 e^{Ct} \right), \quad\forall\, t\ge 0.\end{aligned}$$
This follows directly from Theorem \[thm\_local\] and Lemma \[lem301\] (see for the growth estimate of $L^p_x$-norm for any $T>0$). Since $u^\epsilon \to u$ uniformly in $C([0,T];
H_x^{2k})$, $u(t)$ satisfies the same estimate . Similarly if $u_0\ge 0$, then the estimate follows from the corresponding estimate for $u^\epsilon$ (see and in the proof of Lemma \[lem301B\]). Now that the $L^2$ norm of the constructed local solution remains finite for any $T>0$, it can be continued as long as $\| \partial_x u(t) \|_\infty$ does not blow up.
We are now ready to complete the
By Theorem \[thm\_local\], $u_0 \in \bigcap_{m=0}^\infty H_x^m ({{\mathbb{R}}})$ implies that $u \in \bigcap_{k=2}^\infty D^k_{T_0}$. Therefore $u\in C([0,T_0], H_x^m)$ for any $m\ge 0$. It also follows that $\partial_t u \in \bigcap_{m=1}^\infty
L^2(0,T_0;H^m_x)$. By using , a standard bootstrap argument implies that $u\in C^\infty([0,T_0]\times {{\mathbb{R}}})$. If $u_0\ge
0$, then the positivity of $u$ follows from Lemma \[lem301B\] and the uniform convergence of $u^\epsilon$ to $u$. The theorem is proved.
This follows immediately from Theorem \[thm1\] and Corollary \[cor\_cont\].
This is only slightly different from the proof of Corollary \[cor1\]. Assume $u_0 \in \bigcap_{m=0}^\infty H_x^m({{\mathbb{R}}}) \bigcap L_x^p({{\mathbb{R}}})$ for some $1\le p<2$. Then using , we can bound $$\begin{aligned}
\| u(t) \|_p &{\lesssim}\| u_0\|_p + \int_0^t \left \| \frac 13 \partial_{xx}(u^3) - \partial_x( (\partial_x K*u) u) \right\|_p ds \\
& {\lesssim}\| u_0\|_p + \int_0^t \| u(s) \|_{W^{2,3p}}^3+ \| u(s) \|_{W^{1,2p}}^2 ds \\
& {\lesssim}\| u_0\|_p + \int_0^t \| u(s)\|_{H^3}^3 + \|u(s) \|_{H^2}^2 ds.
\end{aligned}$$ This shows that $u(t) \in L_x^p$ for any $t$. The continuity (including right continuity at $t=0$) follows from similar estimates. We omit the standard details. Finally in the case $u_0\ge 0$, $u_0 \in L_x^1({{\mathbb{R}}})$, the $L^1_x$ preservation follows from direct integration.
Proof of Theorem \[thm4\]
=========================
We argue by contradiction. Let $u_0 \in
C_c^\infty({{\mathbb{R}}})$, $u_0\ge 0$ and assume that the corresponding solution $u$ is global. By Theorem \[thm1\], \[thm2\], \[thm3\], $u \in C([0,T), H_x^m)$ for any $m\ge 0$ and the $L_x^1$-norm of $u$ is preserved. Our intuition of proving the blowup is based on the observation that as time goes on, the boundary of the solution(which in $1$D case consists of two points) will move face to face at a certain speed which has a lower bound independent of time. This clearly will lead to the collapse of the solution. To realize this intuition, we will carry out a detailed analysis on the characterstic lines of the solution which satisfy $$\begin{aligned}
\label{eqXta}
\begin{cases}
\dfrac d {dt} X(t, \alpha) = (K*\partial_x u) (X(t,\alpha),t), \\
X(0,\alpha) = \alpha \in {{\mathbb{R}}}.
\end{cases}\end{aligned}$$ By standard ODE theory and our assumption that $u$ is a smooth global solution, $X(t)$ is well defined and smooth for all time. Moreover, we have the following lemma.
\[lem\_Xta\] The characterstic lines $X(t,\alpha)$ and the solution $u(t,x)$ satisfy the following properties:
1. For any $t\ge 0$, $X(t,\cdot):{{\mathbb{R}}}\to{{\mathbb{R}}}$ is a $C^\infty$ diffeomorphism.
2. $X(t,\cdot)$ is an order-preserving map in the sense that if $\alpha_1<\alpha_2$, then $X(t,\alpha_1)<X(t,\alpha_2)$ for any $t\ge 0$.
3. $X(t,\cdot)$ maps intervals to intervals. More precisely, for any $\alpha_1<\alpha_2$, denote $$X(t,\cdot)([\alpha_1,\alpha_2])=\left\{ X(t,\alpha): \alpha_1 \le \alpha \le \alpha_2 \right\}.$$ Then $X(t,\cdot)([\alpha_1,\alpha_2]) = [X(t,\alpha_1], X(t,\alpha_2)]$.
4. If $u_0(\alpha)=0$, then $u(X(t,\alpha),t)=0$ for any $t\ge 0$.
5. If $supp(u_0(\cdot)) \subset [-L,L]$ for some $L>0$, then $$\begin{aligned}
supp(u(t,\cdot)) \subset [X(t,-L),X(t,L)] \subset [-L,L]\end{aligned}$$ for all $t\ge 0$.
Property (1) is rather standard. Note in particular that $$\begin{aligned}
\frac {\partial X(t,\alpha)} {\partial \alpha} = \exp \left( \int_0^t (\partial_x K*\partial_x u)(X(s,\alpha),s) ds \right)\end{aligned}$$ is smooth and globally invertible due to the fact that $u$ is smooth. Property (2) is also trivial. Assume not true, then by the intermediate value theorem, for some $\tau >0$ we must have $X(\tau,\alpha_1)-X(\tau,\alpha_2)=0$ which is a contradiction to property (1). Property (3) follows immediately from property (1) and (2). For property (4), one simply uses and . Writing $u(t,\alpha) = u(X(t,\alpha),t)$, we have $$\begin{aligned}
\begin{cases}
\frac d {dt }u(t,\alpha) = \left(-\partial_x K*\partial_x u + u \partial_{xx} u+ 2 (\partial_x u)^2 \right) u(t,\alpha), \\
u(0,\alpha) = u_0(\alpha).
\end{cases}\end{aligned}$$ On the RHS of the above equation, one can regard the coefficient of $u(t,\alpha)$ as given functions of $(t,\alpha)$. By direct integration, property (4) now follows immediately. For property (5) we take $y<x(t,-L)$, then by property (1) and (2) we have $y=X(t,\alpha)$ for some $\alpha<-L$. By property (4) since $u_0(\alpha)=0$, we must have $u(t,y)=u(t,X(t,\alpha))=0$. Similarly once can show that if $y>X(t,L)$ then $u(t,y)=0$. This shows that $supp(u(t,\cdot)) \subset [X(t,-L),X(t,L)]$. Now using this fact and , we compute $$\begin{aligned}
\label{eq4459}
\frac d {dt} X(t,-L) & = (\partial_x K*u)(t, X(t,-L)) \notag\\
& = - \int_{{{\mathbb{R}}}} sign(X(t,-L)-y) e^{-|X(t,-L)-y|} u(t,y) dy \notag \\
& = - \int_{y\ge X(t,-L)} sign(X(t,-L)-y) e^{-|X(t,-L)-y|} u(t,y) dy \notag\\
& \ge 0.\end{aligned}$$ By a similar computation one can show that $\frac d {dt} X(t,L) \le 0$. These facts together with property 2 and 3 above easily yield that $[X(t,-L), X(t,L)] \subset [-L,L]$.
Using lemma \[lem\_Xta\], we are able to finish the
We argue by contradiction. Let $u_0 \in C_c^\infty({{\mathbb{R}}})$, $u_0 \ge 0$ with $supp(u_0) \subset [-L,L]$ for some $L>0$. Assume that the corresponding solution $u$ is global. Define the characteristic $X(t,\alpha)$ according to . By Lemma \[lem\_Xta\] we have $$\begin{aligned}
\label{eq5503}
supp(u(t,\cdot)) \subset [X(t,-L),X(t,L)] \subset [-L,L],\quad \forall\, t\ge 0.\end{aligned}$$ By this fact, we then compute using , $$\begin{aligned}
\frac d {dt} X(t,-L) &= - \int_{y\ge X(t,-L)} sign(X(t,-L)-y) e^{-|X(t,-L)-y|} u(t,y) dy \\
& = -\int_{ X(t,-L) \le y \le X(t,L)} sign(X(t,-L)-y) e^{-|X(t,-L)-y|} u(t,y) dy\\
& = \int_{ X(t,-L) \le y \le X(t,L)} e^{-|X(t,-L)-y|} u(t,y) dy\\
& \ge e^{-2L} \|u_0\|_{L^1},\end{aligned}$$ where in the last step we used the fact that the $L^1$ norm of $u$ is preserved. This shows that $X(t,-L)$ grows linearly with time which is contradiction to . Thus we have shown that the solution $u$ with $u_0$ as initial data cannot exist globally. Finally and are easy consequences of Theorem \[thm1\], \[thm2\] and \[thm3\]. The theorem is proved.
Acknowledgements {#acknowledgements .unnumbered}
================
D. Li and X. Zhang are supported under NSF grant DMS-0635607. X. Zhang is also supported under NSF grant No. 10601060, Project 973 and the Knowledge Innovation Program of CAS.
[m]{} Angenent, S.B., Aronson D.G., Betelu, S.I., Lowengrub, J.S., Focusing of an elongated hole in porous medium flow. Phys. D 151(2-4):228–252(2001).
Aronson D.G. and Caffarelli L.A., Optimal regularity for one-dimensional porous medium flow, Rev. Mat. Iberoamericana, 2(1986), no. 4, 357–366.
Aronson D.G. and Graveleau J.A., Self-similar solution to the focusing problem for the porous medium equation. European J. Appl. Math. 4(1):65–81(1993).
Aronson, D.G. Regularity properties of flows through porous media, SIAM J. Appl. Math., 17(1969), 461–467.
Aronson, D.G. The porous medium equation, in “ Nonlinear Diffusion Problems ” (Fasano A. & Primicerio M., eds.), Springer lecture notes in Math., 1224(1986), 1–46.
Aronson D.G. and Vazquez J.L., Eventual $C^\infty$-regularity and concavity for flows in one-dimensional porous media, Arch. Rational Mech. Anal., 99(1987), 329–412.
Bertozzi, A. L. and Brandman J. Finite-time blow-up of $L^\infty$-weak solutions of an aggregation equation. Preprint.
Bénilan, P., Carrillo, J. Wittbold, P., Renormalized entropy solutions of a scalar conservation laws. Ann. Scuola Norm. Sup. Pisa Cl. Sci.(4) 29(2):313–327(2000)
Bénilan, P. Equations d’évolution dan un espace de Banach quelconque et applications. Ph.D. Thesis, Université Orsay, 1972
Benilan P., A strong regularity $L^p$ for solutions of the porous media equation, in " Contributions to Nonlinear Partial Differential Equations“ (Bardos C., Damlamian A., Diaz J.I. and Hernandez J. eds.) Research Notes in Math., 89, Pitman London(1983), 39–58.
Bertozzi, A. L. and Laurent, T. Finite-Time blow up of solutions of an aggregation equation in $\mathbb{R}^n$. Comm in Math. Phys. 274, 717-735 (2007)
Bodnar, M. and Velázquez, J. J. L. Derivation of macroscopic equations for individual cell-based model: a formal approach. Math. Methods Appl. Sci. 28(25), 1757-1779 (2005)
Bodnar, M. and Velázquez, J. J. L. An integrodifferential equation arising as a limit of individual cell-based models. J. Differential Equations. 222 (2) 341-380 (2006)
Burger, M. and Di Francesco M., Large time behaviour of nonlocal aggregation models with nonlinear diffusion. RICAM-Report, Johann Radon Institute for Computational and Applied Mathematics. Austrian Academy of Sciences 15.
Burger, M. Capasso, V. and Morale, D. On an aggregation equation model with long and short range interactions. Nonlinear Anal. Real World Appl. 8 (3), 939-958 (2007).
Brändle C., Vázquez, J.L. Viscosity solutions for quasilinear degenerate parabolic equations of porous medium type. Indiana Univ. Math. J. 54(3):817–860(2005)
Caffarelli, L.A., Cabré, X. Fully nonlinear elliptic equations. Amer. Math. Soc. Colloq. Publ. 43, Amer. Math. Soc., Providence, RI, 1995.
Carrillo, J.A., McCann, R., Viallani, C., Contractions in the $2$-Wasserstein length space and thermalization of granular media. Arch. Rational Mech. Anal. 179(2006), 217–264.
Crandal, M.G., Evans, L.C., Lions, P.L. Some properties of viscosity solutions of Hamilton-Jacobi equations. Trans. Amer. Math. Soc. 282(1984), 487–502.
Caffarelli L.A. and Friedman A., Continuity of the density of a gas flow in a porous medium, Trans. Amer. Math. Soc., 252(1979), 99–113.
Caffarelli, L.A. and Friedman A. Regularity of the free boundary of a gas flow in an $n$-dimensional porous medium. Indiana Univ. Math. J. 29(1980):361–391.
Crandall, M.G. and Liggett, T.M. Generation of semi-groups of nonlinear transformations on general Banach spaces. Amer. J. Math. 93(1971), 265–298.
Caffarelli, L.A., Vázquez, J.L. Viscosity solutions for the porous medium equation. In *Differential equations: La Pietra 1996*(ed. by M. Giaquinta, J. Shatah and S.R.S. Varadhan), Proc. Sympos. Pure Math. 65, Amer. Math. Soc., Providence, RI, 1999, 13–26.
Caffarelli L.A., Vázquez J.L. and Wolanski N.I. Lipschitz continuity of solutions and interfaces of the $N$-dimensional porous medium equation, Indiana Univ. Math. J., 36(1987), no. 2, 373–401.
Dibenedetto E., Regularity results for the porous media equation, Ann. Mat. Pura Appl. (4) 121, 249–262(1979)
Edelstein-Keshet L., Watmough J., and Grünbaum D., Do travelling band solutions describe cohesive swarms? An investigation for migratory locusts, J. Math. Biol., 36, 515–549 (1998)
Flierl G., Grünbaum D., Levin S., and Olson D., From individuals to aggregations: The interplay between behavior and physics, J. Theoret. Biol., 196, 397–45(1999)
Gilding, B.H. Hölder continuity of solutions of parabolic equations, J. London Math. Soc., (2) 13(1976), no. 1, 103–106.
Hosono, Y., Mimura, M. Localized cluster solutions of nonlinear degenerate diffusion equations arising in population dynamics. Siam. J. Math. Anal. 20, 845–869(1989)
Darryl D. Holm and Vakhtang Putkaradze. Aggregation of finite size particles with variable mobility. Phys. Rev. Lett., 95:226106, 2005.
Darryl D. Holm and Vakhtang Putkaradze. Formation of clumps and patches in self-aggregation of finite-size particles. Physica D. 220:183–196, 2006.
Ikeda, T. Stationary solutions of a spatially aggregating population model. Proc. Jpn. Acad. A 60, 46–48(1984)
Ikeda, T. Standing pulse-like solutions of a spacially aggregating population model. Jpn. J. Appl. Math. 2, 111–149(1985)
Ikeda, T., Nagai, T. Stability of localized stationary solutions. Jpn. J. Appl. Math. 4, 73–97(1987)
Kawasaki, K. Diffusion and the formation of spatial distributions. Math. Sci. 16(183), 47–52(1978)
Koch, H., Non-Euclidean singular integrals and the porous medium equation. University of Heidelberg, Habilitation Thesis, 1999.
Kruzhkov, Results on the character of the regularity of solutions of parabolic equations and some of their applications, Math. Notes, 6(1969), 517–523.
Laurent, T. Local and global existence for an aggregation equation. Comm. in PDE 32,1941-1964 (2007)
Lieberman G.L., Second Order Parabolic Differential Equations, World Scientific publishing Co. Pte. Ltd. (1996)
Levine H., Rappel W.J., and Cohen I. Self-organization in systems of self-propelled particles. Phys. Rev. E 63, paper 017101 (2001)
Li D. and Rodrigo J. Finite-time singularities of an aggregation equation in ${{\mathbb{R}}}^n$ with fractional dissipation, to appear in Comm. Math. Phys.
Li D. and Rodrigo J., [Wellposedness and regularity of solutions of an aggregation equation]{}, to appear in Revista Matematica Iberoamericana
Li D. and Rodrigo J., [Refined blowup criteria and nonsymmetric blowup of an aggregation equation]{}. to appear in Adv. Math.
Li D., Rodrigo J. and Zhang X., [Exploding solutions for a nonlocal quadratic evolution problem]{}, to appear in Revista Matematica Iberoamericana.
Li D. and Zhang X., Global wellposedness and blowup of solutions to a nonlocal evolution problem with singular kernels. Submitted.
Mogilner A. and Edelstein-Keshet L., A non-local model for a swarm, J. Math. Biol., 38, 534–570(1999)
Murray J.D., Mathematical Biology I: An Introduction, 3rd ed., Interdiscip. Appl. Math. 17, Springer, New York, 2002.
Muskat M.. The flow of homogeneous fluids through porous media. McGraw-Hill, New York, 1937.
Mimura, M., Yamaguti, M. Pattern formation in interacting and diffusing systems in population biology. Adv. Biophys. 15, 19–65(1982)
Nagai, T., Mimura, M. Asymptotic behavior for a nonlinear degenerate diffusion equation in population dynamics. Siam J. Appl. Math. 43, 449–464(1983)
Okubo A., Diffusion and Ecological Problems, Springer, New York, 1980.
Oleinik, Kalashnikov and Chzou Yui-Lin’, The Cauchy problem and boundary-value problems for equations of unsteady filtration type, Izv. Akad. Nauk SSR Ser. Mat., 22(1958), 667–704.
Otani M. and Sugiyama Y., Gradient estimates of solutions of some non-Newtonian filtration problems, Pitman Research Notes in Mathematics Series, 363(1996), 247–253.
Otani M. and Sugiyama Y., A method of energy estimates in $L^\infty$ and its application to porous medium equations. J. Math. Soc. Japan, 53(4):745–789 (2001).
dal Passo, R., Demotoni, P. Aggregative effects for a reaction-advection equation. J. Math. Biol. 20, 103–112(1984)
Topaz C.M., Bertozzi A.L., and Lewis, M.A., A nonlocal continuum model for biological aggregation. Bull. Math. Bio. 68(7), 1601–1623(2006)
Topaz, C. M. and Bertozzi, A. L. Swarming patterns in a two-dimensional kinematic model for biological groups. SIAM J. Appl Math. 65(1), 152-174 (2004)
Toner J. and Tu Y. Flocks, herds, and schools: A quantitative theory of flocking, Phys. Rev. E 58, 4828–4858(1998)
Vázquez, J.L. The porous medium equation. Mathematical Theory. Oxford Math. Monogr. Oxford University Press, Oxford 2007.
Vicsek T., Czirók A., Farkas I.J., and Helbing D., Application of statistical mechanics to collective motion in biology, Phys. A, 274, 182–189(1999)
Yudovich, V.I. Non-stationary flow of an incompressible liquid. Zh. Vychisl. Mat. Mat. Fiz., 3:1032–1066 (1963).
Zel’dovich, Ya. B. and Raizer, Yu. P. Physics of shock waves and high-temperature hydrodynamic phenomena II, 1966, Academic Press, New York, 1966.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Interest in online rating data has increased in recent years in which ordinal ratings of products or local businesses are provided by users of a website, such as [`Yelp!`]{} or `Amazon`. One source of heterogeneity in ratings is that users apply different standards when supplying their ratings; even if two users benefit from a product the same amount, they may translate their benefit into ratings in different ways. In this article we propose an ordinal data model, which we refer to as a multi-rubric model, which treats the criteria used to convert a latent utility into a rating as user-specific random effects, with the distribution of these random effects being modeled nonparametrically. We demonstrate that this approach is capable of accounting for this type of variability in addition to usual sources of heterogeneity due to item quality, user biases, interactions between items and users, and the spatial structure of the users and items. We apply the model developed here to publicly available data from the website [`Yelp!`]{} and demonstrate that it produces interpretable clusterings of users according to their rating behavior, in addition to providing better predictions of ratings and better summaries of overall item quality.
**Key words and phrases:** Bayesian hierarchical model; data augmentation; nonparametric Bayes; ordinal data; recommender systems; spatial prediction.
author:
- 'Antonio R. Linero[^1] [^2] , Jonathan R. Bradley$^{*}$, and Apurva Desai$^{*}$'
-
-
-
- 'Antonio R. Linero, Jonathan R. Bradley, Apruva S. Desai'
bibliography:
- 'myref.bib'
- 'mybib.bib'
title:
- ' Multi-rubric Models for Ordinal Spatial Data with Application to Online Ratings Data '
- 'Multi-rubric Models for Ordinal Spatial Data with Application to Online Ratings Data'
- Supplementary Material
---
,
Introduction {#sec:introduction}
============
In recent years, the complexity of data used to make decisions has increased dramatically. A prime example of this is the use of online reviews to decide whether to purchase a product or visit a local business; we refer to the objects being reviewed as *items*. Consider data provided by [`Yelp!`]{} (see, <http://www.yelp.com/>), which allows users to rate items, such as restaurants, convenience stores, and so forth, on a discrete scale from one to five “stars.” Additional features of the businesses are also known, such as the spatial location and type of business. Datasets of this type are typically very large and exhibit complex dependencies.
As an example of this complexity, users of [`Yelp!`]{} effectively determine their own standards when rating a local business. We refer to the particular standards a user applies as a *rubric*. We might imagine a latent variable $Y_{iu}$ representing the *utility*, or benefit, user $u$ obtained from item $i$. For a given level of utility, however, different users may still give different ratings due to having different standards for the ratings; for example, one user may rate a restaurant 5 stars as long as it provides a non-offensive experience, a second user might require an exceptional experience to rate the same restaurant 5 stars, and a third user may rate all items with $1$ star in order to “troll” the website. Each of these users are applying different rubrics in translating their utility to a rating for the restaurant. In addition we also expect user-specific selection bias in the sense that some users may rate every restaurant they attend, while other users may only rate restaurants that they feel strongly about.
This article makes several contributions. First, we develop a semiparametric Bayesian model which accounts for the existence of multiple rubrics for ratings data that are observed over multiple locations. To do this, we use a spatial cumulative probit model [e.g., see @Higgs; @BerretProbit; @schliep2015data] in which the break-points are modeled as user-specific random effects. This requires a flexible model for the distribution $F$ of the random effects, which we model as a discrete mixture. A by-product of our approach is that we obtain a clustering of users according to the rubrics they are using.
Second, we use the multi-rubric model to address novel inferential questions. For example, ratings provided to a user might be adjusted to match that user’s rubric, or to provide a distribution for the rating that a user would provide conditional on having a particular rubric. Utilizing this user-specific standardization of ratings may provide users with better intuition for the overall quality of an item.
This adjustment of restaurant quality for the rubrics is similar to, but distinct from, the task of predicting a user’s ratings. Good predictive performance is required for *filtering*, which refers to the task of processing the rating history of a user and producing a list of recommendations [for a review, see @bobadilla2013recommender]. As a third contribution, we show that allowing for multiple rubrics improves predictions.
The model proposed here also has interesting statistical features. A useful feature of our model is that it allows for more accurate comparisons across items. For example, if a user rates all items with $1$ star, then the model discounts this user’s ratings. This behavior is desirable for two reasons. First, if a user genuinely rates all items with $1$ star, then their rating is unhelpful. Second, it down-weights the ratings of users who are exhibiting selection bias and only rating items which they feel strongly about, which is desirable as comparisons across items will be more indicative of true quality if they are based on individuals who are not exhibiting large degrees of selection bias.
Additionally, the rubrics themselves may be of intrinsic interest. We demonstrate that the rubrics learned by our model are highly interpretable. For example, when analyzing the [`Yelp!`]{} dataset in Section \[sec:data-analysis\], we obtain Figure \[fig:rubric-props\] which displays the ratings observed for users assigned to a discrete collection of rubrics and reveals several distinct rating patterns displayed by users.
Other features of our model are also of potentially independent interest. The multi-rubric model can be interpreted as a novel semiparametric random-effects model for ordinal data, even for problems in which the intuition behind the multi-rubric model in terms of latent utility does not hold. Other study designs in which the multi-rubric analogy may be useful include longitudinal survey studies, or more general ordinal repeated-measures designs. Additionally, the cumulative probit model we use to model latent user preferences includes a spatial process to account for spatial dependencies across local businesses. Recovering an underlying spatial process allows for recommending entire regions to visit, rather than singular items. The development of low-rank spatial methodology for large-scale dependent ordinal data is of interest within the spatial literature, as the current spatial literature for ordinal data do not typically address large datasets on a similar order of the [`Yelp!`]{} dataset [e.g., see @de2004simple; @de2000bayesian; @Chen2000; @Cargnoni; @KnorrHeld; @CarlinDiscreteCat; @Higgs; @BerretProbit; @rainfallCat among others]. We model the underlying spatial process using a low-rank approximation [@johan] to a desired Gaussian process [@banerjee; @bradleyMSTM].
Starting from @koren2011ordrec, several works in the recommender systems literature have considered ordinal matrix factorization (OMF) procedures which are similar in many respects to our model (see also @paquet2012hierarchical and @houlsby2014cold). Our work differs non-trivially from these works in that the multi-rubric model treats the break-points as user-specific random effects, with a nonparametric prior used for the random effects distribution $F$. For the [`Yelp!`]{} dataset, this extra flexibility leads to improved predictive performance. Additionally, our focus in this work extends to inferential goals beyond prediction; for example, depending on the distribution of the rubrics of users who rate a given item, the estimate of overall quality for that item can be shrunk to a variety of different centers, producing novel multiple-shrinkage effects. Several works in the Bayesian nonparametric literature have also considered flexible models for random effects in multivariate ordinal models [@kottas2005; @deyoreo2014bayesian; @bao2015bayesian], but do not treat the break-points themselves as random effects.
The paper is organized as follows. In Section \[sec:multi-rubric\], we develop the multi-rubric model, with an eye towards the [`Yelp!`]{} dataset, and provide implementation details. In Section \[sec:simulation\], we illustrate the methodology on synthetic data designed to mirror features of the [`Yelp!`]{} dataset, and demonstrate that we can accurately recover the number and structure of the rubrics when the model holds, as well as effectively estimate the underlying latent utility field. In Section \[sec:data-analysis\], we illustrate the methodology on the [`Yelp!`]{} dataset. We conclude with a discussion in Section \[sec:discussion\]. In supplementary material, we present simulation experiments which demonstrate identifiability of key components of the model.
The Multi-rubric model {#sec:multi-rubric}
======================
Preliminary notation
--------------------
We consider ordinal response variables $Z_{iu}$ taking values in $\{1, \ldots, K\}$. In the context of online ratings data, $Z_{iu}$ represents the rating that user $u$ provides for item $i$. In the context of survey data, on the other hand, $Z_{iu}$ might represent the response subject $u$ gives to question $i$. We do not assume that $Z_{iu}$ is observed for all $(i,u)$ pairs, but instead we observe $(i,u) \in \mathcal S \subseteq \{1, \ldots, I\} \times \{1, \ldots, U\}$, where $U$ is the total number of subjects and $I$ is the total number of items. For fixed $i$ we let ${\mathcal U}_i = \{u : (i,u) \in {\mathcal S}\}$ be the set of users that rate item $i$, and similarly for fixed $u$ we let ${\mathcal I}_u = \{i: (i,u) \in {\mathcal S}\}$ be the set of items that user $u$ rates.
Review of Cumulative Probit Models {#sec:review-probit}
----------------------------------
Cumulative probit models [@chib-93; @albert1997bayesian] provide a convenient framework for modeling ordinal rating data. Consider the univariate setting, with ordinal observations $\{Z_i : 1 \le i \le N\}$ taking values in $\{1,\ldots,K\}$. We assume that $Z_i$ is a rounded version of a latent variable $Y_i$ such that $Z_i = k$ if $\theta_{k-1} \le Y_i < \theta_k$. Here, $-\infty = \theta_0 \le \theta_1 \le \cdots \le \theta_K = \infty$ are unknown break-points. When $Y_i$ has the Gaussian distribution $Y_i \sim {\operatorname{Gau}}(x_i{^{\top}}\gamma, 1)$ this leads to the ordinal probit model, where $\Pr(Z_i = k \mid \theta, \gamma) = \Phi(\theta_k - x_i{^{\top}}\gamma) - \Phi(\theta_{k-1} - x_i{^{\top}}\gamma)$.
We assume ${\operatorname{Var}}(Y_i) = 1$, as the variance of $Y_i$ is confounded with the break-points $\theta = (\theta_1, \ldots, \theta_{K-1})$. Any global intercept term is also confounded with the $\theta$’s; there are two resolutions to this issue. The first is to fix one of the $\theta_k$’s, e.g., $\theta_1 \equiv 0$. The second is to exclude an intercept term from $x_i$. While the former approach is often taken [@albert1997bayesian; @Higgs], it is more convenient in the multi-rubric setting to use the latter approach to avoid placing asymmetric restrictions on the break-points.
The ordinal probit model is convenient for Bayesian inference in part because it admits a simple data augmentation algorithm which iterates between sampling $Y_i {\stackrel{\textnormal{indep}}{\sim}}{\operatorname{TruncGau}}(x_i{^{\top}}\gamma, 1, \theta_{Z_i-1}, \theta_{Z_i})$ for $1 \le i \le N$ and, assuming a flat prior for $\gamma$, sampling $ \gamma
\sim
\Normal\{(X\trans X)^{-1}X\trans \bm Y, (X\trans X)^{-1}\},$ where $X$ has $i^{\text{th}}$ row $x_i{^{\top}}$ and $\bm Y = (Y_1, \ldots,
Y_N)$. Here, ${\operatorname{TruncGau}}(\mu, \sigma^2, a, b)$ denotes the Gaussian distribution truncated to the interval $(a,b)$. Additionally, an update for $\theta$ is needed. Efficient updates for $\theta$ can be implemented by using a Metropolis-within-Gibbs step to update $\theta$ as a block (for details, see [@albert1997bayesian], as well as [@cowles1996accelerating] for alternative MCMC schemes).
Description of the proposed model {#sec:description}
---------------------------------
### The multi-rubric model {#the-multi-rubric-model}
We develop an extension of the cumulative probit model to generic repeated-measures ordinal data $\{Z_{iu} : (i,u) \in \mathcal S\}$. Following @albert1997bayesian we introduce latent utilities $Y_{iu}$, but specify a generic ANOVA model $$\begin{aligned}
\label{eq:anova}
Y_{iu} = f_{iu} + \nu_u + \xi_i + \epsilon_{iu}, \qquad \epsilon_{iu} {\stackrel{\textnormal{iid}}{\sim}}{\operatorname{Gau}}(0, 1),\end{aligned}$$ where $\nu_u$ and $\xi_i$ are main effects and $f_{iu}$ is an interaction effect. The multi-rubric model modifies the cumulative probit model by replacing the break-point parameter $\theta$ with $u$-specific random effects $\theta_u = (\theta_{u0}, \ldots, \theta_{uK})$ with $[\theta_u \mid F] {\stackrel{\textnormal{indep}}{\sim}}F$ for some unknown $F$. As before, we let $Z_{iu} = k$ if $\theta_{u(k-1)} \le Y_{iu} \le \theta_{uk}$. For concreteness, we take $F$ to be a finite mixture $F = \sum_{m = 1}^M \omega_m \delta_{\theta^{(m)}}$ for some large $M$, with $\theta^{(m)} {\stackrel{\textnormal{iid}}{\sim}}H$ and $\omega \sim {\operatorname{Dirichlet}}(a, \ldots, a)$, where $\delta_{\theta{^{(m)}}}$ is a point-mass distribution at $\theta{^{(m)}}$. We note that it is also straight-forward to use a nonparametric prior for $F$ such as a Dirichlet process [@escobarwest1995; @ferguson1973]. We refer to the random effects $\theta^{(1)}, \ldots, \theta^{(M)}$ as *rubrics*. Note that for each subject $u$ there exists a latent class $m$ such that $\theta_u = \theta^{(m)}$.
Figure \[fig:multi-rubric-illustration\] displays the essential idea for the model. Viewing $Y$ as a latent utility, the rubric that the user is associated to leads to different values of the observed rating $Z$. In this example, the second rubric is associated to users who rate many items with a $3$, while the first rubric is associated to users who do not rate many items with a $3$.
![Visualization of the multi-rubric model. The point on the density indicates the realized value of $Y$.[]{data-label="fig:multi-rubric-illustration"}](./Figures/multi-rubric-illustration-1){width=".8\textwidth"}
Treating the break-points as random effects has several benefits. First, it offers additional flexibility over approaches for ordinal data which incorporate a random intercept [@gill2009nonparametric]. Due to the fact that the $\theta_u$’s are confounded with both the location and scale of $Y_{iu}$, treating the break-points as random effects is at least as flexible as treating the location and scale of the distribution of the $Y_{iu}$’s as random effects. We require this additional flexibility, as merely treating the location and scale of the $Y_{iu}$’s as random effects does not allow for the variety of rating behaviors exhibited by users. By treating the break-points as random effects, we are able to capture *any* distribution of ratings in a given rubric (see, e.g., Figure \[fig:rubric-props\]). In addition to flexibility, specifying $F$ as a discrete mixture induces a clustering of users into latent classes. To each user $u$ we associate a latent variable $C_u$ such that $C_u = m$ if $\theta_u = \theta^{(m)}$. As will be demonstrated in Section \[sec:data-analysis\], the latent classes of users discovered in this way are highly interpretable.
### Model for the [`Yelp!`]{} data
Our model for the [`Yelp!`]{} data takes $Y_{iu} \sim {\operatorname{Gau}}(\mu_{iu}, 1)$ where $$\begin{aligned}
\mu_{iu} =
x_i{^{\top}}\gamma +
\alpha_u{^{\top}}\beta_i +
W_i +
b_i,
\qquad
W_i = \psi(s_i){^{\top}}\eta.\end{aligned}$$ This is model with $f_{iu} = \alpha_u{^{\top}}\beta_i$, $\xi_i = x_i{^{\top}}\gamma + W_i + b_i$, and $\nu_u$ removed. This model can be motivated as a combination of the fixed-rank kriging approach of @johan with the probabilistic matrix factorization approach of @salakhutdinov2007probabilistic. The terms $x_i{^{\top}}\gamma$, $W_i$, and $b_i$ are used to account for heterogeneity in the items. The term $x_i{^{\top}}\gamma$ accounts for known covariates $x_i \in {\mathbb R}^p$ associated to each item. The term $W_i$ is used to capture spatial structure, and is modeled with a basis function expansion $W_i = \psi(s_i){^{\top}}\eta$ where $s_i$ denotes the longitude-latitude coordinates associated to the item and $\psi(s) = (\psi_1(s) \ldots, \psi_r(s)){^{\top}}$ is a vector of basis functions. We note that it is straight-forward to replace our low-rank approach for $W_i$ with more elaborate approaches such as the full-scale approach of @sang2012full. The term $b_i$ is an item-specific random effect which is used to capture item heterogeneity which cannot be accounted for by the covariates or the low-rank spatial structure.
The vectors $\alpha_u$ and $\beta_i$ intuitively correspond to unmeasured user-specific and item-specific latent features. The term $\alpha_u{^{\top}}\beta_i$ is large/positive when $\alpha_u$ and $\beta_i$ point in the same direction (i.e., the user’s preferences align with the item’s characteristics), and is large/negative when $\alpha_u$ and $\beta_i$ point in opposite directions. This allows the model to account not only for user-specific biases ($\theta_u$) and item-specific biases $(x_i, W_i, b_i)$, but also interaction effects.
The multi-rubric model can be summarized by the following hierarchical model. For each model, we implicitly assume the statements hold conditionally on all variables in the models below, and that conditional independence holds within each model unless otherwise stated.
Response model:
: $Z_{iu} = k$ with probability $w_{iuk} = \Phi(\theta_{uk} - \mu_{iu}) - \Phi(\theta_{u(k-1)} - \mu_{iu})$ and $\mu_{iu} = x_i{^{\top}}\gamma + \alpha_u{^{\top}}\beta_i + W_i + b_i$.
Random effect model:
: $\theta_u {\stackrel{\textnormal{iid}}{\sim}}F$, $\alpha_{u} \sim {\operatorname{Gau}}(0,\sigma^2_\alpha {\operatorname{I}})$, $\beta_{i} \sim {\operatorname{Gau}}(0,\sigma^2_\beta {\operatorname{I}})$, and $b_i \sim {\operatorname{Gau}}(0, \sigma^2_b)$.
Spatial process model:
: $W_i = \psi(s_i){^{\top}}\eta$ where $\eta \sim {\operatorname{Gau}}(0, \Sigma_\eta)$.
Parameter model:
: $\gamma \sim \operatorname{Flat}$ and $F = \sum_{m=1}^M \omega_m \delta_{\theta^{(m)}}$ where $\omega \sim {\operatorname{Dirichlet}}(a,\ldots,a)$ and $\theta^{(m)} {\stackrel{\textnormal{iid}}{\sim}}H$.
To complete the model we must specify values for the hyperparameters $\sigma_\alpha, \sigma_\beta, \sigma_b, \Sigma_\eta, a$, and $H$, as well as the number of rubrics $M$ and the number of latent factors $L$. In our illustrations we place half-Gaussian priors for the scale parameters, with $(\sigma_\beta, \sigma_b) {\stackrel{\textnormal{iid}}{\sim}}{\operatorname{Gau}}_+(0, 1)$, and $\sigma_\alpha \equiv 1$. We let $\Sigma_\eta = {\operatorname{diag}}(\sigma^2_\eta, \ldots, \sigma^2_\eta)$ and set $\sigma_\eta \sim {\operatorname{Gau}}_+(0,1)$. Here, ${\operatorname{Gau}}_+(0,1)$ denotes a standard Gaussian distribution truncated to the positive reals. For a discussion of prior specification for variance parameters, see @gelman2006prior and @simpson2017penalising.
In our illustrations we use $M = 20$. For the [`Yelp!`]{} dataset, the choice of $M = 20$ rubrics is conservative, and by setting $a = \kappa / M$ for some fixed $\kappa > 0$, we encourage $\omega$ to be nearly-sparse [@ishwaran2002dirichlet; @linero2016bayesian]. This strategy effectively lets the data determine how many rubrics are needed, as the prior encourages $\omega_m \approx 0$ if rubric $m$ is not needed. The prior $H$ for $\theta^{(1)}, \ldots, \theta^{(M)}$ is chosen to have density $h(\theta) = \prod_{k=1}^K {\operatorname{Gau}}(\theta_k \mid 0, \sigma_\theta^2) I(\theta_1 \le \cdots \le \theta_{K-1})$ so that $\theta^{(m)}$ has the distribution of the order statistics of $K-1$ independent ${\operatorname{Gau}}(0, \sigma^2_\theta)$ variables.
Evaluating item quality
-----------------------
A commonly used measure of item quality is the average rating of a user from the population $\lambda_i = E(Z_{iu} \mid x_i, \phi_i, \gamma)$ where $\phi_i = (\beta_i, b_i, W_i)$. This quantity is given by $$\begin{aligned}
\lambda_i
&= \sum_{k = 1}^K k \cdot \Pr(Z_{iu} = k \mid x_i, \phi_i, \gamma)
\\&= \sum_{k=1}^K
\sum_{m=1}^M k \cdot \omega_m \cdot
\int \Pr(Z_{iu} = k \mid x_i, \phi_i, \alpha_u, \gamma, C_u = m)
\, {\operatorname{Gau}}(\alpha \mid 0, \sigma^2_\alpha {\operatorname{I}}) \ d\alpha.\end{aligned}$$ Using properties of the Gaussian distribution, and recalling that $\sigma^2_\alpha = 1$, it can be shown that $$\begin{aligned}
\label{eq:lambda}
\lambda_i
=
\sum_{k=1}^K \sum_{m = 1}^M k \cdot \omega_m \cdot
\left\lbrace
\Phi\left(\frac{\theta^{(m)}_k - \xi_i}{\sqrt{1 + \|\beta_i\|^2}}\right) -
\Phi\left(\frac{\theta^{(m)}_{k-1} - \xi_i}{\sqrt{1 + \|\beta_i\|^2}}\right)
\right\rbrace,\end{aligned}$$ where $\xi_i = x_i{^{\top}}\gamma + b_i + W_i$. In Section \[sec:data-analysis\], we demonstrate the particular users who rated item $i$ exert a strong influence on the $\lambda_i$’s, particularly for restaurants with few ratings.
Rather than focusing on an omnibus measure of overall quality, we can also adjust the overall quality of an item to be rubric-specific. This amounts to calculating $ \lambda_{im} = E(Z_{iu} \mid x_i, \phi_i, \gamma, C_u = m), $ which represents the average rating of item $i$ if all used rubric $m$. Similar to , this quantity can be computed as $$\begin{aligned}
\label{eq:lambda-k}
\lambda_{im}
=
\sum_{k=1}^K k \cdot
\left\lbrace
\Phi\left(\frac{\theta^{(m)}_k - \xi_i}{\sqrt{1 + \|\beta_i\|^2}}\right) -
\Phi\left(\frac{\theta^{(m)}_{k-1} - \xi_i}{\sqrt{1 + \|\beta_i\|^2}}\right)
\right\rbrace.\end{aligned}$$ In Section \[sec:data-analysis\], we use both and to understand the statistical features of the multi-rubric model.
Implementation Details
----------------------
We use the reduced rank model $W = \Psi\eta + b$ where $\Psi\in{\mathbb R}^{I \times r}$ has $i^{\text{th}}$ row given by $\psi(s_i){^{\top}}$. We choose $\Psi$ so that ${\operatorname{Cov}}(\Psi\eta)$ is an optimal low-rank approximation to $\sigma^2_\eta \Xi$ where $\Xi$ is associated to a target positive semi-definite covariogram. This is accomplished by taking $\Psi$ composed of the first $r$ columns of $\Gamma D^{1/2}$ where $\Xi = \Gamma D \Gamma{^{\top}}$ is the spectral decomposition of $\Xi$. The Eckart-Young-Mirsky theorem states that this approximation is optimal with respect to both the operator norm and Frobenius norm [see, e.g., @rasmussen2005gaussian Chapter 8]. A similar strategy is used by @bradleyPCOS [@bradleyMSTM], who use an optimal low-rank approximation of a target covariance structure $\Xi \approx \Psi \Sigma_\eta \Psi{^{\top}}$ where the basis $\Psi$ is held fixed but $\Sigma_{\eta}$ is allowed to vary over all positive-definite $r \times r$ matrices. In our illustrations, we use the squared-exponential covariance, i.e., $\Xi_{ij} = \exp({-{\rho}\|s_i - s_j\|^2})$ [@cressie2015statistics].
To complete the specification of the model, we must specify the bandwidth $\rho$, the number of latent factors $L$, and the number of basis functions $r$. We regard $L$ as a tuning parameter, which can be selected by assessing prediction performance on a held-out subset of the data. In principle, a prior can be placed on $\rho$, however this results in a large computational burden; we instead evaluate several fixed values of $\rho$ chosen according to some rules-of-thumb and select the value with the best performance. For the [`Yelp!`]{} dataset, we selected $\rho = 1000$, which corresponds undersmoothing the spatial field relative to Scott’s rule [see, e.g., @hardle2000multivariate] by roughly a factor of two, and remark that substantively similar results are obtained with other bandwidths. Finally, $r$ can be selected so that the proportion of the variance $\sum_{d=1}^r D_{ii}^2 / \sum_{d=1}^n D_{ii}^2$ in $\Xi$ accounted for by the low-rank approximation exceeds some preset threshold; for the [`Yelp!`]{} dataset, we chose $r = 500$ to account for 99% of the variance in $\Xi$.
When specifying the number of rubrics $M$, we have found that the model is most reliable when $M$ is chosen large and $a = \kappa / M$ for some $\kappa > 0$; under these conditions, the prior for $F$ is approximately a Dirichlet process with concentration $\kappa$ and base measure $H$ (see, e.g., [@teh2006hierarchical]). We recommend choosing $M$ to be conservatively large and allowing the model to remove unneeded rubrics through the sparsity-inducing prior on $\omega$. We have found that taking $M$ large is necessary for good performance even in simulations in which the true number of rubrics is small and known.
We use Markov chain Monte Carlo to approximately sample from the posterior distribution of the parameters. A description of the sampler is given in the appendix.
A note on selection bias
------------------------
Let $\Delta_{iu} =1$ if $(i,u) \in {\mathcal S}$, and $\Delta_{iu} = 0$ otherwise. In not modeling the distribution of $\Delta_{iu}$, we are implicitly modeling the distribution of the $Z_{iu}$’s conditional on $\Delta_{iu} = 1$. When selection bias is present, this may be quite different than the marginal distribution of $Z_{iu}$’s. Experiments due to @marlin2009collaborative provide evidence that selection bias may be present in practice.
A useful feature of the approach presented here is that it naturally down-weights users who are exhibiting selection bias. For example, if user $u$ only rates items they feel negatively about, they will be assigned to a rubric $m$ for which $\theta^{(m)}_{1}$ is very large; this has the effect of ignoring their ratings, as there will be effectively no information in the data about their latent utility. As a result, when estimating overall item quality, the model naturally filters out users who are exhibiting extreme selection bias, which may be desirable.
In the context of prediction, the predictive distribution for $Z_{iu}$ should be understood as being conditional on the event $\Delta_{iu} = 1$; that is, the prediction is made with the additional information that user $u$ chose to rate item $i$. This is the case for nearly all collaborative filtering methods, as correcting for the selection bias necessitates collecting $Z_{iu}$’s for which $\Delta_{iu} = 0$ would occurred naturally; for example, as done by @marlin2009collaborative, we might assess selection bias by conducting a pilot study which forces users to rate items they would not have normally rated. With the understanding that all methods are predicting ratings conditional on $\Delta_{iu} = 1$, the results in Section \[sec:data-analysis\] show that the multi-rubric model leads to increased predictive performance.
Selection bias should also be taken into account when interpreting the latent rubrics produced by our model. Our model naturally provides a clustering of users into latent classes, which we presented as representing differing standards in user ratings; however, we expect that the model is also detecting differences in selection bias across users. We emphasize that our goal is to identify and account for heterogeneity in rating patterns, and we avoid speculating on whether heterogeneity is caused by different rating standards or selection bias. For example, a user who rates items with only one-star or five-stars might be either (i) using a rubric which results in extreme behavior, with most of the break-points very close together; or (ii) actively choosing to rate items which they feel strongly about.
Simulation Study {#sec:simulation}
================
The goal of this simulation is to illustrate that we can accurately learn the existence of multiple rubrics in settings where one would expect it would be difficult to detect them. We consider a situation where the data is generated according to two rubrics that are similar to each other. This allows us to assess the robustness of our model to various “degrees” of the multi-rubric assumption. The performance of our multi-rubric model is assessed relative to the single-rubric model, which is the standard assumption made in the ordinal data literature.
We calibrate components of the simulation model towards the [`Yelp!`]{} dataset to produce realistic simulated data. Specifically, we set $\eta$ and $\sigma^2_b$ equal to the posterior means obtained from fitting the model to the [`Yelp!`]{} dataset in Section \[sec:data-analysis\]. We set $\Sigma_\eta = 0.5{\operatorname{I}}$, corresponding to a much stronger spatial effect than what was observed in the data, and for simplicity we removed the latent-factor aspect of the model by fixing $\sigma^2_\beta \equiv 0$. A two-rubric model is used with $\omega_1 = \omega_2 = 0.5$. We also use the same spatial basis functions and observed values of $(i,u)$ as in the [`Yelp!`]{} analysis in Section \[sec:data-analysis\].
We now describe how the two rubrics $\theta_1$ and $\theta_2$ where chosen. First, $\theta_1$ was selected so that $\{Z_{iu} : C_u = 1, i = 1, \ldots, I\}$ was evenly distributed among the five responses. Associated to $\theta_1$ is a probability vector $p_1 = (0.2, 0.2, 0.2, 0.2, 0.2)$. To specify $\theta_2$, we use the same approach with a difference choice of $p$. Let $p_2 = (0, 0.25, 0.5, 0.25, 0)$. Then $\theta_2^{(\tau)}$ is associated to $\tau p_1 + (1 - \tau) p_2$ in the same manner as $\theta_1$ is associated to $p_1$. Here, $\tau$ indexes the similarity of $\theta_1$ and $\theta_2$, and it can be shown that the total variation distance between the empirical distribution of $\{Z_{iu} : C_u = 1\}$ and $\{Z_{iu} : C_u = 2\}$ is $0.8(1 - \tau)$. Thus, values of $\tau$ near $1$ correspond imply that the rubrics are similar, while values of $\tau$ near $0$ imply that they are dissimilar. Figure \[fig:sim-rubrics\] presents the distribution of the $Z_{iu}$’s with $C_u = 2$ when $\tau = 0, 0.8$, and $1$.
![Empirical distribution of the $Z_{iu}$’s in the simulation model, for $\theta_1$, $\theta^{(0.8)}_2$, and $\theta_2^{(0)}$.[]{data-label="fig:sim-rubrics"}](./Figures/tau-figure-1){width=".8\textwidth"}
We fit a $10$-rubric and single-rubric model for $\tau = 0.0, 0.1, \ldots, 1.0$. Figure \[fig:apurav-sims\] displays the proportion of individuals assigned to each rubric for a given value of $\tau$. If the model is accurately recovering the underlying rubric structure, we expect to see a half of the observations assigned to one rubric, and half to another; due to permutation invariance, which of the 10 rubrics is associated to $\theta_1$ and $\theta_2^{(\tau)}$ vary by simulation. Up to $\tau = 0.9$, the model is capable of accurately recovering the existence of two rubrics. We also see that, even at $\tau = 0.8$, the model accurately recovers the empirical distribution of the $Z_{iu}$’s associated to each rubric.
Next, we assess the benefit of using the multi-rubric model to predict missing values. For each value of $\tau$, we fit a single-rubric and multi-rubric model. Using the same train-test split as in the our real data illustration, we compute the log likelihood on the held-out data $ \text{loglik}_{\text{test}} = \sum_{(i,u) \in \mathcal{S}_{\text{test}}} \log
\Pr(Z_{iu} \mid D),$ which is further discussed in detail in Section \[sec:data-analysis\] . Figure \[fig:lambda-loss\] shows the difference in held-out log likelihood for the single-rubric and multi-rubric model as a function of $\tau$. Up-to $\tau = 0.8$, there is a meaningful increase in the held-out log-likelihood obtained from using the multi-rubric model. The case where $\tau = 1$ is also particularly interesting, as this implies that the data were generated from the single rubric model. Here the predictive performance of our model at missing values appears to be robust to the case when the multiple rubric assumption is incorrect.
Displayed above each point in Figure \[fig:lambda-loss\] is the proportion of observations which are assigned to the correct rubric, where each observation is assigned to their most likely rubric. When the rubrics are far apart the model is capable of accurately assigning observations to rubrics. As the rubrics get closer together, the task of assigning observations to rubrics becomes much more difficult.
This simulation study suggests that the model specified here is able to disentangle the two-rubric structure, even when the rubrics are only subtly different. This leads to clear improvements in predictive performance for small and moderate values for $\tau$. Additionally, when the multi-rubric assumption is negligible, or even incorrect, our model performs as well as the single-rubric model.
Analysis of Yelp data {#sec:data-analysis}
=====================
![Estimate of the underlying spatial field $W(s) = \psi(s){^{\top}}\eta$ at each realized restaurant location using its posterior mean. []{data-label="fig:w-process"}](./Figures/W_process){width=".45\textwidth"}
We now apply the multi-rubric model to the [`Yelp!`]{} dataset, which is publicly available at <https://www.yelp.com/dataset_challenge>. We begin by preprocessing the data to include reviews only between January $1^{\text{st}}$, 2013 and December $31^{\text{st}}$ 2016, and restrict attention to restaurants in Phoenix and its surrounding areas. We further narrow the data to include only users who rated at least 10 restaurants; this filtering is done in an attempt to minimize selection bias, as we believe that “frequent raters” should be less influenced by selection bias.
We first evaluate the performance of the single-rubric and multi-rubric models for various values of the latent factor dimension $L$. We set $M = 20$ and induce sparsity in $\omega$ by setting $\omega \sim {\operatorname{Dirichlet}}(1/20, \ldots, 1/20)$. We divide the indices $(i,u) \in {\mathcal S}$ into a training set ${\mathcal S}_{\text{train}}$ and testing set ${\mathcal S}_{\text{test}}$ of equal sizes by randomly allocating half of the indices to the training set. We evaluate predictions using a held-out log-likelihood criteria $$\begin{aligned}
\label{eq:criteria}
\begin{split}
\text{loglik}_{\text{test}} &= |{\mathcal S}_{\text{test}}|^{-1}\sum_{(i,u) \in {\mathcal S}_{\text{test}}} \log
\Pr(Z_{iu} \mid {\mathcal D}), \\&
\approx |{\mathcal S}_{\text{test}}|^{-1}\sum_{(i,u) \in {\mathcal S}_{\text{test}}} \log
T^{-1} \sum_{t=1}^T \Pr(Z_{iu} \mid C_u^{(t)}, \theta^{(t)}, \mu_{iu}^{(t)}), \end{split}\end{aligned}$$ where ${\mathcal D}= \{Z_{iu} : (i,u) \in {\mathcal S}_{\text{train}}\}$, $\Pr(Z_{iu} \mid {\mathcal D})$ denotes the posterior predictive distribution of $Z_{iu}$, and $t = 1, \ldots, T$ indexes the approximate draws from the posterior obtained by MCMC. Results for the values $L = 1, 3$, and 5, over 10 splits into training and test data, are given in Figure \[fig:heldout-results\]. We also compare our methodology to ordinal matrix factorization [@paquet2012hierarchical] with learned breakpoints and spatial smoothing, and the mixture multinomial model [@marlin2009collaborative] with $10$ mixture components. The multi-rubric model leads to an increase in the held-out data log-likelihood of roughly $5\%$ over ordinal matrix factorization and $8\%$ over the mixture multinomial model. Additionally, we note that the holdout log-likelihood was very stable over replications. The single-rubric model is essentially equivalent to ordinal matrix factorization.
The dimension of the latent factors $\alpha_u$ and $\beta_i$ has little effect on the quality of the model. We attribute this to the fact that $|{\mathcal U}_i|$ and ${\mathcal I}_u|$ are typically small, making it difficult for the model to recover the latent factors. On other datasets where this is not the case, such as the Netflix challenge dataset, latent-factor models represent the state of the art and are likely essential for the multi-rubric model. In the supplementary material we show in simulation experiments that the $\alpha_u$’s, $\beta_i$’s, and $L$ are identified.
![Boxplots of $-2.0 \cdot \text{loglik}_{\text{test}}$ for the mixture multinomial model (MMM, which does not have latent factors), ordinal matrix factorization (OMF), the single rubric model (Single) and the multi-rubric model (Multi), for 10 splits into training and testing data.[]{data-label="fig:heldout-results"}](./Figures/eval-1){width="100.00000%"}
Figure \[fig:w-process\] displays the learned spatial field $\widehat W(s) = \psi(s){^{\top}}\widehat\eta$ where $\widehat \eta$ is the posterior mean of $\eta$. The results suggest that the downtown Phoenix business district and the area surrounding the affluent Paradise Valley possesses a higher concentration of highly-rated restaurants than the rest of the Phoenix area. More sparsely populated areas such as such as Litchfield Park, or areas with lower income such as Guadalupe, seem to have fewer highly-rated restaurants.
![ Top: bar chart giving the number of users assigned to each rubric, where users are assigned to rubrics by minimizing Binder’s loss function. Bottom: bar charts giving the proportions of the observed ratings $Z_{iu}$ for each item-user pair for the top 9 most common rubrics.[]{data-label="fig:rubric-props"}](./Figures/rubric-ratings-all-1){width=".9\textwidth"}
We now examine the individual rubrics. First, we obtain a clustering of users into their rubrics by minimizing Binder’s loss function [@binder1978bayesian] $L(\bm c) = \sum_{u,u'} |\delta_{c_u, c_{u'}} - \Pi_{u,u'}$, where $\delta_{ij} = I(i = j)$ is the Kronecker delta, $\bm c = (c_1, \ldots, c_U)$ is an assignment of users to rubrics, and $\Pi_{u,u'}$ is the posterior probability of $C_u = C_{u'}$. See @fritsch2009improved for additional approaches to clustering objects using samples of the $C_u$’s.
The multi-rubric model produces interesting effects on the overall estimate of restaurant quality. Consider the rubric corresponding to $m = 7$ in Figure \[fig:rubric-props\]. Users assigned to this rubric give the majority of restaurants a rating of five stars. As a result, a rating of 5 stars for the $m = 7$ rubric is less valuable to a restaurant than a rating of 5 stars from a user with the $m = 6$ rubric. Similarly, a rating of $3$ stars from the $m = 7$ rubric is more damaging to the estimate of a restaurant’s quality than a rating of $3$ stars from the $m = 6$ rubric.
For restaurants with a large number of reviews, the effect mentioned above is negligible, as the restaurants typically have a good mix of users from different rubrics. The effect on restaurants with a small number of reviews, however, can be much more pronounced. To illustrate this effect, Figure \[fig:rating-posterior\] displays the posterior distribution of the quantity $\lambda_i$ defined in for the restaurants with $i \in \{ 3356, 3809, 9\}$. Each of these businesses has $4$ reviews total, with empirically averaged ratings of 4.25, 3.75, and 3 stars. For $i = 3809$ and $i = 9$, the users are predominantly from the rubric with $m = 7$; as a consequence, the fact that these restaurants do not have an average rating closer to five stars is quite damaging to the estimate of the restaurant quality. In the case of $i = 3809$, the effect is strong enough that what was ostensibly an above-average restaurant is actually estimated to be below average by the multi-rubric model. Conversely, item $i = 3356$ has ratings of $4, 5, 5$, and $3$ stars, but one of the $5$-star ratings comes from a user assigned to the rubric $m = 2$ which gave a $5$-star rating to only 8% of businesses. As a result, the $5$-star ratings are weighted more heavily than they would otherwise be, causing the distribution of $\lambda_i$ to be shifted slightly upwards.
![Posterior density of $\lambda_i$ for $i \in \{9, 3356, 3809\}$. The dashed line is the empirical average rating of item $i$; the dotted line is the overall average of all ratings. Error bars are centered at the posterior mean with a radius of one standard deviation.[]{data-label="fig:rating-posterior"}](./Figures/rating-posterior-1){width=".8\textwidth"}
Lastly, we consider rescaling the average ratings according to a specific rubric. This may of interest, for example, if one is interested in standardizing the ratings to match a rubric which evenly disperses ratings evenly across the possible stars. To do this, we examine the rubric-adjusted average ratings $\lambda_{im}$ given by . Figure \[fig:rubric-specific\] displays the posterior density of $\lambda_{im}$ for $i = 24$ and $i = 44$, for the $9$ most common rubrics. These two restaurants have over 100 reviews, and so the overall quality can be estimated accurately. We see some expected features; for example, the quality of each restaurant has been adjusted downwards for users of the $m = 10$ rubric, and upwards for the $m = 7$ rubric. The multi-rubric model allows for more nuanced behavior of the adjusted ratings than simple upward/downward shifts. For example, for the mediocre item $i = 44$, we see that little adjustment is made for the $m = 13$ rubric, while for the high-quality item $i = 24$ a substantial downward adjustment is made. This occurs because the model interprets the users with $m = 13$ as requiring a relatively large amount of utility to rate an item 5 stars, so that a downward adjustment is made for the high-quality item; on the other hand, users with $m = 13$ tend to rate things near a $3.5$, so little adjustment needs to be made for the mediocre item.
![Posterior density of $\lambda_{im}$ for $i = 44, 24$. Horizontal lines display the empirical average rating for each item. Rubrics are organized by the average rating assigned to $i = 44$ for visualization purposes.[]{data-label="fig:rubric-specific"}](./Figures/rubric-specific-1){width=".8\textwidth"}
Discussion {#sec:discussion}
==========
In this paper we introduced the multi-rubric model for the analysis of rating data and applied it to public data from the website [`Yelp!`]{}. We found that the multi-rubric model yields improved predictions and induces sophisticated shrinkage effects on the estimated quality of the items. We also showed how the model developed can be used to partition the users into interpretable latent classes.
There are several areas exciting areas for future work. First, while Markov chain Monte Carlo works well for the [`Yelp!`]{} dataset (it took 90 minutes to fit the model of Section \[sec:data-analysis\]), it would be desirable to develop a more scalable procedure, such as stochastic variational inference [@hoffman2013stochastic]. Second, the model described here features limited modeling of the users. Information regarding which items the users have rated has been shown in other settings to improve predictive performance; temporal heterogeneity may also be present in users.
The latent class model described here can also be extended to allow for more flexible models for the $\alpha_u$’s and $\beta_i$’s. For example, a referee pointed out the possibility of inferring about how controversial an item is across latent classes, which could be accomplished naturally by using a mixture model for the $\alpha_u$’s.
A fruitful area for future research is the development of methodology for when MAR fails. One possibility for future work is to extend the model to also model the missing data indicators $\Delta_{iu}$. This is complicated by the fact that, while $\{Y_{iu} : 1 \le i \le I, 1 \le u \le U\}$ is not completely observed, $\{\Delta_{iu} : 1 \le i \le I, 1 \le u \le U\}$ is. As a result, the data becomes much larger when modeling the $\Delta_{iu}$’s.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank Eric Chicken for helpful discussions. This work was partially supported by the Office of the Secretary of Defense under research program \#SOT-FSU-FATs-06 and NSF grants NSF-SES \#1132031 and NSF-DMS \#1712870.
Markov chain Monte Carlo algorithm
==================================
Before describing the algorithm, we define several quantities. First, define $$\begin{aligned}
{3}
R_{iu}^{(\alpha)} &= Y_{iu} - \mu_{iu} + \alpha_i{^{\top}}\beta_u, &&\qquad&
R_{iu}^{(b)} &= Y_{iu} - \mu_{iu} + b_i, \\
R_{iu}^{(\gamma)} &= Y_{iu} - \mu_{iu} + x_i{^{\top}}\gamma, &&\qquad&
R_{iu}^{(\eta)} &= Y_{iu} - \mu_{iu} + \psi(s_i){^{\top}}\eta.\end{aligned}$$ Let $ \bm R^{(\alpha)}_u = \vec(R_{iu}^{(\alpha)} : i \in \sI_i),
\bm R^{(\beta)} = \vec(R_{iu}^{(\alpha)} : u \in \sU_i),
\bm R^{(b)}_i = \vec(R_{iu}^{(b)} : u \in \sU_i),
\bm R^{(\gamma)} = \vec(R_{iu}^{(\gamma)} : (i,u) \in \sS),
\text{ and }
\bm R^{(\eta)}_i = \vec(R_{iu}^{(\eta)} : (i,u) \in \sS).$ Then we can write $$\begin{aligned}
{3}
\bm R^{(\beta)}_i &= \bm A_i \beta_i + \bm \epsilon_i, &&\qquad&
\bm R^{(b)}_i &= {\mathds 1}b_i + \bm \epsilon_i, \\
\bm R^{(\alpha)}_u &= \bm B_u \alpha_i + \bm \epsilon_i, &&\qquad&
\bm R^{(\gamma)}_u &= \bm X \gamma + \bm \epsilon, \\
\bm R^{(\eta)}_u &= \bm \Psi \eta + \bm \epsilon,\end{aligned}$$ where $\bm A_i$ has rows composed of $\alpha$’s associated to users who rated item $i$, $\bm B_u$ has rows composed of $\beta$’s associated to items which were rated by user $u$, and $\bm X$ and $\bm \Psi$ are design matrices associated to the covariates and basis functions respectively. Holding the other parameters fixed, each term above on the left-hand-side is sufficient for its associated parameter on the right-hand-side.
A data augmentation strategy similar to the one proposed by @albert1997bayesian is employed. The updates for the parameters $\eta, \alpha, \beta$, and $\gamma$ use a back-fitting strategy based on the $\bm R$’s above. The Markov chain operates on the state space $ ( C, Y, \theta, b, \alpha, \beta, \gamma, \eta, \omega, \sigma_\alpha, \sigma_\beta, \sigma_\eta )$. We perform the following updates for each iteration of the sampling algorithm, where each step is understood to be done for each relevant $u$ and $i$.
1. Draw $C_u \sim {\operatorname{Categorical}}(\widehat \omega_{u1}, \ldots, \widehat
\omega_{uM})$ where $\widehat \omega_{um}$ is proportional to $\omega_m
\prod_{i \in \mathcal U_u} [\Phi(\theta_{Z_{iu}}^{(m)} - \mu_{iu}) -
\Phi(\theta_{Z_{iu}-1}^{(m)} - \mu_{iu})]$
2. Draw $Y_{iu} \sim {\operatorname{TruncGau}}(\mu_{iu}, 1, \theta^{(C_u)}_{k-1},
\theta^{(C_u)}_k)$, for $(i,u) \in \mathcal S$.
3. Draw $\alpha_u \sim {\operatorname{Gau}}(\widehat \alpha_u, \widehat
\Sigma_{\alpha_u})$ where $ \widehat \Sigma_{\alpha_u}
=
(\bm B_u\trans \bm B_u + \sigma_\alpha^2 I)^{-1}
$ and $\widehat \alpha_u = \widehat \Sigma_{\alpha_u} \bm B_u{^{\top}}\bm R^{(\alpha)}_u$.
4. Draw $\beta_i \sim {\operatorname{Gau}}(\widehat \beta_i, \widehat \Sigma_{\beta_i})$ where $ \widehat \Sigma_{\beta_i}
=
(\bm A_i\trans \bm A_i + \sigma^2_\beta I)^{-1}
$ and $\widehat \beta_i = \widehat \Sigma_{\beta_i} \bm A_i{^{\top}}\bm R^{(\beta)}_i$.
5. Draw $\gamma \sim {\operatorname{Gau}}(\widehat \gamma, \widehat \Sigma_\gamma)$ where $\widehat \Sigma_\gamma = (\bm X{^{\top}}\bm X)^{-1}$ and $\widehat
\gamma = \widehat \Sigma_\gamma \bm X{^{\top}}\bm R^{(\gamma)}$.
6. Draw $b_i \sim {\operatorname{Gau}}(\widehat b_i, \widehat \sigma^2_{b_i})$ where $ \widehat \sigma^2_{b_i} = (\sigma^{-2}_b + |\sU_i|)^{-1},
$ and $\widehat b_i = \widehat \sigma^2_{b_i} \sum_{u \in {\mathcal U}_i} R_{iu}^{(b)}$.
7. Draw $\eta \sim {\operatorname{Gau}}(\widehat \eta, \widehat \Sigma_\eta)$ where $ \widehat \Sigma_\eta
=
(\bm\Psi\trans\bm\Psi + \Sigma^{-1}_\eta)^{-1}
$ and $\widehat \eta = \widehat \Sigma_\eta \bm \Psi{^{\top}}\bm R^{(\eta)}$.
8. Draw $\omega \sim {\operatorname{Dirichlet}}(\widehat a_1, \ldots, \widehat
a_M)$ where $\widehat a_m = a + \sum_{u : C_u = m} 1$.
9. Make an update to $\sigma^2_b$ which leaves $ \Gam(\sigma^2_b \mid 0.5, 0.5) \prod_{i=1}^I \Normal(b_i \mid 0, \sigma_b^2)
$ invariant.
10. Make an update to $\sigma^2_\beta$ which leaves $ \Gam(\sigma^2_{\beta} \mid 0.5, 0.5) \prod_{i=1}^I \prod_{\ell = 1}^L
\Normal(\beta_{i\ell} \mid 0, \sigma_\beta^2)
$ invariant.
11. Make an update to $\sigma^2_\eta$ which leaves $ \Gam(\sigma^2_{\eta} \mid 0.5, 0.5) \prod_{j=1}^r
\Normal(\eta_j \mid 0, \sigma_\eta^2)
$ invariant.
12. Make an update to $\theta^{(m)}$ which leaves $ \Normal(\theta^{(m)} \mid \zeros, \sigma_\theta^2 \Identity) I(\theta_1^{(m)} < \ldots, < \theta_{K-1}^{(m)})
\cdot \prod_{u : C_u = m}
\prod_{i \in \sI_u}
\log [\Phi(\theta_{Z_{iu}}^{(m)} - \mu_{iu}) - \Phi(\theta^{(m)}_{Z_{iu}-1} - \mu_{iu})],
$ invariant.
In our illustrations, we use slice sampling [@slicesampling] to do updates 9–11. The chain is initialized by simulation from the prior with $a = 1$. The only non-trivial step is constructing an update for the $\theta^{(m)}$’s. We use a modification of the approach outlined in @albert1997bayesian, which uses a Laplace approximation to construct a proposal for the full-conditional of the parameters $\delta^{(m)}_1 = \theta^{(m)}_1$ and $\delta^{(m)}_k = \log(\theta^{(m)}_k - \theta^{(m)}_{k-1})$ for $k = 2,\ldots,K-1$. To alleviate computational burden, the proposal is updated every $50^{\text{th}}$ iteration.
Identifiability
===============
We conduct a brief simulation experiment to illustrate that model proposed in Section \[sec:description\] is capable of (i) identifying the correct number of latent factors $L$ and (ii) capable of accruing evidence about the individual $\alpha_u$’s and $\beta_i$’s. We simulate from the data model, response model, random effect model, spatial process model, and parameter model described in Section \[sec:description\] with the dimension of the latent factor set to $L = 4$. For simplicity, we omit the item-specific covariates given by $x_i$. For the random effects model we set $\sigma_\alpha = 2, \sigma_\beta = 5$, and $\sigma_b = 3$. We set $\eta \sim {\operatorname{Gau}}(0, 3{\operatorname{I}})$ and $r = 20$, with the basis functions $\psi_j(s)$ given by Gaussian radial basis functions with knots at placed uniformly at random throughout the spatial domain. For the parameter model, we used $M = 3$ rubrics with $\omega = (1/3, 1/3, 1/3)$ and selected $\theta^{(m)}$ in the manner described in the simulation of Section \[sec:simulation\]. We set $u = 200$ and $i = 200$, and select user/item pairs for inclusion by sampling uniformly at random, with a total of 3981 ratings.
After simulating this data, we fit the multi-rubric model using the correct $\Psi$ using the default prior described in Section \[sec:description\] with the correct choice of basis functions $\psi_j(s)$, with $M = 10$ and $\omega \sim {\operatorname{Dirichlet}}(1/10, \ldots, 1/10)$.
We first assess whether the model is capable of recovering the true number of latent factors. We fit the model for $L \in \{1, \ldots, 7\}$ and evaluated each value of $L$ by held-out log-likelihood criteria after splitting the data randomly into training and testing sets. Results are given in Figure \[fig:loss\]. We see that the model with the highest held-out log-likelihood corresponds to $L = 4$, the correct number of latent factors.
Next, we assess whether the model is capable of learning the individual $\alpha_u$’s and $\beta_i$’s. First, we note that for any orthonormal matrix $O$ with $O{^{\top}}O = {\operatorname{I}}$ we have $$\begin{aligned}
\alpha_u{^{\top}}\beta_i = (O\alpha_u){^{\top}}(O\beta_u).\end{aligned}$$ Moreover, $\alpha_u$ and $O\alpha_u$ are equal in distribution (as are $\beta_i$ and $O\beta_i$), so the prior does not provide any additional identification. Consequently, we can only expect to identify $\alpha_u$ and $\beta_i$ up-to an arbitrary rotation. While it is possible to impose constraints on the $\alpha_u$ and $\beta_i$ — say, by fixing $\alpha_1, \ldots, \alpha_L$ to know values — this is undesirable because it breaks the symmetry of the prior. In view of this, it is standard in the recommender systems literature to not invoke any constraints on the prior .
With these points in mind, Figure \[fig:learning\] displays the prior distribution of $\zeta_i = \beta_{i1} / \sigma_{\beta}$ along with the posterior distribution of $\zeta_i$’s for 9 randomly selected items. We see that, while the overall distribution of the $\zeta_i$’s is in agreement with the prior when considered as a group, for the individual $\zeta_i$’s the prior and posterior differ considerably. This indicates that the model is capable of detecting differences in the $\beta_i$’s across items.
[^1]: Department of Statistics, Florida State University
[^2]: Email: `[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
---
\
[**J.N. ALONSO ÁLVAREZ$^{1}$, J.M. FERNÁNDEZ VILABOA$^{2}$, R. GONZÁLEZ RODRÍGUEZ$^{3}$**]{}
\
$^{1}$ Departamento de Matemáticas, Universidad de Vigo, Campus Universitario Lagoas-Marcosende, E-36280 Vigo, Spain (e-mail: [email protected]) \
$^{2}$ Departamento de Álxebra, Universidad de Santiago de Compostela. E-15771 Santiago de Compostela, Spain (e-mail: [email protected]) \
$^{3}$ Departamento de Matemática Aplicada II, Universidad de Vigo, Campus Universitario Lagoas-Marcosende, E-36310 Vigo, Spain (e-mail: [email protected]) \
[**Abstract**]{} In this paper we introduce the notion of weak non-asssociative Doi-Hopf module and give the Fundamental Theorem of Hopf modules in this setting. Also we prove that there exists a categorical equivalence that admits as particular instances the ones constructed in the literature for Hopf algebras, weak Hopf algebras, Hopf quasigroups, and weak Hopf quasigroups.
[**Keywords.**]{} Hopf algebra, Weak Hopf algebra, Hopf quasigroup, Weak Hopf quasigroup, Doi-Hopf module, Fundamental Theorem.
[**MSC 2010:**]{} 18D10, 16T05, 17A30, 20N05.
introduction
============
Let ${\Bbb F}$ be a field and ${\mathcal C}={\Bbb F}-Vect$. Let $H$ be a Hopf algebra in ${\mathcal C}$ and let $B$ be a right $H$-comodule algebra with coaction $\rho_{B}:B\rightarrow B{\otimes}H$, $\rho_{B}(b)=b_{(0)}{\otimes}b_{(1)}$, where the unadorned tensor product is the tensor product over ${\Bbb F}$ and for $\rho_{B}(b)$ we used the Sweedler notation. In [@Doi83], Doi introduced the notion of $(H,B)$-Hopf module, as a generalization of the classical notion of Hopf module, defined by Larson and Sweedler in [@Lar-Sweed], in the following way: Let $M$ be a right $B$-module and a right $H$-comodule. If, for all $m\in M$ and $b\in B$, we write $m.b$ for the action and $\rho_{M}(m)=m_{[0]}{\otimes}m_{[1]}$ for the coaction, we will say that $M$ is an $(H,B)$-Hopf module if the equality $$\rho_{M}(m.b)=m_{[0]}.b_{(1)}{\otimes}m_{[1]}b_{(2)}$$ holds, where $m_{[1]}b_{(2)}$ is the product in $H$ of $m_{[1]}$ and $b_{(2)}$. A morphism between two $(H,B)$-Hopf modules is an ${\Bbb F}$-linear map that is $B$-linear and $H$-colinear. Hopf modules and morphisms of Hopf modules constitute the category of $(H,B)$-Hopf modules denoted by ${\mathcal M}^{H}_{B}$. If there exists a right $H$-comodule map $h:H\rightarrow B$ which is an algebra map, and $M^{co H}=\{m\in M \;| \; \rho_{M}(m)=m{\otimes}1_{H}\}$, $B^{co H}=\{b\in B \;| \; \rho_{B}(b)=b{\otimes}1_{H}\}$ are the subobjects of coinvariants, for any $m\in M^{co H}$ and $b\in B^{co H}$ we have that $m.b\in M^{co H}$ and then $M^{co H}$ is a right $B^{co H}$-module. Using this property Doi proved in Theorem 3 of [@Doi83] that $M$ is isomorphic to $M^{co H}{\otimes}_{B^{co H}} B$ as $(H,B)$-Hopf modules. Moreover, if $N$ is a right $B^{co H}$-module, the tensor product $N{\otimes}_{B^{co H}} B$, with the action and coaction induced by the product of $B$ and the coproduct of $H$, is an $(H,B)$-Hopf module. This construction is functorial and then we have a functor, called the induction functor, $F=-{\otimes}_{B^{co H}} B:{\mathcal C}_{B^{co H}}\rightarrow {\mathcal M}^{H}_{B}$. Also, for all $M\in {\mathcal M}^{H}_{B}$, the construction of $M^{co H}$ is functorial and we have a functor of coinvariants $G=(\;\;)^{co H}: {\mathcal M}^{H}_{B}\rightarrow {\mathcal C}_{B^{co H}}$ such that $F\dashv G$. Moreover, $F$ and $G$ induce a categorical equivalence between ${\mathcal M}^{H}_{B}$ and the category of right $B^{co H}$-modules. This categorical equivalence was called by Doi and Takeuchi in [@Doi-Take], the strong structure theorem for ${\mathcal M}^{H}_{B}$, and, for $B=H$ and $h=id_{H}$, contains as a particular instance the equivalence derived of the Fundamental Theorem of Hopf modules proved by Larson and Sweedler (see [@Lar-Sweed], [@Sweedler]).
The categorical equivalence of the previous paragraph remains valid for weak Hopf algebras. For a weak Hopf algebra $H$, Böhm introduced in [@bohm] the category of Hopf modules, denoted by ${\mathcal M}^{H}_{H}$, in the same way that in the Hopf algebra setting. If $M\in {\mathcal M}^{H}_{H}$, the subobject of coinvariants is defined by $M^{co H}=\{m\in M \;| \; \rho_{M}(m)=m_{[0]}{\otimes}\Pi_{H}^{L}(m_{[1]})\}$, where $\Pi_{H}^{L}$ is the target morphism associated to $H$. In [@bohm] we can find the weak version of the Fundamental Theorem of Hopf modules, i.e.: For all Hopf module $M$, $M^{co H}{\otimes}_{H_{L}} H$ is isomorphic to $M$ as Hopf modules, where $H_{L}$ is the image of $\Pi_{H}^{L}$. Moreover, if ${\mathcal C}_{H_{L}}$ is the category of right $H_{L}$-modules, there exist two functors $F=-{\otimes}_{H_{L}} H:{\mathcal C}_{H_{L}}\rightarrow {\mathcal M}^{H}_{H}$ and $G=(\;\;)^{co H}: {\mathcal M}^{H}_{H}\rightarrow {\mathcal C}_{H_{L}}$ such that $F$ is left adjoint of $G$ and they induce a pair of inverse equivalences. Therefore, in the weak setting, ${\mathcal M}^{H}_{H}$ is equivalent to ${\mathcal C}_{H_{L}}$. In this case, the following property is a relevant fact for subsequent generalizations: there is an isomorphism of Hopf modules between the tensor product $M^{co H}{\otimes}_{H_{L}} H$ and $M^{co H}\times H$, where $M^{co H}\times H$ is the image of a suitable idempotent morphism $\nabla_{M}:M^{co H}\otimes H\rightarrow M^{co H}\otimes H$. Later, in [@bohm2], Böhm introduced the notion of weak Doi-Hopf module (or weak $(H,B)$-Doi-Hopf module), associated to a weak Hopf algebra $H$ and a right $H$-comodule algebra $B$, and the category of weak Doi-Hopf modules denoted as in the non-weak setting by ${\mathcal M}^{H}_{B}$. In 2004, Zhang and Zhu proved that for any weak Doi-Hopf module $M$ (also called by these authors weak $(H,B)$-Doi-Hopf module), if there exists a right $H$-comodule map $h:H\rightarrow B$ which is an algebra map, the objects $M$ and $M^{co H}{\otimes}_{B^{co H}} B$ are isomorphic as $(H,B)$-Doi-Hopf modules. In this case $B^{co H}=\{b\in B \;| \; \rho_{B}(b)=b_{(0)}{\otimes}\Pi_{H}^{L}(b_{(1)})\}$ and, if $B=H$ and $h=id_{H}$, they recover the isomorphism constructed by Böhm in [@bohm]. As in the Hopf setting, it is possible to construct the induction functor $F=-{\otimes}_{B^{co H}} B:{\mathcal C}_{B^{co H}}\rightarrow {\mathcal M}^{H}_{B}$ and the functor of coinvariants $G=(\;\;)^{co H}: {\mathcal M}^{H}_{B}\rightarrow {\mathcal C}_{B^{co H}}$. These functors satisfy that $F\dashv G$ and $F$ and $G$ is a pair of inverse equivalences. Therefore, ${\mathcal M}^{H}_{B}$ is equivalent to the category of right $B^{co H}$-modules (see [@Hanna]).
In the two previous paragraphs we wrote about categorical equivalences for categories of Hopf modules connected to associative algebraic structures like Hopf algebras and weak Hopf algebras. An interesting generalization of Hopf algebras are nonassociative Hopf algebras. As in the quasi-Hopf setting, nonassociative Hopf algebras are not associative, but the lack of this property is compensated in this case by some axioms involving the division operation. The notion of nonassociative Hopf algebra in a category of vector spaces was introduced by Pérez- Izquierdo [@PI2] with the aim of to construct the universal enveloping algebra for Sabinin algebras, prove a Poincaré-Birkhoff-Witt Theorem for Sabinin algebras and give a nonassociative version of the Milnor-Moore theorem. Later, Klim and Majid [@Majidesfera], in order to understand the structure and relevant properties of the algebraic 7-sphere, introduced the notion of Hopf quasigroup. Hopf quasigroups are examples of nonassociative Hopf algebras and in recent years, interesting research about its specific structure and its dual has been developed ([@Brz], [@Klim], [@Brz2], [@Brz3], [@ZH], [@ZX], [@FT], [@our1], [@our2]). Moreover, nonassociative Hopf algebras arise naturally related with other structures in various nonassociative contexts like, for example quantum quasigroups in the sense of Smith ([@S1], [@S2], [@S3], [@S4]). Nonassociative Hopf algebras include the example of an enveloping algebra $U(L)$ of a Malcev algebra (see [@PIS], [@Majidesfera], [@TV]) as well as the notion of the loop algebra RL of a loop L (see [@BMP-I12], [@MPIS14]). Then, nonassociative Hopf algebras unify Moufang loops and Malcev algebras, and, more generally, formal loops and Sabinin algebras, in the same way that Hopf algebras unify groups and Lie algebras.
For a of Hopf quasigroup in the sense of [@Majidesfera], Brzeziński defined in [@Brz] the notion of Hopf module obtaining a categorical equivalence as in the associative context. In this case, the main difference appears in the definition of the category of Hopf modules ${\mathcal M}^{H}_{H}$. Firstly, because the notion of Hopf module reflects the non-associativity of the product defined on $H$. Secondly, the morphisms are $H$-quasilinear and $H$-colinear (see Definition 3.4 of [@Brz]). In Lemma 3.5 of [@Brz], we can find that, if $M\in {\mathcal M}^{H}_{H}$ and $M^{co H}$ is defined like in the Hopf algebra setting, $M$ is isomorphic to $M^{co H}{\otimes}H$ as Hopf modules. Therefore the Fundamental Theorem of Hopf modules also holds for Hopf quasigroups. Moreover, there exist two functors $F=-{\otimes}H:{\mathcal C}\rightarrow {\mathcal M}^{H}_{H}$ and $G=(\;\;)^{co H}: {\mathcal M}^{H}_{H}\rightarrow {\mathcal C}$ such that $F\dashv G$, and they induce a pair of inverse equivalences. Thus, as it occurs in the Hopf algebra ambit, ${\mathcal M}^{H}_{H}$ is equivalent to the category of ${\Bbb F}$-vector spaces.
Hopf quasigroups admit a generalization to the weak seetting. The new notion, called weak Hopf quasigroup, was introduced in [@Asian] in a monoidal context and a family of non trivial examples can be obtained by working with bigroupoids, i.e., bicategories where every $1$-cell is an equivalence and every $2$-cell is an isomorphism (see Example 2.3 of [@Asian]). In [@MJM] we described these algebraic objects in terms of fusion morphisms and in [@Asian], for a weak Hopf quasigroup $H$ in a braided monoidal category ${\mathcal C}$ with tensor product ${\otimes}$, using the ideas proposed by Brzeziński for Hopf quasigroups, we introduce the notion of Hopf module and the category of Hopf modules ${\mathcal M}^{H}_{H}$. In this case, if we define $M^{co H}$ in the same way that in the weak Hopf algebra setting, we obtain the weak nonassociative version of the Fundamental Theorem of Hopf modules in the following way: every Hopf module $M$ is isomorphic to $M^{co H}\times H$ as Hopf modules, where $M^{co H}\times H$ is the image of the same idempotent $\nabla_{M}$ used for Hopf modules associated to a weak Hopf algebra. Moreover, in [@JPAA] we proved that $H_{L}$, the image of the target morphism, is a monoid, and then it is possible to take into consideration the category ${\mathcal C}_{H_{L}}$ to construct the tensor product $M^{co H}{\otimes}_{H_{L}} H$, and, if the functor $-{\otimes}H$ preserves coequalizers, to endow this object with a Hopf module structure. Unfortunately, unlike the case of weak Hopf algebras, it is not possible to assure in general that $M^{co H}{\otimes}_{H_{L}} H$ is isomorphic to $M^{co H}\times H$. In order to find sufficient conditions under which these objects are isomorphic in ${\mathcal M}^{H}_{H}$, we introduce in [@Strong] the category of strong Hopf modules, denoted by ${\mathcal SM}^{H}_{H}$ and obtain that there exist two functors $F=-{\otimes}_{H_{L}} H:{\mathcal C}_{H_{L}}\rightarrow {\mathcal SM}^{H}_{H}$ and $G=(\;\;)^{co H}: {\mathcal SM}^{H}_{H}\rightarrow {\mathcal C}_{H_{L}}$ such that $F$ is left adjoint of $G$ and they induce a pair of inverse equivalences. In the Hopf quasigroup setting every Hopf module is strong, and then our results are the ones proved by Brzeziński in [@Brz]. The same happens in the weak Hopf case and then we generalize the theorem proved by Böhm, Nill and Szlachányi in [@bohm].
Let ${\mathcal C}$ be a braided monoidal category with tensor product ${\otimes}$. Then for a weak Hopf quasigroup $H$ in ${\mathcal C}$ and a right $H$-comodule magma $B$ (see [@JPAA] for the definition), a question naturally arises: Is it possible to define a general category of $(H,B)$-Hopf modules and to prove a general theorem that permit to recover as particular instances the categorical equivalences cited in the previous paragraphs? The main contribution of this paper is to give a positive answer to this question.
Now, we describe the paper in detail. After this introduction, for a weak Hopf quasigroup $H$ and a right $H$-comodule magma $B$ in a strict braided monoidal category $\mathcal C$ where every idempotent morphism splits, in the second section we introduce the notion of anchor morphism $h:H\rightarrow B$ as an $H$-comodule morphism such that it is a morphism of unital magmas satisfying two suitable conditions. For an anchor morphism $h$, in Definition \[H-D-mod\], we define the notion of strong $(H,B,h)$-Hopf module and prove some properties of these modules. We also find the condition under which the subobject of coinvariants of $B$, defined as in the weak Hopf algebra context, i.e., $B^{co H}=\{b\in B \;| \; \rho_{B}(b)=b_{(0)}{\otimes}\Pi_{H}^{L}(b_{(1)})\}$, is a monoid, and construct the new category of strong $(H,B,h)$-Hopf modules, denoted by ${\mathcal S M}^{H}_{B}$. Moreover, if the category ${\mathcal C}$ admits coequalizers and the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers, we prove in Theorem \[main0\] that the Fundamental Theorem of Hopf Modules holds. In other words, for any strong $(H,B,h)$-Hopf module the objects $M$ and $M^{co H}{\otimes}_{B^{co H}} B$ are isomorphic in ${\mathcal S M}^{H}_{B}$. This result admits as particular instances the results with the same name cited in the previous paragraphs for associative and nonassociative (weak) Hopf structures. Finally, in the last section, we define the induction functor $F=-{\otimes}_{B^{co H}} B:{\mathcal C}_{B^{co H}}\rightarrow {\mathcal S M}^{H}_{B}$ and the functor of coinvariants $G=(\;\;)^{co H}: {\mathcal S M}^{H}_{B}\rightarrow {\mathcal C}_{B^{co H}}$, proving that $F\dashv G$. Also, $F$ and $G$ is a pair of inverse equivalences and, therefore, ${\mathcal S M}^{H}_{B}$ is equivalent to the category of right $B^{co H}$-modules.
Throughout this paper $\mathcal C$ denotes a strict braided monoidal category with tensor product ${\otimes}$, unit object $K$ and braid $c$. Without loss of generality, by the coherence theorems, we can assume the monoidal structure of ${\mathcal C}$ strict. Then, in this paper, we omit explicitly the associativity and unit constraints. For each object $M$ in $ {\mathcal C}$, we denote the identity morphism by $id_{M}:M\rightarrow M$ and, for simplicity of notation, given objects $M$, $N$ and $P$ in ${\mathcal C}$ and a morphism $f:M\rightarrow N$, we write $P{\otimes}f$ for $id_{P}{\otimes}f$ and $f {\otimes}P$ for $f{\otimes}id_{P}$. We also assume that every idempotent morphism in $ {\mathcal C}$ splits, i.e., if $\nabla:Y\rightarrow Y$ is such that $\nabla=\nabla{\circ}\nabla$, there exist an object $Z$, called the image of $p$, and morphisms $i:Z\rightarrow Y$ and $p:Y\rightarrow Z$ such that $\nabla=i{\circ}p$ and $p{\circ}i =id_{Z}$. The morphisms $p$ and $i$ will be called a factorization of $q$. Note that $Z$, $p$ and $i$ are unique up to isomorphism. The categories satisfying this property constitute a broad class that includes, among others, the categories with epi-monic decomposition for morphisms and categories with equalizers or coequalizers. For example, complete bornological spaces is a symmetric monoidal closed category that is not abelian, but it does have coequalizers (see [@Meyer]). On the other hand, let [**Hilb**]{} be the category whose objects are complex Hilbert spaces and whose morphisms are the continuous linear maps. Then [**Hilb**]{} is not an abelian and closed category but it is a symmetric monoidal category (see [@Kad]) with coequalizers.
As for prerequisites, the reader is expected to be familiar with the notions of (co)unital (co)magma, (co)monoid, and morphism of (co)unital (co)magmas. By a unital magma in ${\mathcal C}$ we understand a triple $A=(A, \eta_{A}, \mu_{A})$ where $A$ is an object in ${\mathcal C}$ and $\eta_{A}:K\rightarrow A$ (unit), $\mu_{A}:A{\otimes}A \rightarrow A$ (product) are morphisms in ${\mathcal C}$ such that $\mu_{A}{\circ}(A{\otimes}\eta_{A})=id_{A}=\mu_{A}{\circ}(\eta_{A}{\otimes}A)$. If $\mu_{A}$ is associative, that is, $\mu_{A}{\circ}(A{\otimes}\mu_{A})=\mu_{A}{\circ}(\mu_{A}{\otimes}A)$, the unital magma will be called a monoid in ${\mathcal C}$. Given two unital magmas (monoids) $A= (A, \eta_{A}, \mu_{A})$ and $B=(B, \eta_{B}, \mu_{B})$, $f:A\rightarrow B$ is a morphism of unital magmas (monoids) if $\mu_{B}{\circ}(f{\otimes}f)=f{\circ}\mu_{A}$ and $ f{\circ}\eta_{A}= \eta_{B}$.
By duality, a counital comagma in ${\mathcal C}$ is a triple ${D} = (D, \varepsilon_{D}, \delta_{D})$ where $D$ is an object in ${\mathcal C}$ and $\varepsilon_{D}: D\rightarrow K$ (counit), $\delta_{D}:D\rightarrow D{\otimes}D$ (coproduct) are morphisms in ${\mathcal C}$ such that $(\varepsilon_{D}{\otimes}D){\circ}\delta_{D}= id_{D}=(D{\otimes}\varepsilon_{D}){\circ}\delta_{D}$. If $\delta_{D}$ is coassociative, that is, $(\delta_{D}{\otimes}D){\circ}\delta_{D}= (D{\otimes}\delta_{D}){\circ}\delta_{D}$, the counital comagma will be called a comonoid. If ${D} = (D, \varepsilon_{D}, \delta_{D})$ and ${ E} = (E, \varepsilon_{E}, \delta_{E})$ are counital comagmas (comonoids), $f:D\rightarrow E$ is a morphism of counital comagmas (comonoids) if $(f{\otimes}f){\circ}\delta_{D} =\delta_{E}{\circ}f$ and $\varepsilon_{E}{\circ}f =\varepsilon_{D}.$
If $A$, $B$ are unital magmas (monoids) in ${\mathcal C}$, the object $A{\otimes}B$ is a unital magma (monoid) in ${\mathcal C}$ where $\eta_{A{\otimes}B}=\eta_{A}{\otimes}\eta_{B}$ and $\mu_{A{\otimes}B}=(\mu_{A}{\otimes}\mu_{B}){\circ}(A{\otimes}c_{B,A}{\otimes}B).$ In a dual way, if $D$, $E$ are counital comagmas (comonoids) in ${\mathcal C}$, $D{\otimes}E$ is a counital comagma (comonoid) in ${\mathcal C}$ where $\varepsilon_{D{\otimes}E}=\varepsilon_{D}{\otimes}\varepsilon_{E}$ and $\delta_{D{\otimes}E}=(D{\otimes}c_{D,E}{\otimes}E){\circ}( \delta_{D}{\otimes}\delta_{E}).$
Let $A$ be a monoid. The pair $(M,\phi_{M})$ is a right $A$-module if $M$ is an object in ${\mathcal C}$ and $\phi_{M}:M\otimes A\rightarrow M$ is a morphism in ${\mathcal C}$ satisfying $\phi_{M}\circ(M\otimes
\eta_{A})=id_{M}$, $\phi_{M}\circ (\phi_{M}\otimes A)=\phi_{M}\circ
(M\otimes \mu_{A})$. Given two right ${A}$-modules $(M,\phi_{M})$ and $(N,\phi_{N})$, $f:M\rightarrow N$ is a morphism of right ${A}$-modules if $\phi_{N}\circ (f\otimes A)=f\circ \phi_{M}$. If $D$ is a comonoid, the pair $(M,\rho_{M})$ is a right $D$-comodule if $M$ is an object in ${\mathcal C}$ and $\rho_{M}:M\rightarrow M{\otimes}D$ is a morphism in ${\mathcal C}$ satisfying $(M{\otimes}\varepsilon_{D}){\circ}\rho_{M}=id_{M}$, $(\rho_{M}{\otimes}D){\circ}\rho_{M}=(M{\otimes}\delta_{D}){\circ}\rho_{M}$. Given two right ${D}$-comodules $(M,\rho_{M})$ and $(N,\rho_{N})$, $f:M\rightarrow N$ is a morphism of right ${D}$-comodules if $(f\otimes D){\circ}\rho_{M}=\rho_{N}{\circ}f$.
Finally, if $D$ is a comagma and $A$ a magma, given two morphisms $f,g:D\rightarrow A$ we will denote by $f\ast g$ its convolution product in ${\mathcal C}$, that is $$f\ast g=\mu_{A}{\circ}(f{\otimes}g){\circ}\delta_{D}.$$
Doi-Hopf modules for weak Hopf quasigroups
==========================================
We begin this section by recalling the notion of weak Hopf quasigroup in a braided monoidal category introduced in [@Asian]. In this reference the interested reader can find an exhaustive list of properties of weak Hopf quasigroups, that we will need along the paper.
\[W-H-quasi\]
A weak Hopf quasigroup $H$ in ${\mathcal C}$ is a unital magma $(H, \eta_H, \mu_H)$ and a comonoid $(H,\varepsilon_H, \delta_H)$ such that the following axioms hold:
- $\delta_{H}{\circ}\mu_{H}=(\mu_{H}{\otimes}\mu_{H}){\circ}\delta_{H{\otimes}H}.$
- $\varepsilon_{H}{\circ}\mu_{H}{\circ}(\mu_{H}{\otimes}H)=\varepsilon_{H}{\circ}\mu_{H}{\circ}(H{\otimes}\mu_{H})$
- $= ((\varepsilon_{H}{\circ}\mu_{H}){\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(H{\otimes}\delta_{H}{\otimes}H)$
- $=((\varepsilon_{H}{\circ}\mu_{H}){\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(H{\otimes}(c_{H,H}^{-1}{\circ}\delta_{H}){\otimes}H).$
- $(\delta_{H}{\otimes}H){\circ}\delta_{H}{\circ}\eta_{H}=(H{\otimes}\mu_{H}{\otimes}H){\circ}((\delta_{H}{\circ}\eta_{H}) {\otimes}(\delta_{H}{\circ}\eta_{H}))$
- $=(H{\otimes}(\mu_{H}{\circ}c_{H,H}^{-1}){\otimes}H){\circ}((\delta_{H}{\circ}\eta_{H}) {\otimes}(\delta_{H}{\circ}\eta_{H})).$
- There exists $\lambda_{H}:H\rightarrow H$ in ${\mathcal C}$ (called the antipode of $H$) such that, if we denote the morphisms $id_{H}\ast \lambda_{H}$ by $\Pi_{H}^{L}$ (target morphism) and $\lambda_{H}\ast id_{H}$ by $\Pi_{H}^{R}$ (source morphism),
- $\Pi_{H}^{L}=((\varepsilon_{H}{\circ}\mu_{H}){\otimes}H){\circ}(H{\otimes}c_{H,H}){\circ}((\delta_{H}{\circ}\eta_{H}){\otimes}H).$
- $\Pi_{H}^{R}=(H{\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(c_{H,H}{\otimes}H){\circ}(H{\otimes}(\delta_{H}{\circ}\eta_{H})).$
- $\lambda_{H}\ast \Pi_{H}^{L}=\Pi_{H}^{R}\ast \lambda_{H}= \lambda_{H}.$
- $\mu_H{\circ}(\lambda_H{\otimes}\mu_H){\circ}(\delta_H{\otimes}H)=\mu_{H}{\circ}(\Pi_{H}^{R}{\otimes}H).$
- $\mu_H{\circ}(H{\otimes}\mu_H){\circ}(H{\otimes}\lambda_H{\otimes}H){\circ}(\delta_H{\otimes}H)=\mu_{H}{\circ}(\Pi_{H}^{L}{\otimes}H).$
- $\mu_H{\circ}(\mu_H{\otimes}\lambda_H){\circ}(H{\otimes}\delta_H)=\mu_{H}{\circ}(H{\otimes}\Pi_{H}^{L}).$
- $\mu_H{\circ}(\mu_H{\otimes}H){\circ}(H{\otimes}\lambda_H{\otimes}H){\circ}(H{\otimes}\delta_H)=\mu_{H}{\circ}(H{\otimes}\Pi_{H}^{R}).$
Note that, if in the previous definition the triple $(H, \eta_H, \mu_H)$ is a monoid, we obtain the notion of weak Hopf algebra in a symmetric monoidal category. Then, if ${\mathcal C}$ is the category of vector spaces over a field ${\Bbb F}$, we have the original definition of weak Hopf algebra introduced by Böhm, Nill and Szlachányi in [@bohm]. On the other hand, under these conditions, if $\varepsilon_H$ and $\delta_H$ are morphisms of unital magmas (equivalently, $\eta_{H}$, $\mu_{H}$ are morphisms of counital comagmas), $\Pi_{H}^{L}=\Pi_{H}^{R}=\eta_{H}{\otimes}\varepsilon_{H}$. As a consequence, conditions (a2), (a3), (a4-1)-(a4-3) trivialize, and we get the notion of Hopf quasigroup defined by Klim and Majid in [@Majidesfera]. More concretely, a Hopf quasigroup $H$ in ${\mathcal C}$ is a unital magma $(H,\eta_H,\mu_H)$ and a comonoid $(H,\varepsilon_H,\delta_H)$ satisfying that $\varepsilon_H$ and $\delta_H$ are morphisms of unital magmas (equivalently, $\eta_H$ and $\mu_H$ are morphisms of counital comagmas), and such that there exists a morphism $\lambda_{H}:H\rightarrow H$ in ${\mathcal C}$, called the antipode of $H$, for which $$\label{leftHqg}
\mu_H\circ (\lambda_H{\otimes}\mu_H)\circ (\delta_H{\otimes}H)=
\varepsilon_H{\otimes}H=
\mu_H\circ (H{\otimes}\mu_H)\circ (H{\otimes}\lambda_H{\otimes}H)\circ (\delta_H{\otimes}H)$$ and $$\label{rightHqg}
\mu_H\circ (\mu_H{\otimes}H)\circ (H{\otimes}\lambda_H{\otimes}H)\circ (H{\otimes}\delta_H)=
H{\otimes}\varepsilon_H=
\mu_H\circ(\mu_H{\otimes}\lambda_H)\circ (H{\otimes}\delta_H)$$ hold. Then, as a consequence, we have (a1) and the following identities $$\label{e-m-d-eps}
\varepsilon_{H}{\circ}\eta_{H}=id_{K},\;\;
\varepsilon_{H}{\circ}\mu_{H}=\varepsilon_{H}{\otimes}\varepsilon_{H},\;\;
\delta_{H}{\circ}\eta_{H}=\eta_{H}{\otimes}\eta_{H}.$$
By Proposition 3.2 of [@Asian] we know that the antipode of a weak Hopf quasigroup is unique, and satisfies that $\lambda_{H}{\circ}\eta_{H}=\eta_{H}$, $\varepsilon_{H}{\circ}\lambda_{H}=\varepsilon_{H}$. Also, by Theorem 3.19 of [@Asian], we have that it is antimultiplicative and anticomultiplicative. Moreover, if we define the morphisms $\overline{\Pi}_{H}^{L}$ and $\overline{\Pi}_{H}^{R}$ by $$\overline{\Pi}_{H}^{L}=(H{\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}((\delta_{H}{\circ}\eta_{H}){\otimes}H),\;\;\;\overline{\Pi}_{H}^{R}=((\varepsilon_{H}{\circ}\mu_{H}){\otimes}H){\circ}(H{\otimes}(\delta_{H}{\circ}\eta_{H})),$$ we proved in Proposition 3.4 of [@Asian], that $\Pi_{H}^{L}$, $\Pi_{H}^{R}$, $\overline{\Pi}_{H}^{L}$ and $\overline{\Pi}_{H}^{R}$ are idempotent.
\[monoid-hl\] Let $H$ be a weak Hopf quasigroup and $\Pi\in\{\Pi_{H}^{L},\Pi_{H}^{R},\overline{\Pi}_{H}^{L}, \overline{\Pi}_{H}^{R}\}$ . The following identities hold: $$\label{monoid-hl-1}
\mu_{H}{\circ}((\mu_{H}{\circ}(\Pi{\otimes}H)){\otimes}H)=\mu_{H}{\circ}(\Pi{\otimes}\mu_{H}),$$ $$\label{monoid-hl-2}
\mu_{H}{\circ}(H{\otimes}(\mu_{H}{\circ}(\Pi{\otimes}H)))=\mu_{H}{\circ}((\mu_{H}{\circ}(H{\otimes}\Pi)){\otimes}H),$$ $$\label{monoid-hl-3}
\mu_{H}{\circ}(H{\otimes}(\mu_{H}{\circ}(H{\otimes}\Pi)))=\mu_{H}{\circ}(\mu_{H}{\otimes}\Pi).$$
The proof for $\Pi_{H}^{L}$ is in Proposition 2.4 of [@JPAA] and in a similar way we can prove the result for $\Pi_{H}^{R}$ (see also Proposition 2.3 of [@MJM]). The equalities for $\overline{\Pi}_{H}^{L}$ and $\overline{\Pi}_{H}^{R}$ follow from Proposition 3.11 of [@Asian].
If $H_{L}$ is the image of the idempotent morphism $\Pi_{H}^{L}$, and $p_{L}:H\rightarrow H_{L}$, $i_{L}:H_{L}\rightarrow H$ are the morphisms such that $\Pi_{H}^{L}=i_{L}{\circ}p_{L}$ and $p_{L}{\circ}i_{L}=id_{H_{L}}$, by Proposition 3.13 of [@Asian], $i_{L}$ is the equalizer of $\delta_{H}$ and $(H{\otimes}\Pi_{H}^{L}) {\circ}\delta_{H}$ and $p_{L}$ is the coequalizer of $\mu_{H}$ and $\mu_{H}{\circ}(H{\otimes}\Pi_{H}^{L})$. Then the triple $(H_{L}, \varepsilon_{H_{L}}=\varepsilon_{H}{\circ}i_{L}, \delta_{H}=(p_{L}{\otimes}p_{L}){\circ}\delta_{H}{\circ}i_{L})$ is a comonoid in ${\mathcal C}$, and as a consequence of Lemma \[monoid-hl\], $(H_{L},
\eta_{H_{L}}=p_{L}{\circ}\eta_{H}, \mu_{H_{L}}=p_{L}{\circ}\mu_{H}{\circ}(i_{L}{\otimes}i_{L}))$ is a monoid in ${\mathcal C}$. Following Remark 3.15 of [@Asian], we have similar results for the image of the idempotent morphism $\Pi_{H}^{R}$ denoted by $H_{R}$.
\[H-comodmag\]
Let $H$ be a weak Hopf quasigroup and let $B$ be a unital magma, which is also a right $H$-comodule with coaction $\rho_{B}:B\rightarrow B{\otimes}H$ such that $$\label{chmagma}
\mu_{B{\otimes}H}{\circ}(\rho_{B}{\otimes}\rho_{B})=\rho_{B}{\circ}\mu_{B}.$$ We will say that $(B,\rho_{B})$ is a right $H$-comodule magma if any of the following equivalent conditions hold:
- $(\rho_{B}{\otimes}H){\circ}\rho_{B}{\circ}\eta_{B}=(B{\otimes}(\mu_{H}{\circ}c_{H,H}^{-1}){\otimes}H){\circ}((\rho_{B}{\circ}\eta_{B}){\otimes}(\delta_{H}{\circ}\eta_{H})). $
- $(\rho_{B}{\otimes}H){\circ}\rho_{B}{\circ}\eta_{B}=(B{\otimes}\mu_{H}{\otimes}H){\circ}((\rho_{B}{\circ}\eta_{B}){\otimes}(\delta_{H}{\circ}\eta_{H})). $
- $(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}=(\mu_{B}{\otimes}H){\circ}(B{\otimes}(\rho_{B}{\circ}\eta_{B})),$
- $(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}= ((\mu_{B}{\circ}c_{B,B}^{-1}){\otimes}H){\circ}(B{\otimes}(\rho_{B}{\circ}\eta_{B})).$
- $(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}\eta_{B}=\rho_{B}{\circ}\eta_{B}.$
- $ (B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}{\circ}\eta_{B}=\rho_{B}{\circ}\eta_{B}.$
This definition is similar to the notion of right $H$-comodule monoid in the weak Hopf algebra setting and the proof for the equivalence of (b1)-(b6) also follows in a similar way.
Note that, if $H$ is a Hopf quasigroup and $B$ is a unital magma which is also a right $H$-comodule with coaction $\rho_{B}:B\rightarrow B{\otimes}H$, we will say that $(B,\rho_{B})$ is a right $H$-comodule magma if it satisfies (\[chmagma\]) and $\rho_{B}{\circ}\eta_{B}=\eta_{H}{\otimes}\eta_{B}$. In this case (b1)-(b6) trivialize.
\[ex-hcm\]
1\. If $H$ is a (weak) Hopf quasigroup, $(H, \delta_H)$ is a right $H$-comodule magma.
2\. Let $H$ be a cocommutative weak Hopf quasigrop and assume that ${\mathcal C}$ is symmetric. Then, $c_{H,H}{\circ}\delta_{H}=\delta_{H}$ and, by Theorem 3.22 of [@Asian], $\lambda_{H}^{2}=id_{H}$. If we denote by $H^{op}$ the unital magma $H^{op}=(H, \eta_{H^{op}}=\eta_{H}, \mu_{H^{op}}=\mu_{H}{\circ}c_{H,H})$, we have that $(H^{op}, \rho_{H^{op}}=(H{\otimes}\lambda_{H}){\circ}\delta_{H})$ is an example of right $H$-comodule magma. Indeed, first note that $H^{op}$ is a unital magma. Also $(H{\otimes}\varepsilon_{H}){\circ}\rho_{H^{op}}=id_{H}$ because $\lambda_{H}$ presevers the counit. On the other hand, if $H$ is cocommutative, the equality $$\label{2ex}
\delta_{H}{\circ}\lambda_{H}=(\lambda_{H}{\otimes}\lambda_{H}){\circ}\delta_{H}$$ holds. Then, by the coassociativity of $\delta_{H}$ and (\[2ex\]), we obtain that $(\rho_{H^{op}}{\otimes}H){\circ}\rho_{H^{op}}=(H{\otimes}\delta_{H}){\circ}\rho_{H^{op}}$, an we have that $(H^{op}, \rho_{H^{op}})$ is a right $H$-comodule. Finally,
- $\hspace{0.38cm} \rho_{H^{op}}{\circ}\mu_{H^{op}}$
- $= (\mu_{H}{\otimes}(\lambda_{H}{\circ}\mu_{H})){\circ}(c_{H,H}{\otimes}c_{H,H}){\circ}\delta_{H{\otimes}H} $ [([naturality of $c$ and $c^{2}=id_{H}$]{})]{}
- $=(\mu_{H^{op}}{\otimes}(\mu_{H}{\circ}(\lambda_{H}{\otimes}\lambda_{H}))){\circ}\delta_{H{\otimes}H} $ [([(53) of [@Asian] and $c^{2}=id_{H}$]{})]{}
- $=(\mu_{H^{op}}{\otimes}\mu_{H}){\circ}(H{\otimes}c_{H,H}{\otimes}H){\circ}(\rho_{H^{op}}{\otimes}\rho_{H^{op}})$ [([naturality of $c$]{})]{}
and
- $\hspace{0.38cm} \rho_{H^{op}}{\circ}\eta_{H^{op}}$
- $= (H{\otimes}(\lambda_{H}{\circ}\Pi_{H}^{L})) {\circ}\delta_{H}{\circ}\eta_{H} $ [([(21) of [@Asian]]{})]{}
- $=(H{\otimes}(\lambda_{H}{\circ}\overline{\Pi}_{H}^{L})) {\circ}\delta_{H}{\circ}\eta_{H} $ [([if $H$ is cocommutative $\Pi_{H}^{L}=\overline{\Pi}_{H}^{L}$]{})]{}
- $= (H{\otimes}\Pi_{H}^{L}) {\circ}\delta_{H}{\circ}\eta_{H} $ [([(39) of [@Asian]]{})]{}
- $=(H{\otimes}(\Pi_{H}^{L}{\circ}\Pi_{H}^{L})) {\circ}\delta_{H}{\circ}\eta_{H} $ [([$\Pi_{H}^{L}$ is idempotent]{})]{}
- $= (H{\otimes}\Pi_{H}^{L}) {\circ}\rho_{H^{op}}{\circ}\eta_{H} $ [([(21) of [@Asian]]{})]{}.
Therefore, $(H^{op}, \rho_{H^{op}})$ is a right $H$-comodule magma. Note that, if $H$ is a Hopf quasigroup we have the same example.
3\. By Example 3.1 of [@our2] we have the following. Let $H$ be a Hopf quasigroup and $A$ a unital magma in ${\mathcal C}$. If there exists a morphism $\varphi_A:H{\otimes}A\to A$ such that $$\label{et1}
\varphi_A{\circ}(\eta_H{\otimes}A)= id_A,$$ $$\label{et2}
\varphi_A{\circ}(H{\otimes}\eta_A)=\varepsilon_H{\otimes}\eta_A,$$ hold, then the smash product $A\sharp H=(A\otimes H,\eta_{A\sharp H},\mu_{A\sharp H})$ defined by $$\eta_{A\sharp H}=\eta_A{\otimes}\eta_H,\;\;
\mu_{A\sharp H}=(\mu_A{\otimes}\mu_H){\circ}(A{\otimes}\psi_{H}^{A}{\otimes}H),$$ where $$\psi_{H}^{A}=(\varphi_{A}{\otimes}H){\circ}(H{\otimes}c_{H,A}){\circ}(\delta_{H}{\otimes}A),$$ is a right $H$-comodule magma with comodule structure given by $$\varrho_{A\sharp H}=A{\otimes}\delta_H.$$
4\. Let $H$, $B$ two Hopf quasigroups. Assume that there exists a morphism of Hopf quasigroups $g:B\rightarrow H$, i.e., a morphism of unital magmas and comonoids. Then, $(B,\rho_{B}=(B{\otimes}g){\circ}\delta_{B})$ is an example of right $H$-comodule magma.
Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. We will say that $h:H\rightarrow B$ is an integral if it is a morphism of right $H$-comodules. The integral will be called total if $h{\circ}\eta_{H}=\eta_{B}$.
\[idem-B\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be a total integral. The endomorphism $q_{B}:=\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H})){\circ}\rho_{B}:B\rightarrow B$ satisfies $$\label{idemp-0}
\rho_{B}{\circ}q_{B}=(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}{\circ}q_{B},$$ $$\label{idemp-1}
\rho_{B}{\circ}q_{B}=(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}q_{B},$$ and, as a consequence, $q_{B}$ is an idempotent morphism. Moreover, if $B^{co H}$ (object of coinvariants) is the image of $q_{B}$ and $p_{B}:B\rightarrow B^{co H}$, $i_{B}:B^{co H}\rightarrow B$ are the morphisms such that $q_{B}=i_{B}{\circ}p_{B}$ and $id_{B^{co H}}=p_{B}{\circ}i_{B}$, $$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$B^{co H}$}}
\put(9,2){\makebox(0,0){$B$}} \put(24,2){\makebox(0,0){$B{\otimes}H,$}}
\put(5.5,3){\makebox(0,0){$i_{B}$}}
\put(16,3.5){\makebox(0,0){$ \rho_{B}$}}
\put(16,0.15){\makebox(0,0){$(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}$}}
\end{picture}$$
$$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$B^{co H}$}}
\put(9,2){\makebox(0,0){$B$}} \put(24,2){\makebox(0,0){$B{\otimes}H,$}}
\put(5.5,3){\makebox(0,0){$i_{B}$}}
\put(16,3.25){\makebox(0,0){$ \rho_{B}$}}
\put(16,0.15){\makebox(0,0){$(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}$}}
\end{picture}$$ are equalizer diagrams.
First note that, by the naturality of $c$, the condition of right $H$-comodule morphism for $h$, and the equality $h{\circ}\eta_{H}=\eta_{B}$, we obtain that $$\label{idemp-2}
h{\circ}\Pi_{H}^{R}=(B{\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(c_{H,B}{\otimes}H){\circ}(H{\otimes}(\rho_{B}{\circ}\eta_{B}))$$ holds. Then, as a consequence, by (\[chmagma\]) and the properties of $\varepsilon_{H}$ and $\eta_{B}$, we have $$\label{idemp-3}
\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{B}=id_{B}.$$
Also,
- $\hspace{0.38cm}\rho_{B}{\circ}q_{B}$
- $= \mu_{B{\otimes}H}{\circ}(\rho_{B}{\otimes}(\rho_{B}{\circ}h{\circ}\lambda_{H})){\circ}\rho_{B}$ [([(\[chmagma\])]{})]{}
- $=(\mu_{B}{\otimes}H){\circ}(B{\otimes}((h{\otimes}\mu_{H}){\circ}(c_{H,H}{\otimes}H){\circ}(H{\otimes}(\delta_{H}{\circ}\lambda_{H})){\circ}\delta_{H})){\circ}\rho_{B}$ [([comodule condition for $B$ and ]{}]{}
- $\hspace{0.38cm} $ [)]{}
- $=(\mu_{B}{\otimes}H){\circ}(B{\otimes}((h{\otimes}\mu_{H}){\circ}(c_{H,H}{\otimes}H){\circ}(H{\otimes}c_{H,H}){\circ}(((H{\otimes}\lambda_{H}){\circ}\delta_{H}){\otimes}\lambda_{H}){\circ}\delta_{H})){\circ}\rho_{B}$
- $\hspace{0.38cm} $[([anticomultiplicativity of $\lambda_{H}$ and coassociativity of $\delta_{H}$]{})]{}
- $=((\mu_{B}{\circ}(B{\otimes}h)){\otimes}\Pi_{H}^{L}){\circ}(B{\otimes}(c_{H,H}{\circ}(H{\otimes}\lambda_{H}){\circ}\delta_{H})){\circ}\rho_{B}$ [([naturality of $c$]{})]{},
and then using that $\Pi_{H}^{L}$ is an idempotent morphism we obtain (\[idemp-0\]). Therefore, by (34) of [@Asian], we have (\[idemp-1\]) because $$(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}q_{B}=(B{\otimes}(\overline{\Pi}_{H}^{R}{\circ}\Pi_{H}^{L})){\circ}\rho_{B}{\circ}q_{B}=(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}{\circ}q_{B}=\rho_{B}{\circ}q_{B}.$$
Then, $q_{B}$ is an idempotent morphism. Indeed,
- $\hspace{0.38cm} q_{B}{\circ}q_{B}$
- $= \mu_{H} {\circ}(B{\otimes}(h{\circ}\lambda_{H}{\circ}\overline{\Pi}_{H}^{R}) ){\circ}\rho_{B}{\circ}q_{B}$[([(\[idemp-1\])]{})]{}
- $=\mu_{B} {\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{B}{\circ}q_{B}$ [([(40) of [@Asian]]{})]{}
- $=q_{B}$ [([(\[idemp-3\])]{})]{}.
On the other hand, by (40) of [@Asian] and (\[idemp-3\]), $$\mu_{B} {\circ}(B{\otimes}(h{\circ}\lambda_{H}{\circ}\overline{\Pi}_{H}^{R}) ){\circ}\rho_{B}=\mu_{B} {\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R})) ){\circ}\rho_{B}=id_{B}.$$
Then, $$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$B^{co H}$}}
\put(9,2){\makebox(0,0){$B$}} \put(24,2){\makebox(0,0){$B{\otimes}H$}}
\put(5.5,3){\makebox(0,0){$i_{B}$}}
\put(16,3.5){\makebox(0,0){$ \rho_{B}$}}
\put(16,0.15){\makebox(0,0){$(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}$}}
\end{picture}$$ is a split cofork (see [@Mac]) and thus an equalizer diagram. As a consequence, by (33) of [@Asian], we have $$(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}{\circ}q_{B}=(B{\otimes}(\Pi_{H}^{L}{\circ}\overline{\Pi}_{H}^{R})){\circ}\rho_{B}{\circ}q_{B}=(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}q_{B}=\rho_{B}{\circ}q_{B}$$ and $q_{B}$ equalizes $\rho_{B}$ and $(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}$. Therefore, $i_{B}$ equalizes $\rho_{B}$ and $(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}$ because $p_{B}{\circ}i_{B}=id_{B^{coH}}$. Moreover, if $t:D\rightarrow B$ is such that $(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}{\circ}t=\rho_{B}{\circ}t$, composing with $B{\otimes}\overline{\Pi}_{H}^{R}$, we obtain that $(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}t=\rho_{B}{\circ}t$. Thus, there exists a unique morphism $t^{\prime}:D\rightarrow B^{coH}$ such that $i_{B}{\circ}t^{\prime}=t$. Then, $$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$B^{co H}$}}
\put(9,2){\makebox(0,0){$B$}} \put(24,2){\makebox(0,0){$B{\otimes}H$}}
\put(5.5,3){\makebox(0,0){$i_{B}$}}
\put(16,3.5){\makebox(0,0){$ \rho_{B}$}}
\put(16,0.15){\makebox(0,0){$(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}$}}
\end{picture}$$ is an equalizer diagram.
Note that, under the conditions of the previous proposition, the object of coinvariants is independent of the total integral $h$ because it is the equalizer object of $\rho_{B}$ and $(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}$ (or equivalently, $\rho_{B}$ and $(B{\otimes}\Pi_{H}^{L}){\circ}\rho_{B}$).
Moreover, (see [@JPAA]) the triple $(B^{co H}, \eta_{B^{co H}}, \mu_{B^{co H}})$ is a unital magma (the submagma of coinvariants of $B$), where $\eta_{B^{co H}}:K\rightarrow B^{co H}$, $\mu_{B^{co H}}:B^{co H}{\otimes}B^{co H}\rightarrow B^{co H}$ are the factorizations through $i_{B}$ of the morphisms $\eta_B$ and $\mu_B{\circ}(i_B{\otimes}i_B)$, respectively. Therefore, $\eta_{B^{co H}}$ is the unique morphism such that $$\label{etaH}
\eta_{B}=i_{B}{\circ}\eta_{B^{co H}},$$ and $\mu_{B^{co H}}$ is the unique morphism satisfying $$\label{muH}
\mu_{B}{\circ}(i_{B}{\otimes}i_{B})=i_{B}{\circ}\mu_{B^{co H}}.$$
Thus, $$\label{etaH1}
\eta_{B^{co H}}=p_{B}{\circ}\eta_{B},$$ and $$\label{muH1}
\mu_{B^{co H}}=p_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}i_{B}).$$
In what follows, the object of coinvariants $B^{co H}$ will be called the submagma of coinvariants of $B$. Note that, if $B=H$ and $\rho_{B}=\delta_{H}$, the submagma of coinvariants is $H^{coH}=H_{L}$ and then, in this case, it is a monoid.
\[anchor\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. We will say that $h:H\rightarrow B$ is an anchor morphism if it is a multiplicative total integral (i.e., a right $H$-comodule morphism such that it is a morphism of unital magmas) and the following equalities hold:
- $\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(B{\otimes}\delta_{H})=\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{L})).$
- $\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(B{\otimes}\delta_{H})=\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R})). $
Note that, if the product on $B$ is associative, every multiplicative total integral $h$ satisfies (c1)-(c2) and therefore is an anchor morphism. Also, using that $h$ is a comodule morphism, the condition (c1) can be rewritten as $$\label{re-c1}
\mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\lambda_{H})){\circ}(B{\otimes}(\rho_{B}{\circ}h))=\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{L})).$$
Finally, if $H$ is a Hopf quasigroup, (c1) and (c2) are $$\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(B{\otimes}\delta_{H})=B{\otimes}\varepsilon_{H}=\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(B{\otimes}\delta_{H}),$$ or, in an equivalent way, $B$ is a cleft right $H$-comodule algebra in the sense of Definition 3.1 of [@our2].
\[anch-ex\]
1\. By the definition of weak Hopf quasigroup and (40) of [@Asian], the identity morphism $id_{H}$ is an anchor morphism for the right $H$-comodule magma $(H,\delta_{H})$. Also, if $H$ is a Hopf quasigrup, the identity of $H$ is an anchor morphism.
2\. By the second point of Example \[ex-hcm\] we know that, if $H$ is a cocommutative weak Hopf quasigroup and ${\mathcal C}$ is symmetric, $(H^{op}, \rho_{H^{op}}=(H{\otimes}\lambda_{H}){\circ}\delta_{H})$ is an example of right $H$-comodule magma. Note that in this case we have that $\lambda_{H}{\circ}\eta_{H}=\eta_{H}$ and
- $\hspace{0.38cm} \rho_{H^{op}}{\circ}\lambda_{H}$
- $= (\lambda_{H}{\otimes}\lambda_{H}^2){\circ}\delta_{H} $[([(\[2ex\])]{})]{}
- $=(\lambda_{H}{\otimes}H){\circ}\delta_{H} $ [([Theorem 3.22 of [@Asian], i.e., $\lambda_{H}^2=id_{H}$]{})]{}
Therefore $\lambda_{H}$ is a total integral. Moreover, by (52) of [@Asian], we obtain that $\mu_{H^{op}}{\circ}(\lambda_{H}{\otimes}\lambda_{H})=\lambda_{H}{\circ}\mu_{H}$. Finally,
- $\hspace{0.38cm} \mu_{H^{op}}{\circ}((\mu_{H^{op}}{\circ}(H{\otimes}\lambda_{H})){\otimes}(\lambda_{H}{\circ}\lambda_{H})){\circ}(H{\otimes}\delta_{H})$
- $=\mu_{H}{\circ}(H{\otimes}\mu_{H}){\circ}(c_{H,H}{\otimes}H){\circ}(H{\otimes}c_{H,H}){\circ}(c_{H,H}{\otimes}H){\circ}(H{\otimes}c_{H,H}) {\circ}(H{\otimes}((H{\otimes}\lambda_{H}){\circ}\delta_{H}))$
- $\hspace{0.38cm}$[([Theorem 3.22 of [@Asian], naturality of $c$ and coassociativity of $\delta_{H}$]{})]{}
- $=\mu_H{\circ}(H{\otimes}\mu_H){\circ}(H{\otimes}\lambda_H{\otimes}H){\circ}(\delta_H{\otimes}H){\circ}c_{H,H} $ [([naturality of $c$ and $c^2=id_{H}$]{})]{}
- $=\mu_{H}{\circ}(\Pi_{H}^{L}{\otimes}H){\circ}c_{H,H} $ [([(a4-5) of Definition \[W-H-quasi\]]{})]{}
- $= \mu_{H^{op}}{\circ}(H{\otimes}\Pi_{H}^{L}) $[([naturality of $c$]{})]{}
- $=\mu_{H^{op}}{\circ}(H{\otimes}(\lambda_{H}{\circ}\overline{\Pi}_{H}^{L}))$ [([(39) of [@Asian]]{})]{}
- $=\mu_{H^{op}}{\circ}(H{\otimes}(\lambda_{H}{\circ}\Pi_{H}^{L}))$ [([$H$ is cocommutative, $\Pi_{H}^{L}=\overline{\Pi}_{H}^{L}$]{})]{},
and, by similar arguments, $$\mu_{H^{op}}{\circ}((\mu_{H^{op}}{\circ}(H{\otimes}(\lambda_{H}{\circ}\lambda_{H}))){\otimes}\lambda_{H}){\circ}(H{\otimes}\delta_{H})=\mu_{H^{op}}{\circ}(H{\otimes}(\lambda_{H}{\circ}\Pi_{H}^{R})).$$
Therefore, $\lambda_{H}$ is an anchor morphism for $(H^{op}, \rho_{H^{op}})$. Of course, the same result holds for cocommutative Hopf quasigroups.
3\. In the third point of Example \[ex-hcm\] we saw that if $H$ is a Hopf quasigroup, $A$ is a unital magma in ${\mathcal C}$, and there exists a morphism $\varphi_A:H{\otimes}A\to A$ satisfying (\[et1\]), (\[et2\]), the smash product $A\sharp H$ is a right $H$-comodule magma with coaction $\rho_{A\sharp H}=A{\otimes}\delta_{H}$. By Example 3.1 of [@our2], we have that $h=\eta_{A}{\otimes}H:H\rightarrow A\sharp H$ is a total integral. On the other hand,
- $\hspace{0.38cm} \mu_{A\sharp H}{\circ}(h{\otimes}h)$
- $= ((\varphi_{A}{\circ}(H{\otimes}\eta_{A})){\otimes}\mu_{H}){\circ}(\delta_{H}{\otimes}H)$[([unit properties and naturality of $c$]{})]{}
- $= h{\circ}\mu_{H}$ [([(\[et2\]) and counit properties]{})]{}
and then $h$ is multiplicative. Also, by similar arguments, we have $$\label{nwx}
\mu_{A\sharp H}{\circ}(A{\otimes}H{\otimes}h)=A{\otimes}\mu_{H}.$$
Thus, by (\[nwx\]) and (\[rightHqg\]), the identities $$\mu_{A\sharp H}{\circ}((\mu_{A\sharp H}{\circ}(A{\otimes}H{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(A{\otimes}H{\otimes}\delta_{H})=A{\otimes}(\mu_H\circ(\mu_H{\otimes}\lambda_H)\circ (H{\otimes}\delta_H))=A{\otimes}H{\otimes}\varepsilon_{H},$$ hold. Similarly, by (\[nwx\]) and (\[leftHqg\]) we obtain that $$\mu_{A\sharp H}{\circ}((\mu_{A\sharp H}{\circ}(A{\otimes}H{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(A{\otimes}H{\otimes}\delta_{H})=A{\otimes}H{\otimes}\varepsilon_{H}.$$
Therefore, $h=\eta_{A}{\otimes}H$ is an anchor morphism.
4\. Assume that $H$ and $B$ are Hopf quasigrous in ${\mathcal C}$. Let $g:B\to H$, $f:H\to B$ be morphisms of Hopf quasigroups such that $g{\circ}f=id_H$. Consider the right $H$-comodule structure on $B$ defined in fourth point of Example \[ex-hcm\]. Then, $f$ is an anchor morphism. Indeed, first note that $f{\circ}\eta_{H}=\eta_{B}$, $f{\circ}\mu_{H}=\mu_{B}{\circ}(f{\otimes}f)$ hold because $f$ is a morphism of unital magmas. Also, $f$ is a comodule morphism, i.e., $\rho_{B}{\circ}f=(f{\otimes}H){\circ}\delta_{H}$, because $f$ is a comonoid morphism and $g{\circ}f=id_H$. On the other hand, by Lemma 1.4 of [@our1], we have that $$\label{elambda}
\lambda_{B}{\circ}f=f{\circ}\lambda_{H}$$ holds, and then, using that $f$ is a comonoid morphism, we get $$\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}f)){\otimes}(f{\circ}\lambda_{H})){\circ}(B{\otimes}\delta_{H})=\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}f)){\otimes}(\lambda_{B}{\circ}f)){\circ}(B{\otimes}\delta_{H})$$ $$=\mu_B\circ(\mu_B{\otimes}\lambda_B)\circ (B{\otimes}(\delta_B{\circ}f))=B{\otimes}(\varepsilon_{B}{\circ}f)=B{\otimes}\varepsilon_{H}$$
Similarly, by the same arguments, $$\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(f{\circ}\lambda_{H}))){\otimes}f){\circ}(B{\otimes}\delta_{H})=B{\otimes}\varepsilon_{H},$$ and this implies that $f$ is an anchor morphism.
As a consequence, the examples of strong projections that we can find in [@our1] and [@cocy] provide examples of anchor morphisms.
If $(B,\rho_{B})$ is a right $H$-comodule monoid the following identity $$\label{Eti-1}
(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}\mu_{B}=(\mu_{B}{\otimes}\overline{\Pi}_{H}^{R}){\circ}(B{\otimes}\rho_{B}).$$ holds. Indeed:
- $\hspace{0.38cm}(B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}\mu_{B}$
- $=(\mu_{B}{\otimes}(\varepsilon_{H}{\circ}\mu_{H}{\circ}(\mu_{H}{\otimes}H)){\otimes}H){\circ}(B{\otimes}c_{H,B}{\otimes}H{\otimes}(\delta_{H}{\circ}\eta_{H})){\circ}(\rho_{B}{\otimes}\rho_{B}) $[([(\[chmagma\]) and definition of $\overline{\Pi}_{H}^{R}$]{})]{}
- $=(\mu_{B}{\otimes}(((\varepsilon_{H}{\circ}\mu_{H}){\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(H{\otimes}\delta_{H}{\otimes}H)){\otimes}H){\circ}(B{\otimes}c_{H,B}{\otimes}H{\otimes}(\delta_{H}{\circ}\eta_{H})){\circ}(\rho_{B}{\otimes}\rho_{B}) $
- $\hspace{0.38cm}$[([(a2) of Definition \[W-H-quasi\]]{})]{}
- $= (B{\otimes}\varepsilon_{H} {\otimes}\varepsilon_{H}{\otimes}H){\circ}(\mu_{B{\otimes}H}{\otimes}\mu_{H}{\otimes}H){\circ}(\rho_{B}{\otimes}\rho_{B}{\otimes}H{\otimes}(\delta_{H}{\circ}\eta_{H})){\circ}(B{\otimes}\rho_{B})$ [([comodule condition ]{}]{}
- $\hspace{0.38cm}$[)]{}
- $=(((B{\otimes}\varepsilon_{H}){\circ}\rho_{B}{\circ}\mu_{B}){\otimes}\overline{\Pi}_{H}^{R})){\circ}(B{\otimes}\rho_{B}) $ [([(\[chmagma\]) and definition of $\overline{\Pi}_{H}^{R}$]{})]{}
- $=(\mu_{B}{\otimes}\overline{\Pi}_{H}^{R}){\circ}(B{\otimes}\rho_{B})$ [([comodule condition for $B$]{}).]{}
As a consequence, if $h:H\rightarrow B$ is a total integral, the equality $$\label{c3}
\mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}(B{\otimes}\rho_{B})=\mu_{B}$$ holds because
- $\hspace{0.38cm}\mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}(B{\otimes}\rho_{B})$
- $=\mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\lambda_{H}{\circ}\overline{\Pi}_{H}^{R})){\circ}(B{\otimes}\rho_{B})$ [([(40) of [@Asian]]{})]{}
- $= \mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\lambda_{H}{\circ}\overline{\Pi}_{H}^{R})){\circ}\rho_{B}{\circ}\mu_{B}$ [([(\[Eti-1\])]{})]{}
- $= \mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{B}{\circ}\mu_{B}$ [([(40) of [@Asian]]{})]{}
- $=\mu_{B}$ [([(\[idemp-3\])]{}).]{}
Also, if $h:H\rightarrow B$ is a multiplicative morphism of right $H$-comodules, the following identity holds $$\label{anchor1}
h{\circ}\Pi_{H}^{L}=q_{B}{\circ}h,$$ because
- $\hspace{0.38cm}h{\circ}\Pi_{H}^{L}$
- $= \mu_{B}{\circ}(h{\otimes}(h{\circ}\lambda_{H})){\circ}\delta_{H}$ [([$h$ is multiplicative]{})]{}
- $=q_{B}{\circ}h$ [([$h$ is a comodule morphism]{}).]{}
If $h:H\rightarrow B$ is an anchor morphism we have that the equality $$\label{c4}
\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{L}))){\otimes}h){\circ}(B{\otimes}\delta_{H})=\mu_{B}{\circ}(B{\otimes}h)$$ holds. Indeed:
- $\hspace{0.38cm} \mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{L}))){\otimes}h){\circ}(B{\otimes}\delta_{H})$
- $= \mu_{B}{\circ}((\mu_{B}{\circ}(\mu_{B}{\otimes}(h{\circ}\lambda_{H})){\circ}(B{\otimes}\rho_{B})){\otimes}h){\circ}(B{\otimes}(\rho_{B}{\circ}h))$ [([(\[re-c1\]) and comodule morphism condition ]{}]{}
- $\hspace{0.38cm}$[)]{}
- $=\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(B{\otimes}\delta_{H})){\circ}(\mu_{B}{\otimes}H){\circ}(B{\otimes}(\rho_{B}{\circ}h))$ [([comodule condition for $B$]{})]{}
- $=\mu_{B}{\circ}((\mu_{B}{\circ}(h{\circ}\Pi_{H}^{R})){\circ}(B{\otimes}(\rho_{B}{\circ}h))$ [([(c2) of Definition \[anchor\]]{})]{}
- $=\mu_{B}{\circ}(B{\otimes}h) $ [([(\[c3\])]{}).]{}
Then, by (\[anchor1\]) and the comodule morphism condition for $h$, (\[c4\]) is equivalent to $$\label{anchor2}
\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}q_{B})){\otimes}h){\circ}(B{\otimes}(\rho_{B}{\circ}h))=\mu_{B}{\circ}(B{\otimes}h).$$
Finally, for an anchor morphism $h:H\rightarrow B$ the equality $$\label{anchor21}
\mu_{B}{\circ}(q_{B}{\otimes}h){\circ}\rho_{B}=id_{B}$$ holds, because
- $\hspace{0.38cm}\mu_{B}{\circ}(q_{B}{\otimes}h){\circ}\rho_{B}$
- $=\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(B{\otimes}\delta_{H}){\circ}\rho_{B} $ [([comodule condition for $B$]{})]{}
- $=\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{B} $ [([(c2) of Definition \[anchor\]]{})]{}
- $= id_{B}$ [([(\[idemp-3\])]{}).]{}
\[H-D-mod\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism and let $M$ be an object in ${\mathcal
C}$. We say that $(M, \phi_{M}, \rho_{M})$ is a strong $(H,B,h)$-Hopf module if the following axioms hold:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}B\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{B})=id_{M}.$
- $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{B})){\otimes}B)=\phi_{M}{\circ}(M{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))).$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M}{\otimes}\rho_{B})$.
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}h)){\otimes}(h{\circ}\lambda_H))\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{L})).$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{R})).$
For example, the triple $(H, \mu_{H}, \delta_{H})$ is a strong $(H,H,id_{H})$-Hopf module. Also, if the following equality $$\label{strong-1}
\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}i_{B})){\otimes}B)=\mu_{B}{\circ}(B{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B)))$$ holds, the triple $(B, \mu_{B}, \rho_{B})$ is a strong $(H,B, h)$-Hopf module.
Let $(M, \phi_{M}, \rho_{M})$ be a strong $(H,B,h)$-Hopf module. In a similar way to (\[Eti-1\]), but using (d2-3) of Definition \[H-D-mod\] instead of (\[chmagma\]), it is easy to see that $$\label{Eti-1-1}
(M{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\overline{\Pi}_{H}^{R}){\circ}(M{\otimes}\rho_{M})$$ holds. Moreover, by (\[idemp-2\]), (d2-3), (d2-1), and the condition of comodule for $M$, we also have the equality $$\label{idemp-3M}
\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{M}=id_{M}.$$
As a consequence, we can obtain the Hopf module version of (\[c3\]), i.e., $$\label{d2-6}
\phi_{M}\circ (\phi_{M}{\otimes}(h{\circ}\Pi_{H}^{R}))\circ (M{\otimes}\rho_{B})=\phi_{M}.$$
Moreover, in this setting, the equality $$\label{d2-7}\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{L}))){\otimes}h)\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}h)$$ holds because
- $\hspace{0.38cm}\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{L}))){\otimes}h)\circ (M{\otimes}\delta_{H}) $
- $=\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(q_{B}{\circ}h))){\otimes}h)\circ (M{\otimes}\delta_{H}) $ [([(\[anchor1\])]{})]{}
- $= \phi_{M}\circ (M{\otimes}(\mu_{B}{\circ}((q_{B}{\circ}h){\otimes}h){\circ}\delta_{H}))$ [([(d2-2) of Definition \[H-D-mod\]]{})]{}
- $=\phi_{M}\circ (M{\otimes}(\mu_{B}{\circ}((h{\circ}\Pi_{H}^{L}){\otimes}h){\circ}\delta_{H})) $ [([(\[anchor1\])]{})]{}
- $=\phi_{M}\circ (M{\otimes}(h{\circ}(\Pi_{H}^{L}\ast id_{H}))) $ [([$h$ is multiplicative]{})]{}
- $=\phi_{M}{\circ}(M{\otimes}h) $ [([(4) of [@Asian]]{}).]{}
Finally, by (\[anchor1\]) the equality (\[d2-7\]) is equivalent to $$\label{anchor3}
\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}q_{B})){\otimes}h)\circ (M{\otimes}(\rho_{B}{\circ}h))=\phi_{M}{\circ}(M{\otimes}h).$$
\[idem-M\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism and let $(M, \phi_{M}, \rho_{M})$ be a strong $(H,B,h)$-Hopf module. The endomorphism $q_{M}:=\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H})){\circ}\rho_{M}:M\rightarrow M$ satisfies $$\label{idemp-0M}
\rho_{M}{\circ}q_{M}=(M{\otimes}\Pi_{H}^{L}){\circ}\rho_{M}{\circ}q_{M},$$ $$\label{idemp-1M}
\rho_{M}{\circ}q_{M}=(M{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{M}{\circ}q_{M},$$ and, as a consequence, $q_{M}$ is an idempotent. Moreover, if $M^{co H}$ (object of coinvariants) is the image of $q_{M}$ and $p_{M}:M\rightarrow M^{co H}$, $i_{M}:M^{co H}\rightarrow M$ are the morphisms such that $q_{M}=i_{M}{\circ}p_{M}$ and $id_{M^{co H}}=p_{M}{\circ}i_{M}$, $$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$M^{co H}$}}
\put(9,2){\makebox(0,0){$M$}} \put(24,2){\makebox(0,0){$M{\otimes}H,$}}
\put(5.5,3){\makebox(0,0){$i_{M}$}}
\put(16,3.5){\makebox(0,0){$ \rho_{M}$}}
\put(16,0.15){\makebox(0,0){$(M{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{M}$}}
\end{picture}$$ $$\setlength{\unitlength}{3mm}
\begin{picture}(30,4)
\put(3,2){\vector(1,0){4}} \put(11,2.5){\vector(1,0){10}}
\put(11,1.5){\vector(1,0){10}} \put(1,2){\makebox(0,0){$M^{co H}$}}
\put(9,2){\makebox(0,0){$M$}} \put(24,2){\makebox(0,0){$M{\otimes}H,$}}
\put(5.5,3){\makebox(0,0){$i_{M}$}}
\put(16,3.25){\makebox(0,0){$ \rho_{M}$}}
\put(16,0.15){\makebox(0,0){$(M{\otimes}\Pi_{H}^{L}){\circ}\rho_{M}$}}
\end{picture}$$ are equalizer diagrams.
The proof is similar to the one developed in Proposition \[idem-B\] but using (d2-3) instead of (\[chmagma\]) and the comodule condition for $M$ instead of the comodule condition for $B$.
Note that, as in the case of $B$, the object of coinvariants $M^{coH}$ is independent of the anchor morphism $h$ because it is the equalizer object of $\rho_{M}$ and $(M{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{M}$ (or $\rho_{M}$ and $(M{\otimes}\Pi_{H}^{L}){\circ}\rho_{M}$).
\[l-idem\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism and let $(M, \phi_{M}, \rho_{M})$ be a strong $(H,B,h)$-Hopf module. The following equalities hold: $$\label{l-idem0}
\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)=(\phi_{M}{\otimes}H){\circ}(i_{M}{\otimes}\rho_{B}),$$ $$\label{l-idem2}
q_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)=\phi_{M}{\circ}(i_{M}{\otimes}q_{B}),$$ $$\label{l-idem1}
(q_{M}{\otimes}H){\circ}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)=((\phi_{M}{\circ}(M{\otimes}q_{B})){\otimes}H){\circ}(i_{M}{\otimes}\rho_{B}),$$ $$\label{l-idem1-1}
(p_{M}{\otimes}H){\circ}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)=((p_{M}{\circ}\phi_{M}{\circ}(M{\otimes}q_{B})){\otimes}H){\circ}(i_{M}{\otimes}\rho_{B}),$$ $$\label{l-idem2-1}
q_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})=\phi_{M}{\circ}(i_{M}{\otimes}i_{B}),$$ $$\label{l-idem3}
p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)=p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}q_{B}).$$ $$\label{l-idem4}
\phi_{M}{\circ}(q_{M}{\otimes}h){\circ}\rho_{M}=id_{M}.$$
The proof for the first equality is the following:
- $\hspace{0.38cm} \rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)$
- $= (\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}((\rho_{M}{\circ}i_{M}){\otimes}\rho_{B}) $ [([(d2-3) of Definition (\[H-D-mod\])]{})]{}
- $=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(((M{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{M}{\circ}i_{M}){\otimes}\rho_{B}) $ [([(\[idemp-1M\])]{})]{}
- $= (\phi_{M}{\otimes}(\mu_{H}{\circ}( \overline{\Pi}_{H}^{R}{\otimes}H))){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}((\rho_{M}{\circ}i_{M}){\otimes}\rho_{B}) $ [([naturality of $c$]{})]{}
- $=(\phi_{M}{\otimes}(\varepsilon_{H}{\circ}\mu_{H}){\otimes}H){\circ}(M{\otimes}c_{H,B}{\otimes}H{\otimes}H){\circ}((\rho_{M}{\circ}i_{M}){\otimes}((B{\otimes}\delta_{H}){\circ}\rho_{B})) $ [([(10) of [@Asian]]{})]{}
- $= (((\phi_{M}{\otimes}(\varepsilon_{H}{\circ}\mu_{H})){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M} {\otimes}\rho_{B})){\otimes}H){\circ}(i_{M}{\otimes}\rho_{B})$ [([comodule condition for $B$]{})]{}
- $= (((M{\otimes}\varepsilon_{H}){\circ}\rho_{M}{\circ}\phi_{M}){\otimes}H){\circ}(i_{M}{\otimes}\rho_{B}) $ [([(d2-3) of Definition (\[H-D-mod\])]{})]{}
- $= (\phi_{M}{\otimes}H){\circ}(i_{M}{\otimes}\rho_{B}) $ [([comodule condition for $B$]{}).]{}
The equality (\[l-idem2\]) follows by
- $\hspace{0.38cm} q_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)$
- $= \phi_{M}\circ (\phi_{M}{\otimes}(h{\circ}\lambda_H))\circ (i_{M}{\otimes}\rho_{B}) $ [([(\[l-idem0\])]{})]{}
- $=\phi_{M}{\circ}(i_{M}{\otimes}q_{B})$ [([(d2-4) of Definition (\[H-D-mod\])]{}).]{}
Thus, composing in (\[l-idem0\]) with $q_{M}{\otimes}H$, we obtain (\[l-idem1\]). Also, composing in (\[l-idem1\]) with $p_{M}{\otimes}B$, we have (\[l-idem1-1\]). Moreover, composing with $M^{coH}{\otimes}i_{B}$ in (\[l-idem2\]), we obtain (\[l-idem2-1\]), and doing the same with $p_{M}$ we get (\[l-idem3\]). Finally, (\[l-idem4\]) holds because
- $\hspace{0.38cm}\phi_{M}{\circ}(q_{M}{\otimes}h){\circ}\rho_{M}$
- $=\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H}))){\otimes}h){\circ}(M{\otimes}\delta_{H}){\circ}\rho_{M} $ [([comodule condition for $M$]{})]{}
- $=\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{R})){\circ}\rho_{M} $ [([(d2-5) of Definition (\[H-D-mod\])]{})]{}
- $= id_{M}$ [([(\[idemp-3M\])]{}).]{}
\[B-coinv\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism. If (\[strong-1\]) holds, the submagma of coinvariants $(B^{co H}, \eta_{B^{co H}}, \mu_{B^{co H}})$ is a monoid.
Firstly, remember that if (\[strong-1\]) holds, the triple $(B,\mu_{B},\rho_{B})$ is a strong $(H,B,h)$-Hopf module. Then by (\[l-idem2\]), $$\label{l-idemB}
q_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}i_{B})=\mu_{B}{\circ}(i_{B}{\otimes}i_{B}).$$
As a consequence, $B^{co H}$ is a monoid because,
- $\hspace{0.38cm} i_{B}{\circ}\mu_{B^{co H}} {\circ}(B^{coH}{\otimes}\mu_{B^{co H}})$
- $= q_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}(q_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}i_{B})))$ [([(\[muH1\])]{})]{}
- $=q_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}( \mu_{B}{\circ}(i_{B}{\otimes}i_{B})))$ [([(\[l-idemB\])]{})]{}
- $= q_{B}{\circ}\mu_{B}{\circ}(( \mu_{B}{\circ}(i_{B}{\otimes}i_{B})){\otimes}i_{B}) $ [([(\[strong-1\])]{})]{}
- $=q_{B}{\circ}\mu_{B}{\circ}(( q_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}i_{B})){\otimes}i_{B})$ [([(\[l-idemB\])]{})]{}
- $=i_{B}{\circ}\mu_{B^{co H}} {\circ}(\mu_{B^{co H}}{\otimes}B^{co H})$ [([(\[muH1\])]{}).]{}
\[M-coinv\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism. If (\[strong-1\]) holds, for all strong $(H,B,h)$-Hopf module $(M, \phi_{M}, \rho_{M})$, the object of coinvariants $M^{coH}$ is a right $B^{coH}$-module.
First note that, by Proposition \[B-coinv\], the object $B^{co H}$ is a monoid. Now we will show that there exists an action $\phi_{M^{coH}}:M^{coH}{\otimes}B^{coH}\rightarrow M^{coH}$ such that $\phi_{M^{coH}}{\circ}(M^{coH}{\otimes}\eta_{B^{co H}})=id_{M^{coH}},$ and $\phi_{M^{coH}}{\circ}(\phi_{M^{coH}}{\otimes}B^{co H})=\phi_{M^{coH}}{\circ}(M^{coH}{\otimes}\mu_{B^{co H}})$. To define the action, we begin by proving that $$\label{phiMcoH}
\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})=(M{\otimes}\Pi_{H}^{L}){\circ}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B}).$$
Indeed,
- $\hspace{0.38cm} \rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})$
- $=(\phi_{M}{\otimes}H){\circ}(i_{M}{\otimes}(\rho_{B}{\circ}i_{B}))$ [([(\[l-idem0\])]{})]{}
- $= (\phi_{M}{\otimes}\Pi_{H}^{L}){\circ}(i_{M}{\otimes}(\rho_{B}{\circ}i_{B}))$ [([(\[idemp-0\])]{})]{}
- $= (M{\otimes}\Pi_{H}^{L}){\circ}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})$ [([(\[l-idem0\])]{})]{}
Then, there exists a unique morphism $\phi_{M^{coH}}:M^{coH}{\otimes}B^{coH}\rightarrow M^{coH}$ such that $$\label{PhiBcoH}
\phi_{M}{\circ}(i_{M}{\otimes}i_{B})=i_{M}{\circ}\phi_{M^{coH}}.$$
Therefore, $$\label{PhiBcoH1}
\phi_{M^{coH}}=p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B}).$$
The pair $(M^{coH}, \phi_{M^{coH}})$ satisfies the conditions of right $B^{coH}$-module because $$\phi_{M^{coH}}{\circ}(M^{coH}{\otimes}\eta_{B^{co H}})\stackrel{{\scriptsize \blue (\ref{etaH})}}{=}
p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}\eta_{B})\stackrel{{\scriptsize \blue ({\rm d2-1})}}{=}p_{M}{\circ}i_{M}=id_{M^{coH}},$$ and
- $\hspace{0.38cm}\phi_{M^{coH}}{\circ}(\phi_{M^{coH}}{\otimes}B^{co H})$
- $=p_{M}{\circ}\phi_{M}{\circ}((q_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})){\otimes}i_{B})$ [([(\[PhiBcoH1\])]{})]{}
- $= p_{M}{\circ}\phi_{M}{\circ}((\phi_{M}{\circ}(i_{M}{\otimes}i_{B})){\otimes}i_{B})$ [([(\[l-idem2-1\])]{})]{}
- $=p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}i_{B}))) $ [([(d2-2) of Definition (\[H-D-mod\])]{})]{}
- $=p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}(i_{B}{\circ}\mu_{B^{coH}})) $ [([(\[muH\])]{})]{}
- $=\phi_{M^{coH}}{\circ}(M^{coH}{\otimes}\mu_{B^{co H}}) $ [([(\[PhiBcoH1\])]{})]{}.
\[tensor\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism. Assume that (\[strong-1\]) and $$\label{strong2}
\mu_{B}{\circ}(i_{B}{\otimes}\mu_{B})=\mu_{B}{\circ}((\mu_{B}{\circ}(i_{B}{\otimes}B)){\otimes}B)$$ hold. Then if the category ${\mathcal C}$ admits coequalizers and the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers, for all strong $(H,B,h)$-Hopf module $(M, \phi_{M}, \rho_{M})$, the object $M^{coH}{\otimes}_{B^{coH}}B$, defined by the coequalizer of $T_{M}^{1}=\phi_{M^{coH}}{\otimes}B$ and $T_{M}^{2}=M^{coH}{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))$, is a strong $(H,B,h)$-Hopf module. Moreover, $M^{coH}{\otimes}_{B^{coH}}B$ and $M$ are isomorphic as right $H$-comodules.
In the first step, we begin by proving the existence of an action and a coaction for $M^{coH}{\otimes}_{B^{coH}}H$. If the object $M^{coH}{\otimes}_{B^{coH}}H$ is defined by the coequalizer diagram
$$\label{coeq-1}
\setlength{\unitlength}{1mm}
\begin{picture}(101.00,10.00)
\put(22.50,8.00){\vector(1,0){20.00}}
\put(22.50,4.00){\vector(1,0){20.00}}
\put(63.00,6.00){\vector(1,0){20.00}}
\put(32.00,11.00){\makebox(0,0)[cc]{$T_{M}^{1}$ }}
\put(32.00,1.00){\makebox(0,0)[cc]{$T_{M}^{2}$}}
\put(74.00,9.00){\makebox(0,0)[cc]{$n_{M^{coH}}$ }}
\put(7.00,6.00){\makebox(0,0)[cc]{$ M^{coH}\otimes B^{coH}{\otimes}B$ }}
\put(54.00,6.00){\makebox(0,0)[cc]{$ M^{coH}{\otimes}B$ }}
\put(96.00,6.00){\makebox(0,0)[cc]{$M^{coH}{\otimes}_{B^{coH}} B, $ }}
\end{picture}$$
we have that $$\label{coeq-2}
\setlength{\unitlength}{1mm}
\begin{picture}(101.00,10.00)
\put(19.50,8.00){\vector(1,0){20.00}}
\put(19.50,4.00){\vector(1,0){20.00}}
\put(67.00,6.00){\vector(1,0){15.00}}
\put(30.00,11.00){\makebox(0,0)[cc]{$T_{M}^{1}{\otimes}B$ }}
\put(30.00,1.00){\makebox(0,0)[cc]{$T_{M}^{2}{\otimes}B$}}
\put(74.00,9.00){\makebox(0,0)[cc]{$n_{M^{coH}}{\otimes}B$ }}
\put(0.00,6.00){\makebox(0,0)[cc]{$ M^{coH}\otimes B^{coH}{\otimes}B{\otimes}B$ }}
\put(54.00,6.00){\makebox(0,0)[cc]{$ M^{coH}{\otimes}B{\otimes}B$ }}
\put(101.00,6.00){\makebox(0,0)[cc]{$M^{coH}{\otimes}_{B^{coH}} B {\otimes}B$ }}
\end{picture}$$ is also a coequalizer diagram because the functor $-{\otimes}B$ preserves coequalizers. Consider the morphism $n_{M^{coH}}{\circ}(M^{coH}{\otimes}\mu_{B}): M^{coH}{\otimes}B{\otimes}B\rightarrow M^{coH}{\otimes}_{B^{coH}} B$. Then, $$n_{M^{coH}}{\circ}(\phi_{M^{coH}}{\otimes}\mu_{B})\stackrel{{\scriptsize \blue (\ref{coeq-1})}}{=}
n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}\mu_{B})))\stackrel{{\scriptsize \blue (\ref{strong2})}}{=}n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}((\mu_{B}{\circ}(i_{B}{\otimes}B)){\otimes}B))),$$ and, as a consequence, there exists a unique morphism $$\phi_{M^{coH}{\otimes}_{B^{coH}} B}:M^{coH}{\otimes}_{B^{coH}} B {\otimes}B\rightarrow M^{coH}{\otimes}_{B^{coH}} B$$ such that $$\label{phi-ten}
\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}B)=n_{M^{coH}}{\circ}(M^{co H}{\otimes}\mu_{B}).$$
On the other hand, consider the morphism $$(n_{M^{coH}}{\otimes}H){\circ}(M^{co H}{\otimes}\rho_{B}): M^{coH}{\otimes}B\rightarrow M^{coH}{\otimes}_{B^{coH}} B{\otimes}H.$$
Then, taking into account that, by (\[strong-1\]), $(B,\mu_{B},\rho_{B})$ is a strong $(H,B,h)$-Hopf module,
- $\hspace{0.38cm} (n_{M^{coH}}{\otimes}H){\circ}(\phi_{M^{co H}}{\otimes}\rho_{B})$
- $= (n_{M^{coH}}{\otimes}H){\circ}(M^{coH}{\otimes}((\mu_{B}{\otimes}H){\circ}(i_{B}{\otimes}\rho_{B})))$ [([(\[coeq-1\])]{})]{}
- $= (n_{M^{coH}}{\otimes}H){\circ}(M^{coH}{\otimes}(\rho_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}B))) $ [([(\[l-idem0\])]{})]{}.
Thus, there exists a unique morphism $$\rho_{M^{coH}{\otimes}_{B^{coH}} B}:M^{coH}{\otimes}_{B^{coH}} B \rightarrow M^{coH}{\otimes}_{B^{coH}} B{\otimes}H$$ such that $$\label{rho-ten}
\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}=(n_{M^{coH}}{\otimes}H){\circ}(M^{co H}{\otimes}\rho_{B}).$$
We proceed to show that $(M^{coH}{\otimes}_{B^{coH}} B, \phi_{M^{coH}{\otimes}_{B^{coH}} B}, \rho_{M^{coH}{\otimes}_{B^{coH}} B})$ is a strong $(H,B,h)$-Hopf module. Indeed, first note that by (\[rho-ten\]) and the comodule condition for $B$, $$(M^{coH}{\otimes}_{B^{coH}} B{\otimes}\varepsilon_{H}){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}=
(n_{M^{coH}}{\otimes}\varepsilon_{H}){\circ}(M^{co H}{\otimes}\rho_{B})=n_{M^{coH}}.$$
As a consequence, and using that $n_{M^{coH}}$ is a coequalizer, we obtain that $$(M^{coH}{\otimes}_{B^{coH}} B{\otimes}\varepsilon_{H}){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}=id_{M^{coH}{\otimes}_{B^{coH}} B}.$$
By similar arguments we have $$(\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\otimes}H){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}=
(n_{M^{coH}}{\otimes}H{\otimes}H){\circ}(M^{co H}{\otimes}((\rho_{B}{\otimes}H){\circ}\rho_{B}))$$ $$=(n_{M^{coH}}{\otimes}H{\otimes}H){\circ}(M^{co H}{\otimes}((B{\otimes}\delta_{H}){\circ}\rho_{B}))=(M^{coH}{\otimes}_{B^{coH}} B{\otimes}\delta_{H}){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}.$$
Therefore, the pair $(M^{coH}{\otimes}_{B^{coH}} B, \rho_{M^{coH}{\otimes}_{B^{coH}} B})$ is a right $H$-comodule and we have (d1) of Definition \[H-D-mod\].
On the other hand, by (\[phi-ten\]), $$\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}\eta_{B})=n_{M^{coH}},$$ and then, (d2-1) of Definition \[H-D-mod\], i.e., $\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}\eta_{B})=id_{M^{coH}{\otimes}_{B^{coH}} B}$, holds. Moreover,
- $\hspace{0.38cm} \phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}((\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}q_{B})){\otimes}B)$
- $= n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}((\mu_{B}{\circ}(B{\otimes}q_{B})){\otimes}B)))$ [([(\[phi-ten\])]{})]{}
- $=n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(B{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}B)))))$ [([(\[strong-1\])]{})]{}
- $=\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}B)))$ [([(\[phi-ten\])]{})]{}
and, as a consequence, $$\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}((\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}q_{B})){\otimes}B)$$ $$=\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}B)))$$ because $n_{M^{coH}}{\otimes}B{\otimes}B$ is a coequalizer. Therefore (d2-2) of Definition \[H-D-mod\] follows because $p_{B}$ is a projection.
The proof for (d2-3) of Definition \[H-D-mod\] is the following: Composing with the coequalizer $n_{M^{coH}}{\otimes}B$ we obtain that
- $\hspace{0.38cm} \rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}B)$
- $= (n_{M^{coH}}{\otimes}H){\circ}(M^{coH}{\otimes}(\rho_{B}{\circ}\mu_{B}))$ [([(\[phi-ten\]), (\[rho-ten\])]{})]{}
- $=(n_{M^{coH}}{\otimes}H){\circ}(M^{coH}{\otimes}(\mu_{B{\otimes}H}{\circ}(\rho_{B}{\otimes}\rho_{B}))) $ [([(\[chmagma\])]{})]{}
- $= (\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\otimes}\mu_{H}){\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}c_{H,B}{\otimes}H){\circ}((\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}){\otimes}\rho_{B})$ [([(\[phi-ten\]), (\[rho-ten\])]{})]{}
and thus $$\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}\phi_{M^{coH}{\otimes}_{B^{coH}} B}=(\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\otimes}\mu_{H}){\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\otimes}\rho_{B}).$$
Using that $-{\otimes}H$ preserves coequalizers, the equality (d2-4) of Definition \[H-D-mod\] follows from
- $\hspace{0.38cm} \phi_{M^{coH}{\otimes}_{B^{coH}} B}\circ (\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\otimes}B){\circ}(n_{M^{coH}}{\otimes}((h{\otimes}(h{\circ}\lambda_{H})){\circ}\delta_{H}))$
- $= n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}\circ (\mu_{B}{\otimes}B){\circ}(B{\otimes}((h{\otimes}(h{\circ}\lambda_{H})){\circ}\delta_{H}))))$ [([(\[phi-ten\])]{})]{}
- $=n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{L}))))$ [([(c1) of Definition \[anchor\]]{})]{}
- $= \phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}(h{\circ}\Pi_{H}^{L}))$ [([(\[phi-ten\])]{})]{}.
Also, we have
- $\hspace{0.38cm} \phi_{M^{coH}{\otimes}_{B^{coH}} B}\circ ((\phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(M^{coH}{\otimes}_{B^{coH}} B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (n_{M^{coH}}{\otimes}\delta_{H})$
- $= n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}\circ ((\mu_{B}{\circ}(B{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (B{\otimes}\delta_{H})))$ [([(\[phi-ten\])]{})]{}
- $=n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(B{\otimes}(h{\circ}\Pi_{H}^{R}))))$ [([(c2) of Definition \[anchor\]]{})]{}
- $= \phi_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}(n_{M^{coH}}{\otimes}(h{\circ}\Pi_{H}^{R}))$ [([(\[phi-ten\])]{})]{},
and then (d2-5) of Definition \[H-D-mod\] holds.
Consider the morphism $\phi_{M}{\circ}(i_{M}{\otimes}B):M^{coH}{\otimes}B\rightarrow M$. By (\[PhiBcoH\]) and (d2-2) of Definition \[H-D-mod\], we obtain that $\phi_{M}{\circ}(i_{M}{\otimes}B){\circ}T_{M}^{1}=\phi_{M}{\circ}(i_{M}{\otimes}B){\circ}T_{M}^{2}$ and, as a consequence, there exists a unique morphism $\omega_{M}:M^{coH}{\otimes}_{B^{coH}} B\rightarrow M$ such that $$\label{omeg}
\omega_{M}{\circ}n_{M^{coH}}=\phi_{M}{\circ}(i_{M}{\otimes}B).$$
The morphism $\omega_{M}$ is a morphism of right $H$-comodules because $$(\omega_{M}{\otimes}H){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}{\circ}n_{M^{coH}}\stackrel{{\scriptsize \blue (\ref{rho-ten})}}{=}((\omega_{M}{\circ}n_{M^{coH}}){\otimes}H){\circ}(M^{coH}{\otimes}\rho_{B})
\stackrel{{\scriptsize \blue (\ref{omeg})}}{=}(\phi_{M}{\otimes}H){\circ}(i_{M}{\otimes}\rho_{B})$$$$\stackrel{{\scriptsize \blue (\ref{l-idem0})}}{=}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)\stackrel{{\scriptsize \blue (\ref{omeg})}}{=}\rho_{M}{\circ}\omega_{M}{\circ}n_{M^{coH}}$$ and then $(\omega_{M}{\otimes}H){\circ}\rho_{M^{coH}{\otimes}_{B^{coH}} B}=\rho_{M}{\circ}\omega_{M}.$
Finally, $\omega_{M}$ is an isomorphism with inverse $\omega_{M}^{\prime}=n_{M^{coH}}{\circ}(p_{M}{\otimes}h){\circ}\rho_{M}$. Indeed, firstly note that, by (\[l-idem4\]), we have that $\omega_{M}{\circ}\omega_{M}^{\prime}=id_{M}.$ On the other hand, composing with $n_{M^{coH}}$ we have
- $\hspace{0.38cm}\omega_{M}^{\prime} {\circ}\omega_{M} \circ n_{M^{coH}}$
- $= n_{M^{coH}}{\circ}(p_{M}{\otimes}h){\circ}\rho_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B) $ [([(\[omeg\])]{})]{}
- $= n_{M^{coH}}{\circ}((p_{M}{\circ}\phi_{M}{\circ}(M{\otimes}q_{B})){\otimes}h){\circ}(i_{M}{\otimes}\rho_{B})$ [ ([(\[l-idem1-1\])]{})]{}
- $=n_{M^{coH}}{\circ}(( \phi_{M^{coH}}{\circ}(M^{coH}{\otimes}p_{B})){\otimes}h){\circ}(M^{coH}{\otimes}\rho_{B}) $ [([(\[PhiBcoH1\])]{})]{}
- $= n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}h){\circ}\rho_{B})) $ [([(\[coeq-1\])]{})]{}
- $= n_{M^{coH}} $ [([(\[anchor21\])]{})]{},
and, as a consequence, $\omega_{M}^{\prime} {\circ}\omega_{M}=id_{M^{coH}{\otimes}_{B^{coH}} B}$.
Note that in the previous proposition the existence of the comodule structure on $M^{coH}{\otimes}_{B^{coH}} B$ does not depend on the preservation of coequalizers by the functors $-{\otimes}B$ and $-{\otimes}H$.
\[isomorp\] Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism and let $(P, \phi_{P}, \rho_{P})$, $(Q, \phi_{Q}, \rho_{Q})$ be strong $(H,B,h)$-Hopf modules. If there exists a right $H$-comodule isomorphism $\omega :Q\rightarrow P$, the triple $(P,\phi_{P}^{\omega}=\omega{\circ}\phi_{Q}{\circ}(\omega^{-1}{\otimes}B), \rho_{P})$, called the $\omega$-deformation of $(P, \phi_{P}, \rho_{P})$, is a strong $(H,B,h)$-Hopf module.
The proof follows easily because, if $\omega$ is a right $H$-comodule isomorphism, $\rho_{P}=(\omega{\otimes}H){\circ}\rho_{Q}{\circ}\omega^{-1}$.
Let $H$ be a weak Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma. Let $h:H\rightarrow B$ be an anchor morphism. Assume that (\[strong-1\]) and (\[strong2\]) hold. Then if the category ${\mathcal C}$ admits coequalizers and the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers, we define the category of strong $(H,B,h)$-Hopf modules as the one whose objects are strong $(H,B,h)$-Hopf modules, and whose morphisms $f:M\rightarrow N$ are morphisms of right $H$-comodules and $B$-quasilinear, i.e. $$\label{quasilineal}
\phi_{N}^{\omega_{N}}{\circ}(f{\otimes}B)=f{\circ}\phi_{M}^{\omega_{M}},$$ where $\omega_{M}:M^{coH}{\otimes}_{B^{coH}} B\rightarrow M$, $\omega_{N}:N^{coH}{\otimes}_{B^{coH}} B\rightarrow N$ are the isomorphisms of right $H$-comodules obtained in the proof of Proposition \[tensor\]. This category will be denoted by ${\mathcal S}{\mathcal M}^{H}_{B}(h)$.
As particular instances of ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ we will describe in detail some interesting examples of Hopf module categories associated to Hopf algebras, weak Hopf algebras, Hopf quasigroups and weak Hopf quasigroups.
: If $H$ is a Hopf algebra and $B=H$, $\rho_{B}=\delta_{H}$, $h=id_{H}$, the equalities (\[strong-1\]) and (\[strong2\]) hold trivially because the product on $H$ is associative. In this case the category ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ is the one whose objects are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}H\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{H})=id_{M},$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,H}{\otimes}H){\circ}(\rho_{M}{\otimes}\delta_{H})$,
- $\phi_{M}\circ (\phi_{M}{\otimes}\lambda_H))\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}\lambda_{H})){\otimes}H)\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
because in this case $\Pi_{H}^{L}=\Pi_{H}^{R}=q_{H}=\eta_{H}{\otimes}\varepsilon_{H}$, $H^{coH}=K$, and $i_{H}=\eta_{H}$. Then, if $(M, \phi_{M}, \rho_{M})$ is a classical Hopf module in the sense of Larson and Sweedler [@Lar-Sweed] (see also [@Sweedler]), we have that $(M, \phi_{M}, \rho_{M})$ is an object in ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ because the identity $$\label{LS-asocH}
\phi_{M}{\circ}(\phi_{M}{\otimes}H)=\phi_{M}{\circ}(M{\otimes}\mu_{H})$$ holds. Moreover, the morphisms of ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ are morphisms of right $H$-comodules and $H$-quasilinear, i.e., satisfying (\[quasilineal\]), where $\omega_{M}=\phi_{M}{\circ}(i_{M}{\otimes}H):M^{co H}{\otimes}H\rightarrow M$ is the associated isomorphism with inverse $\omega_{M}^{-1}=(p_{M}{\otimes}H){\circ}\rho_{M}$, and $\phi_{M^{coH}{\otimes}H}=M^{coH}{\otimes}\mu_{H}$ and $\rho_{M^{coH}{\otimes}H}=M^{co H}{\otimes}\delta_{H}$. Then any morphism $f:M\rightarrow N$ of right $H$-comodules and linear in the classical sense, i.e., such that $f{\circ}\phi_{M}=\phi_{N}{\circ}(f{\otimes}H)$, satisfies (\[quasilineal\]). Therefore, the Larson-Sweedler category of Hopf modules, denoted by ${\mathcal M}^{H}_{H}$, is a subcategory of ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$.
Also, in the Hopf algebra setting we have the following more general example: Let $(B,\rho_{B})$ be a right $H$-comodule monoid such that the functors $-{\otimes}B$, $-{\otimes}H$ preserve coequalizers. In this case, the equalities (\[strong-1\]) and (\[strong2\]) hold trivially because the product on $B$ is associative. For any $h$ multiplicative total integral, therefore an anchor morphism, the category ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ is the one whose objects are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}B\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{B})=id_{M},$
- $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{B})){\otimes}B)=\phi_{M}{\circ}(M{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))),$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M}{\otimes}\rho_{B})$,
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(M{\otimes}\delta_{H})
=M{\otimes}\varepsilon_{H},$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
because in this case $\Pi_{H}^{L}=\Pi_{H}^{R}=\eta_{H}{\otimes}\varepsilon_{H}$. The morphisms in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ are morphisms of right $H$-comodules and $B$-quasilinear where $\omega_{M}:M^{{\circ}H}{\otimes}_{B^{coH}} H\rightarrow M$ is the associated isomorphism of right $H$-comodules defined in the proof of Proposition \[tensor\]. Then, if $(M, \phi_{M}, \rho_{M})$ is an $(H,B)$-Hopf module in the sense of Doi [@Doi83], we have that $(M, \phi_{M}, \rho_{M})$ is an object in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ because the identity $$\label{LS-asocB}
\phi_{M}{\circ}(\phi_{M}{\otimes}B)=\phi_{M}{\circ}(M{\otimes}\mu_{B})$$ holds. The morphisms in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ are colinear morphisms satisfying (\[quasilineal\]), for the action $\phi_{M^{{\circ}H}{\otimes}_{B^{coH}} H}$ and the coaction $\rho_{M^{{\circ}H}{\otimes}_{B^{coH}} H}$ defined in the proof of Proposition \[tensor\]. Then any morphism $f:M\rightarrow N$ of right $H$-comodules and right $B$-linear, i.e., such that $f{\circ}\phi_{M}=\phi_{N}{\circ}(f{\otimes}B)$, satisfies (\[quasilineal\]). Therefore, the category of right $(H,B)$-Hopf modules, denoted by ${\mathcal M}^{H}_{B}$, is a subcategory of ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ (for all multiplicative total integral $h$).
: Let $H$ be a weak Hopf algebra. Let $(B,\rho_{B})$ be a right $H$-comodule monoid such that the functors $-{\otimes}B$, $-{\otimes}H$ preserve coequalizers. As in the Hopf algebra setting, the equalities (\[strong-1\]) and (\[strong2\]) hold trivially because the product on $B$ is associative. Also, in this case, a multiplicative total integral $h$ is an anchor morphism because the product on $B$ is associative, (\[idemp-2\]) holds and $\Pi_{H}^{L}\ast id_{H}=id_{H}$. Then, the category ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ is the one whose objects are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}B\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{B})=id_{M},$
- $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{B})){\otimes}B)=\phi_{M}{\circ}(M{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))),$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M}{\otimes}\rho_{B})$,
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{L})),$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}(h{\circ}\Pi_{H}^{R})).$
Then, if $(M, \phi_{M}, \rho_{M})$ is an $(H,B)$-Hopf module in the sense Böhm [@bohm2] (see also [@ZZ]), the triple $(M, \phi_{M}, \rho_{M})$ is an object in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ because the identity (\[LS-asocB\]) holds.
As in the Hopf case, the morphisms in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ are colinear morphisms satisfying (\[quasilineal\]), for $\phi_{M^{{\circ}H}{\otimes}_{B^{coH}} H}$ and $\rho_{M^{{\circ}H}{\otimes}_{B^{coH}} H}$ the action and the coaction defined in the proof of Proposition \[tensor\]. Then any morphism $f:M\rightarrow N$ of right $H$-comodules and right $B$-linear satisfies (\[quasilineal\]). Therefore, as in the classical context, the category of right $(H,B)$-Hopf modules, denoted by ${\mathcal M}^{H}_{B}$, is a subcategory of ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ (for all multiplicative total integral $h$). As a consequence, if $H=B$, $\rho_{B}=\delta_{H}$ and $h=id_{H}$, we obtain that the category ${\mathcal M}^{H}_{H}$, i.e., the category of Hopf modules associated to $H$, is a subcategory of ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$. Note that in this case the objects of ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}H\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{H})=id_{M},$
- $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{L})){\otimes}H)=\phi_{M}{\circ}(M{\otimes}(\mu_{H}{\circ}(i_{L}{\otimes}H))),$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,H}{\otimes}H){\circ}(\rho_{M}{\otimes}\delta_{H})$,
- $\phi_{M}\circ (\phi_{M}{\otimes}\lambda_H)\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}\Pi_{H}^{L}).$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}\lambda_{H})){\otimes}H)\circ (M{\otimes}\delta_{H})=\phi_{M}{\circ}(M{\otimes}\Pi_{H}^{R}),$
because in this case $B^{co H}=H_{L}$ and $i_{B}=i_{L}$.
: The following example comes from the nonassociative setting. Let $H$ be a Hopf quasigroup and assume that $B=H$, $\rho_{B}=\delta_{H}$, $h=id_{H}$. In this case the equalities (\[strong-1\]) and (\[strong2\]) hold because $q_{H}=\eta_{H}{\otimes}\varepsilon_{H}$, $H^{coH}=K$, and $i_{H}=\eta_{H}$. Then the category ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ is the one whose objects are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}H\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{H})=id_{M},$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,H}{\otimes}H){\circ}(\rho_{M}{\otimes}\delta_{H}),$
- $\phi_{M}\circ (\phi_{M}{\otimes}(M{\circ}\lambda_H))\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
- $\phi_{M}\circ (\phi_{M}{\otimes}\lambda_{H})\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H}.$
Note that in this setting the equality $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{H})){\otimes}H)=\phi_{M}{\circ}(M{\otimes}(\mu_{H}{\circ}(i_{H}{\otimes}H)))$ holds because $i_{H}=\eta_{H}$. The morphisms of ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ are morphisms of right $H$-comodules and $H$-quasilinear, i.e., satisfying (\[quasilineal\]), where $\omega_{M}=\phi_{M}{\circ}(i_{M}{\otimes}H):M^{co H}{\otimes}H\rightarrow M$ is the associated isomorphism with inverse $\omega_{M}^{-1}=(p_{M}{\otimes}H){\circ}\rho_{M}$, and $\phi_{M^{coH}{\otimes}H}=M^{coH}{\otimes}\mu_{H}$ and $\rho_{M^{coH}{\otimes}H}=M^{co H}{\otimes}\delta_{H}$. Therefore ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ is the category of Hopf modules introduced by Brzeziński in [@Brz].
In the previous nonassociative setting, let $(B,\rho_{B})$ be a right $H$-comodule magma such that the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers, and such that the equalities (\[strong-1\]) and (\[strong2\]) hold. For any anchor morphism $h$, the category ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ is the one whose objects are triples $(M, \phi_{M}, \rho_{M})$ where:
- The pair $(M, \rho_{M})$ is a right $H$-comodule.
- The morphism $\phi_{M}:M{\otimes}B\rightarrow M$ satisfies:
- $\phi_{M}{\circ}(M{\otimes}\eta_{B})=id_{M},$
- $\phi_{M}{\circ}((\phi_{M}{\circ}(M{\otimes}i_{B})){\otimes}B)=\phi_{M}{\circ}(M{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))),$
- $\rho_{M}{\circ}\phi_{M}=(\phi_{M}{\otimes}\mu_{H}){\circ}(M{\otimes}c_{H,B}{\otimes}H){\circ}(\rho_{M}{\otimes}\rho_{B})$,
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}h)){\otimes}(h{\circ}\lambda_{H})){\circ}(M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
- $\phi_{M}\circ ((\phi_{M}{\circ}(M{\otimes}(h{\circ}\lambda_{H}))){\otimes}h)\circ (M{\otimes}\delta_{H})=M{\otimes}\varepsilon_{H},$
because in this case $\Pi_{H}^{L}=\Pi_{H}^{R}=\eta_{H}{\otimes}\varepsilon_{H}$. The morphisms of ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ are morphisms of right $H$-comodules and $B$-quasilinear where $\omega_{M}:M^{co H}{\otimes}_{B^{coH}} H\rightarrow M$ is the associated isomorphism of right $H$-comodules defined in the proof of Proposition \[tensor\].
: If $H$ is a weak Hopf quasigroup ${\mathcal S}{\mathcal M}^{H}_{H}(id_{H})$ is the category of strong Hopf modules defined in [@Strong] and denoted by ${\mathcal S}{\mathcal M}^{H}_{H}$. Note that, as a consequence of (\[d2-7\]), we obtain that $$\phi_{M}\circ (\phi_{M} {\otimes}H){\circ}(M{\otimes}\Pi_{H}^{L}{\otimes}H)\circ (M{\otimes}\delta_{H})=\phi_{M}$$ is a superfluous identity in the definition of strong Hopf module introduced in [@Strong].
\[struc-tim-2\] Assume that the conditions of Proposition \[tensor\] hold. Let $(M, \phi_{M}, \rho_{M})$ be an object in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$. Let $\omega_{M}$ be the isomorphism of right $H$-comodules between $M^{coH}{\otimes}_{B^{coH}}B$ and $M$. Then the triple $(M, \phi_{M}^{\omega_{M}}, \rho_{M})$ is a strong $(H,B,h)$-Hopf module with the same object of coinvariants that $(M, \phi_{M}, \rho_{M})$. Also, the identity $$\label{action-w}
\phi_{M}^{\omega_{M}}=\phi_{M}{\circ}(q_{M}{\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B)$$ holds and $$\label{q-w}
q_{M}^{\omega_{M}}=q_{M},$$ where $q_{M}^{\omega_{M}}=\phi_{M}^{\omega_{M}}{\circ}(M{\otimes}(h{\circ}\lambda_{H})){\circ}\rho_{M}$ is the idempotent morphism associated to the Hopf module $(M, \phi_{M}^{\omega_{M}}, \rho_{M})$. Moreover, for $(M, \phi_{M}^{\omega_{M}}, \rho_{M})$, the associated isomorphism of right $H$-comodules between $M^{coH}{\otimes}_{B^{coH}}B$ and $M$ is $\omega_{M}$, and the equality $$\label{phi-square}
(\phi_{M}^{\omega_{M}})^{\omega_{M}}=
\phi_{M}^{\omega_{M}}$$ holds. Finally, there exists an idempotent functor $$D:{\mathcal S}{\mathcal M}^{H}_{B}(h)\rightarrow {\mathcal S}{\mathcal M}^{H}_{B}(h),$$ called the $\omega$-deformation functor, defined on objects by $D((M,\phi_{M},\rho_{M}))=(M, \phi_{M}^{\omega_{M}}, \rho_{M})$ and on morphisms by the identity.
By Proposition \[isomorp\], $(M, \phi_{M}^{\omega_{M}}, \rho_{M})$ is a strong $(H,B,h)$-Hopf module. Moreover, note that (\[action-w\]) follows by (\[phi-ten\]) and (\[omeg\]). Then $q_{M}^{\omega_{M}}=q_{M}$ because
- $\hspace{0.38cm} q_{M}^{\omega_{M}}$
- $=\phi_{M}{\circ}(q_{M}{\otimes}(\mu_{B}{\circ}(h{\otimes}h))){\circ}(\rho_{M}{\otimes}\lambda_{H}){\circ}\rho_{M} $ [([(\[action-w\])]{})]{}
- $= \phi_{M}{\circ}(q_{M}{\otimes}(h{\circ}\Pi_{H}^{L})){\circ}\rho_{M} $ [([comodule condition for $M$ and multiplicative condition for $h$]{})]{}
- $= \phi_{M}\circ (\phi_{M}{\otimes}B){\circ}(q_{M}{\otimes}h{\otimes}(h{\circ}\lambda_{H})){\circ}(M{\otimes}\delta_{H}){\circ}\rho_{M} $ [([(d2-4) of Definition \[H-D-mod\]]{})]{}
- $= \phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}h){\circ}\rho_{M}){\otimes}(h{\circ}\lambda_{H})){\circ}\rho_{M} $ [([comodule condition for $M$]{})]{}
- $= q_{M} $[([(\[l-idem4\])]{})]{}.
Therefore, $(M, \phi_{M}^{\omega_{M}}, \rho_{M})$ has the same object of coinvariants that $(M, \phi_{M}, \rho_{M})$. On the other hand,
- $\hspace{0.38cm} (\phi_{M}^{\omega_{M}})^{\omega_{M}} $
- $= \phi_{M}{\circ}(q_{M}{\otimes}\mu_{B}){\circ}(((M{\otimes}h){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(\[action-w\])]{})]{}
- $= \phi_{M}{\circ}(q_{M}{\otimes}\mu_{B}){\circ}(((M{\otimes}(h{\circ}\Pi_{H}^{L})){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(\[idemp-0M\])]{})]{}
- $=\phi_{M}{\circ}(q_{M}{\otimes}\mu_{B}){\circ}(((M{\otimes}(q_{B}{\circ}h)){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(\[anchor1\])]{})]{}
- $= \phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}(q_{B}{\circ}h)){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(d2-2) of Definition \[H-D-mod\]]{})]{}
- $= \phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}(h{\circ}\Pi_{H}^{L})){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(\[anchor1\])]{})]{}
- $=\phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}h){\circ}\rho_{M}{\circ}q_{M}){\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{M}{\otimes}B) $ [([(\[idemp-0M\])]{})]{}
- $=\phi_{M}^{\omega_{M}}$ [([(\[l-idem4\])]{})]{}.
Finally, by (\[q-w\]) and (\[phi-square\]), it is easy to show that $D$ is a well defined idempotent endofunctor.
Assume that the conditions of Proposition \[tensor\] hold. Let $(M, \phi_{M}, \rho_{M})$ be an object in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$. The following identity holds: $$\label{iphi}
\phi_{M}^{\omega_{M}}{\circ}(i_{M}{\otimes}B)=\phi_{M}{\circ}(i_{M}{\otimes}B).$$
Indeed,
- $\hspace{0.38cm} \phi_{M}^{\omega_{M}}{\circ}(i_{M}{\otimes}B) $
- $= \phi_{M}{\circ}(q_{M}{\otimes}\mu_{B}){\circ}(((M{\otimes}(h{\circ}\Pi_{H}^{L})){\circ}\rho_{M}{\circ}i_{M}){\otimes}B) $ [([(\[idemp-0M\])]{})]{}
- $= \phi_{M}{\circ}(q_{M}{\otimes}\mu_{B}){\circ}(((M{\otimes}(q_{B}{\circ}h)){\circ}\rho_{M}{\circ}i_{M}){\otimes}B) $ [([(\[anchor1\])]{})]{}
- $= \phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}(q_{B}{\circ}h)){\circ}\rho_{M}{\circ}i_{M}){\otimes}B) $ [([(d2-2) of Definition \[H-D-mod\]]{})]{}
- $=\phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}(h{\circ}\Pi_{H}^{L})){\circ}\rho_{M}{\circ}i_{M}){\otimes}B) $ [([(\[anchor1\])]{})]{}
- $= \phi_{M}{\circ}((\phi_{M}{\circ}(q_{M}{\otimes}h){\circ}\rho_{M}{\circ}i_{M}){\otimes}B) $ [([(\[idemp-0M\])]{})]{}
- $=\phi_{M}{\circ}(i_{M}{\otimes}B)$ [([(\[l-idem4\])]{})]{}.
\[inv-defor\] Assume that the conditions of Proposition \[tensor\] hold. For any object $(M, \phi_{M}, \rho_{M})$ in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$, the strong $(H,B,h)$-Hopf module $(M^{coH}{\otimes}_{B^{coH}}B, \phi_{M^{coH}{\otimes}_{B^{coH}}B}, \rho_{M^{coH}{\otimes}_{B^{coH}}B})$, constructed in Proposition \[tensor\], is invariant for the $\omega$-deformation functor, i.e., $$D((M^{coH}{\otimes}_{B^{coH}}B, \phi_{M^{coH}{\otimes}_{B^{coH}}B}, \rho_{M^{coH}{\otimes}_{B^{coH}}B}))=(M^{coH}{\otimes}_{B^{coH}}B, \phi_{M^{coH}{\otimes}_{B^{coH}}B}, \rho_{M^{coH}{\otimes}_{B^{coH}}B}).$$
To prove the proposition we only need to show that $$\label{w-MB-inv}
\phi_{M^{coH}{\otimes}_{B^{coH}}B}^{\omega_{M^{coH}{\otimes}_{B^{coH}}B}}=\phi_{M^{coH}{\otimes}_{B^{coH}}B}.$$
Indeed, composing with the coequalizer $n_{M^{coH}}{\otimes}B$ we have $$\phi_{M^{coH}{\otimes}_{B^{coH}}B}^{\omega_{M^{coH}{\otimes}_{B^{coH}}B}}{\circ}(n_{M^{coH}}{\otimes}B)\stackrel{{\scriptsize \blue (\ref{rho-ten}), (\ref{phi-ten})}}{=} n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}(\mu_{B}{\circ}(h{\otimes}B))){\circ}(\rho_{B}{\otimes}B)))$$ $$\stackrel{{\scriptsize \blue (\ref{strong2})}}{=} n_{M^{coH}}{\circ}(M^{coH}{\otimes}(\mu_{B}{\circ}((\mu_{B}{\circ}(q_{B}{\otimes}h){\circ}\rho_{B}){\otimes}B)))\stackrel{{\scriptsize \blue (\ref{anchor21})}}{=} n_{M^{coH}}{\circ}(M^{coH}{\otimes}\mu_{B})\stackrel{{\scriptsize \blue (\ref{phi-ten})}}{=}\phi_{M^{coH}{\otimes}_{B^{coH}}B}{\circ}(n_{M^{coH}}{\otimes}B).$$ Therefore, (\[w-MB-inv\]) holds.
The following result is the nonassociative general version of the Fundamental Theorem of Hopf Modules. The proof follows by the properties of the morphism $\omega_{M}$, obtained in Proposition \[tensor\], and by (\[w-MB-inv\]).
([Fundamental Theorem of Hopf modules]{}) \[main0\] Assume that the conditions of Proposition \[tensor\] hold. Let $(M, \phi_{M}, \rho_{M})$ be an object in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$. The objects $(M^{coH}{\otimes}_{B^{coH}}B, \phi_{M^{coH}{\otimes}_{B^{coH}}B}, \rho_{M^{coH}{\otimes}_{B^{coH}}B})$ and $(M, \phi_{M}, \rho_{M})$ are isomorphic in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$.
The previous theorem is a generalization of the one proved by Sweedler [@Sweedler] for Hopf modules over an ordinary Hopf algebra. It also contains the Fundamental Theorem of relative Hopf modules (or $(H,B)$-Hopf modules, or Doi-Hopf modules) given by Doi and Takeuchi in [@Doi-Take]. On the other hand, in the weak setting, Theorem \[main0\] is a generalization of the one obtained by Böhm, Nill and Szlachányi [@bohm], for Hopf modules over a weak Hopf algebra $H$, and the one proved by Zhang and Zhu [@ZZ], for $(H,B)$-Hopf modules associated to a weak right $H$-comodule algebra $B$. Moreover, in the nonassociative context, it generalizes the result obtained by Brzeziński [@Brz] for Hopf modules associated to a Hopf quasigroup. Finally, for weak Hopf quasigroups, Theorem \[main0\] is a generalization of the Fundamental Theorem of Hopf modules proved in [@Asian] (see also [@Strong]).
Categorical equivalences for strong $(H,B,h)$-Hopf modules
==========================================================
As for prerequisites, in this section we will assume that the conditions of Proposition \[tensor\] hold. Then, in the following $H$ is a weak Hopf quasigroup and $(B,\rho_{B})$ is a right $H$-comodule magma. Also, $h:H\rightarrow B$ is an anchor morphism such that (\[strong-1\]) and (\[strong2\]) hold. Finally, the category ${\mathcal C}$ admits coequalizers and the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers. With ${\mathcal C}_{B^{coH}}$ we will denote the category of right $B^{coH}$-modules.
The main target of this section is to prove that there exists an equivalence between ${\mathcal C}_{B^{coH}}$ and the category of strong $(H,B,h)$-Hopf modules.
Let $(N,\psi_{N})$ be an object in ${\mathcal C}_{B^{coH}}$ and consider the coequalizer diagram
$$\label{coeqN-1}
\setlength{\unitlength}{1mm}
\begin{picture}(101.00,10.00)
\put(20.00,8.00){\vector(1,0){25.00}}
\put(20.00,4.00){\vector(1,0){25.00}}
\put(62.00,6.00){\vector(1,0){21.00}}
\put(32.00,11.00){\makebox(0,0)[cc]{$\psi_{N}{\otimes}B$ }}
\put(33.00,0.5){\makebox(0,0)[cc]{$N{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B))$}}
\put(73.00,9.00){\makebox(0,0)[cc]{$n_{N}$ }}
\put(7.00,6.00){\makebox(0,0)[cc]{$ N\otimes B^{coH}{\otimes}B$ }}
\put(54.00,6.00){\makebox(0,0)[cc]{$ N{\otimes}B$ }}
\put(96.00,6.00){\makebox(0,0)[cc]{$N{\otimes}_{B^{coH}} B. $ }}
\end{picture}$$
Then, $$(n_{N}{\otimes}B){\circ}(\psi_{N}{\otimes}\rho_{B})\stackrel{{\scriptsize \blue (\ref{coeqN-1})}}{=}(n_{N}{\otimes}B){\circ}(N{\otimes}((\mu_{B}{\otimes}B){\circ}(i_{B}{\otimes}\rho_{B})))\stackrel{{\scriptsize \blue (\ref{l-idem0})}}{=}
(n_{N}{\otimes}B){\circ}(N{\otimes}(\rho_{B}{\circ}(\mu_{B}{\circ}(i_{B}{\otimes}B))))$$ and, as a consequence, there exists a unique morphism $\rho_{N{\otimes}_{B^{coH}} B}:N{\otimes}_{B^{coH}} B\rightarrow N{\otimes}_{B^{coH}} B{\otimes}H$ such that $$\label{comodN}
\rho_{N{\otimes}_{B^{coH}} B}{\circ}n_{N}=(n_{N}{\otimes}B){\circ}(N{\otimes}\rho_{B}).$$
On the other hand, by (\[strong2\]), we have $$n_{N}{\circ}(\psi_{N}{\otimes}\mu_{B})=n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}\mu_{B})))=n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}((\mu_{B}{\circ}(i_{B}{\otimes}B)){\otimes}B))),$$ and then,using that the functor $-{\otimes}B$ preserves coequalizers, there exists a unique morphism $$\phi_{N{\otimes}_{B^{coH}} B}:N{\otimes}_{B^{coH}} B{\otimes}B\rightarrow N{\otimes}_{B^{coH}} B$$ such that $$\label{quasi-modN}
\phi_{N{\otimes}_{B^{coH}} B}{\circ}(n_{N}{\otimes}B)=n_{N}{\circ}(N{\otimes}\mu_{B}).$$
By a similar proof to the one used for $M^{coH}{\otimes}_{B^{coH}} B$ in Proposition \[tensor\], we can prove that $$(N{\otimes}_{B^{coH}} B, \phi_{N{\otimes}_{B^{coH}} B}, \rho_{N{\otimes}_{B^{coH}} B})$$ is a strong $(H,B,h)$-Hopf module such that (the proof follows the ideas given in Proposition \[inv-defor\] for $M^{coH}{\otimes}_{B^{coH}} B$) $$\label{w-d2}
\phi_{N{\otimes}_{B^{coH}} B}^{\omega_{N{\otimes}_{B^{coH}} B}}=\phi_{N{\otimes}_{B^{coH}} B}.$$
On the other hand, if $f:N\rightarrow P$ is a morphism in ${\mathcal C}_{B^{coH}}$, we have $$n_{P}{\circ}(f{\otimes}B){\circ}(\psi_{N}{\otimes}B)=n_{P}{\circ}(f{\otimes}B){\circ}(N{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}B)))$$ and then there exists a unique morphism $f{\otimes}_{B^{coH}}B:N{\otimes}_{B^{coH}} B\rightarrow P{\otimes}_{B^{coH}} B$ such that $$\label{mor-induc}
n_{P}{\circ}(f{\otimes}B)=(f{\otimes}_{B^{coH}}B){\circ}n_{N}.$$
The morphism $f{\otimes}_{B^{coH}}B$ is a morphism in ${\mathcal S}{\mathcal M}^{H}_{B}(h)$ because $$\rho_{P{\otimes}_{B^{coH}} B}{\circ}f{\otimes}_{B^{coH}} B {\circ}n_{N}\stackrel{{\scriptsize \blue (\ref{comodN}), (\ref{mor-induc})}}{=}(n_{P}{\otimes}B){\circ}(f{\otimes}\rho_{B})\stackrel{{\scriptsize \blue (\ref{comodN})}}{=}\rho_{P{\otimes}_{B^{coH}} B}{\circ}n_{P}{\circ}(f{\otimes}B)$$ $$\stackrel{{\scriptsize \blue (\ref{mor-induc})}}{=}(f{\otimes}_{B^{coH}} B{\otimes}B){\circ}\rho_{N{\otimes}_{B^{coH}} B}{\circ}n_{N}$$ and $$\phi_{P{\otimes}_{B^{coH}} B}^{\omega_{P{\otimes}_{B^{coH}} B}}{\circ}(f{\otimes}_{B^{coH}} B{\otimes}B){\circ}(n_{N}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{w-d2})}}{=}
\phi_{P{\otimes}_{B^{coH}} B}{\circ}(f{\otimes}_{B^{coH}} B{\otimes}B){\circ}(n_{N}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{quasi-modN}), (\ref{mor-induc})}}{=}
n_{P}{\circ}(f{\otimes}\mu_{B})$$ $$\stackrel{{\scriptsize \blue (\ref{mor-induc})}}{=}
f{\otimes}_{B^{coH}} B{\circ}n_{N}{\circ}(N{\otimes}\mu_{B})
\stackrel{{\scriptsize \blue (\ref{quasi-modN})}}{=}
f{\otimes}_{B^{coH}} B{\circ}\phi_{N{\otimes}_{B^{coH}} B}{\circ}(n_{N}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{w-d2})}}{=}
f{\otimes}_{B^{coH}} B{\circ}\phi_{N{\otimes}_{B^{coH}} B}^{\omega_{N{\otimes}_{B^{coH}} B}}{\circ}(n_{N}{\otimes}B).$$
Summarizing, we have the following proposition:
\[I-functor\] There exists a functor $F:{\mathcal C}_{B^{coH}}\rightarrow {\mathcal S}{\mathcal M}^{H}_{B}(h)$, called the induction functor, defined on objects by $F((N,\psi_{N}))=(N{\otimes}_{B^{coH}} B, \phi_{N{\otimes}_{B^{coH}} B}, \rho_{N{\otimes}_{B^{coH}} B})$ and on morphisms by $F(f)=f{\otimes}_{B^{coH}}B$.
Now let $(M,\phi_{M}, \rho_{M})$ be an object in $ {\mathcal S}{\mathcal M}^{H}_{B}(h)$. By Proposition \[M-coinv\] we have that the object of coinvariants $M^{coH}$ is a right $B^{coH}$-module where $\phi_{M^{coH} }=p_{M}{\circ}\phi_{M}{\circ}(i_{M}{\otimes}i_{B})$ (see (\[PhiBcoH1\])). Let $g:M\rightarrow Q$ be a morphism in $ {\mathcal S}{\mathcal M}^{H}_{B}(h)$. Using the comodule morphism condition we obtain that $\rho_{Q}\circ g\circ i_{M}=(Q{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{Q}\circ g\circ i_{M}$ and this implies that there exists a unique morphism $g^{coH}:M^{coH}\rightarrow Q^{coH}$ such that $$\label{coi-morph}
i_{Q}{\circ}g^{coH}=g{\circ}i_{M}.$$ Also, using that $g$ is $B$-quasilinear, $H$ colinear, and (\[q-w\]) we have $$\label{Bq-Hc}
g{\circ}q_{M}^{\omega_{M}}=g{\circ}q_{M}=q_{T}{\circ}g=q_{T}^{\alpha_{T}}{\circ}g.$$
Then, $$i_{Q}{\circ}g^{coH}{\circ}p_{M}
\stackrel{{\scriptsize \blue (\ref{coi-morph})}}{=}
g{\circ}q_{M}
\stackrel{{\scriptsize \blue (\ref{Bq-Hc})}}{=}
q_{Q}{\circ}g$$ and, as a consequence, $$\label{coinv-m-1}
g^{coH}{\circ}p_{M}=p_{Q}{\circ}g$$ holds.
The morphism $g^{coH}$ is a morphism in ${\mathcal C}_{B^{coH}}$ because
- $\hspace{0.38cm} \phi_{Q^{coH}}{\circ}(g^{coH}{\otimes}B^{coH}) $
- $= p_{Q}{\circ}\phi_{Q}{\circ}((i_{Q}{\circ}g^{coH}){\otimes}i_{B}) $ [([(\[PhiBcoH1\])]{})]{}
- $=p_{Q}{\circ}\phi_{Q}^{\omega_{Q}}{\circ}((i_{Q}{\circ}g^{coH}){\otimes}i_{B}) $ [([(\[iphi\])]{})]{}
- $= p_{Q}{\circ}\phi_{Q}^{\omega_{Q}}{\circ}((g{\circ}i_{M}){\otimes}i_{B}) $ [([(\[coi-morph\])]{})]{}
- $= p_{Q}{\circ}g{\circ}\phi_{M}^{\omega_{M}}{\circ}( i_{M}{\otimes}i_{B}) $ [([(\[quasilineal\])]{})]{}
- $= p_{Q}{\circ}g{\circ}\phi_{M}{\circ}( i_{M}{\otimes}i_{B}) $ [([(\[iphi\])]{})]{}
- $=g^{coH}{\circ}\phi_{M^{coH}}$ [([(\[coinv-m-1\])]{})]{}.
Thus, we have the following result.
\[c-functor\] There exists a functor $G:{\mathcal S}{\mathcal M}^{H}_{B}(h)\rightarrow {\mathcal C}_{B^{coH}},$ called the functor of coinvariants, defined on objects by $G((M,\phi_{M},\rho_{M}))=(M^{coH}, \psi_{M^{coH}})$ and on morphisms by $G(g)=g^{coH}$.
\[p-paper\] The categories ${\mathcal SM}^{H}_{B}(h)$ and ${\mathcal C}_{B^{coH}}$ are equivalent.
To prove the theorem, we firstly obtain that the induction functor $F$, introduced in Proposition \[I-functor\], is left adjoint to the functor of coinvariants $G$ introduced in Proposition \[c-functor\]. Later, we show that the unit and counit associated to this adjunction are natural isomorphisms. Then, we proceed as in the proof of Theorem 3.10 of [@Strong] by dividing the proof in three steps.
[**]{} In this step we define the unit of the adjunction. For any right $B^{coH}$-module $(N, \psi_{N})$, consider $$\alpha_{N}:N\rightarrow GF(N)=(N{\otimes}_{B^{coH}}B)^{coH}$$ as the unique morphism such that $$\label{unit}
i_{N{\otimes}_{B^{coH}}B}{\circ}\alpha_{N}=n_{N}{\circ}(N{\otimes}\eta_{B}).$$
This morphism exists and is unique because $$((N{\otimes}_{B^{coH}}B){\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{N{\otimes}_{B^{coH}}B}{\circ}n_{N}{\circ}(N{\otimes}\eta_{B})
\stackrel{{\scriptsize \blue (\ref{comodN})}}{=}
(n_{N}{\otimes}H){\circ}(N{\otimes}((B{\otimes}\overline{\Pi}_{H}^{R}){\circ}\rho_{B}{\circ}\eta_{B}))$$ $$\stackrel{{\scriptsize \blue ({\rm b5})}}{=}
(n_{N}{\otimes}H){\circ}(N{\otimes}(\rho_{B}{\circ}\eta_{B}))
\stackrel{{\scriptsize \blue (\ref{comodN})}}{=}
\rho_{N{\otimes}_{B^{coH}}B}{\circ}n_{N}{\circ}(N{\otimes}\eta_{B}).$$
Also, $\alpha_{N}$ is a morphism in ${\mathcal C}_{B^{coH}}$. Indeed: Composing with the equalizer $i_{N{\otimes}_{B^{coH}}B}$ we have
- $\hspace{0.38cm} i_{N{\otimes}_{B^{coH}}B} {\circ}\psi_{(N{\otimes}_{B^{coH}}B)^{coH}}{\circ}(\alpha_{N}{\otimes}B^{coH}) $
- $= q_{N{\otimes}_{B^{coH}}B} {\circ}\phi_{N{\otimes}_{B^{coH}}B}{\circ}((i_{N{\otimes}_{B^{coH}}B}{\circ}\alpha_{N}){\otimes}i_{B})$ [([(\[PhiBcoH1\])]{})]{}
- $= q_{N{\otimes}_{B^{coH}}B} {\circ}\phi_{N{\otimes}_{B^{coH}}B}{\circ}((n_{N}{\circ}(N{\otimes}\eta_{B})){\otimes}i_{B})$ [([(\[unit\])]{})]{}
- $= q_{N{\otimes}_{B^{coH}}B} {\circ}n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}(\eta_{B}{\otimes}i_{B})))$ [([(\[quasi-modN\])]{})]{}
- $= n_{N}{\circ}(N{\otimes}(q_{B}{\circ}i_{B}))$ [([properties of $\eta_{B}$, (\[comodN\]), and (\[quasi-modN\])]{})]{}
- $=n_{N}{\circ}(N{\otimes}i_{B}) $ [([properties of $i_{B}$]{})]{}
- $=n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}(i_{B}{\otimes}\eta_{B}))) $ [([properties of $\eta_{B}$]{})]{}
- $=n_{N}{\circ}(\psi_{N}{\otimes}\eta_{B}) $ [([(\[coeqN-1\])]{})]{}
- $=i_{N{\otimes}_{B^{coH}}B} {\circ}\alpha_{N}{\circ}\psi_{N} $ [([(\[unit\])]{})]{}
and therefore $\psi_{N{\otimes}_{B^{coH}}B}{\circ}(\alpha_{N}{\otimes}B^{coH})=\alpha_{N}{\circ}\psi_{N}.$ On the other hand, the morphism $\alpha_{N}$ is natural in $N$ because if $f:N\rightarrow P$ is a morphism in ${\mathcal C}_{B^{coH}}$ $$i_{P{\otimes}_{B^{coH}}B}{\circ}(f{\otimes}_{B^{coH}}B)^{coH}{\circ}\alpha_{N}
\stackrel{{\scriptsize \blue (\ref{coi-morph})}}{=}
(f{\otimes}_{B^{coH}}B){\circ}i_{N{\otimes}_{B^{coH}}B}{\circ}\alpha_{N}
\stackrel{{\scriptsize \blue (\ref{unit})}}{=}
(f{\otimes}_{B^{coH}}B){\circ}n_{N}{\circ}(N{\otimes}\eta_{B})$$ $$\stackrel{{\scriptsize \blue (\ref{mor-induc})}}{=}
n_{P}{\circ}(f{\otimes}\eta_{B})
\stackrel{{\scriptsize \blue (\ref{unit})}}{=}
i_{P{\otimes}_{B^{coH}}B}{\circ}\alpha_{P}{\circ}f,$$ and then $(f{\otimes}_{B^{coH}}B)^{coH}{\circ}\alpha_{N}=\alpha_{P}{\circ}f$.
Finally, we prove that $\alpha_{N}$ is an isomorphism for all right $B^{coH}$-module $N$. First note that, under the conditions of this theorem, the triple $(B, \mu_{B}, \delta_{B})$ is a strong $(H,B, h)$-Hopf module, and then $$\psi_{N}{\circ}(\psi_{N}{\otimes}p_{B})=\psi_{N}{\circ}(N{\otimes}(p_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}B)))$$ because
- $\hspace{0.38cm} \psi_{N}{\circ}(\psi_{N}{\otimes}p_{B})$
- $=\psi_{N}{\circ}(N{\otimes}(\mu_{B^{coH}}{\circ}(B^{coH}{\otimes}p_{B}))) $ [([module condition for $N$]{})]{}
- $=\psi_{N}{\circ}(N{\otimes}(p_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}q_{B})) $ [([(\[muH1\])]{})]{}
- $= \psi_{N}{\circ}(N{\otimes}(p_{B}{\circ}\mu_{B}{\circ}(i_{B}{\otimes}B))) $ [([(\[l-idem3\])]{})]{}.
Therefore, there exists a unique morphism $m_{N}:N{\otimes}_{B^{coH}}B\rightarrow N$ such that $$\label{mn}
m_{N}{\circ}n_{N}=\psi_{N}{\circ}(N{\otimes}p_{B}).$$
Now define the morphism $x_{N}:(N{\otimes}_{B^{coH}}B)^{coH}\rightarrow N$ by $x_{N}=m_{N}{\circ}i_{N{\otimes}_{B^{coH}}B}$. Then, $$x_{N}{\circ}\alpha_{N}
\stackrel{{\scriptsize \blue (\ref{unit})}}{=}
m_{N}{\circ}n_{N}{\circ}(N{\otimes}\eta_{B})
\stackrel{{\scriptsize \blue (\ref{mn})}}{=}
\psi_{N}{\circ}(N{\otimes}(p_{B}{\circ}\eta_{B}))
\stackrel{{\scriptsize \blue (\ref{etaH})}}{=}
\psi_{N}{\circ}(N{\otimes}\eta_{B^{coH}})=
id_{N}.$$
On the other hand, composing with $i_{N{\otimes}_{B^{coH}}B}$ and $p_{N{\otimes}_{B^{coH}}B}{\circ}n_{N}$ we have
- $\hspace{0.38cm} i_{N{\otimes}_{B^{coH}}B} {\circ}\alpha_{N}{\circ}x_{N}{\circ}p_{N{\otimes}_{B^{coH}}N}{\circ}n_{N}$
- $= n_{N}{\circ}((m_{N}{\circ}q_{N{\otimes}_{B^{coH}}B}{\circ}n_{N}){\otimes}\eta_{B}) $ [([(\[unit\])]{})]{}
- $= n_{N}{\circ}((\psi_{N}{\circ}(N{\otimes}(p_{B}{\circ}q_{B}))){\otimes}\eta_{B}) $ [([(\[comodN\]), (\[quasi-modN\]), (\[mn\])]{})]{}
- $=n_{N}{\circ}((\psi_{N}{\circ}(N{\otimes}p_{B})){\otimes}\eta_{B}) $ [([properties of $q_{B}$]{})]{}
- $= n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}(q_{B}{\otimes}\eta_{B}))) $ [([(\[coeqN-1\])]{})]{}
- $= n_{N}{\circ}(N{\otimes}q_{B}) $ [([properties of $\eta_{B}$]{})]{}
- $= q_{N{\otimes}_{B^{coH}}B} {\circ}n_{N} $ [([(\[comodN\]), (\[quasi-modN\])]{})]{}.
Therefore, $\alpha_{N}{\circ}x_{N}=id_{(N{\otimes}_{B^{coH}}B)^{coH}}$ and, as a consequence, $\alpha_{N}$ is an isomorphism.
[**]{} For any $(M,\phi_{M},\rho_{M})\in {\mathcal SM}^{H}_{B}(h)$ the counit is defined as $ \beta_{M}=\omega_{M}:M^{coH}{\otimes}_{B^{coH}}B\rightarrow M,$ where $\omega_{M}$ is the isomorphism satisfying (\[omeg\]). By Theorem \[main0\], we know that $\omega_{M}$ is an isomorphism in ${\mathcal SM}^{H}_{B}(h)$, and it is natural because if $g:M\rightarrow Q$ is a morphism in ${\mathcal SM}^{H}_{B}(h)$ we have $$\beta_{Q}{\circ}(g^{coH}{\otimes}_{B^{coH}} {\otimes}B){\circ}n_{M^{coH}}
\stackrel{{\scriptsize \blue (\ref{mor-induc})}}{=}
\beta_{Q}{\circ}n_{Q^{coH}}{\circ}(g^{coH}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{omeg})}}{=}
\phi_{Q}{\circ}((i_{Q}{\circ}g^{coH}){\otimes}B)
\stackrel{{\scriptsize \blue (\ref{iphi})}}{=}
\phi_{Q}^{\omega_{Q}}{\circ}((i_{Q}{\circ}g^{coH}){\otimes}B)$$ $$\stackrel{{\scriptsize \blue (\ref{coi-morph})}}{=}
\phi_{Q}^{\omega_{Q}}{\circ}((g{\circ}i_{Q}){\otimes}B)
\stackrel{{\scriptsize \blue (\ref{quasilineal})}}{=}
g{\circ}\phi_{M}^{\omega_{M}}{\circ}(i_{M}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{iphi})}}{=}
g{\circ}\phi_{M}{\circ}(i_{M}{\otimes}B)
\stackrel{{\scriptsize \blue (\ref{omeg})}}{=}
g{\circ}\beta_{M}{\circ}n_{M^{coH}}$$ and thus $\beta_{Q}{\circ}g^{coH}{\otimes}_{B^{coH}}{\otimes}B=g{\circ}\beta_{M}.$
[**]{} Now we prove the triangular identities for the unit and counit previously defined. Indeed: The first triangular identity holds because composing with $n_{N}$ we have
- $\hspace{0.38cm} \beta_{N{\otimes}_{B^{coH}}B}{\circ}(\alpha_{N}{\otimes}_{B^{coH}}B){\circ}n_{N} $
- $=\beta_{N{\otimes}_{B^{coH}}B}{\circ}n_{ (N{\otimes}_{B^{coH}}B)^{coH}}{\circ}(\alpha_{N}{\otimes}B) $ [([by (\[mor-induc\])]{})]{}
- $= \phi_{N{\otimes}_{B^{coH}}B}{\circ}((i_{N{\otimes}_{B^{coH}}B}{\circ}\alpha_{N}){\otimes}B)$ [([by (\[omeg\])]{})]{}
- $= \phi_{N{\otimes}_{B^{coH}}B}{\circ}((n_{N}{\circ}(N{\otimes}\eta_{B})){\otimes}B)$ [([by (\[unit\])]{})]{}
- $= n_{N}{\circ}(N{\otimes}(\mu_{B}{\circ}(\eta_{B}{\otimes}B))) $ [([by (\[quasi-modN\])]{})]{}
- $=n_{N} $ [([by the unit properties]{})]{}.
Finally, if we compose with $i_{M}$, $$i_{M}{\circ}\beta_{M}^{coH}{\circ}\alpha_{M^{coH}}
\stackrel{{\scriptsize \blue (\ref{coi-morph})}}{=}
\beta_{M}{\circ}i_{M^{coH}{\otimes}_{B^{coH}}B}{\circ}\alpha_{M^{coH}}
\stackrel{{\scriptsize \blue (\ref{unit})}}{=}
\beta_{M}{\circ}n_{M^{coH}}{\circ}( M^{coH}{\otimes}\eta_{B})
\stackrel{{\scriptsize \blue (\ref{omeg})}}{=}
\phi_{M}{\circ}(i_{M}{\otimes}\eta_{B})
\stackrel{{\scriptsize \blue (d2-1)}}{=}
i_{M},$$ and then $\beta_{M}^{coH}{\circ}\alpha_{M^{coH}} =id_{M^{coH}}.$
As a consequence, we obtain the following particular instances of our main theorem.
The following assertions hold:
- Let $H$ be a Hopf algebra and let $(B,\rho_{B})$ be a right $H$-comodule monoid such the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers. Then, if there exists a multiplicative total integral $h:H\rightarrow B$, the categories of right $(H,B)$-Hopf modules, denoted by ${\mathcal M}^{H}_{B}$ and introduced by Doi in [@Doi83], the category ${\mathcal SM}_{B}^{H}(h)$ of strong $(H,B,h)$-Hopf modules, and the category of right $B^{coH}$-modules ${\mathcal C}_{B^{coH}}$ are equivalent. In particular, if $B=H$ and $\rho_{B}=\delta_{H}$, the Sweedler category of Hopf modules ${\mathcal M}^{H}_{H}$, the category ${\mathcal SM}_{H}^{H}(id_{H})$ of strong $(H,H,id_{H})$-Hopf modules, and the category ${\mathcal C}$ are equivalent.
- Let $H$ be a weak Hopf algebra and let $(B,\rho_{B})$ be a right $H$-comodule monoid such that the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers. Then, if there exists a multiplicative total integral $h:H\rightarrow B$, the categories of right $(H,B)$-Hopf modules, denoted by ${\mathcal M}^{H}_{B}$ and introduced by Böhm in [@bohm2] (see also [@ZZ] and [@Hanna] for the categorical equivalence), the category ${\mathcal SM}_{B}^{H}(h)$ of strong $(H,B,h)$-Hopf modules, and the category of right $B^{coH}$-modules ${\mathcal C}_{B^{coH}}$ are equivalent. In particular, if $B=H$ and $\rho_{B}=\delta_{H}$, the category of Hopf modules ${\mathcal M}^{H}_{H}$, the category ${\mathcal SM}_{H}^{H}(id_{H})$ of strong $(H,H,id_{H})$-Hopf modules, and the category ${\mathcal C}_{H_{L}}$ of right $H_{L}$-modules are equivalent.
- Let $H$ be a Hopf quasigroup and let $(B,\rho_{B})$ be a right $H$-comodule magma such the functors $-{\otimes}B$ and $-{\otimes}H$ preserve coequalizers. Then, if there exists an anchor morphism $h:H\rightarrow B$, the categories ${\mathcal SM}_{B}^{H}(h)$ of strong $(H,B,h)$-Hopf modules, and the category of right $B^{coH}$-modules ${\mathcal C}_{B^{coH}}$ are equivalent. In particular, if $B=H$ and $\rho_{B}=\delta_{H}$, we obtain the result proved by Brzeziński in [@Brz]: the category of Hopf modules ${\mathcal M}^{H}_{H}={\mathcal SM}_{H}^{H}(id_{H})$ and the category ${\mathcal C}$ are equivalent.
- Let $H$ be a weak Hopf quasigroup such the functor $-{\otimes}H$ preserves coequalizers. The category of strong Hopf modules ${\mathcal SM}^{H}_{H}={\mathcal SM}_{H}^{H}(id_{H})$, and the category ${\mathcal C}_{H_{L}}$ of right $H_{L}$-modules are equivalent (this is the main result proved in [@Strong]).
1\. Consider $H$ a Hopf quasigroup, $A$ a unital magma in ${\mathcal C}$, and t $\varphi_A:H{\otimes}A\to A$ a morphism satisfying (\[et1\]), (\[et2\]). By the third points of Examples \[ex-hcm\] and \[anch-ex\] we know that the smash product $A\sharp H$ is a right $H$-comodule magma with coaction $\rho_{A\sharp H}=A{\otimes}\delta_{H}$ and $h=\eta_{A}{\otimes}H:H\rightarrow A\sharp H$ is an anchor morphism.
Moreover, if $A$ is a monoid and the equality $$\label{last-eqt}
\mu_{A}{\circ}(\varphi_{A}{\otimes}\varphi_{A}){\circ}(H{\otimes}c_{H,A}{\otimes}A){\circ}(\delta_{H}{\otimes}A{\otimes}A)=\varphi_{A}{\circ}(H{\otimes}\mu_{A}),$$ holds, then so hold (\[strong-1\]) and (\[strong2\]). Indeed, first note that
- $\hspace{0.38cm} q_{A\sharp H}$
- $= A{\otimes}(\mu_{H}{\circ}(H{\otimes}\lambda_{H}){\circ}\delta_{H}$ [([(\[nwx\])]{})]{}
- $= A{\otimes}\eta_{H}{\otimes}\varepsilon_{H}.$ [([(\[rightHqg\])]{})]{}.
Therefore $(A\sharp H)^{coH}=A$, $p_{A\sharp H}=A{\otimes}\varepsilon_{H}$ and $i_{A\sharp H}=A{\otimes}\eta_{H}$. As a consequence, we have
- $\hspace{0.38cm} \mu_{A\sharp H}{\circ}((\mu_{A\sharp H}{\circ}(A {\otimes}H{\otimes}i_{A\sharp H})){\otimes}A{\otimes}H)$
- $= (\mu_{A}{\otimes}\mu_{H}){\circ}(A{\otimes}(((\mu_{A}{\circ}(\varphi_{A}{\otimes}\varphi_{A})){\otimes}H){\circ}(H{\otimes}A{\otimes}H{\otimes}c_{H,A}){\circ}(H{\otimes}A{\otimes}\delta_{H}{\otimes}A){\circ}(H{\otimes}c_{H,A}{\otimes}A)$
- $\hspace{0.38cm}{\circ}(\delta_{H}{\otimes}A{\otimes}A) ){\otimes}H)$ [([unit properties and associativity of $\mu_{A}$]{})]{}
- $=(\mu_{A}{\otimes}\mu_{H}){\circ}(A{\otimes}(((\mu_{A}{\circ}(\varphi_{A}{\otimes}\varphi_{A}){\circ}(H{\otimes}c_{H,A}{\otimes}A){\circ}(\delta_{H}{\otimes}A{\otimes}A)){\otimes}H){\circ}(H{\otimes}A{\otimes}c_{H,A})$
- $\hspace{0.38cm}{\circ}(H{\otimes}c_{H,A}{\otimes}A){\circ}(\delta_{H}{\otimes}A{\otimes}A)){\otimes}H)$ [([coassociativity of $\delta_{H}$ and naturality of $c$ ]{})]{}
- $=(\mu_{A}{\otimes}\mu_{H}){\circ}(A{\otimes}(((\varphi_{A}{\circ}(H{\otimes}\mu_{A})){\otimes}H){\circ}(H{\otimes}A{\otimes}c_{H,A}){\circ}(H{\otimes}c_{H,A}{\otimes}A){\circ}(\delta_{H}{\otimes}A{\otimes}A)){\otimes}H)$
- $\hspace{0.38cm}$[([(\[last-eqt\])]{})]{}
- $= \mu_{A\sharp H}{\circ}(A{\otimes}H{\otimes}\mu_{A}{\otimes}H)$ [([naturality of $c$]{})]{}
- $= \mu_{A\sharp H}{\circ}(A{\otimes}H{\otimes}(\mu_{A\sharp H}{\circ}(i_{A\sharp H}{\otimes}A{\otimes}H)))$ [([(\[e-m-d-eps\]), naturality of $c$, unit properties, and (\[et1\])]{})]{},
and, on the other hand,
- $\hspace{0.38cm} \mu_{A\sharp H}{\circ}(i_{A\sharp H}{\otimes}\mu_{A\sharp H})$
- $= (\mu_{A}{\otimes}H){\circ}(A{\otimes}\mu_{A\sharp H})$ [([(\[e-m-d-eps\]), naturality of $c$, unit properties, and (\[et1\])]{})]{}
- $=\mu_{A\sharp H}{\circ}(\mu_{A}{\otimes}H{\otimes}A{\otimes}H)$ [([associativity of $\mu_{A}$]{})]{}
- $= \mu_{A\sharp H}{\circ}((\mu_{A\sharp H}{\circ}(i_{A\sharp H}{\otimes}A{\otimes}H)){\otimes}A{\otimes}H)$ [([(\[e-m-d-eps\]), naturality of $c$, unit properties, and (\[et1\])]{})]{}.
Therefore, if $-{\otimes}A$ and $-{\otimes}H$ preserve coequalizers, by (iii) of the previous Corollary, we have that the categories ${\mathcal SM}_{A\sharp H}^{H}(h)$ and ${\mathcal C}_{A}$ are equivalent.
An interesting example of this case can be found using the theory developed in [@Majidesfera]. Let ${\Bbb K}$ be a field and let ${\mathcal C}$ be the symmetric monoidal category of vector spaces over ${\Bbb K}$. Let $G$ be the abelian group ${\Bbb Z}_{2}^{n}$ and let $F:G\times G\rightarrow {\Bbb K}^{\ast}$ be a $2$-cochain, i.e., $F$ is a morphism such that $F(\theta, a)=F(a,\theta)=1$ for all $a\in G$ where $\theta$ is the group identity. The group algebra of $G$, denoted by ${\Bbb K}G$, is a ${\Bbb K}$-vector space with basis $\{e_{a}\;;\; a\in G\}$ and also a unital magma with the product (see [@AMajid]): $$e_{a}e_{b}=F(a,b)e_{a+b}.$$
In what follows we will denote this magma by ${\Bbb K}_{F}G$. As was pointed in [@Majidesfera], this algebraic object lives in the symmetric monoidal category of $G$-graded spaces with associator defined by the $3$-cocycle $\phi(a,b,c)=F(a,b)F(a+b,c)F(b,c)^{-1}F(a,b+c)^{-1}$ and symmetry defined by ${\mathcal R}(a,b)=F(a,b)F(b,a)^{-1}$. For example, the choice of $G={\Bbb Z}_{2}^{3}$ and certain $F$ gives the octonions. Moreover, ${\Bbb K}_{F}G$ is a composition algebra with respect to the Euclidean norm in basis $G$ if two suitable conditions hold (see (2.1) and (2.2) of [@Majidesfera]). This means that the norm $q(\displaystyle\sum_{a}u_{a}e_{a})=\sum_{a}u_{a}^2$ is multiplicative. Then $${\mathcal S}^{2^n-1}=\{\displaystyle\sum_{a}u_{a}e_{a})\;,\; \sum_{a}u_{a}^2=1_{\Bbb K}\}$$ is closed under the product in ${\Bbb K}_{F}G$. By Proposition 3.6 of [@Majidesfera] we know that ${\mathcal S}^{2^n-1}$ is an IP loop, that becomes an usual sphere if we work over ${\Bbb R}$, and then its loop algebra, denoted by ${\Bbb K}{\mathcal S}^{2^n-1}$ is a cocommutative Hopf quasigroup (see Proposition 4.7 of [@Majidesfera]). Let $H$ be ${\Bbb K}{\mathcal S}^{2^n-1}$ and let $A$ be the group algebra of $G$. Then, $A$ is a monoid (it is a cocommutative Hopf algebra) and we have an action $\varphi_{A}:H\otimes A\rightarrow A$, where ${\otimes}=\otimes_{{\Bbb K}}$, defined by $$\varphi_{A}(e_{a}\otimes e_{b})=(-1)^{a.b}e_{b}.$$
It is easy to see that $\varphi_{A}$ satisfies (\[et1\]), (\[et2\]) and (\[last-eqt\]) and, as a consequence of the general theory, we have a categorical equivalence between $\displaystyle{\mathcal SM}_{{\Bbb K}{\Bbb Z}_{2}^{n}\sharp {\Bbb K}{\mathcal S}^{2^n-1}}^{{\Bbb K}{\mathcal S}^{2^n-1}}(h)$ and $\displaystyle{\mathcal C}_{{\Bbb K}{\Bbb Z}_{2}^{n}}$ for $\displaystyle h=\eta_{{\Bbb K}{\Bbb Z}_{2}^{n}}{\otimes}id_{{\Bbb K}{\mathcal S}^{2^n-1}}$.
2\. By the second points of Examples \[ex-hcm\] and \[anch-ex\] we know that, if $H$ is a cocommutative weak Hopf quasigroup and ${\mathcal C}$ is symmetric, $(H^{op}, \rho_{H^{op}}=(H{\otimes}\lambda_{H}){\circ}\delta_{H})$ is an example of right $H$-comodule magma and $\lambda_{H}$ is an anchor morphism. Then, by Theorem 3.22 of [@Asian] and the cocommutativity of $H$, we have the equality: $$q_{H^{op}}=\Pi_{H}^{L}.$$ Therefore, $i_{H^{op}}=i_{L}$, $p_{H^{op}}=p_{L}$ and $(H^{op})^{coH}=H_{L}$. On the other hand, by the naturality of $c$ and (\[monoid-hl-2\]) we obtain that $$\mu_{H^{op}}{\circ}((\mu_{H^{op}}{\circ}(H{\otimes}i_{L})){\otimes}H)=\mu_{H^{op}}{\circ}(H{\otimes}(\mu_{H^{op}}{\circ}(i_{L}{\otimes}H))),$$ and by the naturality of $c$ and (\[monoid-hl-3\]), $$\mu_{H^{op}}{\circ}(i_{L}{\otimes}\mu_{H^{op}})=\mu_{H^{op}}{\circ}((\mu_{H^{op}}{\circ}(i_{L}{\otimes}H)){\otimes}H).$$
Therefore, we have (\[strong-1\]) and (\[strong2\]). As a consequence, if the category ${\mathcal C}$ admits coequalizers and the functor $-{\otimes}H$ preserves coequalizers, by Theorem \[p-paper\] we obtain an equivalence between the categories ${\mathcal SM}^{H}_{H^{op}}(\lambda_{H})$ and ${\mathcal C}_{H_{L}}$. If $H$ is a Hopf quasigroup, we have a similar result that asserts the following: The categories ${\mathcal SM}^{H}_{H^{op}}(\lambda_{H})$ and ${\mathcal C}$ are equivalent.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors were supported by Ministerio de Economía y Competitividad (Spain), grant MTM2016-79661-P. AEI/FEDER, UE, support included (Homología, homotopía e invariantes categóricos en grupos y álgebras no asociativas).
[99]{}
Albuquerque H., Majid S. Quasialgebra structure of the octonions. J. Algebra, 1999, 220:188-224.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R., Soneira Calvo, C. Projections and Yetter-Drinfel’d modules over Hopf (co)quasigroups. J. Algebra, 2015, 443: 153-199.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R., Soneira Calvo, C. Cleft comodules over Hopf quasigroups. Commun. Contemp. Math., 2015, 17: 1550007.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R. Weak Hopf quasigroups. Asian J. of Math., 2016, 20: 665-694.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R. Cleft and Galois extensions associated to a weak Hopf quasigroup. J. Pure Appl. Algebra, 2016, 220: 1002-1034.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R. A characterization of weak Hopf (co)quasigroups. Mediterr. J. Math., 2016, 13: 3747-3764.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R. Strong Hopf modules for weak Hopf quasigroups. Colloq. Math. Warsaw 2017, 148: 231-246.
Alonso Álvarez J.N., Fernández Vilaboa J.M., González Rodríguez R. Multiplication alteration by two-cocycles. The nonassociative version. arXiv:1703.01829 (2017).
Benkart G., Madariaga S., Pérez-Izquierdo, J.M. Hopf algebras with triality. Trans. Amer. Math. Soc., 2012, 365: 1001-1023.
Böhm G., Nill F., Szlachányi K. Weak Hopf algebras, I. Integral theory and $C^{\ast}$-structure. J. Algebra, 1999, 221: 385-438.
Böhm G. Doi-Hopf modules over weak Hopf algebras. Comm. Algebra, 2000, 10: 4687-4698.
Brzeziński T. Hopf modules and the fundamental theorem for Hopf (co)quasigroups. Internat. Elec. J. Algebra, 2010, 8: 114-128.
Brzeziński T., Jiao Z. R-smash products of Hopf quasigroups. Arab. J. Math., 2012, 1: 39-46.
Brzeziński T., Jiao Z. Actions of Hopf quasigroups. Comm. Algebra, 2012, 40: 681-696.
Doi Y. On the structure of relative Hopf modules. Comm. Algebra, 1983, 11: 243-255.
Doi Y., Takeuchi M. Hopf-Galois extensions of algebras, the Miyashita-Ulbrich action, and Azumaya algebras, J. Algebra, 1989, 121: 488-516.
Fang X., Torrecillas B. Twisted smash products and L-R smash products for biquasimodule Hopf quasigroups. Comm. Algebra, 2014, 42:4204-4234.
Henker H. Module categories over quasi-Hopf algebras and weak Hopf algebras and the projectivity of Hopf modules. Dissertation for the Doctoral Degree. Munich: LMU Munich, 2011 (available in http://edoc.ub.uni-muenchen.de/13148/)
Klim J. Bicrossproduct Hopf quasigroups. Comment. Math. Univ. Carolinae, 2010, 51: 287-304.
Jiao Z., Meng H. Ore extensions of Hopf coquasigroups, Math. Notes, 2014, 95: 338-345.
Jiao Z., Zhao X. Almost cocommutative and quasitriangular Hopf coquasigroups. J. Algebra Appl., 2014, 13: 1450010.
Kadison R.V., Ringrose J.R. Fundamentals of the theory of operator algebras, vol. I. Academic Press, New York, 1983. 13
Klim J., Majid S. Hopf quasigroups and the algebraic 7-sphere. J. Algebra, 2010, 323: 3067-3110.
Larson R.G., Sweedler M.E. An associative orthogonal bilinear form for Hopf algebras. Amer. J. Math., 1969, 91: 75-93.
Mac Lane S. Categories for the Working Mathematician. GTM 5, New York: Springer-Verlag, New-York, 1971.
Mostovoy J., Pérez-Izquierdo J.M., Shestakov I.P. Hopf algebras in non associative Lie theory. Bull. Math. Sci., 2014, 4: 129-173.
Meyer R., Local and Analytic Cyclic Homology. EMS Tracts in Mathematics Vol. 3, 2007. 1, 13, 42
Pérez-Izquierdo J M, Shestakov I P. An envelope for Malcev algebras. J. Algebra, 2004, 272: 379-393.
Pérez-Izquierdo J.M. Algebras, hyperalgebras, nonassociative bialgebras and loops. Adv. Math., 2007, 208: 834-876.
Smith J.D.H. One-sided Quantum quasigroups and loops. Demonstr. Math., 2015, 48: 620-636.
Smith J.D.H. Quantum quasigroups and loops. J. Algebra, 2016, 456: 46-75.
Smith J.D.H. Quantum idempotence, distributivity, and the Yang-Baxter equation. Comment. Math. Univ. Carolin., 2016, 57: 567-583.
Smith J.D.H. Quantum Quasigroups and the Quantum Yang-Baxter Equation. Axioms, 2016, 5: 25.
Sweedler M.E. Hopf algebras. New York: Benjamin, New York 1969.
Tvalavadze M.V, Bremner M.R. Enveloping algebras of solvable Malcev algebras of dimension five. Comm. in Algebra, 2011, 39-1532-4125.
Takeuchi M. Relative Hopf-modules-equivalences and freeness criteria. J. Algebra, 1979, 60: 452-471.
Zhang L., Zhu S. Fundamental theorems of weak Doi-Hopf modules and semisimple weak smash product Hopf algebras. Comm. Algebra, 2004, 32:3403-3415.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we provide a simple method to generate higher order position solutions and rogue wave solutions for the derivative nonlinear Schrödinger equation. The formulae of these higher order solutions are given in terms of determinants. The dynamics and structures of solutions generated by this method are studied.'
author:
- 'Yongshuai Zhang, Lijuan Guo, Shuwei Xu, Zhiwei Wu and Jingsong HE'
title: The hierarchy of higher order solutions of the derivative nonlinear Schrödinger equation
---
[[**Keywords**]{}: Higher-order positon. Higher-order rogue wave. Darboux transformation. Derivative nonlinear Schrödinger equation.]{}
[**Introduction**]{}
====================
The derivative nonlinear Schrödinger equation (DNLS) $$\label{dnls}
{\rm i}q_{t}-q_{xx}+{\rm i}(|q|^2q)_{x}=0.$$ plays an important role in plasma physics and nonlinear optics. It not only dominates the evolution of small-amplitude Alfén waves in a low-$\beta$ plasma[@JPSJ41265; @JPP16321; @PS40227; @NWCSP], but also is used to describe the behavior of large-amplitude magnetohydrodynamic (MHD) waves in a high-$\beta$ plasma [@JPP67271; @PLA3726107]. On the other hand, the DNLS equation governs the transmission of sub-picosecond in single mode optical fibers [@PRA231266; @PRA271393; @NFO].
For the DNLS equation with vanishing boundary condition (VBC), Kaup and Newell (KN)[@JMP19798] first found the one-soliton solution by inverse scattering transformation (IST). On the basis of bilinear transformation, the first $N$-solition formula was obtained by Nakamuro and Chen [@JPSJ49813]. Determinant expression of the $N$-soliton solution can be established via applying the Darboux transformation [@JPA23439]. In the case of the non-vanishing boundary condition(NVBC), Kawata and Inoue developed an IST for the DNLS equation and obtained a breather-type soliton (paired soliton) [@JPSJ441968]. Wadati [*et al*]{} derived the stationary solution for the DNLS equation under the plane wave boundary [@JPSJ48279]. Chen and Lam [@PRE69066604] revised the IST for the DNLS by introducing an affine parameter, and derived a breather solution, which can be degenerated to both dark soliton and bright soliton.
Recently, rogue wave, an emerging phenomenon, is passionately discussed. The concept of rogue wave was first proposed in the studies of deep ocean waves [@NOWATIST; @EOW], and gradually extended to other fields such as optics fibre[@Nature450; @PRL101233902; @OE163644], Bose-Einstein condensates [@PRA80033610], capillary phenomena [@PRL104104503], and so on. Rogue wave, “which appears from nowhere and disappears without a trace (WANDT)” [@PLA373675], possesses the following two remarkable characteristics: i) locates in both space and time, ii) exhibits a dominant peak.
The first order rogue wave was found in 1983 by Peregrine [@JAMSS2516], which is a solution of the NLS equation. It is usually called the Peregrine soliton, and has been observed experimentally in fiber [@NP6790], water tank [@PRL106204502] and multi-component plasma [@PRL107255005]. The first order rogue wave solution of the DNLS equation was first found by Xu and coworkers [@JPA44305203] by the Darboux transformation and certain limit technique. Recently, Guo et al [@Guo] obtained two kinds of generalized Darboux transformations, and got the formulae of higher order solutions for both the VBC and NVBC. Moveover, Guo showed two patterns (fundamental and triangular) of the second order rogue wave from a special seed solution $q={\rm exp}(-{\rm i}x)$, which are similar to the case of the NLS equation. Therefore, it is nature to ask whether the rogue wave solution of the DNLS equation possesses new structures that have not been found in other soliton equations such as the NLS equation.
The Darboux transformation, generated from the work of Darboux in 1882 for the Sturm-Liouville equation, has been an important method in generating solutions of integrable systems. To get the rogue wave solution, we need to iterate Darboux transformation at the same eigenvalue, but it does not work in this case. So we must modify the Darboux transformation to get the solutions at the same eigenvalue. In this paper, we adopt the Taylor expansion to deal with this defect, and obtain positon solutions, rational traveling solutions and rogue wave solutions.
The organization of this paper is as follows. In section 2, we provide a new method to generate solutions at the same eigenvalue base on the method of Darboux transformation and Taylor expansion, and display the formula of $N$-th order solution in terms of determinant. As applications, several explicitly analytic expressions are given, which include positon solutions, rational traveling solutions and rogue wave solutions. In section 3, we obtain the multi-rogue wave solutions by altering the mixed coefficients of eigenfunctions, which contain several free parameters. With the help of those parameters, we consider the dynamics of multi-rogue wave. Moreover, three kinds of new structures: *modified-triangular* structure, *ring-triangle* structure and *multi-ring* structure are given. The conclusion is given in the last section.
[**The solutions of the DNLS equation**]{}
==========================================
The Kaup and Newell (KN) system[@JMP19798]: $$\label{KN}
\left\{
\begin{aligned}
r_t-{\rm i}r_{xx}-(r^2q)_x&=0,\\
q_t+{\rm i}q_{xx}-(rq^2)_x&=0.
\end{aligned}
\right.$$ can be represented as the integrability condition of the following Kaup and Newell spectral system (Lax pair)[@JMP19798; @JPSJ68355]:
$$\label{lax}
\left\{
\begin{aligned}
\Psi_{x}&=M\Psi=(J\lambda^2+Q\lambda)\Psi,\\
\Psi_{t}&=N\Psi=(2J\lambda^4+V_{3}\lambda^3+V_{2}\lambda^2+V_{1}\lambda)\Psi,
\end{aligned}
\right.$$
with
$${J=\left( \begin{array}{cc}
{\rm i} &0 \\
0 &-{\rm i}
\end{array}
\right),
\quad Q=\left( \begin{array}{cc}
0 &q\\
r &0
\end{array}
\right),} \nonumber$$
$$V_{3}=2Q, \quad V_{2}=Jqr,
\quad V_{1}=\left(\begin{array}{cc}
0 &-{\rm i}q_{x}+q^2r \\
{\rm i}r_{x}+r^2q &0
\end{array}
\right), \nonumber$$
Where $\lambda\in\mathbb{C}$, $\Psi\in\mathbb{C}^2$, $\Psi$ is called the eigenfunction of the spectral problem corresponding to eigenvalue $\lambda$ .
When $$\label{reduction}
r=-q^*,$$ the KN system can be reduced to the DNLS equation, asterisk denotes complex conjugation.
[**2.1 Determinant expression**]{}\
The $N$-th Darboux transformation of the KN system in terms of determinant was obtained In [@JPA44305203]. And the formulae for $N$-th order solutions were given as following:
[@JPA44305203] Let $\Psi_{i}=\left(\begin{array}{c}
f_i\\
g_i
\end{array}
\right)$ $\left(i=1,2,\cdots,n\right)$ be distinct solutions related to $\lambda_i$ of the spectral problem (\[lax\]), then ($q^{[n]}$,$r^{[n]}$) given by the following formulae are new solutions of the KN system . $$\label{q[n]}
q^{[n]}=\frac{\Omega_{11}^2}{\Omega_{21}^2}q+2{\rm i}\frac{\Omega_{11}\Omega_{12}}{\Omega_{21}^2},\quad
r^{[n]}=\frac{\Omega_{21}^2}{\Omega_{11}^2}r-2{\rm i}\frac{\Omega_{21}\Omega_{22}}{\Omega_{11}^2}.$$ with $n=2k$, $$\quad\Omega_{11}=\begin{vmatrix}
\lambda_{1}^{n-1}g_1&\lambda_{1}^{n-2}f_{1}&\lambda_{1}^{n-3}g_{1}&\cdots&\lambda_{1}g_{1}&f_{1}\\
\lambda_{2}^{n-1}g_{2}&\lambda_{2}^{n-2}f_{2}&\lambda_{2}^{n-3}g_{2}&\cdots&\lambda_{2}g_{2}&f_{2}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
%\lambda_{n-1}^{n-1}g_{n-1}&\lambda_{n-1}^{n-2}f_{n-1}&\lambda_{n-1}^{n-3}g_{n-1}&\cdots&\lambda_{n-1}g_{n-1}&f_{n-1}\\
\lambda_{n}^{n-1}g_{n}&\lambda_{n}^{n-2}f_{n}&\lambda_{n}^{n-3}g_{n}&\cdots&\lambda_{n}g_{n}&f_{n} \nonumber
\end{vmatrix},$$ $$\quad\Omega_{12}=\begin{vmatrix}
\lambda_{1}^{n}f_1&\lambda_{1}^{n-2}f_{1}&\lambda_{1}^{n-3}g_{1}&\cdots&\lambda_{1}g_{1}&f_{1}\\
\lambda_{2}^{n}f_{2}&\lambda_{2}^{n-2}f_{2}&\lambda_{2}^{n-3}g_{2}&\cdots&\lambda_{2}g_{2}&f_{2}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
%\lambda_{n-1}^{n}f_{n-1}&\lambda_{n-1}^{n-2}f_{n-1}&\lambda_{n-1}^{n-3}g_{n-1}&\cdots&\lambda_{n-1}g_{n-1}&f_{n-1}\\
\lambda_{n}^{n}f_{n}&\lambda_{n}^{n-2}f_{n}&\lambda_{n}^{n-3}g_{n}&\cdots&\lambda_{n}g_{n}&f_{n} \nonumber
\end{vmatrix},$$ $$\quad\Omega_{21}=\begin{vmatrix}
\lambda_{1}^{n-1}f_1&\lambda_{1}^{n-2}g_{1}&\lambda_{1}^{n-3}f_{1}&\cdots&\lambda_{1}f_{1}&g_{1}\\
\lambda_{2}^{n-1}f_{2}&\lambda_{2}^{n-2}g_{2}&\lambda_{2}^{n-3}f_{2}&\cdots&\lambda_{2}f_{2}&g_{2}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
%\lambda_{n-1}^{n-1}f_{n-1}&\lambda_{n-1}^{n-2}g_{n-1}&\lambda_{n-1}^{n-3}f_{n-1}&\cdots&\lambda_{n-1}f_{n-1}&g_{n-1}\\
\lambda_{n}^{n-1}f_{n}&\lambda_{n}^{n-2}g_{n}&\lambda_{n}^{n-3}f_{n}&\cdots&\lambda_{n}f_{n}&g_{n} \nonumber
\end{vmatrix},$$ $${\label{even1}}
\Omega_{22}=\begin{vmatrix}\nonumber
\lambda_{1}^{n}g_1&\lambda_{1}^{n-2}g_{1}&\lambda_{1}^{n-3}f_{1}&\cdots&\lambda_{1}f_{1}&g_{1}\\
\lambda_{2}^{n}g_{2}&\lambda_{2}^{n-2}g_{2}&\lambda_{2}^{n-3}f_{2}&\cdots&\lambda_{2}f_{2}&g_{2}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
%\lambda_{n-1}^{n}g_{n-1}&\lambda_{n-1}^{n-2}g_{n-1}&\lambda_{n-1}^{n-3}f_{n-1}&\cdots&\lambda_{n-1}g_{n-1}&f_{n-1}\\
\lambda_{n}^{n}g_{n}&\lambda_{n}^{n-2}g_{n}&\lambda_{n}^{n-3}f_{n}&\cdots&\lambda_{n}f_{n}&g_{n}
\end{vmatrix}.$$ Here $\lambda_{2l}=-\lambda_{2l-1}^*$ and $\Psi_{2l}=\left(\begin{array}{c}
f_{2l}\\
g_{2l}
\end{array} \right)=\left(\begin{array}{c}
g_{2l-1}^*\\
f_{2l-1}^*
\end{array}\right)$.
It is trivial to check $r^{[2k]}=-{q^{[2k]}}^*$.\
[**2.2 Solutions from vacuum**]{}\
Let us consider the trivial case. When $q=r=0$, the following $\Psi$ is an eigenfunction for $\lambda$, $$\label{fun1}
\Psi=\left(\begin{array}{c}
f\\
g
\end{array}
\right),\quad
f={\rm exp}({\rm i}(\lambda^{2}x+2\lambda^{4}t)),\quad
g={\rm exp}(-{\rm i}(\lambda^{2}x+2\lambda^{4}t)).$$\
By applying the above formulae , we can get $N$-soliton solution of the DNLS equation from vacuum. To get new kinds of solutions, we set the eigenvalues share the same value, i.e., iterating the Darboux transformation at the same eigenvalue. However, the formulae will be ineffective in this case. Next, we will use the Taylor expansion to generate the Darboux transformation and get the formula of $q^{[n]}$ at the same eigenvalue as we have done for the case of the NLS equation[@arxiv12093742].
At first, we define new functions $\Psi[i,j,k]$ for a general solution $\Psi=\Psi(\lambda)$ corresponding to $\lambda$ as following: $$\lambda^{j}\Psi=\Psi[i,j,0]+\Psi[i,j,1]\epsilon+\Psi[i,j,2]\epsilon^2+\cdots+\Psi[i,j,k]\epsilon^k+\cdots,$$ with $$\Psi[i,j,k]=\frac{1}{k!}\frac{\partial^k(\lambda_i^j\Psi(\lambda_i))}{\partial\lambda_i^k}.$$ In particular $$\Psi[1,1,0]=\lambda_1\Psi(\lambda_1),\qquad
\Psi[i,j,0]=\lambda_i^j\Psi(\lambda_i).$$
[ Let $\lambda_1=\alpha_1+{\rm i}\beta_1$, $\lambda_2=-\lambda_1^*$, then the following formula is the $n$-th($n=2k$) solution of the DNLS equation generated at the same eigenvalue. $$\label{nposition}
q^{[n]}=2{\rm i}\frac{\delta_{11}\delta_{12}}{\delta_{21}^2}$$ where $$\delta_{11}=\begin{vmatrix}
g[1,n-1,0]&f[1,n-2,0]&g[1,n-3,0]&\cdots&g[1,1,0]&f[1,0,0]\\
g[2,n-1,0]&f[2,n-2,0]&g[2,n-3,0]&\cdots&g[2,1,0]&f[2,0,0]\\
g[1,n-1,1]&f[1,n-2,1]&g[1,n-3,1]&\cdots&g[1,1,1]&f[1,0,1]\\
g[2,n-1,1]&f[2,n-2,1]&g[2,n-3,1]&\cdots&g[2,1,1]&f[2,0,1]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
g[1,n-1,k-1]&f[1,n-2,k-1]&g[1,n-3,k-1]&\cdots&g[1,1,k-1]&f[1,0,k-1]\\
g[2,n-1,k-1]&f[2,n-2,k-1]&g[2,n-3,k-1]&\cdots&g[2,1,k-1]&f[2,0,k-1]\nonumber
\end{vmatrix},$$ $$\delta_{12}=\begin{vmatrix}
f[1,n,0]&f[1,n-2,0]&g[1,n-3,0]&\cdots&g[1,1,0]&f[1,0,0]\\
f[2,n,0]&f[2,n-2,0]&g[2,n-3,0]&\cdots&g[2,1,0]&f[2,0,0]\\
f[1,n,1]&f[1,n-2,1]&g[1,n-3,1]&\cdots&g[1,1,1]&f[1,0,1]\\
f[2,n,1]&f[2,n-2,1]&g[2,n-3,1]&\cdots&g[2,1,1]&f[2,0,1]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
f[1,n,k-1]&f[1,n-2,k-1]&g[1,n-3,k-1]&\cdots&g[1,1,k-1]&f[1,0,k-1]\\
f[2,n,k-1]&f[2,n-2,k-1]&g[2,n-3,k-1]&\cdots&g[2,1,k-1]&f[2,0,k-1]\\\nonumber
\end{vmatrix},$$ $$\delta_{21}=\begin{vmatrix}
f[1,n-1,0]&g[1,n-2,0]&f[1,n-3,0]&\cdots&f[1,1,0]&g[1,0,0]\\
f[2,n-1,0]&g[2,n-2,0]&f[2,n-3,0]&\cdots&f[2,1,0]&g[2,0,0]\\
f[1,n-1,1]&g[1,n-2,1]&f[1,n-3,1]&\cdots&f[1,1,1]&g[1,0,1]\\
f[2,n-1,1]&g[2,n-2,1]&f[2,n-3,1]&\cdots&f[2,1,1]&g[2,0,1]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
f[1,n-1,k-1]&g[1,n-2,k-1]&f[1,n-3,k-1]&\cdots&f[1,1,k-1]&g[1,0,k-1]\\
f[2,n-1,k-1]&g[2,n-2,k-1]&f[2,n-3,k-1]&\cdots&f[2,1,k-1]&g[2,0,k-1]
\end{vmatrix}.\nonumber$$]{}
For the entries in the first column of $\Omega_{11}$ , $$\begin{aligned}
\lambda_1^{n-1}g_1&=g[1,n-1,0],\\
\lambda_2^{n-1}g_2&=g[2,n-1,0],\\
\lambda_3^{n-1}g_3&=g[1,n-1,0]+g[1,n-1,1]\epsilon,\\
\lambda_4^{n-1}g_4&=g[2,n-1,0]+g[2,n-1,1]\epsilon,\\
&\vdots\\
\lambda_{n-1}^{n-1}g_{n-1}&=g[1,n-1,0]+g[1,n-1,1]\epsilon+\cdots+g[1,n-1,k-1]\epsilon^{k-1},\\
\lambda_n^{n-1}g_n&=g[2,n-1,0]+g[2,n-1,1]\epsilon+\cdots+g[2,n-1,k-1]\epsilon^{k-1}.\\
\end{aligned}
%\end{gathered}$$ Taking the similar procedure to the other entries in $\Omega_{11}$, $\Omega_{12}$, and $\Omega_{21}$. Finally, the $q^{[n]}$ can be obtained through simple calculation.
For example, when $n=4$, $$\label{q[41]}
q^{[4]}=\frac{\delta_{11}^2}{\delta_{21}^2}q+2{\rm i}\frac{\delta_{11}\delta_{12}}{\delta_{21}^2}=2{\rm i}\frac{\delta_{11}\delta_{12}}{\delta_{21}^2},$$ where $$\delta_{11}=\begin{vmatrix}
g[1,3,0]&f[1,2,0]&g[1,1,0]&f[1,0,0]\\
g[2,3,0]&f[2,2,0]&g[2,1,0]&f[2,0,0]\\
g[1,3,1]&f[1,2,1]&g[1,1,1]&f[1,0,1]\\
g[2,3,1]&f[2,2,1]&g[2,1,1]&f[2,0,1]
\end{vmatrix},$$ $$\delta_{12}=\begin{vmatrix}
f[1,4,0]&f[1,2,0]&g[1,1,0]&f[1,0,0]\\
f[2,4,0]&f[2,2,0]&g[2,1,0]&f[2,0,0]\\
f[1,4,1]&f[1,2,1]&g[1,1,1]&f[1,0,1]\\
f[2,4,1]&f[2,2,1]&g[2,1,1]&f[2,0,1]
\end{vmatrix},$$ $$\delta_{21}=\begin{vmatrix}
f[1,3,0]&g[1,2,0]&f[1,1,0]&g[1,0,0]\\
f[2,3,0]&g[2,2,0]&f[2,1,0]&g[2,0,0]\\
f[1,3,1]&g[1,2,1]&f[1,1,1]&g[1,0,1]\\
f[2,3,1]&g[2,2,1]&f[2,1,1]&g[2,0,1]
\end{vmatrix}.$$ Substituting the eigenfunction into the formula , we obtain the positon solution $$\label{posoliton1}
q_{positon}^{[4]}=\frac{L_1^*L_2}{L_1^2},$$ where $$\nonumber
\begin{split}
L_1=&G_1-{\rm i}G_2,\\
L_2=&-16{\rm i}\alpha_1\beta_1(\cos(F_2)+{\rm i}\sin(F_2))\left(\left(\beta_1^3+4{\rm i}\alpha_1^4\beta_1x+4{\rm i}\alpha_1^2\beta_1^3x
-32{\rm i}\alpha_1^4\beta_1^3t+16{\rm i}\alpha_1^6\beta_1t-48{\rm i}\alpha_1^2 \right.\right.\\
&\left.\times\beta_1^5t\right)\sinh(F_1)-\left.({\rm i}\alpha_1^3+4\alpha_1^3\beta_1^2x+32\alpha_1^3\beta_1^4t+48\alpha_1^5\beta_1^2t
+4\beta_1^4\alpha_1x-16\beta_1^6\alpha_1t)\cosh(F_1)\right),\\
G_1=&{\alpha_{{1}}}^{4}+{\beta_{{1}}}^{4}+256\,{\alpha_{{1}}}^{8}{\beta_{{
1}}}^{2}xt-256\,{\alpha_{{1}}}^{4}{\beta_{{1}}}^{6}xt+256\,{\alpha_{{1}
}}^{6}{\beta_{{1}}}^{4}xt-256\,{\alpha_{{1}}}^{2}{\beta_{{1}}}^{8}xt+
512\,{\alpha_{{1}}}^{2}{\beta_{{1}}}^{10}{t}^{2}\\
&+32\,{\alpha_{{1}}}^{2
}{\beta_{{1}}}^{6}{x}^{2}+32\,{\alpha_{{1}}}^{6}{\beta_{{1}}}^{2}{x}^{
2}+512\,{\alpha_{{1}}}^{10}{\beta_{{1}}}^{2}{t}^{2}+2048\,{\alpha_{{1}
}}^{8}{\beta_{{1}}}^{4}{t}^{2}+3072\,{\alpha_{{1}}}^{6}{\beta_{{1}}}^{
6}{t}^{2}\\
&+64\,{\alpha_{{1}}}^{4}{\beta_{{1}}}^{4}{x}^{2}+2048\,{\alpha
_{{1}}}^{4}{\beta_{{1}}}^{8}{t}^{2}+({\alpha_{{1}}}^{4}-{\beta_{{1}}}^{4})\cosh \left( 2F_1 \right), \\
G_2=&-16\,{\alpha_{{1}}}^{2}{\beta_{{1}}}^{4}x-384\,{\alpha_{{1}}}^{4}{
\beta_{{1}}}^{4}t+64\,{\alpha_{{1}}}^{2}{\beta_{{1}}}^{6}t+16\,{\alpha
_{{1}}}^{4}{\beta_{{1}}}^{2}x+(2\,{\alpha_{{1}}}^{3}\beta_{{1}}+2\,\alpha_{{1}}{\beta_{{1}}}^{3})\sinh
\left( 2F_1 \right) \\
&+64\,{\alpha_{{1}}}^{6}{\beta_{{1}
}}^{2}t,\\
F_1=&4\,\alpha_{{1}}\beta_{{1}} \left( 4\,t{\alpha_{{1}}}^{2}-4\,t{\beta_{{1}}}^{2}+x \right),\\
F_2=&2\,{\alpha_{{1}}}^{2}x+4\,{\alpha_{{1}}}^{4}t-24\,t{\alpha_{{1}}}^{2}{\beta_{{1}}}^{2}-2\,
{\beta_{{1}}}^{2}x+4\,{\beta_{{1}}}^{4}t.
\end{split}$$
when $x\rightarrow\pm\infty$, $|q^{[4]}|=0$, when $x=0$, $t=0$, $|q^{[4]}|^2=64\beta_1^2$. A simple analysis shows that it possesses phase shift compared with $2$-rd soliton when $t\rightarrow\pm\infty$. After taking values as $\alpha_1=0.5$, $\beta_1=0.5$, the evolution of positon solution of the DNLS equation is shown in Fig. \[positon\].
Next, if we set $\alpha_1\rightarrow0$ in above procedure, we will get the the second order rational traveling solution. With these parameters, we find that the general solution can be given in the same form as , but with the values for $L_1$ and $L_2$ written by
$$\label{posoliton12}
\begin{split}
L_1=&G_1-{\rm i}G_2,\\
L_2=&-8{\rm i}\beta_1{\rm exp}(2{\rm i}\beta_1^2(2\beta_1^2t-x))(-3{\rm i}-12\beta_1^2x-48\beta_1^4t+48{\rm i}
\beta_1^4x^2+2304{\rm i}\beta_1^8t^2-768{\rm i}\beta_1^6xt\\
&-64\beta_1^6x^3+4096\beta_1^{12}t^3-3072\beta_1^{10}xt^2+768\beta_1^8tx^2),\\
G_1=&3+4096\,{\beta_{{1}}}^{10}t{x}^{3}-24576\,{\beta_{{1}}}^{12}{t}^{2}{x}
^{2}+65536\,x{\beta_{{1}}}^{14}{t}^{3}-768\,{\beta_{{1}}}^{6}xt+96\,{
\beta_{{1}}}^{4}{x}^{2}-65536\,{\beta_{{1}}}^{16}{t}^{4}\\
&+4608\,{\beta_
{{1}}}^{8}{t}^{2}-256\,{\beta_{{1}}}^{8}{x}^{4},\\
G_2=&576\,{\beta_{{1}}}^{4}t-48\,{\beta_{{1}}}^{2}x+3072\,{\beta_{{1}}}^{8}
t{x}^{2}-256\,{\beta_{{1}}}^{6}{x}^{3}+16384\,{\beta_{{1}}}^{12}{t}^{3
}-12288\,{\beta_{{1}}}^{10}x{t}^{2}.
\end{split}\nonumber$$
The dynamics of rational travelling solution of the DNLS equation with $\beta_1=0.3$ are shown in Fig. \[ration\]. Actually, it represents the interaction of two rational traveling solitons, and shares same properties with positon.\
[**2.3. Solutions from periodic solution**]{}\
Here, we will apply the method discussed above to generate solutions from periodic seed solution with the same eigenvalue. Moreover, we generate a hierarchy of rogue wave solutions.
We start with a general periodic solution $$\label{seed}
q=c e^{({\rm i}(ax+bt))},\quad
b=a(-c^2+a),\quad a,c\in\mathbb{R}.$$ Substituting into the spectral problem , we obtain the eigenfunction $\Psi=\left(
\begin{array}{c}
f\\
g
\end{array}
\right)$ corresponding to the eigenvalue $\lambda$ via applying the method of separation of variables and superposition principle.
$$\label{eigenfun}
\left(
\begin{array}{c}
f(x,t,\lambda)\\
g(x,t,\lambda)
\end{array}
\right)=\left(
\begin{array}{c}
D_1\omega^1_{11}(x,t,\lambda)+D_2\omega^2_{11}(x,t,\lambda)+D_1{\omega^1_{12}}^\ast(x,t,-\lambda^\ast)+D_2{\omega^2_{12}}^\ast(x,t,-\lambda^\ast)\\
D_1\omega^1_{12}(x,t,\lambda)+D_2\omega^2_{12}(x,t,\lambda)+D_1{\omega^1_{12}}^\ast(x,t,-\lambda^\ast)+D_2{\omega^2_{12}}^\ast(x,t,-\lambda^\ast)
\end{array}
\right).$$
where $$\label{D1}
\left\{
\begin{aligned}
D_1&=1,\\
D_2&=1.
\end{aligned}
\right.$$
$$\left(
\begin{array}{c}
\omega^1_{11}(x,t,\lambda)\\
\omega^1_{12}(x,t,\lambda)
\end{array}
\right)=\left(
\begin{array}{c}
\exp({c_1(x+2\lambda^2t+(-c^2+a)t)+\dfrac{1}{2}({\rm i}(ax+bt))})\\
\dfrac{{\rm i}a-2{\rm i}\lambda^2+2c_1}{2{\lambda}c}\exp({c_1(x+2\lambda^2t+bt)-\dfrac{1}{2}({\rm i}(ax+bt))})
\end{array}
\right),\nonumber$$
$$\left(\begin{array}{c}
\omega^2_{11}(x,t,\lambda)\\
\omega^2_{12}(x,t,\lambda)
\end{array}
\right)=\left(\begin{array}{c}
\exp(-c_1(x+2\lambda^2t+(-c^2+a)t)+\dfrac{1}{2}({\rm i}(ax+bt)))\\
\dfrac{{\rm i}a-2{\rm i}\lambda^2-2c_1}{2{\lambda}c}\exp(-c_1(x+2\lambda^2t+bt)-\dfrac{1}{2}({\rm i}(ax+bt)))
\end{array}
\right),\nonumber$$
$$\omega^1(x,t,\lambda)=\left(\begin{array}{c}
\omega^1_{11}(x,t,\lambda)\\
\omega^1_{12}(x,t,\lambda)\nonumber
\end{array}
\right),
\quad
\omega^2(x,t,\lambda)=\left(\begin{array}{c}
\omega^2_{11}(x,t,\lambda)\\
\omega^2_{12}(x,t,\lambda)\nonumber
\end{array}
\right),$$
$$c_1=\dfrac{\sqrt{-a^2-4\lambda^4-4\lambda^2(c^2-a)}}{2}.\nonumber$$
To get the rogue wave solutions, the value of $n$ in formula (\[q\[n\]\]) must be even. When $n=2k$, we obtain the new expression of $q^{[n]}$.
Assuming $\lambda_1=\frac{1}{2}\sqrt {-{c}^{2}+2\,a}-\frac{1}{2}\,{\rm i}c$, $\lambda_2=-\lambda_1^*$, then $q^{[n]}$ given by following formula is the $k$-th rogue wave solution for the DNLS equation. $$\label{RW}
q^{[n]}=\frac{\delta_{11}^2}{\delta_{21}^2}q+2{\rm i}\frac{\delta_{11}\delta_{12}}{\delta_{21}^2};$$
where $$\delta_{11}=\begin{vmatrix}
g[1,n-1,1]&f[1,n-2,1]&g[1,n-3,1]&\cdots&g[1,1,1]&f[1,0,1]\\
g[2,n-1,1]&f[2,n-2,1]&g[2,n-3,1]&\cdots&g[2,1,1]&f[2,0,1]\\
g[1,n-1,2]&f[1,n-2,2]&g[1,n-3,2]&\cdots&g[1,1,2]&f[1,0,2]\\
g[2,n-1,2]&f[2,n-2,2]&g[2,n-3,2]&\cdots&g[2,1,2]&f[2,0,2]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
g[1,n-1,k]&f[1,n-2,k]&g[1,n-3,k]&\cdots&g[1,1,k]&f[1,0,k]\\
g[2,n-1,k]&f[2,n-2,k]&g[2,n-3,k]&\cdots&g[2,1,k]&f[2,0,k]\nonumber
\end{vmatrix},$$ $$\delta_{12}=\begin{vmatrix}
f[1,n,1]&f[1,n-2,1]&g[1,n-3,1]&\cdots&g[1,1,1]&f[1,0,1]\\
f[2,n,1]&f[2,n-2,1]&g[2,n-3,1]&\cdots&g[2,1,1]&f[2,0,1]\\
f[1,n,2]&f[1,n-2,2]&g[1,n-3,2]&\cdots&g[1,1,2]&f[1,0,2]\\
f[2,n,2]&f[2,n-2,2]&g[2,n-3,2]&\cdots&g[2,1,2]&f[2,0,2]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
f[1,n,k]&f[1,n-2,k]&g[1,n-3,k]&\cdots&g[1,1,k]&f[1,0,k]\\
f[2,n,k]&f[2,n-2,k]&g[2,n-3,k]&\cdots&g[2,1,k]&f[2,0,k]\nonumber
\end{vmatrix},$$ $$\delta_{21}=\begin{vmatrix}
f[1,n-1,1]&g[1,n-2,1]&f[1,n-3,1]&\cdots&f[1,1,1]&g[1,0,1]\\
f[2,n-1,1]&g[2,n-2,1]&f[2,n-3,1]&\cdots&f[2,1,1]&g[2,0,1]\\
f[1,n-1,2]&g[1,n-2,2]&f[1,n-3,2]&\cdots&f[1,1,2]&g[1,0,2]\\
f[2,n-1,2]&g[2,n-2,2]&f[2,n-3,2]&\cdots&f[2,1,2]&g[2,0,2]\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
f[1,n-1,k]&g[1,n-2,k]&f[1,n-3,k]&\cdots&f[1,1,k]&g[1,0,k]\\
f[2,n-1,k]&g[2,n-2,k]&f[2,n-3,k]&\cdots&f[2,1,k]&g[2,0,k]
\end{vmatrix}.\nonumber$$
For the entries in the first column of $\Omega_{11}$ , $$\begin{aligned}
\lambda_1^{n-1}g_1=&g[1,n-1,1]\epsilon,\\
\lambda_2^{n-1}g_2=&g[2,n-1,1]\epsilon,\\
\lambda_3^{n-1}g_3=&g[1,n-1,1]\epsilon+g[1,n-1,2]\epsilon^2,\\
\lambda_4^{n-1}g_4=&g[2,n-1,1]\epsilon+g[2,n-1,2]\epsilon^2,\\
&\vdots\\
\lambda_{n-1}^{n-1}g_{n-1}=&g[1,n-1,1]\epsilon+g[1,n-1,2]\epsilon^2+\cdots+g[1,n-1,k]\epsilon^{k},\\
\lambda_n^{n-1}g_n=&g[2,n-1,1]\epsilon+g[2,n-1,2]\epsilon^2+\cdots+g[2,n-1,k]\epsilon^{k}.\\
\end{aligned}$$ Taking the similar procedure to the other entries in $\Omega_{11}$, $\Omega_{12}$, and $\Omega_{21}$. Finally, the $q^{[n]}$ can be obtained through simple calculation.
Here we point out that the formula is different with the result of [@Guo].
Next, we present some special examples, which have different structures.
- The first order rogue wave solution
For $n=2$, we get the first order rogue wave solution according the above formulae $$\label{RW1}
q_{1_{rw}}^{[2]}=\frac{L_1^*L_2}{L_1^2}c{{\rm exp}^{{\rm i}a \left( x-t{c}^{2}+ta \right) }},$$ where $$\begin{split}
L_1=&e_1+{\rm i}e_2,\quad
L_2=e_3+{\rm i}e_4,\\
e_1=&-8\,{t}^{2}{c}^{2}{a}^{3}+12\,{t}^{2}{c}^{4}{a}^{2}-8\,x{c}^{2}t{a}^{2
}+8\,x{c}^{4}ta-2\,a{x}^{2}{c}^{2}-6\,{t}^{2}{c}^{6}a-1,\\
e_2=&4\,at{c}^{2}-6\,t{c}^{4}+2\,x{c}^{2},\\
e_3=&8\,{t}^{2}{c}^{2}{a}^{3}+8\,x{c}^{2}t{a}^{2}-12\,{t}^{2}{c}^{4}{a}^{2}
+2\,a{x}^{2}{c}^{2}-8\,x{c}^{4}ta+6\,{t}^{2}{c}^{6}a-3, \\
e_4=&12\,at{c}^{2}-6\,t{c}^{4}+2\,x{c}^{2}.\\
\end{split}$$
A direct analysis shows when $x\rightarrow \infty$, $t\rightarrow \infty$, $q^{[2]}\rightarrow c^2$, the maximum module of $\left|q^{[2]}\right|^2$ is equal to $9c^2$ and locates at the origin. A $1$-rogue wave with particular parameters is shown in Fig. \[fig.1rw\]
- High order rogue wave solutions
Generally, the expression of rogue wave becomes more complicated with increasing $n$ (an analytic expression of the second rogue wave solution is displayed in appendix A). Therefore, we use numerical simulations to discuss the high order rogue wave for convenient. We set $a=1$ and $c=1$ in the following.
When $n=4$, we can obtain the second order rogue wave solution of the DNLS equation according to the formula (\[RW\]). $$\label{RW2}
q_{2_{rw}}^{[4]}=\frac{L_1^*L_2}{L_1^2}{\rm exp}({\rm i}x),$$ with $$\begin{split}
L_1=&9+90\,{x}^{2}-12\,{x}^{4}+666\,{t}^{2}+180\,{t}^{4}+8\,{x}^{6}+8\,{t}^
{6}-54\,{\rm i}x+24\,{\rm i}t{x}^{4}-216\,{x}^{2}{t}^{2}\\
&-72\,xt+24\,{x}^{4}{t}^{2}+48\,{x}^{3}t+48\,x{t}^{3}+288\,{\rm i}{t}^{2}x+24\,{x}^{2}{t}^{4}
-24\,{\rm i}{t}^{4}x-24\,{\rm i}{x}^{5}\\
&-48\,{\rm i}{x}^{3}-48\,{\rm i}{t}^{2}{x}^{3}+198\,{\rm i}t+24\,{\rm i}{t}^{5}+48\,{\rm i}{t}^{3}{x}^{2}+336\,{\rm i}{t}^{3},\\
L_2=&45-198\,{x}^{2}-60\,{x}^{4}-486\,{t}^{2}-60\,{t}^{4}-48\,{\rm i}{x}^{3}+528
\,{\rm i}{t}^{3}+72\,{\rm i}{t}^{5}-414\,{\rm i}t+8\,{x}^{6}\\
&+8\,{t}^{6}+72\,{\rm i}t{x}^{4}+144\,{\rm i}{t}^{3}{x}^{2}+24\,{\rm i}{t}^{4}x+48\,{\rm i}{t}^{2}{x}^{3}-576\,{\rm i}{t}^{2}x-
288\,{\rm i}{x}^{2}t-90\,{\rm i}x\\&
-504\,{x}^{2}{t}^{2}
+504\,xt+24\,{x}^{4}{t}^{2}-
144\,{x}^{3}t-144\,x{t}^{3}+24\,{\rm i}{x}^{5}+24\,{x}^{2}{t}^{4}.
\end{split}$$
Besides, we succeed in reaching the $7$-order rogue wave by applying the above formula (\[RW\]). Nevertheless, the analytical expression is too tedious, we omit it here. Their dynamical evolutions are shown in Fig. \[fig.nrw\]. From the figures, we find that the maximum height of the $k$-th order rogue wave is $(2k+1)^2$. As remarked in [@arxiv12093742; @PRE84056611], there are $\frac{k(k+1)}{2}-1$ local maxima on each side of the $x=0$ line for $k$-order rogue wave of the NLS equation. However, there are only $k$ small peak on the each side of the $t=0$ line in Fig. \[fig.nrw\] for $k$-order rogue wave solution of the DNLS equation. We may make a conjecture here that the central peak of rogue wave of the DNLS equation contain more energy than the NLS.
[**The dynamics of rogue wave with parameters**]{}
==================================================
In above section, we assume that $D_1=1$ and $D_2=1$ . Actually, both $D_1$ and $D_2$ can be assumed as some new constants on the premise that is the eigenfunction of spectral system . In this section, we set $D_1$ and $D_2$ as following: $$\label{D3}
\left\{
\begin{aligned}
D_1=&\exp\left(-{\rm i}c_1(S_0+S_1\epsilon+S_2\epsilon^2+S_3\epsilon^3+\cdots+S_{k-1}\epsilon^{k-1})\right),\\
D_2=&\exp\left({\rm i}c_1(S_0+S_1\epsilon+S_2\epsilon^2+S_3\epsilon^3+\cdots+S_{k-1}\epsilon^{k-1})\right).
\end{aligned}
\right.$$ Here $S_0,S_1,S_2,S_3,\cdots,S_{k-1}\in \mathbb{C}$ . Although the terms with nonzero orders of $\epsilon$ in eq.(\[D3\]) vanish in the $\epsilon\rightarrow 0$ limit, their coefficients $S_i$($i=0,1,2,\cdots,k-1$) have a crucial effect on the structure of higher order rogue wave. Depending on these parameters, we can obtain a variety of solutions of the same order. Finding these relative positions in terms of $S_i$ ($i=0,1,2,\cdots,k-1$) is the subject of our analysis below.\
[**3.1. Solutions with one parameter**]{}\
In this subsection, we will discuss the dynamics of high-order rogue wave in detail. In the case that only one of the parameters is nonzero, four typical models are obtained: *fundamental pattern*, *triangular* structure, *ring* structure, and *modified-triangular* structure. Moreover, the *modified-triangular* structure has never been found in other equations.
- Fundamental pattern
For $n=2$, the first order rogue wave possesses only one parameter $S_0$, which is shown in Fig. \[fig.1rws0\]. We observe that it is a translation of the solution in Fig. \[fig.1rw\] relative to the origin. Actually, it can shift the $1$-order rogue wave solution to an arbitrary position on the ($x$,$t$)-plane, but it is trivial. So we omit this case for higher order solutions.\
- Triangular structure
In this subsection, we set $S_i=0$ except $S_1$. The resulting wave functions for orders $k=2,3,4,5,6,7$ are shown in Fig. \[fig.triangle\]. Remarkably, all higher order solutions display *triangular* structure. We observe triangle with three peaks in Fig. \[fig.triangle\](a)(It has been obtained in Ref [@Guo].), ten peaks in Fig. \[fig.triangle\](b), fifteen peaks in Fig. \[fig.triangle\](c), etc. All peaks within the triangle are first order rogue waves. So we can conclude that the *triangular* structure of an order $k$ rogue wave solution is composed of $\frac{k(k+1)}{2}$ first order rogue waves, and it can be observed that successive $k$ rows possessing $k$, $k-1$, $k-2$, $\cdots$, $1$ peaks respectively. Evidently, the structure of the second order rogue wave is same as the result of the NLS equation, which is called triplet [@PLA3752782].
Another remarkable feature of these solutions is that, for the triangle of $k$-order($k>2$) rogue wave solutions, the outer triangle is composed of $3k-3$ first order rogue waves, and the inner triangle contains $\frac{k^2-5k+6}{2}$ first order rogue waves which is similar to the *triangular* structure of $(k-3)$-th order rogue wave solution. For example, the $7$-th order rogue wave solution in Fig. \[fig.triangle\](f) is composed of $28$ first order rogue waves, $18$ first order rogue waves locating on the outer shell, and the inner is similar to the triangle of the forth order rogue wave containing $10$ first order rogue waves.
- Modified-triangular structure
Actually, the inner triangle structure can form a higher order rogue wave inversely by changing the appearance of (\[D3\]). For instance, when $k=5$, if we set
$$\label{eigenfun1}
\left(
\begin{array}{c}
f(x,t,\lambda)\\
g(x,t,\lambda)
\end{array}
\right)=\left(
\begin{array}{c}
D_1\omega^1_{11}(x,t,\lambda)+D_1\omega^2_{11}(x,t,\lambda)+D_2{\omega^1_{12}}^\ast(x,t,-\lambda^\ast)+D_2{\omega^2_{12}}^\ast(x,t,-\lambda^\ast)\\
D_1\omega^1_{12}(x,t,\lambda)+D_1\omega^2_{12}(x,t,\lambda)+D_2{\omega^1_{12}}^\ast(x,t,-\lambda^\ast)+D_2{\omega^2_{12}}^\ast(x,t,-\lambda^\ast)
\end{array}
\right),\nonumber$$
with
$$\nonumber
\left\{
\begin{aligned}
D_1=&\exp(-{\rm i}c_1^2(S_0+S_1\epsilon+S_2\epsilon^2+S_3\epsilon^3+\cdots+S_{k-1}\epsilon^{k-1})),\\
D_2=&\exp({\rm i}c_1^2(S_0+S_1\epsilon+S_2\epsilon^2+S_3\epsilon^3+\cdots+S_{k-1}\epsilon^{k-1})).
\end{aligned}
\right.$$
we will get a triangular structure with an second order rogue wave located in the center. It is remarkable that this structure has never been given before in nonlinear science, which is called *modified-triangular* structure. One with special parameter is shown in Fig. \[fig.newtriangle\].
- Ring structure
If we assume $S_i=0$ except $S_{k-1}$, we can get *ring* structures, which are shown in Fig. \[fig.circle\]. They possess $1$-order rogue waves and higher order rogue waves. Peaks locating on the outer shell of the ring are all first order rogue waves, and locating in the center of the ring are higher order rogue waves. Besides, the number of the first order rogue wave and the order of the inner higher order rogue wave increase according to the order of rogue wave. From those figures, we can conclude that there are $2k-1$ first order rogue waves locating on the outer shell of *ring* structure of the $k$-th order rogue wave solution, and a WANDT of order $k-2$ locates in the center of ring. Notably, this structure has never been displayed for the DNLS equation.
[**3.2. Solutions with more than one parameter**]{}\
Generally, there are $k-1$ free parameters for $k$-order rogue wave solution. As we discussed in previous subsection, the four basic models are depending on particular parameters. If there are two or more parameters which are non-zero, new models will be obtained. Moreover, higher order WANDT locating in the center of ring structure can be split into lower order waves.\
- Ring-triangle
When $k=4$ with parameter $S_3\neq0$, we have gotten a *ring* structure with $7$ $1$-order rogue wave solution located on the outer shell, and a $2$-order rogue wave locating in the center in above section. Further more, if the parameter $S_1$ is also non-zero, the central higher order peak is split into a *triangular* structure and the outer shell remains the same. When $k=5,6,7$, the similar structure is also displayed. These phenomenons are displayed in Fig. \[fig.circletriangle\]. Therefore, we are able to conclude that the central higher order rogue wave will be split into a *triangular* structure if $S_{k-1}\gg0$ and $S_1\neq0$ for $k$-order WANDT.\
- Multi-ring
Similarly, the central higher order WANDT in *ring* structure can also be split into *ring* structure. For instance, when $k=5$, we have observe a *ring* structure in Fig. \[fig.circle\](b). In this case, if we set $S_2\neq0$, the inner $3$-order rogue wave can be split into a *ring* model. Its dynamics are shown in Fig. \[fig.5rwmulticircle\]. When $k=6$, if we assume $S_5=1\times10^8$ and $S_4=1\times10^6$, the inner higher order rogue wave is split into a *ring* structure with a second order rogue wave locating in the center. Its evolution is shown in Fig. \[fig.6rwmulticircle1\]. Indeed, we can continuing decomposing the inner structure with the help of another parameters. A new *multi-ring* model of the six order is displayed in Fig.\[fig.6rwmulticircle2\]. From these figures, we find that both the outer shell and the middle shell are circular, and the inner shell is triangular (or circular). Naturally, the $7$-order rogue wave solution possesses the same character except for the difference that the central peak can be split into both a triangle model and a ring pattern. They are shown in Fig. \[7rwmulticircle1\] and Fig. \[7rwmulticircle2\].
[**Conclusions**]{}
===================
In this paper, we generate the formulae of higher order positon solution in proposition 1 and higher order rogue wave solution in proposition 2 for the DNLS equation at the same eigenvalue with the method of Taylor expansion and limit technique. By applying these formulae, we get positon solutions, rational traveling solutions and rogue wave solutions. These formulae are given in terms of determinants explicitly. Remarkably, the formula for rogue wave solutions is really effective in achieving the analytic expression and computer simulation of $k$-order rogue wave. Further more, we give rise to solutions with different structures are obtained by adjusting the free parameters $D_1$ and $D_2$ in our formula. With the help of these parameters, we study the dynamics of higher order rogue wave solutions. Overall, there are four basic models for higher order rogue wave. By choosing proper parameters, combination structures can be obtained. For example, *fundamental pattern*, *triangular* structure, *ring* structure,*modified-triangular* structure, *ring-triangle* structure, and *multi-ring* structure. The last three models have never been given before.
In the last part of this paper, we make a classification of higher order rogue wave of the DNLS equation. We found out that some basis structures (*fundamental pattern*, *triangular* structure, *ring* structure) also appear in other equations such as NLS. But the *modified-triangular* structure is unique to the DNLS equation.
Our results give an essential understanding of the relation of shift parameters with relative positions, which will be useful in other integrable equations such as Hirota equation, Gerdjikov-Ivanov equation, the Davey-Stewartson equation, and so on.
[**Acknowledgments**]{}
[This work is supported by the NSF of China under Grant No.11271210, No.10971109 and K. C. Wong Magna Fund in Ningbo University. Jingsong He is also supported by Natural Science Foundation of Ningbo under Grant No. 2011A610179. We want to thank Prof. Yishen Li (USTC, Hefei, China) for his long time support and useful suggestions.]{}
The expression of the second order rogue wave solution
======================================================
Here we present the expression of the second order rogue wave solution $$q_{2_{rw}}=-\frac{L_1^*L_2}{L_1^2}{\rm exp}({{\rm i} \left( ax+ \left( -{c}^{2}+a \right) at \right) }),$$ where $$\nonumber
\begin{split}
L_1=&e_1+{\rm i}e_2,\qquad
L_2=e_3+{\rm i}e_4,\\
e_1=&-9+72\,a{c}^{4}xt-216\,{a}^{2}{c}^{2}xt-108\,{a}^{2}{c}^{4}{t}^{2}-18
\,a{c}^{6}{t}^{2}+216\,{c}^{6}xt-54\,a{c}^{2}{x}^{2}-216\,{a}^{3}{c}^{2}{t}^{2}\\
&+1088\,{a}^{3}{c}^{12}{x}^{3}{t}^{3}-1536\,{a}^{8}{c}^{6}x{t}^{5}
+6528\,{a}^{6}{c}^{8}{x}^{2}{t}^{4}-3264\,{a}^{4}{c}^{10}{x}^{3}{t}^{3}
-1920\,{a}^{7}{c}^{6}{x}^{2}{t}^{4}\\
&+3456\,{a}^{5}{c}^{8}{x}^{3}{t}^{3}-456\,{a}^{3}{c}^{10}{x}^{4}{t}^{2}+648\,a{c}^{14}{t}^{4}
-1280\,{a}^{6}{c}^{6}{x}^{3}{t}^{3}+912\,{a}^{4}{c}^{8}{x}^{4}{t}^{2}-2268\,{a
}^{2}{c}^{12}{t}^{4}\\
&-480\,{a}^{5}{c}^{6}{x}^{4}{t}^{2}+4272\,{a}^{3}{c
}^{10}{t}^{4}+96\,{a}^{3}{c}^{8}{x}^{5}t-1296\,a{c}^{12}x{t}^{3}-4368
\,{a}^{4}{c}^{8}{t}^{4}-96\,{a}^{4}{c}^{6}{x}^{5}t\\
&+3456\,{a}^{2}{c}^{
10}x{t}^{3}+1728\,{a}^{5}{c}^{6}{t}^{4}-4320\,{a}^{3}{c}^{8}x{t}^{3}-8
\,{a}^{3}{c}^{6}{x}^{6}+864\,a{c}^{10}{x}^{2}{t}^{2}-216\,{a}^{3}{c}^{
18}{t}^{6}\\
&+1296\,{a}^{4}{c}^{16}{t}^{6}-3456\,{a}^{5}{c}^{14}{t}^{6}+
864\,{a}^{3}{c}^{16}x{t}^{5}+5184\,{a}^{6}{c}^{12}{t}^{6}-4320\,{a}^{4
}{c}^{14}x{t}^{5}-4608\,{a}^{7}{c}^{10}{t}^{6}\\
&+9216\,{a}^{5}{c}^{12}x{
t}^{5}-1368\,{a}^{3}{c}^{14}{x}^{2}{t}^{4}+2304\,{a}^{8}{c}^{8}{t}^{6}
-10368\,{a}^{6}{c}^{10}x{t}^{5}+5472\,{a}^{4}{c}^{12}{x}^{2}{t}^{4}-
512\,{a}^{9}{c}^{6}{t}^{6}\\
&+6144\,{a}^{7}{c}^{8}x{t}^{5}-8736\,{a}^{5}{
c}^{10}{x}^{2}{t}^{4}-96\,{a}^{3}{c}^{4}{x}^{3}t+24\,a{c}^{6}{x}^{4}-
12\,{a}^{2}{c}^{4}{x}^{4}-36\,{c}^{4}{x}^{2}-324\,{c}^{8}{t}^{2}\\
&-192\,
{a}^{6}{c}^{4}{t}^{4}+2496\,{a}^{4}{c}^{6}x{t}^{3}-1656\,{a}^{2}{c}^{8
}{x}^{2}{t}^{2}-384\,{a}^{5}{c}^{4}x{t}^{3}+1296\,{a}^{3}{c}^{6}{x}^{2
}{t}^{2}-240\,a{c}^{8}{x}^{3}t\\
&-288\,{a}^{4}{c}^{4}{x}^{2}{t}^{2}+288\,
{a}^{2}{c}^{6}{x}^{3}t,\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
e_2=&6\,{c}^{2} ( 9\,x+12\,{c}^{4}{x}^{2}ta-24\,{c}^{4}x{t}^{2}{a}^{2}
+48\,x{t}^{2}{c}^{2}{a}^{3}+24\,{x}^{2}t{a}^{2}{c}^{2}-180\,{c}^{6}{t}
^{2}xa+1776\,{c}^{8}{a}^{4}{t}^{4}x\\
&-360\,{c}^{10}{a}^{2}{t}^{3}{x}^{2}
-51\,t{c}^{2}+18\,ta+4\,{c}^{4}{a}^{2}{x}^{5}+108\,{c}^{8}{t}^{2}x-
1008\,{c}^{10}{a}^{4}{t}^{5}-240\,{c}^{6}{t}^{3}{a}^{2}-\\
&108\,{c}^{14}{
a}^{2}{t}^{5}-576\,{c}^{6}{a}^{6}{t}^{5}-64\,{c}^{4}{t}^{3}{a}^{3}+128
\,{c}^{4}{a}^{7}{t}^{5}+1056\,{c}^{8}{a}^{5}{t}^{5}+32\,{t}^{3}{c}^{2}
{a}^{4}+324\,{c}^{8}{t}^{3}a\\
&-36\,{c}^{6}t{x}^{2}+504\,{c}^{12}{a}^{3}{
t}^{5}+40\,{c}^{4}{a}^{3}t{x}^{4}+324\,{c}^{12}{a}^{2}{t}^{4}x-1200\,{
c}^{10}{a}^{3}{t}^{4}x-336\,{c}^{6}{a}^{3}{t}^{2}{x}^{3}\\
&+160\,{c}^{4}{
a}^{4}{t}^{2}{x}^{3}-44\,{c}^{6}{a}^{2}t{x}^{4}+4\,{c}^{4}{x}^{3}-108
\,{c}^{10}{t}^{3}-960\,{c}^{6}{a}^{4}{t}^{3}{x}^{2}+184\,{c}^{8}{a}^{2
}{t}^{2}{x}^{3}\\
&+320\,{c}^{4}{a}^{5}{t}^{3}{x}^{2}+4\,{x}^{3}{c}^{2}a-
1216\,{c}^{6}{a}^{5}{t}^{4}x+320\,{c}^{4}{a}^{6}{t}^{4}x+992\,{c}^{8}{
a}^{3}{t}^{3}{x}^{2} ),\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
e_3=&(e_5-e_1)c,\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
e_4=&6\,{c}^{3} ( -15\,x+108\,{c}^{4}{x}^{2}ta+456\,{c}^{4}x{t}^{2}{a}
^{2}-336\,x{t}^{2}{c}^{2}{a}^{3}-120\,{x}^{2}t{a}^{2}{c}^{2}-324\,{c}^
{6}{t}^{2}xa+2928\,{c}^{8}{a}^{4}{t}^{4}x\\
&-360\,{c}^{10}{a}^{2}{t}^{3}{
x}^{2}+21\,t{c}^{2}-90\,ta+4\,{c}^{4}{a}^{2}{x}^{5}+108\,{c}^{8}{t}^{2
}x-1584\,{c}^{10}{a}^{4}{t}^{5}-384\,{c}^{6}{t}^{3}{a}^{2}-108\,{c}^{
14}{a}^{2}{t}^{5}\\
&-1344\,{c}^{6}{a}^{6}{t}^{5}+544\,{c}^{4}{t}^{3}{a}^{
3}+384\,{c}^{4}{a}^{7}{t}^{5}+2016\,{c}^{8}{a}^{5}{t}^{5}-288\,{t}^{3}
{c}^{2}{a}^{4}+324\,{c}^{8}{t}^{3}a-36\,{c}^{6}t{x}^{2}\\
&+648\,{c}^{12}{
a}^{3}{t}^{5}+56\,{c}^{4}{a}^{3}t{x}^{4}+324\,{c}^{12}{a}^{2}{t}^{4}x-
1584\,{c}^{10}{a}^{3}{t}^{4}x-464\,{c}^{6}{a}^{3}{t}^{2}{x}^{3}+288\,{
c}^{4}{a}^{4}{t}^{2}{x}^{3}\\
&-44\,{c}^{6}{a}^{2}t{x}^{4}
+4\,{c}^{4}{x}^{
3}-108\,{c}^{10}{t}^{3}-1664\,{c}^{6}{a}^{4}{t}^{3}{x}^{2}+184\,{c}^{8
}{a}^{2}{t}^{2}{x}^{3}+704\,{c}^{4}{a}^{5}{t}^{3}{x}^{2}-12\,{x}^{3}{c
}^{2}a\\
&-2496\,{c}^{6}{a}^{5}{t}^{4}x+832\,{c}^{4}{a}^{6}{t}^{4}x+1344\,
{c}^{8}{a}^{3}{t}^{3}{x}^{2}),\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
e_5=&36+288\,a{c}^{4}xt-576\,{a}^{2}{c}^{2}xt+288\,{a}^{2}{c}^{4}{t}^{2}+
432\,a{c}^{6}{t}^{2}+864\,{c}^{6}xt-144\,a{c}^{2}{x}^{2}-576\,{a}^{3}{
c}^{2}{t}^{2}\\
&+1296\,{a}^{2}{c}^{12}{t}^{4}-4032\,{a}^{3}{c}^{10}{t}^{4
}+4032\,{a}^{4}{c}^{8}{t}^{4}-1728\,{a}^{2}{c}^{10}x{t}^{3}-768\,{a}^{
5}{c}^{6}{t}^{4}+3072\,{a}^{3}{c}^{8}x{t}^{3}\\
&-384\,{a}^{3}{c}^{4}{x}^{
3}t-48\,{a}^{2}{c}^{4}{x}^{4}-144\,{c}^{4}{x}^{2}-1296\,{c}^{8}{t}^{2}
-768\,{a}^{6}{c}^{4}{t}^{4}+288\,{a}^{2}{c}^{8}{x}^{2}{t}^{2}-1536\,{a
}^{5}{c}^{4}x{t}^{3}\\
&+576\,{a}^{3}{c}^{6}{x}^{2}{t}^{2}-1152\,{a}^{4}{c
}^{4}{x}^{2}{t}^{2}+192\,{a}^{2}{c}^{6}{x}^{3}t.
\end{split}$$
With the help the formula , we can also obtain the expression of $k$-th (k=3,4,5,6,7) order rogue wave. Since they are too complicated to write down, we omit them.
[200]{} Mio K, Ogino T, Minami K and Taketa S 1976 [ J. Phys. Soc. Japan]{} [**41**]{}, 265–271.
Mj$\phi$lhus E 1976 [ J. Plasma Phys.]{} [**16**]{}, 321–334.
Mj$\phi$lhus E 1989 [ Phys. Scr]{}. [**40**]{}, 227–237.
Mj$\phi$lhus E and Hada T 1997 [*Nonlinear Waves and Chaos in Space Plasma*]{} ed T Hada and H Matsumoto (Tokyo: Terra Sci.) p, 171.
Ruderman M S 2002 [ J. Plasma Phys]{}. [**67**]{}, 271–276.
Fedun V, Ruderman M S and Erdélyi R 2008 [ Phys. Lett]{}. A [**372**]{}, 6107–6110.
Tzoar N and Jain M 1981 [ Phys. Rev]{}. A [**23**]{}, 1266–1270.
Anderson D and Lisak M 1983 [ Phys. Rev]{}. A [**27**]{}, 1393–1398.
Govind P A 2001 [*Nonlinear Fibers Optics*]{} 3rd edn (New York: Academic).
Kaup D J and Newell A C 1978 [ J. Math. Phys]{}. [**19**]{}, 798–801.
Nakamura A and Chen H H 1980 [J. Phys. Soc. Japan]{} [**49**]{}, 813–816.
Huang N N and Chen Z Y 1900 [J. Phys. A: Math. Gen]{} [**23**]{}, 439–453.
Kawata T and Inoue H 1978 [J. Phys. Soc. Japan]{} [**44**]{}, 1968–1976.
Ichikawa Y. Konno K, Wadati M and Sanuki H 1980 [J. Phys. Soc. Japan]{} [**48**]{}, 279–286.
Chen X J and Lam W K 2004 [ Phys. Rev.]{} E [**69**]{}, 066604.
A R Osborne, [*Nonlinear Ocean Waves And The Inverse Scattering Transform*]{} (Academic Press, New York 2010).
E Pelinovsky and C Kharif, [*Extreme Ocean Waves*]{} (Springer, Berlin, Heidelberg, 2008).
D R Solli, C Ropers, P Koonath, and B Jalali 2007 Nature (London) [**450**]{}, 1054–1058.
D R Solli, C Ropers, and B Jalali 2008 Phys. Rev. Lett. [**101**]{}, 233902.
J M Dudley, G Genty, and B J Eggleton 2008 Opt. Express [**16**]{}, 3644–3651.
Yu V Bludov, V V Konotop and N Akhmediev 2009 [Phys. Rev.]{} A [**80**]{}, 033610.
M Shats, H Punzmann, H Xia 2010 Phys. Rev. Lett. [**104**]{}, 104503.
Akhmediev N, Ankiewicz A, and Taki M 2009 [Phys. Lett.]{} A [**373**]{}, 675–678.
D H Peregrine 1983 J. Austral. Math. Soc. Ser. B [**25**]{}, 16–43.
B Kibler, J Fatome, C Finot, G Millot, F Dias, G Genty, N Akhmediev, J M Dudley 2010 Nature Phys. [**6**]{}, 790–795. A Chabchoub, N P Hoffmann, and N Akhmediev 2011 Phys. Rev. Lett. [**106**]{}, 204502.
H Bailung, S K Sharma, and Y Nakamura 2011 Phys. Rev. Lett. [**107**]{}, 255005.
S Xu, J He and L Wang 2011 [J. Phys. A: Math. Theor.]{} [**44**]{}, 305203.
Boling Guo, Liming Ling and Q. P. Liu 2012 Stud. Appl. Math. DOI: 10.1111/j.1467–9590.2012.00568.x.
Kenji Imai 1999 J. Phys. Soc. Japan [**68**]{}, 355–359.
J S He, H R Zhang, L H Wang, K Porsezian, A S Fokas 2012 arXiv:1209.3742v3 \[nlin.SI\].
D J Kedziora, A Ankiewicz, and N Akhmediev 2011 Phys. Rev. E [**84**]{}, 056611.
A Ankiewizc, D J Kedziora and N Akhmediev 2011 [Phys. Lett.]{} A [**375**]{}, 2782–2785.
![1-rogue wave solution with $a=1$, $c=1$.[]{data-label="fig.1rw"}](1RW.eps "fig:"){width="4cm" height="4cm"} ![1-rogue wave solution with $a=1$, $c=1$.[]{data-label="fig.1rw"}](1RWd.eps "fig:"){width="4cm" height="4cm"}
\
\
\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the context of the standard model of particle physics, there is a definite upper limit to the density of stable compact stars. However, if there is a deeper layer of constituents, below that of quarks and leptons, stability may be re-established far beyond this limiting density and a new class of compact stars could exist. These objects would cause gravitational lensing of white dwarfs and gamma-ray bursts, which might be observable as a diffraction pattern in the spectrum. Such observations could provide means for obtaining new clues about the fundamental particles and the origin of cold dark matter.'
author:
- 'F. Sandin[^1]'
title: 'Compact stars in the standard model - and beyond'
---
Introduction
============
The different types of compact objects traditionally considered in astrophysics are white dwarfs, neutron stars (including quark and hybrid stars), and black holes. The first two classes are supported by Fermi degeneracy pressure from their constituent particles. For white dwarfs, electrons provide the pressure counterbalancing gravity. In neutron stars, the neutrons play this role. For black holes, the degeneracy pressure is overcome by gravity and the object collapses to a singularity, or at least to the Planck scale ($\rho \sim 10^{93}$ g/cm$^3$). For a recent review of neutron stars, hybrid stars, and quark stars, see, [*e.g.*]{}, [@Weber:04] and references therein.
The distinct classes of degenerate compact stars originate directly from the properties of gravity, as was made clear by a theorem of Wheeler and collaborators in the mid 1960s [@Harrison:65]. This theorem states that for the solutions to the stellar structure equations, whether Newtonian or relativistic, there is a change in stability of one radial mode of vibration whenever the mass reaches a local maximum or minimum as a function of the central density. The theorem assures that distinct classes of stars, such as white dwarfs and neutron stars, are separated in central density by a region in which there are no stable configurations.
In the standard model of quarks and leptons (SM), the theory of the strong interaction between quarks and gluons predicts that with increasing energy and density, the coupling between quarks asymptotically fades away [@Gross:73; @Politzer:73]. As a consequence of this so-called asymptotic freedom, matter is expected to behave as a gas of free fermions at sufficiently high densities. This puts a definite upper limit to the density of stable compact stars, since the solutions to the stellar equations end up in a never-ending sequence of unstable configurations, with increasing central density. Thus, in the light of the standard model, the densest stars likely to exist are neutron stars, quark stars, or the slightly more dense hybrid stars [@Gerlach:68; @Glendenning:00; @Schertler:00]. However, if there is a deeper layer of constituents, below that of quarks and leptons, the possibility of a new class of compact stars opens up [@Hansson:04].
Though being a quantitatively successful theory, the SM consists of a large number of exogenous [*ad hoc*]{} rules and parameters, which were introduced solely to fit the experimental data. The SM provides no explanation for the deeper meaning of these rules. At a closer look, however, the SM seems to be full of hints to its deeper background. By considering these rules from a historical point of view, a “simple" and appealing explanation is [*compositeness*]{} [@Fredriksson:03], [*i.e.*]{}, that the quarks, leptons, and gauge bosons are composite particles, built out of more elementary [*preons*]{} [@dSouza:92]. Preons provide natural explanations for the particle families of the SM and phenomena such as neutrino oscillations, mixing of the weak gauge bosons, and quarks of different flavour.
Over the last decades, many papers have been written about preons, but so far there are no direct evidence for (or against!) their existence. In the late 1970s, a number of consistency conditions were formulated by ’t Hooft [@tHooft:79]. In the same work, a vector-like non-Abelian $SU(3)$ gauge group was considered, but no solution to the consistency conditions was found. Later, it was shown that with another choice for the gauge group and the flavour structure of preons, [*e.g.*]{}, three different preon flavours, the consistency conditions are satisfied [@Barbieri:80]. For a more detailed discussion of preon models, see [@Fredriksson:03; @dSouza:92; @PreonTrinity:02] and references therein.
Not all clues favour preon models, but the existence of preons is still an open question, and as a consequence so is the question whether a new class of compact stars exists or not. This paper is based on the ideas and results presented in [@Hansson:04]. Assuming that quarks and leptons are composite particles, built out of more elementary preons, I will:
[.]{}
[Give an estimate for the mass and radius of stars composed of preons.]{}
[Show that for some particular equations of state, stable solutions to the general relativistic stellar equations do exist, with densities far beyond the maximum density in stars composed of quarks and leptons.]{}
[Briefly discuss some potential astrophysical consequences and how these objects could be observed. Herein lies the potential importance of this qualitative speculation, since these objects are candidates for cold dark matter and could be found as, [*e.g.*]{}, gravitational “femtolenses".]{}
Compact stars and the maximum density prophecy
==============================================
In order to explain why there is a maximum density for stars composed of quarks and leptons, or any other composite particle composed of these two species, [*e.g.*]{}, nucleons and $^{56}$Fe, some basic knowledge about the theory of compact stars is needed. In the following, I give a short introduction and a summary of the main points.
Due to the high density and large mass of compact stars, a general relativistic treatment of the equilibrium configurations is necessary. This is especially important for the analysis of stability when a star is subject to radial oscillations. Such oscillations are excited to some extent, and for a star to be stable the amplitude of the oscillations must not grow spontaneously with time. The starting point for a general relativistic consideration of compact stars is the Oppenheimer-Volkoff (OV) equations [@OV] for hydrostatic, spherically symmetric equilibrium: $$\begin{aligned}
\frac{dp}{dr} &=& -\frac{G\left(p+\rho c^2\right)\left(m c^2 + 4 \pi r^3 p\right)}
{r \left(r c^4 - 2 G m c^2 \right)}, \label{ov1} \\
\frac{dm}{dr} &=& 4 \pi r^2 \rho. \label{ov2}\end{aligned}$$ Here $p$ is the pressure, $\rho$ the density and $m=m(r)$ the mass within the radial coordinate $r$. The total mass of the star is: $$M = 4\pi \int_0^R r^2 \rho\,dr,$$ where $R$ is the coordinate radius of the star. Combined with an equation of state (EOS) $p = p(\rho)$ obtained from some microscopic (quantum field) theory, the OV solutions give the possible equilibrium states of spherically symmetric stars.
As an example, I show two sequences of compact star configurations. One is composed of nuclear matter (neutron stars) and the other of a deconfined quark matter core with a nuclear matter crust (hybrid stars), see [Fig. \[ordinary\_stars\]]{}.
These configurations were obtained by solving the OV equations [(\[ov1\])-(\[ov2\])]{} numerically. The low-density part of the nuclear matter EOS was extracted from [@NV:71] and the high-density part comes from [@APR:98]. For the deconfined quark matter phase an unpaired massless quark approximation $\rho c^2 = 3p + 4B$ was used, and the “bag pressure" $B$ was fitted such that the transition from quark matter to nuclear matter occurs at $0.01$ times the nuclear saturation density, $n_0 \sim 0.16$ fm$^{-3}$. The exact density where cold nuclear matter decompose into quark matter is unknown, so the transition density used here serves as an example only.
The composition of matter at neutron star densities is an open question and many different models for the EOS exist, [*e.g.*]{}, EOSs for nuclear matter, matter with hyperons, and superconducting quark matter. Regardless of the specific model, the maximum mass and corresponding radius are roughly one solar mass $M_\odot \simeq 2\times 10^{30}$ kg and $10$ km. No substantially more dense configurations composed of quarks and leptons are possible. The motivation goes roughly like this: At white dwarf densities, the nucleons occupy nuclei that contribute little to the pressure, and electrons provide the pressure counterbalancing gravity. With increasing density, the pressure rises and the electrons become more energetic. Eventually, the electrons are captured by protons and the pressure drops. As a consequence, the white dwarf sequence becomes unstable and terminates. At roughly six to seven orders of magnitude higher density than in the maximum-mass white dwarf, nuclei dissolve and the Fermi pressure of nucleons (in nuclear matter) and quarks (in quark matter) stabilize the next sequence of stable stars. The maximum mass of this sequence is of the order of one solar mass, for all compositions (nuclear matter, strange matter, hyperons etc.).
The reason why this is the limiting mass of stable compact stars composed of quarks and leptons is simply that there is no particle that may stabilize another sequence of stars. Each quark flavour is accompanied by an extra Fermi sea that relieves the growth of pressure, and quark Fermi pressure is only won at the expense of pressure from other species. Also, the chemical potential in stable hybrid stars is much smaller than the charm mass, so stars composed of quarks heavier than the strange quark do not exist [@Kettner:95; @Prisznyak:94].
Hence, beyond the very rich and beautiful landscape of structures composed of quarks and leptons, at $\sim 10^{16}$ g/cm$^3$, there is again a desert of instability, just like there are no stable stellar configurations in-between white dwarfs and neutron stars. The question is now if the desert ends before the Planck scale.
Compact stars beyond the desert
===============================
A definite upper limit to the density of any static spherically symmetric star can be obtained from the Schwarzschild radius, $$R = 2GM/c^2,
\label{schwarz}$$ since any object more dense than this would collapse into a black hole. By using the expression for the Schwarzschild radius and the relations: $$\begin{aligned}
M &\sim& m A, \label{msimple} \\
R &\sim& d_0 A^{1/3}, \label{rsimple}\end{aligned}$$ where A is the number of constituent particles, $m$ their mass and $d_0$ the distance between adjacent particles, an order of magnitude estimate for the maximum mass and radius of the corresponding class of compact stars can be calculated. For an idealized neutron star composed of nucleons of mass $m_n \simeq 939$ MeV$/c^2$ and size $d_n\simeq 0.5\times 10^{-15}$ m [(\[schwarz\])-(\[rsimple\])]{} give $A\sim 3\times 10^{57}$ baryons, $R\sim 7$ km and $M \sim 5 \times 10^{30}$ kg$\,\sim 2.5$ M$_\odot$. In reality, a somewhat larger radius and smaller mass are expected, since the density is non-uniform in the star, say $R \sim 10$ km and $M \sim 2$ M$_\odot$. In any case, the correct order of magnitude for the maximum mass and corresponding radius of a neutron star is obtained. The average density is $\bar{\rho}\simeq 10^{15}$ g/cm$^3$.
Since the Schwarzschild limit is almost reached already for the most massive neutron stars, it is reasonable to assume that this should be the case also for a more dense class of compact stars. Then, in order to provide similar estimates for the mass and radius of a star composed of preons, something must be known about the mass and “size" of preons. Before trying to achieve this, it should be emphasized that we know nothing about preons, not even if they exist. So whatever method used, the result is a speculative order of magnitude estimate. But as I will show, it is still possible to reach some qualitative results.
Guided by the observation that the density of nuclear matter is roughly of the same order of magnitude as for deconfined quark matter, I assume that the density of preon matter is roughly of the same order of magnitude as for a closely spaced lattice of some “fundamental" particle of the SM. In this case the problem is simplified to finding a fundamental SM particle, with known mass and maximum spatial extension. The simplest and least ambiguous choice seems to be the electron, since the mass of an electron is well known, and from scattering experiments it is known that electrons do not have any visible substructure down to a scale of $\sim \hbar c/$TeV$\,\sim 10^{-19}$ m. Using the electron mass $m_e \simeq 511$ keV$/c^2$ and an upper estimate for its radius $r_e \sim 10^{-19}$ m, the maximum mass and radius of a star composed of preons is found to be of the order $M\sim 10^2$ M$_\oplus$ and $R\sim 1$ m. Here M$_\oplus\simeq 6\times10^{24}$ kg is the mass of the Earth. The average density is of the order $\sim 10^{23}$ g/cm$^3$.
This crude estimate gives metre-sized objects that are a hundred times more massive than the Earth. Now, I will try to be a bit more specific. Especially, it would be interesting to see whether such objects could be stable or not. In order to do this, I extrapolate an effective model for hadrons, the so-called MIT bag model [@Chodos:74]. In its simplest form, the MIT bag is a gas of massless fermions (partons), enclosed in a region (the bag) and subject to an external pressure $B$ (the bag constant). The EOS for a gas of massless fermions is $\rho c^2=3 p$ and by including $B$ one obtains: $$\rho c^2=3p+4B.$$ This result does not depend on the degeneracy factor, [*i.e.*]{}, the number of fermion species, spin, etc. For a single hadron the pressure is practically zero, so that $\rho c^2=4B$ and the total energy $E$ of a hadron is [@Chodos:74]: $$E=4B{\langleV\rangle},$$ where ${\langleV\rangle}$ is the time-averaged volume of the bag (hadron). Hence, the bag pressure $B$ must be of the order of $1$ GeV/fm$^3\sim 10^{35}$ Pa for hadrons. This is in agreement with experiments and other independent methods of calculating light-quark hadronic masses; Most of the mass-energy is not due to the “bare mass" of the constituents, but the confining interactions.
The MIT bag model is frequently used for the description of deconfined quark matter and applications to compact stars. Its usefulness in this regime originates in asymptotic freedom, simplicity and the possibility to include pertubative corrections. The bag pressure $B$ is introduced in order to confine partons, it is a phenomenological parametrization of the strong interactions that confine quarks into hadrons. These interactions are present also in deconfined quark matter, so the “bag model" should be applicable also in this regime. However, the value of the bag pressure is different, since the density is higher and the interactions weaker. Thus, the so-called bag constant $B$ is not really a constant, but a density dependent parameter. For strange quark matter, the bag constant is roughly $B^{1/4} \sim 150$ MeV$/(\hbar c)^{3/4}$ [@Glendenning:97] and the corresponding contribution to the energy density is $4B \sim 260$ MeV/fm$^3$. This means that a considerable fraction of the density in quark matter, roughly $10^{15}$ g/cm$^3$, is due to the bag constant, [*i.e.*]{}, interactions.
Now, the fundamental assumption here is that preons exist and are fermions. Since preons constitute light particles, such as neutrinos and electrons, the “bare" preon mass should be fairly small. Then a massless fermion approximation, $\rho c^2=3p$, can be used. This EOS does not allow for stable super-dense stars, however, so something more is needed. And that ‘something’ is dynamics, the preon interactions that give mass-energy to the particles of the SM. The question is how, since there is so far no quantitative model for preon interactions. Indeed, a fundamental problem in preon models is to find a suitable dynamics, capable of binding preons into fermions with masses essentially negligible with respect to their inverse radius. With this in mind, the principle of parsimony (“Occam’s razor") seems to be the only guidance. A simple solution is to include the dynamics in terms of a bag constant [@Hansson:04], which roughly reproduces the minimum energy density of an electron, $$\begin{aligned}
B &=& \frac{E}{4{\langleV\rangle}} \sim \frac{3\times 511\,{\rm keV}}{16\pi(10^{-19}\,{\rm m})^3}
\sim 10^{4}\,{\rm TeV/fm}^{3} \nonumber \\
&& \Longrightarrow\, B^{1/4} \sim 10\,{\rm GeV}/(\hbar c)^{3/4}.\end{aligned}$$ The very high density contribution from the bag constant $4B/c^2 \sim 10^{5}$ TeV$\,c^{-2}fm^{-3}\sim10^{23}$ g/cm$^3$ might seem a bit peculiar. But then it should be kept in mind that the density contribution from the bag constant in deconfined quark matter is $\sim 10^{15}$ g/cm$^3$, which is a large fraction of the maximum density in any type of star composed of deconfined quark matter. So the high density is not that peculiar. On the contrary, if something is to be expected, it should be that $B$ is much higher for preon matter than for quark matter, since the typical “preon bag" is much smaller and more dense than a hadron. In the following, for simplicity, I put $\hbar c=1$ for the bag constant and express $B^{1/4}$ in eV.
The density introduced by the bag constant is of the same order of magnitude as the density used in the mass-radius estimate above. The improvement here is the transition to a proper EOS for fermions; the possibility to apply the EOS in a general relativistic framework, for the analysis of mass-radius relations and stability. In addition to the full general relativistic analysis, the mass and radius can be estimated from first principles as a function of the bag constant [@Banerjee:00]. The result is somewhat similar to the original Chandrasekhar limit, but the role of the fermion mass is replaced by the bag constant $B$, $$\begin{aligned}
M &=& \frac{16 \pi}{3 c^2} B R^3, \label{mchandra} \\
R &=& \frac{3 c^2}{16 \sqrt{\pi G B}}. \label{rchandra}\end{aligned}$$ Inserting $B^{1/4} \sim 10$ GeV in [(\[mchandra\])-(\[rchandra\])]{}, an estimate for the (maximum) mass $M \simeq 10^2$ M$_\oplus$ and radius $R \sim 1$ m is obtained. This is consistent with the somewhat simpler mass-radius estimate given above.
Since $B^{1/4}\sim 10$ GeV is only an order of magnitude estimate for the lower limit, the bag constant is considered as a free parameter of the model, constrained by a lower limit of $B^{1/4}=10$ GeV and an upper limit chosen as $B^{1/4}=1$ TeV. The latter value corresponds to a spatial extension of the electron of the order $\sim\hbar c/10^3\,$TeV$\,\sim 10^{-22}$ m. In [Figs. \[maxmass\]-\[maxradius\]]{} the maximum mass and radius of a preon star are plotted vs. the bag constant.
A necessary, but not sufficient, condition for stability of a compact star, is that the total mass is an increasing function of the central density, ${dM/d\rho_c>0}$. This condition is a consequence of a generic microscopic relation known as Le Chatelier’s principle. Roughly, this condition implies that a slight compression or expansion of a star will result in a less favourable state, with higher total energy. Obviously, this is a necessary condition for a stable equilibrium configuration. Equally important, a star must be stable when subject to (small) radial oscillations, in the sense that the amplitude of the oscillations must not grow spontaneously with time. Otherwise a small perturbation would bring about a collapse of the star.
The equation for the analysis of such radial modes of oscillation is due to Chandrasekhar [@Chandra:64]. An overview of the theory, and some applications, can be found in [@Misner:73]. A catalogue of various numerical methods for solving the original set of equations can be found in [@Bardeen:66]. However, a far more practical form of the oscillation equations has been derived by Gondek [*et al.*]{} [@Gondek:97]. The details of the stability analysis can be found in [@Hansson:04]. Here I summarise only the main points.
Assuming a time dependence of the radial displacement of fluid elements of the form $e^{i\omega t}$, the equation governing the radial oscillations is found to be a Sturm-Liouville eigenvalue equation for $\omega^2$. A necessary, and sufficient condition for stability is that all $\omega_i^2$ are positive, since imaginary frequencies give exponentially increasing amplitudes. Furthermore, since $\omega_i^2$ are eigenvalues of a Sturm-Liouville equation, it turns out that it is sufficient to prove that the fundamental (nodeless) mode, $\omega^2_0$, is positive for a star to be stable. In [Fig. \[b100gev\]]{}, the first three oscillation frequencies, $f_i=\omega_i/2\pi$, for various stellar
configurations with $B^{1/4}=100$ GeV are plotted. In agreement with the theorem of Wheeler [*et al.*]{} [@Harrison:65] the onset of instability is the point of maximum mass, as $\omega_0^2$ becomes negative for higher central densities. Thus, for $B^{1/4}=100$ GeV, preon stars are stable up to the maximum mass configuration. The same is true for other values of $B$ [@Hansson:04].
Despite the large uncertainty regarding preon interactions, here manifested as a large uncertainty in the bag constant, preon stars should have central densities beyond $10^{23}$ g/cm$^3$. This makes preon stars fundamentally different from the traditional types of compact stars, since such high central densities implies that the stars must be very small and light in order to be stable, see [Fig. \[classes\]]{}.
Formation, detection and astrophysical consequences
===================================================
The list of possible connections between the properties of the fundamental particles and the large scale structures in the universe is long. However, beyond a density of $\sim 10^{23}$ g/cm$^3$, not much has been proposed, since there are strong arguments against the existence of stable objects beyond $\rho \sim 10^{16}$ g/cm$^3$. That is, if quarks and leptons are fundamental entities.
If preons exist and objects composed of preon matter as small and light as suggested here are stable, density fluctuations in the early universe might have produced primordial preon stars (or “nuggets"). As this material did not take part in the ensuing nucleosynthesis, the abundance of preon nuggets is not constrained by the hot big bang model bounds on baryonic matter. Also, preon nuggets are immune to Hawking radiation [@Hawking:75], which rapidly evaporates small primordial black holes, making it possible for preon nuggets to survive to our epoch. They can therefore serve as the mysterious dark matter needed in many dynamical contexts in astrophysics and cosmology [@Turner:00; @Bergstrom:00]. The idea that preons could be connected to dark matter is already recognized in the literature [@Burdyuzha:98; @Lalakulich:98], but the picture presented here is rather different.
The Friedmann equation for the early universe is $H^2(t)=8\pi G\rho/3$ and in the radiation-dominated epoch $\rho c^2 \sim T^4$ (Boltzmann’s law). When including the number of internal degrees of freedom $g_{eff}$, an expression for the Hubble parameter $H$, in units where $\hbar=c=k_B=1$, is [@Bergstrom:04]: $$H \simeq 1.66 \sqrt{g_{eff}}\frac{T^2}{m_{pl}}.$$ Here $T$ is the temperature in eV and $m_{pl} \simeq 1.2 \times 10^{19}$ GeV/$c^2$ the Planck mass. For the SM, the fermions and the gauge and Higgs bosons give $g_{eff}(T=1\,{\rm TeV}) = 106.75$. In the preon phase, this number should be smaller, say $g_{eff} \sim 10$ for simplicity. Then the Hubble radius at a temperature of $1$ TeV is $H^{-1} \sim 1$ mm and the mass within the Horizon (a causally connected region) is $\rho H^{-3} \sim 10^{-1}$ M$_\oplus$. This is the maximum mass of any structure that could have formed in this early epoch. Hence, the maximum mass within causally connected regions, at the minimum temperature when deconfined preon matter might have formed preon nuggets (and the particles of the SM), is of the correct order of magnitude for stable configurations.
A potential problem is that the Jeans length, which defines the minimum length scale of regions that can contract gravitationally, was roughly of the same order of magnitude as the Hubble radius at that temperature. The Jeans length $\lambda_J$ is [@Bergstrom:04]: $$\lambda_J = v_s\sqrt{\frac{\pi}{G\rho_0}},$$ where $v_s$ is the speed of sound and $\rho_0$ the average background density. For a relativistic fluid with EOS $\rho c^2 = 3p + 4B$, the speed of sound is $v_s=c/\sqrt{3}$ and $\lambda_J \sim 1$ mm $\sim H^{-1}$. However, considering the high level of approximation used here, this is not yet a serious problem. It merely shows that the numbers are in the correct intervals.
But, perhaps it will be the other way around. After all, Popper’s idea that we make progress by falsifying theories is not always true. By utilizing gamma-ray bursts (GRB) or white dwarfs in the large magellanic cloud as light sources, gravitational lenses with very small masses produce a diffraction signal in the spectrum, which might be observable [@Gould:92; @Stanek:93; @Ulmer:95; @Nemiroff:95]. For a lens with mass within the range $10^{-16}$ M$_\odot \leq M \leq 10^{-11}$ M$_\odot$, the angular separation of images would be in the femto-arcsec range (femtolensing), and for more massive lenses, $M \leq 10^{-7}$ M$_\odot$, the angular separation is in the pico-arcsec range (picolensing). The mass within the Hubble radius at $T=1$ TeV is $\sim 10^{-1}$ M$_\oplus\sim 10^{-7}$ M$_\odot$. This roughly defines the maximum mass of preon nuggets that could be abundant enough to be observed as gravitational lenses. Hence, preon nuggets fall in the correct mass range for picolensing and femtolensing.
In [Fig. \[magnification\]]{} the magnification of a distant light source due to gravitational lensing by an intermediate preon star is plotted as a function of the dimensionless frequency: $$\nu = \frac{\tilde{\nu}(1+z_L)2GM}{c^3}.$$ Here, $M$ is the mass of the lens, $z_L$ the redshift of (distance to) the lens and $\tilde{\nu}$ the frequency of light. This result was calculated with a physical-optics model, as described in [@Ulmer:95]. In principle, the time dependent amplitude due to a single light pulse from the source was calculated, and then the power spectrum was obtained by a Fourier transform of the amplitude. The magnification is normalized to a unit flux in the absence of a lens, [*i.e.*]{}, the signal is the observed flux of the source times the magnification. The exact shape of the magnification curve depends on the relative position of the source and lens. Here the source is slightly off-axis (corresponding to $\theta=0.2$ in [@Ulmer:95]).
As mentioned in [@Hansson:04], preon stars could also form in the collapse of normal massive stars, if the collapse is slightly too powerful for the core to stabilize as a neutron star, but not sufficiently violent for the formation of a black hole. Due to the potentially very large magnetic field and rapid rotation of preon stars formed in this way, the astrophysical consequences could be important, [*e.g.*]{}, for acceleration of ultra-high energy (UHE) cosmic rays. However, even though the idea might be appealing at first sight, the possibility to expel such a large fraction of the mass of the progenitor star needs to be better understood. What should be noted here is merely a potential connection to UHE cosmic rays, which might provide a second means for locating and observing preon stars.
Conclusions
===========
If there is a deeper layer of fermionic constituents (preons), below that of quarks and leptons, a new class of stable compact stars could exist. By fitting a simple equation of state for fermions to the minimum energy density of an electron, the maximum mass for stars composed of preons can be estimated to $\sim 10^{2}$ Earth masses and the maximum radius to $\sim 1$ m. The minimum central density is of the order of $\sim 10^{23}$ g/cm$^3$. Preon stars with a maximum mass of $\sim 10^{-1}$ Earth masses and radius $\sim 1$ mm could have been formed by the primordial density fluctuations in the early universe. By utilizing gamma-ray bursts, or white dwarfs in the large magellanic cloud as light sources, an intermediate preon star would produce a diffraction signal in the spectrum, which might be observable. An observation of an object as dense as a preon star would be a direct evidence for a new state of matter, which is not composed of quarks and leptons. Due to the need for observational clues in the cold dark matter sector, this could prove compositeness plausible, without much dedicated effort. This approach might complement direct tests of preon models [@PreonTrinity:02] performed at particle accelerators.
Acknowledgements
================
I acknowledge support from the Swedish National Graduate School of Space Technology. I thank S. Fredriksson and J. Hansson for several useful discussions and for reading the manuscript. I thank M. Alford for providing me with the data tables of the high density nuclear matter equations of state, J. Goodman for useful discussions regarding the femtolensing signature, and J. Bourjaily, D. Casadei, A. Geiser, A. Giazotto, and J. Kamenik for interesting discussions about the astrophysical consequences of preon stars. Finally, I thank G. ’t Hooft and A. Zichichi for organizing an excellent 42nd course of the international school of subnuclear physics, and for the award that was designated this work.
[5]{}
F. Weber, [*Strange Quark Matter and Compact Stars*]{}, astro-ph/0407155 (submitted).
B.K. Harrison, K.S. Thorne, M. Wakano & J.A. Wheeler, [*Gravitation Theory and Gravitational Collapse*]{} (University of Chicago Press, Chicago, 1965).
D.J. Gross & F. Wilczek, Phys. Rev. Lett. [**30**]{}, 1323 (1973).
H.D. Politzer, Phys. Rev. Lett. [**30**]{}, 1346 (1973).
U.H. Gerlach, Phys. Rev. [**172**]{}, 1325 (1968),\
U.H. Gerlach, Ph.D. thesis, Princeton University, 1968.
N.K. Glendenning & C. Kettner, Astron. Astrophys. [**353**]{}, L9 (2000).
K. Schertler [*et al.*]{}, Nucl. Phys. [**A 677**]{}, 463 (2000), astro-ph/0001467.
J. Hansson & F. Sandin, [*Preon stars: a new class of cosmic compact objects*]{}, 2004, astro-ph/0410417.
S. Fredriksson, in [*Proc. of the Fourth Tegernsee Int. Conf. on Particle Physics Beyond the Standard Model, 2004*]{}, ed. by H.-V. Klapdor-Kleingrothaus (Springer-Verlag, Heidelberg, 2004), p. 211, hep-ph/0309213.
I.A. D’Souza & C.S. Kalman, [*Preons*]{} (World Scientific, Singapore, 1992).
G. ’t Hooft, Cargèse Lecture Notes, 1979.
R. Barbieri, L. Maiani & R. Petronzio, Phys. Lett. [**B 96**]{}, 63 (1980).
J.-J. Dugne, S. Fredriksson & J. Hansson, Europhys. Lett. [**57**]{}, 188 (2002).
J.R. Oppenheimer & G. Volkoff, Phys. Rev. [**55**]{}, 374 (1939).
J.W. Negele & D. Vautherin, Nucl. Phys. [**A 207**]{}, 298 (1973).
A. Akmal, V.R. Pandharipande & D.G. Ravenhall, Phys. Rev. [**C 58**]{}, 1804 (1998), nucl-th/9804027.
C. Kettner [*et al.*]{}, Phys. Rev. [**D 51**]{}, 1440 (1995).
M. Prisznyak, B. Lukacs & P. Levai, [*Are There Top Quarks in Superdense Hybrid Stars?*]{}, astro-ph/9412052.
A. Chodos [*et al.*]{}, Phys. Rev. [**D 9**]{}, 3471 (1974).
N.K. Glendenning, [*Compact Stars*]{} (Springer-Verlag, New York, 1997).
S. Banerjee, S.K. Ghosh & S. Raha, J. Phys. [**G 26**]{}, L1 (2000), astro-ph/0001246.
S. Chandrasekhar, Phys. Rev. Lett. [**12**]{}, 114 (1964).
C.W. Misner, K.S. Thorne & J.A. Wheeler, [*Gravitation*]{} (Freeman and Co., San Francisco, 1973).
J.M. Bardeen, K.S. Thorne & D.W. Meltzer, Astrophys. J. [**145**]{}, 505 (1966).
D. Gondek, P. Haensel & J.L. Zdunik, Astron. Astrophys. [**325**]{}, 217 (1997), astro-ph/9705157.
S.W. Hawking, Commun. Math. Phys. [**43**]{}, 199 (1975).
M.S. Turner, Phys. Rep. [**333**]{}, 619 (2000).
L. Bergström, Rep. Progr. Phys. [**63**]{}, 793 (2000).
V. Burdyuzha [*et al.*]{}, in [*Proc. of the Second Int. Workshop on Particle Physics and the Early Universe, 1999*]{}, ed. by D.O. Caldwell (Springer-Verlag, 1999), p. 392, astro-ph/9912555.
O. Lalakulich & G. Vereshkov, in [*Proc. of the Second Int. Conf. on Dark Matter in Astro and Particle Physics, 1999*]{}, ed. by H.-V. Klapdor-Kleingrothaus and L. Baudis (IOP Publishing, Great Yarmouth, 1999), p. 668.
L. Bergström & A. Goobar, [*Cosmology and Particle Astrophysics*]{}, 2nd ed. (Springer-Verlag, Germany, 2004).
A. Gould, ApJ [**386**]{}, L5 (1992).
K.Z. Stanek, B. Paczyński & J. Goodman, ApJ [**413**]{}, L7 (1993).
A. Ulmer & J. Goodman, ApJ [**442**]{}, 67 (1995), astro-ph/9406042.
R.J. Nemiroff & A. Gould, [*Probing for MACHOs of Mass $10^{-15}$M$_\odot - 10^{-7}$M$_\odot$ with Gamma-Ray Burst Parallax Spacecraft*]{}, astro-ph/9505019.
[^1]: e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present numerical and analytical results for the swiching times of magnetic nanoparticles with uniaxial anisotropy at elevated temperatures, including the vicinity of $T_{c}.$ The consideration is based in the Landau-Lifshitz-Bloch equation that includes the relaxation of the magnetization magnitude $M$. The resulting switching times are shorter than those following from the naive Landau-Lifshitz equation due to (i) additional barrier lowering because of the reduction of $M$ at the barrier and (ii) critical divergence of the damping parameters.'
author:
- 'D. A. Garanin$^1$ and O. Chubykalo-Fesenko$^2$'
title: 'Thermal fluctuations and longitudinal relaxation of single-domain magnetic particles at elevated temperatures'
---
The theory of thermal fluctuations of small magnetic particles is one of the fundamental issues of modern micromagnetics. The conditions at which the particle becomes superparamagnetic define the thermal stability of the magnetized state and, therefore, is also valuable for technological application such as magnetic recording [@Weller]. The basis of the theory has been introduced by Brown [@Brown] who suggested to include thermal fluctuations into the Landau-Lifshitz (LL) dynamical equation as formal random fields whose properties are defined by the equilibrium solution of the correspondent Fokker-Planck (FP) equation. He also derived the Arrhenius-Néel formula to describe the relaxation rate in the axially symmetric case of single-domain particles which was lately generalized to the presence of external field [@Aharoni-Coffey; @Kennedy]. Since the paper of Chantrell and Lyberatos [@Lyberatos1], this Langevin-dynamics approach of Brown has been brought to numerical micromagnetics to model the thermal properties of an ensemble of interacting particles and, more generally, of ferromagnetic materials, interpreting the micromagnetic discretization elements as small particles. Generally speaking, the LL equation used in these simulations is a low-temperature approximation only. Recently a generalized Landau-Lifshitz-Bloch (LLB) equation for a ferromagnet [@Panina; @LLB] was derived which is valid for all temperatures and includes the longitudinal relaxation. The deviations of the LLB dynamics from the LL dynamics should be pronounced at high temperatures, especially close to the Curie temperature $T_{c}$. The validity of this approach has been confirmed by the measurements of the domain-wall mobility in crystals of Ba-and Sr-hexaferrites close to $T_{c}$ [@LLB-DW].
Since the proposal of the heat-assisted magnetic recording (HAMR) [@HAMR] the problem of high-temperature thermal magnetization dynamics has become of large practical importance. The basic idea of HAMR is to write bits of information at elevated temperature (close to the Curie temperature, where the switching field is small) and store the information at room temperature. To achieve a significant areal density advantage, the use of high-anisotropy intermetallics such as $Ll_{0}$ FePt has been suggested [@Sun]. Therefore, from both fundamental and applied points of view it is necessary to consider the micromagnetics of small particles (or magnetic grains) at elevated temperatures. The straightforward approach [@Lyberatos2] uses the formalism of the standard stochastic LL equation, however with the temperature-dependent parameters, i.e., the equilibrium magnetization $M_{e}(T)$ introduced through the mean-field approximation (MFA) involving the Brillouin function, and the uniaxial anisotropy $K(T)$ through the scaling relation $K(T)\sim M_{e}^{2}(T)$. However, this approach becomes invalid at elevated temperatures as it does not incorporate the longitudinal relaxational effects. The purpose of this Letter is to introduce the theoretical formalism of the thermal fluctuations of single-domain particles on the basis of the LLB equation which should be valid at all temperatures. As a practical example, we consider analytically and numerically the thermal switching of a FePt particle and discuss the conditions at which the differences between the two formalisms emerge.
We start with the LLB equation [@LLB; @LLBS] augmented by the white-noise fields $\mathbf{\zeta ,}$ $\mathbf{\zeta }_{1},$ and $\mathbf{\zeta }_{2}$ in the form $$\begin{aligned}
&&\mathbf{\dot{n}}=\gamma \lbrack \mathbf{n}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }\right) ]+\frac{\gamma \alpha _{1}}{n^{2}}(\mathbf{n\cdot }\left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }_{1}\right) )\mathbf{n} \notag \\
&&\qquad {}-\frac{\gamma \alpha _{2}}{n^{2}}[\mathbf{n}\times \lbrack
\mathbf{n}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }_{2}\right) ]], \label{LLB}\end{aligned}$$ where $\mathbf{n\equiv M/}M_{e}(T),$ is the reduced magnetization, $\gamma $ is the gyromagnetic ratio, $\alpha _{1}$ and $\alpha _{2}$ are dimensionless longitudinal and transverse damping parameters. The effective field $\mathbf{H}_{\mathrm{eff}}$ is given by $$\mathbf{H}_{\mathrm{eff}}=-\frac{\delta F}{\delta \mathbf{M}}=\mathbf{H}+\mathbf{H}_{A}+(M_{e}/\chi _{\Vert })\left( 1-n^{2}\right) \mathbf{n,}
\label{HeffDef}$$ where $F$ is the free energy density of the single-domain particle (cf. Ref. [@LLB]), $\mathbf{H}$ and $\mathbf{H}_{A}$ are applied and anisotropy fields, and $\chi _{\Vert }=\partial M_{e}/\partial H$ is the longitudinal susceptibility. Parameters $M_{e},$ $\chi _{\Vert },$ and $\alpha _{1,2}$ depend on temperature and they are singular at $T_{c}.$ Within the MFA-based model, one can use Eq. (4.9) of Ref. [@LLBS] with $K_{1}=K_{2}$ and $\gamma _{1,2}\Rightarrow \alpha _{1,2},$ rearranged to the form similar to that of of Ref. [@LLB]: $$\alpha _{1}=\frac{\lambda }{m_{e}}\frac{2T}{3T_{c}}\frac{2q}{\sinh (2q)},\quad \alpha _{2}=\frac{\lambda }{m_{e}}\left[ \frac{\tanh q}{q}-\frac{T}{3T_{c}}\right] . \label{alphas}$$ Here $\lambda $ is a microscopic damping parameter that is temperature dependent but noncritical at $T_{c}$, $$m_{e}\equiv M_{e}(T)/M_{e}(T=0) \label{meDef}$$ is the reduced magnetization, and $q=3T_{c}m_{e}/[2(S+1)T]$. One can see that in the vicinity of $T_{c}$ the relaxational parameters diverge: $\alpha
_{1}\cong \alpha _{2}\varpropto 1/M_{e}(T)$. In accordance with this theoretical prediction, ferromagnetic resonance measurements on permalloy have shown sharp increase of the damping close to the Curie temperature [@Puzlei].
The stochastic fields in the LLB-Langevin equation above can be, in fact, introduced in many different ways. For instance, one can consider all three fields as uncorrelated, set some of them to zero, or identify some of them with each other. In all these cases one obtains different LLB-Langevin equations and different stochastic trajectories. The physical quantities obtained by averaging over realizations of $\mathbf{\zeta ,}$ $\mathbf{\zeta
}_{1},$ $\mathbf{\zeta }_{2}$ are, however, the same for all models. The reason is that in all cases one obtains the same Fokker-Planck equation (FPE), as was shown in Ref. [@Garcia-Palacios] for the LL-Langevin equation. The FPE corresponding to Eq. (\[LLB\]) can be obtained in the same way as that for the LL equation (see Appendix of Ref. [@LLB]). The result has the form of the conservation law $$\frac{\partial f}{\partial t}+\frac{\partial }{\partial \mathbf{n}}\cdot
\mathbf{J}=0, \label{FPE-LLB-Red}$$ where $f\equiv f(\mathbf{n,}t)$ is the probability density and the probability current $\mathbf{J}$ reads $$\begin{aligned}
\mathbf{J} &=&[\mathbf{n}\times \mathbf{H}_{\mathrm{eff}}]f+\frac{\alpha _{1}}{n^{2}}\mathbf{n}\left( \mathbf{n\cdot }\left( \mathbf{H}_{\mathrm{eff}}-\frac{T}{VM_{e}}\frac{\partial }{\partial \mathbf{n}}\right) \right) f
\notag \\
&&-\frac{\alpha _{2}}{n^{2}}\left[ \mathbf{n}\times \left[ \mathbf{n}\times
\left( \mathbf{H}_{\mathrm{eff}}-\frac{T}{VM_{e}}\frac{\partial }{\partial
\mathbf{n}}\right) \right] \right] f. \label{JFPELLB}\end{aligned}$$
We will illustrate the stochastic dynamics of single-domain magnetic particles for the model with the $z$-uniaxial anisotropy, $$F_{A}=\left( M_{x}^{2}+M_{y}^{2}\right) /(2\chi _{\bot }), \label{FADef}$$ where $\chi _{\bot }$ is the transverse susceptibility that is a constant within the MFA. We use Eq. (\[FADef\]) rather than $F_{A}=-M_{z}^{2}/(2\chi _{\bot })$ to make $T_{c}$ independent of the anisotropy and thus to simplify our formalism. Eq. (\[FADef\]) could be rewritten using a generalization of the micromagnetic anisotropy $K$ as $F_{A}=\left( \nu
_{x}^{2}+\nu _{y}^{2}\right) K$ (or as $F_{A}=-\nu _{z}^{2}K$), where $\mathbf{\nu }$ is the magnetization direction vector, $\mathbf{\nu }\equiv
\mathbf{M}/M$ (see Ref. [@LLB]). This is not helpful, however, within the approach based on the LLB equation. The problem is that $K=M^{2}/(2\chi
_{\bot })$ is not a constant and even not a function of temperature (cf. $K=M_{e}^{2}(T)/(2\chi _{\bot })$ used in Ref. [@Lyberatos2]), since the magnetization magnitude $M$ can change dynamically during the thermal escape process. It is convenient to scale the free energy density as $F=\left(
M_{e}^{2}/\chi _{\bot }\right) \tilde{F}$ with $\tilde{F}$ given by[@Kachkachi] $$\tilde{F}=-\mathbf{n\cdot h+}\frac{1}{2}\left( n_{x}^{2}+n_{y}^{2}\right) +\frac{1}{4a}\left( 1-n^{2}\right) ^{2}, \label{Ftilde}$$ where $\mathbf{h\equiv }\left( \chi _{\bot }/M_{e}\right) \mathbf{H}$ and $a\equiv 2\chi _{\Vert }/\chi _{\bot }.$ We also define the temperature variable $\sigma $ similarly to Ref. [@Brown]: $$VF/T\equiv 2\sigma \tilde{F},\qquad \sigma \equiv VM_{e}^{2}/\left( 2\chi
_{\bot }T\right) . \label{betatilDef}$$ We will restrict our consideration to the case $H=0.$ In this case the minima of $\tilde{F}$ are located at $n_{x}=n_{y}=0,$ $n_{z}=\pm 1,$ and $\tilde{F}_{\min }=0.$ The saddle point of $\tilde{F}$ is $n_{z}=0$ and $n_{\bot }\equiv \sqrt{n_{x}^{2}+n_{y}^{2}}=n_{s},$ where $$n_{s}=\left\{
\begin{array}{cc}
\sqrt{1-a}, & a\leq 1 \\
0, & a\geq 1,
\end{array}
\right. \quad \tilde{F}_{\mathrm{sad}}=\left\{
\begin{array}{cc}
\left( 2-a\right) /4, & a\leq 1 \\
1/(4a), & a\geq 1.
\end{array}
\right. \label{nsRes}$$ In the limit $T\rightarrow 0$ (i.e., $\chi _{\Vert }\rightarrow 0$ and thus $a\rightarrow 0)$ Eq. (\[Ftilde\]) confines the vector $\mathbf{n}$ to the sphere $n\equiv |\mathbf{n}|=1,$ and the standard formalism based on the LL equation is recovered. At nonzero temperatures, $a>0,$ the magnetization changes its magnitude. In our model this effect is maximal at the saddle point where the magnetization vector is perpendicular to the easy axis. One can visualize the trajectory of this vector in the process of thermal activation, after averaging over fluctuations, as an ellipsis going via the saddle point. At $T=T^{\ast }$ defined by $a=1$ the ellipsis degenerates into a line. In the range $T^{\ast }\leq T<T_{c}$ one has $a>1,$ and the situation is qualitatively different. Here $\mathbf{n}$ contracts preserving its direction along the $z$ axis and goes through zero at the saddle point, then it grows in the opposite direction. These scenarios are very similar to the transformation of the domain wall structure with temperature [@LLB-DW]. Obviously the process of thermal activation of single-domain magnetic particles cannot be described on the basis of the LL equation at elevated temperatures. The crucial process of the longitudinal relaxation is captured by Eqs. (\[LLB\])–(\[JFPELLB\]) based on the LLB equation.
The escape rate in the case $T\ll \Delta U$ has the form $$\Gamma =\Gamma _{0}\exp \left( -\frac{\Delta U}{T}\right) ,\qquad \frac{\Delta U}{T}=\frac{VF}{T}\equiv 2\sigma \tilde{F}_{\mathrm{sad}}.
\label{GammaGeneral}$$ In addition to the dependence $\sigma \varpropto M_{e}^{2}(T)$ in Eq. (\[betatilDef\]) that is responsible for the barrier lowering at elevated temperatures, there is another source of the barrier lowering described by $\tilde{F}_{\mathrm{sad}}$ in Eq. (\[nsRes\]). In particular, the value of $\tilde{F}_{\mathrm{sad}}$ at $a=1$ is two times smaller than its low-temperature value, $a\rightarrow 0.$ The prefactor $\Gamma _{0}$ in Eq. (\[GammaGeneral\]) can be obtained by solving the FPE, Eq. (\[FPE-LLB-Red\]), similarly to the derivation in Ref. [@Kennedy]. For $a\lesssim 1$ the result depends on $\alpha _{2}$ only$,$ since in this case the barrier is being crossed by the pure rotation of the magnetization vector. For $a\gtrsim 1$ both the longitudinal and transverse relaxation becomes important, and it is difficult to find an analytical solution to the FPE. Fortunately, in this temperature range $\alpha _{1}$ and $\alpha _{2}$ given by Eq. (\[alphas\]) already become very close to each other, so that one can set $\alpha _{1}\Rightarrow \alpha _{2}$ everywhere. Then the calculation yields $$\begin{aligned}
\Gamma _{0} &=&\alpha _{2}\omega _{1}\sqrt{\frac{\sigma }{\pi }}\sqrt{\frac{1-n_{s}^{2}}{a}}\exp \left[ \frac{a\sigma }{2}\left( 1-\frac{1}{a}\right)
^{2}\theta (a-1)\right] \notag \\
&&\qquad \times \text{erfc}\left[ \sqrt{\frac{a\sigma }{2}}\left( 1-\frac{1}{a}\right) \right] . \label{Gamma0General}\end{aligned}$$ Here $\theta (x)$ is the step function and $\omega _{1}=\gamma M_{e}/(2\chi
_{\bot })$ is the ferromagnetic-resonance frequency. For $a\lesssim 1$ the total rate simplifies to $$\Gamma \cong 2\alpha _{2}\omega _{1}\sqrt{\frac{\sigma }{\pi }}\exp \left[
-\sigma \left( 1-\frac{a}{2}\right) \right] . \label{GammaBelow}$$ This reduces to the Brown’s formula $\Gamma =2\alpha _{2}\omega _{1}\sqrt{\sigma /\pi }e^{-\sigma }$ [@Brown] in the limit $a\rightarrow 0.$ Exactly at $a=1,$ Eqs. (\[GammaGeneral\]) and (\[Gamma0General\]) yield $\Gamma =\alpha _{2}\omega _{1}\sqrt{\sigma /\pi }e^{-\sigma /2}.$ Just below $T_{c}$ according to Eqs. (\[Ftilde\]) and (\[betatilDef\]) one has $a\sigma \varpropto \left( T_{c}-T\right) ^{2\beta -\gamma },$ where $\beta $ and $\gamma $ are the critical indices for the magnetization and susceptibility. Within the MFA $a\sigma $ remains finite at $T_{c},$ whereas for more realistic models it diverges. It makes, however, little sense to work out the appropriate limiting expressions for $\Gamma $ because near $T_{c}$ the high-barrier approximation $\Delta U\gg T$ becomes invalid. In fact, the prefactor $\Gamma _{0}$ in Eq. (\[GammaGeneral\]) does not strongly depend on $a.$ The main difference of our result from the Brown’s formula with a temperature-dependent barrier, $\Delta U\varpropto
M_{e}^{2}(T)$ is described by the two factors: (i) additional lowering of the barrier because of the non-rigid magnetization, Eqs. (\[nsRes\]) and (\[GammaGeneral\]); (ii) crytical divergence of the damping at $T_{c},$ Eq. (\[alphas\]).
Brown has obtained the $1/\sigma $ correction to the escape rate for $\sigma
\gg 1$ in the form $\Gamma =2\alpha _{2}\omega _{1}\sqrt{\sigma /\pi }e^{-\sigma }\left( 1-1/\sigma \right) $ [@Brown-2]. Within the LLB approach finding this correction in the whole range $0\leq a\leq \infty $ is a complicated task. For $a\lesssim 1$ the correction factor in Eq. (\[Gamma0General\]) simplifies to $$\left[ 1-\frac{1}{2\sigma }\left( 1+\frac{1}{n_{s}^{2}}\right) \right] .
\label{CorrFactLLB}$$
To illustrate the practical implication of our approach, we consider thermal switching of a FePt particle (magnetic grain) at high temperature. The LLB-Langevin equation, Eq. (\[LLB\]) has been integrated numerically with $\mathbf{\zeta }=0,$ $\left\langle \zeta _{i}^{\nu }\right\rangle =0,$ and $$\left\langle \zeta _{i}^{\mu }(t)\zeta _{j}^{\nu }(t^{\prime })\right\rangle
=\frac{2k_{B}T}{\gamma M_{e}\alpha _{i}}\delta _{ij}\delta _{\mu \nu }\delta
(t-t^{\prime }), \label{fluctfield}$$ where $i,j=1,2$ and $\mu ,\nu =x,y,z$.
The microscopic relaxation parameter $\lambda $ in Eq. (\[alphas\]) has been found analytically for a spin-phonon interaction[@LLBS]. However it is difficult to obtain reliable theoretical results for $\lambda $ in general, as well as to extract $\lambda $ from experiments. For our illustration below we just set $\lambda =0.1$, neglecting its temperature dependence. The values of $m_{e}(T)$ of Eq. (\[meDef\]) can be measured or obtained from the Curie-Weiss equation $m_{e}=B_{S}(m_{e}\tilde{\beta})$, where $\tilde{\beta}\equiv S^{2}J_{0}/(k_{B}T)$, $B_{S}(x)$ is the Brillouin function, and $J_{0}$ is related to the experimentally measured $T_{c}$ via $T_{c}=(1/3)S(S+1)J_{0}$ within the MFA. For FePt $T_{c}=750$ K, and the best fit for $m_{e}(T)$ is obtained with $S=3/2$ [@Lyberatos2]. For FePt we take $M_{e}(T=0)=1100$ emu/cm$^{3}$, $K(T=0)=1.24\times 10^{7}$ erg/cm$^{3}$, so that $\chi _{\bot }=M_{e}^{2}(T=0)/\left[ 2K(T=0)\right] =0.0488$ Oe cm$^{3}$/emu. In the same way, the longitudinal susceptibility $\chi
_{||}=\partial M/\partial H$ can be measured or found analytically from the MFA: $$\chi _{||}=\frac{v_{0}M_{e}^{2}(T=0)}{S^{2}J_{0}}\frac{\tilde{\beta}B_{S}^{\prime }(m_{e}\tilde{\beta})}{1-\tilde{\beta}B_{S}^{\prime }(m_{e}\tilde{\beta})}, \label{suscept}$$ where $v_{0}=6.4\times 10^{-23}$ cm$^{3}$ is the unit-cell volume and $B_{S}^{\prime }(x)\equiv dB_{S}(x)/dx.$
To integrate the LLB-Langevin equation, the Heun numerical scheme [@Garcia-Palacios] has been used. The physically reasonable interpretation of the stochastical process is that in the sense of Stratonovich, as was first stressed for the LL equation in Ref. [@Garcia-Palacios]. Lately, it has been shown [@Berkov] that even a naive Euler scheme which yields the Ito solution, would converge to the proper averaged physical value, if the magnetization is normalized at each time step, reflecting the conservation of the magnetization length. However, in the case of LLB equation the magnetization length is also a stochastic fluctuating variable, so that the choice of the integration scheme should explicitly include the Stratonovich interpretation.
The spins were prepared in the state $n_{z}=-1,$ and mean first-passage time (MFPT) time was evaluated as the time elapsed until the magnetization reached the boundary value beyond the barrier, $n_{z}=0.5$. The exact position of this boundary only slightly changes the MFPT. Alternatively, one can set the boundary at the top of the barrier, $n_{z}=0$. In this case one has to multiply the time by 2, since in 50% of all realizations the spin crosses the barrier and in 50% of all realizations it falls back [@Haenggi]. In the high-barrier case, $T\ll \Delta U,$ the MFPT should coincide with the relaxation time $\Gamma ^{-1}$ following from the FPE.
For a comparison, we also solved the (naive) LL-Langevin equation with a constant but thermally reduced magnetization length, $$\mathbf{\dot{n}}=\gamma \lbrack \mathbf{n}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }\right) ]-\gamma \alpha _{2}[\mathbf{n}\times \lbrack
\mathbf{n}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }_{2}\right)
]], \label{LL-Langevin}$$ where $\mathbf{H}_{\mathrm{eff}}$ is given by Eq. (\[HeffDef\]) without the last term. The temperature dependence enters this equation, as in the LLB case, via the scaling of the anisotropy energy with $M_{e}^{2}(T)$ \[Eq. (\[FADef\]) and Eq. (\[Ftilde\]) without the last term\]. The non-rigorous derivation of Eq. (\[LL-Langevin\]) starts with the equation $\mathbf{\dot{s}}=\gamma \lbrack \mathbf{s}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }\right) ]-\gamma \lambda \lbrack \mathbf{s}\times \lbrack \mathbf{s}\times \left( \mathbf{H}_{\mathrm{eff}}+\mathbf{\zeta }_{2}\right) ]]$ for the spin vector of unit length $\mathbf{s}$. \[The same starting equation is used for the derivation of the LLB equation, Eq. (\[LLB\]), in the classical case\]. Replacing $\mathbf{s}$ in this equation by its thermal average, $\mathbf{s\Rightarrow m\equiv }\left\langle \mathbf{s}\right\rangle
,$ and rescaling $\mathbf{m}$ as $\mathbf{m\equiv n}m_{e}$ yields Eq. (\[LL-Langevin\]) with $\alpha _{2}=\lambda m_{e}.$ The latter is in a striking contradiction with the rigorous LLB expressions for the damping parameters, Eq. (\[alphas\]). This difference becomes very important at elevated temperatures, and there is no easy way to improve the naive LL results. In our simulations of Eq. (\[LL-Langevin\]) we just use the constant $\lambda
=0.1$ instead of $\alpha _{2},$ to conform with existing publications, e.g., with Ref. [@Lyberatos2]. Using $\alpha _{2}=\lambda m_{e}$ leads to even more pronounced difference between the LL and LLB results.
\[Fig-Gamma\]
Fig. \[Fig-Gamma\] shows the MFPT of a 8 nm one-domain FePt particle as a function of temperature, calculated numerically from the LL-Langevin and LLB-Langevin equations. These numerical results are compared in Fig. \[Fig-Gamma\] with Brown’s analytical formula for the relaxation time $\Gamma
^{-1}$ [@Brown] and the result of Eq. (\[Gamma0General\]). The switching time calculated within the LLB approach is always lower than that of the LL approach due to the additional lowering of the energy barrier (\[nsRes\]) and the critical growth of the damping at $T_{c}$. We have also shown the temperature $T^{\ast }\simeq 738$ K at which $a=1$ and the geometry of the barrier changes. For a given particle’s volume, our high-barrier approximation leading to Eq. (\[GammaGeneral\]) becomes invalid for $T\gtrsim T^{\ast }.$ We cannot increase the volume, however, without violating the single-domain requirement.
Both Brown’s formula for the LL model and Eq. (\[GammaGeneral\]) for the LLB model describing the Arrhenius dynamics are only valid for $T\ll \Delta
U $. Switching times showing an unphysical divergence near $T_{c}$ is the signature of their breakdown. For the LL model, there is a better analytical approach describing the thermally activated escape in terms of the integral relaxation time (IRT) [@Panina; @LLB; @IRT] that is valid in the whole temperature range. The IRT for the LL model is also plotted in Fig. \[Fig-Gamma\] showing a good agreement with the numerical data at all temperatures. The possibility to find the IRT analytically results from the spatial one-dimentionality of the FPE in the axially-symmetric case: $f=f(\theta ,t).$ For the LLB model there are two spatial coordinates, $\left\{ \theta ,n\right\} $ or $\left\{ n_{z},n_{\bot }=\sqrt{n_{x}^{2}+n_{y}^{2}}\right\} ,$ and a rigorous analytical solution for the IRT seems to be impossible.
In conclusion we have introduced the formalism of the temperature fluctuations within the mean field approach based on the Landau-Lifshitz-Bloch equation. The new Langevin equation could serve as a basis for the temperature-dependent micromagnetic approach for small particles (or discretization elements) at high temperature, similar to standard temperature-dependent micromagnetics but valid in the whole temperature range. This new micromagnetics will have practical importance for the heat-assisted magnetic recording applications.
[99]{} A. Moser and D. Weller, in *The Physics of Ultra-High-Density Magnetic Recording*, Springer, New York, 2001, p. 145
W. F. Brown, Phys. Rev. **130**, 1677 (1963)
Amikam Aharoni, Phys. Rev. **177,** 793 (1965); W. T. Coffey, D. S. F. Crothers, J. L. Dormann, L. J. Geoghan and E. C. Kennedy, Phys. Rev. B **58,** 3249 (1998).
D. A. Garanin, E. C. Kennedy, D. S. F. Crothers and W. T. Coffey, Phys. Rev. B **60,** 6499 (1999).
A. Lyberatos and R. W. Chantrell, J. Appl. Phys. **73,** 6501 (1993).
D. A. Garanin, V. V. Ishchenko, and L. V. Panina, Theor. Math. Phys. **82**, 169 (1990).
D. A. Garanin, Physica A, **172**, 470 (1991).
D. A. Garanin, Phys. Rev. B **55,** 3050 (1997).
J. Kötzler, D. A. Garanin, M. Hartl, and L. Jahn, Phys. Rev. Lett. **71**, 177 (1993).
S. Cumpson, P. Hidding, and R. Coehoorn, IEEE Trans. Magn. **36**, 2271 (2000); J. J. M. Ruigrok, R. Coehoorn, S. R. Cumpson, and H. W. Kesteren, J. Appl. Phys. **87**, 5298 (2000).
S. Sun, C. B. Murray, D. Weller, L. Folks, and A. Moser, Science **287**, 1989 (2000).
A. Lyberatos and K. Yu. Guslienko, J. Appl. Phys. **94**, 1119 (2003).
I. Puzei and V. S. Pokatilov, Sov. Phys. Solid State **16**, 671 (1974).
J. L. Garcia-Palacios and F. J. Lazaro, Phys. Rev. B **58,** 14937 (1998).
H. Kachkachi and D. A. Garanin, Physica A, **291**, 485 (2001).
W. F. Brown, Physica B+C, **86-88**, 1423 (1977).
D. V. Berkov and N. L. Gorn, J. Phys.: Condens. Matter. **14,** 281** ** (2002); E. Martinez, L. Lopez-Diaz, L. Torres, and O. Alejos, Physica B 334, 252 (2004).
P. Hänggi, P. Talkner, and M. Borkovec, Rev. Mod. Phys. **62**, 251 (1990).
D. A. Garanin, Phys. Rev. E **54,** 3250 (1996); Europhys. Lett. **48**, 486 (1999).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have observed a new, complete, cooling-core sample with the VLA, in order to understand how the massive black hole in the central galaxy interacts with the local cluster plasma. We find that every cooling core is currently being energized by an active radio jet, which has probably been destabilized by its interaction with the cooling core. We argue that current models of cooling-core radio galaxies need to be improved before they can be used to determine the rate at which the jet is heating the cooling core. We also argue that the extended radio haloes we see in many cooling-core clusters need extended, [*in situ*]{} re-energization, which cannot be supplied solely by the central galaxy.'
author:
- 'Jean A. Eilek'
- 'Frazer N. Owen'
title: 'Radio Galaxies in Cooling Cores: Insights from a Complete Sample'
---
What heats cooling cores? The spotlight has turned on active galactic nuclei (AGN), driven by massive black holes in the heart of the galaxy at the center of the cooling core (CC). The jets in a few bright, well-studied radio galaxies (RGs) ([*e.g.*]{}, M87 [@OEK]; A2052 [@Blanton]; Perseus A [@Fabian]) seem to be pouring out more than enough energy to offset radiative cooling in the CC in these clusters. But is this not the full answer; questions remain. Does every CC have a central RG? Is a typical cooling-core radio galaxy (CCRG) strong enough to offset local cooling? Does the energy carried by the jet couple effectively to the intracluster medium (ICM)? How can we use radio and X-ray data to estimate the jet power and energy input to the CC? To answer these questions, we must study more than the brightest few CCRGs, and must also look critically at dynamical models of the RG. We therefore carried out deep radio observations of a complete sample of CCRGs. In this paper we summarize our results and speculate on how to extend current models; more details will be given in [@EOM].
The Data: What We Did
=====================
We formed a complete, X-ray selected sample of CCs in nearby Abell clusters. We started with ROSAT All-Sky Survey images of nearby ($z < 0.09$) Abell clusters [@Ledlow]. We identified clusters which are X-ray bright ($L_x > 3 \times 10^{43}$erg/s within a 500 kpc aperture), centered on a massive galaxy, and with a centrally concentrated X-ray atmosphere (ratio of flux within 500 and 62.5 kpc apertures no larger than $\sim 14$). These criteria correlate well with strong CCs found in other, more detailed deprojection analyses ([ *e.g.*]{}, [@Peres].) From these we selected clusters favorably placed in the sky for nightime VLA observations in 2002. This procedure gave us 22 clusters: A85, A133, A193, A426, A496, A780, A1644, A1650, A1651, A1668, A1795, A1927, A2029, A2052, A2063, A2142, A2199, A2428, A2495, A2597, A2626 and A2670. Good, deep radio data already exist for 3 of these (A85; A426, the Perseus A cluster; A780, the Hydra A cluster). We observed the rest with the VLA. Because high-resolution radio data exist for many of these objects, we designed our observations to detect faint, extended radio emission [@Tomislav]. We note that M87 is not in our formal sample, because it is not in an Abell cluster, and its CC is on the weak side. We include it in much of our analysis, however, because it is so well studied [@OEK; @Forman], and it is an important example of the interaction between CCs and their embedded RGs. In addition, nine of our clusters were also included in our VLA search [@Tomislav] for cluster-scale radio haloes, giving us additional information on extended emission from these objects.
The Data: what we have learned
==============================
Our data show that the story is more complicated than has been thought. We find evidence every that cooling core is being disrupted, and probably energized, by an AGN. Our data also suggest that the radio-loud plasma does mix with the ICM, at least on large scales, and that the AGN may well [*not*]{} be the only driver for the ICM in a cooling-core cluster.
Every Cooling Core Contains a Radio Source
------------------------------------------
Every cooling core in our sample contains a currently-active radio core (some too faint to have been detected in previous work). This means that the central AGN are active 100% of the time; they do not have any “off” periods. However, if currrent dynamical estimates of source ages ($\sim 100
$ Myr) are close to correct, the central AGN is probably variable, cycling through high-power and low-power states. The jet and inner halo of M87 [@Dean] appear to be an example of a recently “reinvigorated” AGN.
It follows that every RG in a cooling-core is currently being driven by an active jet. In particular, our deep radio images sometimes reveal faint jets connecting the central AGN to what were previously thought to be offset “relics” (e.g., A133, A2199). We therefore argue that very few CCRGs are simply passive, buoyant bubbles. The situation is more complex; we hope the data can guide us toward improved models.
The Radio Galaxies are Unusually Disturbed
------------------------------------------
Cooling-core RGs are characterized by unusual morphologies. Most of them are neither Fanaroff-Riley Type I (tailed), nor Type II (classical doubles). Sixteen CCRGs in our sample are well enough imaged to reveal their structure (the remaining 7 are too faint and too small). Three of these 16 are standard tailed sources (including Hydra A). The rest are diffuse and amorphous. Such a large fraction (80% of the set) is far too many to be “normal” RGs seen end-on. Furthermore, amorphous sources such as these are rare in the general radio-galaxy population; only 5 of the $\sim 200$ well-imaged cluster radio galaxies in the Owen-Ledlow set [@OL] are amorphous, and [*all*]{} of those sit in the centers of strong cooling cores.
These data show that an unusually strong interaction occurs between the radio jet and the dense cooling-core into which it tries to propagate. The interaction seems to destabilize the jet, on a scale of only a few kpc (the short jets in M87, Perseus A and A2052 are good examples here). It follows that the evolution of a cluster-core RG is not governed by directed momentum flux, in a large-scale jet, as is the case with most RGs. Instead, isotropized energy flow from a disrupted jet creates the amorphous haloes that we see. The strong RG-CC interaction is also suggestive of an effective energy transfer between the jet and the local cooling core; however the details of the process remain unclear.
The Radio Haloes Extend to Large Scales
---------------------------------------
Many of our CCRG are much larger than was previously known. Most of the amorphous sources have two scales of radio emission: a smaller, brighter source (often previously studied in higher resolution observations) is embedded within a larger, faint, extended mini-halo. Typical sizes of these mini-haloes range from $\sim 70$ to $ 200$ kpc. For instance, Per A [@Burns] and A2029 [@Tomislav] can be traced to $\sim 200$ kpc, nearly as large as the long radio tails of Hyd A [@Lane]. If the AGN is always “on”, but cycles between strong and weak states, the mini-haloes may be relics of previous activity cycles. In addition, we have detected Mpc-scale radio haloes in two clusters, A2328 and A2495, which refutes the current idea that Mpc-scale haloes avoid cooling cores.
To put these sizes in context, recall that the size of the cooling core is typically $\sim50\!-\!100$ kpc. The size of the cluster’s potential well, as measured by the Navarro-Frenk-White scale radius, is only a few hundred kpc in CC clusters. Furthermore, the haloes do not obviously have clear edges. The sizes we measure are limited by the sensitivity of the observations, and may not be the true extent of the radio emission. With such large scales, the radio haloes may better be regarded as part of the entire ICM, not just a byproduct of the central AGN.
The Radio and X-ray Plasmas Must Mix
------------------------------------
Our sample contains a variety of mixing states. Some of our clusters have small, clear X-ray cavities in the inner CC, approximately coincident with the RG. These cavities are very likely filled by radio-loud plasma, with little or no X-ray plasma. In other clusters ([*e.g.*]{}, M87 [@Forman], A1795 [@Ettori]), the interaction between the two plasmas is more complex. The X-ray plasma is clearly interacting with, but not evacuated by, the radio plasma. In still others, the X-ray plasma appears smooth and undisturbed on the scale of the RG, at the best current X-ray resolution and sensitivity. We suspect the two plasmas have at least partially mixed in these CCs. In addition, because the ICM is not dramatically disturbed on hundred-kpc scales, the larger radio haloes must be effectively mixed with the ICM.
It seems, therefore, that the radio and X-ray plasmas mix effectively during the lifetime of the CCRG. Just how this occurs is not clear, given the stabilizing effects of even a small magnetic field in the ICM [@JdeY]. Deep, high resolution radio images ([*e.g.*]{}, M87, [@OEK]; A2199, in preparation) may provide hints. We see radio-loud filaments in these sources that appear to be escaping from the RG and penetrating the CC plasma. These filaments may be similar to the magnetic flux ropes which are known to penetrate the terrestrial magnetopause. Such flux ropes, once formed, may decouple from the main body of the radio source and rise bouyantly through the ICM, giving rise to an extended radio halo coincident with a relatively smooth X-ray atmosphere.
The Models: Where to Go Next
============================
The important question in the context of this meeting is, how much energy does a “typical” central AGN deposit in the cooling core plasma? To answer this, we must determine the mean jet power, $P_j$, averaged over the lifetime of a typical CCRG, and how effectively that power is deposited in the local plasma. We emphasize that we have [*no*]{} direct measure of $P_j$. The radio power of the CCRG is a poor tracer of the jet power [@JAE]. The best we can do directly from observations is to use minimum-pressure arguments, which are possible if the jet is resolved ([*e.g.*]{} M87 [@OEK]). This gives us a lower bound on $P_j$. To go further, we must choose a dynamical model for the RG and its interaction with the local ICM. This sounds simple, but the devil is in the details.
Calorimetry
-----------
The simplest cases are cooling cores which have clear X-ray cavities that coincide with an extended RG. In these the mean jet power can be found from the energy within the cavity and the age of the source. This is a simple, attractive appproach, which has been applied by various authors ([*e.g.*]{}, [@Birzan; @DeY]). However, it has complications. One is that measuring the energy content of the cavity is not straightforward, because it is hard to know the extent to which the radio and X-ray plasmas have mixed in most of these clusters. A second concern is how to estimate the age of the source. Most authors currently assume the radio source is passive, having been previously inflated by an AGN which has since turned off. If this holds, the source age is its size divided by the bouyant speed, $v_b$. But what is $v_b$? The sound speed is no more than an optimistic upper limit, because $v_b$ is quite subsonic for small structures. In addition, magnetic tension from even a very small intracluster field can exceed hydrodynamic drag, and reduce $v_b$ even more [@JdeY].
A more serious concern, however, is the evidence from our work that every CCRG is currently being driven by jets from an active AGN, and that large-scale radio haloes have mixed with the ICM. It follows that very few CCRGs are well described as isolated, passive, buoyant bubbles (although buoyancy surely plays some role in the evolution of the RG). New models are needed.
Possible Dynamical Models
-------------------------
As a first step towards such new models, we suggest that CCRG evolution can be broken into two stages. We envision an early stage in which young, driven sources interact with and expand into the ICM, and a later stage in which the RGs have grown into extended mini-haloes mixed with the ICM. If AGN activity is cyclic, an older mini-halo could coexist with a younger, restarted inner core. We note that our ideas here are no more than toy models; they need to be developed and tested against real, well-observed CCRGs.
Because the observations show that the AGN in every CCRG is “alive”, and because CCRGs are often amorphous, we suggest that small, young sources are being driven by a quasi-isotropic energy flux (as from an unstable jet). Such evolution can be approximated by a self-similar analysis [@Falle]. However, because the edges of the X-ray cavities are not strong shocks ([*e.g.*]{}, [@Fabian]), we know the expansion is slow; this suggests the expansion proceeds at approxiate pressure balance [@OEK]. Such a model predicts the source size $R(t) \propto (P_j t)^x$, where $x$ depends on the ambient pressure gradient. We emphasize that $P_j$ and $t$ cannot be determined separately in this model; the best we can do is the limit $\dot R < c_s$, which gives an upper limit to $P_j$.
Because the data also show the radio and X-ray plasmas are well mixed for larger RGs, we further suggest that CCRGs eventually fragment and mix with the ICM. The fragmentation may occur [*via*]{} MHD surface effects (such as the tearing-mode instability) which create magnetic filaments or flux ropes. Alternatively, the small-scale flux ropes which we know exist in MHD turbulence may retain coherence and diffuse into the extended ICM in late stages of CCRG evolution. (The ubiquity of filaments in well-imaged RGs suggests such structures are common in general; why should CCRGs be different?) We expect the flux ropes to rise slowly under buoyancy, and to retain their identity for awhile, after which they probably dissipate and merge with the local ICM. In principle, $P_j$ could be estimated for such a source from the energy content of the radio plasma and its buoyant rise time, but uncertainties in filling factors and flux rope sizes limit the quantitative usefulness of this approach.
The Large Radio Haloes
----------------------
Some of our radio haloes are large enough to raise the question, where does “CCRG” end and “cluster halo” begin? That is, on what scale is the physics of the full cluster more important than the influence of the AGN? The synchrotron size is one criterion: how large can the radio halo can be without needing [*in situ*]{} energization? We are skeptical of simple synchrotron-aging estimates, because magnetic fields in the radio source and the ICM are almost certainly inhomogeneous. One can, however, derive a useful limit. The lowest loss rate for the radio-loud electrons is that of inverse Compton losses on the cosmic microwave background. If the electrons spend most of their time in sub-$\mu$G magnetic fields, and occasionally migrate into high-field regions (probably a few $\mu$G) where they become radio-loud, we can find an upper limit to their synchrotron life.
This cartoon predicts the radio plasma in a buoyant flux rope can reach $\sim 100$ kpc before it fades away. Radio sources larger than this must be undergoing extended, [*in situ*]{} re-energization. It follows that some driver other than the AGN must exist on large scales. Ongoing minor mergers are thought to support radio haloes in large, non-CC clusters. They may be important in CC clusters as well ([*e.g.*]{}, [@Motl]), and may be driving the larger haloes. But then, if we admit the need for non-AGN heating of CC clusters on large scales, can we be sure that the cooling core itself is heated only by the AGN?
[99.]{} L. Birzan, D. Rafferty, B. McNamara et al: ApJ, **607**, 800 (2004)
E. Blanton, C. Sarazin, B. McNamara etal: ApJ, **558**, 15 (2001)
J. Burns, M. Sulkanen, G. Gisler et al: ApJ, **388**, 49 (1992).
D. De Young: to appear in ApJ (a-ph 0605734) (2006)
J. Eilek: Radio Jets in Cooling Cores. In: *The Riddle of Cooling Flows in Galaxies and Clusters of Galaxies*, ed. by T. Reiprich et al, published electronically at [http://www.astro.virginia.edu/coolflow]{} (2004)
J. Eilek, F. Owen, T. Marković: ApJ, to be submitted (2007)
S. Ettori, A. Fabian, S. Allen et al: MNRAS, **331**, 635 (2002)
S. Falle: MNRAS, **250**, 581 (1991).
A. Fabian, A. Celotti, K. Blundell et al: MNRAS, **331**, 369 (2002)
W. Forman, P. Nulsen, S. Heinz et al: ApJ, **635**, 894 (2005)
D. Hines, F. Owen, J. Eilek: ApJ, **347**, 713 (1989)
T. Jones, D. De Young: ApJ, **624**, 586 (2005)
W. Lane, T. Clarke, G. Taylor et al: AJ, **127**, 48 (2004)
M. Ledlow, W. Voges, F. Owen etal: AJ, **126**, 2740 (2003)
T. Marković, Dynamics of the Thermal and Nonthermal Components of the Intra-Cluster Medium, PhD thesis, New Mexico Tech (2004)
P. Motl, J. Burns, C. Loken et al: ApJ, **606**, 635 (2004)
F. Owen, J. Eilek, N. Kassim: ApJ, **543**, 611 (2000).
F. Owen, M. Ledlow, ApJS, **493**, 73 (1997)
C. Peres, A. Fabian, A. Edge et al: MNRAS, **298**, 416 (1998)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Many machine learning systems today are trained on large amounts of human-annotated data. Data annotation tasks that require a high level of competency make data acquisition expensive, while the resulting labels are often subjective, inconsistent, and may contain a variety of human biases. To improve the data quality, practitioners often need to collect multiple annotations per example and aggregate them before training models. Such a multi-stage approach results in redundant annotations and may often produce imperfect “ground truth” that may limit the potential of training accurate machine learning models. We propose a new end-to-end framework that enables us to: (i) merge the aggregation step with model training, thus allowing deep learning systems to learn to predict ground truth estimates directly from the available data, and (ii) model difficulties of examples and learn representations of the annotators that allow us to estimate and take into account their competencies. Our approach is general and has many applications, including training more accurate models on crowdsourced data, ensemble learning, as well as classifier accuracy estimation from unlabeled data. We conduct an extensive experimental evaluation of our method on 5 crowdsourcing datasets of varied difficulty and show accuracy gains of up to 25% over the current state-of-the-art approaches for aggregating annotations, as well as significant reductions in the required annotation redundancy.'
bibliography:
- 'references.bib'
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the Double Coset Membership problem for permutation groups possesses perfect zero-knowledge proofs.'
author:
- |
Oleg Verbitsky\
[Department of Algebra]{}\
[Faculty of Mechanics & Mathematics]{}\
[Kyiv National University]{}\
[Volodymyrska 60]{}\
[01033 Kyiv, Ukraine]{}
title: |
\
\
On the Double Coset Membership Problem\
for Permutation Groups
---
= 45
\[section\] \[section\] \[section\] \[section\] \[section\]
Introduction
============
Definition of the problem
-------------------------
Let $S_m$ be a symmetric group of order $m$. We suppose that an element of $S_m$, a permutation of an $m$-element set, is encoded by a binary string of length $n=\lceil\log_2m!\rceil$, $m(\log_2m-O(1))\le n\le m\log_2m$. Whenever we refer to a [*permutation group*]{} $G$, we mean that $G$ is a subgroup of $S_m$ for some $m$. Throughout the paper we assume that permutation groups are given by a list of their generators.
In this paper we address the following algorithmic problem considered first by Luks [@Luk].
[[DCM]{}]{} ([Double Coset Membership]{})\
[*Given:*]{} two permutations $\sigma$ and $\tau$ and two permutation groups $G$ and $H$, all of the same order.\
[*Recognize if:*]{} $\sigma\in G\tau H$.
Current complexity status
-------------------------
For the background on computational complexity theory the reader is referred to [@GJo].
[[DCM]{}]{} is in the class NP by the Babai-Szemerédy Reachability Theorem [@BSz]. This theorem says that, given any set $S$ of generators of a finite group $G$ and any $g\in G$, there exists a sequence of elements $u_1,\ldots,u_l$ of $G$ such that the following conditions are met.
1. Each $u_i$ either belongs to $S$ or is obtained by the inversion or the group operation from one or two previous elements of the sequence.
2. $u_l=g$.
3. $l\le (1+\log_2|G|)^2$.
As $\sigma\in G\tau H$ iff $\tau^{-1}\sigma\in (\tau^{-1}G\tau)H$, [[DCM]{}]{} admits the following reformulation.
[[DCM]{}]{} (An equivalent formulation)\
[*Given:*]{} a permutation $s$ and two permutation groups $G$ and $H$, all of the same order.\
[*Recognize if:*]{} $s\in GH$.
Consider two related problems, the first one easier and the second one harder than [[DCM]{}]{}.
[[Membership in a Permutation Group]{}]{}\
[*Given:*]{} a permutation $s$ and a permutation group $G$ of the same order.\
[*Recognize if:*]{} $s\in G$.
[Membership in a 3-fold Group Product]{}\
[*Given:*]{} a permutation $s$ and three permutation groups $G$, $H$, and $K$, all of the same order.\
[*Recognize if:*]{} $s\in GHK$.
It is known that the former problem is solvable in polynomial time [@Sim; @FHL] and that the latter problem is NP-complete [@Luk2]. There are evidences that the complexity of [[DCM]{}]{} is strictly in between. On the one hand, the problem of recognition if two given graphs are isomorphic is polynomial-time reducible to [[DCM]{}]{} [@Luk], see also Proposition \[prop:gi\] below. [[DCM]{}]{} is therefore not expected to be solvable in polynomial time as long as the Graph Isomorphism problem is not solved in polynomial time (the currently best algorithm due to Luks and Zemplyachenko runs in time $\exp(O(\sqrt{n\log n}))$ for graphs on $n$ vertices, see [@BLu]). On the other hand, [[DCM]{}]{} belongs to the complexity class coAM (see Subsection \[ss:zkdefn\] for the definition). By [@BHZ], if NP is a subclass of coAM, then the polynomial-time hierarchy of complexity classes collapses to its second level, i.e., $\Sigma_2^P=\Pi_2^P$ (see [@GJo]). As the latter consequence is widely considered unlikely, it is unlikely that [[DCM]{}]{} is NP-complete.
Like the membership in coAM, some other complexity-theoretic results known for Graph Isomorphism also generalize to [[DCM]{}]{}. Both the problems have [*program checkers*]{} [@BKa], and both are [*low*]{} for the complexity class PP [@KST2].
It is worth noting that several other group-theoretic problems are polynomial-time equivalent with [[DCM]{}]{}. We mention a few examples from the list of such problems compiled in [@Luk; @Hof2]: Given permutation groups $G$, $H$ and permutations $\sigma$, $\tau$, (a) find generators for $G\cap H$; (b) recognize if $G\sigma$ and $H\tau$ intersect; (c) if $\sigma\in G$, find the centralizer of $\sigma$ in $G$; (d) if $\sigma,\tau\in G$, recognize if the centralizer of $\tau$ in $S_m$ intersects $G\sigma$. In [@BKa] it is shown that [[DCM]{}]{} is equivalent with the problem, given $s\in GH$, to find a factorization $s=gh$ with $g\in G$ and $h\in H$.
Our result
----------
A natural question to ask about an NP problem whose polynomial-time solvability and NP-completeness are unknown is if it possesses a perfect or a statistical zero-knowledge interactive proof system. Informally speaking, a zero-knowledge proof system for a recognition problem of a language $L$ is a protocol for two parties, the prover and the verifier, that allows the prover to convince the verifier that a given input belongs to $L$, with high confidence but without communicating the verifier any information (the rigorous definitions are in Subsection \[ss:zkdefn\]).
The concept of a zero-knowledge proof has notable applications in designing cryptographic protocols and in estimating the computational complexity of a language recognition problem. Namely, by [@AHa] the class PZK of languages having perfect zero-knowledge proof systems is a subclass of coAM. Thus, the existence of a perfect zero-knowledge proof of the membership in $L$ not only has a cryptographic meaning but also implies that $L$ is in coAM and hence cannot be NP-complete unless the polynomial-time hierarchy collapses.
For the Graph Isomorphism problem, its membership in coAM was proven directly in [@Scho] and its membership in PZK was proven in [@GMW]. For [[DCM]{}]{}, the proof of its membership in coAM given in [@BMo] is direct. In the present paper we prove that [[DCM]{}]{} is also in PZK. We therefore extend the list of problems in PZK that currently includes Graph Isomorphism [@GMW], Quadratic Residuosity [@GMR], a problem equivalent to Discrete Logarithm [@GKu], and approximate versions of the Shortest Vector and Closest Vector problems for integer lattices [@GGo].
Background on zero-knowledge proofs
===================================
Definitions {#ss:zkdefn}
-----------
We denote the [*length*]{} of a binary word $w$ by $|w|$. We consider [*languages*]{} over the binary alphabet which are subsets of $\{0,1\}^*$. The [*complement*]{} of $L$ is the language $\bar L=\{0,1\}^*\setminus L$. Note that the [[DCM]{}]{} problem can be represented as a recognition problem for the language $L={\left\{ \hspace{0.5mm} (s,G,H) :
\hspace{0.5mm} s\in GH \right\}}$, where $(s,G,H)$ is a suitable binary encoding of the triplet consisting of a permutation $s$ and the lists of generators for permutation groups $G$ and $H$.
We use the standard computational model of a deterministic [*Turing machine*]{}, abbreviated further on as TM. We assume that a TM has three tapes, namely, the input tape, the output tape, and the work tape where all computations are performed.
A [*probabilistic*]{} TM, abbreviated further on as PTM, in addition has the fourth tape containing a potentially infinite random binary string. Assuming that a PTM halts on input $w$ and random string $r$, we denote its running time by $t(w,r)$. A PTM is [*polynomial-time*]{} if $t(w,r)$ is bounded by a polynomial in $|w|$ for all $w$ and $r$. Assuming that a PTM halts on $w$ for almost all $r$, the function $t(w,r)$ for a fixed $w$ can be considered as a random variable on the probability space $\{0,1\}^{ {\bf N} }$ of all random strings. A PTM is [*expected polynomial-time*]{} on $L\subseteq\{0,1\}^*$ if for all $w\in L$ the expectation of $t(w,r)$ is bounded by a polynomial in $|w|$.
An [*interactive proof system*]{} $\langle V,P\rangle$, further on abbreviated as IPS, consists of two PTMs, a polynomial-time $V$ called [*the verifier*]{} and a computationally unlimited $P$ called [*the prover*]{}. The input tape is common for the verifier and the prover. The verifier and the prover also share a communication tape which allows message exchange between them. The system works as follows. First both the machines $V$ and $P$ are given an input $w$ and each of them is given an individual random string, $r_V$ for $V$ and $r_P$ for $P$. Then $P$ and $V$ alternatingly write messages to one another in the communication tape. $V$ computes its $i$-th message $a_i$ to $P$ based on the input $w$, the random string $r_V$, and all previous messages from $P$ to $V$. $P$ computes its $i$-th message $b_i$ to $V$ based on the input $w$, the random string $r_P$, and all previous messages from $V$ to $P$. After a number of message exchanges $V$ terminates interaction and computes an output based on $w$, $r_V$, and all $b_i$. The output is denoted by $\langle V,P\rangle(w)$. Note that, for a fixed $w$, $\langle V,P\rangle(w)$ is a random variable depending on both random strings $r_V$ and $r_P$.
Let $\epsilon(n)$ be a function of a natural argument taking on positive real values. We call $\epsilon(n)$ [*negligible*]{} if $\epsilon(n)<n^{-c}$ for every $c$ and all $n$ starting from some $n_0(c)$. For example, an [*exponentially small*]{} function $\epsilon(n)=d^{-n}$, where $d>1$, is negligible.
We say that $\langle V,P\rangle$ is an [*IPS for a language $L$ with error $\epsilon(n)$*]{} if the following two conditions are fulfilled.
[*Completeness.*]{} If $w\in L$, then $\langle V,P\rangle(w)=1$ with probability at least $1-\epsilon(|w|)$.\
[*Soundness.*]{} If $w\notin L$, then, for an arbitrary interacting PTM $P^*$, $\langle V,P^*\rangle(w)=1$ with probability at most $\epsilon(|w|)$.
We will call any prover $P^*$ interacting with $P$ on input $w\notin L$ [*cheating*]{}. If in the completeness condition we have $\langle V,P\rangle(w)=1$ with probability 1, we say that $\langle V,P\rangle$ has [*one-sided error*]{} $\epsilon(n)$.
We say that $\langle V,P\rangle$ is an [*IPS for a language $L$*]{} if $\langle V,P\rangle$ is an IPS for $L$ with negligible error.
An IPS is [*public-coin*]{} if the concatenation $a_1\ldots a_k$ of the verifier’s messages is a prefix of his random string $r_V$. A [*round*]{} is sending one message from the verifier to the prover or from the prover to the verifier. The class AM consists of those languages having IPSs with error $1/3$ and with number of rounds bounded by a constant for all inputs. A language $L$ belongs to the class coAM iff its complement $\bar L$ belongs to AM.
Given an IPS $\langle V,P\rangle$ and an input $w$, let ${\mbox{view}}_{V,P}(w)=(r'_V,a_1,b_1,\ldots,a_k,b_k)$ where $r'_V$ is a part of $r_V$ scanned by $V$ during work on $w$ and $a_1,b_1,\ldots,a_k,b_k$ are all messages from $P$ to $V$ and from $V$ to $P$ ($a_1$ may be empty if the first message is sent by $P$). Note that the verifier’s messages $a_1,\ldots,a_k$ could be excluded because they are efficiently computable from the other components. For a fixed $w$, ${\mbox{view}}_{V,P}(w)$ is a random variable depending on $r_V$ and $r_P$.
An IPS $\langle V,P\rangle$ is [*perfect zero-knowledge on $L$*]{} if for every interacting polynomial-time PTM $V^*$ there is a PTM $M_{V^*}$, called a [*simulator*]{}, that on every input $w\in L$ runs in expected polynomial time and produces output $M_{V^*}(w)$ which, if considered as a random variable depending on a random string of $M_{V^*}$, is distributed identically with ${\mbox{view}}_{V^*,P}(w)$. The latter condition means that $${ {\bf P} \left[ M_{V^*}(w)=z \right] }={ {\bf P} \left[ {\mbox{view}}_{V^*,P}(w)=z \right] }\mbox{ for all }z.$$ If only a weaker condition that $$\sum_z{\left| { {\bf P} \left[ M_{V^*}(w)=z \right] }-{ {\bf P} \left[ {\mbox{view}}_{V^*,P}(w)=z \right] } \right|}
\mbox{ is negligible}$$ is true, we call $\langle V,P\rangle$ [*statistical zero-knowledge*]{}. These notions formalize the claim that the verifier gets no information during interaction with the prover: Everything that the verifier gets he can get without the prover by running the simulator.
According to the definition the verifier learns nothing even if he deviates from the original program and follows an arbitrary probabilistic polynomial-time program $V^*$. We will call the verifier $V$ [*honest*]{} and all other verifiers $V^*$ [*cheating*]{}. If the existence of a simulator is claimed only for the honest verifier, we call such a proof system [*honest-verifier perfect ([*or*]{} statistical) zero-knowledge*]{}.
The class of languages $L$ having IPSs that are perfect (resp. statistical) zero-knowledge on $L$ is denoted by PZK (resp. SZK). Recall that the error here is supposed negligible.
The [*$k(n)$-fold sequential composition*]{} of an IPS $\langle V,P\rangle$ is the IPS $\langle V',P'\rangle$ in which $V'$ and $P'$ on input $w$ execute the programs of $V$ and $P$ sequentially $k(|w|)$ times, each time with independent choice of random strings $r_V$ and $r_P$. At the end of interaction $V'$ outputs 1 iff $\langle V,P\rangle(w)=1$ in all $k(|w|)$ executions. The initial system $\langle V,P\rangle$ is called [*atomic*]{}.
In the [*$k(n)$-fold parallel composition*]{} $\langle V'',P''\rangle$ of $\langle V,P\rangle$, the program of $\langle V,P\rangle$ is executed $k(|w|)$ times in parallel, that is, in each round all $k(|w|)$ versions of a message are sent from one machine to another at once as a long single message. In every parallel execution $V''$ and $P''$ use independent copies of $r_V$ and $r_P$. At the end of interaction $V'$ outputs 1 iff $\langle V,P\rangle(w)=1$ in all $k(|w|)$ executions.
Known results on zero-knowledge proofs
--------------------------------------
We first notice a simple property of sequential composition of IPSs.
\[prop:seqrep\] If $\langle V,P\rangle$ is an IPS for a language $L$ with one-sided constant error $\epsilon$, then the $k(n)$-fold sequential composition of $\langle V,P\rangle$ is an IPS for $L$ with one-sided error $\epsilon^{k(n)}$.
Parallel composition obviously preserves the number of rounds, the public-coin property, and the property of error to be one-sided. It is not hard to prove that $k$-fold parallel composition reduces the one-sided error $\epsilon$ to $\epsilon^k$. It is also not hard to prove that parallel composition preserves perfect and statistical zero-knowledge for the honest verifier. These observations are summarized in the next proposition.
\[prop:parrep\] Assume that $\langle V,P\rangle$ is a honest-verifier perfect zero-knowledge public-coin IPS for a language $L$ that on all inputs works in a constant $c$ rounds with one-sided constant error $\epsilon$. Then $k(n)$-fold parallel composition of $\langle V,P\rangle$ is a honest-verifier perfect zero-knowledge IPS for $L$ that works in $c$ rounds with error $\epsilon^{k(n)}$.
We also refer to the following deep results in the theory of zero-knowledge proofs.
\[prop:aha\] $\mbox{SZK}\subseteq\mbox{coAM}$.
\[prop:oka\]
1. Every honest-verifier statistical zero-knowledge IPS for a language $L$ can be transformed in an honest-verifier statistical zero-knowledge public-coin IPS for $L$.
2. If $L$ has an honest-verifier statistical zero-knowledge public-coin IPS, then $\bar L$ has a honest-verifier statistical zero-knowledge constant-round IPS.
Note that the item 2 of Proposition \[prop:oka\] strengthens Proposition \[prop:aha\] because by [@GSi] every IPS can be made public-coin at cost of decreasing the number of rounds in 2.
\[prop:gsv\] Every honest-verifier statistical zero-knowledge public-coin IPS for a language $L$ can be transformed in a general statistical zero-knowledge public-coin IPS for $L$. If the error of the initial IPS is one-sided, so is the error of the resulting IPS.
Note that, to achieve the negligible error, the transformation of Proposition \[prop:gsv\] makes the number of rounds increasing with the input size increasing, even if the initial IPS is constant-round. A transformation preserving the constant number of rounds is known only under an unproven assumption about the hardness of the Discrete Logarithm problem (the formal statement of the assumption can be found in [@BMOs]).
\[prop:bmo\] Suppose that a language $L$ has an honest-verifier statistical zero-knowledge IPS that on every input $w$ works in $c(|w|)$ rounds with error at most $1/3$. Then, under the assumption on the hardness of Discrete Logarithm, $L$ has a general statistical zero-knowledge IPS that on input $w$ works in $O(c(|w|))$ rounds with exponentially small error.
Background on permutation groups
================================
Given a finite set $X$, by a [*random element*]{} of $X$ we mean a random variable uniformly distributed over $X$.
\[prop:sim\]
1. There is a polynomial-time algorithm for recognizing the [[Membership in a Permutation Group]{}]{}.
2. There is a probabilistic polynomial-time algorithm that, given a list of generators for a permutation group $G$, outputs a random element of $G$.
The [[DCM]{}]{} problem is at least as hard as testing isomorphism of two given graphs.
\[prop:gi\] The Graph Isomorphism problem is polynomial-time reducible to [[DCM]{}]{}.
We include a proof for the sake of completeness.
Consider two graphs of order $n$ with adjacency matrices $A=(a_{ij})$ and $B=(b_{ij})$. Let $S_1={\left\{ \hspace{0.5mm} (i,j) :
\hspace{0.5mm} a_{ij}=1 \right\}}$ and $S_2={\left\{ \hspace{0.5mm} (i,j) :
\hspace{0.5mm} b_{ij}=1 \right\}}$.
Let $G$ be the group of permutations of the square $\{1,\ldots,n\}^2$ generated by simultaneous transpositions of $i$-th and $j$-th rows and $i$-th and $j$-th columns for all $1\le i<j\le n$. The graphs are isomorphic iff $G$ contains a permutation $\sigma$ such that $\sigma(S_1)=S_2$.
Let $H$ be the group of permutations $\tau$ such that $\tau(S_1)=S_1$ and $s$ be an arbitrary permutation such that $s(S_1)=S_2$. As easily seen, a permutation $\sigma$ as above exists iff $s\in GH$.
Note that the reduction described allows one to transform any zero-knowledge proof system for [[DCM]{}]{} in a zero-knowledge proof system for Graph Isomorphism.
Zero-knowledge proofs for [[DCM]{}]{}
=====================================
\[thm:1\] The [[DCM]{}]{} problem has an honest-verifier perfect zero-knowledge three-round public-coin IPS with one-sided error $1/2$.
On input $(s,G,H)$ such that $s\in GH$ the IPS $\langle V,P\rangle$ proceeds as follows.
[*1st round.*]{}
$P$ generates random elements $g\in G$ and $h\in H$, computes $t=gsh$, and sends $t$ to $V$. $V$ checks if $t$ is a permutation of the given order and if not (this is possible in the case of a cheating prover) halts and outputs 1.
[*2nd round.*]{}
$V$ chooses a random bit $b\in\{0,1\}$ and sends it to $P$.
[*3rd round.*]{}
[*Case $b=0$.*]{} $P$ sends $V$ permutations $g$ and $h$. $V$ checks if $g\in G$, $h\in H$, and $t=gsh$.
[*Case $b\ne 0$*]{} (this includes the possibility of a message $b\notin\{0,1\}$ produced by a cheating verifier). $P$ decomposes $s$ into the product $s=g_0h_0$ with $g_0\in G$ and $h_0\in H$, computes $g_1=gg_0$ and $h_1=h_0h$, and sends $g_1$ and $h_1$ to $V$. $V$ checks if $g_1\in G$, $h_1\in H$, and $t=g_1h_1$.
$V$ halts and outputs 1 if the conditions are checked successfully and 0 otherwise.
This IPS is obviously public-coin. We need to check that this is indeed an IPS for [[DCM]{}]{} with one-sided error 1/2 and, moreover, that this is a honest-verifier perfect zero-knowledge IPS.
[*Completeness. *]{}
If $s\in GH$, then it is clear that $V$ outputs 1 with probability 1.
[*Soundness. *]{}
Assume that $s\notin GH$ and consider an arbitrary cheating prover $P^*$. Observe that if both $t=gsh$, $g\in G$, $h\in H$ and $t=g_1h_1$, $g_1\in G$, $h_1\in H$, then $s\in GH$. It follows that, for at least one value of $b$, $V$ outputs 0 and therefore $V$ outputs 1 with probability at most 1/2.
[*Zero-knowledge. *]{}
Assume that $s\in GH$. During interaction with $P$, $V$ sees ${\mbox{view}}_{V,P}(s,G,H)=(b,t,b,g',h')$ where $g'$ and $h'$ are received by $V$ in the 3rd round. If $b=0$, then $t=gsh$, $g'=g$, and $h'=h$. If $b=1$, then $t=g'h'$, $g'=gg_0$, and $h'=h_0h$. In both the cases $g'$ and $h'$ are random elements of $G$ and $H$ respectively. The random variable ${\mbox{view}}_{V,P}(s,G,H)$ can be therefore generated by the following simulator: Generate a random bit $b$ and random elements $g'\in G$ and $h'\in H$; If $b=0$, set $t=g'sh'$; If $b=1$, set $t=g'h'$.
\[cor:1\] The [[DCM]{}]{} problem has an honest-verifier perfect zero-knowledge three-round public-coin IPS with one-sided error $2^{-n}$.
By Proposition \[prop:parrep\] the $n$-fold parallel composition of the IPS from Theorem \[thm:1\] reduces the error to $2^{-n}$ and preserves the properties of the atomic system.
Let [Double Coset Non-Membership]{}, abbreviated as [[DCNM]{}]{}, be the problem opposite to [[DCM]{}]{}, that is, given a permutation $s$ and two permutation groups $G$ and $H$, to recognize if $s\notin GH$. The [[DCNM]{}]{} problem is clearly polynomial-time equivalent with recognition of the set-theoretic complement of [[DCM]{}]{}, where the latter is encoded as a language in the binary alphabet.
\[cor:2\] [[DCNM]{}]{} has an honest-verifier statistical zero-knowledge constant-round IPS.
The corollary follows from Corollary \[cor:1\] by Proposition \[prop:oka\].
We also give an alternative direct proof of this claim describing an honest-verifier perfect zero-knowledge two-round IPS $\langle V,P\rangle$ for [[DCNM]{}]{} with one-sided error 1/2. This system, for the case of permutation groups, generalizes the IPS suggested in [@Bab3] for the problem of testing the membership in a finite group given by a list of generators and an oracle access to the group operation.
On input $(s,G,H)$ such that $s\notin GH$ the system works as follows.
[*1st round.*]{}
$V$ chooses a random bit $b$ to be the first bit of a random string $r_V$ and, based on the subsequent bits of $r_V$, generates random elements $g\in G$ and $h\in H$. If $b=0$, $V$ computes $t=gh$; If $b=1$, $V$ computes $t=gsh$. Then $V$ sends $t$ to $P$.
[*2nd round.*]{}
$P$ recognizes if $t\in GH$. If so, $P$ sets $a=0$; If not, $P$ sets $a=1$. Then $P$ sends $a$ to $V$.
$V$ checks if $a=b$ and halts. If the equality is true, $V$ outputs 1; Otherwise $V$ outputs 0.
[*Completeness. *]{}
Assume that $s\notin GH$. In the first round, $t\in GH$ if $b=0$ and $t\notin GH$ if $b=1$. Therefore $V$ outputs 1 with probability 1.
[*Soundness. *]{}
Assume that $s\in GH$. Then $t\in GH$ regardless of the value of $b$. Moreover, $t$ is the product of random elements of $G$ and $H$ and, as a random variable, is independent of the random variable $b$. It follows that in the second round a message from the cheating prover $P^*$ to $V$, which is a function of $s$, $G$, $H$, $r_{P^*}$, and $t$, is equal to $b$ with probability at most 1/2. Hence $\langle V,P^*\rangle(s,G,H)=1$ with probability at most 1/2.
[*Zero-knowledge. *]{}
Assume that $s\notin GH$. During interaction with $P$, $V$ sees ${\mbox{view}}_{V,P}(s,G,H)=(r'_V,t,a)$, where $a$ equals the first bit $b$ of $r'_V$. The simulator therefore just generates a random string $r_V$, extracts the first bit $b$ from it, sets $a=b$, based on the remaining bits of $r_V$ computes $g$ and $h$, based on $b$, $g$, and $h$ computes $t$, and sets $r'_V$ to be the prefix of $r_V$ that was actually used for these purposes.
\[cor:3\] [[DCM]{}]{} is in SZK. Moreover, [[DCM]{}]{} has a statistical zero-knowledge public-coin IPS with one-sided error.
Apply the transformation from Proposition \[prop:gsv\] to the IPS from Corollary \[cor:1\].
Note that another proof of the membership of [[DCM]{}]{} in SZK can be given by applying Propositions \[prop:oka\] and \[prop:gsv\] to the IPS in the alternative proof of Corollary \[cor:2\].
\[cor:4\] [[DCM]{}]{} is in coAM. Therefore [[DCM]{}]{} is not NP-complete unless the polynomial-time hierarchy collapses at the second level.
This is an immediate consequence of Corollary \[cor:2\] or a consequence of Corollary \[cor:3\] based on Proposition \[prop:aha\].
\[cor:5\] Under the assumption on the hardness of Discrete Logarithm, [[DCM]{}]{} has a constant-round statistical zero-knowledge IPS with exponentially small error.
The corollary follows from Theorem \[thm:1\] by Proposition \[prop:bmo\].
\[thm:2\] The $n$-fold sequential composition of the IPS in Theorem \[thm:1\] is a perfect zero-knowledge public-coin IPS for [[DCM]{}]{} with exponentially small error. Hence [[DCM]{}]{} is in PZK.
Denote the composed IPS by $\langle V,P\rangle$. As the atomic system is public-coin, so is $\langle V,P\rangle$. By Proposition \[prop:seqrep\] $\langle V,P\rangle$ is an IPS for [[DCM]{}]{} with one-sided error $2^{-n}$. We have to prove that $\langle V,P\rangle$ is perfect zero-knowledge.
For each verifier $V^*$ interacting with $P$ we describe a probabilistic expected polynomial-time simulator $M_{V^*}$. The $M_{V^*}$ uses the program of $V^*$ as a subroutine. Assume that the running time of $V^*$ is bounded by a polynomial $q(n)$ in the input size. On input $w$, $M_{V^*}$ will run the program of $V^*$ on input $w$ with random string $r$, where $r$ is the prefix of $M_{V^*}$’s random string of length $q(|w|)$. In all other cases $M_{V^*}$ will use the remaining part of its random string.
Work of $M_{V^*}$ on input $w=(s,G,H)$ consists of $|w|$ stages, where a stage corresponds to an iteration of the atomic system.
[*Stage $i$.*]{}
$M_{V^*}$ chooses random elements $g_i\in G$ and $h_i\in H$ and a random bit $a\in\{0,1\}$. If $a=0$, $M_{V^*}$ computes $t_i=g_ish_i$; If $a=1$, it computes $t_i=g_ih_i$. Then $M_{V^*}$ computes $b_i=V^*(w,r,t_1,g_1,h_1,\ldots,t_{i-1},g_{i-1},h_{i-1},t_i)$, the message that $V^*(w,r)$ sends $P$ in the $i$-th sequential iteration of the atomic system after receiving $P$’s message $t_i$ and under the condition that in the preceding iterations $P$’s messages were $t_1,g_1,h_1,\ldots,t_{i-1},g_{i-1},h_{i-1}$. If $b_i$ and $a$ are simultaneously equal to or different from 0, then $M_{V^*}$ puts $v_i=(t_i,b_i,g_i,h_i)$ and proceeds to the $(i+1)$-th stage. If exactly one of $b_i$ and $a$ is equal to 0, then $M_{V^*}$ restarts the same $i$-th stage with new independent choice of $a$, $g_i$, $h_i$.
After all stages are completed, $M_{V^*}$ halts and outputs $(r',v_1,\ldots,v_{|w|})$, where $r'$ is the prefix of $r$ actually used by $V^*$ during interaction on input $w$ with the prover sending the messages $t_1,g_1,h_1,\ldots,t_{|w|},g_{|w|},h_{|w|}$. Notice that it might happen that in unsuccessful attempts to pass some stage $V^*$ used a prefix of $r$ longer than $r'$.
We first check that $M_{V^*}$ terminates in expected polynomial time whenever $s\in GH$. Since $V^*$ is polynomial-time, one attempt to pass Stage $i$, $i\le|w|$, takes time bounded by a polynomial in $|w|$. Recall that $M_{V^*}$ is programmed so that $a$ and $r$ are independent. Furthermore, $a$ and $t_i$ are independent. Indeed, if $a=1$, then $t_i=g_ih_i$ is the product of random elements of $G$ and $H$. If $a=0$, then $t_i=(g_ig_0)(h_0h_i)$ is such a product as well. Here $g_0\in G$ and $h_0\in H$ are elements of an arbitrary decomposition $s=g_0h_0$. It follows that $a$ and $b_i$ are independent and therefore an execution of the stage is successful with probability 1/2. We conclude that on average each stage consists of 2 executions. Thus, on average $M_{V^*}$ makes $2|w|$ polynomial-time executions and this takes expected polynomial time.
We finally need to check that, whenever $s\in GH$, the output $M_{V^*}(w)$ is distributed identically with ${\mbox{view}}_{V^*,P}(w)$. Notice that both the random variables depend on $V^*$’s random string $r$. It therefore suffices to show that the distributions are identical when conditioned on an arbitrary fixed $r$. For $0\le i\le|w|$, let $D^i_M(w,r)$ denote the probability distribution of $(r',v_1,\ldots,v_i)$ conditioned on $r$, and $D^i_{V^*,P}(w,r)$ denote the distribution of the part of ${\mbox{view}}_{V^*,P}(w)$ formed up to the $i$-th sequential iteration. With this notation, we have to prove that $D^{|w|}_M(w,r)=D^{|w|}_{V^*,P}(w,r)$. Using the induction on $i$, we prove that $D^{i}_M(w,r)=D^{i}_{V^*,P}(w,r)$ for every $0\le i\le|w|$.
The base case of $i=0$ is trivial. Let $i\ge 1$ and assume that $$\label{eq:1}
{ {\bf P} \left[ D^{i-1}_M(w,r)=u_{i-1} \right] }={ {\bf P} \left[ D^{i-1}_{V^*,P}(w,r)=u_{i-1} \right] }$$ for every value $u_{i-1}$. Given $u_{i-1}$, assume now that both $D^{i-1}_M(w,r)=u_{i-1}$ and $D^{i-1}_{V^*,P}(w,r)=u_{i-1}$, and under these conditions consider how the $i$-th components $v_i=(t_i,b_i,g_i,h_i)$ are distributed in $u_i=u_{i-1}v_i$ according to $D^{i}_M(w,r)$ and $D^{i}_{V^*,P}(w,r)$. We will show that $$\label{eq:2}
\parbox{.9\textwidth}{
\begin{eqnarray*}
{ {\bf P} \left[ \hspace{0.5mm} D^{i}_M(w,r)=u_{i-1}v_i
\left| \hspace{0.5mm} D^{i-1}_M(w,r)=u_{i-1} \right. \right] }\hspace{53mm}&&\\
={ {\bf P} \left[ \hspace{0.5mm} D^{i}_{V^*,P}(w,r)=u_{i-1}v_i
\left| \hspace{0.5mm} D^{i-1}_{V^*,P}(w,r)=u_{i-1} \right. \right] }&&
\end{eqnarray*}
}$$ for every value $v_i$. Together with [(\[eq:1\])]{} this will imply the identity of $D^{i}_M(w,r)$ and $D^{i}_{V^*,P}(w,r)$.
To prove [(\[eq:2\])]{}, we will show that according to the both conditional distributions $v_i$ is uniformly distributed on the set $$\begin{aligned}
S=\Bigl\{ \hspace{0.5mm} (t,b,g,h) & : &
t\in GH,\ b=V^*(w,r,u_{i-1},t),\ g\in G,\ h\in H,\\
&& t=gsh\mbox{ if }b=0\mbox{ and }t=gh\mbox{ if }b\ne 0\Bigr. \Bigr\}.\end{aligned}$$
Given $t$ and $s$, define sets $R(t)={\left\{ \hspace{0.5mm} (g,h) :
\hspace{0.5mm} g\in G,\, h\in H,\, gh=t \right\}}$ and $R_s(t)={\left\{ \hspace{0.5mm} (g,h) :
\hspace{0.5mm} g\in G,\, h\in H,\, gsh=t \right\}}$. The first claim of the following lemma appeared in [@Hof1].
\[lem:\] Let $k=|G\cap H|$. Assume that $s=g_0h_0$ with $g_0\in G$ and $h_0\in H$. Then the following statements are true.
1. Every $t\in GH$ has $k$ representations $t=gh$ with $g\in G$ and $h\in H$, i.e., $|R(t)|=k$. If $t=g_1h_1$, then all other representations are $$\label{eq:f}
t=(g_1f)(f^{-1}h_1),$$ where $f$ ranges over group $G\cap H$.
2. For every $t$, the mapping $\alpha(g,h)=(gg_0,h_0h)$ is one-to-one from $R_s(t)$ to $R(t)$.
3. Every $t\in GH$ has $k$ representations $t=gsh$ with $g\in G$ and $h\in H$, i.e., $|R_s(t)|=k$.
4. If $\phi{:G\times H \rightarrow GH}$ is defined by $\phi(g,h)=gh$, then $|\phi^{-1}(t)|=k$ for every $t\in GH$.
5. If $\psi{:G\times H \rightarrow GH}$ is defined by $\psi(g,h)=gsh$, then $|\psi^{-1}(t)|=k$ for every $t\in GH$.
6. If $t=gh$ is the product of uniformly distributed random elements $g\in G$ and $h\in H$, then $t$ is uniformly distributed on $GH$.
7. If a uniformly distributed random pair $(g,h)\in G\times H$ is conditioned on $gh=t$ for an arbitrary fixed $t\in GH$, then $(g,h)$ is uniformly distributed on $R(t)$.
8. If $t=gsh$ and $g\in G$ and $h\in H$ are uniformly distributed random elements, then $t$ is uniformly distributed on $GH$.
9. If a uniformly distributed random pair $(g,h)\in G\times H$ is conditioned on $gsh=t$ for an arbitrary fixed $t\in GH$, then $(g,h)$ is uniformly distributed on $R_s(t)$.
We first prove Item 1. Let $e$ denote the identity permutation. Clearly that we have at least $k$ representations of the form [(\[eq:f\])]{}. On the other hand, every representation $t=gh$ is of this form. Indeed, we have $(g^{-1}g_1)(h_1h^{-1})=e$ and hence both $g^{-1}g_1$ and $h_1h^{-1}$ are simultaneously in $G$ and in $H$.
To prove Item 2, observe that $\alpha$ is indeed from $R_s(t)$ to $R(t)$. The map $\alpha'(g,h)=(gg_0^{-1},h_0^{-1}h)$ is easily seen to be from $R(t)$ to $R_s(t)$ and inverse to $\alpha$.
Items 1 and 2 imply Item 3, Item 3 implies Item 5, and Item 5 implies Item 8. Item 1 implies Item 4, and Item 4 implies Item 6. Items 7 and 9 are true by the definition of $R(t)$ and $R_s(t)$.
The distribution $D^i_{V^*,P}(w,r)$ conditioned on $D^{i-1}_{V^*,P}(w,r)=u_{i-1}$ is samplable as follows. Choose random elements $g\in G$ and $h\in H$. Compute $t_i=gsh$ and $b_i=V^*(w,r,u_{i-1},t_i)$. If $b_i=0$, set $g_i=g$ and $h_i=h$, otherwise set $g_i=gg_0$ and $h_i=h_0h$. Clearly, this distribution of $(t_i,b_i,g_i,h_i)$ is over $S$.
By Item 8 of Lemma \[lem:\], $t_i$ is uniformly distributed on $GH$. If $b_i=0$, then by Item 9 of Lemma \[lem:\], for every fixed $t_i$, the pair $(g_i,h_i)$ is uniformly distributed on $R_s(t)$. If $b_i\ne 0$, then by Item 2 of Lemma \[lem:\], for every fixed $t_i$, the pair $(g_i,h_i)$ is uniformly distributed on $R(t)$. It follows that $D^i_{V^*,P}(w,r)$ conditioned on $D^{i-1}_{V^*,P}(w,r)=u_{i-1}$ is uniform on $S$.
Consider now the sampling procedure for the distribution $D^i_{M}(w,r)$ conditioned on $D^{i-1}_{M}(w,r)=u_{i-1}$ as in the description of the simulator $M_{V^*}$. Under the condition that $a=0$, by Items 8 and 9 of Lemma \[lem:\], $t_i$ is distributed uniformly over $GH$ and for every fixed value of $t_i$, the pair $(g_i,h_i)$ is uniformly distributed over $R_s(t)$. Under the condition that $a=1$, by Items 6 and 7 of Lemma \[lem:\], $t_i$ is distributed uniformly over $GH$ and for every fixed value of $t_i$, the pair $(g_i,h_i)$ is uniformly distributed over $R(t)$. This leads to an equivalent sampling procedure: Choose a random $t_i\in GH$, compute $b_i=V^*(w,r,u_{i-1},t_i)$; If $b_i=0$, choose a random pair $(g_i,h_i)$ in $R_s(t_i)$, otherwise in $R(t)$. It follows that $D^i_{M}(w,r)$ conditioned on $D^{i-1}_{M}(w,r)=u_{i-1}$ is uniform on $S$.
The simulator in the proof of Theorem \[thm:2\] is [*black-box*]{}, that is, for each $V^*$ it follows the same program that uses the strategy of $V^*$ as a subroutine. It should be noted that by [@GKr] the parallel composition of the IPS in Theorem \[thm:1\] is [*not*]{} zero-knowledge with black-box simulator unless [[DCM]{}]{} is decidable in probabilistic polynomial time.
Future work
===========
A natural question arises if our results can be extended to [*matrix groups*]{} over finite fields. One of the reasons why this case is more complicated is that, unlike permutation groups, no efficient test of membership for matrix groups is known. We intend to tackle this question in a subsequent paper.
[10]{}
B. Aiello and J. Håstad. Perfect zero-knowledge languages can be recognized in two rounds. In [*Proc. of the [28]{}th IEEE Ann. Symp. on Foundations of Computer Science (FOCS)*]{}, pages 439–448, 1987.
L. Babai. Local expansion of vertex-transitive graphs and random generation in finite groups. In [*Proc. of the [23]{}rd ACM Ann. Symp. on Theory of Computing (STOC)*]{}, pages 164–174, 1991.
L. Babai and E.M.Luks. Canonical labeling of graphs. In [*Proc. of the [15]{}th ACM Ann. Symp. on the Theory of Computing (STOC)*]{}, pages 171–183, 1983.
L. Babai and S. Moran. Arthur-Merlin games: a randomized proof system, and a hierarchy of complexity classes. , 36:254–276, 1988.
L. Babai and E. Szemerédi. On the complexity of matrix group problems. In [*Proc. of the [25]{}th IEEE Ann. Symp. on Foundations of Computer Science (FOCS)*]{}, pages 229–240, 1984.
M. Bellare, S. Micali, and R. Ostrovsky. The (true) complexity of statistical zero knowledge. In [*Proc. of the [22]{}nd ACM Ann. Symp. on Theory of Computing (STOC)*]{}, pages 494–502, 1990.
M. Blum and S. Kannan. Designing programs that check their work. , 42(1):269–291, 1995.
R. B. Boppana, J. Håstad, and S. Zachos. Does co-NP have short interactive proofs? , 25:127–132, 1987.
M. L. Furst, J. Hopcroft, and E. M. Luks. Polynomial-time algorithms for permutation groups. In [*Proc. of the [21]{}st IEEE Ann. Symp. on Foundations of Computer Science (FOCS)*]{}, pages 36–41, 1980.
M. R. Garey and D. S. Johnson. . W. H. Freeman, 1979 (a Russian translation available).
O. Goldreich and S. Goldwasser. On the limits on the non-approximability of lattice problems. In [*Proc. of the [30]{}th ACM Ann. Symp. on Theory of Computing (STOC)*]{}, pages 1–9, 1998.
O. Goldreich and H. Krawczyk. On the composition of zero-knowledge proof systems. , 25(1):169–192, 1996.
O. Goldreich and E. Kushilevitz. A perfect zero-knowledge proof for a decision problem equivalent to [D]{}iscrete [L]{}ogarithm. , 6:97–116, 1993.
O. Goldreich, S. Micali, and A. Wigderson. Proofs that yield nothing but their validity or all languages in NP have zero-knowledge proof systems. , 38(3):691–729, 1991.
O. Goldreich, A. Sahai, and S. Vadhan. Honest-verifier statistical zero-knowledge equals general statistical zero-knowledge. In [*Proc. of the [30]{}th ACM Ann. Symp. on Theory of Computing (STOC)*]{}, pages 399–408, 1998.
S. Goldwasser, S. Micali, and C. Rackoff. The knowledge complexity of interactive proof systems. , 18(1):186–208, 1989.
S. Goldwasser and M. Sipser. Private coins versus public coins in interactive proof systems. In [*Proc. of the [18]{}th ACM Ann. Symp. on the Theory of Computing (STOC)*]{}, pages 59–68, 1986.
C. Hoffmann. , volume 136 of [*Lecture Notes in Computer Science*]{}. Springer Verlag, 1982.
C. Hoffmann. Subcomplete generalizations of [G]{}raph [I]{}somorphism. , 25:332–359, 1982.
J. Köbler, U. Schöning, and J. Torán. raph [I]{}somorphism is low for [PP]{}. In [*Symposium on Theoretical Aspects of Computer Science*]{}, volume 577 of [*Lecture Notes in Computer Science*]{}, pages 401–411. Springer Verlag, 1992.
E. M. Luks. Isomorphism of graphs of bounded valence can be tested in polynomial time. , 25:42–65, 1982.
E. M. Luks. L. Babai. Automorphism groups, isomorphism, reconstruction. , pages 1447–1540. Elsevier Publ., 1995.
T. Okamoto. On relationships between statistical zero-knowledge proofs. In [*Proc. of the [28]{}th ACM Ann. Symp. on Theory of Computing (STOC)*]{}, pages 649–658, 1996.
U. Schöning. Graph isomorphism is in the low hierarchy. In [*Proceedings of the STACS’87, Lecture Notes in Computer Science, [247]{}*]{}, pages 114–124, New York/Berlin, 1987. Springer-Verlag.
C. C. Sims. , volume 697 of [*Lecture Notes in Computer Science*]{}, pages 108–124. Springer Verlag, Berlin, 1978.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the fact that the contribution of the states with isospin $I=0$ in the difference of the amplitudes of the processes $\ggc$ and $\gg0$ is very small, we have analyzed the dispersion sum rules for the difference between the dipole polarizabilities of the charged and neutral pions as a function of the $\s$ meson parameters. Then taken into account the current perturbation value of $\amn=-1.9$, we have found $\amc=9.4 \div 8.2$ for values of the $\s$ meson parameter within the region: $\ms=400 \div 550$ MeV, $\G_{\s}=400\div 600$, $\G_{\s\to\g\g}=0 \div 3$ keV. It has been shown that the value of the decay width of the $h_1(1170)$ meson into $\g\pi^0$ can be found if the difference $\amc$ is reliably determined from the experiment. Estimation of the optimal value of the decay width $\s\to\g\g$ has given $\G_{\s\to\g\g}\lesssim 0.7$keV.'
author:
- 'L.V. Fil’kov'
title: Relation between the dipole polarizabilities of charged and neutral pions
---
0.5cm
0.3cm
Introduction
============
Pion polarizabilities are the fundamental structure parameters characterizing the behavior of the pion in an external electromagnetic field. Dipole polarizabilities arise as ${\cal O}(\nu_1 \nu_2)$ terms in the expansion of the non-Born amplitudes of Compton scattering in powers of the initial and final photon energies $\nu_1$ and $\nu_2$. In terms of the electric $\alpha_1$ and magnetic $\beta_1$ dipole polarizabilities, the corresponding effective interaction has the form: $$H^{(2)}_{eff}=-\frac12\,4\pi\,(\alpha_1\,\vec{E}^2+\beta_1\,\vec{H}^2).$$
The dipole polarizabilities measure the response of the hadron to quasistatic electric and magnetic fields. In what follows, these parameters are given in units $\unit$.
The values of the pion polarizabilities are very sensitive to the predictions of different theoretical models. Therefore, an accurate experimental determination of them is very important for testing the validity of such models.
At present, the value of the difference of the charged pion dipole polarizabilities found from radiative $\pi^+$ meson photoproduction from protons [@mainz] is equal to $11.6\pm 1.5_{stat}\pm 3.0_{syst}\pm 0.5_{mod}$ and close to the value obtained from scattering of high energy $\pi^-$ mesons off the Coulomb field of heavy nuclei in Serpukhov [@antip] and equal to $13.6\pm 2.8\pm 2.4$. On the other hand, these values differ from the prediction of the chiral perturbation theory (ChPT) (($4.7\div 6.7$) [@gasser]). The experiment of the Lebedev Physical Institute on radiative pion photoproduction from protons [@lebed] has given $\alpha_{1\pi^+}=20\pm 12$. This value has large error bars but nevertheless shows a large discrepancy with regard to the ChPT predictions, as well.
The preliminary result of the COMPASS collaboration $\amc=5.0\pm 3.4_{stat}\pm 1.2_{syst}$ has been found by studying the $\pi^-$ meson scattering off the Coulomb field of heavy nuclei [@compass]. This result is more close to the ChPT prediction. However to obtain this result the authors used very big values of momentum transfer $Q^2_{max}\approx 5\times 10^{-3}$(Gev/c)$^2$. In this region an interference between the Coulomb and nuclear amplitudes should be taken into account [@fil4; @fil-PS; @walch]. It should be noted that the authors of work [@antip] chose $Q^2<6\times 10^{-4}$ (GeV/c)$^2$ to guarantee that the contribution of the strong interaction below the Coulomb peak is negligible.
The charged pion polarizabilities can be also found by studying the process $\ggc$. Investigation of the process $\ggn$ at low and middle energies was carried out in the framework of different theoretical models and, in particular, in the frame of dispersion relations (DR). In Ref. [@fil1; @fil2; @fil3] we have analyzed the processes $\gg0$ and $\ggc$ using DRs with subtractions for the invariant amplitudes $\mpp$ and $\mm$ without partial-wave expansions. The subtraction constants have been uniquely determined in these works through the pion polarizabilities. The values of the polarizabilities have been found from the fit of the experimental data of the processes $\ggc$ and $\gg0$ up to 2500 MeV and 2250 MeV, respectively. As a result, we have found $\amc=13.0^{+2.6}_{1.9}$ and $\amn=-1.6\pm 2.2$. The result for $\amc$ is in good agreement with the values obtained in Ref.[@mainz; @antip; @lebed] whereas it is at variance with the ChPT prediction.
In the works [@bab; @holst; @kal; @garcia] the dipole polarizabilities of charged pions have been determined from the experimental data of the process $\ggc$ in the full energy region $\sqrt{t}<700$ MeV, (where $t$ is the square of the total energy in $\g\g$ c.m. system). The results obtained in these works are close to ChPT predictions [@gasser; @burgi]. However, the values of the experimental cross section of the process $\ggc$ in this region [@pluto; @dm1; @dm2; @mark] are very ambiguous, and, as has been shown in Ref. [@holst; @fil3], even changes of these values by more than 100% are still compatible with the present error bars.
Therefore, it is necessary to consider other additional possibilities of the $\amc$ determination.
Such an information could be obtained from the dispersion sum rules (DSR) for these parameters. However, the main contribution to DSR for $\amc$ is given by the $\s$ meson, which is very wide and this causes additional uncertainties in the DSR calculation.
On the other hand, if we consider the DSR for the difference between the charged and neutral meson polarizabilities $\Delta(\am)=(\amc-\amn)$, then the contribution of mesons with isotopic spin $I=0$ in the $t$ channel to this difference would be equal 0 [@lvov1; @fgr], when the masses of charged and neutral $\pi$ mesons are equal each other. As a result, a model dependence for $\amc$ should be decreased essentially.
In the present work we investigate DSR for $\dcn$ as a function of the decay width of $\sigma\to\g\g$, when the masses of the charged and neutral $\pi$ mesons are not equal each other. As will be shown, the contribution of the $\sigma$ meson is small in this case too, and we can find a realistic limit on the value of $\amc$.
It has been shown that the value of the decay width of the $h_1(1170)$ meson into $\g\pi^0$ can be found if the value $\dcn$ is reliably determined from experiment.
Dispersion sum rules for the pion polarizabilities
==================================================
We will consider the helicity amplitudes $\mpp$ and $\mm$. These amplitudes have no kinematical singularities or zeros [@aber]. The relations between the amplitudes $\ggc$, $\gg0$ and the ones with isotopic spins $I=0$ and $I=2$ read F\_C&=&(F\^0 + F\^2),\
F\_N&=&(F\^0 - F\^2). \[iso\]
The dipole ($\alpha_1$ and $\beta_1$) polarizabilities are defined [@rad; @fil2] through expansion of the non-Born helicity amplitudes of Compton scattering on the pion in powers of $t$ at fixed $s=\m$ (s=,t)&=&2+[O]{}(t),\
(s=,t)&=&2+[O]{}(t), \[mpm\] where $\m$ is the $\pi$ meson mass (different for $\pi^0$ and $\pi^{\pm}$), $t+s+u=2\m$.
The dispersion sum rules for the difference of the dipole polarizabilities was obtained in Ref. [@fil1] using DRs at fixed $u=\m$ without subtractions for the amplitude $\mpp$. In this case, the Regge-pole model allows the use of DR without subtractions [@aber]. Such a DSR is &=& {\_[4]{}\^ .\
& & .+\_[4]{}\^ }. \[dsr1m\]
As is evident from Eq.(\[iso\]), the contribution of the isoscalar mesons to the difference $\dcn$ equals 0 (if the masses of the charged and neutral pions are equal). We will study this difference when these masses do not equal each other.
The DSRs for the charged pions are saturated by the contributions of the $\rho(770)$, $b_1(1235)$, $a_1(1260)$, and $a_2(1320)$ mesons in the $s$-channel and $\s$, $f_0(980)$, $f_0^{\i}(1370)$ in the $t$-channel. For the $\pi^0$ meson the contribution of the $\rho$, $\omega(782)$, $\phi$, $h_1(1170)$, and $b_1(1235)$ mesons are considered in the $s$-channel and the same mesons as for the charged pions in the $t$-channel. Besides, we take into account a nonresonant S-wave contribution of two charged pions in the t channel.
The parameters of the $\rho$, $\omega$, $\phi$, $b_1$ and $a_2$ mesons are given by the Particle Data Group [@pdg]. For the $a_1(1260)$ meson we took $m_{a_1}=1230$ MeV [@pdg], $\G_{a_1}=$425 MeV (the average value of the PDG estimate [@pdg]), $\G_{a_1\to \g\pi^{\pm}}=0.64$ MeV [@zel].
The parameters of the $f_0(980)$ and $f_0^{\i}(1370)$ mesons are taken as follows:
$f_0(980)$: $m_{f_0}=980$ MeV [@pdg], $\G_{f_0}=70$ MeV (the average of the PDG [@pdg] estimate), $\G_{f_0\to \g\g}=0.56\times 10^{-3}$ MeV, $\G_{f_0\to \pi\pi}=0.84\,\G_{f_0}$ [@anis], $\G_{f_0\to K\bar{K}}=0.16\,\G_{f_0}$;
$f_0^{\i}(1370)$: $m_{f_0^{\i}(1370)}=1430$ MeV, $\G_{f_0^{\i}(1370)}=145$ MeV, $\G_{f_0^{\i}(1370)\to \g\g}=0.54\times 10^{-5}$ MeV [@morg], $\G_{f_0^{\i}(1370)\to \pi\pi}=0.26\,\G_{f_0^{\i}(1370)}$ [@bugg].
The mass and the total decay width of the $h_1(1170)$ meson are taken from PDG: $m_{h_1}=1170$ MeV, $\G_{h_1}=360$ MeV. The decay $h_1\to \g\pi^0$ has not yet been observed. Therefore we use this decay width according the work [@garcia]: \_[h\_1\^0]{}=C\_[h\_1]{}, where the coefficient $C_{h_1}$ can be estimated using nonet symmetry [@garcia; @lvov2]: C\_[h\_1(1170)]{}9 C\_[b\_1(1235)]{}0.45. As a result we have $\G_{h_1\to\g\pi^0}\simeq 1.6845\pm 0.44$ MeV.
Recently, a lot of works have been devoted to the study of the $\s$ meson (see, for example [@capr1; @capr2; @mous; @kamin; @pela]). An average of the most advanced data on the $\s$ meson gives ([@pela]) m\_=4466, \_/2=2765. In our analysis we use the values of the mass of the $\s$ meson and its total decay width in the following intervals:
$m_{\s}= 425 \div 550$ MeV, $\Gamma_{\s}=400\div 600$ MeV.
The values of the decay width of $\s\to\g\g$ we consider from 0 up to 3 keV.
Expressions for the imaginary parts of the resonances under consideration are given in Appendix.
Besides the contribution of the $\s$, $f_0(980)$, and $f_0^{\i}(1370)$ mesons we have taken into account a nonresonant contribution of the $S$-waves with the isospin $I=0$ and 2 according to the diagrams of Fig. 1.
It is worth noting that the vertexes of the $\s$ and $f_0$ meson poles in the dispersion approach include the full dynamics of the transitions on the mass shell. In this case there is no need to consider direct and rescattering mechanisms of transition separately.
According to the unitarity condition, the imaginary part of the amplitude $\mpp$ for the $\pi^+\pi^-$-loop diagram in Fig. 1 can be written as Im \^[(s)]{}=[B]{} Re T\_[\^+\^-]{}, where ${\cal B}\equiv {\cal B}(\ggc)$ is the contribution of the Born amplitude to the $S$-wave of the $\ggc$ amplitude and equal to \[born\] [B]{}=16()(), $q (q_0)$ is the momentum (energy) of the $\pi$ meson. The Born amplitude can be expressed in terms of the $I=0$ and $I=2$ isospin amplitudes as \[bornI\] [B]{}= B\^[(I=0)]{}+ B\^[(I=2)]{}. Taking into account that (0)=- B\^[(I=0)]{}+ B\^[(I=2)]{}=0 we have [@bogl] B\^[(I=0)]{}=, B\^[(I=2)]{}=.
The amplitudes of $\pi\pi$ scattering are expressed through the amplitudes in the isotopic space $T^{(I=0)}$ and $T^{(2)}$ as follows: \[pipi\] T\_[\^+\^-\^+\^-]{}&=&(T\^[(0)]{}+ T\^[(2)]{}),\
T\_[\^+\^-\^0\^0]{}&=&(T\^[(0)]{}- T\^[(2)]{}).
According to the relations (\[bornI\]) and (\[pipi\]) the imaginary parts of the $\pi\pi$ loop contributions to the $S$-wave of the amplitude $\mpp$ are equal to Im \^[(s)]{}(0)= Re(T\^[(0)]{} + T\^[(2)]{}), Im \^[(s)]{}()= Re(4 T\^[(0)]{} + T\^[(2)]{} ).
The amplitudes $T^{(0)}$ and $T^{(2)}$ can be presented as ReT\^[(I)]{}=\_I (t) (t), where $\delta_0^I(t)$ is the phase-shift of the $S$-wave of $\pi\pi$ scattering with isospin $I$ and $\eta_I$ is the inelasticity.
The expression for the phase-shift $\delta_0(t)$ has been determined using the parameterization of Ref. [@garcia1]. At low energy $t\lesssim 4\mk$ we have &=& { +B\_0+B\_1 w(t) .\
& & +.B\_2 w(t)\^2+B\_3 w(t)\^3 }, \[d01\] where
$$w(t)=\dfrac{\sqrt{t}-\sqrt{4\mk-t}}{\sqrt{t}+\sqrt{4\mk-t}},$$ and $\eta_0^0$ is equal to 1.
For the energy $4\mk <t<$ (1.42 GeV)$^2$ we use [@garcia1] \_0\^0(t) = d\_0+B+C +D(t-4), \[d02\] \_0\^0(t)&=&exp, \[eta\] where $q_k=\sqrt{t/4-\mk}$ and $q_\eta=\sqrt{t/4-\me}$; $m_k$ and $m_{\eta}$ are the masses of $K$ and $\eta$ mesons, respectively.
The parameters in Eqs(\[d01\]-\[eta\]) are listed in Table 1.
$B_0$ $7.26\pm 0.23$ $d_0$ $(227.1 \pm 1.3)^{\circ}$ $\ep_1$ $4.7\pm 0.2$
------- ---------------- ------- --------------------------- --------- ----------------
$B_1$ $-25.3\pm 0.5$ $B $ $(94.0\pm 2.3)^{\circ}$ $\ep_2$ $-15.0\pm 0.8$
$B_2$ $-33.1\pm 1.2$ $C $ $(40.4\pm 2.9)^{\circ}$ $\ep_3$ $4.7\pm 2.6$
$B_3$ $-26.6\pm 2.3$ $D $ $(-86.9\pm 4.0)^{\circ}$ $\ep_4$ $0.38\pm 0.34$
: The values of the coefficients in Eqs. (\[d01\],\[eta\]). [@garcia1]
To describe the phase-shift $\delta_0^2(t)$ we use Schenk’s parameterization [@schenk] in the energy region up to 1.5 GeV, assuming that $\eta_0^2=1$ [@ananth] ={A0\^2+B\_0\^2 q\^2 +C\_0\^2 q\^4+D\_0\^2 q\^6}(), where && A\_0\^2=-0.044, B\_0\^2=-0.0855, C\_0\^2=-0.00754\^4,\
&& D\_0\^2=0.000199\^6, s\_0\^2=-11.9.
Calculation of $\amc$
=====================
The value of $\amn$ has been determined from the investigation of the process $\gg0$ in the works [@fil1; @kal; @garcia]: $-1.6\pm 2.2$, $-0.6\pm 1.8$, $-1.25\pm 0.08\pm 0.15$.
These values are in good agreement with the prediction of ChPT [@bell] $\amn^{ChPT}=-1.9\pm 0.2$. Therefore, in order to determine $\amc$ we have added the value of $\amn^{ChPT}$ to the results of the calculations of $\Delta(\am)$ with help of DSR (\[dsr1m\]) at different values of the decay width of $\sigma\to\g\g$, when the mass and the total decay width of the $\s$ meson vary within the following values: $\ms=400\div 550$ MeV, $\G_{\s}=400\div 600$ MeV.
The results of the calculation are shown in Fig.2. Line (1) corresponds to calculations with $\sqrt{t_{\s}}=\ms-i 1/2\G_{\s}=446-i 276$. Lines (2) and (3) correspond to $\ms=$400 MeV, $\G_{\s}=$600 MeV and $\ms=$550 MeV, $\G_{\s}=$400 MeV, respectively. As is evident from this Figure the values of $\amc$ weakly depend on the mass and the total decay width of the $\s$ meson in the region under consideration. The values of $\amc$ obtained are within $9.4\div 8.2$.
The greatest contribution to $\dcn$ is given by the $\omega$ and $h_1$ mesons. The parameters of the $\omega$ meson and so its contribution are well known. On the other hand the experimental data on the $h_1$ meson are very poor. In particular, the decay of this meson into $\g\pi^0$ was not observed yet still in the experiment. Therefore, a reliable experimental determination of $\dcn$ will allow to determine the real value of the decay width $\G_{h_1\to\g\pi^0}$. For example, if the result of work [@mainz] $\amc=11.6$ is confirmed, then $\G_{h_1\to\g\pi^0}=0.875$ MeV.
Line (4) in Fig.1 is the result of the calculations of DSR (\[dsr1m\]) for $\amc$ at $\ms=446$MeV and $\G_{\s}=552$MeV. This result strongly depends on the decay width $\G_{\s\to\g\g}$ and indicates that realistic values of $\amc$ can be obtained if $\G_{\s\to\g\g}\lesssim 0.7$keV.
The influence of the upper integration limit ($\Lambda$) in the DSR (\[dsr1m\]) on the results of the calculation was investigated. They are not practically changed for $\Lambda$ more than (6 GeV)$^2$. In the present work we performed the integration up to (20 GeV)$^2$.
Conclusions
===========
Using the fact that the contribution of the state with isospin $I=0$ to the difference $\Delta(\am)=\amc -\amn$ is very small we have analyzed DSR for this difference at the real values of the pion masses. DSR has been calculated for the $\s$ meson parameters within the intervals: $\ms=400\div 550$ MeV, $\G_{\s}=400\div 600$, $\G_{\s\to\g\g}=0\div 3$ keV. In order to determine $\amc$ we have added $\amn^{ChPT}$=-1.9 to $\dcn$. The values of $\amc$ found weakly depend on the $\s$ meson parameters and are in the range $\amc=9.4\div 8.2$. This result is in agreement with the experimental values obtained in work [@mainz], whereas it is at variance with the calculations in the framework of ChPT [@gasser].
It has been shown that further experimental investigation of $\dcn$ can be an opportunity to determine the decay width $\G_{h_1\to\g\pi^0}$.
Besides, the analysis of DSR for $\amc$ showed that more realistic values of this parameter ($\amc< 15$) can be obtained with help of DSR (\[dsr1m\]) if the decay width $\G_{\s\to\g\g}\lesssim 0.7$keV. The values $\G_{\s\to\g\g}\lesssim 1$ keV were obtained early in works [@fil1; @fil2; @dubn; @achas] also. Results with $\G_{\s\to\g\g}>1$ keV quoted in the resent literature are listed in [@mous; @hofer]
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank H.J. Arends, A. Thomas, Th. Walcher, V.L. Kashevarov, and A.I. L’vov for useful discussions. This work was supported by the Deutsche Forschungsgemeinschaft (SFB 443).
Appendix A {#appendix-a .unnumbered}
==========
The contributions of the vector and axial-vector mesons $(\rho, \omega,
\phi, a_1$, and $b_1 )$ to $Im \mpp(s,u=\m)$ are calculated with the help of the expression Im\^[(V)]{}(s,u=)=4g\_V\^2 s where $m_V$ is the meson mass, the sign “+” corresponds to the contribution of the $a_1$ and $b_1$ mesons and g\_V\^2=6()\^3 \_[V]{} D\_1(m\_V\^2)/D\_1(s), \_0=()\^[32]{}D\_1(m\_V\^2)/D\_1(s)\_V Here $D_1$ is connected with the centrifugal potential and equal to $D_1=1+(q_i r)^2$ [@blatt], $r=1$fm is an effective interaction radius, $\G_V$ and $\G_{V\to \g\pi}$ are the total decay width and the decay width into $\g\pi$ of these mesons. The momentums $q_i^2$ for $(\rho,\,\omega,\,\phi,\,a_1$, and $b_1 )$ mesons are equal to $(s-4\m)/4$, $(s-9\m)/4$, $(s-4\mk)/4$, $(s-(m_{\rho}+\m)/)4$, and $(s-16\m)/4$, respectively.
Appendix B {#appendix-b .unnumbered}
==========
The amplitude of the contribution of a scalar meson to the process $\ggn$ can be written as \[tampl\] T=. Then it is easy to show that the imaginary part of the amplitude $Im\mpp^{\s}(t)$ of the $\s$ meson contributions to the process under consideration could be presented as \[Iampl\] Im\^(t)=, where \[gampl\] g\_=,
\[g0ampl\] \_0\^= ( )\^[1/2]{}\_ . These expressions (\[Iampl\])-(\[g0ampl\]) can be very useful to describe scaler mesons with large decay widths.
As the two $K$ mesons give a big contribution to the decay width of the $f_0(980)$ meson and the threshold of the reaction $\g\g\to K\overline{K}$ is very close to the mass of the $f_0(980)$ meson, we consider Flatt[é]{}’s expression [@flatte] for the $f_0(980)$ meson contribution to the process $\ggn$.
For $t>4\mk$: Im \^[f\_0]{}=g\_[f\_0]{}, where 0 &=&. For $t<4\mk$: &&Im =g\_[f\_0]{} 0(\^2 +0\^2 )\^[-1]{}, 0=()\^[1/2]{}.
[99]{} J. Ahrens [*et al.*]{}, Eur. Phys. J. A [**23**]{}, 113 (2005). Yu.M. Antipov [*at al.*]{}, Phys. Lett. B [**121**]{}, 445 (1983); Yu.M. Antipov [*at al.*]{}, Z. Phys. C [**26**]{}, 495 (1983). J. Gasser, M.A. Ivanov, and M.E. Sainio, Nucl. Phys. [**B745**]{}, 84 (2006). T.A. Aibergenov [*at al.*]{}, Czech. J. Phys. B [**36**]{}, 948 (1986). A. Guskov (COMPASS Collaboration), Fizika B [**17**]{}, 313 (2008). L.V. Fil’kov and V.L. Kashevarov, Proceed. of “NSTAR 2007”, Bonn, Germany, 05 – 08 September (2007), p.179; arXiv:0802.0965 \[nucl-th\]. L.V. Fil’kov and V.L. Kashevarov, Proceed. of 6th Intern. Workshop on Chiral Dynamics, 06 – 10 July (2009), Bern, Switzerland, PoS (CD09) 036; arXiv:0909.4849 \[hep-ph\]. Th. Walcher, Progr. Part. Nucl. Phys. [**61**]{}, 106 (2008). L.V. Fil’kov and V.L. Kashevarov, Eur. Phys. J. A [**5**]{}, 285 (1999). L.V. Fil’kov and V.L. Kashevarov, Phys. Rev. C [**72**]{}, 035211 (2005). L.V. Fil’kov and V.L. Kashevarov, Phys. Rev. C [**73**]{}, 035210 (2006). D. Babusci [*et al*]{}., Phys. Lett. [**B277**]{}, 158 (1992). J.F. Donoghue and B.R. Holstein, Phys. Rev. D [**48**]{}, 137 (1993). A.E. Kaloshin and V.V. Serebryakov, Z. Phys. C [**64**]{}, 689 (1994). R. Garcia-Martin and B. Moussallam, Eur. Phys. J. C [**70**]{}, 155 (2010). U. Bürgi, Nucl. Phys. [**B479**]{}, 392 (1997). C. Berger [*et al.*]{}, (PLUTO Collaboration), Z. Phys. C [**26**]{} 199 (1984). A. Courau [*et al.*]{}, (DM1 Collaboration), Nucl. Phys. B [**271**]{}, 1 (1986). Z. Ajaltoni [*et al.*]{}, (DM2 Collaboration), Phys. Lett. B [**194**]{}, 573 (1987). J. Boyer [*et al.*]{}, (Mark II Collaboration), Phys. Rev. D [**42**]{}, 1350 (1990). A.I. L’vov and V.A. Petrun’kin, Preprint of P.N. Lebedev Physical Institute, No. 170 (1977). L.V. Filkov, I. Guiaşu, and E.E. Radescu, Phys. Rev. D [**26**]{}, 3146 (1982). H.A. Abarbanel and M.L. Goldberger, Phys. Rev. [**165**]{}, 1594 (1968). I. Guiaşu and E.E. Radescu, Ann. Phys. (NY) [**122**]{}, 436 (1979). J. Beringer [*et al.*]{}, (PDG), Phys. Rev. D [**86**]{}, 010001 (2012). M. Zielinski [*et al.*]{}, Phys. Rev. Lett. [**52**]{}, 1195 (1984). V.V. Anisovich [*et al.*]{}, Phys. At. Nucl. [**65**]{}, 1545 (2002). D. Morgan and M.R. Pennington, Z. Phys. [**C48**]{}, 623 (1990). D.V. Bugg, A.V. Sarantsev, and B.S. Zou, Nucl. Phys. [**B471**]{}, 59 (1990). A.I. L’vov and V.A. Petrun’kin, Sov. Physics-Lebedev Inst. Reports, No 12 (1985). I. Caprini, G. Colangelo, and H. Leutwiler, Phys. Rev. Lett. [**96**]{}, 132001 (2006); hep-ph/0512364. I. Caprini, Phys. Rev. D [**77**]{}, 114019 (2008); arXiv:0804.3504. B. Moussallam, Eur. Phys. J. C[**71**]{}, 1814 (2011). R. Kamiński, R. Garcia-Martin, J.R. Peláes, and J. Ruiz de Elvira, Nucl. Phys. Proc. Suppl. [**B234**]{}, 253 (2012); arXiv:1211.2617 \[hep-ph\]. J.R. Peláez, J. Nebreda, G. Rios, and J. Ruiz de Elvira, Acta Phys. Polon. Supp. [**6**]{}, 735 (2013); arXiv:1304.5121 \[hep-ph\]. M. Boglione, M.R. Pennington, Eur. Phys. J. C [**9**]{}, 11 (1999). R. Garcia-Martin, R. Kamiński J.R. Peláez [*et al.*]{}, Phys. Rev. D [**83**]{}, 074004 (2011). A. Schenk, Nucl. Phys. [**B363**]{}, 97 (1991). B. Ananthanarayan [*et al.*]{}, Phys. Report [**353**]{}, 207 (2001). S. Bellucci, J. Gasser, and M.E. Sainio, Nucl. Phys. [**B423**]{}, 80 (1994). S. Dubni$\breve{c}$ka, A.Z. Dubni$\breve{c}$kov$\acute{a}$, and M. Se$\breve{c}$ansk$\acute{y}$, Acta Phys. Slov. [**55**]{}, 25 (2005). N.N. Achasov and G.N. Shestakov, JETP Lett. [**88**]{}, 295 (2008). M. Hoferichter, D.R. Philips, and C. Schat, Eur. Phys. J., C [**71**]{}, 1743 (2011). J.M. Blatt and V.F. Weisskopf, Theoretical Nuclear Physics, Springer-Verlag, New York, p.359–365, 386–389 (1979). S.M. Flatté, Phys. Lett. [**B63**]{}, 228 (1976).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A strategy to evaluate the distribution of the largest Schmidt eigenvalue for entangled random pure states of bipartite systems is proposed. We point out that the multiple integral defining the sought quantity for a bipartition of sizes $N,M$ is formally identical (upon simple algebraic manipulations) to the one providing the probability density of Landauer conductance in open chaotic cavities supporting $N$ and $M$ electronic channels in the two leads. Known results about the latter can then be straightforwardly employed in the former problem for both systems with broken ($\beta=2$) and preserved ($\beta=1$) time reversal symmetry. The analytical results, yielding a continuous but not everywhere analytic distribution, are in excellent agreement with numerical simulations.'
address: |
Abdus Salam International Centre for Theoretical Physics\
Strada Costiera 11, 34151 Trieste, Italy
author:
- Pierpaolo Vivo
title: 'Largest Schmidt eigenvalue of entangled random pure states and conductance distribution in chaotic cavities.'
---
Introduction
============
Consider two sets of $n$ correlated random variables in $[0,1]$, $\{\lambda_i\}$ and $\{T_i\}$ ($i=1,\ldots,n$), respectively distributed according to the following joint probability densities (jpd): $$\begin{aligned}
\mathcal{P}_1(\lambda_1,\ldots,\lambda_n) &=C_{n,\alpha}^{(\beta)}\delta\left(\sum_{i=1}^n\lambda_i -1\right)\prod_{i=1}^n
\lambda_i^{\alpha}\prod_{j<k}|\lambda_j-\lambda_k|^\beta\label{P1}\\
\mathcal{P}_2(T_1,\ldots, T_n) &=K_{n,\alpha^\prime}^{(\beta)}\prod_{i=1}^n
T_i^{\alpha^\prime}\prod_{j<k}|T_j-T_k|^\beta\label{P2}\end{aligned}$$ where $C_{n,\alpha}^{(\beta)}$ and $K_{n,\alpha^\prime}^{(\beta)}$ are known normalization constants and $\beta=1,2$. In the following two subsection we will provide physical motivations for considering such sets, namely set $1$ corresponds to the distribution of Schmidt eigenvalues for entangled random pure states in bipartite systems (see subsection \[ent\]), while set $2$ corresponds to the distribution of transmission eigenvalues of an open cavity in the chaotic regime (see subsection \[condcav\]).
Consider now the following statistical quantities:
- The cumulative distribution $Q_n(x)=\mathrm{Prob}[\lambda_{\mathrm{max}}\leq x]$ of the largest member of set $1$, $\lambda_{\mathrm{max}}=\max_i\{\lambda_i\}$. By definition, this is given by the following $n$-fold integral: $$\label{cum1}
Q_n(x)=\int_{[0,x]^n}d\lambda_1\cdots d\lambda_n \mathcal{P}_1(\lambda_1,\ldots,\lambda_n)$$ Differentiating $Q_n(x)$, one obtains the *probability density* of $\lambda_{\mathrm{max}}$, p\_n(x)=Q\_n(x)\[densq\]
- The probability density $\mathcal{P}_G(y)=\mathrm{Prob}[y\leq G\leq y+dy]$ of the quantity $G=\sum_{i=1}^n T_i$, which is given by the following $n$-fold integral [^1]: $$\mathcal{P}_G(y)=\int_{[0,1]^n}dT_1\cdots dT_n \mathcal{P}_2(T_1,\ldots, T_n)\delta\left(y-\sum_{i=1}^n T_i\right)$$
Simple algebraic manipulations, summarized in Appendix \[appA\] lead to the following relation between the two quantities above: $$\label{main}
\boxed{Q_n(x)=\frac{C_{n,\alpha}^{(\beta)}}{K_{n,\alpha}^{(\beta)}}
x^{n+\alpha n+\frac{\beta}{2}n(n-1)-1}\mathcal{P}_G\left(\frac{1}{x}\right),\qquad 1/n\leq x\leq 1,\ \alpha=\alpha^\prime}$$ The identity in eq. is the main result of this paper[^2]. Notwithstanding its remarkable simplicity, eq. actually permits an exact evaluation of $Q_n(x)$, the so far unavailable distribution of the largest Schmidt eigenvalue for [*finite*]{} $n$ (see subsection \[ent\]), in terms of $\mathcal{P}_G(y)$ (the probability density of Landauer conductance, see subsection \[condcav\]) about which much more is known.
The plan of the paper is as follows. In the next two subsections, we give a rather detailed introduction to the physics of entangled random pure states in bipartite systems (related to the set $1$ above) and about Landauer conductance in chaotic mesoscopic cavities supporting a finite number of electronic channels in the two attached leads (related to the set $2$ above). In section \[mainresults\] we exploit the identity to derive analytically the cumulative distribution and the density of $\lambda_{\mathrm{max}}$ in a few illustrative cases. These results are then compared with numerical simulations with excellent agreement. Eventually we present concluding remarks in section \[concl\] and technical developments in the three appendices.
Entangled random pure states. {#ent}
-----------------------------
Entanglement of pure bipartite systems is one of the most active areas of research nowadays, due to possible applications to quantum information and quantum computation problems [@NeilsenBook; @PeresBook]. It is also probably the simplest setting where well-behaved entanglement quantifiers can be defined, such as the von-Neumann or Rényi entropies of either subsystem [@PeresBook], the so called concurrence for two-qubit systems [@Wootters] or other entanglement monotones [@cappellini; @gour].
*Typical* properties of such states are best addressed by considering *random* pure states (see e.g. [@majreview] for an excellent review). More precisely, consider a bipartition of a $NM$-dimensional Hilbert space ${\cal
H}^{(NM)}$ as ${\cal H}^{(NM)}={\cal H}^{(N)}_A \otimes {\cal H}^{(M)}_B$, where we assume without loss of generality that $N\le M$. For example, $A$ may be taken as a given system (say a set of spins) living in an external environment (e.g., a heat bath) $B$. A quantum state $|\psi\kt$ of the composite system can be expanded as a linear combination $$\label{psi111}
|\psi\kt= \sum_{i=1}^N\sum_{\alpha=1}^M x_{i,\alpha}\,
|i^A\kt\otimes |\alpha^B\kt$$ where $|i^A\kt$ and $|\alpha^B\kt$ are two complete basis of ${\cal H}^{(N)}_A $ and ${\cal H}^{(M)}_B$ respectively. The coefficients $x_{i,\alpha}$’s of this expansion form the entries of a rectangular $(N\times M)$ matrix $\mathcal{X}$.
We consider here *entangled random pure* states $|\psi\kt$. This means that:
1. $|\psi\kt$ *cannot* be expressed as a direct product of two states belonging to the two subsystems $A$ and $B$.
2. the expansion coefficients $x_{i,\alpha}$ are random variables drawn from a certain probability distribution.
3. the density matrix of the composite system is simply given by $\rho=|\psi\kt \bra
\psi|$ with the constraint ${\rm Tr}[\rho]=1$, or equivalently $\bra \psi|\psi\kt=1$.
More precisely, the density matrix of $|\psi\kt$ can then be straightforwardly expressed as = \_[i,]{}\_[j,]{} x\_[i,]{} x\_[j,]{}\^\* |i\^Aj\^A| |\^B \^B|, \[dem2\] where the Roman indices $i$ and $j$ run from $1$ to $N$ and the Greek indices $\alpha$ and $\beta$ run from $1$ to $M$.
The *reduced* density matrix $\rho_A= {\rm Tr}_B[\rho]$ is obtained by tracing out the environmental degrees of freedom (i.e. those of the subsystem $B$): \_A = [Tr]{}\_B\[\]=\_[=1]{}\^M \^B||\^B. \[rdm1\] Using the expansion in Eq. one gets \_A = \_[i,j=1]{}\^N \_[=1]{}\^M x\_[i,]{} x\_[j,]{}\^\* |i\^Aj\^A|=\_[i,j=1]{}\^N W\_[ij]{} |i\^Aj\^A| \[rdm2\] where $W_{ij}$’s are the entries of the $N\times N$ matrix $\mathcal{W}=\mathcal{X} \mathcal{X}^{\dagger}$. In a similar way, one could obtain the reduced density matrix $\rho_B={\rm Tr}_A[\rho]$ of the subsystem $B$ in terms of the $M\times M$ matrix $\mathcal{W}^\prime=\mathcal{X}^\dagger \mathcal{X}$. The two matrices $\mathcal{W}$ and $\mathcal{W}^\prime$ share the same set of nonzero (positive) real eigenvalues $\{\lambda_1,\lambda_2,\ldots,\lambda_N\}$. In the diagonal basis, one can express $\rho_A$ as \_A= \_[i=1]{}\^N \_i |\^A\_i \^A\_i| \[diagA\] where $|\ld^A_i \kt$’s are the normalized eigenvectors of $\mathcal{W}=\mathcal{X}\mathcal{X}^{\dagger}$ and similarly for $\rho_B$. The original composite state $|\psi\kt$ in this diagonal basis reads: |= \_[i=1]{}\^[N]{} |\_i\^A|\^B\_i \[Sch1\] Eq. is known as the Schmidt decomposition, and the normalization condition $\bra \psi|\psi\kt=1$, or equivalently ${\rm Tr}[\rho]=1$, imposes the constraint on the sum of eigenvalues, $\sum_{i=1}^N \ld_i=1$.
For *random* pure states, the expansion coefficients in eq. can be typically drawn from an unbiased (so called *Hilbert-Schmidt*) distribution (real or complex) $$\label{HS}
{\rm Prob}[\mathcal{X}]\propto \delta\left( {\rm Tr}(\mathcal{X} \mathcal{X}^{\dagger})-1\right)$$ where the Dyson index $\beta=1,2$ corresponds respectively to real and complex $\mathcal{X}$ matrices[^3]. The meaning of eq. is clear: all normalized density matrices compatible with unitary invariance are sampled with equal probability, which corresponds to having minimal *a priori* information about the quantum state under consideration. This in turn induces nontrivial correlations among the Schmidt eigenvalues (which are now real random variables between $0$ and $1$ whose sum is $1$) and makes the investigation of several statistical quantities about such states quite interesting. Here we present a quick summary of known results:
- the joint probability density (jpd) of Schmidt eigenvalues, derived by Lloyd and Pagels [@LP], which is precisely given by $\mathcal{P}_1(\lambda_1,\ldots,\lambda_N)$ in eq. , with $n=N$ and $\alpha=(\beta/2)(N-M+1)-1$. Note that the delta function there guarantees that ${\rm Tr}[\rho]=1$. The normalization constant in this case reads [@ZS]: C\_[N,=(/2)(N-M+1)-1]{}\^[()]{}=
- the average von Neumann entropy for large $N,M$ (computed by Page [@Page95] for $\beta=2$ and extended in [@Arul1] to the case $\beta=1$);
- the average von Neumann entropy for [*finite*]{} $N,M$ and $\beta=2$, conjectured by Page [@Page95] and independently proven by many researchers soon after [@proofspage] also in a non-extensive setting [@malacarne];
- density of Schmidt eigenvalues (one-point function) for finite $(N,M)$, derived independently in [@densitybeta2] and [@adachi] for $\beta=2$ and in [@vivodens] for $\beta=1$;
- universality of eigenvalue correlations for $\beta=2$ [@liu];
- average fidelity between quantum states [@zyc] and distribution of so-called $G$-concurrence [@cappellini] for $\beta=2$;
- distribution of so-called [*purity*]{} for small $N$ [@giraud], and phase transitions in its Laplace transform for large $N$ [@scard1];
- full distribution of Rényi entropies (including large deviation tails), computed in [@majnadal] for large $N=M$ and all $\beta$s using a Coulomb gas method. As a byproduct, the authors also obtain in [@majnadal] the average and variance of Rényi entropy valid for large $N=M$, and the density of Schmidt eigenvalues for all $\beta$’s and $N=cM$ large;
- distribution of smallest eigenvalue (related to so-called *Demmel condition number* [@demmel]) for $\beta=1,2$ and finite $M=N$, derived independently in [@edeldemmel] and [@majbohi]. In the latter paper, a conjecture by Znidaric [@Znd] was proven [^4] (see also [@chen] for an extension of these results to the case $N\neq M$);
- distribution of largest eigenvalue for *large* $N=M$ and all $\beta$s [@majnadal], including small and large deviation laws. Typical fluctuations around the mean $\approx 4/N$ (once properly scaled) are found to follow the Tracy-Widom distribution (see also [@nadler] for a related result). No results seem to be available for the case of *finite* $N,M$.
Given the current interest in the distribution of extreme Schmidt eigenvalues, the reader may on the one hand wonder whether they really encode useful information, and on the other why the largest eigenvalue distribution for finite $N,M$ is much harder to obtain via the same strategy used for the smallest one [@majbohi].
In order to answer the first question, first note that due to the constraint $\sum_{i=1}^N \ld_i=1$ and the fact that all eigenvalues are nonnegative, it follows that[^5] $1/N\le \ld_{\rm max}\le 1$ and $0\le \ld_{\rm min} \le 1/N$. Now consider the following limiting situations. Suppose that the largest eigenvalue $\lambda_{\rm max}=\max_i\{\lambda_i\}$ takes its maximum allowed value $1$. Then it follows immediately that all the remaining $(N-1)$ eigenvalues must be identically $0$. In this situation eq. (\[Sch1\]) tells us that $|\psi\kt$ is fully [*unentangled*]{} (completely separable). On the other hand, if $\lambda_{\rm max}=1/N$ (i.e., it takes its lowest allowed value), all the eigenvalues must have the same value, $\lambda_i=1/N$ for all $i$. In this case, the pure state $|\psi \kt$ is [*maximally*]{} entangled, as this state maximizes the von Neumann entropy $S_{\mathrm{VN}}=-\sum_{i=1}^N\lambda_i\ln\lambda_i=\ln (N)$. In other words, the knowledge of the largest eigenvalue distribution really provides essential information about how entangled a random pure state is.
A discussion about the asymmetry in the treatment of smallest and largest Schmidt eigenvalues is included in Appendix \[appB\].
Landauer conductance in open cavities {#condcav}
-------------------------------------
Consider a cavity of submicron dimensions etched in a semiconductor and connected to the external world by two leads supporting $M$ and $N$ electronic channels. It is well established that the electrical current flowing through such a cavity when brought out of equilibrium by an applied external voltage presents time-dependent fluctuations which persist down to zero temperature [@beenakker] and are thus associated with the granularity of the electron charge $e$. Typical features observed in experiments include weak localization [@chang], universality in conductance fluctuations [@marcus] and constant Fano factor [@oberholzer]. The Landauer-Büttiker scattering approach [@beenakker; @landauer; @buttikerPRL] is rather successful in describing the statistics of quantum transport: it amounts to relating the wave function coefficients of the incoming and outgoing electrons through the unitary scattering matrix $\mathcal{S}$ ($2N_0\times 2N_0$, if $N_0=N+M$): $$\label{ScatteringMatrix S}
\mathcal{S}=
\begin{pmatrix}
\mathbf{r} & \mathbf{t}^\prime \\
\mathbf{t} & \mathbf{r}^\prime
\end{pmatrix}$$ where the transmission ($\mathbf{t},\mathbf{t}^\prime$) and reflection $(\mathbf{r},\mathbf{r}^\prime)$ blocks are submatrices encoding the transmission and reflection coefficients among different channels[^6]. Experimental observables can be extracted from the eigenvalues of the hermitian transport matrix $\mathbf{T}=\mathbf{t} \mathbf{t}^\dagger$: for example, the dimensionless conductance and the shot noise are given respectively by $G=\Tr(\mathbf{T})$ [@landauer] and $P=\Tr[\mathbf{T}(\mathbf{1}-\mathbf{T})]$ [@ya].
Random Matrix Theory (RMT) has been very successful in describing the statistics of universal fluctuations in such systems, when the corresponding classical dynamics is chaotic: the scattering matrix $\mathcal{S}$ is drawn from a suitable ensemble of random matrices, with the overall constraint of unitarity [@muttalib; @stone; @mellopereira]. A maximum entropy approach (under the assumption of ballistic point contacts [@beenakker]) forces the probability distribution of $\mathcal{S}$ to be uniform within the unitary group, i.e. $\mathcal{S}$ belongs to one of Dyson’s Circular Ensembles [@Mehta; @Dys:new].
From the uniformity of the probability density of $\mathcal{S}$ within the unitary group, the jpd of the transmission eigenvalues $\{T_i\}$ of the matrix $\mathbf{T}$, from which the statistics of interesting experimental quantities could be in principle derived, is precisely given by eq. , with $n=\min(N,M)$ and $\alpha^\prime = \frac{\beta}{2}(|N-M|+1)-1$ [@beenakker; @mellopereira; @forrcond]. There, the Dyson index $\beta$ characterizes different symmetry classes ($\beta=1,2$ according to the presence or absence of time-reversal symmetry). The eigenvalues $T_i$ are thus correlated real random variables between $0$ and $1$. The normalization constant $K_{n,\alpha^\prime}^{(\beta)}$ is explicitly known from the celebrated Selberg’s integral as: K\_[n,\^]{}\^[()]{}=\_[j=0]{}\^[n-1]{}
From , in principle the statistics of many observables of interest can be calculated. In particular, we focus here on the dimensionless Landauer conductance $G=\Tr(\mathbf{T})=\sum_{i=1}^n T_i$, which satisfies the bounds $0\leq G\leq n$. Mean and variance of $G$ and other quantities have been known for a long time when $N,M$ are large [@beenakker; @brouwer; @novaes], and recently also for finite $N,M$ [@novaes; @savin; @savin2; @vivovivo].
Stimulated by some recent experimental progresses [@hemmady], which made eventually possible to explore the [*full distribution*]{} of the conductance (and not just its mean and variance), a lot of effort has been devoted to its theoretical characterization. An explicit expression was first obtained for $n=1,2$ [@baranger; @jalabert; @garcia], while more results were available in the case of quasi one-dimensional wires [@muttalibwolfle] and 3D insulators [@muttalib3D]. For the shot noise, the full distribution was known only for $N=M=1$ [@pedersen]. Very recently, Sommers [*et al.*]{} [@sommers] announced two formulas for the distribution of conductance and shot noise, valid at arbitrary number of open channels and for any $\beta$, which are based on Fourier expansions. In [@savinnew], a complete and systematic approach to the full statistics of conductance and shot noise was brought forward. This is based on symmetric function expansions, and is valid for $\beta=1,2$ and arbitrary $M,N$. In addition, the authors of [@savinnew] were able to provide general formulae in terms of determinants or Pfaffians for the full probability distribution of conductance and shot noise at quantized $\beta$ and general $N,M$ (see below). In [@Kanz], the integrable theory of quantum transport in chaotic cavities (based on Painlevé transcendent) for $\beta=2$ was formulated, and recursion formulae for the efficient computation of conductance and shot noise cumulants have been derived. In two recent publications [@vivoPRL], the distributions of conductance, shot noise and integer moments were computed in the limit $N=M\gg 1$ using a Coulomb gas method, and long [*power-law*]{} tails were detected in the distributions[^7], a result confirmed by extensive numerical simulations. In [@kumar], exact results (basically equivalent to the general formulae in [@savinnew]) for the Laplace transform $\tilde{\mathcal{P}}_G(s)$ of the conductance distribution for any $N,M$ and all $\beta$s are formally given as follows:
- : $\tilde{\mathcal{P}}_G(s)\propto\mathrm{Pf}[\Psi^{(1)}_{j,k}(s)]_{j,k=0,\ldots,N-1}$ for $N$ even, and $\tilde{\mathcal{P}}_G(s)\propto\mathrm{Pf}\left[\begin{array}{cr}
\Psi^{(1)}_{j,k}(s) & \Phi^{(1)}_{j}(s) \\
-\Phi^{(1)}_{k}(s) & 0
\end{array}\right]_{j,k=0,\ldots,N-1}$ for $N$ odd, where $\mathrm{Pf}[\mathcal{A}]$ is the Pfaffian of the even-dimensional antisymmetric matrix $\mathcal{A}$ [@Mehta] and the proportionality constants are known explicitly. The arrays $\Psi^{(1)}_{j,k}(s)$ and $\Phi^{(1)}_{j}(s)$ are given by: $$\begin{aligned}
\Psi^{(1)}_{j,k}(s) &=\int_0^1\int_0^1\mathrm{sgn}(x-y)e^{-sx}e^{-sy}x^{\alpha^\prime +j}y^{\alpha^\prime +k}dx\ dy\\
\Phi^{(1)}_{j}(s) &=\int_0^1 e^{-sx}x^{\alpha^\prime +j}dx
\end{aligned}$$
- : in this case, we have a representation in terms of a Hankel determinant \_G(s)\_[j,k=0,…,N-1]{}\[hankel\]where: $$\Psi^{(2)}_{j,k}(s)=\int_0^1 e^{-s x}x^{\alpha^\prime +j+k}dx$$ that was first derived in [@Kanz].
Explicit inversions of the Laplace transforms above are always possible on a case-by-case basis, and a catalogue of such evaluations for few interesting cases is provided in ref. [@kumar]. In next section and in Appendix \[appC\], we will combine such explicit formulas with the identity in eq. to illustrate the validity of our approach to the distribution of largest Schmidt eigenvalue.
Two applications of the main identity {#mainresults}
=====================================
For illustrative purposes, we consider two applications of the main identity eq. . A third one is discussed in great detail in appendix \[appC\].
1. : in this case, the probability density of Landauer conductance has been derived explicitly in [@kumar] as: $$\begin{aligned}
\nonumber \mathcal{P}_G(y) &=\frac{3}{8}\left[y^5-(y-1)^3 (y^2-12 y+51)\theta(y-1)-(y-2)^3\right.\\
&\left.\times (y^2+6 y+24)
\theta(y-2)\right]\theta(3-y)\label{pptt}\end{aligned}$$ where $\theta(x)$ is Heaviside step function.
2. : again, the probability density of Landauer conductance in this case has been derived explicitly in [@kumar] as: $$\mathcal{P}_G(y) =
\begin{cases}
\frac{20}{143}y^{13/2} & \mbox{for }0\leq y\leq 1\\
\frac{5}{2288}\left[3003y^5-21021y^4+55770y^3\right. & \\
-70070y^2+42315y-9933 &\\
-32(y-2)^{7/2}(2 y^3+14 y^2+63 y+231) &\\
\left.\times\theta(y-2)\right] &\mbox{for }1\leq y\leq 3
\end{cases}\label{pptt1}$$
Note that:
- since the conductance density has always a compact support $[0,n]$ (where $n=\min(N,M)$), it follows immediately from eq. that the cumulative distribution of the largest Schmidt eigenvalue $Q_n(x)$ and its density $p_n(x)$ have compact support $1/n\leq x\leq 1$, as expected.
- since the conductance density is known to be continuous but not everywhere analytic [@sommers; @savinnew] (i.e. it displays ’critical’ points at which higher derivatives are discontinuous), the cumulative distribution and the density of the largest Schmidt eigenvalue enjoy this property too.
In fig. \[cumfig\] we plot the cumulative distributions $Q_n(x)$ corresponding to the two cases above. The curves are obtained via eq. , where $\mathcal{P}_G(y)$ is respectively given by and . The distributions are increasing functions of the argument $x$, as they should, and such that $Q_n(1)=1$. Differentiation of the analytical formulas provide the density of the largest eigenvalue, $p_n(x)$, which is plotted in figs. \[densityfig1\] and \[densityfig2\] along with numerical simulations. These are obtained as follows [@ZS; @ZyckBook]:
1. we generate $\kappa\simeq 10^4,10^5$ [*real*]{} Gaussian $M\times N$ matrices $\mathcal{X}$.
2. for each instance we construct the Wishart matrix $\mathcal{W}=\mathcal{X}^T \mathcal{X}$.
3. we diagonalize $\mathcal{W}$ and collect its $N$ real and non-negative eigenvalues $\{ \tilde{\lambda}_1,\ldots,\tilde{\lambda}_N\}$.
4. we define a new set of variables $0\leq \lambda_i\leq 1$ as $\lambda_i =\tilde{\lambda}_i /\sum_{i=1}^N \tilde{\lambda}_i$, for $i=1,\ldots,N$. The set of variables $\lambda_i$ is guaranteed to be sampled according to the measure (\[P1\]).
5. we construct a normalized histogram of $\lambda_{\rm{max}}=\max_i\{\lambda_i\}$.
The agreement between theory and simulations is excellent.
We also evaluated exactly the average of $\langle\lambda_{\rm max}\rangle=\int_{1/n}^1 dx\ x\ p_n(x)$ for the two cases above and found: $$\begin{aligned}
\langle\lambda_{\rm max}\rangle_1 &=\frac{25}{36} \approx 0.694444...\\
\langle\lambda_{\rm max}\rangle_2 &=\frac{1}{810}(378+89\sqrt{3}) \approx 0.656978...\end{aligned}$$ However, a general formula for $\langle\lambda_{\rm max}\rangle$ is still elusive (unlike $\langle\lambda_{\rm min}\rangle$ [@majbohi]) and may well attract further researches (see also Appendix \[appC\]).
![Cumulative distribution $Q_n(x)$ of the largest Schmidt eigenvalue via eq. for $n=N=3,M=4,\beta=1$ (violet) and $n=N=3,M=5,\beta=1$ (blue). This distribution is identically zero for $x\leq 1/n$ and $Q_n(1)=1$ as it should.[]{data-label="cumfig"}](cumulatives.eps){width=".7\hsize"}
Conclusions {#concl}
===========
We have presented an exact identity relating two statistical quantities which arise in different contexts: the cumulative distribution of the largest Schmidt eigenvalue for entangled random pure states in bipartite systems of sizes $(M,N)$ and the probability density of Landauer conductance in chaotic cavities supporting $N$ and $M$ electronic channels in the two external leads. Recent analytical results for the latter are exploited to derive (so far unavailable) exact formulas for the former quantity at [*finite*]{} $N,M$ (while large $N,M$ results are already available [@majnadal]), which is of interest in order to quantify the degree of entanglement of random pure states. A detailed introduction to the physics involved has been provided, along with a precise discussion of the asymmetry in the treatment of the smallest and largest Schmidt eigenvalue distributions. A general formula for $\langle\lambda_{\rm{max}}\rangle$, the average of the largest eigenvalue, valid for arbitrary $N,M$ is unfortunately still lacking and certainly deserves further investigations.\
\
[**Acknowledgments:**]{} it is a pleasure to thank Satya N. Majumdar, Céline Nadal, Oriol Bohigas and Gernot Akemann for collaborations on related projects and many interesting discussions, and Boaz Nadler for useful correspondence. I am also grateful to Dima Savin for a careful reading of the manuscript and many helpful advices.
Derivation of main identity. {#appA}
============================
Consider the cumulative distribution $Q_n(x)$ of $\lambda_{\mathrm{max}}=\max_i\{\lambda_i\}$ (eq. ): $$Q_n(x)=C_{n,\alpha}^{(\beta)}\int_{[0,x]^n}d\lambda_1\cdots d\lambda_n \delta\left(\sum_{i=1}^n\lambda_i -1\right)\prod_{i=1}^n
\lambda_i^{\alpha}\prod_{j<k}|\lambda_j-\lambda_k|^\beta$$ A change of variables $\lambda_i=x T_i$ leads to: $$\begin{aligned}
\nonumber Q_n(x) &=C_{n,\alpha}^{(\beta)} x^{n+\alpha n+\frac{\beta}{2}n (n-1)}\times\\
&\times\int_{[0,1]^n}d T_1\cdots d T_n
\delta\left(x\sum_{i=1}^n T_i -1\right)\prod_{i=1}^n T_i^{\alpha}\prod_{j<k}|T_j-T_k|^\beta\end{aligned}$$ Using the property of delta functions $\delta(\gamma Y)=\delta(Y)/|\gamma|$, we obtain straighforwardly eq. .
Asymmetry in the treatment of smallest and largest Schmidt eigenvalues {#appB}
======================================================================
Consider the cumulative distribution $\Theta_n(x)$ of the *smallest* member of set $1$, $\lambda_{\rm min}=\min_i \{\lambda_i\}$ (i.e. the distribution of the smallest Schmidt eigenvalue for entangled random pure states). By definition it is given by: $$\label{cummin}
\Theta_n(x)=C_{n,\alpha}^{(\beta)}\int_{[x,\infty]^n}d\lambda_1\cdots d\lambda_n \delta\left(\sum_{i=1}^n\lambda_i -1\right)\prod_{i=1}^n
\lambda_i^{\alpha}\prod_{j<k}|\lambda_j-\lambda_k|^\beta$$ where the upper limit of integration can be safely extended up to $\infty$ in view of the unit norm constraint. In the case $\alpha=0$ and $\beta=2$, which we focus on here for illustrative purposes, the evaluation of this multiple integral proceeds via the auxiliary function $\Theta_n(x,t)$ [@majbohi]: $$\label{cummint}
\Theta_n(x,t)=C_{n,0}^{(2)}\int_{[x,\infty]^n}d\lambda_1\cdots d\lambda_n \delta\left(\sum_{i=1}^n\lambda_i -t\right)
\prod_{j<k}|\lambda_j-\lambda_k|^2$$ such that $\Theta_n(x)\equiv \Theta_n(x,1)$. Next, one takes the Laplace transform of $\Theta_n(x,t)$: $$\label{cummint}
\int_0^\infty dt\Theta_n(x,t)e^{-st}=C_{n,0}^{(2)}\int_{[x,\infty]^n}d\lambda_1\cdots d\lambda_n e^{-s\sum_{i=1}^n\lambda_i }
\prod_{j<k}|\lambda_j-\lambda_k|^2$$ and in the r.h.s performs a linear shift $y_i=s(\lambda_i-x)$ (this is the crucial technical step), to get: $$\label{cummint}
\int_0^\infty dt\Theta_n(x,t)e^{-st}=\frac{e^{-sNx}}{s^{N^2}}C_{n,0}^{(2)}\underbrace{\int_{[0,\infty]^n}d y_1\cdots d y_n e^{-\sum_{i=1}^n y_i }
\prod_{j<k}|y_j-y_k|^2}_{\propto\mathcal{Z}_{\rm{WL}}}$$ Note that, thanks to the linear shift, the dependence on the Laplace variable $s$ has been entirely transferred outside the $n$-fold integral: this is now proportional to the partition function $\mathcal{Z}_{\rm{WL}}$ of an associated Wishart-Laguerre (WL) ensemble of random covariance matrices [@Mehta; @Wishart] of the form $\mathcal{W}=\mathcal{X}^\dagger\mathcal{X}$, where $\mathcal{X}$ is a Gaussian rectangular matrix with real or complex entries. The joint distribution of the $n$ nonnegative eigenvalues of $\mathcal{W}$ is known [@James] \^[([WL]{})]{}(\_1,…,\_n) =\_[n,]{}\^[()]{} e\^[-\_[i=1]{}\^n \_i]{} \_[i=1]{}\^n \_i\^ \_[j<k]{} |\_j-\_k|\^ \[wishart1\] where $\mathcal{N}_{n,\alpha}^{(\beta)}$ is a known normalization constant. Therefore, the jpd of Schmidt eigenvalues can be seen as a [*fixed-trace*]{} (microcanonical) version of the Wishart-Laguerre (canonical) ensemble[^8].
Now, in the case of the cumulative distribution of the *largest* eigenvalue $Q_n(x)$, the integrals on the r.h.s. run over $[0,x]$ instead of $[x,\infty]$, making the aforementioned linear shift less useful. One could keep pursuing the Laplace-transform route (introducing an auxiliary function $Q_n(x,t)$) with the change of variables $s\lambda_i=y_i$ obtaining: $$\int_0^\infty dt Q_n(x,t)e^{-st}=\frac{C_{n,0}^{(2)}}{ s^{n^2}}\underbrace{\int_{[0,sx]^n}d y_1\cdots d y_n e^{-\sum_{i=1}^n y_i }
\prod_{j<k}|y_j-y_k|^2}_{\propto E_{\rm{WL}}[(sx,\infty)]}$$ but the integral on the r.h.s. does not permit this time a friendly Laplace-inversion (see however next appendix). Indeed, this integral is readily recognized as proportional to $E_{\rm{WL}}[(sx,\infty)]$, where $E_{\rm{WL}}[(a,b)]$ is the *gap probability* for a Wishart-Laguerre ensemble, i.e. the probability that the interval $(a,b)$ on the real axis is free of eigenvalues. This quantity is exactly known in terms of Painlevé V [@TW], with the consequence that an explicit Laplace inversion formula is not available to date.
In summary, the asymmetry in the treatment of the smallest and largest Schmidt eigenvalues arises because the linear shift that works in the former case fails in the latter, and this fact calls for the alternative approach developed in this paper.
A third application of main identity for $\beta=2$ and $N=M$ {#appC}
============================================================
In this appendix we discuss in more detail a third application of the main identity eq. . In the case $\beta=2$ and $M=N$, the Hankel determinant representation in eq. actually allows a more systematic (and more easily automatized) treatment of the cumulative distribution of the largest Schmidt eigenvalue. Following the Laplace transform route outlined in the previous appendix, one can easily write down the following equation: \_0\^dt Q\_n(x,t)e\^[-st/x]{}=C\_[n,0]{}\^[(2)]{} x\^[n\^2]{}\_[n! \_[j,k=0,…, n-1]{}]{} where we have used the Hankel determinant representation for the $n$-fold integral on the r.h.s. provided in [@Kanz] (fully equivalent to ), with $\mathcal{F}_1(s)=(1-e^{-s})/s$.
A change of variable $t=x\tau$ on the l.h.s. followed by a (formal) Laplace inversion, leads to the final formula for $Q_n(x)$ in this case[^9]: Q\_n(x)=C\_[n,0]{}\^[(2)]{} n! x\^[n\^2-1]{}\^[-1]{}(1/x)\[invlap\] where $\mathcal{L}^{-1}[f(s)](t)$ is the inverse Laplace transform of $f(s)$ with parameter $t$. The r.h.s. of can be systematically evaluated in $^{\textregistered}$ and its derivative ($p_n(x)$) has been plotted in fig. \[densityapp\] for $n=N=M=3,4$.
![Probability density $p_n(x)=\frac{d}{dx}Q_n(x)$ of the largest Schmidt eigenvalue via eq. for $n=N=M=3$ (blue) and $n=N=M=4$ (violet) and $\beta=2$. []{data-label="densityapp"}](pnbeta2.eps){width=".7\hsize"}
Also the average of the largest eigenvalue $\langle\lambda_{\rm{max}}\rangle=\int_{1/n}^1 dx\ x\ p_n(x)$ can be evaluated exactly in $^{\textregistered}$ for a given $n$, and we provide a few evaluations in the following table. Unlike the average of the smallest eigenvalue, which in the same circumstances ($\beta=2$ and $M=N$) has the attractively simple form $\langle\lambda_{\rm{min}}\rangle = 1/N^3$ exactly for all $N$ [@majbohi], the situation for the largest eigenvalue seems more complicated and it is much harder to even conjecture a possible formula for the sought average valid for all $N=M$. We leave this as a challenging open problem, noticing [*en passant*]{} that the asymptotic value $\langle\lambda_{\rm{max}}\rangle \sim 4/N$ derived in [@majnadal] is approached [*very slowly*]{} as $N$ increases [@vivodens]. This fact makes the investigation of [*finite*]{} (small) $N$ results very much called for.
$N=M$ $\langle\lambda_{\rm{max}}\rangle$ (exact) $\langle\lambda_{\rm{max}}\rangle$ (approx.)
------- ----------------------------------------------------------------------------------- ----------------------------------------------
$2$ $\frac{7}{8}$ $0.875$
$3$ $\frac{313}{432}$ $0.724537$
$4$ $\frac{1 367 807}{2 239 488}$ $0.610768$
$5$ $\frac{4 581 882 694 877}{8 707 129 344 000}$ $0.526222$
$6$ $\frac{225 128 892 964 655 720 357 665 283}{487 487 792 008 396 800 000 000 000}$ $0.461814$
[99]{}
M. A. Nielsen and I. L. Chuang, [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge, 2000).
A. Peres, [*Quantum Theory: Concepts and Methods*]{} (Kluwer Academic Publishers, Dordrecht, 1993).
S. Hill and W. K. Wootters, Entanglement of a pair of quantum bits, [*Phys. Rev. Lett.*]{} [**78**]{}, 5022 (1997); W. K. Wootters, Entanglement of formation of an arbitrary state of two qubits, [*Phys. Rev. Lett.*]{} [**80**]{}, 2245 (1998).
V. Cappellini, H.-J. Sommers, and K. Życzkowski, Distribution of $G$ concurrence of random pure states, [*Phys. Rev. A*]{} [**74**]{}, 062322 (2006).
G. Gour, Family of concurrence monotones and its applications, [*Phys. Rev. A*]{} [**71**]{}, 012318 (2005).
S. N. Majumdar, Extreme eigenvalues of Wishart matrices: application to entangled bipartite system, [*Preprint*]{} \[arXiv:1005.4515\] (2010).
S. Lloyd and H. Pagels, Complexity as thermodynamic depth, [*Annals of Physics*]{} [**188**]{}, 186 (1988); see also E. Lubkin, Entropy of an $n$-system from its correlation with a $k$-reservoir, [*J. Math. Phys.*]{} [**19**]{}, 1028 (1978) and M.J.W. Hall, Random quantum correlations and density operator distributions, [*Phys. Lett. A*]{} [**242**]{}, 123 (1998).
K. Życzkowski and H.-J. Sommers, Induced measures in the space of mixed quantum states, [*J. Phys. A: Math. Gen.*]{} [**34**]{}, 7111 (2001).
D. N. Page, Average entropy of a subsystem, [*Phys. Rev. Lett.*]{} [**71**]{}, 1291 (1995).
J. N. Bandyopadhyay and A. Lakshminarayan, Testing statistical bounds on entanglement using quantum chaos, [*Phys. Rev. Lett.*]{} [**89**]{}, 060402 (2002).
S. K. Foong and S. Kanno, Proof of Page’s conjecture on the average entropy of a subsystem, [*Phys. Rev. Lett.*]{} [**72**]{}, 1148 (1994); J. Sánchez-Ruiz, Simple proof of Page’s conjecture on the average entropy of a subsystem, [*Phys. Rev. E*]{} [**52**]{}, 5653 (1995); S. Sen, Average entropy of a quantum subsystem, [*Phys. Rev. Lett.*]{} [**77**]{}, 1 (1996).
L. C. Malacarne, R. S. Mendes, and E. K. Lenzi, Average entropy of a subsystem from its average Tsallis entropy, [*Phys. Rev. E*]{} [**65**]{}, 046131 (2002).
H.-J. Sommers and K. Życzkowski, Statistical properties of random density matrices, [*J. Phys. A: Math. Theor.*]{} [**37**]{}, 8457 (2004).
H. Kubotani, S. Adachi, and M. Toda, Exact formula of the distribution of Schmidt eigenvalues for dynamical formation of entanglement in quantum chaos, [*Phys. Rev. Lett.*]{} [**100**]{}, 240501 (2008); S. Adachi, M. Toda, and H. Kubotani, Random matrix theory of singular values of rectangular complex matrices I: Exact formula of one-body distribution function in fixed-trace ensemble, [*Annals of Physics*]{} [**324**]{}, 2278 (2009).
P. Vivo, Entangled random pure states with orthogonal symmetry: exact results, [*J. Phys. A: Math. Theor.*]{} [**43**]{}, 405206 (2010).
D.-Z. Liu and D.-S. Zhou, Local statistical properties of Schmidt eigenvalues of bipartite entanglement for a random pure state, [*International Mathematics Research Notices*]{}, doi:10.1093/imrn/rnq091 (2010).
K. Życzkowski and H.-J. Sommers, Average fidelity between random quantum states, [*Phys. Rev. A*]{} [**71**]{}, 032313 (2005).
O. Giraud, Purity distribution for bipartite random pure states, [*J. Phys. A: Math. Theor.*]{} [**40**]{}, F1053 (2007); Distribution of bipartite entanglement for random pure states, [*J. Phys. A: Math. Theor.*]{} [**40**]{}, 2793 (2007).
P. Facchi, U. Marzolino, G. Parisi, S. Pascazio, and A. Scardicchio, Phase transitions of bipartite entanglement, [*Phys. Rev. Lett.*]{} [**101**]{}, 050502 (2008); A. De Pasquale, P. Facchi, G. Parisi, S. Pascazio, and A. Scardicchio, Phase transitions and metastability in the distribution of the bipartite entanglement of a large quantum system, [*Phys. Rev. A*]{} [**81**]{}, 052324 (2010).
C. Nadal, S. N. Majumdar, and M. Vergassola, Phase transitions in the distribution of bipartite entanglement of a random pure state, [*Phys. Rev. Lett.*]{} [**104**]{}, 110501 (2010); Statistical distribution of quantum entanglement for a random bipartite state, [*Preprint*]{} \[arXiv:1006.4091\] (2010).
J.W. Demmel, The probability that a numerical analysis problem is difficult, [*Math. Comput.*]{} [**50**]{}, 449 (1988).
A. Edelman, On the distribution of a scaled condition number, [*Math. Comput.*]{} [**58**]{}, 185 (1992).
S. N. Majumdar, O. Bohigas, and A. Lakshminarayan, Exact minimum eigenvalue distribution of an entangled random pure state, [*J. Stat. Phys.*]{} [**131**]{}, 33 (2008).
M. Znidaric, Entanglement of random vectors, [*J. Phys. A: Math. Theor.*]{} [**40**]{}, F105 (2007).
Y. Chen, D.-Z. Liu, and D.-S. Zhou, Smallest eigenvalue distribution of the fixed-trace Laguerre beta-ensemble, [*J. Phys. A: Math. Theor.*]{} [**43**]{}, 315303 (2010).
B. Nadler, On the distribution of the ratio of the largest eigenvalue to the trace of a Wishart matrix, J. Multivar. Anal. in press (2010).
C. W. J. Beenakker, Random-matrix theory of quantum transport, [*Rev. Mod. Phys.*]{} [**69**]{}, 731 (1997).
A. M. Chang, H. U. Baranger, L. N. Pfeiffer, and K.W. West, Weak Localization in chaotic versus nonchaotic cavities: a striking difference in the line shape, [*Phys. Rev. Lett.*]{} [**73**]{}, 2111 (1994).
C. M. Marcus, A. J. Rimberg, R. M. Westervelt, P. F. Hopkins, and A. C. Gossard, Conductance fluctuations and chaotic scattering in ballistic microstructures, [*Phys. Rev. Lett.*]{} [**69**]{}, 506 (1992).
S. Oberholzer, E. V. Sukhorukov, C. Strunk, C. Schönenberger, T. Heinzel, and M. Holland, Shot noise by quantum scattering in chaotic cavities, [*Phys. Rev. Lett.*]{} [**86**]{}, 2114 (2001).
R. Landauer, Spatial variation of currents and fields due to localized scatterers in metallic conduction, [*IBM J. Res. Dev.*]{} [**1**]{}, 223 (1957); Electrical resistance of disordered one-dimensional lattices, [*Phil. Mag.*]{} [**21**]{}, 863 (1970); D. S. Fisher and P. A. Lee, Relation between conductivity and transmission matrix, [*Phys. Rev. B*]{} [**23**]{}, 6851 (1981).
M. Büttiker, Four-terminal phase-coherent conductance, [*Phys. Rev. Lett.*]{} [**57**]{}, 1761 (1986).
Ya. M. Blanter and M. Büttiker, Shot noise in mesoscopic conductors, [*Phys. Rep.*]{} [**336**]{}, 1 (2000).
K. A. Muttalib, J. L. Pichard, and A. D. Stone, Random-Matrix Theory and universal statistics for disordered quantum conductors, [*Phys. Rev. Lett.*]{} [**59**]{}, 2475 (1987).
A. D. Stone, P. A. Mello, K. A. Muttalib, and J. L. Pichard, in [*Mesoscopic Phenomena in Solids*]{}, edited by B. L. Altshuler, P. A. Lee and R. A. Webb (North Holland, Amsterdam, 1991).
P. A. Mello, P. Pereyra, and N. Kumar, Macroscopic approach to multichannel disordered conductors, [*Annals of Physics*]{} [**181**]{}, 290 (1988).
M. L. Mehta, [*Random Matrices*]{}, 3rd Edition (Elsevier-Academic Press, 2004).
F. J. Dyson, Statistical theory of the energy levels of complex systems, [*J. Math. Phys.*]{} [**3**]{}, 140 (1962); [**3**]{}, 157 (1962); [**3**]{}, 166 (1962).
P. J. Forrester, Quantum conductance problems and the Jacobi ensemble, [*J. Phys. A: Math. Gen.*]{} [**39**]{}, 6861 (2006).
P. W. Brouwer and C. W. J. Beenakker, Diagrammatic method of integration over the unitary group, with applications to quantum transport in mesoscopic systems, [*J. Math. Phys.*]{} [**37**]{}, 4904 (1996).
M. Novaes, Full counting statistics of chaotic cavities with many open channels, [*Phys. Rev. B*]{} [**75**]{}, 073304 (2007); Statistics of quantum transport in chaotic cavities with broken time-reversal symmetry, [*Phys. Rev. B*]{} [**78**]{}, 035337 (2008).
D. V. Savin and H.-J. Sommers, Shot noise in chaotic cavities with an arbitrary number of open channels, [*Phys. Rev. B*]{} [**73**]{}, 081307(R) (2006).
D. V. Savin, H.-J. Sommers, and W. Wieczorek, Nonlinear statistics of quantum transport in chaotic cavities, [*Phys. Rev. B*]{} [**77**]{}, 125332 (2008); J.-G. Luque and P. Vivo, Nonlinear Random Matrix statistics, symmetric functions and hyperdeterminants, [*J. Phys. A: Math. Theor.*]{} [**43**]{}, 085213 (2010).
P. Vivo and E. Vivo, Transmission eigenvalue densities and moments in chaotic cavities from random matrix theory, [*J. Phys. A: Math. Theor.*]{} [**41**]{}, 122004 (2008).
S. Hemmady, J. Hart, X. Zheng, T. M. Antonsen, E. Ott, and S. M. Anlage, Experimental test of universal conductance fluctuations by means of wave-chaotic microwave cavities, [*Phys. Rev. B*]{} [**74**]{}, 195326 (2006).
H. U. Baranger and P. A. Mello, Mesoscopic transport through chaotic cavities: a random S-matrix theory approach, [*Phys. Rev. Lett.*]{} [**73**]{}, 142 (1994).
R. A. Jalabert, J.-L. Pichard, and C. W. J. Beenakker, Universal quantum signatures of chaos in ballistic transport, [*Europhys. Lett.*]{} [**27**]{}, 255 (1994).
A. García-Martín and J. J. Sáenz, Universal conductance distributions in the crossover between diffusive and localization regimes, [*Phys. Rev. Lett.*]{} [**87**]{}, 116603 (2001).
K. A. Muttalib and P. Wölfle, ’One-sided’ log-normal distribution of conductances for a disordered quantum wire, [*Phys. Rev. Lett.*]{} [**83**]{}, 3013 (1999); A. Cresti, R. Farchioni, and G. Grosso, Conductance distributions at the metal-insulator crossover in quasi 1-D pseudorandom wires, [*Eur. Phys. J. B*]{} [**46**]{}, 133 (2005); K. A. Muttalib, P. Wölfle, A. García-Martín, and V. A. Gopar, Nonanalyticity in the distribution of conductances in quasi-one-dimensional wires, [*Europhys. Lett.*]{} [**61**]{}, 95 (2003); L. S. Froufe-Pérez, P. García-Mochales, P. A. Serena, P. A. Mello, and J. J. Sáenz, Conductance distributions in quasi-one-dimensional disordered wires, [*Phys. Rev. Lett.*]{} [**89**]{}, 246403 (2002); V. A. Gopar, K. A. Muttalib, and P. Wölfle, Conductance distribution in disordered quantum wires: crossover between the metallic and insulating regimes, [*Phys. Rev. B*]{} [**66**]{}, 174204 (2002).
K. A. Muttalib, P. Markoš, and P. Wölfle, Conductance distribution in strongly disordered mesoscopic systems in three dimensions, [*Phys. Rev. B*]{} [**72**]{}, 125317 (2005); P. Markoš, Dimension dependence of the conductance distribution in the nonmetallic regimes, [*Phys. Rev. B*]{} [**65**]{}, 104207 (2002).
M. H. Pedersen, S. A. van Langen, and M. Büttiker, Charge fluctuations in quantum point contacts and chaotic cavities in the presence of transport, [*Phys. Rev. B*]{} [**57**]{}, 1838 (1998).
H.-J. Sommers, W. Wieczorek, and D. V. Savin, Statistics of conductance and shot-noise power for chaotic cavities, [*Acta Phys. Pol. A*]{} [**112**]{}, 691 (2007).
B. A. Khoruzhenko, D. V. Savin, and H.-J. Sommers, Systematic approach to statistics of conductance and shot-noise in chaotic cavities, [*Phys. Rev. B*]{} [**80**]{}, 125301 (2009).
V. Al. Osipov and E. Kanzieper, Integrable theory of quantum transport in chaotic cavities, [*Phys. Rev. Lett.*]{} [**101**]{}, 176804 (2008); Statistics of thermal to shot noise crossover in chaotic cavities, [*J. Phys. A: Math. Theor.*]{} [**42**]{}, 475101 (2009).
P. Vivo, S. N. Majumdar, and O. Bohigas, Distributions of conductance and shot noise and associated phase transitions, [*Phys. Rev. Lett.*]{} [**101**]{}, 216809 (2008); Probability distributions of linear statistics in chaotic cavities and associated phase transitions, [*Phys. Rev. B*]{} [**81**]{}, 104202 (2010).
S. Kumar and A. Pandey, Conductance distributions in chaotic mesoscopic cavities, [*J. Phys. A: Math. Theor.*]{} [**43**]{}, 285101 (2010).
I. Bengtsson and K. Życzkowski, [*Geometry of Quantum States*]{}, (Cambridge Univ. Press, New York, 2006).
J. Wishart, The generalised product moment distribution in samples from a normal multivariate population, [*Biometrika*]{} [**20A**]{}, 32 (1928).
A. T. James, Distribution of matrix variates and latent roots derived from normal samples, [*Ann. Math. Stat.*]{} [**35**]{}, 475 (1964).
G. Akemann, G. M. Cicuta, L. Molinari, and G. Vernizzi, Compact support probability distributions in random matrix theory, [*Phys. Rev. E*]{} [**59**]{}, 1489 (1999); Non-universality of compact support probability distributions in random matrix theory, [*Phys. Rev. E*]{} [**60**]{}, 5287 (1999).
A. Lakshminarayan, S. Tomsovic, O. Bohigas, and S. N. Majumdar, Extreme statistics of complex random and quantum chaotic states, [*Phys. Rev. Lett.*]{} [**100**]{}, 044103 (2008).
C. A. Tracy and H. Widom, Fredholm determinants, differential equations and matrix models, [*Commun. Math. Phys.*]{} [**163**]{}, 33 (1994); P. J. Forrester and N. S. Witte, Application of the $\tau$-function theory of Painlevé equations to random matrices: PV, PIII, the LUE, JUE, and CUE, [*Commun. Pure Appl. Math.*]{} [**55**]{}, 679 (2002).
[^1]: One has the following normalizations $\int_{1/n}^1 dx\ p_n(x)=\int_0^n dy\ \mathcal{P}_G(y)=1$.
[^2]: The bound $x\geq 1/n$ will be discussed in detail later on.
[^3]: These two cases in turn correspond to quantum systems whose Hamiltonians preserve ($\beta=1$) or break ($\beta=2$) time-reversal symmetry.
[^4]: For $\beta=2$, $\langle\lambda_{\rm{min}}\rangle=1/N^3$ exactly for all $N=M$, while for $\beta=1$ $\langle\lambda_{\rm{min}}\rangle\sim c_1/N^3$ for large $N=M$, where the constant $c_1$ is precisely known [@majbohi].
[^5]: This means that both the smallest and the largest eigenvalue distributions have compact supports and justifies the bound $x\geq 1/n$ in eq. .
[^6]: ($\mathbf{t},\mathbf{t}^\prime$) are respectively of size $N\times M$ and $M\times N$, while $(\mathbf{r},\mathbf{r}^\prime)$ are of size $M\times M$ and $N\times N$.
[^7]: A careful asymptotic analysis of formulae in [@Kanz] leads to the same findings. In particular, there are no long exponential tails in the distributions, as originally claimed in [@Kanz].
[^8]: Note that the presence of a fixed-trace constraint has crucial consequences on the spectral properties of random matrix ensembles [@akemann; @ASOS].
[^9]: One could have started directly from eq. and would have been led to the very same eq. upon noticing that $\mathcal{P}_G(1/x)=(n!/K_{n,0}^{(2)})\mathcal{L}^{-1}\left[\det[(-\partial_s)^{j+k}\mathcal{F}_1(s)]_{j,k=0,\ldots, n-1}\right](1/x)$ [@Kanz].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The paper aims to adopt the complex quaternion and octonion to formulate the field equations for electromagnetic and gravitational fields. Applying the octonionic representation enables one single definition to combine some physics contents of two fields, which were considered to be independent of each other in the past. J. C. Maxwell applied simultaneously the vector terminology and the quaternion analysis to depict the electromagnetic theory. This method edified the paper to introduce the quaternion and octonion spaces into the field theory, in order to describe the physical feature of electromagnetic and gravitational fields, while their coordinates are able to be the complex number. The octonion space can be separated into two subspaces, the quaternion space and the S-quaternion space. In the quaternion space, it is able to infer the field potential, field strength, field source, field equations, and so forth, in the gravitational field. In the S-quaternion space, it is able to deduce the field potential, field strength, field source, and so forth, in the electromagnetic field. The results reveal that the quaternion space is appropriate to describe the gravitational features; meanwhile the S-quaternion space is proper to depict the electromagnetic features.'
author:
- 'Zi-Hua Weng'
title: Field Equations in the Complex Quaternion Spaces
---
\[sec:level1\]INTRODUCTION
==========================
J. C. Maxwell described the physical feature of electromagnetic field with the vector as well as the quaternion. In 1843 W. R. Hamilton invented the quaternion, and J. T. Graves invented the octonion. Two years later, A. Cayley reinvented the octonion independently in 1845. During two decades from that time on, the scientists and engineers separated the quaternion into the scalar part and vector part, to facilitate its application in the engineering. In 1873 Maxwell mingled naturally the quaternion analysis and vector terminology to depict the electromagnetic feature in his works. Recently some scholars begin to study the physics feature of gravitational field with the algebra of quaternions.
The ordered couple of quaternions compose the octonion. On the contrary, the octonion can be separated into two parts. This case is similar to the complex number, which can be separated into the real number and the imaginary number. The octonion is able to be divided into two parts as well, the quaternion and the $S$-quaternion (short for the second quaternion), and their coordinates are able to be complex numbers. For the convenience of description, the quaternion described in the following context sometimes includes not only the quaternion but also the $S$-quaternion.
Since the vector terminology can describe the electromagnetic and gravitational theories, the quaternion should be able to depict these two theories also. In the paper, the quaternion space is suitable to describe the gravitational features, and the $S$-quaternion space is proper to depict the electromagnetic features. And it may be one approach to figure out some puzzles in the electromagnetic and gravitational theories described with the vector.
In recent years applying the quaternion to study the electromagnetic feature has been becoming one significant research orientation, and it continues the development trend of gradual deepening and expanding. Increasingly, the focus is being placed on the depiction discrepancy of electromagnetic features between the quaternion and vector. The related research into the discrepancy is coming out all the time.
Some scholars have been applying the quaternion analysis to study the electromagnetic and gravitational theories, trying to promote the further progress of these two field theories. V. Majernik [@majernik1] described the electromagnetic theory with the complex quaternion. W. M. Honig [@honig] and A. Singh [@singh] applied respectively the complex quaternion to deduce directly the Maxwell’s equations in the classical electromagnetic theory. S. Demir [@demir1] etc studied the electromagnetic theory with the hyperbolic quaternion. K. Morita [@morita] researched the quaternion field theory. S. M. Grusky [@grusky] etc utilized the quaternion to investigate the time-dependent electromagnetic field. H. T. Anastassiu [@anastassiu] etc applied the quaternion to describe the electromagnetic feature. J. G. Winans [@winans] described the physics quantities with the quaternion. Meanwhile the complex quaternion has certainly piqued scholars’ interest in researching the gravitational theory. J. Edmonds [@edmonds] utilized the quaternion to depict the wave equation and gravitational theory in the curved space-time. F. A. Doria [@doria] adopted the quaternion to research the gravitational theory. A. S. Rawat [@rawat] etc discussed the gravitational field equation with the quaternion treatment. Moreover a few scholars applied the octonion analysis to study the electromagnetic and gravitational theories. M. Gogberashvili [@gogberashvili] discussed the electromagnetic theory with the octonion. V. L. Mironov [@mironov] etc applied the octonion analysis to represent the Maxwell’s equations and relevant physics features. S. Demir [@demir2] etc applied the algebra of octonions to study the gravitational field equations. In the paper, we shall focus on the application of the complex quaternion on the electromagnetic and gravitational theory field.
At present there are mainly three description methods for the electromagnetic theory. (1) Vector analysis. It is ripe enough the electromagnetic theory described with the three-dimensional vector. However the theory has a questionable logicality, because the deduction of Maxwell’s equations is dealt with the participation of the current continuity equation. Logically, the electromagnetic strength determines the electromagnetic source via the Maxwell’s equations, and then the electromagnetic source concludes the current continuity equation. On the contrary, the electromagnetic source decides the displacement current in the Maxwell’s equations via the current continuity equation. This description method results in the problem of logic circulation. The puzzle urges other scholars to look for the new description method. (2) Real quaternion analysis. In the electromagnetic theory described with the real quaternion, it is able to deduce directly the electromagnetic field equations. But its displacement current’s direction and the gauge condition of field potential are respectively different to that in the classical electromagnetic theory. The limited progress arouses the scholar’s enthusiasm to reapply the complex quaternion to depict the electromagnetic theory. (3) Complex quaternion analysis. In the electromagnetic theory described with the complex quaternion, it is able to deduce directly the Maxwell’s equations in the classical electromagnetic theory, without the help of the current continuity equation. Similarly some scholars have been applying the complex quaternion to research the gravitational theory. A comparison of the two description methods indicates that there are quite a number of sameness between the field theory described by the complex quaternion and the classical field theory described by the vector terminology, except for a little discrepancy.
In the existing studies dealt with the quaternion field theory up to now, the most of researches introduce the quaternion to rephrase simplistically the physics concept and deduction in the classical field theory. They considered the quaternion as one ordinary substitution for the complex number or the vector in the theoretical applications. This is a far cry from the expectation that an entirely new method can bring some new conclusions. In the past, the application of one new mathematic description method each time usually enlarged the range of definition of some physics concepts, bringing in new perspectives and inferences. Obviously the existing quaternion studies in the field theory so far have not achieved the expected outcome.
Making use of the comparison and analysis, it is found a few primal problems from the preceding studies. (1) The preceding studies were not able to describe simultaneously the electromagnetic and gravitational fields. These existing researches divide the electromagnetic and gravitational fields into two isolated parts, and then describe respectively the two fields. In the paper, the electromagnetic field and gravitational field can combine together to become one unitary field in the theoretical description, depicting the physics features of two fields simultaneously. (2) The preceding studies were difficult to unify similar physics quantities of two fields into the single definition. In the existing researches, the field potential (or field strength, field source) of gravitational field is different to that of electromagnetic field, and it is not able to unify them into the single definition. But the paper is able to unify the field potential (or field strength, field source) of electromagnetic and gravitational fields into the single definition.
In the paper, the author explores one new description method, introducing the complex quaternion space into the field theory, to describe the physical feature of electromagnetic and gravitational fields. This method’s inferences cover the most of conclusions of electromagnetic and gravitational fields described with the vector.
\[sec:level1\]Field Equations
=============================
The octonion space $\mathbb{O}$ can be separated into two orthogonal subspaces, the quaternion space $\mathbb{H}_g$ and the $S$-quaternion space $\mathbb{H}_e$ . And the quaternion space $\mathbb{H}_g$ is independent of the $S$-quaternion space $\mathbb{H}_e$ . The quaternion space $\mathbb{H}_g$ is suitable to describe the feature of gravitational field, while the $S$-quaternion space $\mathbb{H}_e$ is proper to depict the property of electromagnetic field.
In the quaternion space $\mathbb{H}_g$ for the gravitational field, the basis vector is $\mathbb{H}_g = ( \emph{\textbf{i}}_0, \emph{\textbf{i}}_1, \emph{\textbf{i}}_2, \emph{\textbf{i}}_3 )$, the radius vector is $\mathbb{R}_g = \emph{i} r_0 \emph{\textbf{i}}_0 + \Sigma r_k \emph{\textbf{i}}_k$, and the velocity is $\mathbb{V}_g = \emph{i} v_0 \emph{\textbf{i}}_0 + \Sigma v_k \emph{\textbf{i}}_k$. The gravitational potential is $\mathbb{A}_g = \emph{i} a_0 \emph{\textbf{i}}_0 + \Sigma a_k \emph{\textbf{i}}_k$, the gravitational strength is $\mathbb{F}_g = f_0 \emph{\textbf{i}}_0 + \Sigma f_k \emph{\textbf{i}}_k$, and the gravitational source is $\mathbb{S}_g = \emph{i} s_0 \emph{\textbf{i}}_0 + \Sigma s_k \emph{\textbf{i}}_k$. In the $\emph{S}$-quaternion space $\mathbb{H}_e$ for the electromagnetic field, the basis vector is $\mathbb{H}_e = ( \emph{\textbf{I}}_0, \emph{\textbf{I}}_1, \emph{\textbf{I}}_2, \emph{\textbf{I}}_3 )$, the radius vector is $\mathbb{R}_e = \emph{i} R_0 \emph{\textbf{I}}_0 + \Sigma R_k \emph{\textbf{I}}_k$, and the velocity is $\mathbb{V}_e = \emph{i} V_0 \emph{\textbf{I}}_0 + \Sigma V_k \emph{\textbf{I}}_k$. The electromagnetic potential is $\mathbb{A}_e = \emph{i} A_0 \emph{\textbf{I}}_0 + \Sigma A_k \emph{\textbf{I}}_k$, the electromagnetic strength is $\mathbb{F}_e = F_0 \emph{\textbf{I}}_0 + \Sigma F_k \emph{\textbf{I}}_k$, and the electromagnetic source is $\mathbb{S}_e = i S_0 \emph{\textbf{I}}_0 + \Sigma S_k \emph{\textbf{I}}_k$. Herein $\mathbb{H}_e = \mathbb{H}_g \circ \emph{\textbf{I}}_0$. The symbol $\circ$ denotes the octonion multiplication. $r_j, ~v_j, ~a_j, ~s_j, ~R_j, ~V_j, ~A_j, ~S_j, ~f_0$, and $F_0$ are all real. $f_k$ and $F_k$ are the complex numbers. $i$ is the imaginary unit. $\emph{\textbf{i}}_0 = 1$. $j = 0, ~1, ~2, ~3$. $k = 1, ~2, ~3$.
These two orthogonal quaternion spaces, $\mathbb{H}_g$ and $\mathbb{H}_e$, composes one octonion space, $\mathbb{O} = \mathbb{H}_g + \mathbb{H}_e$. In the octonion space $\mathbb{O}$ for the electromagnetic and gravitational fields, the octonion radius vector is $\mathbb{R} = \mathbb{R}_g + k_{eg} \mathbb{R}_e$, the octonion velocity is $\mathbb{V} = \mathbb{V}_g + k_{eg} \mathbb{V}_e$ , with $k_{eg}$ being one coefficient. Meanwhile the octonion field potential is $\mathbb{A} = \mathbb{A}_g + k_{eg} \mathbb{A}_e$, the octonion field strength is $\mathbb{F} = \mathbb{F}_g + k_{eg} \mathbb{F}_e$. From here on out, some physics quantities are extended from the quaternion functions to the octonion functions, according to the characteristics of octonion. Apparently $\mathbb{V}$, $\mathbb{A}$, and their differential coefficients are all octonion functions of $\mathbb{R}$.
The octonion definition of field strength is, $$\mathbb{F} = \square \circ \mathbb{A} ~,$$ where $\mathbb{F}_g = \square \circ \mathbb{A}_g$, $\mathbb{F}_e = \square \circ \mathbb{A}_e$. The gauge conditions of field potential are chosen as $f_0 = 0$ and $F_0 = 0$. The operator is $ \square = \emph{i} \emph{\textbf{i}}_0 \partial_0 + \Sigma \emph{\textbf{i}}_k \partial_k$. $ \nabla = \Sigma \emph{\textbf{i}}_k \partial_k$, $ \partial_j = \partial / \partial r_j$. $v_0$ is the speed of light.
The octonion field source $\mathbb{S}$ of the electromagnetic and gravitational fields can be defined as, $$\begin{aligned}
\mu \mathbb{S} && = - ( \emph{i} \mathbb{F} / v_0 + \square )^* \circ \mathbb{F}
\nonumber \\
&& = \mu_g \mathbb{S}_g + k_{eg} \mu_e \mathbb{S}_e - ( \emph{i} \mathbb{F} / v_0 )^* \circ \mathbb{F} ~,\end{aligned}$$ where $\mu$, $\mu_g$, and $\mu_e$ are coefficients. $\mu_g < 0$, and $\mu_e > 0$. $*$ denotes the conjugation of octonion. In general, the contribution of the tiny term, $( \emph{i} \mathbb{F}^* \circ \mathbb{F}/ v_0 )$, in the above could be neglected.
According to the coefficient $k_{eg}$ and the basis vectors, $\mathbb{H}_g$ and $\mathbb{H}_e$, the octonion definition of the field source can be separated into two parts, $$\begin{aligned}
&& \mu_g \mathbb{S}_g = - \square^* \circ \mathbb{F}_g ~,
\\
&& \mu_e \mathbb{S}_e = - \square^* \circ \mathbb{F}_e ~,\end{aligned}$$ where Eq.(3) is the definition of gravitational source, while Eq.(4) is the definition of electromagnetic source.
The octonion velocity $\mathbb{V}$ is defined as, $$\mathbb{V} = v_0 \partial_0 \mathbb{R} ~,$$ where in the case for single one particle, a comparison with the classical electromagnetic (and gravitational) theory reveals that, $\mathbb{S}_g = m \mathbb{V}_g$, and $\mathbb{S}_e = q \mathbb{V}_e$. $m$ is the mass density, while $q$ is the density of electric charge.
$ $ $1$ $\emph{\textbf{i}}_1$ $\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_3$ $\emph{\textbf{I}}_0$ $\emph{\textbf{I}}_1$ $\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_3$
----------------------- ----------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------
$1$ $1$ $\emph{\textbf{i}}_1$ $\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_3$ $\emph{\textbf{I}}_0$ $\emph{\textbf{I}}_1$ $\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_3$
$\emph{\textbf{i}}_1$ $\emph{\textbf{i}}_1$ $-1$ $\emph{\textbf{i}}_3$ $-\emph{\textbf{i}}_2$ $\emph{\textbf{I}}_1$ $-\emph{\textbf{I}}_0$ $-\emph{\textbf{I}}_3$ $\emph{\textbf{I}}_2$
$\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_2$ $-\emph{\textbf{i}}_3$ $-1$ $\emph{\textbf{i}}_1$ $\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_3$ $-\emph{\textbf{I}}_0$ $-\emph{\textbf{I}}_1$
$\emph{\textbf{i}}_3$ $\emph{\textbf{i}}_3$ $\emph{\textbf{i}}_2$ $-\emph{\textbf{i}}_1$ $-1$ $\emph{\textbf{I}}_3$ $-\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_1$ $-\emph{\textbf{I}}_0$
$\emph{\textbf{I}}_0$ $\emph{\textbf{I}}_0$ $-\emph{\textbf{I}}_1$ $-\emph{\textbf{I}}_2$ $-\emph{\textbf{I}}_3$ $-1$ $\emph{\textbf{i}}_1$ $\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_3$
$\emph{\textbf{I}}_1$ $\emph{\textbf{I}}_1$ $\emph{\textbf{I}}_0$ $-\emph{\textbf{I}}_3$ $\emph{\textbf{I}}_2$ $-\emph{\textbf{i}}_1$ $-1$ $-\emph{\textbf{i}}_3$ $\emph{\textbf{i}}_2$
$\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_3$ $\emph{\textbf{I}}_0$ $-\emph{\textbf{I}}_1$ $-\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_3$ $-1$ $-\emph{\textbf{i}}_1$
$\emph{\textbf{I}}_3$ $\emph{\textbf{I}}_3$ $-\emph{\textbf{I}}_2$ $\emph{\textbf{I}}_1$ $\emph{\textbf{I}}_0$ $-\emph{\textbf{i}}_3$ $-\emph{\textbf{i}}_2$ $\emph{\textbf{i}}_1$ $-1$
: The multiplication table of octonion.[]{data-label="tab:table3"}
\[sec:level2\]Gravitational field equations
-------------------------------------------
In the quaternion space $\mathbb{H}_g = ( 1, \emph{\textbf{i}}_1, \emph{\textbf{i}}_2, \emph{\textbf{i}}_3 )$, the gravitational potential is $\mathbb{A}_g = \emph{i} a_0 + \textbf{a}$, the operator is $ \square = \emph{i} \partial_0 + \nabla $, the gravitational strength is $\mathbb{F}_g = \square \circ \mathbb{A}_g$. The definition of gravitational strength can be expressed as, $$\begin{aligned}
\mathbb{F}_g = ( - \partial_0 a_0 + \nabla \cdot \textbf{a} ) + \emph{i} ( \partial_0 \textbf{a} + \nabla a_0 ) + \nabla \times \textbf{a} ~,\end{aligned}$$ where $\textbf{a} = \Sigma a_k \emph{\textbf{i}}_k$; $a_0 = \varphi / v_0$, with $\varphi$ being the scalar potential of gravitational field.
For the sake of convenience the paper adopts the gauge condition, $f_0 = - \partial_0 a_0 + \nabla \cdot \textbf{a} = 0$. Further the above can be written as $\textbf{f} = \emph{i} \textbf{g} / v_0 + \textbf{b}$, with $\textbf{f} = \Sigma f_k \emph{\textbf{i}}_k$. One component of gravitational strength is the gravitational acceleration, $\textbf{g} / v_0 = \partial_0 \textbf{a} + \nabla a_0$. The other is $\textbf{b} = \nabla \times \textbf{a}$, which is similar to the magnetic flux density.
In the definition of gravitational source Eq.(3), the gravitational source is, $\mathbb{S}_g = \emph{i} s_0 + \textbf{s}$. Substituting the gravitational source $\mathbb{S}_g$ and gravitational strength $\mathbb{F}_g$ into the definition of gravitational source can express the definition as, $$- \mu_g ( \emph{i} s_0 + \textbf{s} ) = ( \emph{i} \partial_0 + \nabla )^* \circ ( \emph{i} \textbf{g} / v_0 + \textbf{b} ) ~,$$ where the gravitational coefficient is $\mu_g = 1 / ( \varepsilon_g v_0^2)$, and $\varepsilon_g = - 1 / (4 \pi G)$, with $G$ being the gravitational constant. The density of linear momentum is, $\textbf{s} = \Sigma s_k \emph{\textbf{i}}_k$.
The above can be rewritten as, $$\begin{aligned}
- \mu_g ( \emph{i} s_0 + \textbf{s} ) = \emph{i} ( \partial_0 \textbf{b} + \nabla^* \times \textbf{g} / v_0 ) + \emph{i} ( \nabla^* \cdot \textbf{g} / v_0 )
+ ( \nabla^* \times \textbf{b} - \partial_0 \textbf{g} / v_0 ) + \nabla^* \cdot \textbf{b} ~.\end{aligned}$$
Comparing both sides of the equal sign in the above will yield, $$\begin{aligned}
&& \nabla^* \cdot \textbf{b} = 0 ~,
\\
&& \partial_0 \textbf{b} + \nabla^* \times \textbf{g} / v_0 = 0 ~,
\\
&& \nabla^* \cdot \textbf{g} / v_0 = - \mu_g s_0 ~,
\\
&& \nabla^* \times \textbf{b} - \partial_0 \textbf{g} / v_0 = - \mu_g \textbf{s} ~,\end{aligned}$$ where $s_0 = m v_0$. $\textbf{g} = \Sigma g_k \emph{\textbf{i}}_k$, $\textbf{b} = \Sigma b_k \emph{\textbf{i}}_k$.
Eqs.(9)-(12) are the gravitational field equations. Because the gravitational constant $G$ is weak and the velocity ratio $\textbf{v} / v_0$ is tiny, the gravity produced by the linear momentum $\textbf{s}$ can be ignored in general. When $\textbf{b} = 0$ and $\textbf{a} = 0$, Eq.(11) can be degenerated into the Newton’s law of universal gravitation in the classical gravitational theory.
By all appearances the quaternion operator $\square$, gravitational potential $\mathbb{A}_g$, gravitational source $\mathbb{S}_g$, radius vector $\mathbb{R}_g$, and velocity $\mathbb{V}_g$ etc are restricted by the component selection in the definitions. That is, the coordinates with the basis vector $\emph{\textbf{i}}_0$ are the imaginary numbers, while the coordinates with the basis vector $\emph{\textbf{i}}_k$ are all real. In the paper they can be written as the preceding description form in the context, in order to deduce the gravitational field equations. Those restriction conditions are the indispensable components of the classical gravitational theory.
The deduction approach of the gravitational field equations can be used as a reference to be extended to that of electromagnetic field equations.
------------------------------------------------------------------------------------------------------------------------------
definition expression meaning
----------------------------- ------------------------------------------------------------------------------------------------
$\nabla \cdot \textbf{a}$ $-(\partial_1 a_1 + \partial_2 a_2 + \partial_3 a_3)$
$\nabla \times \textbf{a}$ $\emph{\textbf{i}}_1 ( \partial_2 a_3
- \partial_3 a_2 ) + \emph{\textbf{i}}_2 ( \partial_3 a_1
- \partial_1 a_3 ) + \emph{\textbf{i}}_3 ( \partial_1 a_2
- \partial_2 a_1 )$
$\nabla a_0$ $\emph{\textbf{i}}_1 \partial_1 a_0
+ \emph{\textbf{i}}_2 \partial_2 a_0
+ \emph{\textbf{i}}_3 \partial_3 a_0 $
$\partial_0 \textbf{a}$ $\emph{\textbf{i}}_1 \partial_0 a_1
+ \emph{\textbf{i}}_2 \partial_0 a_2
+ \emph{\textbf{i}}_3 \partial_0 a_3 $
$\nabla \cdot \textbf{A}$ $-(\partial_1 A_1 + \partial_2 A_2 + \partial_3 A_3) \emph{\textbf{I}}_0 $
$\nabla \times \textbf{A}$ $-\emph{\textbf{I}}_1 ( \partial_2
A_3 - \partial_3 A_2 ) - \emph{\textbf{I}}_2 ( \partial_3 A_1
- \partial_1 A_3 ) - \emph{\textbf{I}}_3 ( \partial_1 A_2
- \partial_2 A_1 )$
$\nabla \circ \textbf{A}_0$ $\emph{\textbf{I}}_1 \partial_1 A_0
+ \emph{\textbf{I}}_2 \partial_2 A_0
+ \emph{\textbf{I}}_3 \partial_3 A_0 $
$\partial_0 \textbf{A}$ $\emph{\textbf{I}}_1 \partial_0 A_1
+ \emph{\textbf{I}}_2 \partial_0 A_2
+ \emph{\textbf{I}}_3 \partial_0 A_3 $
------------------------------------------------------------------------------------------------------------------------------
: The multiplication of the operator and octonion physics quantity.[]{data-label="tab:table2"}
\[sec:level2\]Electromagnetic field equations
---------------------------------------------
In the electromagnetic theory described with the complex quaternion, it is able to deduce directly the Maxwell’s equations in the classical electromagnetic theory. In this approach, substituting the $S$-quaternion for the quaternion, one can obtain the same conclusions still.
In the $\emph{S}$-quaternion space $\mathbb{H}_e = ( \emph{\textbf{I}}_0, \emph{\textbf{I}}_1, \emph{\textbf{I}}_2, \emph{\textbf{I}}_3 )$, the electromagnetic potential is $\mathbb{A}_e = \emph{i} \textbf{A}_0 + \textbf{A}$, the electromagnetic strength is $\mathbb{F}_e = \square \circ \mathbb{A}_e$. The definition of electromagnetic strength can be expressed as $$\begin{aligned}
\mathbb{F}_e = ( - \partial_0 \textbf{A}_0 + \nabla \cdot \textbf{A} ) + \emph{i} ( \partial_0 \textbf{A} + \nabla \circ \textbf{A}_0 ) + \nabla \times \textbf{A} ~,\end{aligned}$$ where the vector potential of electromagnetic field is $\textbf{A} = \Sigma A_k \emph{\textbf{I}}_k$. $\textbf{A}_0 = A_0 \emph{\textbf{I}}_0$; $A_0 = \phi / v_0$, with $\phi$ being the scalar potential of electromagnetic field.
For the sake of convenience the paper adopts the gauge condition, $\textbf{F}_0 = - \partial_0 \textbf{A}_0 + \nabla \cdot \textbf{A} = 0$. Therefore the above can be written as $\textbf{F} = \emph{i} \textbf{E} / v_0 + \textbf{B}$. Herein the electric field intensity is $\textbf{E} / v_0 = \partial_0 \textbf{A} + \nabla \circ \textbf{A}_0$, and the magnetic flux density is $\textbf{B} = \nabla \times \textbf{A}$. $\textbf{F}_0 = F_0 \emph{\textbf{I}}_0$. $\textbf{F} = \Sigma F_k \emph{\textbf{I}}_k$.
In the definition of electromagnetic source Eq.(4), the electromagnetic source is, $\mathbb{S}_e = \emph{i} \textbf{S}_0 + \textbf{S}$. Substituting the electromagnetic source $\mathbb{S}_e$ and electromagnetic strength $\mathbb{F}_e$ into the definition of electromagnetic source can separate the definition as, $$- \mu_e ( \emph{i} \textbf{S}_0 + \textbf{S} ) = ( \emph{i} \partial_0 + \nabla )^* \circ ( \emph{i} \textbf{E} / v_0 + \textbf{B} ) ~,$$ where the electromagnetic coefficient is $\mu_e = 1 / ( \varepsilon_e v_0^2)$, with $\varepsilon_e$ being the permittivity.
The above can be rewritten as, $$\begin{aligned}
- \mu_e ( \emph{i} \textbf{S}_0 + \textbf{S} ) = \emph{i} ( \partial_0 \textbf{B} + \nabla^* \times \textbf{E} / v_0 ) + \emph{i} ( \nabla^* \cdot \textbf{E} / v_0 )
+ ( \nabla^* \times \textbf{B} - \partial_0 \textbf{E} / v_0 ) + \nabla^* \cdot \textbf{B} ~.\end{aligned}$$
Comparing the variables on both sides of the equal sign in the above will reason out the Maxwell’s equations as follows, $$\begin{aligned}
&& \nabla^* \cdot \textbf{B} = 0 ~,
\\
&& \partial_0 \textbf{B} + \nabla^* \times \textbf{E} / v_0 = 0 ~,
\\
&& \nabla^* \cdot \textbf{E} / v_0 = - \mu_e \textbf{S}_0 ~,
\\
&& \nabla^* \times \textbf{B} - \partial_0 \textbf{E} / v_0 = - \mu_e \textbf{S} ~,\end{aligned}$$ where $\textbf{S}_0 = S_0 \emph{\textbf{I}}_0$, $S_0 = q V_0$. $\textbf{E} = \Sigma E_k \emph{\textbf{I}}_k$, $\textbf{B} = \Sigma B_k \emph{\textbf{I}}_k$. The density of electric current is $\textbf{S} = \Sigma S_k \emph{\textbf{I}}_k$. For the charged particles, there may be, $\mathbb{V}_e = \mathbb{V}_g \circ \textbf{I} ( \emph{\textbf{I}}_j )$. And the unit $\textbf{I} ( \emph{\textbf{I}}_j )$ is one function of $\emph{\textbf{I}}_j$, with $\textbf{I} ( \emph{\textbf{I}}_j )^* \circ \textbf{I} ( \emph{\textbf{I}}_j ) = 1$.
In a similar way the electromagnetic potential $\mathbb{A}_e$, electromagnetic source $\mathbb{S}_e$, radius vector $\mathbb{R}_e$ , and velocity $\mathbb{V}_e$ etc are restricted by the component selection of the definitions. That is, the coordinates with the basis vector $\emph{\textbf{I}}_0$ are the imaginary numbers, while the coordinates with the basis vector $\emph{\textbf{I}}_k$ are all real. In the paper they can be written as the preceding description form in the context, in order to deduce the Maxwell’s equations. Those restriction conditions are the indispensable components of the classical electromagnetic theory.
On the analogy of the coordinate definition of complex coordinate system, one can define the coordinate of octonion, which involves the quaternion and $\emph{S}$-quaternion simultaneously. In the octonion coordinate system, the octonion physics quantity can be defined as $ \{ ( \emph{i} c_0 + \emph{i} d_0 \textbf{\emph{I}}_0 ) \circ \emph{\textbf{i}}_0 + \Sigma ( c_k + d_k \textbf{\emph{I}}_0^* ) \circ \emph{\textbf{i}}_k \} $. It means that there are the quaternion coordinate $c_k$ and the $S$-quaternion coordinate $d_k \emph{\textbf{I}}_0^*$ for the basis vector $\emph{\textbf{i}}_k$, while the quaternion coordinate $c_0$ and the $S$-quaternion coordinate $d_0 \emph{\textbf{I}}_0$ for the basis vector $\emph{\textbf{i}}_0$. Herein $c_j$ and $d_j$ are all real.
\[sec:level1\]Equivalent Transformation
=======================================
\[sec:level2\]Gravitational Field Equations
-------------------------------------------
Making use of the transforming of the basis vectors of $\nabla$ in the quaternion operator $\square$, it is able to translate further the gravitational field equations, Eqs.(9)-(12), from the quaternion space $\mathbb{H}_g$ into that in the three-dimensional vector space. In the three-dimensional vector space $(\emph{\textbf{j}}_1 ,~\emph{\textbf{j}}_2 ,~\emph{\textbf{j}}_3)$, the operator is $\blacktriangledown = \Sigma \emph{\textbf{j}}_k \partial_k$, with $\emph{\textbf{j}}_k^{~2} = 1$.
In the gravitational field equations, the operator $\nabla$ should be substituted by the operator $\blacktriangledown$. Meanwhile $\textbf{b}$, $\textbf{g}$, and $\textbf{s}$ are substituted by $\textbf{b}^\prime = \Sigma b_k \emph{\textbf{j}}_k$, $\textbf{g}^\prime = \Sigma g_k \emph{\textbf{j}}_k$, and $\textbf{s}^\prime = \Sigma s_k \emph{\textbf{j}}_k$ respectively. The cross and dot products of quaternion are substituted by that of vector respectively. And the gravitational field equations, Eqs.(9)-(12), can be transformed into, $$\begin{aligned}
&& - \blacktriangledown \cdot \textbf{b}^\prime = 0 ~,
\\
&& \partial_0 \textbf{b}^\prime - \blacktriangledown \times \textbf{g}^\prime / v_0 = 0 ~,
\\
&& \blacktriangledown \cdot \textbf{g}^\prime / v_0 = - \mu_g s_0 ~,
\\
&& \blacktriangledown \times \textbf{b}^\prime + \partial_0 \textbf{g}^\prime / v_0 = \mu_g \textbf{s}^\prime ~,\end{aligned}$$ where $\textbf{b}^\prime = \blacktriangledown \times \textbf{a}^\prime$, $\textbf{g}^\prime / v_0 = \partial_0 \textbf{a}^\prime + \blacktriangledown a_0$. $\textbf{a}^\prime = \Sigma a_k \emph{\textbf{j}}_k$.
Expressing the above into the scalar equations will reveal that Eqs.(9)-(12) and Eqs.(20)-(23) are equivalent to each other. And the definition of Eq.(6) requires to substitute $\textbf{g}^\prime$ by $\textbf{g}^{\prime\prime} = - \textbf{g}^\prime$, in order to approximate to the Newton’s law of universal gravitation in the classical gravitational theory as near as possible. Therefore the gravitational field equations, Eqs.(20)-(23), are able to be transformed into that in the three-dimensional vector space, $$\begin{aligned}
&& \blacktriangledown \cdot \textbf{b}^\prime = 0 ~,
\\
&& \partial_0 \textbf{b}^\prime + \blacktriangledown \times \textbf{g}^{\prime\prime} / v_0 = 0 ~,
\\
&& \blacktriangledown \cdot \textbf{g}^{\prime\prime} / v_0 = \mu_g s_0 ~,
\\
&& \blacktriangledown \times \textbf{b}^\prime - \partial_0 \textbf{g}^{\prime\prime} / v_0 = \mu_g \textbf{s}^\prime ~.\end{aligned}$$
Comparing Eq.(26) with Poisson equation for the Newton’s law of universal gravitation reveals that those two equations are the same formally. In case the gravitational strength component $\textbf{b}$ is weak and the velocity ratio $\textbf{v} / v_0$ is quite tiny, Eq.(26) will be reduced into the Newton’s law of universal gravitation in the classical gravitational theory.
\[sec:level2\]Electromagnetic field equations
---------------------------------------------
By means of the transformation of basis vectors of $\nabla$ in the quaternion operator, it is able to translate further the electromagnetic field equations, Eqs.(16)-(19), from the $\emph{S}$-quaternion space $\mathbb{H}_e$ into that in the three-dimensional vector space $(\emph{\textbf{j}}_1 ,~\emph{\textbf{j}}_2 ,~\emph{\textbf{j}}_3)$. In the electromagnetic field equations, the operator $\nabla$ is substituted by $\blacktriangledown$. And $\textbf{B}$, $\textbf{E}$, and $\textbf{S}$ are substituted by $\textbf{B}^\prime$, $\textbf{E}^\prime$, and $\textbf{S}^\prime$ respectively. Therefore the electromagnetic field equations, Eqs.(16)-(19), can be transformed into, $$\begin{aligned}
&& \blacktriangledown \cdot \textbf{B}^\prime = 0 ~,
\\
&& \partial_0 \textbf{B}^\prime + \blacktriangledown \times \textbf{E}^\prime / v_0 = 0 ~,
\\
&& \blacktriangledown \cdot \textbf{E}^\prime / v_0 = - \mu_e S_0 ~,
\\
&& \blacktriangledown \times \textbf{B}^\prime - \partial_0 \textbf{E}^\prime / v_0 = - \mu_e \textbf{S}^\prime ~,\end{aligned}$$ where $\textbf{B}^\prime = - \blacktriangledown \times \textbf{A}^\prime$, $\textbf{E}^\prime / v_0 = \partial_0 \textbf{A}^\prime + \blacktriangledown A_0$. $\textbf{A}^\prime = \Sigma A_k \emph{\textbf{j}}_k$, $\textbf{S}^\prime = \Sigma S_k \emph{\textbf{j}}_k$.
Expressing the above into the scalar equations will reveal that Eqs.(16)-(19) and Eqs.(28)-(31) are equivalent to each other. And the definition of Eq.(13) requires to substitute $\textbf{E}^\prime$ and $\textbf{B}^\prime$ by $\textbf{E}^{\prime\prime} = - \textbf{E}^\prime$ and $\textbf{B}^{\prime\prime} = - \textbf{B}^\prime¡ä$ respectively, in order to approximate to the Maxwell’s equations in the classical electromagnetic theory as near as possible. Therefore the electromagnetic field equations, Eqs.(28)-(31), are able to be transformed into that in the three-dimensional vector space, $$\begin{aligned}
&& \blacktriangledown \cdot \textbf{B}^{\prime\prime} = 0 ~,
\\
&& \partial_0 \textbf{B}^{\prime\prime} + \blacktriangledown \times \textbf{E}^{\prime\prime} / v_0 = 0 ~,
\\
&& \blacktriangledown \cdot \textbf{E}^{\prime\prime} / v_0 = \mu_e S_0 ~,
\\
&& \blacktriangledown \times \textbf{B}^{\prime\prime} - \partial_0 \textbf{E}^{\prime\prime} / v_0 = \mu_e \textbf{S}^\prime ~.\end{aligned}$$
Comparing Eqs.(32)-(35) with the Maxwell’s equations in the classical electromagnetic theory states that those two field equations are equivalent to each other. It means that in the electromagnetic theory described with the quaternion, without the participation of the current continuity equation, it is still able to reason out the Maxwell’s equations in the classical electromagnetic theory.
\[sec:level1\]Conclusions and Discussions
=========================================
Applying the complex quaternion space, the paper is able to describe simultaneously the electromagnetic field equations and the gravitational field equations. The spaces of gravitational and electromagnetic fields both can be chosen as the quaternion spaces. Furthermore some coordinates of those two quaternion spaces and relevant physics quantities may be the imaginary numbers.
In the definitions of the quaternion operator, field potential, field source, radius vector, and velocity etc, the coordinates with the basis vector $\emph{\textbf{I}}_0$ or $\emph{\textbf{i}}_0$ are all imaginary numbers. However this simple case is still able to result in many components of other physics quantities to become the imaginary numbers even the complex numbers. In the definitions, only by means of the confinement of the component selection can we achieve the quaternion field theory which approximating to the classical field theory, enabling it to cover the classical field theory. Therefore it is able to be said that the component selection are the essential ingredients for the field theory too.
In the quaternion space, it is able to deduce the gravitational field equations as well as the gauge condition of gravitational potential etc. The Newton’s law of universal gravitation in the classical gravitational theory can be derived from the gravitational field equations described with the quaternion. Two components of the gravitational strength are the counterpart of the linear acceleration and of the precession angular velocity. In general, the component of the gravitational strength, which is corresponding to the precession angular velocity of the gyroscopic torque dealt with the angular momentum, is comparatively tiny.
In the $S$-quaternion space, it is able to reason out directly the electromagnetic field equations as well as the gauge condition of electromagnetic potential etc. The electromagnetic field equations described with the $S$-quaternion can be equivalently translated into the Maxwell’s equations of the classical electromagnetic theory in the three-dimensional vector space. Moreover the definition of electromagnetic strength and the gauge condition of electromagnetic potential both can be equivalently translated respectively into that in the classical electromagnetic theory.
It should be noted that the paper discussed only the electromagnetic field equations and the gravitational field equations etc described with the complex quaternion, and the equivalent transformation of those field equations into that in the three-dimensional vector space. However it clearly states that the introducing of the complex quaternion space is able to availably describe the physics features of electromagnetic and gravitational fields. This will afford the theoretical basis for the further relevant theoretical analysis, and is helpful to apply the complex quaternion analysis to study the angular momentum, electric/magnetic dipole moment, torque, energy, and force etc in the electromagnetic and gravitational fields.
The author is indebted to the anonymous referee for their valuable and constructive comments on the previous manuscript. This project was supported partially by the National Natural Science Foundation of China under grant number 60677039.
V. Majernik, “Quaternionic formulation of the classical fields", [*Advances in Applied Clifford Algebras*]{}, vol. 9, no. 1, pp.119–130, 1999.
W. M. Honig, “Quaternionic electromagnetic wave equation and a dual charge-filled space", [*Lettere al Nuovo Cimento*]{}, vol. 19, no. 4, pp.137–140, 1977.
A. Singh, “Quaternionic form of the electromagnetic-current equations with magnetic monopoles", [*Lettere al Nuovo Cimento*]{}, vol. 31, no. 5, pp.145–148, 1981.
S. Demir, M. Tanisli, and N. Candemir, “Hyperbolic Quaternion Formulation of Electromagnetism", [*Advances in Applied Clifford Algebras*]{}, vol. 20, no. 3-4, pp.547–563, 2010.
K. Morita, “Quaternions, Lorentz group and the Dirac theory", [*Progress of Theoretical Physics*]{}, vol. 117, no. 3, pp.501–532, 2007.
S. M. Grusky, K. V. Khmelnytskaya, and V. V. Kravchenko, “On a quaternionic Maxwell equation for the time-dependent electromagnetic field in a chiral medium", [*Journal of Physics A*]{}, vol. 37, no. 16, pp.4641–4647, 2004.
H. T. Anastassiu, P. E. Atlamazoglou, and D. I. Kaklamani, “Application of bicomplex (quaternion) algebra to fundamental electromagnetics: A lower order alternative to the Helmholtz equation", [*IEEE Transactions on Antennas and Propagation*]{}, vol. 51, no. 8, pp.2130–2136, 2003.
J. G. Winans, “Quaternion physical quantities", [*Foundations of Physics*]{}, vol. 7, no. 5-6, pp.341–349, 1977.
J. D. Edmonds, “Quaternion wave equations in curved space-time", [*International Journal of Theoretical Physics*]{}, vol. 10, no. 2, pp.115–122, 1974.
F. A. Doria, “A Weyl-like equation for the gravitational field", [*Lettere al Nuovo Cimento*]{}, vol. 14, no. 13, pp.480–482, 1975.
A. S. Rawat and O. P. S. Negi, “Quaternion Gravi-Electromagnetism", [*International Journal of Theoretical Physics*]{}, vol. 51, no. 3, pp.738–745, 2012.
M. Gogberashvili, “Octonionic Electrodynamics", [*Journal of Physics A*]{}, vol. 39, no. 22, pp.7099–7104, 2006.
V. L. Mironov and S. V. Mironov, “Octonic representation of electromagnetic field equations", [*Journal of Mathematical Physics*]{}, vol. 50, no. 1, Article ID 012901, 10 pages, 2009.
M. Tanisli, M. E. Kansu, and S. Demir, “Reformulation Electromagnetic and Gravito-electromagnetic Equations for Lorentz System with Octonion Algebra", [*General Relativity and Gravitation*]{}, vol. 46, no. 5, Article ID 1739, 21 pages, 2014.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We describe Concurrent C0, a type-safe C-like language with contracts and session-typed communication over channels. Concurrent C0 supports an operation called forwarding which allows channels to be combined in a well-defined way. The language’s type system enables elegant expression of session types and message-passing concurrent programs. We provide a Go-based implementation with language based optimizations that outperforms traditional message passing techniques.'
author:
- Max Willsey
- Rokhini Prabhu
- 'Frank Pfenning [^1]'
bibliography:
- 'sources.bib'
date: 'Last updated: '
title: 'Design and Implementation of Concurrent C0 [^2]'
---
\[\]
\#1 {#section .unnumbered}
===
[^1]: This work is partially funded by the FCT (Portuguese Foundation for Science and Technology) through the Carnegie Mellon Portugal Program.
[^2]: An extended version of this paper can be found at <http://maxwillsey.com/papers/cc0-thesis.pdf>
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Dominique LECOMTE
date: '* Topology Appl. 154,1 (2007), 241-262*'
title: 'On minimal non-potentially closed subsets of the plane.'
---
[ We study the Borel subsets of the plane that can be made closed by refining the Polish topology on the real line. These sets are called potentially closed. We first compare Borel subsets of the plane using products of continuous functions. We show the existence of a perfect antichain made of minimal sets among non-potentially closed sets. We apply this result to graphs, quasi-orders and partial orders. We also give a non-potentially closed set minimum for another notion of comparison. Finally, we show that we cannot have injectivity in the Kechris-Solecki-Todorčevi' c dichotomy about analytic graphs.]{}
**[1 Introduction.]{}**
The reader should see \[K\] for the descriptive set theoretic notation used in this paper. This work is the continuation of a study made in \[L1\]-\[L4\]. The usual way of comparing Borel equivalence relations $E\subseteq
X\times X$ and ${E'\subseteq X'\times X'}$ on Polish spaces is the Borel reducibility quasi-order: $$E\leq_B E'~\Leftrightarrow ~\exists u\!:\! X\!\rightarrow\!
X'~\hbox{\rm Borel~with}~E\! =\! (u\!\times\! u)^{-1}(E')$$ (recall that a quasi-order is a reflexive and transitive relation). Note that this makes sense even if $E$, $E'$ are not equivalence relations. It is known that if $(B_n)$ is a sequence of Borel subsets of $X$, then we can find a finer Polish topology on $X$ making the $B_n$’s clopen (see Exercise 13.5 in \[K\]). So assume that $E\leq_B E'$ and let $\sigma$ be a finer Polish topology on $X$ making $u$ continuous. If $E'$ is in some Baire class $\Gamma$, then $E\in\Gamma ([X,\sigma ]^2)$. This motivates the following (see \[Lo2\]):
(Louveau) Let $X,Y$ be Polish spaces, $A$ a Borel subset of $X \times Y$, and $\Gamma$ a Baire class. We say that $A$ is $potentially~in~\Gamma$ $\big($denoted $A\in \mbox{pot}(\Gamma)\big)$ iff we can find a finer Polish topology $\sigma$ $($resp., $\tau )$ on $X$ $($resp., $Y)$ such that $A\in\Gamma ([X, \sigma ]\times [Y, \tau ])$.
This notion is a natural invariant for $\leq_B$: if $E'\in \mbox{pot}(\Gamma)$ and $E\leq_B E'$, then $E\in \mbox{pot}(\Gamma)$. Using this notion, A. Louveau showed that the collection of ${{\bf\Sigma}^{0}_{\xi}}$ equivalence relations is not cofinal for $\leq_B$, and deduces from this the inexistence of a maximum Borel equivalence relation for $\leq_B$.
A. Louveau has also more recently noticed that one can associate a quasi-order $R_{A}\subseteq (X\times 2)^2$ to $A\subseteq X^2$ as follows: $$(x,i)~R_{A}~(y,j)~~\Leftrightarrow ~~(x,i)=(y,j)~~\mbox{or}~~
[(x,y)\in A~~\mbox{and}~~(i,j)=(0,1)].$$ Note that $R_{A}$ is also antisymmetric, so that it is actually a partial order.
A. Louveau noticed the following facts, using the following notion of comparison between Borel subsets $A\!\subseteq\! X\times Y$, ${A'\!\subseteq\! X'\!\!\times\! Y'}$ of products of two Polish spaces: $$A\sqsubseteq^r_{B} A'~\Leftrightarrow ~\exists u\!:\! X\!\rightarrow\! X'\
\exists v\!:\! Y\!\rightarrow\! Y'~\hbox{\rm one-to-one\ Borel~with}~A\! =\!
(u\!\times\! v)^{-1}(A').$$ Here the letter $r$ means “rectangle" ($u$ and $v$ may be different).
- Assume that $A\subseteq X^2$ has full projections, and that $A'\subseteq (X')^2$. Then $A\sqsubseteq^r_{B}A'$ is equivalent to $R_{A}\leq_B R_{A'}$.
- If $A\subseteq X^{2}$ is $\sqsubseteq^r_{B}$-minimal among non-potentially closed sets, then $R_{A}$ is $\leq_{B}$-minimal among non-potentially closed partial orders.
- Conversely, if $R_{A}$ is $\leq_{B}$-minimal among non-potentially closed partial orders and if $A$ has full projections, then $A$ is $\sqsubseteq^r_{B}$-minimal among non-potentially closed sets.
These facts show that, from the point of view of Borel reducibility, the study of Borel partial orders is essentially the study of arbitrary Borel subsets of the plane. This strengthens the motivation for studying arbitrary Borel subsets of the plane, from the point of view of potential complexity.
$\bullet$ A standard way to see that a set is complicated is to notice that it is more complicated than a well-known example. For instance, we have the following result (see \[SR\]):
(Hurewicz) Let $P_{f}:=\{\alpha\!\in\!
2^\omega /\exists n\!\in\!\omega ~\forall m\!\geq\! n~~\alpha (m)\!
=\! 0\}$, $X$ be a Polish space and $A$ a Borel subset of $X$. Then exactly one of the following holds:
\(a) The set $A$ is ${{\bf\Pi}^{0}_{2}}(X)$.
\(b) There is a continuous injection $u:2^\omega\rightarrow X$ such that $P_{f}=u^{-1}(A)$.
This theorem has been generalized to all Baire classes in \[Lo-SR\]. We try to adapt this result to the Borel subsets of the plane. In this direction, we have the following result for equivalence relations (see \[H-K-Lo\]):
(Harrington-Kechris-Louveau) Let $X$ be a Polish space, $E$ a Borel equivalence relation on $X$, and $E_{0}\!:=\!\{(\alpha ,\beta )\!\in\! 2^\omega\times 2^\omega
/\exists n\!\in\!\omega ~\forall m\!\geq\! n~~\alpha (m)\! =\!\beta
(m)\}$. Then exactly one of the following holds:
\(a) The relation $E$ is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
\(b) We have $E_{0}\leq_B E$ (with $u$ continuous and one-to-one).
$\bullet$ We will study structures other than equivalence relations (for example quasi-orders), and even arbitrary Borel subsets of the plane. We need some other notions of comparison.
Recall that Wadge’s quasi-order $\leq_{W}$ on Borel subsets of $\omega^\omega$ is defined by $$A\leq_W A'~~\Leftrightarrow ~~\exists
u\!:\!\omega^\omega\!\rightarrow\!
\omega^\omega ~\hbox{\rm continuous~with}~A\! =\! u^{-1}(A').$$ It is known that this quasi-order is well-founded (in the sense that there is no sequence $(B_{n})$ with $B_{n+1}\leq_{W}B_{n}$ and $B_{n}\not\leq_{W}B_{n+1}$ for each $n$). Moreover, any $\leq_{W}$-antichain is of cardinality at most 2 (in fact of the form $\{A,\neg A\}$). It follows that any class ${{\bf\Delta}^{1}_{1}}\setminus{{\bf\Pi}^{0}_{\xi}}$ admits a unique (up to the equivalence associated to $\leq_{W}$) minimal element.
$\bullet$ There are several natural ways of comparing Borel subsets $A\!\subseteq\! X\times Y$, ${A'\!\subseteq\! X'\!\!\times\! Y'}$ of products of two Polish spaces. All of them will have the same behavior here. The one we will use is the following: $$A\leq^r_{c} A'~\Leftrightarrow ~\exists u\!:\! X\!\rightarrow\!
X'\ \exists
v\!:\! Y\!\rightarrow\! Y'~\hbox{\rm continuous~with}~A\! =\!
(u\!\times\! v)^{-1}(A').$$ Here the letter $c$ is for “continuous". We have the following (see \[L1\]):
Let $\Delta (2^\omega )\!:=\!
\{(\alpha ,\beta )\!\in\! 2^\omega\!\times 2^\omega\! /\alpha\!
=\!\beta\}$, ${L_0\!:=\!\{(\alpha ,\beta )\!\in\! 2^\omega\!\!\times\! 2^\omega\!
/\alpha\! <_{\mbox{lex}}\!\beta\}}$, $X$ and $Y$ be Polish spaces, and $A$ a $pot\big(\check D_2({{\bf\Sigma}^{0}_{1}})\big)$ subset of $X\times Y$. Then exactly one of the following holds:
\(a) The set $A$ is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
\(b) $\neg\Delta (2^\omega )\leq^r_{c} A$ or $L_{0}\leq^r_{c} A$ (with $u$ and $v$ one-to-one).
$\bullet$ Things become much more complicated at the level $D_{2}({{\bf\Sigma}^{0}_{1}})$ (differences of two open sets; $\check D_2({{\bf\Sigma}^{0}_{1}})$ is the dual Wadge class of unions of a closed set and an open set; notice that we can extend Definition 1 to the class $\check D_2({{\bf\Sigma}^{0}_{1}})$). We will show the following:
There is a perfect $\leq^r_{c}$-antichain $(A_{\alpha})_{\alpha\in 2^\omega}\subseteq D_{2}({{\bf\Sigma}^{0}_{1}})
(2^\omega\times 2^\omega )$ such that $A_{\alpha}$ is $\leq^r_{c}$-minimal among ${{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets, for any $\alpha\!\in\! 2^\omega$.
In particular, unlike for classical Baire classes and $\leq_{W}$, one cannot characterize non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets by an obstruction condition involving only one (or even countably many) set(s). We will also show that ${[D_{2}({{\bf\Sigma}^{0}_{1}})\setminus
\mbox{pot}({{\bf\Pi}^{0}_{1}}),\leq^r_{c}]}$ is ill-founded.
Theorem 5 can be applied to structures. We will show the following:
There is a perfect $\leq_{B}$-antichain $(R_{\alpha})_{\alpha\in 2^\omega}\subseteq D_{2}({{\bf\Sigma}^{0}_{1}})
\big((2^\omega\times 2)^2\big)$ such that $R_{\alpha}$ is $\leq_{B}$-minimal among ${{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets, for any $\alpha\in 2^\omega$. Moreover, $(R_{\alpha})_{\alpha\in 2^\omega}$ can be taken to be a subclass of any of the following classes:
- Directed graphs (i.e., irreflexive relations).
- Graphs (i.e., irreflexive and symmetric relations).
- Oriented graphs (i.e., irreflexive and antisymmetric relations).
- Quasi-orders.
- Strict quasi-orders (i.e., irreflexive and transitive relations).
- Partial orders.
- Strict partial orders (i.e., irreflexive, antisymmetric and transitive relations).
**Remarks. (a) Theorem 6 shows that Harrington, Kechris and Louveau’s Theorem is very specific, and that the combination of symmetry and transitivity is very strong.**
\(b) We produce concrete examples of such antichains. These examples must be in any complete family of minimal sets, up to bi-reducibility.
$\bullet$ Theorem 5 shows that any complete family of minimal sets for $[{{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}}),\leq^r_{c}]$ has size continuum. So we must find another notion of comparison. In [\[K-S-T\]]{}, the following notion is defined. Let $X$, $X'$ be Polish spaces, and $A\!\subseteq\! X\!\times\! X$, $A'\!\subseteq
X'\!\times\! X'$ be analytic sets. We set $$(X,A)\preceq_{c}(X',A')\Leftrightarrow\exists u\!:\!
X\!\rightarrow\! X'~\hbox{\rm continuous~with}~A\!\subseteq\!
(u\!\times\! u)^{-1}(A').$$ When $u$ is Borel we write $\preceq_{B}$ instead of $\preceq_{c}$.
Let $\psi:\omega\rightarrow 2^{<\omega}$ be the natural bijection ($\psi (0)=\emptyset$, $\psi (1)=0$, $\psi (2)=1$, $\psi (3)=0^2$, $\psi (4)=01$, $\psi (5)=10$, $\psi (6)=1^2$, $\ldots$). Note that $|\psi (n)|\leq n$, so that we can define $s_{n}:=\psi (n)0^{n-|\psi
(n)|}$. The crucial properties of $(s_{n})$ are that it is $dense$ (there is $n$ such that $t\prec s_{n}$, for each $t\in 2^{<\omega}$), and that $|s_{n}|=n$. We set $$A_{1}\!:=\!\{ (s_{n}0\gamma ,s_{n}1\gamma )/n\!\in\!\omega
~~\mbox{and}~\ \gamma\!\in\! 2^\omega\}.$$ The symmetric set $s(A_1)$ generated by $A_1$ is considered in \[K-S-T\], where the following is essentially proved:
(Kechris, Solecki, Todorčevi' c) Let $X$ be a Polish space and $A$ an analytic subset of $X\times X$. Then exactly one of the following holds:
\(a) $(X,A)\preceq_{B}(\omega ,\not=)$.
\(b) $(2^\omega ,A_{1})\preceq_{c}(X,A)$.
Actually, the original statement in \[K-S-T\] is when $A$ is a graph, and with $s(A_1)$ instead of $A_1$. But we can get Theorem 7 without any change in the proof in \[K-S-T\].
$\bullet$ In \[L3\] the following is shown (see Theorem 2.9):
Let $X$, $Y$ be Polish spaces, and $A$ a $\mbox{pot}({{\bf\Delta}^{0}_{3}})$ subset of $X \times Y$. Then exactly one of the following holds:
\(a) The set $A$ is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
\(b) There are $u:2^\omega\rightarrow X$, $v:2^\omega\rightarrow Y$ continuous with ${A_{1}\! =\! (u\!\times\! v)^{-1}(A)\cap\overline{A_{1}}}$.
(We can replace $A_1$ in \[L3\] by what we call $A_{1}$ here.) We generalize this result to arbitrary Borel subsets of $X\times Y$:
Let $X$, $Y$ be Polish spaces, and $A$, $B$ be disjoint analytic subsets of $X \times Y$. Then exactly one of the following holds:
\(a) The set $A$ is separable from $B$ by a $\mbox{pot}({{\bf\Pi}^{0}_{1}})$ set.
\(b) There are $u:2^\omega\rightarrow X$ and $v:2^\omega\rightarrow Y$ continuous such that the inclusions $A_{1}\subseteq (u\times v)^{-1}(A)$ and $\overline{A_{1}}\setminus A_1\subseteq (u\times v)^{-1}(B)$ hold.
Moreover, we can neither replace $\overline{A_{1}}\setminus A_1$ with $(2^\omega\times 2^\omega)\setminus A_1$, nor ensure that $u$ and $v$ are one-to-one.
So we get a minimum non-potentially closed set if we do not ask for a reduction on the whole product.
$\bullet$ In \[K-S-T\], it is conjectured that we can have $u$ one-to-one in Theorem 7.(b). This is not the case:
There is no graph $(X_{0},R_{0})$ with $X_{0}$ Polish and $R_{0}\in{{\bf\Sigma}^{1}_{1}}(X_{0}^2)$ such that for every graph $(X,A)$ of the same type, exactly one of the following holds:
(a) $(X,A)\preceq_{B}(\omega ,\not= )$.
(b) $(X_{0},R_{0})\preceq_{c,1-1}(X,A)$.
The proof is based on the counterexample used in \[L3\] to show that we cannot have injectivity in Theorem 2.9.
$\bullet$ The paper is organized as follows.
- In Section 2, we prove Theorem 9.
- In Section 3, we prove Theorem 10.
- In Section 4, we give a sufficient condition for minimality among non-potentially closed sets. We use it to prove Theorems 5 and 6.
- In Section 5, we give conditions on $A$ which allow us to replace $\overline{A_{1}}\setminus A_1$ with $(2^\omega\times
2^\omega)\setminus A_1$ in Theorem 9 (and therefore come back to $\leq^r_{c}$). We can write ${A_{1}\! =\!\bigcup_{n}~\mbox{Gr}(f_{n})}$, where ${f_{n}(s_{n}0\gamma
)\!:=\! s_{n}1\gamma}$. Roughly speaking, we require that the $f_{n}$’s do not induce cycles. This is really the key property making the $A_{\alpha}$’s appearing in the statement of Theorem 5 pairwise orthogonal. We will deduce from this the minimality of $A_1$ among non-potentially closed sets for $\leq^r_{c}$, using the sufficient condition for minimality in Section 4.
**[2 A minimum non-potentially closed set.]{}**
We will prove Theorem 9. The proof illustrates the link between the dichotomy results in \[K-S-T\] and the notion of potential Baire class. We will see another link in Section 3. The next lemma is essentially Lemma 3.5 in \[L1\], and the crucial point of its proof.
Let $X$ be a nonempty Polish space, $n$ be an integer, $D_{f_n}$ and $f_n[D_{f_n}]$ be dense $G_{\delta}$ subsets of some open subsets of $X$, and $f_{n}:D_{f_n}\rightarrow f_n[D_{f_n}]$ a continuous and open map.
\(a) Let $G$ be a dense $G_{\delta}$ subset of $X$. Then $\mbox{Gr}(f_{n})\subseteq\overline{\mbox{Gr}(f_{n})\cap G^{2}}$, for each $n$.
\(b) Let $A:=\bigcup_{n}~\mbox{Gr}(f_{n})$. If $\Delta
(X)\subseteq\overline{A}\setminus A$, then $A$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
[**Proof.**]{} (a) Let $U$ (resp., $V$) be an open neighborhood of $x\in D_{f_{n}}$ (resp., $f_{n}(x)$). Then $f_{n}[D_{f_{n}}]\cap
V\cap G$ is a dense $G_{\delta}$ subset of $f_{n}[D_{f_{n}}]\cap V$, thus $f_n^{-1}(V\cap G)$ is a dense $G_{\delta}$ subset of $f_n^{-1}(V)$. Therefore $G\cap
f_n^{-1}(V)$ and $G\cap f_n^{-1}(V\cap G)$ are dense $G_{\delta}$ subsets of $f_n^{-1}(V)$. So we can find $$y\in U\cap G\cap f_n^{-1}(V\cap G).$$ Now $\big(y,f_n(y)\big)$ is in the intersection $(U\times V)\cap \hbox{Gr}(f_n)\cap
G^2$, so this set is non-empty.
\(b) We argue by contradiction: we can find a finer Polish topology on $X$ such that $A$ becomes closed. By 15.2, 11.5 and 8.38 in \[K\], the new topology and the old one agree on a dense $G_{\delta}$ subset of $X$, say $G$: $A\cap G^2\in{{\bf\Pi}^{0}_{1}}(G^2)$. Let $x\in G$. We have $(x,x)\in G^2\cap\overline{A}\setminus A$. By (a) we get $\overline{A}\subseteq\overline{A\cap G^2}$. Thus $(x,x)\in G^2\cap\overline{A\cap G^2}\setminus (A\cap G^{2})$, which is absurd.
The set $A_{1}=\overline{A_{1}}\setminus\Delta (2^\omega )$ is $D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$, and $\overline{A_{1}}=A_{1}\cup\Delta (2^\omega )$.
[**Proof.**]{} As we saw in the introduction, we can write ${A_{1}\! =\!\bigcup_{n}~\hbox{Gr}(f_{n})}$, where ${f_{n}(s_{n}0\gamma )\!:=\! s_{n}1\gamma}$. Notice that $f_{n}$ is a partial homeomorphism with clopen domain and range. Moreover, we have $$\Delta (2^\omega )\subseteq\overline{A_{1}}\setminus A_{1}$$ (in fact, the equality holds). Indeed, if $t\in 2^{<\omega}$, we have $(s_{\psi^{-1}(t)}0^\infty ,
s_{\psi^{-1}(t)}10^\infty )\in N^2_{t}\cap A_{1}$. Thus $A_{1}=\overline{A_{1}}\setminus\Delta (2^\omega )$ is $D_{2}({{\bf\Sigma}^{0}_{1}})$, and the corollary follows from Lemma 11.
[**Proof of Theorem 9.**]{} We cannot have (a) and (b) simultaneously. For if $D$ is potentially closed and separates $A$ from $B$, then we get $A_1=(u\times v)^{-1}(D)\cap\overline{A_1}$, thus $A_{1}\in
\mbox{pot}({{\bf\Pi}^{0}_{1}})$, which contradicts Corollary 12.
$\bullet$ Let $f:\omega^\omega\rightarrow X\times Y$ be a continuous map with $f[\omega^\omega ]=B$, and $f_{0}$ (resp., $f_{1}$) be the first (resp., second) coordinate of $f$, so that ${(f_{0}\times f_{1})[\Delta (\omega^\omega )]=B}$. We set $R:=(f_{0}\times f_{1})^{-1}(A)$, which is an irreflexive analytic relation on $\omega^\omega$. By Theorem 7, either there exists a Borel map ${c:\omega^\omega\rightarrow\omega}$ such that $(\alpha ,\beta )\in R$ implies ${c(\alpha )\not= c(\beta )}$, or there is a continuous map $u_{0}:2^\omega\rightarrow\omega^\omega$ such that ${(\alpha ,\beta )\in A_{1}}$ implies ${\big(u_{0}(\alpha ),u_{0}(\beta )\big)\in R}$.
$\bullet$ In the first case, we define ${C_{n}:=c^{-1}(\{ n\})}$. We get ${\Delta (\omega^\omega )\subseteq\bigcup_{n} C_{n}^2\subseteq\neg R}$, so that $${B\subseteq\bigcup_{n} f_{0}[C_{n}]\times f_{1}[C_{n}]\subseteq\neg A}.$$ By a standard reflection argument there is a sequence $(X_n)$ (resp., $(Y_n)$) of Borel subsets of $X$ (resp., $Y$) with $$\bigcup_{n} f_{0}[C_{n}]\times f_{1}[C_{n}]\subseteq\bigcup_{n} X_{n}\times Y_n\subseteq\neg A.$$ But $\bigcup_{n}\ X_{n}\times Y_n$ is $\mbox{pot}({{\bf\Sigma}^{0}_{1}})$, so we are in the case (a).
$\bullet$ In the second case, let $u:=f_{0}\circ u_{0}$, $v:=f_{1}\circ u_{0}$. These maps satisfy the conclusion of condition (b) because $\overline{A_{1}}\setminus A_{1}\subseteq\Delta (2^\omega )$, by Corollary 12.
$\bullet$ By the results in \[L3\], we can neither replace $\overline{A_{1}}\setminus A_1$ with $(2^\omega\times
2^\omega)\setminus A_1$, nor can we ensure that $u$ and $v$ are one-to-one.
**Remarks. (a) In Theorem 9, we cannot ensure that $u=v$ when $X=Y$: take $X:=2^\omega$, $$A:=\{ (\alpha ,\beta)\in N_{0}\times N_{1}/\alpha <_{\mbox{lex}}\beta\}$$ and $B:=(N_{0}\times N_{1})\setminus A$.**
\(b) This proof cannot be generalized, in the sense that we used the fact that the range of a countable union of Borel rectangles (a $\mbox{pot}({{\bf\Sigma}^{0}_{1}})$ set) by a product function is still a countable union of rectangles, so more or less a $\mbox{pot}({{\bf\Sigma}^{0}_{1}})$ set. This fails completely for the dual level. Indeed, we saw that the range of the diagonal (which is closed) by a product function can be any analytic set. So in view of generalizations, it is better to have another proof of Theorem 9.
**[3 The non-injectivity in the Kechris-Solecki-Todorčevi' c dichotomy.]{}**
Now we will prove Theorem 10. The proof we give is not the original one, which used effective descriptive set theory, and a reflection argument. The proof we give here is due to B. D. Miller, and is a simplification of the original proof.
**Notation. If $A\subseteq X^2$, $A^{-1}:=\{(y,x)\in X^2/(x,y)\in A\}$ and $s(A):=A\cup A^{-1}$ is the symmetric set generated by $A$.**
$\bullet$ Fix sets $S_{0}\supseteq S_{1}\supseteq\ldots$ of natural numbers such that
\(1) $S_{n}\setminus S_{n+1}$ is infinite for each integer $n$.
\(2) $\bigcap_{n\in\omega}\ S_{n}=\emptyset$.
$\bullet$ For each $n\in\omega$, fix $f_{n}:S_{n}\rightarrow S_{n}\setminus S_{n+1}$ injective, and define $g_{n}:2^\omega\rightarrow 2^\omega$ by $$[g_{n}(\alpha )](k):=\left\{\!\!\!\!\!\!\!\!
\begin{array}{ll}
& \alpha [f_{n}(k)]~\hbox{\rm if}~k\in S_{n}\mbox{,}\cr\cr
& \alpha (k)~\hbox{\rm otherwise.}
\end{array}
\right.$$ $\bullet$ It is clear that each of the closed sets $M_{n}:=\{ \alpha \in 2^\omega /g_{n}(\alpha )=\alpha\}$ is meager, and since each $g_{n}$ is continuous and open, it follows that the $F_{\sigma}$ set $$M:=\bigcup_{s\in\omega^{<\omega}, n\in\omega}\ (g_{s(0)}\circ\ldots\circ
g_{s(\vert s\vert -1)})^{-1}(M_{n})$$ is also meager, so that $X:=2^\omega\setminus M$ is a comeager, dense $G_{\delta}$ set which is invariant with respect to each $g_{n}$. Put $G_{1}:=\bigcup_{n\in\omega}\ s[\hbox{Gr}({g_{n}}\vert_{X})]$.
[**Proof of Theorem 10.**]{} We argue by contradiction: this gives $(X_{0},R_{0})$.
[**Claim 1.**]{}* Let $X$ be a Polish space, and $g_{0},g_{1},\ldots :X\rightarrow X$ fixed-point free Borel functions such that $g_{m}\circ g_{n}=g_{m}$ if $m<n$. Then every locally countable Borel directed subgraph of the Borel directed graph $G:=\bigcup_{n\in\omega}\ \hbox{Gr}(g_{n})$ has countable Borel chromatic number, i.e., satisfies Condition (a) in Theorem 7.*
Suppose that $H$ is a locally countable Borel directed subgraph of $G$. By the Lusin-Novikov uniformization theorem, there are Borel partial injections $h_{n}$ on $X$ such that $H=\bigcup_{n\in\omega}\ \hbox{Gr}(h_{n})$. By replacing each $h_{n}$ with its restrictions to the sets $\{ x\in D_{h_{n}}/h_{n}(x)=g_{m}(x)\}$, for $m\in\omega$, we can assume that for all $n\in\omega$, there is $k_{n}\in\omega$ such that $h_{n}=g_{k_{n}}\vert_{D_{h_{n}}}$. It is easily seen that the directed graph associated with a Borel function has countable Borel chromatic number (see also Proposition 4.5 of \[K-S-T\]), so by replacing $h_{n}$ with its restriction to countably many Borel sets, we can assume also that for all $n\in\omega$, $D_{h_{n}}^{2}\cap\bigcup_{k\leq k_{n}}\ \hbox{Gr}(g_{k})=\emptyset$. It only remains to note that $D_{h_{n}}^{2}\cap\bigcup_{k>k_{n}}\ \hbox{Gr}(g_{k})=\emptyset$. To see this, simply observe that if $k>k_{n}$ and $x,g_{k}(x)\in D_{h_{n}}$, then $h_{n}(x)=g_{k_{n}}(x)=g_{k_{n}}\circ g_{k}(x)=h_{n}\circ g_{k}(x)$, which contradicts the fact that $h_{n}$ is a partial injection. This proves the claim.
[**Claim 2.**]{}* The Borel graph $G_{1}$ has uncountable Borel chromatic number, but if $H\subseteq G_{1}$ is a locally countable Borel directed graph, then $H$ has countable Borel chromatic number.*
Condition (1) implies that $g_{m}\circ g_{n}=g_{m}$ if $m<n$, so Claim 1 ensures that if $H\subseteq G_{1}$ is a locally countable Borel directed graph, then $H$ has countable Borel chromatic number.
To see that $G_{1}$ has uncountable Borel chromatic number, it is enough to show that if $B\in{{\bf\Delta}^{1}_{1}}(2^\omega )$ is non-meager, then $B\cap G_{1}^{2}\not=\emptyset$. Let $s\in 2^{<\omega}$ such that $B$ is comeager in $N_{s}$. It follows from condition (2) that there is $n\in\omega$ such that $\vert s\vert <k$ for each $k\in S_{n}$. Then $g_{n}$ is a continuous, open map which sends $N_{s}$ into itself, thus $B\cap X\cap N_{s}\cap g_{n}^{-1}(B\cap X\cap N_{s})$ is comeager in $N_{s}$. Letting $x$ be any element of this set, it follows that $x$, $g_{n}(x)$ are $G_{1}$-related elements of $B$.
We are now ready to prove the theorem: as $(X_{0},R_{0})$ satisfies (b), it does not satisfy (a). Therefore $R_{0}$ has uncountable Borel chromatic number. As $s(A_{1})$ and $G_{1}$ have uncountable Borel chromatic number, we get $(X_{0},R_{0})\preceq_{c,1-1}[2^\omega ,s(A_{1})]$ and $(X_{0},R_{0})\preceq_{c,1-1}(2^\omega ,G_{1})$ (with witness $\pi$). As $s(A_{1})$ is locally countable, $R_{0}$ is also locally countable. Therefore $(\pi\times\pi )[R_{0}]$ is a locally countable Borel subgraph of $G_{1}$ with uncountable Borel chromatic number, which contradicts Claim 2.
**Remark. This proof also shows a similar theorem for irreflexive analytic relations, by considering $\bigcup_{n\in\omega}\ \hbox{\rm Gr}(g_{n}\vert_{X})$ (resp., $A_{1}$) instead of $G_{1}$ (resp., $s(A_{1})$).**
**[4 Perfect antichains made of sets minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets.]{}**
As mentioned in the introduction, a great variety of very different examples appear at level $D_{2}({{\bf\Sigma}^{0}_{1}})$, all of the same type. Let us make this more specific.
We say that $\big(X,(f_{n})\big)$ is a $converging~situation$ if
\(a) $X$ is a nonempty 0-dimensional perfect Polish space.
\(b) $f_{n}$ is a partial homeomorphism with ${{\bf\Delta}^{0}_{1}}(X)$ domain and range.
\(c) The diagonal $\Delta (X)=\overline{A^f}\setminus A^f$, where $A^f:=\bigcup_{n}~\mbox{Gr}(f_{n})$.
This kind of situation plays an important role in the theory of potential complexity (see, for example, Definition 2.4 in \[L3\]).
[**Remarks.**]{} (a) Note that if $\big(X,(f_{n})\big)$ is a converging situation, then Lemma 11 ensures that $A^f$ is $D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$, since ${A^f=\overline{A^f}\setminus\Delta (X)}$.
\(b) It is clear that an analytic graph $(X,A)$ has countable Borel chromatic number if and only if $A$ is separable from $\Delta (X)$ by a $\mbox{pot}({{\bf\Delta}^{0}_{1}})$ set. By Remark (a), this implies that $\big(2^\omega ,s(A^f)\big)$ does not have countable Borel chromatic number if $\big(X,(f_{n})\big)$ is a converging situation.
**Notation. In the sequel, we set $f^{B}_{n}:={f_{n}\vert}_{B\cap f_{n}^{-1}(B)}$ if $B\subseteq X$ and $\big(X,(f_{n})\big)$ is a converging situation, so that $\hbox{Gr}(f^{B}_{n})=\hbox{Gr}(f_{n})\cap B^2$.**
The reader should see \[Mo\] for the basic notions of effective descriptive set theory. Let $Z$ be a recursively presented Polish space.
$\bullet$ The topology ${\it\Delta}_{Z}$ is the topology on $Z$ generated by ${{\it\Delta}^{1}_{1}}(Z)$. This topology is Polish (see the proof of Theorem 3.4 in \[Lo2\]).
$\bullet$ The Gandy-Harrington topology ${\it\Sigma}_{Z}$ on $Z$ is generated by ${{\it\Sigma}^{1}_{1}}(Z)$. Recall that $$\Omega_{Z}:=\{z\in Z/\omega_1^z =\omega_1^{\mbox{CK}}\}$$ is Borel and ${{\it\Sigma}^{1}_{1}}$, and $[\Omega_{Z},{\it\Sigma}_{Z}]$ is a $0$-dimensional Polish space (in fact, the intersection of $\Omega_{Z}$ with any nonempty ${{\it\Sigma}^{1}_{1}}$ set is a nonempty clopen subset of $[\Omega_{Z},{\it\Sigma}_{Z}]$-see \[L1\]).
Let $\big(X,(f_{n})\big)$ be a converging situation, $P$ a Borel subset of $X$ such that $A^f\cap P^2$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$, and $\sigma$ a finer Polish topology on $P$. Then we can find a Borel subset $S$ of $P$ and a topology $\tau$ on $S$ finer than $\sigma$ such that $\big([S,\tau],(f^{S}_{n})_{n}\big)$ is a converging situation.
[**Proof.**]{} We may assume that $[P,\sigma ]$ is recursively presented and $f^{P}_{n}$, $A^f\cap P^2$ are ${{\it\Delta}^{1}_{1}}$. We set ${D\!:=\!\{x\!\in\! P/x\!\in\! {\it\Delta }^1_1\}}$, and ${S\!:=\!\{ x\!\in\! P/(x,x)\!\in\!\overline{A^{f}\cap P^2}^
{{\it\Delta}_{P}^2}\}\cap\Omega_{P}\!\setminus\! D}$. As ${S\!\in\!{{\it\Sigma}^{1}_{1}}}$, $[S,{\it\Sigma }_{P}]$ is a 0-dimensional perfect Polish space. We set ${E:=A^{f}\cap (P\setminus D)^{2}}$. Note that $D$ is countable. By Remark 2.1 in \[L1\], $E$ is not potentially closed since $$A^{f}\cap P^{2}=[A^{f}\cap \big( (P\cap D)\times P\big)]\cup
[A^{f}\cap\big( P\times (P\cap D)\big)]\cup E.$$ Therefore ${\overline{E}^{{\it\Delta}_{P}^2}\setminus E}$ is a nonempty subset of ${(P\setminus D)^{2}\cap\overline{A^{f}}\setminus
A^{f}\subseteq\Delta (X)}$. Thus ${S\not=\emptyset}$. Note also that ${(x,x)\!\in\!\overline{A^{f}\cap P^2}^{{\it\Delta}_{P}^2}\cap S^2\!
=\!\overline{A^{f}\cap P^2}^{{\it\Sigma}_{P}^2}\cap S^2\! =\!
\overline{A^{f}\cap S^2}^{[S,{\it\Sigma}_{P} ]^2}}$ if $x\!\in\! S$. Conversely, we have ${\overline{A^{f}\cap S^2}^{[S,{\it\Sigma}_{P}]^2}
\!\setminus\!\big(A^{f}\cap S^2\big)\!\subseteq\!
S^2\cap\overline{A^{f}}\!
\setminus\! A^{f}\!\subseteq\!\Delta (S)}$. We have proved that $S$ is a Borel subset of $P$ such that $\big([S,{\it\Sigma}_{P}],(f^{S}_{n})_{n}\big)$ is a converging situation.
Let $Y$, $Y'$ be Polish spaces, $A\in{{\bf\Delta}^{1}_{1}}(Y\times Y')$, $\big(X,(f_{n})\big)$ a converging situation. We assume that $A\leq^r_{c} A^f$. Then exactly one of the following holds:
\(a) The set $A$ is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
\(b) We can find a Borel subset $B$ of $X$ and a finer topology $\tau$ on $B$ such that $\big([B,\tau],(f^{B}_{n})_{n}\big)$ is a converging situation and $A^f\cap B^{2}\leq^r_{c}A$.
[**Proof.**]{} Let $u$ and $v$ be continuous functions such that $A=(u\times v)^{-1}(A^f)$. We assume that $A$ is not potentially closed. By Theorem 9 we can find continuous maps $u':2^\omega\rightarrow Y$ and $v':2^\omega\rightarrow Y'$ such that $A_{1}=(u'\times v')^{-1}(A)\cap\overline{A_{1}}$. We set $H:=u\big[u'[2^\omega ]\big]$, $K:=v\big[v'[2^\omega ]\big]$ and $P:=H\cap K$. Then $H$, $K$ and $P$ are compact and $A^{f}\cap (H\times K)$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$ since $$A_{1}=[(u~\circ ~u')\times (v~\circ ~v')]^{-1}\big(A^{f}\cap (H\times K)\big)\cap\overline{A_{1}}$$ (we have $A_{1}\notin \mbox{pot}({{\bf\Pi}^{0}_{1}})$ by Corollary 12). Therefore $A^{f}\cap P^{2}$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$, since $$\begin{array}{ll}
A^{f}\cap (H\times K)\!\!\!\!
& = [A^{f}\cap\big( (H\setminus K)\times K\big)]\cup
[A^{f}\cap\big( H\times (K\setminus H)\big)]\cup [A^{f}\cap P^{2}]\cr
& = [\overline{A^{f}}\cap\big( (H\setminus K)\times K\big)]\cup
[\overline{A^{f}}\cap\big( H\times (K\setminus H)\big)]\cup
[A^{f}\cap P^{2}].
\end{array}$$ By Lemma 14 we can find a Borel subset $S$ of $P$ and a finer topology $\sigma$ on $S$ such that $\big([S,\sigma ],(f^{S}_{n})_{n}\big)$ is a converging situation.
By the Jankov-von Neumann Theorem there is $f':S\rightarrow
u^{-1}(S)$ (respectively, ${g':S\!\rightarrow\! v^{-1}(S)}$) Baire measurable such that $u\big(f'(x)\big)\! =\! x$ (respectively, ${v\big(g'(x)\big)\!
=\! x}$), for each $x\in S$. Notice that $f'$ and $g'$ are one-to-one. Let $G$ be a dense $G_{\delta}$ subset of $S$ such that $f'\vert_{G}$ and $g'\vert_{G}$ are continuous. These functions are witnesses to the inequality $A^f\cap G^2\leq^r_{c}A$. By Lemma 11, we get ${\hbox{Gr}(f^{S}_{n})\subseteq\overline{\hbox{Gr}(f^{S}_{n})\cap G^2}}$. Therefore ${\overline{A^f\cap S^2}=\overline{A^f\cap G^2}}$, $\Delta
(G)=G^{2}\cap\overline{A^f
\cap G^2}\setminus (A^f\cap G^2)$, and $A^f\cap G^2$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$ by Lemma 11.
By Lemma 14 we can find a Borel subset $B$ of $G$, equipped with some topology $\tau$ finer than $\sigma$, such that $\big([B,\tau
],(f^{B}_{n})_{n}\big)$ is a converging situation.
Let $\big(X,(f_{n})\big)$ be a converging situation. The following statements are equivalent:
\(a) $A^f$ is $\leq^r_{c}$-minimal among ${{\bf\Delta}^{1}_{1}}\setminus
\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets.
\(b) For any Borel subset $B$ of $X$ and any finer Polish topology $\tau$ on $B$, $A^f \leq^r_{c} A^f\cap B^2$ if $A^f\cap B^2$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
\(c) For any Borel subset $B$ of $X$ and for each finer topology $\tau$ on $B$, $A^f \leq^r_{c} A^f\cap B^2$ if $\big([B,\tau],(f^{B}_{n})_{n}\big)$ is a converging situation.
[**Proof.**]{} (a) $\Rightarrow$ (b) and (b) $\Rightarrow$ (c) are obvious. So let us show that (c) $\Rightarrow$ (a). Let $Y$, $Y'$ be Polish spaces, $A\in{{\bf\Delta}^{1}_{1}}(Y\times Y')\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$. We assume that $A\leq^r_{c} A^f$. By Theorem 15 we get a Borel subset $B$ of $X$ and a finer topology $\tau$ on $B$ such that $\big([B,\tau],(f^{B}_{n})_{n}\big)$ is a converging situation and $A^f\cap B^{2}\leq^r_{c}A$. By (c) we get $A^f \leq^r_{c} A^f\cap B^2$. Therefore $A^f \leq^r_{c}
A$.
This is the sufficient condition for minimality that we mentioned in the introduction. The following definitions, notation and facts will be used here and in Section 5 to build the reduction functions in the minimality results that we want to show.
Let $R$ be a relation on a set $E$.
$\bullet$ An $R\! -\! path$ is a finite sequence $(e_{i})_{i\leq n}\!\subseteq\! E$ such that $(e_{i},e_{i+1})\!\in\! R$ for $i\! <\! n$.
$\bullet$ We say that $E$ is $R\! -\! connected$ if there is an $R$-path $(e_{i})_{i\leq n}$ with $e_0=e$ and $e_n=e'$ for each $e,\ e'\in E$.
$\bullet$ An $R\! -\! cycle$ is an $R$-path $(e_{i})_{i\leq n}$ such that $n\!\geq\! 3$ and $$[0\!\leq\! i\!\not=\! j\!\leq\! n~\hbox{\rm and}~e_{i}\! =\!
e_{j}]~\Leftrightarrow ~\{ i,j\}\! =\!\{ 0,n\}.$$ $\bullet$ We say that $R$ is $acyclic$ if there is no $R$-cycle.
Recall that if $R$ is symmetric and acyclic, $e$, $e'\!\in\! E$ and $(e_{i})_{i\leq n}$ is an $R$-path with $e_{0}=e$ and $e_{n}=e'$, then we can find a unique $R$-path $p_{e,e'}:=(f_{j})_{j\leq m}$ without repetition with $f_{0}=e$ and $f_{m}=e'$. We will write $|p_{e,e'}|=m+1$.
[**Notation.**]{} Let $\Theta:= (\theta_n)\subseteq
2^{<\omega}$ with $|\theta_n|=n$. We will use two examples of such $\Theta$’s: $\theta_n = 0^n$ and $\theta_n = s_n$ (where $s_n$ has been defined in the introduction to build $A_1$). We define a tree ${\mathfrak R}_\Theta$ on $2\times 2$: $${\mathfrak R}_\Theta:=\{ (e,e')\!\in\! (2\!\times\! 2)^{<\omega}/e\!
=\! e'\ \ \hbox{\rm or}\ \ \exists n\!\in\!\omega\ \exists w\!\in\!
2^{<\omega}\ \ (e,e')\! =\! (\theta_n0w,\theta_n1w)\}.$$ Recall that $s({\mathfrak R}_\Theta )$ is the symmetric set generated by ${\mathfrak R}_\Theta $.
\(a) $\big( 2^n,s({\mathfrak R}_\Theta )\big)$ is connected, for each $n\in\omega$.
\(b) The relation $s({\mathfrak R}_\Theta )$ is acyclic.
\(c) If $e,\ e'\in 2^n$ and $l<n$ is maximal with $e(l)\not=
e'(l)$, the coordinate $l$ is changed only once in $p_{e,e'}$, and the other changed coordinates are at a level less than $l$.
[**Proof.**]{} (a) We argue by induction on $n$. As $(\emptyset )$ is an $s({\mathfrak R}_\Theta )$-path from $\emptyset$ to $\emptyset$, the statement is true for $n=0$. Assume that it is true at the level $n$, and let $e$, $e'\in 2^{n+1}$. We can write $e=s\epsilon$ and $e'=s'\epsilon'$, where $s,t\in 2^n$ and $\epsilon ,\epsilon'\in 2$. If $\epsilon =\epsilon'$, then let $(f_i)_{i\leq m}$ be an $s({\mathfrak R}_\Theta )$-path with $f_0=s$ and $f_m=s'$. Let $e_i:=f_i\epsilon$. Then $(e_i)_{i\leq m}$ is an $s({\mathfrak R}_\Theta
)$-path with $e_0=e$ and $e_m=e'$. If $\epsilon\not=\epsilon'$, then let $(f_i)_{i\leq m}$ be an $s({\mathfrak R}_\Theta )$-path with $f_0=s$ and $f_m=\theta_n$, and $(g_j)_{j\leq p}$ be an $s({\mathfrak R}_\Theta )$-path with $g_0=\theta_n$ and $g_p=s'$. We set $e_i:=f_i\epsilon$ if $i\leq m$, $g_{i-m-1}\epsilon'$ if $m<i\leq m+p+1$. Then $(e_i)_{i\leq m+p+1}$ is an $s({\mathfrak R}_\Theta )$-path with $e_0=e$ and $e_{m+p+1}=e'$.
\(b) We argue by contradiction. Let $(e_i)_{i\leq n}$ be an $s({\mathfrak R}_\Theta )$-cycle, $p>0$ be the common length of the $e_i$’s, and $l<p$ maximal such that the sequence $(e_i(l))_{i\leq
n}$ is not constant. We can find $i_1$ minimal with $e_{i_1}(l)\not=
e_{i_1+1}(l)$. We have ${e_{i_1}(l)\! =\! e_0(l)\! =\! e_n(l)}$. We can find $i_2>i_1+1$ minimal with ${e_{i_1+1}(l)\not= e_{i_2}(l)}$. Then ${e_{i_1}(l)=e_{i_2}(l)}$ and $e_{i_1}=e_{i_2}$, because $|\theta_l|=l$. Thus $i_1=0$ and $i_2=n$. But $e_{i_1+1} = e_{i_2-1}$, which is absurd. Note that this proof of (b) is essentially in \[L3\], Theorem 2.7.
\(c) This follows from (b) and the proof of (a).
Now we come to some examples of converging situations, with some cycle relations involved.
[**Notation.**]{} Let $S\subseteq\omega$, and $$A^S:=\{ (s0\gamma ,s1\gamma )/~s\!\in\! 2^{<\omega}\hbox{\rm and\
Card}(s)\!\in\! S~\hbox{\rm and}~\gamma\!\in\! 2^\omega\}.$$ ($\hbox{\rm Card}(s)$ is the number of ones in $s$.) We define partial homeomorphisms $${f^S_{n}:\bigcup_{s\in 2^n,~\hbox{\rm Card}(s)\in S}~N_{s0}\rightarrow
\bigcup_{s\in 2^n,~\hbox{\rm Card}(s)\in S}~N_{s1}}$$ by ${f^S_{n}(s0\gamma ):=s1\gamma}$. Notice that $A^S=A^{f^S}$ is Borel. One can show the existence of ${\mathfrak A}:2^{\omega}\rightarrow
2^\omega$ continuous such that ${\mathfrak A}(S)$ is a Borel code for $A^S$, for each $S\subseteq\omega$. Notice that $\big(2^\omega ,(f^S_{n})_n\big)$ is a converging situation if and only if $S$ is infinite. This is also equivalent to $A^{S}\notin \mbox{pot}({{\bf\Pi}^{0}_{1}})$. Indeed, if $S$ is finite, $\overline{A^{S}}\setminus A^{S}$ is a countable subset of $\Delta
(2^\omega )$.** So in the sequel we will assume that $S$ is infinite.**
Let $n_{S}:=\hbox{\rm min}~S$, and $S':=\{ n\! -\! n_{S}/n\!\in\! S\}$. Then $0\in S'$ and the maps $u$ and $v$ defined by $u(\alpha )=v(\alpha
):=1^{n_{S}}\alpha$ are witnesses to $A^{S'}\leq^r_{c} A^{S}$.** So in the sequel we will also assume that $0\in S$.**
$\bullet$ If $S\subseteq\omega$ and $t\in\omega^{<\omega}\setminus\{\emptyset\}$, then we set $f^S_{t}:=f^S_{t(0)}\ldots f^S_{t(|t|-1)}$, when it makes sense. We will also use the following tree ${\mathfrak R}$ on $2\times 2$. If $s,t\in 2^{<\omega}$, then we set $$s~{\mathfrak R}~t~\Leftrightarrow ~|s|\! =\! |t|~\mbox{and}~
(N_{s}\times N_{t})\cap A^S\!\not=\!\emptyset .$$ In particular, if $n_0\! <\! n_1$ and $1\!\in\! S$, then we get ${f^{S}_{<n_0,n_1>}(0^\infty )=f^{S}_{<n_1,n_0>}(0^\infty )}$. This is the kind of cycle relation we mentioned in the introduction. In this case $s({\mathfrak R})$ is not acyclic since $<0^{n_1+1},0^{n_0}10^{n_1-n_0},0^{n_0}10^{n_1-n_0-1}1,0^{n_1}1,0^{n_1+1}>$ is an $s({\mathfrak R})$-cycle. We set $f^{C}_{n}:={f^{S}_{n}}\vert_{C\cap
{f^{S}_{n}}^{-1}(C)}$ for each Borel subset $C$ of $2^\omega$, when $S$ is fixed.
$\bullet$ Let $(H)$ be the following hypothesis on $S$: $$(H)~\left\{\!\!\!\!\!\!
\begin{array}{ll}
& ~~~~\mbox{Let}~C\in{{\bf\Delta}^{1}_{1}}(2^\omega )\mbox{,}~\sigma ~
\mbox{be~a~finer~topology~on}~C~\mbox{such~that}~\big([C,\sigma ],
(f^{C}_{n})_{n}\big)\cr
& \mbox{is~a~converging~situation,}~l\mbox{, }p\in\omega\mbox{.~Then~we~can~find}~n\geq l~
\mbox{and}~\gamma\in D_{f^{C}_{n}}\cr
& \mbox{with}~\mbox{Card}(\gamma\lceil n)+\big(S\cap [0,p]\big)=S\cap
\big(\mbox{Card}(\gamma\lceil n)+[0,p]\big).
\end{array}
\right.$$ The next result will lead to a combinatorial condition on $S$ implying the minimality of $A^{S}$ among non-potentially closed sets.
Let $S$ satisfy $(H)$, $B\!\in\!{{\bf\Delta}^{1}_{1}}(2^\omega )$, and $\tau$ a finer topology on $B$ such that $\big([B,\tau ],(f^{B}_{n})_{n}\big)$ is a converging situation. Then $A^{S}\leq^r_{c} A^{S}\cap B^2$.
[**Proof.**]{} Let $X:=[B,\tau ]$, $f_{n}:=f^{B}_{n}$. We are trying to build continuous maps $u,~v:2^\omega\rightarrow X$ such that $A^S=(u\times v)^{-1}(A^f)$. We will actually have more: $u=v$ will be one-to-one. We set $s\wedge t\!:=\! s\lceil
\mbox{max}\{n\!\in\!\omega /s\lceil n\! =\! t\lceil n\}$, for $s,t\in
2^{<\omega}$.
$\bullet$ We construct a sequence $(U_{s})_{s\in
2^{<\omega}}$ of nonempty clopen subsets of $X$, $\phi:
\omega\rightarrow\omega$ strictly increasing, and $\theta:
\omega\rightarrow\omega$ such that $$\begin{array}{ll}
(i) & ~~U_{s^\frown i }\subseteq U_{s}.\cr
(ii) & ~\hbox{\rm diam}(U_{s})\leq 1/|s|~\mbox{if}~s\not=\emptyset .\cr
(iii) & (s~{\mathfrak R}~t~\hbox{\rm and}~s\!\not=\! t)\Rightarrow\left\{\!\!\!\!\!\!
\begin{array}{ll}
& U_{t}\! =\! f_{\phi (| s\wedge t| )}[U_{s}]\mbox{,}\cr
& \theta (| s\!\wedge\! t| )\! +\!\big(S\cap [0,| s\!\wedge\! t|
]\big)\! =\! S\cap\big(
\theta (| s\!\wedge\! t| )\! +\! [0,| s\!\wedge\! t| ]\big)\mbox{,}\cr
& \forall z\!\in\! U_{s}~~\mbox{Card}\big(z\lceil\phi(| s\!\wedge\! t|
)\big)\!
=\!\theta (|s\!\wedge\! t|)\! +\!\mbox{Card}(s\lceil |s\!\wedge\! t|).
\end{array}
\right.\cr
(iv) & ~(\neg ~s~{\mathfrak R}~t~\hbox{\rm and}~|s|=|t|)~\Rightarrow
~(U_{s}\times U_{t})\cap [\bigcup_{q<|s|}
\hbox{Gr}(f_{q})\cup\Delta (X)]=\emptyset .
\end{array}$$ $\bullet$ First we show that this construction is sufficient to get the theorem. We define a continuous map ${u:2^\omega\rightarrow X}$ by $\{ u(\alpha )\}:=\bigcap_{n}U_{\alpha\lceil n}$. If $\alpha <_{\mbox{lex}}\beta$, then we have $\neg\beta\lceil r~{\mathfrak R}~\alpha\lceil r$ if $r$ is big enough, thus by condition (iv), $\big( u(\beta ), u(\alpha )\big)$ is in ${U_{\beta\lceil r}
\times U_{\alpha\lceil r}\subseteq X^2\setminus\Delta (X)}$. Therefore $u$ is one-to-one. If ${(\alpha ,\beta )\in A^{S}}$, fix $n$ such that $\beta =f^{S}_{n}(\alpha )$. Then $\alpha\lceil r$ and $\beta\lceil
r$ satisfy the hypothesis in condition (iii) for each $r>n$. Therefore $u(\beta )=f_{\phi (n)}\big(u(\alpha )\big)$ and $\big(u(\alpha ),u(\beta )\big)\in A^{f}$. If $\alpha =\beta$, then $(\alpha ,\beta )\notin A^{S}$ and ${\big(u(\alpha ),u(\beta )\big)\in\Delta (X)\subseteq\neg A^{f}}$. Otherwise, $(\alpha ,\beta )\notin\overline{A^{S}}$ and there is $r_{0}$ such that $\alpha\lceil r$ and $\beta\lceil r$ satisfy the hypothesis in condition (iv) for $r\geq r_{0}$. This shows that $\big(u(\alpha ),u(\beta )\big)\notin A^{f}$. So it is enough to do the construction.
$\bullet$ We set $U_{\emptyset}:=X$. Suppose that $(U_{s})_{s\in 2^{\leq p}}$, $\big(\phi (j)\big)_{j<p}$ and $\big(\theta (j)\big)_{j<p}$ satisfying conditions (i)-(iv) have been constructed, which is done for $p=0$.
$\bullet$ We will use the relation ${\mathfrak R}_\Theta$ defined before Proposition 18 with $\theta_n\! :=\! 0^n$. Notice that ${\mathfrak R}_\Theta\subseteq {\mathfrak R}$. We set $t_{0}\!:=\!\theta_p0$. We define a partition of $2^{p+1}$ as follows. Using Proposition 18.(b) we set, for $k\!\in\!\omega$, $$H_k\!:=\!\{ t\!\in\! 2^{p+1}/|p_{t,t_0}|\! =\! k\! +\! 1\}.$$ If $H_{k+1}$ is non-empty, then $H_k$ is non-empty. Thus we can find an integer $q$ such that $H_0$, ..., $H_q$ are not empty and $H_k$ is empty if $k>q$. We order $2^{p+1}$ as follows: $t_{0}$, then $H_{1}$ in any order with $\theta_p1$ first, $H_{2}$ in any order, $\ldots$, $H_{q}$ in any order. This gives $t_{0}$, $\ldots$, $t_{2^{p+1}-1}$. Notice that we can find $j<n$ such that $t_{j}~s({\mathfrak R}_\Theta )~t_{n}$ if $0<n<2^{p+1}$. In particular, if $E^n:=\{t_{j}/j\leq n\}$, then $\big( E^n,s({\mathfrak R}_\Theta )\big)$ is connected for each $n<2^{p+1}$.
$\bullet$ We will construct integers $\phi (p)$, $\theta
(p)$ and nonempty clopen subsets $U^n_{k}$ of $X$, for $n<2^{p+1}$ and $k\leq n$, satisfying $$\begin{array}{ll}
(1) & U^n_{k}\subseteq U_{t_{k}\lceil p}.\cr
(2) & \hbox{\rm diam}(U^n_{k})\leq 1/{p+1}.\cr
(3) & (t_{k}~{\mathfrak R}~t_{l}~\hbox{\rm and}~t_{k}\!\not=\!
t_{l})\Rightarrow\cr
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left\{\!\!\!\!\!\!
\begin{array}{ll}
& U^n_{l}\! =\! f_{\phi (|t_{k}\wedge t_{l}|)}[U^n_{k}]\mbox{,}\cr
& \theta (|t_{k}\!\wedge\! t_{l}|)\! +\!\big(S\!\cap\!
[0,|t_{k}\!\wedge\! t_{l}|]\big)\! =\!
S\!\cap\!\big(\theta (|t_{k}\!\wedge\! t_{l}|)\! +\!
[0,|t_{k}\!\wedge\! t_{l}|]\big)\mbox{,}\cr &
\forall z\!\in\! U^n_{k}~~\hbox{\rm Card}\big(z\lceil\phi(|t_{k}\!\wedge\!
t_{l}|)\big)\! =\!
\theta (|t_{k}\!\wedge\! t_{l}|)\! +\!\hbox{\rm
Card}(t_{k}\lceil|t_{k}\!\wedge\!
t_{l}|).
\end{array}\right.\cr
(4) & \neg ~t_{k}~{\mathfrak R}~t_{l}~\Rightarrow ~(U^n_{k}\times
U^n_{l})\cap
[\bigcup_{q\leq p} \hbox{Gr}(f_{q})\cup\Delta (X)]=\emptyset .\cr
(5) & U^{n+1}_{k}\subseteq U^n_{k}.
\end{array}$$ We will then set $U_{t_{k}}:=U^{2^{p+1}-1}_{k}$ for $k<2^{p+1}$, so that conditions (i)-(iv) are fullfilled.
$\bullet$ Let $C\in{{\bf\Delta}^{0}_{1}}(U_{t_{0}\lceil
p})\setminus\{\emptyset\}$ such that $C^{2}\cap\bigcup_{q\leq p}~\hbox{Gr}(f_{q})=\emptyset$. Apply hypothesis $(H)$ to $C$ and $\sigma\!:=\!\tau$. This gives ${n_{0}\!\geq\!\hbox{\rm sup}\{\phi (q)\! +\! 1/q\! <\! p\}}$ and $\gamma\in D_{f^{C}_{n_{0}}}$ such that $$\hbox{\rm Card}(\gamma\lceil n_{0})+\big(S\cap [0,p]\big)=
S\cap\big( \hbox{\rm Card}(\gamma\lceil n_{0})+[0,p]\big).$$ We set $\phi (p):=n_{0}$, $\theta (p):=\hbox{\rm Card}(\gamma\lceil n_{0})$.
We then choose $U^{0}_{0}\in{{\bf\Delta}^{0}_{1}}\big( C\cap
f^{-1}_{n_{0}}(C)\big)\setminus\{\emptyset\}$ with suitable diameter such that $f_{n_{0}}[U^{0}_{0}]\cap U^{0}_{0}=\emptyset$, and $z\lceil n_{0}\! =\!\gamma\lceil n_{0}$ for each $z\!\in\!
U^{0}_{0}$. Assume that ${U^{0}_{0}, ~\ldots,~U^{n-1}_{0},}$ ${\ldots, U^{n-1}_{n-1}}$ satisfying conditions (1)-(5) have been constructed (which has already been accomplished for $n=1$). As $n\geq 1$, we have $t_{n}\not= t_{0}$ and $|p_{t_n,t_0}|\geq 2$. So fix $r<n$ such that $p_{t_n,t_0}(1)=t_{r}$. Notice that $U^{n-1}_{r}$ has been constructed.
[**Case 1.**]{} $t_{n}\lceil p=t_{r}\lceil p$.
- We have $t_{n}\lceil p=\theta_p$, thus $|p_{t_n,t_0}|
=2$, $r=0$, $t_{n}=\theta_p1$ and $n=1$. Moreover, $U^{0}_{0}$ is a subset of ${f^{-1}_{\phi (p)}}(U_{t_{1}\lceil p})$, so we can choose a nonempty clopen subset $U^1_{1}$ of $f_{\phi (p)}[U^{0}_{0}]$ with suitable diameter. Then we set $U^1_{0}\!:=\! f^{-1}_{\phi (p)}(U^1_{1})\!\subseteq\!
U^{0}_{0}$. So ${U^n_{0},\ldots ,U^n_{n}}$ are constructed and fullfill (1)-(3) and (5). It remains to check condition (4).
- Fix ${k,~l\leq 1}$ such that $\neg ~t_{k}~{\mathfrak
R}~t_{l}$. Then $k=1=1-l$. We have $U^1_{1} = f_{\phi(p)}[U^1_{0}]$. Thus $$U^1_{1}\times U^1_{0}=
f_{\phi(p)}[U^1_{0}]\times U^1_{0}=
f_{n_{0}}[U^1_{0}]\times U^1_{0}\subseteq
f_{n_{0}}[U^0_{0}]\times U^0_{0}\subseteq C^{2}\mbox{,}$$ so we are done by the choice of $C$ and $U^{0}_{0}$.
[**Case 2.**]{} $t_{n}\lceil p\not=t_{r}\lceil p$.
2.1. $t_{r}~{\mathfrak R}_{\Theta}~t_{n}$.
- By the induction hypothesis we have $U_{t_{n}\lceil p}=
f_{\phi(|t_{r}\wedge t_{n}|)}[U_{t_{r}\lceil p}]$ and $U^{n-1}_{r}\subseteq U_{t_{r}\lceil p}$. We choose a nonempty clopen subset $U^n_{n}$ of $f_{\phi(| t_{r}\wedge t_{n}| )}[U^{n-1}_{r}]$ with suitable diameter, so that conditions (1)-(5) for $k = l = n$ are fullfilled.
- We then define the $U_{q}^n$’s for $q < n$, by induction on $|p_{t_{q},t_{n}}|$: fix $m\leq n$ with $p_{t_{q},t_{n}}(1) = t_{m}$. Notice that $q=r$ if $m=n$.
2.1.1. $t_{m}~{\mathfrak R}_{\Theta}~t_{q}$.
We have $m<n$ since we cannot have $p_{t_{q},t_{n}}(1)~{\mathfrak
R}_{\Theta}~t_{q}$ and $t_{q}~{\mathfrak R}_{\Theta}~p_{t_{q},t_{n}}(1)$ ($\tilde s \leq_{\hbox{\rm lex}}
\tilde t$ if $\tilde s~{\mathfrak R}_{\Theta}~\tilde t$). So $U^{n-1}_{q} = f_{\phi (|t_{m}\wedge t_{q}|)}[U^{n-1}_{m}]$. We put $$U_{q}^n:= f_{\phi (|t_{m}\wedge t_{q}|)}[U_{m}^n].$$ The set $U^{n}_{q}$ is a nonempty clopen subset of $U^{n-1}_{q}$ since $U^{n}_{m}\subseteq U^{n-1}_{m}$.
2.1.2. $t_{q}~{\mathfrak R}_{\Theta}~t_{m}$.
If $m<n$, then we have $U^{n-1}_{m} = f_{\phi (|t_{q}\wedge
t_{m}|)}[U^{n-1}_{q}]$. We put $$U_{q}^n:= f_{\phi (|t_{q}\wedge t_{m}|)}^{-1}(U_{m}^n)\mbox{,}$$ so that $U^{n}_{q}$ is a nonempty clopen subset of $U^{n-1}_{q}$. If $m=n$, then $q=r$ and the same conclusion holds, by the choice of $U^{n}_{n}$.
- So condition (5) is fullfilled in both cases. Conditions (1) and (2) are fullfilled for $k=q$, too. Let us check that the first part of condition (3) restricted to ${\mathfrak R}_{\Theta}$ is fullfilled. Fix $k\not= l\leq n$ with $t_{k}~{\mathfrak R}_{\Theta}~t_{l}$. If $|p_{t_{k},t_{n}}| = 1$ and $|p_{t_{l},t_{n}}| = 2$, then the link between $t_{k}$ and $t_{l}$ has already been considered. The argument is similar if $|p_{t_{k},t_{n}}| = 2$ and $|p_{t_{l},t_{n}}| = 1$. If $|p_{t_{k},t_{n}}|$ and $|p_{t_{l},t_{n}}|$ are at least 2, then $p_{t_{k},t_{n}}(1) = p_{t_{l},t_{n}}(0)$ or $p_{t_{k},t_{n}}(0) = p_{t_{l},t_{n}}(1)$, by Proposition 18.(b). Here again, the link has already been considered. So condition (3) restricted to ${\mathfrak R}_{\Theta}$ is fullfilled. It remains to check conditions (3) and (4).
- Fix $k\not= l$ such that $t_{k}~{\mathfrak R}~t_{l}$. Then $t_{k}$, $t_{l}$ differ at one coordinate only, and $t_{k}<_{\mbox{lex}}t_{l}$.
[**Claim.**]{} *Assume that $t_{k}$, $t_{l}$ differ at one coordinate only, and that $t_{k}<_{\mbox{lex}}t_{l}$. Then $$\mbox{Card}\big(z\lceil\phi(|t_{k}\wedge t_{l}|)\big)=\theta (|t_{k}\wedge
t_{l}|)+\mbox{Card}(t_{k}\lceil|t_{k}\wedge t_{l}|)$$ for each $z\in U^n_{k}$.*
We can write $$\begin{array}{ll}
t_{k} &:=0^{n_{0}}10^{n_{1}}1\ldots 0^{n_{j-1}}10^{n_{j}}1
0^{n_{j+1}}1\ldots 0^{n_{q-1}}10^{n_{q}}\mbox{,}\cr\cr
t_{l} &:=0^{n_{0}}10^{n_{1}}1\ldots 0^{n_{j-1}}10^{n'_{j}}10^m
10^{n_{j+1}}1\ldots 0^{n_{q-1}}10^{n_{q}}~~(n'_{j}+1+m=n_{j}).
\end{array}$$ By construction we have
Notice that the length of $t_{k}\!\wedge\! t_{l}$ is equal to ${\Sigma_{r\leq j-1}~(n_{r}\! +\! 1)\! +\! (n'_{j}\! +\! 1)\! -\!
1}$. Set $$f:=f_{\phi (n_{0})}\!\ldots\! f_{\phi (\Sigma_{r\leq j-1}~(n_{r}+1)-1)}.$$ Then $U^n_{l}=ff_{\phi (|t_{k}\wedge t_{l}|)}f^{-1}(U^n_{k})$. Fix $k'\not= l'\leq n$ such that $$\begin{array}{ll}
t_{k'}
& := 0^{\Sigma_{r\leq j-1}~(n_{r}+1)+n_{j}}10^{n_{j+1}}1\ldots 0^{n_{q-1}}10^{n_{q}},\cr\cr
t_{l'}
& := 0^{\Sigma_{r\leq j-1}~(n_{r}+1)+n'_{j}}10^m10^{n_{j+1}}1\ldots 0^{n_{q-1}}10^{n_{q}}.
\end{array}$$ Note that ${\hbox{\rm Card}\big(y\lceil\phi(|t_{k}\wedge t_{l}|)\big) =
\theta (|t_{k}\wedge t_{l}|)+\hbox{\rm Card}(t_{k'}\lceil |t_{k}\wedge
t_{l}|)}$, for each $y\in U^n_{k'}$, since $t_{k'}~{\mathfrak R}_{\Theta}~t_{l'}$. But ${\hbox{\rm Card}\big(z\lceil\phi(|t_{k}\wedge t_{l}|)\big) =
\hbox{\rm Card}\big(y\lceil\phi(|t_{k}\wedge t_{l}|)\big)+j}$, for each $z$ in $U^n_{k}=f[U^n_{k'}]$. As $$\hbox{\rm Card}(t_{k'}\lceil |t_{k}\wedge t_{l}|) =
\hbox{\rm Card}(t_{k}\lceil |t_{k}\wedge t_{l}|)-j\mbox{,}$$ we get $$\hbox{\rm Card}\big(z\lceil\phi(|t_{k}\wedge t_{l}|)\big)=\theta
(|t_{k}\wedge t_{l}|)+\hbox{\rm Card}(t_{k}\lceil |t_{k}\wedge
t_{l}|).\leqno{(+)}$$ This proves the claim.
- The second assertion in condition (3) is clearly fullfilled since $|t_{k}\wedge t_{l}|=|t_{k'}\wedge t_{l'}|$. As $t_{k}~{\mathfrak
R}~t_{l}$ and $t_{k}\not= t_{l}$ we get $\hbox{\rm Card}(t_{k}\lceil| t_{k}\wedge t_{l}| )\in
S$. This implies that $S$ contains ${\theta (|t_{k}\wedge t_{l}|)+\hbox{\rm Card}(t_{k}\lceil |t_{k}\wedge
t_{l}|)}$. By the claim we get $$U^n_{l}=f_{\phi(|t_{k}\wedge t_{l}|)}ff^{-1}(U^n_{k})=
f_{\phi(|t_{k}\wedge t_{l}|)}[U^n_{k}]$$ (the compositions $ff_{\phi(|t_{k}\wedge t_{l}|)}f^{-1}$ and $f_{\phi(|t_{k}\wedge t_{l}|)}ff^{-1}$ are defined on $U^n_{k}$, so they are equal on this set). Thus condition (3) is fullfilled.
- To get condition (4), fix $k,l\leq n$ with $\neg
~t_{k}~{\mathfrak R}~t_{l}$, $v(i):=|p_{t_k,t_l}(i)\wedge p_{t_k,t_l}(i+1)|$, and $\varepsilon
(i):=1$ (resp., $-1$) if $p_{t_k,t_l}(i)~{\mathfrak R}_{\Theta}~p_{t_k,t_l}(i+1)$ (resp., $p_{t_k,t_l}(i+1)~{\mathfrak R}_{\Theta}~p_{t_k,t_l}(i)$), for $i+1<|p_{t_k,t_l}|$. We set ${f^\varepsilon_{v}:={f_{\phi
(v(|v|-1))}^{\varepsilon (|v|-1)}\ldots
f_{\phi (v(0))}^{\varepsilon (0)}}}$, so that $U^n_{l}=f^\varepsilon_{v}(U^n_{k})$. Let $m$ be maximal such that $t_{k}(m)\not= t_{l}(m)$. As $\phi$ is strictly increasing, we get $(U^n_{k}\times U^n_{l})\cap\Delta (X)=\emptyset$, by Proposition 18.(c).
- If $t_{k}$, $t_{l}$ differ in at least two coordinates $m\not= m'$, then the number of appearances of $m$ and $m'$ in $v$ is odd. As $\phi$ is strictly increasing, this is also true for $\phi (m)\not=\phi (m')$ in $\{\phi\big( v(i)\big)/i<|v|\}$. This implies that $(U^n_{k}\times U^n_{l})\cap [\bigcup_{q\leq p}~\hbox{Gr}(f_{q})]=\emptyset$.
- If $t_{k}$, $t_{l}$ differ at only one coordinate $m$ and $t_{k}>_{\hbox{\rm lex}}t_{l}$, then ${\alpha\big(\phi
(m)\big)>\beta\big(\phi (m)\big)}$ if $(\alpha ,\beta )$ is in $U^n_{k}\times U^n_{l}$, and $(U^n_{k}\times U^n_{l})\cap [\bigcup_{q\leq
p}~\hbox{Gr}(f_{q})]=\emptyset$.
- So we may assume that $t_{k}$, $t_{l}$ differ only at coordinate $\phi^{-1}(q)$, and that $t_{k}<_{\hbox{\rm lex}}t_{l}$. By the Claim we have $(+)$ for each $z\in U^n_{k}$. But $\hbox{\rm Card}(t_{k}\lceil |t_{k}\wedge t_{l}|)\notin
S$, since $\neg ~t_{k}~{\mathfrak R}~t_{l}$. So $\hbox{\rm Card}(z\lceil q)\notin S$ if $z\in
U^n_{k}$, and $f_{q}$ is not defined on $U^n_{k}$.
2.2. $t_{n}~{\mathfrak R}_{\Theta}~t_{r}$.
This cannot hold since $t_{r}{\mathfrak R}_{\Theta}t_{n}$. Indeed, if $t_{n}\! =\! 0^{n_{0}}10^{n_{1}}1\ldots 0^{n_{q-1}}10^{n_{q}}$, then $$\begin{array}{ll}
p_{t_n,t_0}(1)
& = 0^{n_{0}+n_{1}+1}1\ldots 0^{n_{q-1}}10^{n_{q}}\mbox{,}\cr
& \ \ .\cr
& \ \ .\cr
& \ \ .\cr
p_{t_n,t_0}(|p_{t_n,t_0}|-2)\!\!\!\!
& = 0^{n_{0}+n_{1}+\ldots +n_{q-1}+q-1}10^{n_{q}}.
\end{array}$$ This finishes the proof.
The set $S$ satisfies hypothesis $(H)$ if the following is fullfilled: $$\forall p\!\in\!\omega~\exists k\!\in\!\omega~\forall
q\!\in\!\omega~\exists c\!\in\!\omega\!\cap\!
[q,q\! +\! k]~~c\! +\!\big(S\!\cap\! [0,p]\big)\! =\! S\!\cap\!\big(
c\! +\! [0,p]\big).\leqno{(M)}$$ In particular, condition (M) implies that $A^{S}$ is minimal among non-potentially closed sets for $\leq^r_{c}$.
[**Proof.**]{} Note that $\emptyset\not=\Delta (C)
\subseteq\overline{\bigcup_{q\geq l}~\hbox{Gr}(f^{C}_{q})}$, since $\big([C,\sigma ],
(f^{C}_{n})_{n}\big)$ is a converging situation. So fix $q_{0}\geq
l$ such that $D_{f^{C}_{q_{0}}}\not=\emptyset$, and $O_{0}:=D_{f^{C}_{q_{0}}}$. Assume that $q_{r}$ and $O_{r}$ have been constructed. We then choose $q_{r+1}>q_{r}$ such that ${O_{r}\cap
(f^{C}_{q_{r+1}})^{-1}(O_{r})\not=\emptyset}$, and we define ${O_{r+1}:=O_{r}\cap (f^{C}_{q_{r+1}})^{-1}(O_{r})}$. This gives $(q_{r})_{r<M}$ and $(O_{r})_{r<M}$, where $M:=p+k$.
$\bullet$ For $t\in\omega^{<\omega}$, we let $f^C_{t}:=f^C_{t(0)}\ldots f^C_{t(|t|-1)}$, when it makes sense. We choose $${n\geq\hbox{\rm max}(\hbox{\rm sup}\{q_{r}+1/r<M\},l)}$$ with ${f^{C}_{q_{0},\ldots ,q_{M-1}}[O_{M-1}]\cap {f^{C}_{n}}^{-1}(f^{C}_{q_{0},\ldots
,q_{M-1}}[O_{M-1}])}\not=\emptyset$. Let $\beta:=f^{C}_{q_{0},\ldots ,q_{M-1}}(\alpha )$ be in the intersection. Notice that $q:=\hbox{\rm Card}(\beta\lceil n)-M\in\omega$. This gives $c$ in $\omega\cap [q,q+k]$. As $0\in S$, there is $j\leq k$ with ${c=\hbox{\rm Card}(\beta\lceil n)-p-j\in S}$. Notice that $\beta = f^{C}_{q_{0},\ldots ,q_{p+j-1}}(\gamma )$, where $\gamma = f^{C}_{q_{p+j},\ldots ,q_{M-1}}(\alpha )$. As ${\hbox{\rm Card}(\gamma\lceil n) =c}$, $f^{S}_{n}(\gamma )$ is defined. But $f^{C}_{n}(\beta )$ is in $f^{C}_{q_{0},\ldots
,q_{M-1}}[O_{M-1}]$ and $f^{S}_{n}(\gamma )$ is in $C$. So $f^{C}_{n}(\gamma )$ is defined.
$\bullet$ The lemma now follows from Corollary 16 and Theorem 19.
[**Example.**]{} We set ${S_{m,F}:=\{n\!\in\!\omega /
n~(\hbox{\rm mod}~m)\!\in\!\{ 0\}\!\cup\! F\}}$, where $m\in\omega\setminus\{ 0\}$ and ${F\subseteq m\setminus\{ 0\}}$. Then $S_{m,F}$ fullfills condition (M). In particular, $A^{\omega}$ is minimal. But this gives only countably many examples. To get more, we need some more notation:
[**Notation.**]{} For $\beta\in \omega^\omega$, we set $S_{\beta}:=\{\Sigma_{i<l}~\big(1+\beta (i)\big)/l\!\in\!\omega\}$. Notice that $0\in S_{\beta}$, $S_{\beta}$ is infinite, and that any infinite $S$ containing $0$ is of this form. Moreover, the map $\beta\mapsto S_{\beta}$ is continuous since $n\!\in\! S_{\beta}\Leftrightarrow\exists l\!\leq\! n~~n\! =\!
\Sigma_{i<l}~\big(1\! +\!\beta (i)\big)$. We will define a family $(\beta_{\alpha})_{\alpha\in 2^\omega}$. Actually, we can find at least two examples:
$\bullet$ The original example is the following. For $\alpha\in
2^\omega$, we recursively define a sequence $(s_{\alpha ,n})_{n}\subseteq 2^{<\omega}$ as follows: $s_{\alpha ,0}:=0$, $s_{\alpha ,1}:=1$, $s_{\alpha ,n+2}:=
s_{\alpha ,n}^{\alpha (n)+1}s_{\alpha ,n+1}^{\alpha (n+1)+1}$. Notice that $s_{\alpha ,n}\prec_{\not=}s_{\alpha ,n+2}$, so that $\beta_{\alpha}:= \hbox{\rm lim}_{n\rightarrow\infty}~s_{\alpha ,2n}\in
2^\omega$ is defined.
$\bullet$ A. Louveau found another example for which it is simpler to check property (M) (and ($\perp$) later), and in the sequel we will work with it. For $\alpha\in 2^\omega$, $n\in\omega$ and $\varepsilon\in 2$, we set $$\gamma_{\alpha}(4n+2\varepsilon ):=\varepsilon\mbox{,}$$ $\gamma_{\alpha}(2n+1):=\alpha (n)$ (so that $\gamma_{\alpha}$ has infinitely many zeros and ones, and the map $\alpha\mapsto\gamma_{\alpha}$ is continuous). For $i\in\omega$, we then set $(i)_{0}:=\hbox{\rm max}\{m\!\in\!\omega /2^m~\hbox{\rm divides}~i+1\}$. Finally, we put $\beta_{\alpha}(i):=\gamma_{\alpha}\big((i)_{0}\big)$.
Notice that the map $\alpha\mapsto\beta_{\alpha}$ is continuous, so that the map $\alpha\mapsto A^{S_{\beta_{\alpha}}}$ is continuous in the codes.
Let $\alpha\in 2^\omega$. Then $S_{\beta_{\alpha}}$ satisfies condition (M). In particular, $A^{S_{\beta_{\alpha}}}$ is minimal among non-potentially closed sets for $\leq^r_{c}$.
[**Proof.**]{} First notice that it is enough to show that the following is fullfilled: $$\forall P\!\in\!\omega~\exists K\!\in\!\omega~
\forall Q\!\in\!\omega~\exists C\!\in\!\omega\!\cap\! [Q,Q\! +\! K]~~
\beta_{\alpha}\lceil P\!\prec\!\beta_{\alpha}\! -\!
\beta_{\alpha}\lceil C.\leqno{(MM)}$$ Indeed, this condition associates $K$ to $P:=p$. Set $k:=2K+1$. For $q\in\omega$, let $Q$ be minimal with $\Sigma_{i<Q}~\big(1+\beta_{\alpha} (i)\big)\geq q$, and fix $C\in\omega\cap
[Q,Q+K]$ such that $\beta_{\alpha}\lceil P\prec\beta_{\alpha}-
\beta_{\alpha}\lceil C$. We put $$c:=\Sigma_{i<C}~\big(1+\beta_{\alpha} (i)\big).$$ Notice that $c\leq q+k$ since $c\!\leq\!\Sigma_{i<Q-1}~\big(1\! +\!\beta_{\alpha} (i)\big)
\! +\!\Sigma_{Q-1\leq i<C}~\big(1\! +\!\beta_{\alpha} (i)\big)\! <\!
q\! +\! 2(C\! -\! Q\! +\! 1)\leq q\! +\! 2(K\! +\! 1)$. Finally, note that $c+\Sigma_{i<j}~\big(1+\beta_{\alpha} (i)\big)=\Sigma_{i<C+j}~\big(1+\beta_{\alpha} (i)\big)$, by induction on $j\leq p$.
Notice that for any integers $n,~i$ and $l$ with $i<2^n-1$, we have $(2^n\cdot l+i)_{0}=(i)_{0}$. Indeed, we can find $N$ with $i=2^{(i)_{0}}(2N+1)-1$, and $(i)_{0}<n$. Thus ${2^n\cdot l+i=2^{(i)_{0}}(2^{n-(i)_{0}}\cdot l+2N+1)-1}$ and $(2^n\cdot l+i)_{0}=(i)_{0}$. Now, if $P\in\omega$, then let $n_{0}$ be minimal with $K:=2^{n_{0}}-1\geq P$. If $Q\in\omega$, then let $l\in\omega\cap [\frac{Q}{2^{n_{0}}},\frac{Q}{2^{n_{0}}}+1[$ and $C:=2^{n_{0}}\cdot l$. If $i<P$, then $i<2^{n_{0}}-1$, so ${(2^{n_{0}}\cdot l+i)_{0}=(i)_{0}}=(C+i)_{0}$. Thus $\beta_{\alpha}(i)=\beta_{\alpha}(C+i)$.
Now we come to the study of the cardinality of complete families of minimal sets.
Let $\big(X,(f_{n})\big),\big(X',(f'_{n})\big)$ be converging situations, and ${u,v:X\!\rightarrow\! X'}$ continuous maps such that $A^{f}=(u\times v)^{-1}(A^{f'})$. Then $u=v$.
[**Proof.**]{} For $x\in X$, fix $x_{k}\in X$ and $n_{k}\in\omega$ such that $\big(x_{k},f_{n_{k}}(x_{k})\big)$ tends to $(x,x)$. Note that $\big(u(x),v(x)\big)\notin A^{f'}$. Moreover, $\big(u[x_{k}],v[f_{n_{k}}(x_{k})]\big)\in A^{f'}$. Thus ${\big(u(x),v(x)\big)\in\overline{A^{f'}}\setminus A^{f'}=\Delta
(X')}$, therefore $u=v$.
Recall that $A^{-1}:=\{(y,x)\in X^2/(x,y)\in A\}$ if $A\subseteq
X^2$.
Fix $S,~S'$ satisfying condition (M). Then
\(a) $A^{S}\perp^r_{c}A^{S'}$, provided that the following condition is fullfilled: $$\exists p\!\in\!\omega ~\forall c\!\in\!\omega ~~c\!
+\!\big(S\!\cap\! [0,p]\big)\!\not=\! S'\!\cap\!\big(c\! +\!
[0,p]\big).\leqno{(\perp )}$$ (b) $A^{S}\perp^r_{c}(A^{S'})^{-1}$, provided that the following condition is fullfilled: $$\exists p\!\in\!\omega ~\forall c\!\in\!\omega ~~c\!
-\!\big(S\!\cap\! [0,p]\big)\!\not=\! S'\!\cap\!\big(c\! -\!
[0,p]\big).\leqno{(\perp^{-1})}$$
[**Proof.**]{} (a) We argue by contradiction: by Lemma 20, we can find continuous maps ${u,~v:2^\omega\!\rightarrow\! 2^\omega}$ such that $A^{S}=(u\times v)^{-1}(A^{S'})$. By Lemma 22, we have $u=v$.
**Claim.* Let $n$, $q$ be integers and $N\in{{\bf\Sigma}^{0}_{1}}(2^\omega)\setminus\{\emptyset\}$. Then we can find integers ${n'> n}$, ${q'\! >\! q}$ and a nonempty open subset $N'$ of $N\cap {f^S_{n'}}^{-1}(N)$ with ${{f^{S'}_{q'}}[u(\gamma )]\! =\! u[f^S_{n'}(\gamma )]}$, and $${\mbox{Card}(\gamma\lceil n')+\big(S\cap [0,p]\big)=
S\cap\big(\mbox{Card}(\gamma\lceil n')+[0,p]\big),}$$ for each $\gamma\in N'$.***
Indeed, let ${\delta\in u[N]}$. As $(\delta ,\delta )$ is not in $\bigcup_{q'\leq q}\hbox{Gr}(f^{S'}_{q'})$, we can find a clopen neighborhood $W$ of $\delta$ such that ${W^2\cap\bigcup_{q'\leq q}
\hbox{Gr}(f^{S'}_{q'})=\emptyset}$. Let $\tilde N\in{{\bf\Delta}^{0}_{1}}(2^\omega )\setminus\{\emptyset\}$ with ${\tilde N\subseteq N\cap u^{-1}(W)}$. By Lemma 20, we can find $n'\! >\! n$ and $\gamma_{0}\!\in\!\tilde N\cap
{f^S_{n'}}^{-1}(\tilde N)$ with $${\hbox{\rm Card}(\gamma_{0}\lceil n')\! +\!\big(S\!\cap\! [0,p]\big)\! =\!
S\!\cap\!\big(\hbox{\rm Card}(\gamma_{0}\lceil n')\! +\! [0,p]\big).}$$ Now there is $q'(\gamma )$ such that ${{f^{S'}_{q'(\gamma
)}}[u(\gamma )] =
u[f^S_{n'}(\gamma )]}$, for ${\gamma\in {\tilde N}\cap
{f^S_{n'}}^{-1}({\tilde N})\cap N_{\gamma_{0}\lceil n'}}$. We have $q'(\gamma )>q$, by the choice of $W$. By Baire’s Theorem we get $q'$ and $N'$.
By the Claim we get $n_1$, $q_{1}$ and $N_{1}\subseteq
D_{f^S_{n_1}}$ with ${f^{S'}_{q_{1}}}[u(\gamma )] = u[f^S_{n_{1}}(\gamma )]$ and $${\hbox{\rm Card}(\gamma\lceil n_{1})+\big(S\cap [0,p]\big)=
S\cap\big(\hbox{\rm Card}(\gamma\lceil n_{1})+[0,p]\big)\mbox{,}}$$ for each ${\gamma\in N_{1}}$.
We then get $n_2>n_{1}$, ${q_{2}>q_1}$, and a nonempty open subset $N_{2}$ of $N_{1}\cap {f^S_{n_{2}}}^{-1}(N_{1})$ with ${{f^{S'}_{q_{2}}}[u(\gamma )] = u[f^S_{n_{2}}(\gamma )]}$ and $${\hbox{\rm Card}(\gamma\lceil n_{2})+\big(S\cap [0,p]\big)=
S\cap\big(\hbox{\rm Card}(\gamma\lceil n_{2})+[0,p]\big)\mbox{,}}$$ for each $\gamma$ in $N_{2}$. We continue in this fashion, until we get $n_{p+1}$, $q_{p+1}$ and $N_{p+1}$. Fix $\gamma\!\in\! N_{p+1}$ and set $c:=\hbox{\rm Card}\big(u(\gamma )\lceil
q_{p+1}\big)$.
$\bullet$ Fix ${m\in S\!\cap\! [0,p]}$. For $t\in\omega^{<\omega}$, we set $f^S_{t}:=f^S_{t(0)}\ldots
f^S_{t(|t|-1)}$, when it makes sense. Notice that ${f^S_{n_{p-m+1},\ldots ,n_{p+1}}(\gamma ) =
f^{S}_{n_{p+1},n_{p-m+1},\ldots ,n_{p}}(\gamma )}$ is defined. Therefore, $A^{S}$ contains $$\big(f^{S}_{n_{p-m+1},\ldots ,n_{p}}(\gamma ), f^{S}_{n_{p-m+1},\ldots ,n_{p+1}}(\gamma )\big)
\mbox{,}$$ which implies that $A^{S'}$ contains ${\big(u[f^{S}_{n_{p-m+1},\ldots ,n_{p}}(\gamma )],
u[f^{S}_{n_{p-m+1},\ldots ,n_{p+1}}(\gamma )]\big)}$. This shows that $A^{S'}$ contains $\big(f^{S'}_{q_{p-m+1},\ldots
,q_{p}}[u(\gamma )],
f^{S'}_{q_{p-m+1},\ldots ,q_{p+1}}[u(\gamma )]\big)$, thus $$f^{S'}_{q_{p-m+1},\ldots ,q_{p+1}}[u(\gamma )]\! =\!
f^{S'}_{q_{p+1},q_{p-m+1},\ldots ,q_{p}}[u(\gamma )]\mbox{,}$$ so ${c+\big( S\cap [0,p]\big)\subseteq
S'\cap\big(c+[0,p]\big)}$.
$\bullet$ Conversely, let $m:=c+m'\in
S'\cap\big(c+[0,p]\big)$. Again $f^S_{n_{p-m'+1},\ldots ,n_{p+1}}(\gamma )$ is defined. Notice that $$u[f^S_{n_{p-m'+1},\ldots ,n_{p+1}}(\gamma )]\! =\!
f^{S'}_{q_{p-m'+1},\ldots ,q_{p+1}}[u(\gamma )]\! =\!
f^{S'}_{q_{p+1},q_{p-m'+1},\ldots ,q_{p}}[u(\gamma )].$$ Therefore $\big(u[f^S_{n_{p-m'+1},\ldots ,n_{p}}(\gamma )],
u[f^S_{n_{p-m'+1},\ldots ,n_{p+1}}(\gamma )]\big)\!\in\! A^{S'}$, $A^{S}$ contains the pair $$\big(f^S_{n_{p-m'+1},\ldots ,n_{p}}(\gamma ),f^S_{n_{p-m'+1},\ldots ,n_{p+1}}(\gamma )\big)\mbox{,}$$ and $f^S_{n_{p-m'+1},\ldots ,n_{p+1}}(\gamma ) = f^S_{n_{p+1},n_{p-m'+1},\ldots ,n_{p}}(\gamma )$. Therefore $\hbox{\rm Card}(\gamma\lceil n_{p+1})+m'\!\in\! S$ and $m'\!\in\! S$, so ${S'\cap\big(c\! +\! [0,p]\big)\subseteq c\! +\!\big( S\cap [0,p]\big)}$. But this contradicts condition [($\perp$)]{} since we actually have the equality.
\(b) The proof is similar to that of (a). This time $A^{S}=(u\times v)^{-1}\big( (A^{S'})^{-1}\big)$. We construct sequences $(n_{j})_{1\leq j\leq p+1}$, $(q_{j})_{1\leq j\leq p+1}$ and $(N_{j})_{1\leq j\leq p+1}$ satisfying the equality ${{f^{S'}_{q_{j}}}^{-1}[u(\gamma )] = u[f^S_{n_{j}}(\gamma )]}$ and $${\hbox{\rm Card}(\gamma\lceil n_{j})+\big(S\cap
[0,p]\big)=S\cap\big(\hbox{\rm Card}(\gamma\lceil n_{j})+[0,p]\big)\mbox{,}}$$ for each ${\gamma\in N_{j}}$. This gives $$(f^{S'}_{q_{p-m+1}})^{-1}\ldots
(f^{S'}_{q_{p+1}})^{-1}[u(\gamma )] =
(f^{S'}_{q_{p+1}})^{-1}
(f^{S'}_{q_{p-m+1}})^{-1}\ldots
(f^{S'}_{q_{p}})^{-1}[u(\gamma )]\mbox{,}$$ thus $c\! -\!\big( S\cap [0,p]\big)\subseteq S'\cap\big(c\! -\! [0,p]
\big)$, and we complete the proof as we did for (a).
Let $\alpha\not= \alpha'\in
2^\omega$. Then $S_{\beta_{\alpha}}$, $S_{\beta_{\alpha'}}$ satisfy conditions (M), ($\perp$) and ($\perp^{-1}$). In particular, $A^{S_{\beta_{\alpha}}}\perp^r_{c}A^{S_{\beta_{\alpha'}}}$ and $A^{S_{\beta_{\alpha}}}\perp^r_{c}(A^{S_{\beta_{\alpha'}}})^{-1}$.
Theorem 5 is a corollary of this result. We saw that the map $\alpha\mapsto A^{S_{\beta_{\alpha}}}$ is continuous in the codes, and it is injective by Corollary 24. This implies that $(A^{S_{\beta_{\alpha}}})_{\alpha\in 2^\omega}$ is a perfect antichain for $\leq^r_{c}$ made of minimal sets (we use Corollaries 21 and 24).
[**Proof.**]{} If $s\in 2^{<\omega}$ and $t\in
2^{\leq\omega}$, we say that $s\subseteq t$ if we can find an integer $l\leq |t|$ such that $s\prec t-t\lceil l$. We define $s^{-1}\in 2^{| s| }$ by $s^{-1}(i):=s(| s| -1-i)$, for $i<| s| $. We say that $s$ is $symmetric$ if $s=s^{-1}$.
$\bullet$ It is enough to prove the following condition: $$\exists P\!\in\!\omega~\ \beta_{\alpha}\lceil
P\!\not\subseteq\!\beta_{\alpha'}\ \ \hbox{\rm and}\ \
(\beta_{\alpha}\lceil
P)^{-1}\!\not\subseteq\!\beta_{\alpha'}.\leqno{(\perp\perp)}$$ Indeed, we will see that $(\perp\perp )$ implies $(\perp )$ and $(\perp^{-1})$ of Theorem 23. Condition $(\perp\perp )$ gives $P>0$. Let $p:=2P$ and $c\in\omega$. We argue by contradiction.
$(\perp )$ Assume that ${c+\big(S_{\beta_{\alpha}}\cap [0,p]\big)=S_{\beta_{\alpha'}}\cap\big(c+[0,p]\big)}$. As $0\in S_{\beta_{\alpha}}$, we can find $l$ with $$c\! =\!\Sigma_{i<l}~\big(1\! +\!\beta_{\alpha'}(i)\big).$$ It is enough to prove that if $n\! <\! P$, then $\beta_{\alpha}(n)\! =\!\beta_{\alpha'}(l\! +\! n)$.
We argue by induction on $n$.
- Notice that $\beta_{\alpha}(0)=0$ is equivalent to $1\in
S_{\beta_{\alpha}}$ and to $\beta_{\alpha'}(l)=0$. Therefore ${\beta_{\alpha}(0)=\beta_{\alpha'}(l)}$.
- Now suppose that $n+1<P$ and $\beta_{\alpha}(m)=\beta_{\alpha'}(l+m)$, for each $m\leq n$. As $$2+\Sigma_{m\leq n}~\big(1+\beta_{\alpha}(m)\big)\leq p\mbox{,}$$ we get $\beta_{\alpha}(n+1)=\beta_{\alpha'}(l+n+1)$.
$(\perp^{-1})$ Assume that ${c-\big(S_{\beta_{\alpha}}\cap
[0,p]\big)=
S_{\beta_{\alpha'}}\cap\big(c-[0,p]\big)}$. Let $l':=l-P$ (as $2P-1$ or $2P$ is in $S_{\beta_{\alpha}}\cap [0,p]$, $c>2P-2$ and $l'\geq
0$). As $(\beta_{\alpha}\lceil P)^{-1}\not\subseteq\beta_{\alpha'}$ we can find $n<P$ such that $\beta_{\alpha}(n)\not=\beta_{\alpha'}(l-1-n)$, since $(\beta_{\alpha}\lceil P)^{-1}\not\prec\beta_{\alpha'}
-\beta_{\alpha'}\lceil l'$. We conclude as in the case $(\perp )$.
$\bullet$ First notice that $\beta_{\alpha}\lceil
(2^n-1)=[\beta_{\alpha}
\lceil (2^n-1)]^{-1}$ for each integer $n$. Indeed, let $i<2^n-1$. It is enough to see that $(i)_{0}=(2^n-2-i)_{0}$. But we have $${2^n-2-i=2^n-2-2^{(i)_{0}}(2N+1)+1=2^{(i)_{0}}(2^{n-(i)_{0}}-2N-1)-1\mbox{,}}$$ so we are done, since $2^{n-(i)_{0}}-2N-1$ is odd and positive. So it is enough to find $n$ such that $\beta_{\alpha}\lceil
(2^n-1)\not\subseteq\beta_{\alpha'}$.
$\bullet$ Let $n_{0}$ minimal with $\gamma_{\alpha} (n_{0})\!\not=\!\gamma_{\alpha'} (n_{0})$, and $n_{1}\! >\! n_{0}\! +\! 1$ with ${\gamma_{\alpha'} (n_{0}+1)\!\not=\!\gamma_{\alpha'} (n_{1})}$. We put $n\!:=\! n_{1} + 2$. We argue by contradiction: we get $l$ with ${\gamma_{\alpha}\big((i)_{0})\! =\!\gamma_{\alpha'}\big((l+i)_{0})}$, for each $i\! <\! 2^n-1$.
$\bullet$ Notice that for each $m<n-1$ we can find $i<2^{n-1}$ with $(l+i)_{0}=m$. Indeed, let $${N\in\omega\cap [2^{-m-1}(l+1)-2^{-1},2^{-m-1}(2^{n-1}+l+1)-2^{-1}[}.$$ It is clear that $i:=2^m(2N+1)-l-1$ is suitable.
$\bullet$ Let $M\geq n_{0}$ and $(\varepsilon_{j})_{j\leq
M}\subseteq 2$ with $l=\Sigma_{j\leq M}~\varepsilon_{j}\cdot 2^j$. For $k\leq n_{0}$ we define $${i_{k}\!:=\!\Sigma_{j<k}~(1\! -\!\varepsilon_{j})\cdot 2^j +\varepsilon_{k}\cdot 2^k}.$$ Note that $i_{k}\! <\! 2^{k+1}$ and ${l\! +\! i_{k}\equiv 2^k\! -\! 1~(\hbox{\rm mod}~2^{k+1})}$. We will show the following, by induction on $k$:
- The sequence $\big(\beta_{\alpha}(i)\big)_{i<2^{n-1},i\equiv
2^k-1~(\hbox{\rm mod}~2^{k+1})}$ is constant with value $\gamma_{\alpha} (k)$, and equal to $${\big(\beta_{\alpha'}(l + i)\big)_{i<2^{n-1},i\equiv
2^k-1~(\hbox{\rm mod}~2^{k+1})}}.$$ - The sequence $\big(\beta_{\alpha'}(l+i)\big)_{i<2^{n-1},i\equiv
i_{k}~(\hbox{\rm mod}~2^{k+1})}$ is constant with value $\gamma_{\alpha'} (k)$.
- The sequence $\big(\beta_{\alpha'}(l+i)\big)_{i<2^{n-1},i\equiv
i_{k}+2^k~(\hbox{\rm mod}~2^{k+1})}$ is not constant.
- $\varepsilon_{k}=0$ and $\gamma_{\alpha} (k)=\gamma_{\alpha'}
(k)$.
This will give the desired contradiction with $k=n_{0}$.
So assume that these facts have been shown for $j<k\leq n_{0}$.
- The first point is clear.
- The second one comes from the fact that $l+i$ is of the form $2^k(2K+1)-1$ if $i\equiv i_{k}~(\hbox{\rm mod}~2^{k+1})$, since ${l+i_{k}\equiv
2^k-1~(\hbox{\rm mod}~2^{k+1})}$.
- To see the third one, choose $i<2^{n-1}$ such that $(l+i)_{0}=n_{0}+1$ (or $n_{1}$). We have to see that $i\equiv i_{k}+2^k~(\hbox{\rm
mod}~2^{k+1})$. We can find $(\eta_{j})_{j<n-1}$ with $i=\Sigma_{j<n-1}~\eta_{j}\cdot 2^j$, so that $$l+i+1\equiv
1+\Sigma_{j<k}~\eta_{j}\cdot 2^j+(\varepsilon_{k}+\eta_{k})\cdot 2^k~
(\hbox{\rm mod}~2^{k+1})\mbox{,}$$ by the induction hypothesis. This inductively shows that $\eta_{j}=1$ if $j<k$ and ${\eta_{k}=1-\varepsilon_{k}}$. Thus $i\equiv
2^k-1+(1-\varepsilon_{k})\cdot 2^k~(\hbox{\rm mod}~2^{k+1})$. But ${i_{k}+2^k\equiv 2^k-1+\varepsilon_{k}\cdot 2^k+2^k~(\hbox{\rm mod}~2^{k+1})}$. Thus $i_{k}+2^k\equiv -1+\varepsilon_{k}\cdot 2^k~(\hbox{\rm mod}~2^{k+1})$. Finally, $2^k-1\equiv i_{k}$ (resp., $i_{k}+2^k$) $(\hbox{\rm mod}~2^{k+1})$ if $\varepsilon_{k}=0$ (resp., $\varepsilon_{k}=1$).
- So $\varepsilon_{k}=0$ and $\gamma_{\alpha} (k)=\gamma_{\alpha'}
(k)$.
This finishes the proof.
Now we prove that $[D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}}),\leq^r_{c}]$ is not well-founded.
[**Notation.**]{} Let ${\cal S}:\omega^\omega\rightarrow\omega^\omega$ be the shift map: ${\cal S}(\alpha )(k):=\alpha (k+1)$, $\beta_{0}$ be the sequence $(0,1,2,\ldots )$, and ${\beta_{n}:={\cal S}^n(\beta
_{0})}$. Notice that $\beta_{n}(i)=i+n$, by induction on $n$. We put $B_{n}:=A^{S_{\beta_{n}}}$.
We have $B_{n+1}\leq^r_{c}B_{n}$ and $B_{n}\not\leq^r_{c}B_{n+1}$ for each integer $n$.
[**Proof.**]{} We define injective continuous maps $u=v:2^\omega\rightarrow 2^\omega$ by ${u(\alpha
):=1^{1+n}\alpha}$. They are clearly witnesses for $B_{n+1}\leq^r_{c}B_{n}$.
$\bullet$ Conversely, we argue by contradiction. This gives continuous maps $u$ and $v$ such that $${B_{n}=(u\times v)^{-1}(B_{n+1})}.$$ By Lemma 22, we have $u=v$. We set ${f^n_{m}:=f^{S_{\beta_{n}}}_{m}}$, and $f^n_{t}:=f^n_{t(0)}\ldots f^n_{t(|t|-1)}$ for $t\in\omega^{<\omega}\setminus\{\emptyset\}$, when it makes sense. Let $\alpha\in N_{0^{n+3}}$, so that ${\alpha = 0^{n+3}\gamma}$.
$\bullet$ If $f^{n}_{t}(\alpha )$ is defined, then fix $m_{t}\!\in\!\omega$ with ${u[f^{{n}}_{t}(\alpha )]\! =\! f^{{n+1}}_{m_{t}}
\big(u[f^{{n}}_{t-t\lceil 1}(\alpha )]\big)}$, and set ${U\!\! :=\! u(\alpha )}$ (with the convention that $f^{n}_{\emptyset}\! :=\!\mbox{Id}_{2^\omega}$). Then ${u[f^{{n}}_{t}(\alpha )]\! =\!
f^{{n+1}}_{m_{t},m_{t-t\lceil 1},\ldots ,m_{t-t\lceil (|t|-1)}}(U)}$. In particular, $$f^{{n+1}}_{m_{(1,\ldots ,n+2)},m_{(2,\ldots ,n+2)},\ldots
,m_{n+2}}(U)=
f^{{n+1}}_{m_{(n+2,1,\ldots ,n+1)},m_{(1,\ldots ,n+1)},\ldots
,m_{n+1}}(U).$$ Therefore $$\{m_{(1,\ldots ,n+2)},m_{(2,\ldots ,n+2)},\ldots ,m_{n+2}\}=
\{m_{(n+2,1,\ldots ,n+1)},m_{(1,\ldots ,n+1)},\ldots ,m_{n+1}\}.$$
If $m_{n+2}=m_{n+1}$, then we get $u(0^{n+2}1\gamma )=u(0^{n+1}10\gamma
)$. As $f^{{n}}_{0}(0^{n+1}10\gamma )=10^n10\gamma$, we get $\big(u(0^{n+1}10\gamma ),v(10^n10\gamma )\big)\in B_{n+1}$ and $(0^{n+2}1\gamma ,10^n10\gamma )\in B_{n}$, which is absurd. Now suppose that $M:=\hbox{\rm max}(m_{(1,\ldots ,n+2)},m_{(2,\ldots ,n+2)},\ldots
,m_{n+2})$ is in $\{m_{n+1},m_{n+2}\}$. Then we can find ${1\leq k\leq n+1}$ such that $$\hbox{\rm Card}\big(U\lceil M\big),\ \hbox{\rm Card}\big(U\lceil M\big)\! +\!
k\in\{\Sigma_{i<l}~\big( 1\!
+\!\beta_{n+1}(i)\big)/l\!\in\!\omega\}.$$ But this is not possible, since ${\Sigma_{i<l+1}~\big( 1\! +\!\beta_{n+1}(i)\big)\!
-\!\Sigma_{i<l}~\big(
1\! +\!\beta_{n+1}(i)\big)\! =\! l\! +\! n\! +\! 2}$.
$\bullet$ We then get the contradiction by induction, since we can remove $M$ from both $$\{m_{(1,\ldots ,n+2)},m_{(2,\ldots ,n+2)},\ldots ,m_{n+2}\}\mbox{,}$$ $\{m_{(n+2,1,\ldots ,n+1)},m_{(1,\ldots ,n+1)},\ldots ,m_{n+1}\}$.
[**Remarks.**]{} (a) We showed that $(A^{S_{\beta_{\alpha}}})_{\alpha\in 2^\omega}$ is a perfect antichain made of sets minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq^r_{c}$. There are other natural notions of reduction. We defined $\leq^r_{c}$ in the introduction. If we moreover ask that $u$ and $v$ are one-to-one, this defines a new quasi-order that we denote $\sqsubseteq^r_{c}$. If $u$ and $v$ are only Borel, we have two other quasi-orders, denoted $\leq^r_{B}$ and $\sqsubseteq^r_{B}$. If $X=Y$, $X'=Y'$ and $u=v$, we get the usual notions $\leq_{c}$, $\sqsubseteq_{c}$, $\leq_{B}$ and $\sqsubseteq_{B}$. Let $\leq$ be any of these eight quasi-orders. Then $(A^{S_{\beta_{\alpha}}})_{\alpha\in 2^\omega}$** is a perfect antichain made of sets minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq$ :**
$\bullet$ Let us go back to Theorem 15 first. Assume this time that $A\leq A^f$. Then in the second case we can have $A^f\cap B^2\sqsubseteq^r_{c}A$ if $\leq$ is rectangular, and $A^f\cap B^2\sqsubseteq_{c}A$ otherwise. The changes to make in the proof are the following. Let $\nu$ (resp., $\nu'$) be a finer Polish topology on $Y$ (resp., $Y'$) making $u$ (resp., $v$) continuous. We get continuous maps $u':2^\omega\rightarrow [Y,\nu ]$ and $v':2^\omega\rightarrow [Y',\nu']$. The proof shows that $f\vert_{G}$ and $g\vert_{G}$ are actually witnesses for $A^f\cap
G^2\sqsubseteq^r_{c}A$ if $\leq$ is rectangular, and $A^f\cap G^2\sqsubseteq_{c}A$ otherwise.
$\bullet$ In Corollary 16, we can replace $\leq^r_{c}$ with $\leq$.
$\bullet$ The proof of Theorem 19 shows that, in its statement, we can write $A^{S}\sqsubseteq_{c} A^{S}\cap B^2$.
$\bullet$ The proof of Lemma 20 shows that, in its statement, we can replace $\leq^r_{c}$ with $\leq$.
$\bullet$ It follows from Corollary 21 that $A^{S_{\beta_{\alpha}}}$ is, in fact, minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq$.
$\bullet$ To see that $(A^{S_{\beta_{\alpha}}})_{\alpha\in 2^\omega}$ is an antichain for $\leq^r_{B}$, it is enough to see that in the statement of Theorem 23, we can replace $\perp^r_{c}$ with $\perp^r_{B}$. We only have to change the beginning of the proof of Theorem 23. This time $u$ and $v$ are Borel. Let $\tau$ be a finer Polish topology on $2^\omega$ making $u$ and $v$ continuous, and $X:=[2^\omega ,\tau ]$. By Lemma 20, $A^{S}$ is $\leq^r_{c}$-minimal, so $(2^\omega, A^{S})\leq^r_{c} (X,
A^{S})\leq^r_{c} A^{S'}$, and we may assume that $u$ and $v$ are continuous.
\(b) We proved that $[D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}}),\leq^r_{c}]$ is not well-founded. Let $\leq$ be any of the eight usual quasi-orders. Then $[D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}}),\leq]$** is not well-founded:**
$\bullet$ The proof of Proposition 25 shows that $B_{n+1}\sqsubseteq_{c}B_{n}$, thus $B_{n+1}\leq B_{n}$.
$\bullet$ We have to see that $B_{n}\not\leq^r_{B}
B_{n+1}$. We argue by contradiction, so that we get $u$ and $v$ Borel.
$\bullet$ Let us show that we can find a dense $G_{\delta}$ subset $G$ of $2^\omega$ such that $u\vert_{G}=v\vert_{G}$ is continuous, and $f^{{n}}_{m}(\alpha )\in G$, for each $\alpha\in G\cap
D_{f^{{n}}_{m}}$.
[**Claim.**]{} *The set ${H:=\{\alpha\!\in\! 2^\omega /\forall
p~\exists m\!\geq\! p\ \ \alpha\!\in\! D_{f^{{n}}_{m}}\}}$ is a dense $G_{\delta}$ subset of $2^\omega$.*
We argue by contradiction. We can find a nonempty clopen set $V$ disjoint from $H$. The set $B_{n}\cap V^2$ has finite sections, so is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$ (see Theorem 3.6 in \[Lo1\]). But $\big(V,({f^n_{m}}|_{V\cap
{f^n_{m}}^{-1}(V)})\big)$ is a converging situation, so that $B_{n}\!\cap\! V^2$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
So we can find a dense $G_{\delta}$ subset $K$ of $2^\omega$ such that $u\vert_{K}$, $v\vert_{K}$ are continuous and $K\subseteq H$. Now let $K_{0}:=K$, $K_{p+1}:=K_{p}\setminus\big(\bigcup_{m}D_{f^{{n}}_{m}}\setminus
{f^{{n}}_{m}}^{-1}(K_{p})\big)$, and ${G:=\bigcap_{p}K_{p}}$. If $\alpha\in K_{1}$, fix $(m_{k})$ infinite such that $\alpha\in\bigcap_{k}D_{f^{{n}}_{m_{k}}}$. We have $f^{{n}}_{m_{k}}(\alpha )\in K_{0}$, so $\big(u(\alpha
),v[f^{{n}}_{m_{k}}
(\alpha )]\big)$ tends to ${\big(u(\alpha ),v(\alpha )\big)\in
\overline{B_{n+1}}\setminus B_{n+1}=\Delta (2^\omega )}$. So $u\vert_{K_{1}}=v\vert_{K_{1}}$. Now it is clear that $G$ is suitable.
$\bullet$ We take $\alpha\in G\cap N_{0^{n+3}}$ and complete the proof as we did for Proposition 25.
[**Proof of Theorem 6.**]{} We will actually prove a stronger statement. We set $$(P_{0},P_{1},P_{2},P_{3},P_{4}):=( \hbox{\rm
reflexive,~irreflexive,~symmetric,~antisymmetric,~transitive}).$$ Let $\sigma\in 2^5\setminus\{\{2,4\},\{0,2,4\}\}$ such that the class $\Gamma$ of ${{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$ relations satisfying $\wedge_{j\in\sigma}~P_{j}$ is not empty. Then we can find a perfect $\leq_{B}$-antichain $(R_{\alpha})_{\alpha\in 2^\omega}$ in $D_{2}({{\bf\Sigma}^{0}_{1}})\cap\Gamma$ such that $R_{\alpha}$ is $\leq_{B}$-minimal among ${{\bf\Delta}^{1}_{1}}\setminus
\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets, for any $\alpha\in 2^\omega$.
$\bullet$ First, notice that if $\{0,1\}\subseteq\sigma$ or $\sigma =\{1,2,4\}$, then every relation satisfying $\wedge_{j\in\sigma} P_{j}$ is empty, thus $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. If $\{2,3\}\subseteq\sigma$, then every Borel relation satisfying $\wedge_{j\in\sigma} P_{j}$ is a subset of the diagonal, and is therefore $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. If $\sigma =\{0,2,4\}$, we are in the case of Borel equivalence relations, and by Harrington, Kechris and Louveau’s Theorem, $E_{0}$ is minimum among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ equivalence relations. If $\sigma =\{2,4\}$, then any Borel relation $A\subseteq X^2$ satisfying $\wedge_{j\in\sigma} P_{j}$ is reflexive on its domain $\{x\in X/(x,x)\in A\}$, which is a Borel set. Thus we are reduced to the case of equivalence relations. In the sequel, we will avoid these cases and show the existence of a perfect antichain made of minimal sets for $[\Gamma
,\leq_{B}]$.
$\bullet$ Let $\mathbb{A}\!:=\!\{A^{S_{\beta_{\alpha}}}/\alpha\!\in\!
2^\omega\}$. In the introduction, we defined $R_{A}$ for $A\subseteq 2^\omega\times 2^\omega$.
[**Claim 1.**]{} *$\{ R_{A}/A\!\in\! {\mathbb{A}}\}$ is a $\leq_B$-antichain.*
Assume that ${A\!\not=\! A'\!\in\! {\mathbb{A}}}$ satisfy $R_{A}\leq_B R_{A'}$. Then there is ${f:2^\omega\times 2\rightarrow 2^\omega\!\times\! 2}$ with $${R_{A}=(f\!\times\! f)^{-1}(R_{A'}).}$$ We set $F_\varepsilon\! :=\!\{x\!\in\! 2^\omega\!\times\! 2/x_1\! =\!\varepsilon\}$ and ${b_{\varepsilon}\! :=\! R_{A}\cap
(F_\varepsilon\!\times\! F_\varepsilon )}$, for $\varepsilon\!\in\! 2$. We then put ${a\!:=\! R_{A}\cap (F_0\!\times\! F_1)}$. We have $R_{A}\! =\! a\cup b_{0}\cup b_{1}$, and ${b_{\varepsilon}\! =\!\{(x,y)\!\in\! F_{\varepsilon}\!\times\!
F_{\varepsilon}/x_{0}\!
=\! y_{0}\}}$ is $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. We set $$F'_\varepsilon:=\{x\in 2^\omega\times 2/f_1(x)=\varepsilon\}$$ and ${b'_{\varepsilon}:=R_{A}\cap (F'_\varepsilon\times F'_\varepsilon )}$, for $\varepsilon\in 2$.
We then put ${a':=R_{A}\cap (F'_0\times F'_1)}$. We have $R_{A}=a'\cup b'_{0}\cup b'_{1}$, and $${b'_{\varepsilon}=(f\vert_{F'_{\varepsilon}}\times f\vert_{F'_{\varepsilon}})^{-1}
\big(\Delta (2^\omega\times 2)\big)\in \mbox{pot}({{\bf\Pi}^{0}_{1}})}.$$ Notice that $A\leq_{B}^r R_{A}$, so that $R_{A}$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. So ${R_{A}=(a\cap a')\cup b_{0}\cup b_{1}\cup b'_{0}\cup b'_{1}}$, and $a\cap a'$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. It remains to define $C:=a\cap a'$, viewed as a subset of ${(F_0\cap F'_0)\times (F_1\cap F'_1)}$. We equip $F_0\cap F'_0$ (resp., $F_1\cap F'_1$) with a finer Polish topology making $f\vert_{F_0\cap F'_0}$ (resp., $f\vert_{F_1\cap F'_1}$) continuous. Then $C\!\leq^r_{c}\!
A$ and $C\!\leq^r_{c}\! A'$, which contradicts Corollaries 21 and 24.
[**Claim 2.**]{} *Let $A\! =\! A^{S_{\beta_{\alpha}}}\!\in\!
{\mathbb{A}}$. Then $R_{A}$ is minimal for $\leq_{B}$ among ${{\bf\Delta}^{1}_{1}}\setminus
\mbox{pot}({{\bf\Pi}^{0}_{1}})$ relations.*
- Assume that $R\leq_{B} R_{A}$. This gives ${f:X\rightarrow 2^\omega\!\times\! 2}$ Borel with ${R\! =\! (f\!\times\! f)^{-1}(R_{A})}$. Again we set ${F_{\varepsilon}:=\{x\in X/f_{1}(x)=\varepsilon\}}$ for $\varepsilon\in 2$, and we see that $R\cap (F_{0}\times F_{1})$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$.
- Let $\tau$ be a finer Polish topology on $X$ making $f$ continuous. By Theorem 9 there are $${u:2^\omega\rightarrow [F_{0},\tau ]\mbox{,}}$$ ${v:2^\omega\rightarrow [F_{1},\tau ]}$ continuous with ${A_{1}\! =\! (u\!\times\! v)^{-1}\big(R\cap (F_{0}\!\times\!
F_{1})\big)\cap\overline{A_{1}}}$. We define ${H:=f_{0}\big[u[2^\omega ]\big]}$, ${K:=f_{0}\big[v[2^\omega ]\big]}$ and $P:=H\cap K$; this defines compact subsets of $2^\omega$. Then ${A\cap (H\times K)}$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$ since $${A_{1}\! =\! [(f_{0}\circ u)\!\times\! (f_{0}\circ
v)]^{-1}\big(A\cap (H\times K)\big)\cap\overline{A_{1}}.}$$ As in the proof of Theorem 15, this implies that ${A\cap P^2}$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. By Lemma 14, we can find a Borel subset $S$ of $P$ and a finer topology $\sigma$ on $S$ such that $\big( [S,\sigma
],(f^{S}_{n})_n\big)$ is a converging situation.
- By 18.3 in \[K\], we can find a Baire measurable map ${g_{\varepsilon}:S\rightarrow f^{-1}(S\times\{\varepsilon\})}$ such that $$f_{0}\big(g_{\varepsilon}(\alpha )\big)=\alpha\mbox{,}$$ for $\alpha$ in $S$ and $\varepsilon\in 2$. Let $G$ be a dense $G_{\delta}$ subset of $S$ such that each ${g_{\varepsilon}\vert}_{G}$ is continuous. Now we define $F:G\times 2\rightarrow X$ by ${F(\alpha ,\varepsilon ):=g_{\varepsilon}(\alpha )}$. Then $R_{A}\cap (G\times 2)^2=(F\times F)^{-1}(R)$, so $R_{A}\cap (G\times 2)^2\leq_{B}R$. As in the proof of Theorem 15, we see that $A\cap G^2$ is not $\mbox{pot}({{\bf\Pi}^{0}_{1}})$. But $A\cap
G^2\sqsubseteq_{c}A$. By Remark (a) above, we get $A\sqsubseteq_{c}A\cap G^2$. Thus $R_{A}\sqsubseteq_{c}R_{A}\cap (G\times 2)^2$ and ${R_{A}\leq_{B}R}$.
Finally, one easily checks the existence of a continuous map $c:2^\omega\rightarrow 2^\omega$ such that $c(\delta )$ is a Borel code for $R_{A}$ if $\delta$ is a Borel code for $A$. So there is a continuous map ${r:2^\omega\rightarrow 2^\omega}$ such that $r(\alpha )$ is a Borel code for $R_{A^{S_{\beta_{\alpha}}}}$. This shows, in particular, the existence of a perfect antichain made of minimal sets for ${[{{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})~ \hbox{\rm quasi-orders},\leq_{B}]}$ and ${[{{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})~ \hbox{\rm partial~orders},\leq_{B}]}$. More generally, this works if $\sigma\subseteq\{0,3,4\}$.
$\bullet$ Similarly, we define, for $A\subseteq X^2$, a strict partial order relation $R'_A$ on $X\times 2$ by $$(x,i)~R'_A~(y,j)~~\Leftrightarrow ~~[(x,y)\in A~ \hbox{\rm and}~
i=0~ \hbox{\rm and}~j=1].$$ The proof of the previous point shows that if $\sigma\subseteq\{1,3,4\}$, then $\{R'_{A}/A\in {\mathbb{A}}\}$ is a perfect antichain made of minimal sets for $[\Gamma ,\leq_{B}]$. Notice that this applies when $\Gamma$ is the class of ${{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$ strict quasi-orders, strict partial orders, directed graphs or oriented graphs.
$\bullet$ Similarly again, we can define, for $A\subseteq
X^2$, $S_A$ reflexive symmetric on $X\times 2$ by $$\begin{array}{ll}
(x,i)~S_A~(y,j)~~\Leftrightarrow
& (x,i)=(y,j)~ \hbox{\rm or}~
[(x,y)\in A~ \hbox{\rm and}~i=0~ \hbox{\rm and}~j=1]~ \hbox{\rm
or}~\cr
& [(y,x)\in A~ \hbox{\rm and}~i=1~ \hbox{\rm and}~j=0].
\end{array}$$ Let $A_{0}:=A$ and $A_{1}:={A}^{-1}$. The proof of Claim 1 shows that if $A\not= A'\in {\mathbb{A}}$ satisfy $S_{A}\leq_B S_{A'}$, then we can find $C\notin \mbox{pot}({{\bf\Pi}^{0}_{1}})$ and $\varepsilon ,~\varepsilon'\in
2$ such that $C\leq^r_{c} A_{\varepsilon}$ and $C\leq^r_{c}
A'_{\varepsilon'}$. But this contradicts Corollaries 21 and 24. This shows that if $\sigma=\{0,2\}$, then $\{S_{A}/A\in {\mathbb{A}}\}$ is a perfect antichain made of minimal sets for $[\Gamma ,\leq_{B}]$.
$\bullet$ Similarly again, we can define, for $A\subseteq
X^2$, a graph relation $S'_A$ on $X\times 2$ by $$(x,i)~S'_A~(y,j)~\Leftrightarrow ~
[(x,y)\!\in\! A~ \hbox{\rm and}~i\! =\! 0~ \hbox{\rm and}~j\! =\! 1]~
\hbox{\rm or}~
[(y,x)\!\in\! A~ \hbox{\rm and}~i\! =\! 1~ \hbox{\rm and}~j\! =\! 0].$$ The proof of the previous point shows that if $\sigma\subseteq\{1,2\}$, then $\{S'_{A}/A\in {\mathbb{A}}\}$ is a perfect antichain made of minimal sets for $[\Gamma ,\leq_{B}]$. Notice that this applies when $\Gamma$ is the class of ${{\bf\Delta}^{1}_{1}}\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$ graphs. This finishes the proof.
[**Remarks.**]{} (a) We showed that $(R_{A})_{A\in {\mathbb{A}}}$ is a perfect antichain made of sets minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq_{B}$. Fix $\leq$ in $\{\leq_{c},\sqsubseteq_{c},\leq_{B},\sqsubseteq_{B}\}$. Then $(R_{A})_{A\in {\mathbb{A}}}$** is a perfect antichain made of sets minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq$. It is enough to check the minimality. The only thing to notice, in the proof of Claim 2 of the proof of Theorem 6, is that we have $R_{A}\cap (G\times 2)^{2}\sqsubseteq_{c}R$ and $R_{A}\sqsubseteq_{c} R$. Similarly, $R'_{A}$, $S^{}_{A}$ and $S'_{A}$ ($A\in {\mathbb{A}}$) are minimal for $\leq_{c}$, $\sqsubseteq_{c}$ and $\sqsubseteq_{B}$.**
\(b) We have $\neg\Delta (2^\omega )\perp^r_{B}L_{0}$. Indeed, assume that $\neg\Delta (2^\omega )=(u\times v)^{-1}(L_0)$. Then $u(\alpha )<_{\hbox{\rm lex}} v(\beta )$ if $\alpha\not=\beta$, and $v(\alpha )\leq_{\hbox{\rm lex}} u(\alpha )$. Thus $$u(\beta )<_{\hbox{\rm lex}} v(\alpha )\leq_{\hbox{\rm lex}} u(\alpha )
<_{\hbox{\rm lex}}v(\beta )\leq_{\hbox{\rm lex}} u(\beta )\mbox{,}$$ which is absurd. Now assume that $L_0=(u\times v)^{-1}\big(\neg\Delta (2^\omega )\big)$. Then $\beta\leq_{\hbox{\rm lex}}\alpha$ implies $u(\alpha ) = v(\beta )$, thus $u=v$ has to be constant. Thus $\alpha<_{\hbox{\rm lex}}\beta$ implies that $u(\alpha )$ and $v(\beta )$ are different and equal.
In the introduction, we saw that $\{\neg\Delta (2^\omega ),L_{0}\}$ is a complete family of minimal sets for $${[\mbox{pot}\big(\check
D_2({{\bf\Sigma}^{0}_{1}})\big)\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}}),\sqsubseteq^r_{c}].}$$ We just saw that $\{\neg\Delta (2^\omega ),L_{0}\}$ is an antichain for $\leq^r_{B}$, and therefore for any of the eight usual quasi-orders. These facts imply that $\neg\Delta (2^\omega )$ and $L_{0}$ are minimal among non-$\mbox{pot}({{\bf\Pi}^{0}_{1}})$ sets for $\leq^r_{c}$, $\sqsubseteq^r_{c}$, $\leq^r_{B}$ and $\sqsubseteq^r_{B}$. But $\neg\Delta (2^\omega )$ and $L_{0}$ are also minimal for $\leq_{c}$, $\sqsubseteq_{c}$, $\leq_{B}$ and $\sqsubseteq_{B}$. Indeed, if $O$ is any of these two open sets, we have $\overline{O}\setminus O=\Delta
(2^\omega )$. This gives $G$ such that $O\cap G^2\sqsubseteq_{c}A$, as in the proof of Theorem 15 (and Remark (a) after Proposition 25). Then any increasing continuous injection $\phi: 2^\omega\rightarrow G$ is a witness to $O\sqsubseteq_{c}O\cap G^{2}$.
**[5 The minimality of $A_{1}$ for the classical notions of comparison.]{}**
As announced in the introduction, we will show a result implying that $A_{1}$ is minimal among non-potentially closed sets. The following definition specifies the meaning of the expression “the $f_{n}$’s do not induce cycles" mentioned in the introduction. This kind of notion has already been used in the theory of potential complexity (see Definition 2.10 in \[L3\]).
We say that $\big(X,(f_{n})\big)$ is an $acyclic\ situation$ if
\(a) $\big(X,(f_{n})\big)$ is a converging situation, with only $\Delta (X)\subseteq\overline{A^f}\setminus A^f$ in condition (c).
\(b) For $v\in\omega^{<\omega}\setminus\{\emptyset\}$ and $\varepsilon\in\{-1,1\}^{|v|}$, the following implication holds:
**Notation. We define $f^1_{n}:N_{s_{n}0}\rightarrow N_{s_{n}1}$ by $f^1_{n}(s_{n}0\gamma ):=s_{n}1\gamma$ (where $s_n$ is as defined in the introduction, to build $A_1=\bigcup_n\ \hbox{Gr}(f^1_n)$).**
Let $\alpha\in 2^\omega$, $v\in\omega^{<\omega}\setminus\{\emptyset\}$ and $\varepsilon\in\{-1,1\}^{|v|}$. Assume that ${v(i)\not= v(i+1)}$ or $\varepsilon (i)\not= -\varepsilon (i+1)$ if $i<|v|-1$. Then ${f^{1}_{v(|v|-1)}}^{\varepsilon (|v|-1)}\ldots {f^{1}_{v(0)}}^{\varepsilon (0)}
(\alpha )$ is either undefined, or of value different than $\alpha$.
**Proof. We argue by contradiction. Let $v$ be a counter-example of minimal length. Note that $|v|\geq 3$. Set $l:=\hbox{\rm max}_{i<|v|}~v(i)$, $e_0:=e_{|v|}:=\alpha\lceil (l\! +\! 1)$, and, for $0<i<|v|$: $$e_i:=[{f^{1}_{v(i-1)}}^{\varepsilon (i-1)}\ldots {f^{1}_{v(0)}}^
{\varepsilon (0)}(\alpha )]~\lceil ~(l\! +\! 1).$$ Set $\Theta :=(\theta_n)$, where $\theta_n:=s_n$. Then $(e_i)_{i\leq |v|}$ is an $s({\mathfrak R}_\Theta )$-cycle, which contradicts Proposition 18.(b).**
**Example. $\big(2^\omega ,(f^1_{n})\big)$ is an acyclic situation. Indeed, $\big(2^\omega ,(f^{1}_n)\big)$ is a converging situation, by Corollary 12. Let us show that condition (b) in the definition of an acyclic situation is true for $\big(2^\omega ,(f^{1}_n)\big)$. The domain $D$ of ${f^{1}_{v(|v|-1)}}^{\varepsilon (|v|-1)}\!\!\!\!\ldots
{f^{1}_{v(0)}}^{\varepsilon (0)}$ is clopen. If $U$ is not included in $D$, then we can take $V:=U\setminus
D$. Otherwise, let $\alpha\!\in\! U$. By Lemma 27, and by continuity, we can find a clopen neighborhood $V$ of $\alpha$ included in $U$ such that ${f^{1}_{v(|v|-1)}}^{\varepsilon (|v|-1)}\ldots
{f^{1}_{v(0)}}^{\varepsilon (0)}[V]\cap V=\emptyset$.**
Let $\big(X,(f_{n})\big)$ be an acyclic situation. Then $A_{1}\leq^r_{c}A^f$.
[**Proof.**]{} It looks like those of Theorems 2.6 and 2.12 in \[L3\]. The main difference is that we want a reduction defined on the whole product. It is also similar to the proof of Theorem 19. Let us indicate the differences with the proof of Theorem 19. We replace $A^S=\bigcup_n\
\hbox{Gr}(f^S_n)$ with $A_1=\bigcup_n\ \hbox{Gr}(f^1_n)$.
$\bullet$ We only construct $(U_{s})_{s\in 2^{<\omega}}$ and $\phi$, so that (iii) becomes $$(iii)\ (s~{\mathfrak R}~t~\hbox{\rm and}~s\!\not=\! t)\Rightarrow
U_{t}\! =\! f_{\phi (|s\wedge t|)}[U_{s}].$$ $\bullet$ Here we choose $\Theta =(\theta_n)$ with $\theta_n:=s_n$. Notice that ${\mathfrak R}_\Theta ={\mathfrak R}$.
$\bullet$ Condition $(3)$ becomes $$(3)\ (t_{k}~{\mathfrak R}~t_{l}~\hbox{\rm and}~t_{k}\!\not=\!
t_{l})\Rightarrow
U^n_{l}\! =\! f_{\phi (|t_{k}\wedge t_{l}|)}[U^n_{k}].$$
$\bullet$ We can find $C\in{{\bf\Delta}^{0}_{1}}(U_{t_{0}\lceil
p})\setminus\{\emptyset\}$ such that $C^{2}\cap\bigcup_{q\leq p}~\hbox{Gr}(f_{q})=\emptyset$, and also $${{n_{0}}\geq\hbox{\rm sup}~\{\phi (q)\! +\! 1/q<p\}}$$ with ${C^{2}\cap \hbox{Gr}(f_{n_{0}})\!\not=\!\emptyset}$, since ${\Delta (X)\subseteq\overline{A^f}\setminus A^f}$. We set $\phi (p):={n_{0}}$. We first construct clopen sets $\tilde U^n_k$ as in the proof of Theorem 19.
[**Case 2.**]{} $t_{n}\lceil p\not=t_{r}\lceil p$.
2.1. $t_{r}~{\mathfrak R}_\Theta ~t_{n}$.
To get condition (4), fix $k,l\!\leq\! n$ with $\neg ~t_{k}~{\mathfrak R}~t_{l}$. Set ${f^\varepsilon_{v}\!:=\! {f_{\phi (v(|v|-1))}^{\varepsilon (|v|-1)}\ldots
f_{\phi (v(0))}^{\varepsilon (0)}}}$, so that ${\tilde U^n_{l}={f^{\varepsilon}_{v}}[\tilde U^n_{k}]}$, and we have $\phi\big( v(i)\big)\not=\phi\big( v(i+1)\big)$, since $\phi$ is strictly increasing. As $\big(X,(f_{n})\big)$ is without cycles, we can find $x\in\tilde
U^n_k$ with $f^{\varepsilon}_{v}(x)\not= x$. We can therefore find a clopen neighborhood $\underline{U}^n_k$ of $x$, included in $\tilde U^n_k$, such that $\underline{U}^n_k\cap f^{\varepsilon}_{v}[\underline{U}^n_k]
=\emptyset$. We construct clopen sets $\underline{U}^n_r$, for $k\not= r\leq n$, as before, ensuring condition (3). Notice that $\underline{U}^n_r\subseteq\tilde
U^n_r$, so that the hereditary conditions (1), (2) and (5) remain fullfilled. In finitely many steps we get $(\underline{U}^n_k\times\underline{U}^n_l)\cap\Delta (X)=\emptyset$, for each pair $(k,l)$. The argument is similar for $\hbox{Gr}(f_{q})$ instead of $\Delta (X)$.
2.2. $t_{n}~{\mathfrak R}_\Theta ~t_{r}$.
This case is similar to case 2.1.
[**Remark.**]{} We actually showed that $A_{1}\sqsubseteq_{c}A^f$.
$A_{1}$ is minimal among non-potentially closed sets for the eight usual quasi-orders.
[**Proof.**]{} Let $B\!\in\!{{\bf\Delta}^{1}_{1}}(2^\omega )$, $\tau$ a finer topology on $B$, $Z\!:=\! [B,\tau ]$ and ${f_{n}\!:=\! {f^{1}_{n}}\vert_{B\cap {f^{1}_{n}}^{-1}(B)}}$. We assume that $\big(Z,(f_{n})\big)$ is a converging situation. By Corollary 16 and Remark (a) after Proposition 25, it is enough to show that ${A_{1}\sqsubseteq_{c}A_{1}\cap Z^{2}=A^f}$. By Theorem 28 and the remark above, it is enough to check that $\big(Z,(f_{n})\big)$ is an acyclic situation, i.e., condition (b). Fix $\alpha\!\in\! U$ and $f_{v}^\varepsilon:=f_{v(|v|-1)}^{\varepsilon (|v|-1)}\ldots
f_{v(0)}^{\varepsilon (0)}$. If $U$ is not included in $D_{f_{v}^\varepsilon}$, then we can take $V:=U\setminus D_{f_{v}^\varepsilon}$, because the domain is a clopen subset of $Z$. As $f_{v}^\varepsilon$ is continuous, it is enough to see that $f_{v}^\varepsilon (\alpha )\not=\alpha$, if $U$ is included in $D_{f_{v}^\varepsilon}$. But this is clear, since ${f^{1}_{v(|v|-1)}}^{\varepsilon (|v|-1)}\ldots
{f^{1}_{v(0)}}^{\varepsilon (0)}(\alpha )$ is different from $\alpha$, by Lemma 27.
[**Remarks.**]{} (a) Theorem 28 is also a consequence of the following result:
(Miller) Let $X$ be a Polish space, and $A$ a locally countable ${{\bf\Sigma}^{1}_{1}}$ oriented graph on $X$ whose symmetrization is acyclic (in the sense of Definition 17). Then exactly one of the following holds:
\(a) $A$ has countable Borel chromatic number.
\(b) $A_{1}\sqsubseteq_{c}A$.
Theorem 30 is actually a corollary of a more general result, motivated by the results of this paper, which gives a basis for locally countable Borel directed graphs of uncountable Borel chromatic number, with respect to $\sqsubseteq_{c}$. The proof of both Theorem 30 and the basis result appear in \[M1\].
\(b) We saw that $A_{1}\sqsubseteq_{c}A^f$ if $\big(X,(f_{n})\big)$ is an acyclic situation. There is another example of a $$D_{2}({{\bf\Sigma}^{0}_{1}})\setminus \mbox{pot}({{\bf\Pi}^{0}_{1}})$$ set, which seems more “natural" than $A_{1}$. It is $$C_{1}\! :=\!\{(\alpha ,\beta )\!\in\! 2^\omega\!\times\! 2^\omega
/\exists s\!\in\! 2^{<\omega}~\exists\gamma\!\in\! 2^\omega ~~
(\alpha ,\beta )\! =\! (s0\gamma ,s1\gamma )\}.$$ Its symmetric version plays an important role in the theory of potential complexity (see for example Theorem 3.7 and Corollary 4.14 in \[L1\]). We wonder what $\{ C_{1}\}$ is a basis for. Roughly speaking, $\{ C_{1}\}$ will be a basis for situations where commuting relations between the $f_{n}$’s are involved. More specifically,
We say that $\big(X,(f_{n})\big)$ is a $commuting~situation$ if
\(a) $X$ is a nonempty perfect closed subset of $\omega^\omega$.
\(b) $f_{n}$ is a partial homeomorphism with disjoint ${{\bf\Delta}^{0}_{1}}(X)$ domain and range. Moreover $\alpha <_{\mbox{lex}} f_{n}(\alpha )$ if $\alpha\in D_{f_{n}}$.
\(c) $\Delta (X)\subseteq\overline{A^f}\setminus A^f$, and $A^f\in{{\bf\Pi}^{0}_{2}}(X^{2})$.
\(d) For each $\alpha\in f_{m}^{-1}(D_{f_{n}})$ we have $\alpha\in f_{n}^{-1}(D_{f_{m}})$ and $f_{m}\big(f_{n}(\alpha )\big)=f_{n}\big(f_{m}(\alpha )\big)$. Moreover the graphs of the $f_{n}$’s are pairwise disjoint.
A $0$-dimensional Polish space is homeomorphic to a closed subset of $\omega^\omega$. So condition (a) is essentially the same as condition (a) of a converging situation. We use this formulation for the last part of condition (b). The disjunction of the domain and the range of $f_{n}$, and the inequality $\alpha <_{\hbox{\rm lex}} f_{n}(\alpha )$ come from symmetry problems. We will come back later to this. We will also come back to the ${{\bf\Pi}^{0}_{2}}$ condition. It is linked with transitivity properties. The first part of condition (d) expresses the commutativity of the functions. One can show the following result, whose proof contains a part quite similar to the proof of Theorems 19 and 28.
Let $\big(X,(f_{n})\big)$ be a commuting situation. Then $C_{1}\sqsubseteq_{c}A^f$.
The proof of this uses the fact that $C_{1}=A^{f}$, where $\big(2^\omega ,(f_{n})\big)$ is a commuting situation. Let $$g_{n}:2^\omega\rightarrow 2^\omega$$ be defined by $g_{n}(\alpha )(k):=\alpha (k)$ if $k\not= n$, $1-\alpha (n)$ otherwise. Then $s(C_{1})=\bigcup_{n}~\hbox{\rm Gr}(g_{n})$, so $\big(2^\omega ,(g_{n})\big)$ is not a commuting situation, since otherwise we would have $C_{1}\sqsubseteq_{C} s(C_{1})$, which is absurd since $s(C_{1})$ is symmetric and $C_{1}$ is not. But the two reasons for that are that $\alpha\not <_{\hbox{\rm lex}} g_{n}(\alpha )$, and that the domain and the range of the bijections $g_{n}$ are not disjoint.
Similarly, let $\phi :\omega\rightarrow P_{f}\setminus\{ 0^\infty\}$ be a bijective map. We let $g'_{n}(\alpha )(p)\! :=\!\alpha (p)$ if $\phi (n)(p)\! =\! 0$, $1$ otherwise. This defines ${g'_{n}\! :\!\{\alpha\!\in\! 2^\omega\! /\forall p~\phi (n)(p)\! =\! 0~or~
\alpha (p)\! =\! 0\}\!\rightarrow\! 2^\omega\! }$. Note that $${E_{0}\!\cap\! L'_{0}\! =\!\bigcup_{q} \hbox{\rm Gr}(g'_{n})\mbox{,}}$$ where $L'_{0}\! :=\!\{ (\alpha ,\beta )\!\in\! 2^\omega\!\times\! 2^\omega\! /
\forall i\!\in\!\omega ~\alpha (i)\!\leq\!\beta
(i)~\mbox{and}~\alpha\!\not=\!\beta\}$.
Then $\big(2^\omega ,(g'_{n})\big)$ is not a commuting situation, since otherwise $C_{1}\sqsubseteq_{C} E_{0}\cap L'_{0}$, which is absurd since $E_{0}\cap L'_{0}$ is transitive and $C_{1}$ is not. But the reason for that is that $E_{0}\cap L'_{0}\notin{{\bf\Pi}^{0}_{2}}$.
B. D. Miller has also a version of Theorem 32 for directed graphs of uncountable Borel chromatic number (in \[M2\]). Its proof uses some methods analogous to those in the proof of Theorem 30. All of this shows the existence of numerous analogies between non potentially closed directed graphs and directed graphs of uncountable Borel chromatic number.
**[6 References.]{}**
\[H-K-Lo\] L. A. Harrington, A. S. Kechris and A. Louveau, *A Glimm-Effros dichotomy for Borel equivalence relations, J. Amer. Math. Soc. 3 (1990), 903-928*
\[K\] A. S. Kechris, *Classical Descriptive Set Theory, Springer-Verlag, 1995*
\[K-S-T\] A. S. Kechris, S. Solecki and S. Todorčević, *Borel chromatic numbers, Adv. Math. 141 (1999), 1-44*
\[L1\] D. Lecomte, *Classes de Wadge potentielles et théorèmes d’uniformisation partielle, Fund. Math. 143 (1993), 231-258*
\[L2\] D. Lecomte, *Uniformisations partielles et critères à la Hurewicz dans le plan, Trans. A.M.S. 347, 11 (1995), 4433-4460*
\[L3\] D. Lecomte, *Tests à la Hurewicz dans le plan, Fund. Math. 156 (1998), 131-165*
\[L4\] D. Lecomte, *Complexité des boréliens à coupes dénombrables, Fund. Math. 165 (2000), 139-174*
\[Lo1\] A. Louveau, *A separation theorem for ${{\it\Sigma}^{1}_{1}}$ sets, Trans. A. M. S. 260 (1980), 363-378*
\[Lo2\] A. Louveau, *Ensembles analytiques et boréliens dans les espaces produit, Astérisque (S. M. F.) 78 (1980)*
\[Lo-SR\] A. Louveau and J. Saint Raymond, *Borel classes and closed games : Wadge-type and Hurewicz-type results, Trans. A. M. S. 304 (1987), 431-467*
\[M1\] B. D. Miller, *Basis theorems for graphs of uncountable Borel chromatic number, pre-print*
\[M2\] B. D. Miller, *A dichotomy theorem for graphs induced by commuting families of Borel injections, pre-print*
\[M\] Y. N. Moschovakis, *Descriptive set theory, North-Holland, 1980*
\[SR\] J. Saint Raymond, *La structure borélienne d’Effros est-elle standard ?, Fund. Math. 100 (1978), 201-210*
**Acknowledgements. I would like to thank A. Louveau for his interest for this work, and also for some nice remarks that can be found in this paper. I thank B. D. Miller who helped me to improve the English in this paper while I was visiting UCLA during June 2005. I also thank B. D. Miller for some nice remarks that improved the quality of this article.**
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We are concerned with the Cauchy problem for the KdV equation for nonsmooth locally integrable initial profiles $q$’s which are, in a certain sense, essentially bounded from below and $q\left( x\right) =O\left(
e^{-cx^{\varepsilon }}\right) ,x\rightarrow +\infty $, with some positive $c$ and $\varepsilon $. Using the inverse scattering transform, we show that the KdV flow turns such initial data into a function which is (1) meromorphic (in the space variable) on the whole complex plane if $\varepsilon >1/2$, (2) meromorphic on a strip around the real line if $\varepsilon =1/2$, and (3) Gevrey regular if $\varepsilon <1/2$. Note that $q$’s need not have any decay or pattern of behavior at $-\infty $.
address:
- University of Alaska Fairbanks
- |
Department of Mathematics and Statistics\
University of Alaska Fairbanks\
PO Box 756660\
Fairbanks, AK 99775
author:
- Alexei Rybkin
bibliography:
- 'bibryb.bib'
date: 'September, 2011'
title: 'Spatial Analyticity of solutions to integrable systems. I. The KdV case'
---
[^1]
Introduction and statements of main results
===========================================
The gain and persistence of regularity effects are important features of many dispersive (linear and nonlinear) partial differential equations (PDEs). The literature on the subject is truly enormous and we make no attempt to give a comprehensive review here. We only mention two recent relevant papers by Himonas et al [@Han_Himonas_2011], [@Himonas_2011] where the interested reader can find further references on analytic and Gevrey regularity properties for KdV-type equations. In fact, we are interested in a much stronger effect of formation of meromorphic solution out of nonsmooth data. More specifically, in the current paper, we are concerned with the following problem.
\[pb1\] Given the Cauchy problem for the KdV equation[^2] $$\begin{cases}
\partial _{t}u-6u\partial _{z}u+\partial _{z}^{3}u=0 \\
u|_{t=0}=q\end{cases}, \label{eq1.1}$$describe the largest possible class of (non-smooth) initial data $q$ which evolve into functions $u(z,t)$ meromorphic with respect to $z$ for any $t>0$.
Meromorphic (or, more generally, analytic) solutions have of course been intensively studied since the boom around integrable systems started in the late 60s. A pure soliton (reflectionless) solution, historically the first explicit solution, is meromorphic on the whole complex plane having infinitely many double poles. This fact is of course a trivial observation immediately following from the explicit formula for multisoliton solutions. We emphasize that how those poles interact is not obvious at all. This question was raised back in earlier 70s by Kruskal and has been followed up by many. We refer the interested reader to a particularly influential 1977 paper [@Air77] by Airault-McKean-Moser and recent Bona-Weissler [Bona09]{} and the literature cited therein. More complicated examples of explicit solutions include algebraric, rational, meromorphic simply periodic, elliptic, etc. (see, e.g. [@AblSatJMP78], [@Birnir87], [@GUW06], [@AktMee06] and the literature cited therein). All these examples are of course very specific and in addition those $q$’s are already meromorphic (i.e. smooth on the real line). Although Problem \[pb1\] is not addressed in those papers but they demonstrate the importance of meromorphic solutions.
Through the paper we deal with initial data subject to
\[hyp1.1\] $q$ is real and $L_{\limfunc{loc}}^{1}$ such that
1. (semiboundedness from below) $$\label{Cond1}
\inf\func{Spec}\left(-\partial_x^2+q(x)\right)=-h_0^2$$ with some $h_0\ge0$.
2. (subexponential decay at $+\infty $) For $x$ large enough $$\int_{x}^{\infty }\left\vert q\right\vert \leq C_{q}e^{-cx^{\varepsilon }}
\label{Cond2}$$with some positive $C_{q},c,\varepsilon $.
We assume that the constants $c,\varepsilon $ in are chosen optimal.
Note that the set of such functions is very large. Indeed, in terms of $q$ itself, Condition is satisfied if $$\limfunc{Sup}\limits_{x}\int_{x-1}^{x}\max \left( -q,0\right) <\infty ,
\label{Cond_q}$$i.e. $q$ is essentially bounded from below [@Glazman66]. The condition (\[Cond\_q\]) cannot be improved since (\[Cond\_q\]) becomes also necessary for (\[Cond1\]) if $q$’s are negative. Therefore, any $q$ subject to Hypothesis \[hyp1.1\] is essentially bounded from below, has subexponential decay at $+\infty $ and arbitrary otherwise. Such functions can grow (arbitrarily fast) at $-\infty $ or look like a stock market (Gaussian white noise on a left half line) but still satisfy our hypothesis as long as they exhibit rapid decay (\[Cond2\]) at $+\infty $. In spectral terms (\[Cond2\]) implies that $\left( 0,\infty \right) $ belongs to the absolutely continuous spectrum of $-\partial _{x}^{2}+q(x)$.
We now state our main results.
\[thm1.2\] Under Hypothesis \[hyp1.1\] with $\varepsilon \geq 1/2$ on the initial data $q$ in , the problem has an analytic in $z$ solution $u(z,t)$ given by $$u(z,t)=-2\partial _{z}^{2}\log \det \left( 1+\mathbb{M}(z,t)\right) ,
\label{det_form}$$where $\mathbb{M}(z,t)$ is a trace class operator-valued function constructed in Proposition \[pr4.1\] below for any $t>0$. Moreover, for any $t>0$
1. If $\varepsilon>1/2$ then $u(z,t)$ is meromorphic on $\mathbb{C}$.
2. If $\varepsilon =1/2$ then $u(z,t)$ is meromorphic in the strip $$\left\vert {\operatorname{Im}}z\right\vert <\frac{9\sqrt{2}}{8}c\sqrt{t} \label{strip}$$where $c$ is as in .
\[thm1.3\] Under Hypothesis \[hyp1.1\] with $0<\varepsilon <1/2$ on the initial data $q$ in , the operator-valued function $\mathbb{M}(x,t)$ given in Proposition \[pr4.1\] is trace class for any real $x$ and $t>0$ and $$\mathbb{M}(x,t)=\mathbb{M}^{\left( 1\right) }(x,t)+\mathbb{M}^{\left(
2\right) }(x,t),$$where $\mathbb{M}^{\left( 1\right) }(x,t)$ is meromorphic in $x$ and $\mathbb{M}^{\left( 2\right) }(x,t)$ is Gevrey $G^{\frac{1}{2\varepsilon }-1}$ regular. If in addition $1+\mathbb{M}(x,t)$ is invertible for any real $x$ and $t>0$ then the problem has a solution $u(x,t)$ given by $$u(x,t)=-2\partial _{x}^{2}\log \det \left( 1+\mathbb{M}(x,t)\right) ,
\label{det_form1}$$belonging to the Gevrey class $G_{\limfunc{loc}}^{\frac{1}{2\varepsilon }-1}$.
Theorems \[thm1.2\] and \[thm1.3\] significantly improve our results in [@Ry11] which in turn improve Tarama [@Tarama04]. Theorems [thm1.2]{} and \[thm1.3\] have also some important corollaries. We will come back to the relevant discussions in the last section when we have the necessary background. We only mention here that our approach is based on the Inverse Scattering Transform (IST) combined with pseudo-analytic continuation techniques developed by E.M. Dyn’kin (see e.g. [@Dyn76], [@BorDyn93]) and we do not believe that any of the statements of Theorem \[thm1.2\] can be obtained by purely PDE techniques.
The paper is organized as follows. In Section 2, for the reader’s convenience we list our main notation and give the relevant preliminaries. In Section 3 we define a suitable reflection coefficient and investigate its properties which will play a central role in our consideration. The results of this section may have some independent interest. In Section 4 we give a brief review of the classical IST stated in terms of Hankel operators and further prepare to prove our main results in Section 5. Section 6, the last one, is devoted to discussions of our results and some corollaries which directly follow from them. It also contains some open problems.
Notation and Preliminaries
==========================
We adhere to standard terminology accepted in Analysis. Namely, $\mathbb{R}_{\pm }\equiv[0,\pm \infty )$, $\mathbb{C}$ is the complex plane, $$\mathbb{C}_{\pm }=\left\{ z\in \mathbb{C}:\pm {\operatorname{Im}}z>0\right\} .$$Through the paper the subscript $\pm $ indicates objects (functions, operators, spaces, etc.) somehow related to $\mathbb{R}_{\pm }$ or $\mathbb{C}_{\pm }$. The bar $\overline{z}$ denotes the complex conjugate of $z$.
When appropriate, we write$$y\eqsim x\text{ in place of \ }y=\limfunc{const}\cdot x$$and similarly whenever convenient $$y\lesssim _{a}x\text{ in place of }y\leq C_{a}x$$with some $C_{a}>0$ dependent on a parameter $a$ but independent of $x$. If $C_{a}$ is an absolute constant we then write $y\lesssim x$. This will help us keep bulky formulas under control.
We use $\left\Vert \cdot \right\Vert _{X}$ to denote the norm in a Banach (Hilbert) space $X$.
We will need the Gevrey classes $G^{\alpha },\alpha >0,$ on $\mathbb{R}$ of all functions $f$:
$$\left\Vert \partial _{x}^{n}f\right\Vert _{L^{\infty }}\lesssim
_{f}Q_{f}^{n}\left( n!\right) ^{1+\alpha },n=0,1,2,...$$
with some $Q_{f}>0.$
By [@BorDyn93], Theorem 3, the statement $f\in G^{\alpha }$ is equivalent to the statement that $f$ admits a pseudo analytic extension to the whole complex plane such that $$\left\vert \partial _{\overline{z}}f\right\vert \lesssim _{f}\exp \left\{
-Q\left\vert {\operatorname{Im}}z\right\vert ^{-\frac{1}{\alpha }}\right\}
\label{lambda_bar}$$with some $Q>0$.
In a similar manner one introduces local Gevrey classes $G_{\limfunc{loc}}^{\alpha }$.
Next, $\mathfrak{S}_{2}$ denotes the Hilbert-Schmidt class$$\mathfrak{S}_{2}=\left\{ A:\left\Vert A\right\Vert _{\mathfrak{S}_{2}}^{2}\equiv\func{tr}\left( A^{\ast }A\right) <\infty \right\}$$and $\mathfrak{S}_{1}$ is the trace class: $$\mathfrak{S}_{1}=\left\{ A:\left\Vert A\right\Vert _{\mathfrak{S}_{1}}\equiv\func{tr}\left( A^{\ast }A\right) ^{1/2}<\infty \right\} .$$ Note that $A\in\mathfrak{S}_1$ if and only if $A=A_1A_2$ with some $A_1,A_2\in\mathfrak{S}_2$.
Some other miscellaneous notation: $\chi _{S}\left( x\right) $ is the characteristic function of a set $S$, i.e. $$\chi _{S}\left( x\right) \equiv\left\{
\begin{array}{c}
1,x\in S \\
0,x\notin S\end{array}\right. .$$In particular $\chi _{\pm }\equiv\chi _{_{\mathbb{R}_{\pm }}}$is the Heaviside function of $\mathbb{R}_{\pm }$. We also write $$\left. f\right\vert _{S}=\chi _{S}f.$$
The notation $H_{q}\equiv -\partial _{x}^{2}+q(x)$ for the Schrödinger operator on $L^{2}\left( \mathbb{R}\right) $ will be frequently used.
The reflection coefficient and its analytic structure
=====================================================
In this section we define a suitable reflection coefficient and investigate its properties which will play a central role in our consideration. The results of this section may have some independent interest.
In the short-range scattering for the full line Schrödinger operator, one typically introduces the right and left reflection coefficients $R(\lambda ),L(\lambda )$ and the transmission coefficient $T(\lambda )$ as functions of the momentum $\lambda $ (see e.g. [@Deift79]). These quantities (also called transition coefficients) can also be properly defined in much larger spectral situations through Wronskians and/or Titchmarsh-Weyl $m$-functions (see e.g. [@GNP97; @GS97]). Such extensions need not be unique. However, in our setting of step-like potentials decaying at $+\infty $, there is a natural candidate for the right reflection coefficient $R(\lambda )$.
\[def3.1\] Let $q(x)$ be real, locally integrable such that $q\in
L^{1}\left( \mathbb{R}_{+}\right) $ and $-\partial _{x}^{2}+q(x)$ is in the limit point case at $-\infty $. Denoting by $R_{n}(\lambda )$ the right reflection coefficient (which is necessarily well defined) from the potential $q_{n}=q|_{(-n,\infty )}$, we call the weak limit (if it exists) $$R(\lambda )\equiv \text{w-}\lim R_{n}(\lambda ),\;n\rightarrow \infty ,
\label{eq3.1}$$the right reflection coefficient from the potential $q$.
Note that one should not expect in pointwise convergence as an explicit counterexample $q=$ $\chi _{-}$ readily shows. Uniform convergence in is not available in general even in the short-range setting [@Deift79].
As shown in [@Ry11], Lemma 5.4, the reflection coefficient introduced this way is well defined. The following statement will play a crucial role in our consideration.
\[pr3.2\] Under Hypothesis \[hyp1.1\], the right reflection coefficient given by exists and admits the representation $$R(\lambda )=A(\lambda )+\frac{S(\lambda )G(\lambda )}{\lambda B(\lambda )}
\label{Rep}$$where functions $A,B,S,G$ have the properties
1. \[it1\] $A$ is an analytic on $\mathbb{C}^+\setminus[0,ih_0]$ function such that $\left\vert A \right\vert \le2$ on $\mathbb{R}$ and $A(\lambda)=o\left(1/\lambda\right)$, $\lambda\to\infty$ along any ray in $\mathbb{C}^+$
2. \[it2\] $B$ is the Blaschke product $$B(\lambda )=\prod_{k=1}^{N}\frac{\lambda -i\varkappa _{k}}{\lambda
+i\varkappa _{k}}$$where real $\varkappa _{k}$’s are such that $\left\{ -\varkappa
_{k}^{2}\right\} _{k=1}^{N}$ is the negative discrete spectrum of $H_{q_{+}}$, $q_{+}\equiv q|_{\mathbb{R}_{+}}$
3. \[it3\] $\left\vert S(\lambda )\right\vert \leq 1$, $\lambda \in
\mathbb{C}^{+}$
4. \[it4\] $G\in G^{\frac{1}{\varepsilon }-1}$
5. \[it5\] $\left\vert S(\lambda )G(\lambda )/\lambda \right\vert
\leq 1$ a.e. on $\mathbb{R}$
6. \[it6\] If $R_{n}$ is as in Definition \[def3.1\] then $$R_{n}(\lambda )=A_{n}(\lambda )+\frac{S(\lambda )G(\lambda )}{\lambda
B(\lambda )}$$and $$A_{n}\rightarrow A,\;n\rightarrow \infty$$uniformly on any compact in $\mathbb{C}^{+}\setminus \lbrack 0,ih_{0}]$.
Most of statements in Proposition \[pr3.2\] (save ) are proven in [@Ry11] and we restrict ourselves to some comments only. Note first that Condition 1 of Hypothesis \[hyp1.1\] implies that $-\partial
_{x}^{2}+q(x)$ is in the limit point case at $-\infty $ (see, e.g. [ClarkGeszt03]{} for complete results on this matter). Splitting $$q=q_{-}+q_{+},\ \ q_{\pm }=q|_{\mathbb{R}_{\pm }} \label{split_q}$$induces the representation $$R=\frac{T_{+}^{2}R_{-}}{1-R_{-}L_{+}}+R_{+}$$where $\pm $ label scattering quantities associated with $q_{\pm }$. The functions $T_{+},L_{+},R_{-}$ can be analytically continued into $\mathbb{C}^{+}$ and $$A\equiv \frac{T_{+}^{2}R_{-}}{1-R_{-}L_{+}}$$has properties , . For $R_{+}$, which is independent of $n$, we use the representation [@Deift79], Theorem 2, $$R_{+}(\lambda )=\frac{T_{+}(\lambda )}{\lambda }G(\lambda )$$where $$G(\lambda )=\frac{1}{2i}\int_{-\infty }^{\infty }e^{-2i\lambda x}g(x)dx
\label{eq6'.1}$$with some $g$ obeying $$\left\vert g(x)\right\vert \leq \left\vert q(x)\right\vert +\limfunc{const}\int_{x}^{\infty }\left\vert q\right\vert . \label{Est_on_g}$$Since $R_{+}(\lambda )$ is a reflection coefficient we have . Since $T_{+}$ is a transmission coefficient, $$T_{+}(\lambda )=\prod_{k=1}^{N}\frac{\lambda +i\varkappa _{k}}{\lambda
-i\varkappa _{k}}\cdot S(\lambda )=B\left( \lambda \right) ^{-1}S(\lambda )$$where $S$ is an outer function of $\mathbb{C}^{+}$: $\left\vert S(\lambda
)\right\vert \leq 1$, $\lambda \in \mathbb{C}^{+}$. This proves and .
The proposition is proven if we show . Due to we should demonstrate that $G$ admits a pseudo analytic extension the whole complex plane such that $$\left\vert \partial _{\overline{\lambda }}G\right\vert \lesssim \exp \left\{
-Q\left\vert {\operatorname{Im}}\lambda \right\vert ^{-\frac{\varepsilon }{1-\varepsilon }}\right\} \label{eq6''.1}$$with some $Q>0$. There are a few explicit ways to construct pseudo analytic continuations (see e.g. [@Dyn76], [@BorDyn93], [@Tarama04]) producing different extensions. We modify the one used in [@Tarama04] to obtain a better $Q$ in \[eq6”.1\]. Note that $$G\left( \lambda \right) \eqsim \widehat{g}\left( 2\lambda \right) \label{Gg}$$where $\widehat{g}$ is the Fourier transform of $g$ which due to satisfies Condition 2 of Hypothesis \[hyp1.1\] with some $\widetilde{c}<c$. I.e. $$\int_{x}^{\infty }\left\vert g\right\vert \lesssim _{g}e^{-\widetilde{c}x^{\varepsilon }}. \label{est_g}$$For any $\lambda \in \mathbb{C}^{+}$ define $$\widetilde{G}\left( \lambda \right) =\sum_{n\geq 1}\theta \left(
r^{\varepsilon +2}x_{n}^{1-\varepsilon }\frac{{\operatorname{Im}}\lambda }{\widetilde{c}}\right) G_{n}(\lambda ), \label{eq6''.1too}$$where $\theta $ is a smooth on $\mathbb{R}_{+}$ function such that: $$\begin{aligned}
\theta (x)& =1,\;x\in \lbrack 0,1], \\
\theta (x)& =0,\;x\geq r,\end{aligned}$$$r>1$, $x_{n}=r^{n}$ and $$G_{n}(\lambda )=\int_{x_{n-1}}^{x_{n}}e^{-i\lambda x}g(x)dx.$$The formula clearly defines an extension of $\widehat{g}\left( \lambda \right) $ to complex $\lambda $. We next show that $\widetilde{G}$ is uniformly bounded on $\mathbb{C}^{+}$. Bound $G_{n}$ first. By $$\left\vert G_{n}(\lambda )\right\vert \lesssim e^{\left\vert {\operatorname{Im}}\lambda
\right\vert \cdot x_{n}}\int_{x_{n-1}}^{x_{n}}\left\vert g\right\vert
\lesssim _{g}\exp \left\{ \left\vert {\operatorname{Im}}\lambda \right\vert \cdot x_{n}-\widetilde{c}x_{n-1}^{\varepsilon }\right\}$$and one has $$\left\vert \widetilde{G}(\lambda )\right\vert \lesssim _{g}\sum_{n\geq
1}\sum_{n\geq 1}\theta \left( r^{\varepsilon +2}x_{n}^{1-\varepsilon }\frac{{\operatorname{Im}}\lambda }{\widetilde{c}}\right) \exp \left\{ \left\vert {\operatorname{Im}}\lambda
\right\vert \cdot x_{n}-\widetilde{c}x_{n-1}^{\varepsilon }\right\} .
\label{eq6'''.1}$$ In many terms are in fact zero and nontrivial ones are subject to $$r^{\varepsilon +2}x_{n}^{1-\varepsilon }\frac{\left\vert {\operatorname{Im}}\lambda
\right\vert }{\widetilde{c}}\leq r.$$I.e. only nonzero terms in are the ones obeying $$x_{n}^{1-\varepsilon }\leq \frac{\widetilde{c}}{r^{\varepsilon +1}}\cdot
\frac{1}{\left\vert {\operatorname{Im}}\lambda \right\vert }. \label{eq6'''.2}$$
Under the condition , for the argument of the exponential in , we have ($1/r<\delta <1$) $$\begin{aligned}
\left\vert {\operatorname{Im}}\lambda \right\vert \cdot x_{n}-\widetilde{c}r^{-\varepsilon
}x_{n}^{\varepsilon }& =\left( \left\vert {\operatorname{Im}}\lambda \right\vert \cdot
x_{n}-\delta \widetilde{c}r^{-\varepsilon }x_{n}^{\varepsilon }\right)
-(1-\delta )r^{-\varepsilon }x_{n}^{\varepsilon } \label{eq6'''.3} \\
& =\left\vert {\operatorname{Im}}\lambda \right\vert x_{n}^{\varepsilon }\left(
x_{n}^{1-\varepsilon }-\delta \frac{\widetilde{c}}{\left\vert {\operatorname{Im}}\lambda
\right\vert }\right) -(1-\delta )r^{-\varepsilon }x_{n}^{\varepsilon }.
\notag\end{aligned}$$By the right hand side of doesn’t exceed $$\begin{aligned}
&&\left\vert {\operatorname{Im}}\lambda \right\vert x_{n}^{\varepsilon }\left( \frac{\widetilde{c}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\frac{1}{r^{\varepsilon
+1}}-\frac{\widetilde{c}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\frac{\delta }{r^{\varepsilon }}\right) -(1-\delta )r^{-\varepsilon }x_{n}^{\varepsilon } \\
&=&-\widetilde{c}\left( \delta -\frac{1}{r}\right) x_{n-1}^{\varepsilon
}-(1-\delta )x_{n-1}^{\varepsilon } \\
&<&-\limfunc{const}x_{n-1}^{\varepsilon }.\end{aligned}$$It follows now from this estimate and that $$\left\vert \widetilde{G}(\lambda )\right\vert \lesssim _{g}\sum_{n\geq
0}\exp \{-\limfunc{const}x_{n}^{\varepsilon }\}<\infty . \label{eq6iv.0}$$Similarly one proves that all derivatives of $G$ are also bounded on $\mathbb{C}^{+}$.
It remains now to show . One has $$\begin{aligned}
\left\vert \partial _{\overline{\lambda }}\widetilde{G}\right\vert & \leq
\sum_{n\geq 1}\theta ^{\prime }\left( r^{\varepsilon +2}x_{n}^{1-\varepsilon
}\frac{\left\vert {\operatorname{Im}}\lambda \right\vert }{\widetilde{c}}\right) \frac{r^{\varepsilon +1}x_{n}^{1-\varepsilon }}{2\widetilde{c}}\left\vert
G_{n}\right\vert \label{eq6iv.1} \\
& \lesssim _{g}\sum_{n\geq 1}x_{n}^{1-\varepsilon }\exp \{\left\vert {\operatorname{Im}}\lambda \right\vert \cdot x_{n}-\widetilde{c}r^{-\varepsilon
}x_{n}^{\varepsilon }\}. \notag\end{aligned}$$ Only terms subject to $$\frac{\widetilde{c}r^{-\varepsilon -2}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\leq x_{n}^{1-\varepsilon }\leq \frac{\widetilde{c}r^{-\varepsilon -1}}{\left\vert {\operatorname{Im}}\lambda \right\vert } \label{eq6iv.2}$$make a non trivial contribution to the series in . The inequality implies $$\begin{aligned}
x_{n}& \geq \left( \frac{\widetilde{c}r^{-\varepsilon -2}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\right) ^{\frac{1}{1-\varepsilon }}, \notag \\
\intertext{or}
x_{n}^{\varepsilon }& \geq \left( \frac{\widetilde{c}r^{-\varepsilon -2}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\right) ^{\frac{\varepsilon }{1-\varepsilon }}. \label{eq6iv.3}\end{aligned}$$
Splitting the argument of the exponential in same way as and using , we have $$\begin{aligned}
\left\vert {\operatorname{Im}}\lambda \right\vert \cdot x_{n}^{\varepsilon }-\frac{\widetilde{c}}{r^{\alpha }}x_{n}^{\alpha }& \leq \left\vert {\operatorname{Im}}\lambda
\right\vert \cdot x_{n}^{\varepsilon }\left( \frac{\widetilde{c}r^{-\varepsilon -1}}{\left\vert {\operatorname{Im}}\lambda \right\vert }-\frac{r^{-\varepsilon }\delta \widetilde{c}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\right) -(1-\delta )\frac{\widetilde{c}x_{n}^{\varepsilon }}{r^{\varepsilon }} \\
& =-x_{n}^{\varepsilon }\widetilde{c}r^{-\varepsilon -1}(r\delta
-1)-(1-\delta )\frac{\widetilde{c}x_{n}^{\varepsilon }}{r^{\varepsilon }} \\
& \leq -\left( \frac{\widetilde{c}r^{-\varepsilon -2}}{\left\vert {\operatorname{Im}}\lambda \right\vert }\right) ^{\frac{\varepsilon }{1+\varepsilon }}\widetilde{c}r^{-\varepsilon -1}(r\delta -1)-(1-\delta )\frac{\widetilde{c}x_{n}^{\varepsilon }}{r^{\varepsilon }} \\
& -\frac{\widetilde{c}^{\frac{1}{1-\varepsilon }}}{r^{\frac{2\varepsilon +1}{1-\varepsilon }}}\frac{r\delta -1}{\left\vert {\operatorname{Im}}\lambda \right\vert ^{\frac{\varepsilon }{1-\varepsilon }}}-(1-\delta )\frac{\widetilde{c}x_{n}^{\varepsilon }}{r^{\varepsilon }}.\end{aligned}$$
Inserting this into we obtain $$\begin{aligned}
\left\vert \partial _{\overline{\lambda }}\widetilde{G}\right\vert &
\lesssim _{g}\left( \sum_{n\geq 0}x_{n}^{1-\varepsilon }\exp \{-\limfunc{const}x_{n}^{\varepsilon }\}\right) \cdot \exp \left\{ -\widetilde{Q}\left\vert {\operatorname{Im}}\lambda \right\vert ^{-\frac{\varepsilon }{1-\varepsilon }}\right\} \label{eq6v.0} \\
\widetilde{Q}& \equiv(r\delta -1)\frac{\widetilde{c}^{\frac{1}{1-\varepsilon
}}}{r^{\frac{2\varepsilon +1}{1-\varepsilon }}}<(r\delta -1)\frac{c^{\frac{1}{1-\varepsilon }}}{r^{\frac{2\varepsilon +1}{1-\varepsilon }}}.
\label{eq6v.1}\end{aligned}$$The series in is convergent and $\widetilde{G}\left( \lambda
\right) $ is an pseudo analytic extension of $\widehat{g}\left( \lambda
\right) $ from the real line to the upper half plane. Due to we have found a pseudo analytic extension of $G$ subject to with $Q=2\widetilde{Q}$. This completes our proof.
The representation is not unique. It depends on the reference point in the splitting of . This flexibility will be used later.
We have also had some flexibility in choosing $r$ and $\delta $ in subject to $r>1$, $1/r<\delta <1$. The range for $Q=2\widetilde{Q}$ given by is $$0<Q<\frac{2(\varepsilon -1)(3\varepsilon )^{\frac{2\varepsilon +1}{1-\varepsilon }}}{(2\varepsilon +1)^{\frac{1+\varepsilon }{1-\varepsilon }}}\cdot c^{\frac{1}{1-\varepsilon }}$$which is inessential to what follows but the borderline case $\varepsilon
=1/2$. For this case, $$0<Q<\frac{3^{3}}{2^{7}}c^{2}. \label{eq7.1}$$
The Marchenko integral operator and the inverse scattering transform
====================================================================
The integral operator we are concerned with in this section appears to have been introduced by Marchenko and received a comprehensive treatment in his classical book [@Mar86]. To acknowledge Marchenko’s profound contribution to the subject, we denote this operator by $\mathbb{M}$ (Marchenko used $\mathbb{F}$) but otherwise try to retain as much of his original notation as possible.
We call the integral operator $\mathbb{M}:L^{2}\left( \mathbb{R}_{+}\right)
\rightarrow L^{2}\left( \mathbb{R}_{+}\right)$ Marchenko if $$\begin{aligned}
\left( \mathbb{M}f\right) (x)& =\int_{0}^{\infty }M(x+y)f(y)dy,\quad f\in
L^{2}\left( \mathbb{R}_{+}\right) , \notag \\
M(\cdot )& =\int_{0}^{\infty }e^{-(\cdot )\lambda }d\rho (\lambda
)+\int_{-\infty }^{\infty }e^{2i(\cdot )\lambda }R(\lambda )\frac{d\lambda }{2\pi }, \label{eq4.1}\end{aligned}$$where $\rho $ is a finite nonnegative measure and $R$ is such that for a.e. $\lambda \in \mathbb{R}$ $$R(-\lambda )=\overline{R(\lambda )}\quad ,\quad \left\vert R(\lambda
)\right\vert \leq 1.$$
The operator $\mathbb{M}$ introduced this way is clearly a Hankel operator (the kernel depends on the sum of the arguments). We say that $\mathbb{M}$ is associated with a potential $q$ if $R$ is a reflection coefficient from $q $ and $\rho $ characterizes the negative spectrum of $H_{q}$. For short-range $q$’s, the measure $\rho $ is purely discrete and ($\delta $ is the Dirac $\delta $-function) $$d\rho (\lambda )=\sum_{n=1}^{N}c_{n}^{2}\delta (\lambda -\varkappa
_{n})d\lambda$$where $\varkappa _{n}$’s are such that $\left\{ -\varkappa _{n}^{2}\right\}
_{n=1}^{N}$ are the negative bound states of $H_{q}$ and $\left\{c_{n}\right\}_{n=1}^N$ are the associated norming constants. Of course, the kernel $M$ (which we also call Marchenko) is nothing but the sum of the Laplace transform of the (finite, positive) measure $\rho $ and the Fourier transform of the (symmetric, bounded) function $R$. This is the main feature of the Marchenko operator resulting in the decomposition $$\mathbb{M}=\mathbb{M}_{1}+\mathbb{M}_{2}, \label{eq4.2}$$where $\mathbb{M}_{1}\geq 0$ and ($\chi =\chi _{+}$, $\mathcal{F}$ is the Fourier transform) $$\mathbb{M}_{2}=\chi \mathcal{F}^{-1}R\mathcal{F}^{-1}$$and is selfadjoint.
Note that the Marchenko operator is not typically studied in the context of Hankel operators. We have found in [@Ry11] that the language of Hankel operators is very convenient in the IST formalism. In this language, $\rho
,R $ are called the measure and symbol of $\mathbb{M}_{1},\mathbb{M}_{2}$ respectively (see e.g. [@Nik2001]). In the language of inverse scattering, $(\rho ,R)$ are commonly referred to as the (right) scattering data[^3].
In the context of the Cauchy problem for the KdV equation, we have a two parametric family of Marchenko operators $\mathbb{M}(z,t)$, where $(z,t)$ are the (real) variables in . Namely, the measure $\rho _{z,t}$ and the symbol $R_{z,t}$ of $\mathbb{M}(z,t)$ are given by $$\begin{aligned}
d\rho _{z,t}(\lambda )& =\zeta _{z,t}(i\lambda )d\rho (\lambda ), \\
R_{z,t}(\lambda )& =\zeta _{z,t}(\lambda )R(\lambda ), \\
\zeta _{z,t}(\lambda )& :\equiv e^{8i\lambda ^{3}t+2i\lambda z}\end{aligned}$$and $(\rho ,R)$ are the scattering data for the profile $q$. Clearly for $\lambda \in \mathbb{R}$ $$\begin{aligned}
\left\vert \zeta _{z,t}(\lambda )\right\vert & =1, \\
\left\vert \zeta _{z,t}(i\lambda )\right\vert & =e^{8\lambda ^{3}t-2\lambda
z}.\end{aligned}$$
The whole point of the IST is that $(\rho _{0,t},R_{0,t})$ are the scattering data for $H_{u(z,t)}$ where $u(z,t)$ solves . The actual mechanism to recover $u(z,t)$ amounts to solving the Marchenko integral equation[^4] or equivalently through the Riemann-Hilbert problem. For our purposes the explicit formula is convenient. Note that is nothing but Cramer’s rule for linear integral equations. For $R\equiv 0$ (reflectionless initial profile), assuming that $q$ is short-range, $\mathbb{M}(z,t)$ turns into a finite rank operator of a very explicit structure. The formula in this particular case was discovered in the 1960s. In the general short-range case ($R\neq 0$), we are not sure whom it should be attributed to but it was systemically studied by Pöppe in the 1980s (see, e.g. [@P84] and also [@P89] for the Kadomtsev-Petviashvili and [@P83] for the sine-Gordon equations. In the context of nondecaying initial data, it appears first in [@Ry11]. Note that the sense in what the determinant in is understood is an issue which doesn’t seem to be fully addressed in the literature. It is typically defined by the Fredholm formula through an absolutely convergent (Fredholm) series. We actually show that $\mathbb{M}(z,t)$ is trace class for any $z$ (even complex) and $t>0$ under Hypothesis \[hyp1.1\]. This means that $\det
\left( 1+\mathbb{M}(z,t)\right) $ is an invariant, i.e. it produces the same value in any basis in $L^{2}\left( \mathbb{R}_{+}\right) $.
In the setting of step-like potentials, the Marchenko operator has been intensively studied in the Kharkov mathematical school by Hruslov, Kotlyarov and their students[^5] (see, e.g. [@Hruslov76], [@KhrKot94]).
We also refer to Cohen [@Cohen1984], Kappeler [@Kappeler86], Venakides [@Ven86] (and the literature cited therein), and recent Egorova-Teschl [@ET11]. In all the above papers save [@ET11], $q$’s are assumed to have a specific type of behavior at $-\infty $ (approaching either a constant or a periodic function) and fall off at $+\infty$. In [ET11]{}, the interesting case of two finite gap potentials fused together is considered.
We summarize important properties of the Marchenko operator in the following (see [@Ry11] for details).
\[pr4.1\] Assuming Hypothesis \[hyp1.1\], let $\mathbb{M}(z,t)$ be the Marchenko operator associated with $q$ and let $A$ be as in Proposition [pr3.2]{}. Then for any $z\in \mathbb{R}$, $t>0$, $$\mathbb{M}(z,t)=\mathbb{M}_{+}(z,t)+\mathbb{A}(z,t), \label{eq4}$$where $\mathbb{M}_{+}(z,t)$ is the Marchenko operator associated with $q_{+}=q|_{\mathbb{R}_{+}}$ and $\mathbb{A}(z,t)$ is a Hankel integral operator with the kernel $$\frac{1}{2\pi }\int_{\mathbb{R}+ih}e^{2i\lambda (\cdot )}\zeta
_{z,t}(\lambda )A(\lambda )d\lambda ,\quad h>h_{0}.$$Furthermore, $\mathbb{A}(z,t)$ is an entire operator-valued function of trace class for any complex $z$ and $t>0$, continuous with respect to $q$ in the following sense: If $q_{1},q_{2}$ are two functions subject to Hypothesis \[hyp1.1\] then $$\left\Vert \mathbb{A}_{1}(z,t)-\mathbb{A}_{2}(z,t)\right\Vert _{\mathfrak{S}_{1}}\leq \frac{1}{4\pi h}\left\Vert \zeta _{z,t}(A_{1}-A_{2})\right\Vert
_{L^{1}(\mathbb{R}+ih)}$$for any $z\in \mathbb{C}$, $t>0$.
Note that $\mathbb{M}(z,t)$ depends on $(z,t)$ through $\zeta _{z,t}$.
Proof of the main results
=========================
With all the preparations done in the previous sections, the actual proofs will be quite short. It is convenient to conduct both proofs at a time. Note first that, by a trivial shifting, we may assume without loss of generality that $H_{q_{+}}$ has at most one bound state $-\varkappa ^{2}$. Consider the problem with $$q_{n}(x)=\begin{cases}
q(x)\quad & ,\quad x\geq -n \\
0 & ,\quad x<-n\end{cases}.$$
It is well-known that for such initial profiles[^6] $$u_{n}(z,t)=-2\partial _{z}^{2}\log \det \left( 1+\mathbb{M}_{n}(z,t)\right) .
\label{eq5.1}$$By Proposition \[pr4.1\] $$\mathbb{M}(z,t)=\mathbb{M}_{+}(z,t)+\mathbb{A}(z,t)+\delta \mathbb{A}(z,t)$$where $\delta \mathbb{A}\equiv\mathbb{A}_{n}-\mathbb{A}$ is meromorphic in $z $ for any $t>0$ and small in the $\mathfrak{S}_{1}$-norm for $n$ large enough. I.e. $$\left\Vert \mathbb{M}_{n}(z,t)-\mathbb{M}(z,t)\right\Vert _{\mathfrak{S}_{1}}\rightarrow 0,\quad n\rightarrow \infty . \label{eq5.2.1}$$Therefore, $\mathbb{M}(z,t)\in \mathfrak{S}_{1}$ is proven if we show that $\mathbb{M}_{+}(z,t)\in \mathfrak{S}_{1}$.
Split $$\mathbb{M}_{+}(z,t)=\mathbb{M}_{1}^{+}(z,t)+\mathbb{M}_{2}^{+}(z,t)$$where $\mathbb{M}_{1}^{+}(z,t),\mathbb{M}_{2}^{+}(z,t)$ are the Hankel operators with the kernels $$c_{0}^{2}\zeta _{z,t}(i\varkappa )e^{-\varkappa (x+y)},$$and $$\frac{1}{2\pi }\int_{-\infty }^{\infty }e^{i\lambda (x+y)}\zeta
_{z,t}(\lambda )R_{+}(\lambda )d\lambda$$respectively. Here $c_{0}$ stands for the norming constant associated with the bound state $-\varkappa ^{2}$.
The operator $\mathbb{M}_{1}(z,t)$ is rank 1 and clearly entire in $z$. Thus we only need to properly control $\partial _{z}^{n}\mathbb{M}_{2}^{+}(z,t)$ in the $\mathfrak{S}_{1}$-norm. Evaluate (so far formally) the kernel of $\partial _{z}^{n}\mathbb{M}_{2}^{+}(z,t),n=0,1,2,...,$ by the Green formula applied to the strip $\mathbb{R}\times (0,\varkappa /2)$ and by Proposition \[pr3.2\] ($\lambda =\alpha +i\beta ,\partial _{\overline{\lambda }}=\frac{1}{2}(\partial _{\alpha }+i\partial _{\beta })$) $$\begin{aligned}
\frac{1}{2\pi }\int_{-\infty }^{\infty }e^{i\lambda (x+y)}(2i\lambda
)^{n}\zeta _{z,t}(\lambda )R_{+}(\lambda )d\lambda & =\int_{\mathbb{R}+i\varkappa /2}e^{i\lambda (x+y)}(2i\lambda )^{n}\left( B^{-1}SG\right)
(\lambda )\frac{d\lambda }{2\pi } \notag \\
& \quad +2i\int_{0}^{\varkappa /2}d\beta \int d\alpha \;e^{i\lambda
(x+y)}F(\alpha ,\beta ) \label{eq5.3.1} \\
& \equiv H_{1}(x+y)+H_{2}(x+y), \notag\end{aligned}$$where $$F(\alpha ,\beta )\equiv \frac{1}{2\pi }\zeta _{z,t}(\lambda )(2i\lambda
)^{n-1}\frac{S(\lambda )}{B(\lambda )}\partial _{\overline{\lambda }}G(\alpha ,\beta ).$$
Due to the rapid decay of $e^{8i\lambda^3t}$ as $\lambda\to\infty$ along $\mathbb{R}+ih$, the function $F(\alpha,\beta)$ is subject to the conditions of Proposition \[pr4.1\] and hence the integral operator with kernel $H_1$ is trace class. Our analysis of the integral operator with kernel $H_2$ is based upon the following lemma.
\[lem5.1\] Let $F(\alpha ,\beta )$ be such that for some $h>0$ $$\int_{0}^{h}\left( \int_{-\infty }^{\infty }\left\vert F\left( \alpha ,\beta
\right) \right\vert d\alpha \right) \frac{d\beta }{\beta }<\infty .
\label{star}$$Then the integral Hankel operator $\mathbb{H}$ with the kernel ($\lambda
=\alpha +i\beta $) $$H(x)=\int_{0}^{h}d\beta \int_{-\infty }^{\infty }\frac{d\alpha }{2\pi }e^{i\lambda x}F(\alpha ,\beta )$$is trace class and $$\left\Vert \mathbb{H}\right\Vert _{\mathfrak{S}_{1}}\leq \frac{1}{2}\int_{0}^{h}\frac{d\beta }{\beta }\int_{-\infty }^{\infty }d\alpha
\left\vert F(\alpha ,\beta )\right\vert .$$
We have $$\begin{aligned}
H(x+y)& \eqsim \int_{0}^{h}d\beta e^{-\beta x}\int_{-\infty }^{\infty
}d\alpha e^{i\alpha x}F(\alpha ,\beta ) \notag \\
& =\int_{0}^{h}e^{-\beta (x+y)}\widehat{F_{\beta }}(x+y)dx,
\label{eq5.2.1too} \\
\widehat{F_{\beta }}(x+y)& \equiv\int_{-\infty }^{\infty }e^{i\alpha
(x+y)}F(\alpha +\beta )d\alpha \notag \\
& \eqsim \widehat{F_{\beta }^{1/2}}\ast \widehat{F_{\beta }^{1/2}}(x+y)
\notag \\
& =\int_{-\infty }^{\infty }F_{\beta }^{1/2}(x+s)F_{\beta }^{1/2}(y-s)ds.
\label{eq5.2.2}\end{aligned}$$ Here we have used the convolution theorem. Inserting into implies that $$\mathbb{H}=\int_{0}^{h}\mathbb{H}_{\beta ,1}\mathbb{H}_{\beta ,2}d\beta,
\label{eq5.2.3}$$where $\mathbb{H}_{\beta ,1}$ and $\mathbb{H}_{\beta ,2}$ and integral (but not Hankel) operators on $L^{2}\left( \mathbb{R}\right) $ with the kernels $$\begin{aligned}
H_{\beta ,1}(x,s)& =\chi (x)e^{-\beta x}\widehat{F_{\beta }^{1/2}}(x+s), \\
H_{\beta ,2}(s,y)& =\chi (y)e^{-\beta y}\widehat{F_{\beta }^{1/2}}(y-s)\end{aligned}$$respectively.
It follows from that $$\label{eq5.3.1too}
\left\Vert \mathbb{H} \right\Vert_{\mathfrak{S}_1} \le \int_0^h \left\Vert
\mathbb{H}_{\beta,1} \right\Vert_{\mathfrak{S}_2} \cdot\left\Vert \mathbb{H}_{\beta,2} \right\Vert_{\mathfrak{S}_2}d\beta.$$
Evaluate now the Hilbert-Schmidt norms of $\mathbb{H}_{\beta ,1}$ and $\mathbb{H}_{\beta ,2}$. By the Plancherel equation we have $$\begin{aligned}
\left\Vert \mathbb{H}_{\beta ,1}\right\Vert _{\mathfrak{S_{2}}}^{2}&
=\int_{0}^{\infty }\int_{-\infty }^{\infty }\left\vert H_{\beta
,1}(x,s)\right\vert ^{2}ds\;dx \\
& =\int_{0}^{\infty }dxe^{-2\beta x}\int_{-\infty }^{\infty }ds\left\vert
\widehat{F_{\beta }^{1/2}}(x+s)\right\vert ^{2} \\
& =\frac{\left\Vert F_{\beta }^{1/2}\right\Vert _{L^{2}\left( \mathbb{R}\right) }^{2}}{2\beta }=\frac{1}{2\beta }\left\Vert F_{\beta }\right\Vert
_{L^{1}\left( \mathbb{R}\right) }.\end{aligned}$$That is $$\left\Vert H_{\beta ,1}\right\Vert _{\mathfrak{S}_{2}}\leq \frac{1}{\sqrt{2\beta }}\left\Vert F_{\beta }\right\Vert _{L^{1}\left( \mathbb{R}\right)
}^{1/2}.$$Similarly, $$\left\Vert H_{\beta ,2}\right\Vert _{\mathfrak{S}_{2}}\leq \frac{1}{\sqrt{2\beta }}\left\Vert F_{\beta }\right\Vert _{L^{1}\left( \mathbb{R}\right)
}^{1/2}$$and yields $$\left\Vert \mathbb{H}\right\Vert _{\mathfrak{S}_{1}}\leq \frac{1}{2}\int_{0}^{h}\left\Vert F_{\beta }\right\Vert _{L^{1}}\frac{d\beta }{\beta }.$$The lemma is proven.
Let us find suitable bounds on $\mathbb{R}\times \lbrack 0,\varkappa /2]$ for the functions involved in $F$: $$\begin{aligned}
\left\vert \zeta _{z,t}(\lambda )\right\vert & =\left\vert e^{8i\lambda
^{3}t+2i\lambda z}\right\vert =e^{8\beta ^{3}t-2\beta {\operatorname{Re}}z}\cdot
e^{-24\beta t\alpha ^{2}-2\alpha {\operatorname{Im}}z} \\
& \leq e^{\varkappa (\varkappa ^{2}t+\left\vert z\right\vert )}\cdot \exp
\left\{ -\left( \sqrt{24\beta t}\alpha +\frac{{\operatorname{Im}}z}{\sqrt{24\beta t}}\right) ^{2}+\frac{{\operatorname{Im}}^{2}z}{24\beta t}\right\} , \\
\left\vert \lambda ^{n-1}B^{-1}(\lambda )S(\lambda )\right\vert & \lesssim
\left( \left\vert \alpha \right\vert +\beta \right) ^{n-1}, \\
\left\vert \partial _{\overline{\lambda }}G\right\vert & \lesssim
_{q_{+}}\exp \left\{ -Q\beta ^{-\frac{\varepsilon }{1-\varepsilon }}\right\}
.\end{aligned}$$ Thus $$\left\vert F(\alpha ,\beta )\right\vert \lesssim _{q_{+}}e^{\varkappa \left(
\varkappa ^{2}t+\left\vert z\right\vert \right) }\left( \left\vert \alpha
\right\vert +\beta \right) ^{n-1}e^{-\left( \sqrt{24\beta t}\alpha +\frac{{\operatorname{Im}}z}{\sqrt{24\beta t}}\right) ^{2}}\exp \left\{ \frac{{\operatorname{Im}}^{2}z}{24t}\beta
^{-1}-Q\beta ^{-\frac{\varepsilon }{1-\varepsilon }}\right\} .
\label{eq5.4.0}$$ To prove Theorem \[thm1.2\] we only need to consider the case $n=1$. We have $$\begin{gathered}
\label{eq5.4.1}
\int_{0}^{\varkappa /2}\frac{d\beta }{\beta }\int_{-\infty }^{\infty
}\left\vert F\left( \alpha ,\beta \right) \right\vert d\alpha \\
\lesssim _{z,t,q_{+}}\int_{0}^{\varkappa /2}\beta ^{-3/2}\exp \left\{ \frac{{\operatorname{Im}}^{2}z}{24t}\beta ^{-1}-Q\beta ^{-\frac{\varepsilon }{1-\varepsilon }}\right\} d\beta .\end{gathered}$$ So $F$ is subject to the condition of Lemma \[lem5.1\] if the integral in converges, which depends on $\varepsilon$ and ${\operatorname{Im}}z$.
- $\varepsilon >1/2$. Then[^7] $\frac{\varepsilon }{1-\varepsilon }>1$ and the right hand side of is finite for any $z\in \mathbb{C}$. This means that $\mathbb{M}_{+}(z,t)$ is an entire $\mathfrak{S}_{1}$-valued function for any $t>0$ and due to , we can pass to the limit in as $n\rightarrow \infty $ by standard properties of infinite determinants (see e.g. [@GGK00]). This proves in Theorem \[thm1.2\].
- $\varepsilon =1/2$. Then $\frac{\varepsilon }{1-\varepsilon }=1$ and the right hand side of converges if and only if $$\frac{{\operatorname{Im}}^{2}z}{24t}-Q<0$$or when $$\left\vert {\operatorname{Im}}z\right\vert <\sqrt{12Q}\cdot \sqrt{t}.$$Choosing the maximum possible value of $Q$ in we get $$\left\vert {\operatorname{Im}}z\right\vert <\frac{9\sqrt{2}}{8}c\sqrt{t}$$and of Theorem \[thm1.2\] follows. Thus, Theorem \[thm1.2\] is proven.
- $0<\varepsilon <1/2$. Then $\frac{\varepsilon }{1-\varepsilon }<1$ and clearly diverges for any ${\operatorname{Im}}z\neq 0$ and our method fails to establish analyticity and we have to go back to and analyze it for any natural $n$. Expanding $\left( \left\vert \alpha \right\vert +\beta \right) ^{n-1}$ in by the binomial formula we have $$\begin{gathered}
\label{eq5.6.1}
\int_{0}^{\varkappa /2}\frac{d\beta }{\beta }\int_{-\infty }^{\infty
}F(\alpha ,\beta )d\alpha \\
\lesssim _{z,t,q_{+}}\sum_{k=0}^{n-1}\binom{n-1}{k}\int_{0}^{\varkappa
/2}d\beta \beta ^{k-1}e^{-Q\beta ^{-\frac{\varepsilon }{1-\varepsilon }}}\int_{0}^{\infty }\alpha ^{n-k-1}e^{-24\beta t\alpha ^{2}}d\alpha .\end{gathered}$$Reducing the inner integral in to the gamma function[^8], $$\begin{aligned}
\eqref{eq5.6.1}& =\sum_{k=0}^{n-1}\binom{n-1}{k}\int_{0}^{\varkappa
/2}d\beta \beta ^{k-1}e^{-Q\beta ^{-\frac{\varepsilon }{1-\varepsilon }}}\cdot \frac{1}{(3\beta t)^{\frac{n-k}{2}}}\Gamma \left( \frac{n-k}{2}\right) \notag \\
& \lesssim \sum_{k=0}^{n}\binom{n}{k}(3t)^{-\frac{n-k}{2}}\Gamma \left(
\frac{n-k}{2}\right) \int_{0}^{1}d\beta \beta ^{\frac{3k-n}{2}-1}e^{-Q\beta
^{-\frac{\varepsilon }{1-\varepsilon }}}. \label{eq5.7.1}\end{aligned}$$Introducing in the last integral the new variable $s=\beta ^{-\frac{\varepsilon }{1-\varepsilon }}$ and setting $\gamma \equiv\frac{1-\varepsilon }{2\varepsilon }>1/2$ we get $$\begin{aligned}
\int_{0}^{1}d\beta \beta ^{\frac{3k-n}{2}-1}e^{-Q\beta ^{-\frac{\varepsilon
}{1-\varepsilon }}}& =\frac{1-\varepsilon }{\varepsilon }\int_{1}^{\infty
}s^{\frac{\varepsilon -1}{\varepsilon }\left( \frac{3k-n}{2}-1\right)
-1}e^{-Qs}ds \notag \\
& \lesssim \int_{1}^{\infty }s^{\gamma (n-3k)-1}e^{-Qs}ds \notag \\
& \lesssim Q^{-\gamma (n-3k)+1}\int_{Q}^{\infty }s^{\gamma (n-3k)}e^{-s}ds.
\label{eq5.7.2}\end{aligned}$$The behavior of the last integral depends on the sign of $\omega
_{k}\equiv\gamma (n-3k)$. If $\omega _{k}\geq 0$, i.e. $3k\leq n$, then $$\begin{aligned}
J_{k}& \equiv \int_{Q}^{\infty }s^{\omega _{k}}e^{-s}ds \\
& \leq \int_{0}^{\infty }s^{\omega _{k}-1}e^{-s}ds=\Gamma (\omega _{k}) \\
& =\Gamma (\gamma (n-3k)).\end{aligned}$$If $\omega _{k}<0$, i.e. $3k>n$, then $$J_{k}\leq Q^{\omega _{k}-1}\int_{Q}^{\infty }e^{-s}ds\leq Q^{\omega _{k}-1}.$$ Splitting the sum in accordingly, we see that the right hand side of is dominated by $$\begin{gathered}
\label{eq5.8.1}
\sum_{0\leq 3k\leq n}\binom{n}{k}(3t)^{-\frac{n-k}{2}}Q^{-\omega _{k}}\Gamma
\left( \frac{n-k}{2}\right) \Gamma \left( \omega _{k}\right) +\sum_{n<3k\leq
3n}\binom{n}{k}(3t)^{-\frac{n-k}{2}}\Gamma \left( \frac{n-k}{2}\right) \\
\equiv S_{1}+S_{2}.\end{gathered}$$
Analyze now $S_{1}$ and $S_{2}$. For $S_{1}$ we have $$\begin{aligned}
S_{1}& \leq \Gamma \left( \frac{n}{2}\right) \Gamma (\gamma n)\sum_{0\leq
3k\leq n}\binom{n}{k}(3t)^{-\frac{n-k}{2}}Q^{-\omega _{k}} \notag \\
& \leq \left( Q^{2\gamma }+\frac{1}{\sqrt{3t}Q^{\gamma }}\right) ^{n}\Gamma
\left( \frac{n}{2}\right) \Gamma (\gamma n). \label{eq5.9.1}\end{aligned}$$For $S_{2}$ we obtain $$S_{2}\leq \left( 1+\frac{1}{\sqrt{3t}}\right) ^{n}\Gamma \left( \frac{n}{3}\right)$$and hence the contribution from $S_{2}$ to produces a real analytic function. On the other hand, as it easily follows from , the contribution from $S_{1}$ produces a function from $G^{\gamma -1/2}=G^{\frac{1}{2\varepsilon }-1}$. Thus we have proven that if $0<\varepsilon <1/2$ then $$\mathbb{M}_{+}(x,t)=\mathbb{M}_{+}^{\left( 1\right) }(x,t)+\mathbb{M}_{+}^{\left( 2\right) }(x,t)$$where $\mathbb{M}_{+}^{\left( 1\right) }(x,t)$ is a real analytic $\mathfrak{S}_{1}$-valued function and $\mathbb{M}_{+}^{\left( 2\right) }(x,t)$ is a $\mathfrak{S}_{1}$-valued function from the Gevrey class $G^{\frac{1}{2\varepsilon }-1}$. Thus, we can pass to the limit as before. The limiting function is from the Gevrey class $G^{\frac{1}{2\varepsilon }-1}$ if $\det
(1+\mathbb{M}(x,t))$ doesn’t vanish for all $x\in \mathbb{R}$. The latter occurs if $1+\mathbb{M}(x,t)$ is invertible on $\mathbb{R}$ for any $t>0$.
Theorem \[thm1.3\] is proven.
Discussions, corollaries, and open problems
===========================================
Discussions
-----------
Theorem \[thm1.2\] improves our main result from [@Ry11] where $\mathbb{M}\left( x,t\right) \in \mathfrak{S}_{1}$ was not proven and only real analyticity of $u\left( x,t\right) $ was obtained. The main idea of [@Ry11] is to put together the analytic continuation arguments of [Ryb10]{} to treat initial data on $\mathbb{R}_{-}$ and Tarama’s approach from [@Tarama04] to handle the data on $\mathbb{R}_{+}$. As far as we know the solution to Problem \[pb1\] given in [@Tarama04] was best known back then. The main result of [@Tarama04] says that $u\left( x,t\right) $ is real analytic under the following conditions: $q$ is real and $L_{\limfunc{loc}}^{2}$ such that $$\int_{-\infty }^{\infty }\left( 1+\left\vert x\right\vert \right) \left\vert
q\left( x\right) \right\vert dx<\infty$$and for $x$ large enough there are positive $C_{q},c$ so that$$\int_{x}^{\infty }\left\vert q\right\vert ^{2}\leq C_{q}e^{-cx^{1/2}}.$$Note that these conditions are much stronger than Hypothesis \[hyp1.1\]. The techniques used in [@Tarama04] are also based upon the (classical) IST but his analysis relies on the properties of the Airy function as opposed to ours which is based on analytic and pseudo-analyitc continuations. The latter appears particularly well-suited for addressing Problem \[pb1\] and consequently significantly less involved.
It is proven in [@Deift79], Theorem 7.2 that if $q$ is analytic in the strip $\left\vert \func{Im}z\right\vert <a$ and has Schwartz decay there, then $u\left( z,t\right) $ is meromorphic in a strip with at most $N$ poles (where $N$ is the number of bound states of $\ H_{q}$) off the real line. By Theorem 7.1 from the same [@Deift79], for the reflection coefficient one then have $R\left( \lambda \right) =O\left( e^{-2a\left\vert \lambda
\right\vert }\right) $ as $\lambda \rightarrow \infty $ which of course need not occur in our case. This implies that our real meromorphic solution $u\left( z,t\right) $ in Theorem \[thm1.2\] has, in general, infinitely many poles for any $t>0$ in any strip around the real line accumulating only to infinity. By general theorems [@Steinberg69] on families of compact meromorphic operators these poles continuously depend on $t$ and hence may appear or disappear only on the boundary of analyticity of $u\left(
x,t\right) $ (including infinity).
Corollaries
-----------
The following statement is a direct consequence of the analyticity of $u\left( z,t\right) $ for $t>0$.
\[Corollary’\]Under conditions of Theorem \[thm1.2\] the solution $u\left( z,t\right) $ can not vanish on an open set for any $t>0$ unless $q$ is identically zero.
This quickly recovers and improves a number of unique continuation results due to Zhang [@Zhang92]. E.g., one of the main results of [@Zhang92] says that $u\left( x,t\right) $ cannot have compact support at two different moments unless it vanishes identically. The techniques of [@Zhang92] rely upon the classical IST (coupled with some Hardy space arguments) and are valid under certain decay and regularity conditions on $q$.
\[Corol2\] The class of (nonsmooth) initial data $q$ such that $$\int_{-\infty }^{\infty }e^{c\left\vert x\right\vert ^{\varepsilon
}}\left\vert q\left( x\right) \right\vert dx<\infty \text{ for some }c,\varepsilon >0\text{ } \label{exp_decay}$$is not preserved under the KdV flow.
Assume that for some $t=t_{0}$ the function $u\left( x,t_{0}\right) $ is subject to (\[exp\_decay\]). Since the KdV equation is invariant under the transformation $\left( x,t\right) \rightarrow \left( -x,-t\right) $, the solution $u_{0}\left( x,t\right) $ to the problem (\[eq1.1\]) with the initial data $q_{0}\left( x\right) =u\left( -x,t_{0}\right) $, by Theorems \[thm1.2\], \[thm1.3\], will be at least smooth for any $t>0$. But $u_{0}\left( x,t_{0}\right) =q\left( x\right) $ forcing original $q$ to be smooth too.
Corollary \[Corol2\], in turn, implies that under the KdV flow neither an exponential decay at $-\infty $ nor smoothness persist in general. Note in this connection that issues related to persistence of regularity are also very important and have been extensively studied but we don’t touch on this here.
The explicit formula (\[det\_form\]), which was used to derive our analyticity results, does have some practical value. E.g. it implies that the large time asymptotic behavior of $u\left( x,t\right) $ is completely determined by the measure $\rho (\lambda )$ in (\[eq4.1\]) alone. This fact is so far rigorously proven for $q$’s tending to a negative constant or a periodic function at $-\infty $ and was used to obtain explicit expressions for the so-called asymptotic solitons (see, e.g. [@Hruslov76], [@Ven86], and [@KhrKot94]). We plan to return to this important issue elsewhere.
Open problems
-------------
1. We believe that under Hypothesis \[hyp1.1\] our solutions $u\left(
x,t\right) $ have no singularities on the real line for any $t>0$. If this held then the problem would be globally well-posed under Hypothesis \[hyp1.1\] only and no blow-up solution could develop. That is to say that $1+\mathbb{M}(x,t)$ is automatically invertible for any real $x$ and $t>0$ under Hypothesis \[hyp1.1\] alone. This fact is quite easy if in (\[eq4.1\]) the support of $\rho \left( \lambda \right) $ is rich enough (a set of uniqueness of an analytic function) or $\left\vert R\left( \lambda
\right) \right\vert <1$ on any set of positive Lebesgue measure (see [Ry11]{}). The situation is much less trivial if $R\left( \lambda \right) $ in (\[eq4.1\]) is unimodular for a.e. real $\lambda $ (i.e. $q$ is completely reflecting). An affirmative answer is given in [@GR11] for the case of $q $ such that $q|_{\mathbb{R}_{+}}=0$ and $H_{q}\geq 0$ (absence of negative spectrum). To address the problem as stated one needs to show that $1+\mathbb{M}(x,t)$ is invertible in the case when in (\[eq4.1\]) $\rho
\left( \lambda \right) $ is supported on a set $\left\{ \lambda _{n}\right\}
\subset \mathbb{R}_{+}$ such that $\dsum \lambda _{n}<\infty $ and $\left\vert R\left( \lambda \right) \right\vert =1$ a.e. on the real line. In term of the Schrodinger operator $H_{q}$ itself this means that the absolutely continuous spectrum of $H_{q}$ is simple and supported on $\mathbb{R}_{+}$ but there is a rich embedded positive singular spectrum. Physically relevant examples can be constructed from the Gaussian white noise, Pearson sparse blocks, Kotani potentials, etc.
2. We do not know much about the Banach (or Hilbert) space of meromorphic function to which $u\left( z,t\right) $ from \[thm1.2\] belongs. It would be very interesting to find such spaces as this would give, among others, important norm estimates for $u\left( z,t\right) $ which our paper lacks.
3. We (cautiously) conjecture that in Theorem \[thm1.3\] $u\left(
x,t\right) $ could be represented for any $t>0$ as a meromorphic function plus a small Gevrey regular function. We can in fact show that the trace norm of $\mathbb{M}^{\left( 2\right) }(x,t)$ from Theorem \[thm1.3\] can be made small but it is not clear if after taking the $\det $ and then $\log
$ the analytic and small Gevrey parts will still be separated. Of course, this question will immediately have an affirmative answer if under conditions of Theorem \[thm1.3\] the solution $u\left( x,t\right) $ happens to be real analytic. Our methods however fail to yield such results.
Acknowledgement {#acknowledgement .unnumbered}
===============
We are grateful to Fritz Gesztesy for valuable discussions.
[^1]: Based on research supported in part by the NSF under grant DMS 1009673.
[^2]: We use $z$ instead of $x$ for the spatial variable as it will frequently be complex.
[^3]: For short-range $q$’s, the so-called left scattering data are also considered, which need not be well-defined in our setting.
[^4]: also known as Gelfand-Levitan-Marchenko.
[^5]: Remark that this school has been greatly infuenced by Marchenko himself and he remains to be its part.
[^6]: So far we only know that the determinant exists in the Fredholm sense.
[^7]: We assume $\varepsilon <1$. If $\varepsilon \geq 1$ then Theorem \[thm1.2\] is trivial.
[^8]: Recall $\Gamma (z)=\displaystyle\int_{0}^{\infty }\alpha ^{z-1}e^{-\alpha
}d\alpha $.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Cooling oxygen-deficient strontium titanate to liquid-helium temperature leads to a decrease in its electrical resistivity by several orders of magnitude. The temperature dependence of resistivity follows a rough T$^{3}$ behavior before becoming T$^{2}$ in the low-temperature limit, as expected in a Fermi liquid. Here, we show that the roughly cubic resistivity above 100K corresponds to a regime where the quasi-particle mean-free-path is shorter than the electron wave-length and the interatomic distance. These criteria define the Mott-Ioffe-Regel limit. Exceeding this limit is the hallmark of strange metallicity, which occurs in strontium titanate well below room temperature, in contrast to other perovskytes. We argue that the T$^{3}$-resistivity cannot be accounted for by electron-phonon scattering à la Bloch-Gruneisen and consider an alternative scheme based on Landauer transmission between individual dopants hosting large polarons. We find a scaling relationship between carrier mobility, the electric permittivity and the frequency of transverse optical soft mode in this temperature range. Providing an account of this observation emerges as a challenge to theory.'
author:
- 'Xiao Lin$^{1}$, Carl Willem Rischau$^{1}$, Lisa Buchauer$^{1}$, Alexandre Jaoui$^{1}$, Benoît Fauqué$^{1,2}$ and Kamran Behnia$^{1}$'
date: 'April 28, 2017'
title: 'Metallicity without quasi-particles in room-temperature strontium titanate'
---
![**Resistivity and mobility in doped SrTiO$_3$:** a) Temperature dependence of resistivity in SrTiO$_{3-\delta}$ as the carrier concentrations is tuned from 10$^{17}$ to 10$^{19}$ cm$^{-3}$. Room-temperature resistivity is several hundred times higher than low-temperature resistivity. b) Hall mobility as a function of temperature. Above 100 K, the mobility does not depend on carrier concentration and is roughly cubic in temperature. c) Assuming that inelastic resistivity follows a power law, i.e. $\rho= \rho_{0}+ AT^{\alpha}$, one can extract the exponent, $\alpha$ by taking a logarithmic derivative: $\alpha=dln(\rho-\rho_{0})/dlnT$, as in the case of cuprates[@Cooper:2009]. At low temperature, $\alpha\simeq 2$. Above 50K, it starts a significant shift upward and exceeds 3 around 150K, before steadily decreasing afterwards. The antiferrodistortive transition[@Tao:2016] is the source of the small anomaly at 105K.[]{data-label="Fig1"}](Fig1V6.pdf){width="45.00000%"}
Introduction
============
The existence of well-defined quasi-particles is taken for granted in the Boltzmann-Drude picture of electronic transport. In this picture, carriers of charge or energy are scattered after traveling a finite distance. The Mott-Ioffe-Regel(MIR) limit is attained when the mean-free-path of a carrier falls below its Fermi wavelength or the interatomic distance[@Hussey:2004]. In most metals, resistivity saturates when this limit is approached. But in “bad”[@Emery:1995] or “strange”[@Sachdev:2015] metals, it continues to increase[@Gunnarsson:2003]. It is indeed strange when the mean-free-path persists to fall after attaining its shortest conceivable magnitude and often unexpected behavior is considered bad.
In most cases, bad metals present a linear temperature dependence beyond the MIR limit. Bruin *et al.*[@Bruin:2013] recently noticed that the T-linear scattering rate in numerous metals, either ordinary or strange, has a similar magnitude. This observation suggests the relevance of a universal Planck timescale ($\tau_{P}\sim\frac{\hbar}{k_{B}T}$, where $k_{B}$ is the Boltzmann and $\hbar$ the reduced Planck constants) to electronic dissipation and motivated a theoretical attempt to quantify an upper limit to the diffusion constant in incoherent metals[@Hartnoll:2014]. During the past years, several theories of charge transport with no need for well-defined quasi-particles were proposed[@Mukerjee:2006; @Lindner:2010; @Andreev:2011; @Davison:2014; @Principi:2015; @Pakhira:2015; @Lim:2015; @Perep:2016]. More recently, direct measurements of thermal diffusivity[@Zhang:2016] have provided additional evidence for quasi-particle-free thermal transport at room temperature in cuprates.
We show here that n-doped SrTiO$_{3}$ is an unnoticed case of strange or bad metallicity. The resistivity of this dilute metal decreases by orders of magnitude upon cooling to cryogenic temperatures[@Fredrikse:1964; @Tufte:1967; @Fredrikse:1967; @Wemple:1969; @Spinelli:2010]. In the low-temperature limit, resistivity follows a T-square behavior[@Vandermarel:2011; @Lin:2015], as expected in a Fermi liquid. Above 100 K, however, the temperature dependence is close to cubic[@Tufte:1967; @Lin:2015]. In this regime, the mean-free-path of carriers becomes shorter than their wavelength. At room temperature, it falls well below the interatomic distance, the lowest conceivable length scale. Strontium titanate is the strangest known metal according to the criteria for strangeness widely used. Since quasi-particles in this system are not well-defined in a time scale long enough to allow the existence of a scattering time, charge transport cannot be reasonably described by a phonon-scattering picture. This was the approach in previous attempts[@Fredrikse:1964; @Tufte:1967; @Fredrikse:1967; @Wemple:1969; @Verma:2014; @Himmetoglu:2014; @Zhou:2016] to understand charge transport. We argue that metallicity is caused by the temperature dependence of the transmission coefficient along Landauer canals between dopants. Such a picture does not require a scattering time or a mean-free-path. We find an empirical link between mobility and permittivity, which gives a reasonable account of the experimental data. Providing a quantitative account of this observation emerges as a specific challenge to transport theory.
![**Breakdown of the Mott-Ioffe-Regel limit:** a) Temperature dependence of the mean-free-path in dilute metallic SrTiO$_{3-\delta}$ sample up to 400 K. Blue circles represent $\ell_{e}$, extracted from resistivity using Eq.1. Red diamonds represent the mean-free-path in the non-degenerate regime: $\ell'_{e}=\ell_{e}\frac{n^{-1/3}}{\Lambda}$. Also shown are the de Broglie thermal wavelength, $\Lambda$, the interelectron distance n$^{-1/3}$, the inverse of the Fermi wavelength ($k_{F}^{-1}$) and the lattice parameter, $a$=0.39 nm. The system remains metallic even at 400 K, in spite of exceeding the MIR limit by all conceivable criteria. b) A color plot of $\ell'_{e}$, extracted from the resistivity data of Fig. 1 in the (T, n) plane. Different crossovers are shown. The MIR limit is exceeded in a high-temperature window narrowing down with increasing concentration.[]{data-label="Fig2"}](Fig2V6.pdf){width="45.00000%"}
Compound k$_{F}$ (nm$^{-1}$) $\rho$(400K) (m$\Omega$cm) $\ell$(400K)(nm) k$_{F}$$\ell$ (400K) T$_{k_F\ell=1}$ (K) Reference
---------------------------------------------- --------------------- ---------------------------- ------------------ ---------------------- --------------------- ----------------------------------
La$_{1.72}$Sr$_{0.18}$Cu0$_{4}$ 6.6 0.32 0.35 2.2 850 [@Takagi:1992; @Gunnarsson:2003]
YBa$_{2}$Cu$_{4}$O$_{8}$ 1.4 0.36 0.8 1.1 450 [@Hussey:1997; @Bangura:2008]
Sr$_{2}$RuO$_{4}$ 5.6 0.2 0.38 2.1 800 [@Tyler:1998]
SrTiO$_{3-\delta}$ (n=4 10$^{17}$ cm$^{-3}$) 0.24 5900 0.2 0.05 100 This work
\[table:1\]
Results
=======
Metallicity beyond the MIR limit in strontium titanate
------------------------------------------------------
Fig. 1a shows the temperature dependence of resistivity in SrTiO$_{3-\delta}$ as the carrier concentration is changed from 2.4 10$^{17}$cm$^{-3}$ to 3.5 10$^{19}$cm$^{-3}$. At low temperature ($T\ll100K$), the system displays a T-square resistivity[@Vandermarel:2011; @Lin:2015] as expected for electron-electron scattering[@Pal:2012] in a Fermi liquid. Plotting resistivity as a function of T$^{2}$ (see Fig. 2 in ref.[@Lin:2015]), one can see that inelastic resistivity becomes T-square in the low-temperature limit. A wide range of dense Fermi liquids display Kadowaki-Woods scaling[@KW] of their T-square resistivity prefactor and their electronic specific heat. Since the latter is proportional to carrier concentration, the case of low-density systems such as doped SrTiO$_{3}$ is different[@Hussey:2005]. However, the scaling between the resistivity prefactor and Fermi energy persists. Indeed, the magnitude of this prefactor decreases by several orders of magnitude with changing carrier concentration[@Lin:2015].
Fig. 1b shows the temperature dependence of the extracted Hall mobility, $\mu=\frac{1}{ne\rho}$. Above 100 K, it displays a temperature dependence close to T$^{-3}$ and barely varies with carrier concentration. Our data is in agreement with the less extensive data reported previously by Tufte and Chapman[@Tufte:1967]. Fig. 1c shows how the exponent of inelastic resistivity evolves as a function of temperature. The total resistivity can be expressed as $\rho=\rho_{0}+AT^{\alpha}$. As seen in Fig. 1c, $\alpha$ is 2 at very low-temperature ($T<50K$) but smoothly rises with warming. One does not expect the T-square resistivity to survive at the Fermi degeneracy temperature. Therefore, the electron-electron scattering origin of the T-square term in the dilute limit has been put into question[@Maslov2017; @Swift2017]. We note, however, that the smooth variation of $\alpha$, makes the detection of the upper boundary of T-square resistivity impossible. As one can see in the figure, $\alpha$ peaks around 150 K at a value slightly exceeding 3 and then decreases with increasing temperature.
In the zero temperature limit, the magnitude of electron mobility, set by defect scattering can become remarkably large in doped strontium titanate[@Son:2010]. As a consequence, quantum oscillations are easily detectable in moderate magnetic fields[@Lin:2013; @Allen:2013; @Lin:2014]. The large magnitude of defect-limited mobility at low temperature and its evolution with carrier concentration can be explained in a simple model[@Behnia:2015] invoking the long effective Bohr radius of the quantum paraelectric[@Muller:1979] parent insulator.
The focus of the present paper is charge transport above 100 K where mobility presents a strong temperature dependence and almost no variation with carrier concentration over two orders of magnitude (Fig. 1b). Early studies[@Fredrikse:1964; @Tufte:1967; @Fredrikse:1967; @Wemple:1969] attributed the T$^{-3}$ variation of mobility to the scattering of carriers by phonons. There was a controversy between Fredrikse *et al.*[@Fredrikse:1967] (who invoked longitudinal optical phonons) and Wemple *et al.*[@Wemple:1969] (who argued in favor of transverse optical phonons). Neither during this early debate[@Fredrikse:1964; @Tufte:1967; @Fredrikse:1967; @Wemple:1969], nor in more recent theoretical accounts of the temperature dependence of mobility in n-doped SrTiO$_{3}$[@Baratoff:1981; @Verma:2014; @Himmetoglu:2014; @Mikheev:2015; @Zhou:2016], the magnitude of the mean-free-path was discussed.
In the Drude-Boltzmann picture, the conductivity, $\sigma$, of a metal depends on fundamental constants and a number of material-related properties. These are either, the scattering time $\tau$, and the effective mass, $m^{*}$, or the Fermi wave-vector, k$_{F}$ and the electron mean-free-path, $\ell_{e}$. Namely: $$\label{1}
\sigma=\frac{ne^{2}\tau}{m^{*}}=\frac{1}{3\pi^{2}}\frac{e^{2}}{\hbar}k_{F}^{2}\ell_{e}$$
The Fermi surface of n-doped SrTiO$_{3}$ in the dilute limit is known. It is located at the center of the Brillouin zone, because a band originating from Ti atoms’ t$_{2g}$ orbitals has a minimum at the $\Gamma-$point[@Vandermarel:2011; @Allen:2013; @Tao:2016]. The threefold degeneracy of this band is lifted by tetragonal distortion and spin-orbit coupling. As the system is doped, the three bands are filled successively and three concentric Fermi surfaces will emerge and grow in size one after the other. This is a picture based on band calculations[@Vandermarel:2011] and confirmed by experimental studies of quantum oscillations[@Lin:2013; @Allen:2013; @Lin:2014]. Below 10$^{18}$cm$^{-3}$, the first critical doping for the emergence of secondary pockets, there is a single Fermi surface with moderate (1.6) anisotropy[@Allen:2013; @Lin:2014; @Tao:2016]. Therefore, k$_F$ can be quantified in a straightforward manner. Fig. 2a shows the temperature dependence of the mean-free-path extracted from Eq. 1 in a dilute sample whose resistivity was measured up to 400 K. The carrier density is the one given by the Hall coefficient. As previously reported in ref. [@Tufte:1967] , the Hall coefficient does not display any notable temperature dependence. Moreover, the carrier density according to the magnitude of the Hall coefficient matches the Fermi surface volume according to the frequency of quantum oscillations[@Lin:2014]. Therefore, there is little ambiguity regarding carrier concentration.
As seen in Fig. 2a, the mean-free-path becomes shorter than the inverse of the Fermi wave-vector k$_{F}^{-1}$ and then the lattice parameter $a$. Now, at this carrier concentration, with an effective mass of m$^{*}$=1.8m$_{e}$, the Fermion degeneracy temperature is as low as 15 K[@Lin:2013]. As one sees in the figure, this is the temperature at which de Broglie thermal wavelength, $\Lambda=\frac{h}{\sqrt{2\pi m^{*}k_{B}T}}$ becomes comparable to the interelectron distance $n^{-1/3}$. Above this temperature, the electrons become non-degenerate and follow a Maxwell-Boltzmann distribution. As a consequence and thanks to thermal excitation, the typical carrier momentum is larger than $\hbar$k$_{F}$. In other words, the typical velocity is not the Fermi velocity, but the thermal velocity, $v_{T}= \sqrt{\frac{2k_{B}T}{m^{*}}}$. Let us define the mean-free-path of electrons with thermal velocity as $\ell'_{e}=\tau v_{T}=\ell_{e}n^{-1/3}/\Lambda$. As seen in the figure, because of $\ell'_{e}>\ell_{e}$, the crossover temperatures shift upward. However, the picture remains qualitatively the same. In the non-degenerate regime, the relevant wavelength is $\Lambda$ and not the Fermi wavelength. As seen in the figure, the mean-free-path becomes shorter than $\Lambda$ too.
A room-temperature resistivity of 2.4 m$\Omega$.cm implies a scattering time of $\tau \simeq 6fs$. This is shorter than the Planck time[@Bruin:2013] ($\tau_{P}=\frac{\hbar}{k_{B}T}$) at room temperature, $\tau_{P}(300K)=25.3 fs $. This is the typical time it takes for a carrier with a thermal velocity, v$_{T}$, to move as far as its de Broglie thermal wavelength ( $\frac{\Lambda}{v_{T}}=1.77\tau_{P}$). By all accounts, room-temperature quasi-particles in dilute metallic strontium titanate do not live long enough to qualify for a scattering-based picture.
A color plot of the mean-free-path in the (T, n) plane produced from the data of Fig.1 and Eq. 1 is presented in Fig. 2b. The MIR limit is exceeded in a high-temperature window, widening with the decrease in carrier concentration. The hitherto unnoticed specificity of SrTiO$_{3}$ is that it becomes a strange metal well below room temperature. The data in Table 1 gives an idea of the “strangeness” or the “badness” of SrTiO$_{3}$ compared to two other perovskytes. As seen in the table, at 400 K, k$_{F}\ell$ in this system is much lower than in cuprates or in Sr$_{2}$RuO$_{4}$. Contrary to the two other systems, the passage beyond the MIR limit occurs well below room temperature implying the necessity of a transport picture with no reference to well-defined quasi-particles. This is the first principal message of this paper.
![**Comparison with other dilute metallic systems:** a) R$_{\Box}$ of a one Fermi-wavelength-thick sheet of SrTiO$_{3-\delta}$ compared with graphene[@Efetov:2010] with comparable areal carrier density. b) Temperature-dependence of resistivity in SrTiO$_{3-\delta}$ and in metallic silicon[@Yamanouchi:1967] with comparable volume carrier concentration. c) Temperature-dependence of resistivity in SrTiO$_{3-\delta}$, metallic PbTe and metallic KTaO$_{3-\delta}$[@Wemple:1965] in similar range of carrier concentration. In silicon and graphene, the change in resistance caused by phonon scattering of electrons is a modest part of the total resistance. In all other systems, which are close to a ferroelectric instability, resistivity changes by several orders of magnitude.[]{data-label="Fig3"}](Fig3V3.pdf){width="42.00000%"}
Comparison with graphene and metallic silicon
---------------------------------------------
The temperature dependence of resistivity above 100 K and the magnitude of inelastic resistivity are very different from dilute metallic systems subject to phonon scattering. Let us compare strontium titanate with two exemplary systems. A recent study of charge transport in graphene[@Efetov:2010] has shown the relevance of the Bloch-Grüneisen picture of electron-phonon scattering in our temperature window of interest. In this picture, above a characteristic temperature (usually, but not always, the Debye temperature) resistivity displays a linear temperature dependence. This happens because all phonon modes capable of scattering electrons are thermally populated and the phase space for scattering is proportional to the thermal width near the Fermi level. Far below the characteristic temperature, the resistivity is expected to display a much stronger T$^{5}$ (in 3D) or T$^{4}$ (in 2D) temperature dependence because the typical wave-vector of the thermally excited phonons becomes smaller with decreasing temperature. As a consequence, the capacity of scattering phonons to decay the momentum flow decays rapidly with temperature. Efetov and Kim[@Efetov:2010] showed that this picture gives a very satisfactory account of inelastic resistivity in graphene as the carrier density is tuned from 1.36 to 10.8 10$^{13}$cm$^{-2}$. Let us note that in gated graphene, inelastic (that is, temperature-dependent) resistivity is far from dominating the total resistivity. This is seen in Fig. 3a, which compares the temperature-dependence of resistivity in n-doped SrTiO$_{3}$ and graphene at an identical interelectron distance. As one can see in the figure, resistivity changes by 200 in SrTiO$_{3}$ and by 1.6 in graphene. Charge conductivity is higher in graphene by a factor of two at low temperature and by two orders of magnitude at room temperature. A comparison with the case of metallic silicon[@Yamanouchi:1967] reveals a similar discrepancy. As one can see in Fig. 3b, cooling phosphorous-doped silicon leads to a modest change in resistivity, much less than what is seen in doped SrTiO$_{3}$. As a consequence, while at low-temperature, strontium titanate has a significantly higher mobility, the opposite is true at room temperature. In contrast to strontium titanate, in both graphene and silicon, one can use Eq. 1 one to extract a meaningful mean-free-path, which increases by a factor of two or less as the system is cooled down from room temperature to helium liquefaction temperature.
It is also instructive to compare SrTiO$_{3}$ with two other dilute metals close to a ferroelectric instability. As one can see in Fig.3c, both PbTe and KTaO$_{3}$ show comparably large changes in their resistivity upon cooling. Reports on KTaO$_{3}$[@Wemple:1965; @Wemple:1969] and on lead chalcogenide salts (PbTe, PbSe and PbS)[@Petritz:1955; @Allagier:1958] indicate that high-temperature mobility does not vary with carrier concentration, but shows a strong temperature dependence ($\mu \propto T^{-\lambda}$, with $\lambda\simeq2.5-3$). Both these features are similar to what we see in strontium titanate. Only in strontium titanate, however, the magnitude of the room temperature mobility is low enough to make it a clearly strange metal below room temperature.
To sum up, in silicon and graphene, the temperature-induced change in resistivity implies a modest change in mean-free-path and follows a Bloch-Gruneisen behavior (a slow T-linear behavior at high temperature and a fast T$^{4}$ or T$^{5}$ behavior at low temperature). On the other hand, doped SrTiO$_{3}$ and other metals close to a ferroelectric instability do not fit within this picture. The apparent mean-free-path of carriers changes by orders of magnitude. It becomes very short at high temperature, and in the case of strontium titanate, too short to be physically meaningful.
Discussion
==========
Conduction as transmission in a doped polar semiconductor
---------------------------------------------------------
![**Two pictures of temperature-dependent conductivity:** a) In a Drude-Boltzmann picture of transport, cooling the system leads to a rarefication of scattering events. Conductivity increases because the time interval (and the spatial distance) between two successive scattering events increases with cooling. Such a picture requires a well-defined quasi-particle across the distance in time and in space separating two scattering events. b) In a Landauer picture of transport, each dopant (each oxygen vacancy here) is a reservoir connected to other adjacent reservoirs by conducting channels. Conductivity increases upon cooling because the transmission coefficient enhances. Such a picture does not require well-defined quasi-particles. []{data-label="Fig4"}](Fig4V2.pdf){width="45.00000%"}
Fig. 4 sketches two alternative descriptions of charge conduction. In the semi-classic picture of metallic charge transport (Fig. 4a), conductivity increases with decreasing temperature, because the scattering events relaxing the momentum of a charge-carrying quasi-particle become less frequent with decreasing temperature. Now, consider an alternative picture (Fig. 4b), where each dopant is hosting a polaron, ‘an electron trapped by digging its own hole’[@Mott:1940] in a polar semiconductor. The polaron picture has been frequently invoked in the case of strontium titanate[@Eagles:1965; @Mechelen:2008].
In this second picture, conductivity is set by transmission among dopant sites and there is no need for well-defined quasi-particles. Can conductivity increase with decreasing temperature in such a picture and generate a metallic behavior? This can happen if cooling increases the transmission probability between adjacent dopant sites. In the case of strontium titanate, static permittivity increases with cooling[@Muller:1979]. Therefore, the lower the temperature, the shallower the donor potential. As a consequence, polarons become less bound with decreasing temperature and conductivity is metallic.
![**Length scales and transmission** a) Temperature dependence of different length scales at two different carrier densities. The de Broglie thermal wavelength, $\Lambda$, was calculated assuming an effective mass of m$^{*}$=1.8m$_{e}$, found by quantum oscillations at low temperature. The effective Bohr radius was calculated using this mass and the permittivity data reported in [@Muller:1979]. b) Transmission coefficient extracted from resistivity data for different carrier concentrations compared to $\Lambda^{2}$.[]{data-label="Fig5"}](Fig5V5.pdf){width="42.00000%"}
Let us now consider different length scales of the system. Fig. 5a shows them in our temperature range of interest where mobility becomes independent of carrier concentration. When carrier density changes from $2.4\times 10^{17}cm^{-3}$ to $1.7\times 10^{19}cm^{-3}$, the distance between electrons varies between 16 to 4 nm. Taking the low-temperature effective mass of m$^{*}$=1.8m$_{e}$ seen by quantum oscillations[@Lin:2013], one finds that the de Broglie thermal wave-length is $\Lambda\simeq 3-6 nm$ in this temperature range. Thus, the system is mostly non-degenerate. Electric permittivity, $\epsilon$, is large ($\epsilon(300K) \simeq300)$[@Muller:1979; @Lowndes:1973] and increases steadily with cooling. As a consequence, the effective Bohr radius (a$^{*}_{B}=\frac{4\pi\epsilon\hbar^{2}}{m^{*}e^{2}}$) varies between 8 and 40 nm. This puts the system above the Mott criterion for metallicity (n$^{1/3}$ a$^{*}_{B}> 0.25$[@Edwards:1978]), which, one shall not forget, is relevant to degenerate electrons[@Mott:1990]. These length scales are in the same range of magnitude. They are all longer than the lattice parameter (0.39nm). The question is how to picture conduction as transmission[@Imry:1999] along a random network of donors.
Since mobility is found to be independent of carrier concentration above 100K, we can express conductivity as:
$$\label{2}
\sigma=\rho^{-1}=G_{0}n <T>$$
Here, $G_{0}=\frac{2e^{2}}{h}$, is the quantum of conductance and $<T>$ represents the average transmission between a dopant site and its immediate neighbors. Note that in three dimensions, $<T>$ is proportional to mobility ($\mu=2\frac{e}{h}<T>$).
The temperature dependence of $<T>$, extracted from the experimental data using Eq. 2 is shown in Fig. 5b. As seen in the figure, a hundred-fold enhancement in carrier concentration does not affect $<T>$. One would have expected that the Landauer transmission between one site and its more or less distant neighbors would depend on the length of the trajectory and the details of the structural landscape. It is therefore surprising that the averaged transmission coefficient, $<T>$, is insensitive to such a large change in the average interdopant distance.
Fig. 5b also compares the magnitude of $<T>$ with $\Lambda^{2}$. the latter represents the areal uncertainty in the spatial localisation of a non-degenerate electron in three dimensions. As seen in the figure, $<T>$ decreases faster than $\Lambda^{2}$. There is little experimental information regarding the effective mass in this temperature range. A larger temperature-dependent effective mass would make $\Lambda^{2}$ smaller and its decrease faster. Thus, one can bridge the gap between $<T>$ and $\Lambda^{2}$ with a hypothetical temperature-dependent polaronic mass. However, there is no experimental evidence for this. At low temperatures, quantum oscillations[@Lin:2014; @Allen:2013] resolve a modest mass enhancement($\leq4$). This is also the case of infrared conductivity[@Mechelen:2008]. Thus, we cannot simply picture conductivity as ballistic charge transport along Landauer wires. Let us now consider a diffusive picture.
Mobility, diffusivity and their link to electric permittivity
-------------------------------------------------------------
In a non-degenerate semiconductor, mobility and diffusion constant, $D$, of carriers are linked through Einstein’s equation[@Ashcroft:1976]:
$$\label{3}
D=\frac{\mu k_{B}T}{e}$$
The temperature dependence of $D$ extracted from the mobility data of the sample with a carrier concentration of 2.4 10$^{17}$ cm$^{-3}$ is shown in Fig. 6a. Note that since mobility does not show any detectable variation with carrier concentration, this curve is representative of $D$ for carrier concentrations in the range of 10$^{17}$-10$^{19}$cm$^{-3}$. The magnitude of $D$ obtained in this way is remarkably low, well below 1cm$^{2}$/s. For comparison, the diffusion constant of electrons in silicon rises from 35 cm$^{2}$/s at room temperature to $\sim$100 cm$^{2}$/s at 100 K[@Brunetti:1981].
Hartnoll recently proposed a lower bound to diffusion constant in incoherent metals[@Hartnoll:2014]:
$$\label{4}
D_{H}=\frac{\hbar v_{F}^{2}}{k_{B}T}$$
As one can see in Fig. 6a, the experimental $D$ decreases faster than T$^{-1}$. Hartnoll boundary was obtained for degenerate systems. Ours is non-degenerate and the typical velocity is thermal velocity and not v$_{F}$. The Fermi velocity changes drastically with carrier density, but the experimental $D$ does not. As defined in Eq. 4 $, D_{H}$, does not appear to be directly relevant to our observation.
![**Diffusivity, ferroelectric soft mode and permittivity** a) Blue solid circles represent the diffusion constant extracted from mobility. The dashed line represents a T$^{-1}$ temperature dependence. Open circles represent the soft mode frequency, $\omega_{TO}^{-3}$ according to ref.[@Fleury:1968]. The two quantities scale with each other in this temperature range ($D\propto\omega_{TO}^{-3}$). b) The inverse of electric permittivity, $\epsilon^{-1}$[@Muller:1979] and $\omega_{TO}^{2}$ in the same temperature range. The scaling is $\epsilon^{-1}\propto\omega_{TO}^{2}$.[]{data-label="Fig5"}](Fig6V4.pdf){width="42.00000%"}
An obvious piece of the puzzle of charge transport in strontium titanate is the fact that the insulating parent is a quantum paraelectric with a strongly temperature-dependent electric permittivity[@Lowndes:1973; @Muller:1979]. We have found that in our temperature range of interest $D$ is proportional to $\epsilon^{1.5} $. This empirical observation is the second principal message of this paper.
The scaling between static electric permittivity, $\epsilon_{0}$, and the frequency of the Transverse Optic (TO) soft mode[@Fleury:1968; @Yamada:1969], $\omega_{TO}$ in strontium titanate is well-known. Yamada and Shirane[@Yamada:1969] showed that $\omega_{TO}=194.4\sqrt{\epsilon^{-1}}$ within experimental resolution. In a polar crystal, there is a theoretical link between the ratio of static, $\epsilon_{0}$, to high-frequency, $\epsilon_{\infty}$, permittivity and the ratio of longitudinal, $\omega_{L}$, to transverse, $\omega_{T}$, phonon frequencies. This is the expression first derived by Lyddane, Sachs, and Teller[@Lyddane:1941; @Cohran:1962]:
$$\label{5}
\frac{\epsilon_{0}}{\epsilon_{\infty}}=(\frac{\omega_{L}}{\omega_{T}})^2$$
When the transverse mode is soft, the system begins to show a large static polarizability. As one can see in Fig. 6b, the available experimental data confirm the scaling between these two distinct measurable quantities. A theory assuming an anisotropy in the polarizability of the oxygen ion[@Migoni:1976], the starting point of the polarization theory of ferroelectricity[@Bilz:1987], explains the temperature dependence of $\omega_{TO}$ in strontium titanate quantitatively. The soft mode and the large permittivity are properties of the insulating parent. However, they are known to survive in presence of dilute metallicity. Indeed, neutron scattering studies have explored the way the soft mode responds to the introduction of mobile electrons[@Bauerle; @Bussmann:1981] and found negligible shift in frequency for carrier densities below 7.8 10$^{19}$cm$^{-3}$.
As one can see in Fig. 6a, comparing our data with those reported by Yamada and Shirane[@Yamada:1969], one finds that: $$\label{7}
\mu k_{B}T \propto \omega_{T0}^{-3}\propto \epsilon/\omega_{T0}$$
An explanation for this empirical link between carrier mobility and lattice properties is beyond the scope of this paper and emerges as a challenge to theory.
Concluding remarks
------------------
The persistence of T-square resistivity in dilute metallic strontium titanate is puzzling[@Lin:2015]. However, its context is familiar. Electrons are degenerate and k$_{F}\ell >>1$. We now see that this low-temperature puzzle is only the tip of an iceberg. Above 100 K, the application of the Drude picture of conductivity yields a paradox: a carrier mean-free-path shorter than all conceivable length scales. Both the magnitude and the functional behavior of the temperature-induced change in resistivity are very different from what is expected according to the Bloch-Grüneisen law and what is seen in silicon and graphene. The aim of this paper is to attract theoretical attention to a hitherto unnoticed case of strange metallicity. It remains to be seen if electron viscosity plays any role in diffusive transport. On the experimental side, extending the available data to higher temperatures and lower doping concentration would map the boundaries of metallicity and the MIR limit in this system as well as other polar semiconductors.
It is a pleasure to thank Sean Hartnoll and Dmitrii Maslov for stimulating discussions. This work has been supported by an *Ile de France* regional grant, by Fonds ESPCI-Paris and by JEIP- Collège de France and funded in part by a QuantEmX grant from ICAM and the Gordon and Betty Moore Foundation through Grant GBMF5305 to KB.
[XX]{} N. E. Hussey, K. Takenaka and H. Takagi. Universality of the Mott–-Ioffe-–Regel limit in metals. Phil. Mag. **84**, 2847 (2004).
V. J. Emery and S. A. Kivelson. Superconductivity in bad metals. Phys. Rev. Lett. **74**, 3253(1995)
S. Sachdev. Bekenstein-Hawking Entropy and Strange Metals Phys. Rev. X **5**, 041025 (2015)
O. Gunnarsson, M. Calandra and J. E. Han. Colloquium: Saturation of electrical resistivity. Rev. Mod. Phys. **75**, 1085 (2003).
J. A. N. Bruin, H. Sakai, R. S. Perry and A. P. Mackenzie. Similarity of scattering rates in metals showing T-linear resistivity. Science **339**, 804 (2013).
S. A. Hartnoll. Theory of universal incoherent metallic transport. Nature Physics **11**, 54 (2015).
S. Mukerjee, V. Oganesyan and D. Huse. Statistical theory of transport by strongly interacting lattice fermions. Phys. Rev. B **73**, 035113 (2006).
N. H. Lindner and A. Auerbach. Conductivity of hard core bosons: A paradigm of a bad metal. Phys. Rev. B **81**, 054512 (2010).
A.V. Andreev, S. A. Kivelson and B. Spivak. Hydrodynamic Description of Transport in Strongly Correlated Electron Systems. Phys. Rev. Lett. **106**, 256804 (2011).
R. A. Davison, K. Schalm and J. Zaanen. Holographic duality and the resistivity of strange metals. Phys. Rev. B **89**, 245116 (2014).
A. Principi and G. Vignale. Violation of the Wiedemann-Franz law in hydrodynamic electron liquids. Phys Rev Lett **115**, 056603 (2015).
N. Pakhira and R. H. McKenzie. Shear viscosity of strongly interacting fermionic quantum fluids. Phys. Rev. B **92**, 125103 (2015).
K. Limtragool and P. Phillips. Power-law optical conductivity from unparticles: Application to the cuprates. Phys. Rev. B 92, 155128 (2015).
E. Perepelitsky, A. Galatas, J. Mravlje, R. Zitko, E. Khatami, B. S. Shastry, and A. Georges. Transport and optical conductivity in the Hubbard model: A high-temperature expansion perspective. Phys. Rev. B **94**, 235115 (2016).
J.-C. Zhang, E. M. Levenson-Falk, B. J. Ramshaw, D. A. Bonn, R. Liang, W. N. Hardy, S. A. Hartnoll, A. Kapitulnik. Anomalous Thermal Diffusivity in Underdoped YBa$_{2}$Cu$_{3}$O$_{6+x}$, arXiv:1610.05845 (2016)
H. P. R. Frederikse W. R. Thurber and W. R. Hosler. Electronic Transpoort in Strontium Titanate. Phys. Rev. **134**, A442 (1964).
O. N. Tufte and P. Chapman. Electron Mobility in Semiconducting Strontium Titanate. Phys. Rev. **155**, 796 (1967).
H. P. R. Frederikse and W. R. Hosler. Hall Mobility in SrTiO$_{3}$. Phys. Rev. **161**, 822 (1967).
S. H. Wemple, M. Didomenico, Jr., and A. Jayaraman. Electron Scattering in Perovskite-Oxide Ferroelectric Semiconductors. Phys. Rev. **180**, 547 (1969).
A. Spinelli, M. A. Torija, C. Liu, C. Jan and C. Leighton. Electronic transport in doped SrTiO$_{3}$: Conduction mechanisms and potential applications, Phys. Rev. B **81**, 155110 (2010).
D. van der Marel, J. L. M. van Mechelen and I. I. Mazin. Common Fermi-liquid origin of T$^{2}$ resistivity and superconductivity in n-type SrTiO$_{3}$. Phys. Rev. B **84**, 205111 (2011).
X. Lin, B. Fauqué and K. Behnia. Scalable T$^{2}$ resistivity in a small single-component Fermi surface. Science **349**, 945 (2015).
A. Baratoff and G. Binnig. Mechanism of superconductivity in SrTiO$_{3}$. Physica B$+$C **108**, 1335 (1981).
A. Verma, A.P. Kajdos, T. A. Cain, S. Stemmer, and D. Jena. Intrinsic Mobility Limiting Mechanisms in Lanthanum-Doped Strontium Titanate. Phys. Rev. Lett. **112**, 216601 (2014).
B. Himmetoglu, A. Janotti, H. Peelaers, A. Alkauskas, and C. G. Van de Walle. First-principles study of the mobility of SrTiO$_{3}$. Phys. Rev. B **90**, 241204(R) (2014).
E. Mikheev, B. Himmetoglu, A. P. Kajdos, P. Moetakef, T. A. Cain, C. G. Van de Walle, and S. Stemmer. Limitations to the room temperature mobility of two- and three-dimensional electron liquids in SrTiO$_{3}$. Appl. Phys. Lett. **106**, 062102 (2015)
W. X. Zhou *et al.* Electron–soft phonon scattering in n -type SrTiO$_{3}$. Phys. Rev. B **94**, 195122 (2016).
H. K. Pal, V. I. Yudson, and D. L. Maslov, Lith. J. Phys. **52**, 142 (2012).
K. Kadowaki, S.B. Woods. Universal relationship of the resistivity and specific heat in heavy-fermion compounds. Solid State Commun. **58**, 507 (1986).
N. E. Hussey. Non-generality of the Kadowaki–Woods Ratio in Correlated Oxides. J. Phys. Soc. Jpn. **74**, 1107 (2005)
M. Swift and C. G. Van de Walle. Conditions for T$^{2}$ resistivity from electron-electron scattering. arXiv:1701.04744
D. L. Maslov and A. V. Chubukov. Optical response of correlated electron systems.Rep. Prog. Phys. **80**, 026503 (2017).
J. Son, P. Moetakef, B. Jalan, O. Bierwagen, N. J. Wright, R. Engel-Herbert and S. Stemmer. Epitaxial SrTiO$_{3}$ films with electron mobilities exceeding 30,000 cm$^{2}V^{-1}s^{-1}$. Nature Materials **9**, 482 (2010).
X. Lin, Z. Zhu, B. Fauqué, and K. Behnia. The Fermi surface of the most dilute superconductor. Phys. Rev. X **3**, 021002 (2013).
S. J. Allen, B. Jalan, S. B. Lee, D. G. Ouellette, G. Khalsa, J. Jaroszynski, S. Stemmer, and A. H. MacDonald. Conduction-band edge and Shubnikov-de-Haas effect in low-electron-density SrTiO$_{3}$. Phys. Rev. B **88**, 045114 (2013).
X. Lin, G. Bridoux, A. Gourgout, G. Seyfarth, S. Krämer, M. Nardone, B. Fauqué and K. Behnia. Critical Doping for the Onset of a Two-Band Superconducting Ground State in SrTiO$_{3 - \delta}$. Phys. Rev. Lett. **112**, 207002 (2014).
K. Behnia, On mobility of electrons in a shallow Fermi sea over a rough seafloor. J. Physics: Condens. Matt. **27**, 375501 (2015).
K. A. Müller and H. Burkard SrTiO$_{3}$: An intrinsic quantum paraelectric below 4 K. Phys. Rev. B **19**, 3593 (1979).
R. A. Cooper Y. Wang, B. Vignolle, O. J. Lipscombe, S. M. Hayden, Y. Tanabe, T. Adachi, Y. Koike, M. Nohara, H. Takagi, C. Proust and N. E. Hussey. Anomalous Criticality in the Electrical Resistivity of La$_{2–x}$Sr$_{x}$CuO$_{4}$. Science **323**, 603 (2008).
Q. Tao, B. Loret, B. Xu, X. Yang, C. W. Rischau, X. Lin, B. Fauqué, M. J. Verstraete, and K. Behnia. Nonmonotonic anisotropy in charge conduction induced by antiferrodistortive transition in metallic SrTiO$_{3}$. Phys. Rev. B **94**, 035111 (2016).
H. Takagi *et al.* Systematic Evolution of Temperature-Dependent Resistivit in La$_{2-x}$Sr$_{x}$Cu0$_{4}$. Phys. Rev. Lett. **69**, 2975 (1992).
N. E. Hussey, K. Nozawa, H. Takagi, S. Adachi and K. Tanabe. Anisotropic resistivity of YBa$_{2}$Cu$_{4}$O$_{8}$: Incoherent-to-metallic crossover in the out-of-plane transport. Phys. Rev. B **56**, R11 423(R) (1997).
A. F. Bangura *et al.* Small Fermi Surface Pockets in Underdoped High Temperature Superconductors: Observation of Shubnikov-–de Haas Oscillations in YBa$_{2}$Cu$_{4}$O$_{8}$. Phys. Rev. Lett. **100**, 047004 (2008).
A. W. Tyler, A. P. Mackenzie, S. NishiZaki, and Y. Maeno. High-temperature resistivity of Sr$_{2}$RuO$_{4}$: Bad metallic transport in a good metal. Phys. Rev. B **58**, R10107(R) (1998)
D. K. Efetov and P. Kim. Controlling Electron-Phonon Interactions in Graphene at Ultrahigh Carrier Densities. Phys. Rev. Lett. **105**, 256805 ( 2010).
C. Yamanouchi, K. Mizuguchi and W. Sasaki. Electric Conduction in Phosphorus Doped Silicon at Low Temperatures. J. Phys. Soc. Jpn. **22**, 859 (1967).
S. H. Wemple. Some Transport Properties of Oxygen-Deficient Single-Crystal Potassium Tantalate (KTaO$_{3}$). Phys. Rev. **137**, A1575 (1965).
R. L. Petritz and W. W. Scanlon. Mobility of Electrons and Holes in the Polar Crystal, PbS. Phys. Rev. **97**, 1620 (1955)
R. S. Allgaier and W. W. Scanlon. Mobility of Electrons and Holes in PbS, PbSe, and PbTe between Room Temperature and 4.2K. Phys. Rev. **111**, 1029 (1958)
N. F. Mott and R. W. Gurney. Electronic Processes in Ionic Crystals. Clarendon Press, Oxford (1940)
P. P. Edwards and M. J. Sienko. Universality aspects of the metal-nonmetal transition in condensed media. Phys. Rev. B **17**, 2575 (1978).
N. F. Mott. Metal-insulator transitions. Second edition, Taylor and Francis, London (1990)
R. P. Lowndes and A. Rastogi. Stabilization of the paraelectric phase of KTaO$_{3}$ and SrTiO$_{3}$ by strong quartic anharmonicity. J. Phys. C: Solid State Phys. **6**, 932 (1973).
N. W. Ashcroft and N. D. Mermin, Solid State Physics, New York : Holt, Rinehart and Winston (1976).
Y. Imry Y. and R. Landauer. Conductance viewed as transmission. Rev. Modern Phys. **71**, S306 (1999).
D. M. Eagles. Polaron coupling constants in SrTiO$_{3}$. J. Phys. Chem. Solids **26**, 672 (1965)
J.L.M van Mechelen, D. van der Marel, C. Grimaldi, A. B. Kuzmenko, N. P. Armitage, N. Reyren, H. Hagemann, and I. I. Mazin. Electron-phonon interaction and charge carrier mass enhancement in SrTiO$_{3}$. Phys. Rev. Lett. **100**, 226403 (2008).
R. Brunetti, C. Jacoboni, F. Nava, L. Reggiani, G. Bosman and R. J. J. Zijlstra. Diffusion coefficient of electrons in silicon. J. Appl. Phys. **52**, 6713 (1981)
P. A. Fleury and J. M. Worlock. Electric-Field-Induced Raman Scattering in SrTiO$_{3}$ and KTaO$_{3}$. Phys. Rev. **174**, 613 (1968)
Y. Yamada, and G. Shirane. Neutron Scattering and Nature of the Soft Optical Phonon in SrTiO$_{3}$. J. Phys. Soc. Jpn. **26**, 396(1969)
R. H. Lyddane, R. G. Sachs, and E. Teller. On the Polar Vibrations of Alkali Halides. Phys. Rev. **59**, 673 (1941)
W. Cochran and R.A. Cowley. Dielectric constants and lattice vibrations. J. Phys. and Chem. Sol. **23**, 447 (1962)
R. Migoni, H. Bilz, and D. D. Bäuerle. Origin of Raman Scattering and Ferroelectricity in Oxidic Perovskites. Phys. Rev. Lett. **37**, 1155 (1976)
H. Bilz, G. Benedek, and A. Bussmann-Holder. Theory of ferroelectricity: The polarizability model. Phys. Rev. B **35**, 4840 (1987)
D. Bäuerle, D. Wagner, M. Wöhlecke, B. Dorner and H. Kraxenberger. Soft modes in semiconducting SrTiO$_{3}$: II. The ferroelectric mode. Z. Physik B - Condensed Matter **38**, 335 (1980)
A. Bussmann-Holder, H. Bilz, D. Bäuerle, D. Wagner. A polarizability model for the ferroelectric mode in semiconducting SrTiO$_{3}$. Z. Physik B - Condensed Matter **41**, 353 (1981)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report on the first successful growth of single crystals of the noncentrosymmetric superconductor Mo$_3$Al$_2$C obtained by means of a cubic-anvil, high-pressure and high-temperature technique. Composition, structure, and normal-state transport properties of the crystals were studied by means of X-ray diffraction, energy-dispersive X-ray spectroscopy, magnetic susceptibility and resistivity measurements as a function of temperature. Variations in critical temperature ($T_c$) between 8.6 and 9.3 K were observed, probably due to the slightly different carbon stoichiometry of the samples. Single-crystal X-ray refinement confirmed the high structural perfection of the grown crystals. Remarkably, the refined Flack parameter values for all the measured crystals using a $P4_132$ space-group model were consistently close to either 0 or 1, hence indicating that the considered crystals belong to two enantiomorphic space groups, $P4_132$ and $P4_332$. An anomaly in the resistivity is observed at $\simeq$ 130 K, most likely associated with the onset of a charge-density-wave phase. The superconducting properties (and in particular the symmetry, the amplitude and the temperature dependence of the superconducting gap) were studied by using – for the first time in this compound – point contact Andreev-reflection spectroscopy. The results confirm that Mo$_3$Al$_2$C is a moderately strongly-coupled superconductor with $2\Delta/k_\mathrm{B}T_c \simeq 4$ and unambiguously prove that the order parameter has an $s$-wave symmetry despite the asymmetric spin-orbit coupling arising from the lack of inversion symmetry.'
author:
- 'N. D. Zhigadlo'
- 'D. Logvinovich'
- 'R. S. Gonnelli'
- 'V. A. Stepanov'
- 'D. Daghero'
title: 'Crystal growth, characterization and advanced study of the noncentrosymmetric superconductor Mo$_3$Al$_2$C'
---
Introduction {#sec:intro}
============
Noncentrosymmetric superconductors (NCS) have recently attracted considerable attention. The discovery of a mixing of spin-singlet and spin-triplet states in the heavy fermion NCS CePt$_3$Si (with critical temperature $T_c \simeq 0.75$ K [@Bauer2004]) has boosted a research aimed at investigating the effects on the superconducting pairing of the lack of inversion symmetry. The latter is reflected in the electronic structure through the spin-orbit coupling; a strong asymmetric spin-orbit coupling indeed results in a spin-splitting of the energy bands [@Mineev2012]. Actually, heavy-fermion compounds also display a strong electronic correlation (not to mention the complications due to the localized magnetic moments arising from the rare-earth $f$ electrons) so that they are not the simplest materials where to study the effects of a lack of inversion symmetry. In the past decade, many other NCS with a strong spin-orbit coupling have been investigated, displaying relatively high critical temperatures ($T_c > 9$ K) and critical magnetic fields, based on transition metals (Mo, Pd, Pt...) such as, for example, Y$_2$C$_3$ [@Amano2004], Li$_2$Pd$_x$Pt$_{3-x}$B [@Togano2004], Mo$_3$Al$_2$C [@Bauer2010] that do not present the aforementioned complications.
Mo$_3$Al$_2$C, in particular, was identified as an interesting candidate for unconventional superconductivity driven by the lack of inversion symmetry [@Bauer2010; @Karki2010]. The compound has a $\beta$-Mn type structure, with space group $P4_132$ and a $T_c$ of ca. 9K. Experimental studies of Mo$_3$Al$_2$C polycrystals have been carried out both in the normal and in the superconducting state. In the normal state, some anomalies in the resistivity, susceptibility, and nuclear-magnetic-resonance (NMR) measurements were interpreted as being due to the opening of a charge-density-wave gap [@Koyama2011; @Koyama2013] at about 130 K. In the superconducting state, early results of NMR and specific-heat measurements were interpreted as arising from a non conventional superconducting coupling, possibly with a nodal gap [@Bauer2010]. However, the low-temperature specific heat behaviour was found to be exponential, as expected for an isotropic gap [@Karki2010], and some of the anomalies at higher temperature were later attributed to minor impurity phases. Penetration-depth measurements confirmed this picture evidencing a temperature-independent penetration depth below 0.5 K [@Bonalde2011]. In polycrystals of improved quality, NMR measurements showed a peak in $1/T_1$ just below the transition and a low-temperature exponential behaviour consistent with a pure $s$-wave order parameter [@Koyama2013]. More recently, the absence of time-reversal symmetry breaking in Mo$_3$Al$_2$C was concluded by $\mu SR$ measurements [@Bauer2014].
A direct measure of the gap by means of a spectroscopic technique such as tunnel spectroscopy or Andreev-reflection spectroscopy is however missing. In particular, the necessity to carry out direct gap measurements in single crystals has been evidenced by the aforementioned, strong effect of the impurity phases on the measured quantities.
Experimental details and crystal growth {#sec:growth}
=======================================
For the growth of Mo$_3$Al$_2$C single crystals, we used the cubic-anvil, high-pressure, and high-temperature technique. The details of experimental set up can be found in our previous publications [@growth1; @growth2]. The apparatus consist of a 1500-ton press, with a hydraulic-oil system comprising a semi-cylindrical multianvil module (Rockland Research Corp.). A set of steel parts transmit the force through six tungsten carbide pistons to the sample in a quasi isostatic way. This method was successfully applied earlier on to grow various compounds, including superconducting and magnetic intermetallic crystals, diamonds, pyrochlores, $Ln$Fe$Pn$O ($ Ln$: lanthanide, $Pn$: pnictogen), oxypnictides, and numerous other compounds [@highpressure].
The main challenges when growing single crystals of Mo$_3$Al$_2$C are the relatively poor reactivity of C and the huge differences in the melting temperatures of components: 2617$^\circ$C for Mo, 660$^\circ$C for Al, and 3827$^\circ$C for C. This is probably the reasons why it is so difficult to synthesize even single-phase polycrystalline samples. As already shown in previous studies, the superconductivity in this material is very sensitive to the details of the heat treatment and to final stoichiometry [@Reith2015]. In order to shed light on these metallurgical problems one needs to perform systematic studies on single crystals. We note here that any high-temperature growth in quartz ampoules that utilizes Al as a flux cannot be performed easily, since Al vapor attacks the quartz, leading to a loss of the protective atmosphere. However, as we will show in the present study, the single crystals of Mo$_3$Al$_2$C can be grown successfully by using a high-pressure closed system.
In order to better understand the origin of variation in $T_c$ we synthesized more than twenty different compositions with different stoichiometries under high-pressure conditions. The most representative stoichiometries and synthesis protocols are depicted schematically in Fig. \[fig:protocol\].
{width="\textwidth"}
A mixture of Mo, Al, and C powders in a molar ratio Mo$_3$Al$_{2.0-2.2}$C$_{0.87-1.1}$ was placed inside a boron nitride (BN) crucible with an inner diameter of 8.0 mm, and a length of 9.0 mm. The heating element is a graphite tube. Six tungsten carbide anvils generate pressure on the whole assembly. In a typical run, a pressure of 3 GPa is applied at room temperature. The mixture was heat-treated according to the synthesis protocols summarized in Fig. \[fig:protocol\] while keeping the pressure constant throughout the growth process. The pressure was released only at the end of the crystal growth. The results of the different loadings and protocols can be summarized as follows:
- The perfectly stoichiometric Mo$_3$Al$_2$C loadings, heat-treated at 1550$^{\circ}$C and then slowly cooled, resulted in single crystals with $T_c$ ranging between 8.55 and 9.25 K (see Fig. \[fig:protocol\], protocols A and B).
- The Al-rich, i.e., Mo$_3$Al$_{2.2}$C loadings (protocol C) yielded samples with $T_c$ = 8.55 - 8.75 K.
- The C-rich Mo$_3$Al$_2$C$_{1.1}$ loadings (protocol D) yielded crystals with $T_c$ = 9.2 - 9.3 K, that however contained also a Mo$_2$C phase as impurity besides the Mo$_3$Al$_2$C phase, as revealed through a kink in the magnetic susceptibility at 3 K.
- The stoichiometric Mo$_3$Al$_2$C and the C deficient Mo$_3$Al$_2$C$_{0.87}$ loadings heat-treated by protocols E and F led to single-phase samples with $T_c = 9.05 - 9.3$ K and $T_c= 9.15 - 9.3$ K, respectively.
The origin of the difference in $T_c$ between different crystals is not clear yet, but current studies suggest that it is most likely due to slightly different sample stoichiometries. Indeed, nominally carbon-deficient crystals (as, e.g., Mo$_3$Al$_2$C$_{0.87}$) feature a higher $T_c$ with respect to nominally stoichiometric ones. This raises the question whether the formation of C vacancies is thermodynamically possible. Considering the well-established crystal structure such an idea looks surprising, as no Mo or Al vacancies could be detected. However, the situation is different for carbon which, as a light atom, makes it difficult to establish experimentally the actual C content. One should also consider that carbides normally exhibit vacancies in their sublattices [@Cvacancies]. We note here that polycrystalline Mo$_3$Al$_2$C samples show a rather wide spread of $T_c =9.0 - 9.3$ K, indicating that such samples are not perfectly stoichiometric. All together, these results indicate that the studied compound might be stabilized by vacancy formation.
The influence of stoichiometry, and in particular of carbon deficiency, on the thermodynamic stability and on the electronic structure of Mo$_3$Al$_2$C has been studied theoretically by means of DFT calculations [@Reith2015]. The electronic density of states at the Fermi level, which is an important parameter that controls the critical temperature, is enhanced by carbon vacancies. The phonon DOS $D(\omega)$ is heavily affected as well: on increasing the concentration of carbon vacancies, the low-energy part of the spectrum becomes increasingly Debye-like \[$D(\omega) \propto \omega^2$\]. Although a detailed evaluation of the effect of such changes on the critical temperature requires a precise knowledge of the electron-phonon spectral function $\alpha^2F(\omega)$ (rather than of the phonon DOS alone), the link between the concentration of C vacancies and the critical temperature is highly plausible.
In our experimental conditions we observed variations in $T_c$ even within individual batches. Therefore, besides the starting nominal composition and synthesis temperature, the temperature gradient across the sample is another important parameter that controls $T_c$. For a cell temperature of $1500-1550^{\circ}$C, we estimate a temperature gradient across the sample of ca. $70^{\circ}$C. The whole assembly produces, therefore, a parabolic temperature profile across the sample, with the maximum temperature being reached at its center (for more details, see Ref. [@growth1]). Thus, one may expect the formation of Mo$_3$Al$_2$C with slightly lower $T_c$s at the top and bottom parts of the BN crucible, whereas crystals with a higher $T_c$, i.e., Mo$_3$Al$_2$C$_{0.87}$, will grow close to the centre of the crucible, where the temperature is highest. Our experimental findings are consistent with DFT calculations reported in Ref. [@Reith2015] and demonstrate that superconducting transition temperature and carbon vacancy concentration can be controlled by the synthesis protocol. It is quite plausible that quenching of the samples from high temperatures will result in the freezing of high concentration of C vacancies ($\simeq$ 0.13 - 0.14), whereas at slow cooling process the vacancy concentration might be reduced since more carbon atoms can be introduced in the lattice.
![Optical microscope images of Mo$_3$Al$_2$C crystals. (a) Top view of as-grown solidified lump after radial crushing. The lump was limited by the crucible and the circular cross section is due to the crucible walls. (b) Mechanically extracted Mo$_3$Al$_2$C crystals.[]{data-label="fig:crystals"}](crystals.eps){width="\columnwidth"}
Figure \[fig:crystals\] shows optical microscope images of Mo$_3$Al$_2$C single crystals. Fig. \[fig:crystals\]a shows a top view of the as-grown solidified lump with a mixture of single-crystalline Mo$_3$Al$_2$C and some flux. The lump was crucible-limited and its circular cross section is due to the crucible walls. After crushing, a large number of crystals with various shapes, and sizes up to $1\,\mathrm{mm}^3$ were found (Fig. \[fig:crystals\]b). The stoichiometry of the grown crystals was checked by energy-dispersive X-ray spectroscopy. For each growth batch, 5 to 7 crystals were analyzed. Their Mo and Al content was always found to be close to the stoichiometric 3:2 composition within experimental resolution 0.1. The search for possible C deficiencies in nominally deficient samples did not prove conclusive due to substantial fluctuations of the experimental points between 0.8 and 1.0.
A magnetic determination of the critical temperature for each growth batch was performed on clean single crystals with sizes of few hundreds of micrometers. Some examples of zero-field-cooled (ZFC) temperature-dependent measurements in low field are shown in Fig. \[fig:magnetization\]. The spread of $T_c$ for each individual batch is indicated in Fig. \[fig:protocol\]. In general, the transition to the superconducting state is very narrow; the “10-90” criterion gives $\delta T_c = T_c^{90}-T_c^{10} \simeq 0.3$ K. The good quality of the grown crystals is confirmed by the resistivity studies presented in section IV.
![Zero-field-cooled magnetization curves for Mo$_3$Al$_2$C crystals extracted from various growth batches. The superconducting transition is rather narrow and the critical temperature varies between 8.6 and 9.3 K.[]{data-label="fig:magnetization"}](magnetization.eps){width="\columnwidth"}
Structure refinement {#sec:structure}
====================
Single-crystal X-ray diffraction experiments were performed with an Oxford Diffraction diffractometer equipped with an Onyx CCD detector, a Mo X-ray tube and a graphite monochromator adjusted to select K lines. Several crystals were measured to obtain representative crystal structures. The measurements at $\lambda=0.71073$ [Å]{} were performed with an oscillation angle of 1$^{\circ}$. Reflection integration and data reduction were performed using the software package CrysAlis from Oxford Diffraction. Structure solution and the determination of atomic positions were performed using the programs SUPERFLIP [@Palatinus2007] and EDMA [@Palatinus2012] respectively. All the refinements were done with SHELXL97 [@Sheldrick2008].
![Reconstructed $h0l$ layer of the reciprocal space. Red circles mark some of reflections that satisfy the reflection condition $h00: h=4n$.[]{data-label="fig:Xray"}](Xray.eps){width="\columnwidth"}
The Fourier spectra of all the measured crystals exhibit an $m\overline{3}m$ Laue symmetry. The reflection conditions we find ($h00$: $h = 4n$) correspond to two enantiomorphic space groups, $P4_132$ and $P4_332$ (Fig.\[fig:Xray\]). These can be distinguished based on the Flack parameter value [@Flack1983], provided the anomalous signal is strong enough, a sufficient number of Friedel pairs is measured, and special attention is paid to the absorption corrections. These factors are critical to achieve a low estimated standard deviation (*e.s.d.*) for the Flack parameter and, hence, a reliable conclusion about the absolute crystal structure. To facilitate the accurate refinement of the Flack parameter, datasets with a high redundancy and a coverage close to 100 were collected for all the crystals. Regardless of all the precautions adopted, the minimal achieved *e.s.d.*’s for the Flack parameter value were about 0.1, which is somewhat higher than the recommended value of less than 0.04. However it is remarkable that Flack parameter values refined for all the measured crystals using the model with the space group $P4_132$ were reproducibly close to either 0 or 1, a strong indication that the studied samples were a mixture of crystals with space groups $P4_132$ and $P4_332$. The results of two representative refinements are reported in Table \[tab:1\] and in Ref. [@suppl]. The refined atomic positions and displacement parameters for the model with space group $P4_132$ are presented in Tables \[tab:2\] and \[tab:3\].
[|l|l|l|]{}\
Chemical formula & Mo$_3$Al$_2$C & Mo$_3$Al$_2$C\
$T_c$ (prot. B) & 8.9 K & 8.9 K\
$M_r$ (g/mol) & 353.79 & 353.79\
$Z$ & 4 & 4\
Cell setting, space group & Cubic, $P4_132$ & Cubic, $P4_332$\
$a$ ([Å]{}) & 6.86715(3) & 6.86793(3)\
$V$ ([Å]{}$^3$) & 323.839(3) & 323.950(3)\
$\mu$ (mm$^{-1}$) & 11.697 & 11.697\
Crystal shape, color & Platelet, black & Platelet, black\
Crystal size (mm$^3$) & $0.08 \times 0.06 \times 0.02$ & $0.06 \times 0.04 \times 0.02$\
Temperature (K) & 298 & 298\
Radiation type & Mo$K \alpha$ & Mo$K \alpha$\
Radiation wavelength ([Å]{}) & 0.71073 & 0.71073\
\
Diffractometer & Goniometer Xcalibur, & Goniometer Xcalibur\
Detector & Onyx CCD & Onyx CCD\
Data collection method & scans & scans\
Scan width (deg) & 1.0 & 1.0\
Absorption correction & Multi-scan & Multi-scan\
$T_{min}$ & 0.42981 & 0.72247\
$T_{max}$ & 1.0000 & 1.0000\
Number of measured reflections & 11189 & 12086\
Number of independent reflections & 293 & 294\
Number of observed reflections & 289 & 289\
Criterion for observed reflections & $ I > 2 \sigma (I)$ & $ I > 2 \sigma (I)$\
$R_{int}$(all), $R \sigma$(all) & 0.0483, 0.090 & 0.0339, 0.0070\
$\theta_{max}$ (deg) & 37.6 & 37.6\
\
Refinement on & $F^2$ & $F^2$\
Weighting scheme & $w = 1/[\sigma^2(F_o^2)+(0.0178 P)^2+0.8687 P]$, & $w = 1/[\sigma^2(F_o^2)+(0.0233 P)^2+0.2695 P]$\
& where $P=(F_o^2+2F_c^2)/3$ & where $P=(F_o^2+2F_c^2)/3$\
$R_{obs}$ / $R_{all}$ / $wR_{obs}$ / $wR_{all}$ & 0.0170/0.0175/0.0378/0.0379 & 0.0157/0.0167/0.0370/0.0373\
Goodness-of-fit (obs/all) & 1.196/1.196 & 1.208/1.208\
No. of parameters & 13 & 13\
$(\Delta/\sigma)_{max}$ & 0.825 & 0.176\
$\Delta \rho_{max}$, $\Delta \rho_{min}$ (e [Å]{}$^{-3}$) & 0.681, -0.778 & 0.842, -0.931\
Extinction correction & SHELXL97 & SHELXL97\
Extinction coefficient & 0.0276(18) & 0.0122(14)\
Absolute structure & Flack(1983) & Flack(1983)\
Flack parameter & -0.07(13) & -0.02(11)\
Atom Wyckoff position $x$ $y$ $z$ $U_{iso}/U_{eq}$ ([Å]{}$^2$)
------ ------------------ ------------- ------------- ------------- ------------------------------
Mo 12$d$ $1/8$ 0.79721(3) 0.04721(3) 0.00933(10)
Al 8$c$ 0.93283(10) 0.43283(10) 0.06717(10) 0.0067(2)
C 4$b$ $3/8$ $5/8$ $1/8$ 0.0104(8)
$U^{11}$ $U^{22}$ $U^{33}$ $U^{23}$ $U^{13}$ $U^{12}$
---- ------------- ------------- ------------- -------------- -------------- -------------
Mo 0.01304(14) 0.00748(10) 0.00748(10) -0.00031(8) -0.00190(6) 0.00190(6)
Al 0.0067(2) 0.0067(2) 0.0067(2) -0.00037(19) -0.00037(19) 0.00037(19)
C 0.0104(8) 0.0104(8) 0.0104(8) -0.0014(9) 0.0014(9) 0.0014(9)
The crystalline structure of Mo$_3$Al$_2$C corresponds to that of $\beta$-Mn: both Mo and Al atoms are in a distorted icosahedron, while C atoms are in a distorted octahedral coordinations built from metal atoms (Fig. \[fig:structure\]). The space groups $P4_132$ and $P4_332$ differ by inversion symmetry, arising from the pair of enantiomorphic axes $4_1$ and $4_3$. The difference between the two axes can be explained with reference to Fig. \[fig:structure\]. By considering the $z$-coordinate of the Al-atom closest to $4_1z$-axis, its symmetry-equivalent is obtained via a counterclockwise rotation by $90^{\circ}$ with subsequent translation by $c/4$ along the $c$-axis. The translation component of the $4_3$ $z$-axis is the same, but the rotation occurs clockwise. Hence, in the former case Al atoms (as well as Mo and C) ascend counterclockwise along the $c$-axis, whereas in the latter case Al atoms ascend clockwise along the $c$-axis. The same is valid for the coordination polyhedra. To further confirm the enantiomorphic purity of the studied crystals, single crystal X-ray diffraction at the other wavelengths or at lower temperature could be performed.
Electric transport measurements
===============================
Transport measurements were carried out as a function of temperature either in vacuum (using a Cryomech$^\mathrm{TM}$ ST-403 3K pulse-tube cryocooler) or in He exchange gas, by mounting the crystals on the cold head of a cryogenic insert. Figure \[fig:res\]a shows the resistivity of crystal B/2 (i.e., obtained with the protocol B, see Fig. \[fig:protocol\]) as a function of temperature (blue circles). The residual resistivity is $\rho_0 = 125 \, \mu \Omega$ cm, i.e., of the same order of magnitude as that reported for some polycrystalline samples [@Bauer2010] but, unlike in polycrystals [@Bauer2010; @Karki2010; @Koyama2013], a rather clear minimum in the resistivity is observed at $T^*=134 \pm 4$ K (see Fig. \[fig:res\]b). It is interesting to note that the opening of a charge density wave (CDW) gap at about 130 K was proposed [@Koyama2011] based on $^{27}\mathrm{Al}$ nuclear magnetic resonance measurements that showed anomalies in the Knight shift and in the spin-lattice relaxation rate [@Kuo2012]. A signature of these anomalies was later found in the resistivity of polycrystals in the form of a slope change in the $\rho(T)$ curve [@Koyama2013].
![(a) Resistivity of the crystal B/2 (blue circles, left-hand vertical axis) and resistance of the crystal F/3 (red circles, right vertical axis) as a function of temperature up to 200 K. Inset: detail of the superconducting transitions. (b) Magnification of the normal-state resistivity of B/2 showing a clear minimum around 130 K.[]{data-label="fig:res"}](resistivity3.eps){width="\columnwidth"}
As clearly shown in Fig. \[fig:res\]a, the superconducting transition of the B/2 crystal is rather sharp, with an onset at $T_{c}^{on}=8.48 \pm 0.02$ K and a zero-resistance state achieved at $T_c^0=8.35\pm 0.02$ K. Note that the critical temperature agrees very well with that determined via magnetization ($T_c^{M,on}=8.55$ K) and, as in all the crystals grown with protocol B, is smaller than that reported in the literature for this compound ($T_c=9.3$ K).
Fig. \[fig:res\]a reports also the resistance of the C-deficient crystal F/3 (grown by using the protocol F) as red symbols; the irregular shape of the crystal did not allow us to reliably estimate the resistivity. In this case $T_c^{on}= 9.27 \pm 0.02$ K and $T_c^0=9.19 \pm 0.02$ K. The critical temperature resulting from magnetization measurements in this case was $T_c^{M,on} = 9.3$ K, again in perfect agreement with the transport results. The $R(T)$ curve in this case is monotonic, i.e., does not exhibit any minimum around 130 K. At this temperature, a smooth slope change is instead observed, as in polycrystals [@Koyama2013].
![(a) I-V characteristics of the B/2 crystal, as a function of temperature. (b) Red solid circles: values of the pseudocritical current $j_m$ (minimum normal-zone existence current density) that marks the onset of an electric field of $1 \mu$V/cm (left scale). The solid line is a fit of the data with a function of the form $j_m=C \sqrt{1-(T/T_c)}$ that gives $T_c=8.48 \pm 0.05$ K. Black open circles: resistivity of the same crystal as a function of temperature (right scale). Note that $T_c^{on}= 8.48 \pm 0.02$ K. []{data-label="fig:critcurr"}](critcurr.eps){width="0.9\columnwidth"}
Figure \[fig:critcurr\]a shows the temperature dependence of the $I$-$V$ characteristics of the B/2 crystal. The measurements were performed by applying short current pulses and measuring the voltage drop across the sample in the same four-probe configuration used for the resistivity measurements. To optimize the thermal exchange between the sample and its environment, allow a proper heat dissipation, and ensure that every single point was taken at the same starting temperature conditions, the sample was mounted in He exchange gas and the current pulses were separated by several seconds of idle time. The curves clearly show the current-induced transition from the superconducting to the resistive state, and then to the normal state, where their slopes corresponds to the normal-state conductance of the sample just above $T_c$ (straight line). Figure \[fig:critcurr\]b shows the temperature dependence of the current density (let us call it $j_m$ for reasons that will become clear below), at which the electric field across the sample is 1$\mu$V/cm. In other words, $j_m$ is the current density at which a non-zero resistance appears. Its temperature dependence is [@gurevich1987]: $$j_m(T) = C \sqrt{1-\frac{T}{T_c}}.$$ The fit of the data points with this function (solid line in Fig. \[fig:critcurr\]b) gives $T_c = 8.48$ K, corresponding to $T_c^{on}$, i.e., to the achievement of the normal state in the resistivity curve (see Fig. \[fig:res\]a). This functional form of the temperature dependence of $j_m$, together with the shape of the I-V curves themselves (the very sharp onset of a finite voltage, the ubiquitous ripples in the transition region, the linear behaviour just above $j_m$) clearly indicate that the transition to the resistive state is dominated by self-heating effects and thus occurs at current densities $j_m$ much smaller than the real $j_c$ [@gurevich1987]. As we will show in the following, this is relevant for the analysis of the results of point-contact Andreev-reflection spectroscopy.
Point-contact Andreev reflection spectroscopy (PCARS)
=====================================================
Point-contact spectroscopy is a simple but very effective technique for the measurement of the amplitude and symmetry of the energy gap in superconductors. It consists in measuring the differential conductance ($dI/dV$) of a ballistic contact between a normal metal (N) and the superconductor (S) under study, as a function of the voltage bias. If the potential barrier at the N/S interface is sufficiently small, the conduction is dominated by Andreev reflection [@Andreev; @BTK], even though the probability of quasiparticle tunnelling is not zero. The spectrum obtained in this way can be fitted to suitable models for the Andreev reflection at the N/S interface [@BTK; @kashiwaya] to obtain quantitative information on the amplitude of the gap and on its symmetry in the reciprocal space. The condition for ballistic conduction (and thus for energy-resolved information to be extracted from the spectra) is that the radius of the contact (schematized as a circular aperture in the otherwise completely opaque barrier between the N and S banks) is smaller than the electronic mean free path in both banks, i.e. $a \ll \mathrm{min}[\ell, \ell']$. When this condition is satisfied, the resistance of the contact (called Sharvin resistance) is determined only by the geometry, i.e., by the number of the quantum-conductance channels: $$R_N = R_S = \frac{2h}{e^2 a^2 k_{F, min} \tau} \label{eq:Sharvin}$$ where $\tau$ is a function of the Fermi velocities in the two banks [@daghero2010] and is often disregarded for approximate estimations. Note that $R_N$ is the resistance of the ballistic contact when both banks are in the normal state, but also coincides with the resistance of the same contact in the superconducting state ($T<T_c$) measured at sufficiently high bias (i.e., $V > 3\Delta$, with $\Delta$ the amplitude of the superconducting gap) [@daghero2010]. Point contacts on Mo$_3$Al$_2$C crystals were made by using the so-called “soft” PCARS technique, i.e., by stretching a thin gold wire across the crystal (that was not subject to any treatment), so that it touches the crystal in a single point. Sometimes we used a drop of Ag paste to make the contact, which ensures a much higher mechanical and thermal stability of the contact itself [@daghero2010]; in other cases we used the gold wire alone. In each case, Mo$_3$Al$_2$C represents the bank with the smaller $k_F$. By using the free-electron approximation one finds that the Sharvin resistance reads: $$R_S = \frac{4 \rho \ell}{3 \pi a^2}$$ where both $\rho$ and $\ell$ refer to Mo$_3$Al$_2$C. The mean free path has been reported to be $\ell = 3$ nm [@Bauer2010] and the low-temperature resistivity is $\rho \simeq 125 \mu\Omega$cm, so that the condition $a < \ell$ means that the resistance of the contact should be larger than 180 $\Omega$ to fulfill the ballistic conditions. Obtaining such a high value of contact resistance is not straightforward because the surface of the crystals is extremely clean.
![(a) Normalized conductance curves of a Au/Mo$_3$Al$_2$C contact (on the crystal B/3) with $R_N =250 \, \Omega$, recorded at different temperatures (symbols) with the relevant fit with a 2D BTK model [@kashiwaya] for a $s$-wave order parameter. (b) Temperature dependence of the energy gap extracted from the fit (red circles) and of the other parameters, $Z$ (blue squares) and $\Gamma$ (black open circles). Note that neither $\Gamma$ or $Z$ depend on temperature, as expected for a perfectly ballistic contact.[]{data-label="fig:goodconductance"}](goodconductance.eps){width="0.9\columnwidth"}
Figure \[fig:goodconductance\]a shows some conductance curves, recorded at different temperatures (starting from 2.6 K, i.e., below $T_c/3$) of a Au/Mo$_3$Al$_2$C contact with $R_N = 250\, \Omega$ in the crystal B/3 that has a critical temperature $T_c^{M,on} = 8.75$ K determined from magnetization. The enhancement of conductance at low bias is due to Andreev reflection; the shape of the curves is typical of an $s$-wave superconductor, without any evidence of other components of the order parameter such as, for instance, a $p$-wave one. This is fully consistent with previous findings from magnetic penetration depth measurements [@Bonalde2011] and muon spin rotation and relaxation studies [@Bauer2014], that showed the absence of time-reversal symmetry breaking in Mo$_3$Al$_2$C. The two symmetric maxima occur roughly at $V= \pm \Delta/e$, but a fit with a suitable model was required to gather more accurate quantitative information. To this aim, the curves were normalized, i.e., divided by $R_N$, and vertically shifted for clarity; the solid lines in the figure represent best fits using the Tanaka-Kashiwaya version [@kashiwaya] of the BTK model [@BTK] for Andreev reflection at a N-S interface. The agreement, especially in the low-bias region, where the noise is smaller, is very good. The model we used contains only three parameters: the gap amplitude $\Delta$, a barrier parameter $Z$, and a smearing parameter $\Gamma$ that accounts for the intrinsic finite lifetime of quasiparticles and for extrinsic (experimental) sources of broadening [@daghero2010]. Note that only $\Delta$ was varied as a function of temperature to adjust the theoretical curves to the experimental spectra, while both $Z$ and $\Gamma$ were determined from the fit of the lowest-temperature curve and then kept constant on increasing the temperature.
The temperature dependence of the gap extracted from the curves is shown in Fig. \[fig:goodconductance\]b. Unfortunately, the contact broke down (because of the different thermal expansion coefficients of the various materials) at 6.6 K. However, the data points could still be fitted to a BCS-like curve: $$\Delta(T) = \Delta(0) \tanh \left(1.74 \sqrt{\frac{T_c}{T}-1}\right)$$ giving $\Delta(0) = 1.728 \pm 0.008$ meV and $T_c=8.86 \pm 0.07$. The corresponding gap ratio $2 \Delta(0) /k_B T_c = 4.53$ lies midway between those obtained from specific-heat measurements (4.028 [@Karki2010] and 4.06 [@Bauer2014]) and that given by penetration depth measurements (5.18 [@Bauer2014]).
In contacts with resistance smaller than 180 $\Omega$, the spectra systematically display anomalous “dips" at intermediate bias values, whose position strongly depends on the resistance of the contact. As an example, Fig. \[fig:conductances\] reports the conductance curves of the *same* Mo$_3$Al$_2$C/Ag paste contact on the crystal F/3 ($T_c^{M,on}=9.3$ K determined from magnetization) for different values of its resistance, progressively reduced by applying short voltage pulses as described elsewhere [@daghero2010]. It is clear that, on decreasing $R_N$, the dips become deeper and deeper and move towards lower voltages. This behavior excludes that such dips have any relation with the superconducting energy gap or with any other intrinsic energy scale of the material under study.
There are two possible mechanisms that can give rise to these dips. The first, in principle, can occur even for ballistic contacts, and reflects simply the fact that, on increasing the current, the critical current density (or, better, $j_m$, as shown in Fig. \[fig:critcurr\]) is reached and the material turns to the normal state. This picture, nevertheless, does not apply to our case since, in purely ballistic contacts (see Fig.\[fig:goodconductance\]), such dips are missing.
The other mechanism for the formation of the dips relates them to the breakdown of the conditions of ballistic conduction through the contact [@Sheet2004; @Daghero2006]. A rough estimate of the contact radius for the curves of Fig. \[fig:conductances\], based on the Sharvin equation, indicates that $a$ increases from $\simeq 7.5$ nm when $R_N = 28.1\, \Omega$ to $\simeq 19$ nm when $R_N = 4.3 \, \Omega$. Since $\ell \simeq 3$ nm, none of these contacts is actually in the Sharvin regime. The contact resistance can thus be approximately expressed [@Wexler1966] as $R_N = R_S + R_M$ where the Maxwell term $R_M$ is equal to $(\rho + \rho')/2a$. In the superconducting state at low $T$, $\rho'=0$ and the Maxwell term only contains the (very small) resistivity of gold. Since $R_S \propto a^{-2}$ and $R_M \propto a^{-1}$, their relative contribution to the total contact resistance depends on $a$. Unless $a$ is very small, $R_M$ is not negligible and this has many consequences.
First, when the critical conditions of temperature and current density are reached, $\rho'$ starts playing a role in the Maxwell term. Since $\rho'$ is again proportional to the derivative of the curves in Fig. \[fig:critcurr\], the dips arise [@Sheet2004]. Clearly, the smaller the contact resistance, the smaller the voltage at which this happens, as shown in Fig. \[fig:conductances\]. When the dips are too close to $\Delta/e$, the Andreev-reflection features look narrower than in higher-resistance contacts; in the limit case, they collapse into a single featureless maximum at zero bias (top curve in Fig. \[fig:conductances\]). Moreover, the dips move towards lower voltages in curves recorded at increasing temperatures. This is shown for example in Fig. \[fig:Tdep\]a that reports the temperature dependence of the conductance curves of a point contact on the crystals B/2, with a resistance $R_N \simeq 24.6\, \Omega$.
Second, once $\rho' \neq 0$, Joule heating in the contact starts to be substantial and rapidly makes the temperature in the contact increase according to the law $T^2= T_0^2 + V^2/4L$, $L$ being the Lorentz number and $T_0$ the bath temperature. Therefore, only the part of the spectrum at $|V|<V_{dip}$ is actually recorded at the bath temperature. Third, while the resistance of the contact at low bias is practically $R_S$ (the term $\rho/4a$ being very small) it becomes equal to $R_S+R_M$ at high bias. Since $R_M/R_S$ increases on increasing $a$, in low-resistance contacts the low-bias part appears enhanced with respect to the tails and when such curves are normalized their amplitude can sometimes exceed the ideal value of 2. The two top curves in Fig.\[fig:conductances\] are close to this condition.
![Two examples of temperature dependencies of PCARS spectra. (a) Au/Mo$_3$Al$_2$C contact on crystal B/2, having $R_N=24.6 \, \Omega$ and a large Maxwell contribution so that the Andreev features are almost completely eroded by the dips, and the zero-bias feature is unnaturally enhanced with respect to the high-bias tails. (b) Ag-paste Mo$_3$Al$_2$C contact on crystal F/3, having $R_N=76.5 \, \Omega$, and a smaller Maxwell contribution to the resistance. The dips here fall sufficiently far from the edges of the Andreev-reflection feature, at least at low temperature. In both (a) and (b) all the curves apart from the top one are vertically offset for clarity. The temperature corresponding to the achievement of the normal state is indicated in bold.[]{data-label="fig:Tdep"}](Tdep.eps){width="0.9\columnwidth"}
The conclusion of this digression is that a reliable determination of the gap amplitudes requires ballistic contacts, but even the spectra of diffusive contacts can provide a reasonably good estimation of the gap amplitude, provided that the dips fall sufficiently far from $\Delta/e$. For example, the curves shown in Fig. \[fig:Tdep\]b are at the limit. At low temperature the dip falls at a voltage high enough to ensure that the central part of the curve (i.e. up to and slightly above the conductance maxima) is still reliable, and its fit can provide a reasonable estimation of the gap.
![(a) The PCARS spectra obtained by normalizing the differential conductance curves of Fig.\[fig:Tdep\]b (symbols) together with the relevant fit within the 2D BTK model. The best fit was obtained by minimizing the SSR in range $\pm 2$ mV. (b) Temperature dependence of the best-fitting parameters $\Delta$ (solid circles) $\Gamma$ (open circles) and $Z$ (solid squares).[]{data-label="fig:Tdepfit"}](Tdepfit.eps){width="0.9\columnwidth"}
The fit of the curves reported in Fig. \[fig:Tdep\]b is shown in Fig. \[fig:Tdepfit\]. In panel (a) the PCARS spectra (i.e. the differential conductance curves normalized to their high-voltage resistance) are compared to the best-fitting 2D-BTK curve. The fitting procedure was the following. First we approximately determined up to which voltage the model for an ideal contact was able to account for the observed shape of the curve. In other words, we excluded from the calculation of the sum of squared residuals (SSR) the part of the curve dominated by the dips. It turned out that a safe range was $\pm 2$ meV. This is enough to fit the region of the zero-bias valley and the two symmetric maxima – which is the most important to assess the possible presence of zero-bias excitations associated to gapless regions of the Fermi surface (like, e.g. when a gap has nodes) or more generally the symmetry of the gap in the $k$ space.
Then, we fitted all the curves by minimizing the SSR in the same range at all temperatures. This was made both by using an automatic fitting procedure using the 1D BTK model [@BTK], and by changing the parameter values by hand, within the 2D BTK model [@kashiwaya; @daghero2010]. In both cases, we used an isotropic $s$-wave gap. Within the range of reliability, there is again no detectable discrepancy between the experimental curves and the theoretical spectra, and hence no sign of a non-$s$-wave component. The values of the gap amplitude $\Delta$ and of the broadening parameter $\Gamma$ resulting from the two procedures were identical ($Z$ values were instead different, as expected, because of the different assumptions of the two models).
Fig. \[fig:Tdepfit\]b shows the fitting parameters resulting from the 2D BTK model. Because of the presence of the dips and of their progressive moving to lower voltages on increasing temperature, the values of the gap start to deviate very soon (at about 6 K) from a BCS-like $\Delta(T)$ curve. Therefore, determining the zero-temperature gap amplitude $\Delta_0$ is not easy: Qualitatively, one can say that only the first two points lie on top of a BCS-like curve with $\Delta_0=1.6$ meV and gap ratio 4.1. Note that the values of $\Gamma$ necessary to obtain a good fit decrease on increasing temperature. This is a clear sign that the Andreev features are progressively narrowing, eroded by the dips. In the ideal case $\Gamma$ should instead increase or at most remain approximately constant (see Fig. \[fig:goodconductance\]b).
Similar behaviour was observed in all the contacts showing dips. Fig. \[fig:gaps\] summarizes the gap values obtained from the fit of the low-temperature curves of various contacts with different resistance, in different crystals. The data show that the resistance of the contact has a very clear effect on the value of the fitting parameter $\Delta$. In contacts with low resistance, where the dips interfere with the Andreev structures (as in Fig. \[fig:Tdep\]a) the value of $\Delta$ that results from the fit is not a good measure of the superconducting gap amplitude. Hence, the large variability in the values of $\Delta$ for $R_N < 50 \, \Omega$ simply reflects the fact that $\Delta$ is sensitive to *extrinsic* effects (e.g. the Maxwell term in the contact resistance) that depend on the contact and not only on the material under study. For $R_N > 50 \,\Omega$ the number of points is much smaller, because of the aforementioned difficulty to keep the contact resistance sufficiently high. However, on increasing $R_N$ the values of $\Delta$ follow a clear trend with a tendency to saturate at 1.6-1.7 meV at the highest values of contact resistance. Note that the data points shown in Fig.\[fig:gaps\]a come from different crystals (see the legend) that have slightly different critical temperatures: $8.48 \pm 0.02$ K for B/2, $8.86 \pm 0.07$ K for B/3, and $9.27 \pm 0.02$ K for F/3. Despite these differences, all the values of $\Delta$ fit in a common trend. Figure \[fig:gaps\]b reports the values of the gap ratio $2\Delta /k_BT_c$ as a function of $R_N$. The use of the gap ratio allows a safe comparison of the results obtained in different crystals. A saturating trend very similar to that of $\Delta$ is clearly seen. At high values of $R_N$, where the conditions for ballistic conduction are definitely fulfilled, the gap ratio $2\Delta /k_BT_c$ tends to the value 4.2-4.3.
Conclusions
===========
In summary, we have reported the successful growth of Mo$_3$Al$_2$C single crystals under high-pressure, high-temperature conditions. Such synthesis conditions allow us to overcome the problems related to the low reactivity of C and the greatly different melting temperatures of the elements. Our results suggest that C-deficient stoichiometries can be achieved in Mo$_3$Al$_2$C under high pressure conditions. The current work could provide a helpful guidance for further experimental synthesis of various C-based intermetallics under high pressure.
Single-crystal X-ray refinement confirmed the high structural perfection of the grown crystals. It is remarkable that the Flack parameter values refined for all the measured crystals by using a $P4_132$ space-group model were reproducibly close to either 0 or 1, hence strongly suggesting that the studied crystals correspond to the two enantiomorphic space groups $P4_132$ and $P4_332$.
The superconducting critical temperature of the crystals was measured by means of magnetization and transport measurements. A certain variability in $T_c$ between 8.6 K and 9.3 K was observed, depending on the growth protocol and probably related to the actual C content. The transition turned out to be very narrow, displaying $\delta T_c \simeq 0.3$ K. Some features in the resistivity (i.e., a clear minimum or a change in slope) were observed around 130 K, possibly associated with the opening of a charge density wave gap. Extensive point-contact Andreev reflection spectroscopy studies confirmed that Mo$_3$Al$_2$C is a moderately strong-coupling superconductor with $2\Delta/k_BT_c \simeq 4$. The order parameter has an $s$-wave symmetry, despite the asymmetric spin-orbit coupling arising from the lack of an inversion symmetry.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank P. Moll and B. Batlogg for collaboration on early stage of this study and for fruitful discussions. We also thank J. Hulliger and T. Shiroka for critically reading the manuscript and helpful comments.
[99]{} E. Bauer, G. Hilscher, H. Michor, Ch. Paul, E. W. Scheidt, A. Gribanov, Yu. Seropegin, H. Noel, M. Sigrist, and P. Rogl, Phys. Rev. Lett. **92**, 027003 (2004). V. P. Mineev and M. Sigrist, in *Non-Centrosymmetric Superconductors: Introduction and Overview* edited by E. Bauer and M. Sigrist, Lecture Notes in Physics Volume 847, Springer-Verlag Berlin Heidelberg 2012. G. Amano, S. Akutagawa, T. Muranaka *et al.* J. Phys. Soc. Jpn. **73** 530 (2004). K. Togano, P. Badica, Y. Nakamori, S. Orimo, H. Takeya, and K. Hirata, Phys. Rev. Lett. **93** 247004 (2004). E. Bauer, G. Rogl, X.-Q. Chen, R. T. Khan, H. Michor, G.Hilscher, E. Royanian, K. Kumagai, D. Z. Li, Y. Y. Li, R. Podolsky, and P. Rogl, Phys. Rev. B **82**, 064511 (2010). A.B. Karki, Y. M. Xiong, I. Vekhter, D. Browne, P. W. Adams, D. P. Young, K. R. Thomas, J. Y. Chan, H. Kim, and R. Prozorov, Phys. Rev. B **82**, 064512 (2010). T. Koyama, Y. Ozaki, K. Ueda, T. Mito, T. Kohara, T. Waki, Y. Tabata, C. Michioka, K. Yoshimura, M.-T. Suzuki, and H. Nakamura, Phys. Rev. B **84**, 212501 (2011). T. Koyama, Y. Maeda, T. Yamazaki, K. Ueda, T. Mito, T. Kohara, T. Waki, Y. Tabata, H. Tsunemi, M. Ito, and H. Nakamura, J. Phys. Soc. Jpn. **82**, 073709 (2013). I. Bonalde *et al.*, Phys. Rev. B **84**, 134506 (2011) E. Bauer, C. Sekine, U. Sai, P. Rogl, P. K. Biswas, and A. Amato, Phys. Rev. B **90**, 054522 (2014). N. D. Zhigadlo *et al.,* Phys. Rev. B **86**, 214509 (2012).
J.Karpinski *et al.*, Supercond. Sci. Technol. **16**, 221 (2003); N. D. Zhigadlo, J. Cryst. Growth **402**, 308 (2014); N. D. Zhigadlo, J. Cryst. Growth **382**, 75 (2013); N. D. Zhigadlo, J. Cryst. Growth **480**, 148 (2017); N. D. Zhigadlo, N. Barbero, T. Shiroka, J. All. Comp. **725**, 1027 (2017); N. D. Zhigadlo, J. Cryst. Growth **455**, 94 (2016); N. D. Zhigadlo, J. Crys. Growth **395**, 1 (2014); S. Katrych, Q. F. Gu, Z. Bukowski, N. D. Zhigadlo, G. Krauss, and J. Karpinski, J. Solid State Chem. **182**, 428 (2009); N. D. Zhigadlo *et al.*, Phys. Rev. B **84**, 134526 (2011).
D. Reith, C. Blaas-Schenner, and R. Podloucky, Phys. Rev. B **86**, 104105 (2012). R. T. Gordon, N. D. Zhigadlo, S. Weyeneth, S. Katrych, and R. Prozorov, Phys. Rev. B **87**, 094520 (2013); D. Ernsting, D. Billington, T.E. Millichamp, R.A. Edwards, H. A. Sparkes, N. D. Zhigadlo, S. R. Giblin, J. W. Taylor, J. A. Duffy, and S. B. Dugdale, Sci. Rep. **7**, 10148 (2017). L. Palatinus, G. Chapuis, J. Appl. Cryst. **40**, 786-790 (2007). L. Palatinus, S. J. Prathapa, and S. van Smaalen, J. Appl. Cryst. **45**, 575-580 (2012). G. M. Sheldrick, Acta Cryst. A**64**, 112-122 (2008). H. D. Flack *et al.* Acta Cryst. A**39**, 876-881 (1983). See Supplement Material at <http://link.aps.org/supplemental/PhysRevB> for the structural details in the crystallographic information files (CIFs) for P$4_132$ (suppl. 1) and P$4_332$ (suppl. 2) space groups. C. N. Kuo, H. F. Liu, and C. S. Lue, Phys. Rev. B **85**, 052501 (2012). A. V. Gurevich and R. G. Mints, Rev. Mod. Phys. **59**, 941 (1987). A. Andreev, Zh. Eksp. Teor. Fiz. **46**, 1823 (1964); A. Andreev, Sov. Phys. JETP **19**, 1228 (1974) (Engl. Transl.); D. Saint-James, J. Phys. **25**, 899 (1964). G.E. Blonder, M. Tinkham, and T. M. Klapwijk, Phys. Rev. B. **25**, 4515 (1982). S. Kashiwaya, Y. Tanaka, M. Koyanagi, and K. Kajimura, Phys. Rev. B **53**, 2667 (1996). D. Daghero and R.S. Gonnelli, Supercond. Sci. Tecnol. **23**, 043001 (2010). G. Sheet, S. Mukhopadhyay, and P. Raychaudhuri, Phys. Rev. B **69** 134507 (2004). D. Daghero, A. Calzolari, G. A. Ummarino, M. Tortello, R. S. Gonnelli, V. A. Stepanov, C. Tarantini, P. Manfrinetti, and E. Lehmann, Phys. Rev. B **74**, 174519 (2006). G. Wexler, Proc. Phys. Soc. **89**, 927 (1966).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The paper concerns the so-called aligned events observed in cosmic rays. The phenomenon of the alignment of the most energetic subcores of gamma-ray–hadron ($\gamma-h$) families (particles of the highest energies in the central EAS core) was firstly found in the “Pamir” emulsion chamber experiment and related to a coplanar particle production at $E_0>10^{16}$ eV. Here a separation distribution (distances between pairs of muons) for aligned events has been analyzed throughout muon groups measured by Baksan Underground Scintillation Telescope (BUST) for threshold energies $0.85 \div 3.2$ TeV during a period of 7.7 years. Only muon groups of multiplicity $m\geq 4$ with inclined trajectories for an interval of zenith angles $50^\circ - 60^\circ$ were selected for the analysis. The analysis has revealed that the distribution complies with the exponential law. Meanwhile the distributions become steeper with the increase of threshold energy. There has been no difference between the lateral distribution of all the groups and the distribution of the aligned groups.'
---
[**LATERAL DISTRIBUTION FOR ALIGNED EVENTS IN MUON GROUPS DEEP UNDERGROUND**]{}[Talk given at the ISVHECRI’2006, Weihai, China]{}\
\
\
\
INTRODUCTION {#introduc}
============
The most essential problem of cosmic ray physics is the study of the energy spectrum of primary cosmic radiation (PCR), their chemical composition and interactionS of PCR particles with atmospheric nuclei at superhigh energies, inaccessible for modern accelerators. Muon groups in underground experiments, being one of the components of extensive air showers (EAS), convey information about characteristics of cosmic rays. Alongside with these data on muon groups embrace the interval of primary energies located between direct and indirect methods of measuring: from $10^{13}$ to $10^{17}$ eV, i.e., they cover the region of the presumed knee of the primary cosmic-ray spectrum.
Besides, an interesting cosmic-ray phenomenon related to the so-called alignment of clusters of close particles or relatively isolated particles is also observed in the knee range. This phenomenon has been first found in experiments on study of gamma-ray–hadron ($\gamma-h$) families (groups of the most energetic particles in the EAS cores) with the use of X-ray emulsion chambers (XREC) in events with an observed energy $\Sigma E_\gamma > 700$ TeV [@Pamir4_1; @Kopenetal]. The alignment is realized as the tendency of the location of particles near a straight line on the target diagram that corresponds in the three-dimensional case to the absence of the azimuthal symmetry with respect to the EAS axis and concentration of particles’ momenta near some plane. The explanation of the whole block of data related to the alignment phenomenon by fluctuations of cascade development in the atmosphere seems to be improbable $(\ll 10^{-20})$ [@JHEP05_2005]. The influence of the Earth’s magnetic field and thunderstorm-cloud electric fields is also negligible [@8ISVHECRIMukh]. It was assumed that the cause of the alignment is trivial diffraction kinematics effects [@Smirnova; @Zhu; @capdev]; generation of quark-gluon strings [@Halzen]; appearance of a new strong interaction and generation of new super-heavy quarks of higher color symmetry [@White]; semihard double diffraction dissociation with a tension of a quark-gluon string between a semihardly scattered constituent quark and spectator quarks of the projectile hadron and its following rupture with creation of secondary particles [@Royzen].
EXPERIMENT
==========
{height="14.0pc"}
\[plot1\]
{height="14.0pc"}
\[plot2\]
{height="22pc"}
\[fig2\]
There are very few experimental works on aligned events study in muon groups deep underground: MACRO [@sioli] and BUST [@npb06v5p473; @29ICRCv9p243]. A search for aligned events has been done using the alignment parameter [@Pamir4_1]:\
$\lambda_m ={\sum\limits_{i\ne j \ne k}^N cos 2\varphi^k_{ij}}/[
{m(m-1)(m-2)}], $ where $\varphi^k_{ij}$ is the angle between the lines connecting the $k$-th particle with the $i$-th and $j$-th particles, $m$ is the number of particles $(m \geq 4)$. Note that $\lambda_m$ decreases from $\lambda_m=1$ for completely aligned events to $\lambda_m \simeq -1/[m(m-1)(m-2)]$ in the case of isotropic events. In [@sioli] the distribution of events as function of parameter $\lambda_m$ is peaked at $\lambda_m =1$ due to the asymmetry of MACRO detector geometry $(12\times 76$ m$^2$). Two variants of simulation (CORSIKA/QGSJET and HEMAS/DPMJET) were made. The experimental data not contradict the simulation. Results of our analysis of BUST experimental data are shown in Figs. \[plot1\] and \[plot2\] for two muon threshold energies $E_{thr}^\mu$ (0.85 and 3.2 TeV). One can see that there is no a significant difference in the lateral distribution for events selected by $\lambda$ parameter. Note that visible change of slope in \[plot2\] for both types of events can be explained by the admixture of $\sim 10$% wrong events corresponding to lower energy threshold due to errors in angular measurements and very high gradient of the mountain profile.
CONCLUSION
==========
It was shown in our previous works [@npb06v5p473; @29ICRCv9p243] that aligned muon events observed in our experiment are pure statistical ones. In this work we shown that there are no statistically significant differences in lateral distributions of “aligned” and all events. The latter confirms once again that our “aligned” muon groups are just a statistical tail of a big number of events recorded by BUST.
Fig. \[fig2\] shows the absence of a significant difference in lateral distributions of muon groups simulated with a QGSJET-like MC0 model and a coplanar particle generation model (CPGM) [@JHEP05_2005]. A key for explanation of this result can be found if one takes into account that a fraction of muons produced by decaying high-energy mesons generated in the first interaction is negligible as compared to muons produced in following generations deeper in the atmosphere.
This work is supported by Russian Foundation for Basic Research (projects 04-02-17083, 05-02-17599, 05-02-16781, 06-02-16606, 06-02-16969), Russian Ministry of Education and Science (projects SS-5573.2006.2 and SS-4580.2006.02) and Russian Federal Agency of science and innovations (state contract 02.445.11.7070).
[9]{} Pamir Collaboration, Proc. 4th ISVHECRI, Beijing (1986) 429.
V.V. Kopenkin [*et al*]{}., Phys.Rev.D 52 (1995) No. 5, 2766.
R.A. Mukhamedshin, J. High Energy Phys. 05 (2005) 049.
R.A. Mukhamedshin, Proc. 8th ISVHECRI, Tokyo (1994) 57.
M.D. Smirnova and Yu.A. Smorodin, Proc. 21st ICRC, Adelaide, 8 (1990) 310.
Q.-Q. Zhu, Y.D. He, and A.X. Huo, Proc. 21st ICRC, Adelaide, 8 (1990) 306.
J.N. Capdevielle [*et al*]{}., Proc. 27th ICRC, Hamburg, 1 (2001) 1410.
F. Halzen and D.A. Morris, Phys.Rev.D 42 (1990) 1435.
A.R. White, Int.J.Mod.Phys.A8 (1993) 4755.
I.I. Royzen, Mod. Phys. Lett. A9 (1994) No. 38 3517.
M. Sioli, PhD thesis. Univ. Degli Studi di Bologna. 1999.
A.L. Tsyabuk [*et al*]{}., Nucl. Phys. B (Proc Suppl.) 151 (2006) 473.
A.L. Tsyabuk [*et al*]{}., Proc. 29th ICRC Pune, 9 (2005) 243
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Xinlun Cheng [^1]'
bibliography:
- 'msRAA20190151R1.bib'
title: 'Search for strong galaxy-galaxy lensing in SDSS-III BOSS'
---
Introduction
============
Strong galaxy-galaxy lensing produces multiple, distorted images of a background galaxy through the gravitational field of the foreground galaxy. It was first suggested by @zwicky for galaxy-clusters, and the first discovery of a galaxy-scale lens was made by @walsh. The development in optics has led to a rapid increase in the number of strong lensing system discovered. As in 2010, the total number of confirmed strong gravitational lensing by galaxies was approximately 200 [@sl_review].
A number of methods have been utilized to discover strong gravitational lensing by galaxies. These include manual or automatic inspection of high-resolution images from optical or radio telescopes [@1996ApJ...467L..73H; @1999AJ....117.2010R; @2003MNRAS.341...13B; @2004ApJ...600L.155F; @2007ApJ...660L..31M; @2008ApJS..176...19F; @2008MNRAS.389.1311J; @2009ApJ...694..924M; @2017MNRAS.468.3757W], employing citizen science [@spacewarp] or machine learning approaches in sky surveys, detection of emission lines in optical spectroscopy with emission from foreground galaxy removed [@1996MNRAS.278..139W; @slacs; @2006AJ....132..999O; @2008AJ....135..512O; @bells; @2017ApJ...851...48S], examination of 2D distribution of redshift in a small region with integral field unit (IFU) or high-resolution spectrograph [@1997ApJ...474L...1Z; @2007NJPh....9..443B; @2012ApJ...758L..17B], and detection of variability in time-domain [@2005ApJ...626..649P]. Compared to other methods, spectroscopic methods have the advantage of being able to obtain the spectroscopic redshift of both the lens and the source galaxies and the velocity dispersion of the lens galaxy, but lensing system discovered by these methods often require imaging confirmations from optical telescopes.
Many physical implementations could be found in strong lensing by galaxies. Strong gravitational lensing is an indispensable tool in the measurement of the projected mass distributions of galaxies, including both dark and luminous matter [@1937ApJ....86..217Z; @2004ApJ...611..739T; @2014MNRAS.439.2494O; @2018MNRAS.480..431L]. By analyzing the fine structures of an Einstein ring, one could study the gravitational perturbation of unseen dark matter substructure, which could reveal the distribution of dark matter halos and distinguish cold dark matter (CDM) model from warm dark matter (WDM) model [@2001ApJ...561...46K; @2002ApJ...572...25D; @2013MNRAS.435L..53P]. If a point source with variable flux under constant $\rm H_0$, such as a quasar, is gravitationally lensed, it can put constraint on cosmological parameters, especially the expansion rate $H(z)$, by the lens time delay measurement . Recent measurements are capable of limiting the Hubble Constant $H_0$ to 3.8% precision [@2017MNRAS.465.4914B]. Since strong lensing always magnifies the source objects, it could be used as cosmic telescope, revealing details of distant galaxies and quasars. It can provide estimation for the mass of super massive black holes for galaxies, which is one of the few methods to estimate the mass of centeral black holes of non-AGN galaxies [@2001MNRAS.323..301M]. Based on the variability of source images, @2008ApJ...673...34P shed light on the spatial structure of the accretion disk of a quasar. @2013ApJ...777L..17V revealed a star-forming galaxy at z=3.42 with the help of a lens galaxy at z=1.53. While supernova is difficult to detect beyond z=1.5, gravitational lensed supernova have been discovered [@2016ApJ...820...50R; @2017Sci...356..291G] and offered an unprecedented view of the high-redshift universe.
The Sloan Digital Sky Survey III (SDSS-III) [@2011AJ....142...72E] initiated the Baryon Oscillation Spectroscopic Survey (BOSS), and with several upgrades to the original equipment, approximately 1.5 million spectra for luminous galaxies with redshift $z<1.0$ were obtained. The upgrades include reducing the fiber diameter to 2 arcseconds, increasing efficiency in grating and CCD [@2013AJ....145...10D]. This provides us with a wonderful opportunity to search for strong galaxy-galaxy lensing systems using spectroscopic methods.
Although fully automatic strong lensing detection using imaging data has been extensively researched , such attempt in spectroscopic detection is rare partially due to the lack of data as training set for an automatic detection program. Recently, several highly-possible catalogs have been composed and confirmed by images from Hubble Space Telescope (HST) [@slacs; @bells]. In this work, we applied a similar method to the entire SDSS-III BOSS sample, fully released in SDSS DR12 [@2015ApJS..219...12A], and produced a catalog of possible strong lensing candidates, in which we confirmed two out of five candidates with CFHT Megacam as a pilot study.
The outline of the paper is as following. In Section \[sec:algo\], we describe our selection and contamination removal algorithm. Results are reported in Section \[sec:res\], and the processing and analysis of CFHT Megacam images are in Section \[sec:cfht\]. We discuss our finding in Section \[sec:dis\]. In Section \[sec:con\], we present our conclusion.
Selection algorithm {#sec:algo}
===================
Algorithm design overview
-------------------------
The kernel of spectroscopic detection is that since both the photons from the lens and the background galaxy are captured in the same fiber, the spectrum from that fiber should be the superposition of two spectra of galaxies with different redshifts. In the BOSS spectroscopic pipe line, which is described in @2012AJ....144..144B, redshift and classification templates for galaxies, quasars, and star classes are constructed by performing a rest-frame principal-component analysis (PCA) of training samples with known redshift. The leading eigenvectors from the PCA are used to define a linear template basis that is used to model the spectra in the redshift analysis. A range of trial galaxy redshifts is explored from redshift -0.01 to 1.00. The combination of redshift and template class that yields the lowest reduced chi-squared is adopted as the pipeline measurement of the redshift and classification of the spectrum. This best fit is then subtracted from the original spectra and yield the reduced spectra for each galaxy, thus removing the foreground emissions from the lens galaxy. However, emission lines from the source galaxy are not removed and should be detected in the reduced spectra.
The selection algorithm can be divided into two separated procedures: primary selection and post-processing. The primary selection is aimed to select objects with features of strong lensing using information from the corresponding fiber and its adjacent fibers, and the post-processing is mainly built to remove false positives produced by various contamination.
Primary selection
-----------------
The basic design of the primary selection is similar to that in [@bells], but with a few tweaks and improvements. The steps are designed as follow:
1. A selection on the type and class of objects is performed to remove QSOs and stars from the sample since we are only interested in galaxy-galaxy lensing.
2. Galaxies whose redshift could not be determined ($\rm zwarning \neq 0$) are removed.
3. Spectra glitches and strong sky emission lines are masked. Residual spectra with unusual high SNR are removed due to possible incomplete template subtraction. A list of all masked lines is provided in .
4. Peak search is performed by maximum-likelihood estimator of a Gaussian fitted to the data [@slacs] with $A=1.2$ and $\sigma=30km/s$. Peaks with $\rm SNR > 6.0$ are accepted.
5. A single Gaussian, as well as double Gaussians with same width are fit at each accepted peak, and their reduced chi-square were compared. The one with smaller value is stored.
6. Peaks with wavelength $\lambda>9200\AA$ are rejected to remove contamination from sky emission lines.
7. Remaining peaks are ranked according to the stored chi-square and only the smallest five ones are kept if there are more. If no doublet remains, the object is labeled as non-lensing system. Otherwise move on to the next step.
8. Emission from adjacent fibers are compared and labeled as non-lensing if contamination is confirmed. See the next paragraph for details.
9. Assuming the detected peak is the \[OII\]$\lambda$3727,3729 doublet, redshift of the source ($z_s$) is inferred and if the difference between that of the lens ($z_l$) is smaller than 0.05, the object is labeled as non-lensing. Otherwise move onto the next step.
10. Supposed positions of other strong emission lines (\[H$\beta$\]$\lambda$4858, \[OIII\]$\lambda$4959 and \[OIII\]$\lambda$5008) are computed. Peak search similar to Step IV is performed to find the exact position of the emission lines. Emission lines are fitted with a single Gaussian and their SNR are calculated and stored for later usage. Only $\rm SNR > 3.0$ is counted as detection. Note that we do not assume a fixed ratio between the flux of \[OIII\]$\lambda$4959 and that of \[OIII\]$\lambda$5008.
11. Residual spectra, including emission line fittings, are plotted for easier inspection and the object is labeled as candidate.
12. Repeat the steps until all plates in BOSS has been checked.
Central wavelength (Å) Maksed Width (Å)
------------------------ ------------------
5577.35 20
6300.31 20
6363.10 20
8344.61 20
9375.98 20
: Masked sky emission lines in Primary selection. These are also the prominent night sky emission lines listed in SDSS DR12 documentation.[]{data-label="tab.sky_sdss"}
In the BOSS spectrometer setup, photons coming from each fiber are stored in different pixels on the CCD. However, a strong emission line from one fiber might affect nearby pixels, causing a spike, which would be mistaken as a peak by the peak detection program, at the same wavelength in the observer frame. Step VIII is designed to minimize the possibility of such false detection. Reduced spectra from the previous and next fiber are extracted and double Gaussians are fit at the same position of the peak in the current fiber. The ratio of flux of emission lines between current one and adjacent ones are computed, and plotted. A number of randomly selected spectra are examined by eye and classified into two categories: contaminated by nearby fibers or not. We decided that a cut at $x + y < 0.6$, where $x$ is the ratio of flux between the current fiber and the previous one and $y$ is that of the next one, is appropriate since we can get rid of $>95\%$ the contamination while still achieving a high completeness.
Post-processing
---------------
Post-processing is traditionally done by human, but due to the size of our data set, it is necessary to automate the process. Analyzing the candidates that pass through the primary selection steps described above, we found out that there are mainly three sources of false positives: sky emission lines and bad foreground template subtraction, low-z objects, and faint objects at given redshift bin. Here are the steps of our post-processing.
1. Remove candidates with detected $[\rm OII]$ doublets falling within over dense regions in observer frame or foreground rest frame. See \[alg:sky\] for detail.
2. Remove candidates with redshift of lens galaxy $z_l<0.1$ and I-band magnitude $\rm I > 22$. See \[alg:mag\] for detail.
3. The sum of the SNR of \[OII\]$\lambda$3727,3729 doublet, \[H$\beta$\]$\lambda$4858, \[OIII\]$\lambda$4959 and \[OIII\]$\lambda$5008 is calculated and saved as a total SN ratio. Remaining candidates are ranked in decreasing order.
4. Candidates with large relative error ($>20\%$) of velocity dispersion are removed from the list. Candidates with $\sigma_v = 0$ or $850\rm\ km\ s^{-1}$ are also removed, since these two values indicate problematic fitting of velocity dispersion in the pipeline of BOSS.
5. Candidates with estimated Einstein radius $R_E<0.5''$ are removed due to high ($>90\%$) false positive rate reported by @2008ApJ...682..964B
6. Remove all previously confirmed lensing system from the list (see \[sec:res\] for more discussions), and save the first 100 candidates into a highly possible candidate list. Visual inspections of template subtraction are preformed on them to ensure that no bad template fitting occurred in BOSS spectroscopic pipe line.
### Sky emission and residual emission from foreground {#alg:sky}
Sky emission lines and residual emission lines from foreground galaxies would cause false detection due to their strong intensity. Although it is hard to remove them with information from a few fibers, they will appear as overdense regions when looking at the histogram of the wavelength of the OII doublets of these candidates in observer rest frame (ORF, ) and lens galaxy rest frame (LRF, ).
In the observer rest frame, the median number of candidates per bin $\rm med=29$ (solid line) with standard deviation $\sigma=33.5$. Therefore, we remove candidates whose wavelength of OII doublet lies in bins with number of candidates $>\rm med+2\sigma=96$ (dash line). The corresponding wavelength range is listed in .
$\lambda_{\rm start}$(Å) $\lambda_{\rm End}$(Å) Reason
-------------------------- ------------------------ -------------------
7080 7120 Over dense in ORF
7700 7720 Over dense in ORF
8000 8020 Over dense in ORF
8760 8780 Over dense in ORF
8880 8980 Over dense in ORF
6600(1+$z_l$) 6740(1+$z_l$) Over dense in LRF
: Contamination removal wavelength range in Post-processing. For candidates with OII doublet wavelength falling into any range listed in this table, they are considered false positives and removed from the candidate list. ORF stands for observer rest frame, and LRF stands for lens galaxy rest frame.[]{data-label="tab:sky"}
In the lens galaxy rest frame, contamination removal is performed by comparing over dense regions with a list of galaxy emission lines. We noticed an extreme dense region between 6660Åand 6740Å. We believe that two reasons might lead to this phenomenon. First, emission lines \[SII\]$\lambda$6716 and \[SII\]$\lambda$6730 are within the region and are likely to be falsely detected as source galaxy OII emission line. Second, some galaxies have really strong and wide \[H$\alpha$\]$\lambda$6562 emission and the BOSS pipe line might have trouble fitting and subtracting it completely, so the leftover emission might be picked up by the Primary Selection program and identified as OII doublet of the source galaxy. The corresponding wavelength range is listed in .
![Histogram of wavelength of detected OII doublet in the observer rest frame. False positives caused by sky emission lines will appear as a higher than normal amount of objects in the bin. Solid line indicates the median number of objects in each bin, and the dash line indicates $2\sigma$ confidence level of number of objects in each bin.[]{data-label="fig:o2_orf"}](msRAA20190151R1fig01.pdf){width="\columnwidth"}
![Histogram of wavelength of detected OII doublet in the lens galaxy rest frame. False positives caused by sky emission lines will appear as a higher than normal amount of objects in the bin. Notice a huge rise between 6600Åand 6740Å. The possible cause is explained in \[alg:sky\][]{data-label="fig:o2_lrf"}](msRAA20190151R1fig02.pdf){width="\columnwidth"}
### Low redshift objects and Faint objects {#alg:mag}
A number of objects targeted by BOSS have redshift $z<0.1$, which are unlikely to become strong lensing systems. However, these objects tend to be bright and since the template could not cover every small emission lines, they take up a large proportion of candidates. Therefore, all candidates with redshift $z<0.1$ are removed in the last step. This could be confirmed by plotting redshift $z$ versus selection ratio (). The rise in the selection ratio at $z>0.8$ is due to the limited number of objects in high redshift. A similar cut in redshift can also be found in both SLACS [@slacs] and BELLS [@bells].
![Redshift of lens galaxy ($z_l$) versus selection ratio. Unusually high selection ratio could be observed at $z<0.1$. Therefore, objects with lens galaxy redshift less than 0.1 are much more likely to be false positives and are removed from the candidate list.[]{data-label="fig:z"}](msRAA20190151R1fig03.pdf){width="\columnwidth"}
![Histogram of I-band magnitude after extinction correction. Faint objects are usually not massive and thus have a lower possibility of gravitational lens the source galaxy.[]{data-label="fig:mag"}](msRAA20190151R1fig04.pdf){width="\columnwidth"}
Faint objects are usually believed to be less massive when compared to bright objects with similar redshift. Therefore, it is less likely for faint foreground objects to act as lens. From , we empirically removed candidates with I-band magnitude after extinction correction $\rm magI - 1.698 ebv > 22$, where $ebv$ is the extinction correction. This threshold is chosen due to two reasons. First, there is a sharp drop in the number of candidates, and secondly, as presented by SLACS and BELLS, no candidate with I-band magnitude fainter than 22 is selected.
Selected Candidates {#sec:res}
===================
A total number of 2388 plates are in SDSS BOSS, and in each plate there are 1000 fibers. Target selection by object type and class returns 1500002 galaxies. @bells have already checked several hundred plates, so these plates are excluded from counting the number of candidates. Thus we would only run the selection algorithm on 1380002 galaxies. After running through the steps in Section \[sec:algo\], the number of candidates is 396. The detailed numbers of candidates remaining and left after each step are presented in .
Key step Candidates rejected Candidates left
------------------------------------------------------------ --------------------- -----------------
Unusually high spectra SNR 12806 1367196
Peak search 1250086 117110
Singlet fit better 72833 44277
$\lambda>9200$ 20711 23566
Adjacent fiber contamination 4114 19452
$z_s - z_l < 0.05$ 9627 9825
Sky and bad subtraction 2495 7330
Low redshift 931 6399
Faint objects 116 6283
Other emission lines 4753 1530
Inaccurate velocity dispersion measurement and $R_E<0.5''$ 1134 396
: Number of candidates remaining and rejected after each key step described in Section \[sec:algo\][]{data-label="tab.boss"}
The distribution of the position on the sky, lens galaxy redshift ($z_l$) and source galaxy redshift ($z_s$) of candidates are shown in , and , respectively. We can conclude that there is no significant bias across the sky. Cutouts from existing surveys (DECals, CFHTLS, and HSC) have also been acquired and examined, but none is capable of reaching a definitive conclusion that whether they are strong lensing systems or not.
![The distribution of candidates on the sky. We can conclude that the distribution is uniform, with no significant bias across the sky.[]{data-label="fig:ra_dec"}](msRAA20190151R1fig05.pdf){width="\columnwidth"}
![The distribution of lens galaxy redshift of the candidates.[]{data-label="fig:zl"}](msRAA20190151R1fig06.pdf){width="\columnwidth"}
![The distribution of source galaxy redshift of the candidates.[]{data-label="fig:zs"}](msRAA20190151R1fig07.pdf){width="\columnwidth"}
Here we compiled a list of highly possible candidates catalog in Appendix A.
Follow-up Observation with CFHT Megacam {#sec:cfht}
=======================================
All candidates have been checked to see if there is any existing HST coverage. However, none of them has available HST images. The final confirmation of strong lensing system requires imaging data from telescopes. Therefore, we applied CFHT Megacam telescope observation time through Telescope Access Program (TAP). Canada-France-Hawaii Telescope (CFHT) is a 3.6m telescope located at Mauna Kea, with average seeing of $\sim 0.6 ''$. Megacam is the wide-field optical imaging facility at CFHT, with 36 CCDs covering 1x1 square degree field-of-view.
Observation details
-------------------
Due to limited observation time, we only chose the 5 candidates that are visible in September 2018. Since the average seeing of CFHT is $\sim 0.6 ''$, all chosen targets have estimated Einstein radius greater than $0.6''$, with three targets close to $1.0''$ and two targets close to $1.5''$. Since no existing high resolution imaging data is available for these targets for exposure time calculation, the brightness of the distorted source images is estimated with data provided in [@bells], given both our and their targets are selected with SDSS-III BOSS spectra with similar methods. We found that exposure time of 150 seconds is enough to achieve $\text{SNR}=10$. 7 images are taken for each target each band, resulting in a total exposure time of $160\text{ sec}\times7=1120\text{ sec}$ per band per target. Since the seeing might not be good enough to clearly separate the arcs and rings from the lens galaxy, we believe that color images might increase our chance of confirmation since there is often color difference between the lens galaxy and the source galaxy (usually bluer). Therefore, we requested images from three bands: $g$, $r$, and $i$. Their spectra, detected emission lines and CFHT false color (gri) images are shown in – . Detailed information about these 5 systems are shown in .
Name Plate - MJD - FiberID $z_{l}$ $z_{s}$ $R_E (arcsec)$ $\sigma_v$ (km/s)
---------------------------- ----------------------- --------- --------- ---------------- -------------------
SDSS J004617.02+252041.5 6286 - 56301 - 265 0.428 0.829 0.841 253.54
SDSS J012832.31$-$023500.0 4352 - 55533 - 120 0.155 0.468 1.093 245.95
SDSS J012903.47+035901.7 4310 - 55508 - 314 0.196 0.915 1.373 248.99
SDSS J015102.97+042013.1 4270 - 55511 - 571 0.434 0.938 0.942 271.45
SDSS J231904.19+193242.4 6123 - 56217 - 216 0.297 0.848 1.408 281.59
: Information about 5 targets proposed for CFHT Megacam observation. The first column is the name of the targets, the second column is the plate, modified Julian Date of spectra and the fiber id. The third column, $z_l$, is the redshift of lens galaxy, and the forth column, $z_s$, is the estimated source galaxy redshift. The fifth column, $R_E$, is the calculated Einstein radius, assuming $H_0=70\ \rm km\ s^{-1} Mpc^{-1}$ and $\Omega_m=0.3$. The last column, $\sigma_v$, is the velocity dispersion from BOSS spectra.[]{data-label="tab.cfht"}
![SDSS J004617.02+252041.5. The left panels display false color (gri) images from CFHT Megacam. The white circles represent BOSS fiber, with diameter of 2”. The middle panels are the detected OII doublet in the residual spectra. The black lines indicate the flux in the current fiber, and the blue and green ones refer to flux in adjacent fibers. The red lines indicate best fit of the detected OII doublet. The right panels are part of the residual spectra near the inferred wavelength of OIII and H$\beta$ of the source galaxy in observer frame. The wavelength of \[H$\beta$\]$\lambda$4858 and \[OIII\]$\lambda$5007 are pointed out with vertical green dash lines, and red lines indicate best fit. The observer frame wavelength of \[OIII\]$\lambda$4959 is not indicated in the plot, but the emission line was still fitted with a Gaussian.[]{data-label="fig.cfht_01"}](msRAA20190151R1fig08.pdf){width="\textwidth"}
![SDSS J012832.31$-$023500.0. See for description of the layout of the figure.[]{data-label="fig.cfht_02"}](msRAA20190151R1fig09.pdf){width="\textwidth"}
![SDSS J012903.47+035901.7. See for description of the layout of the figure.[]{data-label="fig.cfht_03"}](msRAA20190151R1fig10.pdf){width="\textwidth"}
![SDSS J015102.97+042013.1. See for description of the layout of the figure.[]{data-label="fig.cfht_04"}](msRAA20190151R1fig11.pdf){width="\textwidth"}
![SDSS J231904.19+193242.4. See for description of the layout of the figure.[]{data-label="fig.cfht_05"}](msRAA20190151R1fig12.pdf){width="\textwidth"}
CFHT observation results
------------------------
Images from CFHT Megacam are processed by following the steps described below: Star and galaxy Catalog was first generated with SExtractor on each image and then cross-matched with SDSS DR7 catalog with SCAMP (Software for Calibrating AstroMetry and Photometry, @2006ASPC..351..112B) for astrometric calibration and photometric calibration. Then all images of the same band of the same target were coadded with SWarp [@2002ASPC..281..228B]. Cutouts were made and combined into false color (gri) images for visual inspection, as illustrated from to . Each cut out is 5.58x5.58 $\rm arcsec^2$.
From these images, we can conclude that SDSS J012832.31$-$023500.0, SDSS J012903.47+035901.7 and SDSS J015102.97+042013.1 are not strong lensing systems, and SDSS J004617.02+252041.5 or SDSS J231904.19+193242.4 are strong lensing candidates. In the images of these two targets, we clearly detect one distorted source galaxy image, but their counter images are probably out-shined by the central galaxy. Further investigation will be discussed in the next section.
Discussion {#sec:dis}
==========
As discussed in Section \[sec:cfht\], due to limited seeing, none of the 5 systems have an image with clear separation. However, for 2 of them, SDSS J004617.02+252041.5 and SDSS J231904.19+193242.4, we capture one clearly separated source image, which could reveal more information about the system and provide us with a reasonable estimation of the position of the counter-image. Using GALFIT [@2010AJ....139.2097P], we successfully removed the lens galaxies for these two systems and obtained the position of the visible source image with fitting. With a simple Singular Isothermal Sphere (SIS) model, we estimate the separation of their counter-image to their respective central galaxy, shown in . Each pixel in the images corresponds to $0.187''$ on the sky. Since the distorted source images are only slightly larger than ten pixels, the magnitude estimation by GALFIT has very large errors and are not included.
Object Name J004617.02+232041.5 J231904.19+193242.4
---------------------- --------------------- ---------------------
Lens Center (pixel) 14.304, 16.294 16.571, 17.210
Image Center (pixel) 21.771, 15.512 8.297, 13.117
$R_E$ (arcsec) 0.841 1.408
$R_2$ (arcsec) 0.300 1.089
: Measured and estimated information of the two systems. The second row refers to the pixel position of the center of the lens galaxy, obtained with GALFIT. The third row indicates the pixel position of the center of the visible distorted source image, also obtained with GALFIT. The last row $R_2$ refers to the estimated separation of the undetected counter image from the lens galaxy with the SIS model.[]{data-label="tab.lens_info"}
Given the average seeing of these images are 0.55”, it is safe to conclude that the lens galaxy could easily outshine the counter-image, although with better seeing condition, such as HST or Keck, it would be easy to detect distorted counter-image of the source galaxy.
[@bells] published a catalog of 36 confirmed strong lensing candidates confirmed by HST. We cross checked our results with their list, and we recovered 30 of theirs. The other seven are lost since we are more conservative in our selection algorithm. A number of candidates inside the BELLS catalog have \[OII\]$\lambda$3727, 3729 doublet beyond 9200$\AA$ in the observer frame, which are discarded in our selection algorithm for lowering overall contamination rate. Since the final confirmation of strong lensing system always requires long exposure time from best telescopes, we consider the 80% recovery rate as acceptable.
The ratio of discovered strong galaxy-galaxy lensing system candidates discovered by the algorithm is $0.029\%$, which is considerably lower than the theoretical estimation by @2008ApJ...685...57D who gives a value of $0.5\%$ to $1.3\%$, and the lensing probability in SLACS [@slacs], which is about $0.1\%$. However, this value is comparable to the ratio of BELLS, which is $36/133852=0.027\%$ [@bells]. The reasons of this lowered lensing probability compared to previous searches and theoretical estimation are two-folds. Firstly, the diameter of the fiber is reduced from $3''$ to $2''$, and according to @2008ApJ...685...57D, the selection effect introduced by finite fiber size plays an important role in limiting the maximum number of strong galaxy-galaxy lensing system that could be discovered. Secondly, we only considered those candidates with more than one emission line and with a large enough estimated Einstein radius, which would also remove many of the potential strong galaxy-galaxy lensing systems.
Conclusion {#sec:con}
==========
In this work, we applied similar selection algorithm to the entire SDSS-III BOSS galaxies and were able to automate many steps that were used to be manually done in [@bells]. During this process we also generated a list of 100 highly possible candidates. Follow-up CFHT observations were preformed on five targets, with two of them being potential strong lensing systems. Although only one distorted image can be seen and its counter images probably outshined by central galaxy, we are able to confirm that they are two galaxies detected at different redshift. Assuming these are strong lensing systems, we are capable to calculate the position of the source image closer to the central galaxy and they could not be clearly resolved under the seeing with CFHT. Further follow-up with telescopes with higher resolution power, such as Keck or HST, might be able to clearly reveal the true nature of these systems.
Faced with limited seeing, one can extend the list of confirmation of candidates with Hubble Space Telescope and Keck AO. As indicated in @bells, the complete BOSS galaxy sample is estimated to produce at least a few hundred strong lenses, which, if confirmed, would greatly increase the amount of galaxy-galaxy strong lensing systems known. With a larger number and higher quality imaging data, we plan to apply more complicated lens modeling, such as SIE Lens Models with Sersic source models described in @slacs and @bells, to shed light on the evolution of the relationship between stellar mass, luminosity and galaxy structure.
The author appreciate the anonymous referee for his/her helpful suggestions which improved the paper. The author thanks Drs. Shude Mao, Yiping Shu, Dandan Xu and Mr. Yunchong Wang for helpful discussions and contributions to this project, which is partly supported by the National Key Basic Research and Development Program of China (No. 2018YFA0404501 to SM), by the National Science Foundation of China (Grant No. 11821303 and 11761131004 to SM).
This research uses data obtained through the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories, Chinese Academy of Sciences, and the Special Fund for Astronomy from the Ministry of Finance. Based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Science de l’Univers of the Centre National de la Recherche Scientifique (CNRS) of France, and the University of Hawaii.
This paper made use of SDSS-III data. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
Highly Probable Candidate List
==============================
RA (J2000) Dec (J2000) Plate-MJD-FiberID $z_l$ $z_s$ $\sigma_v$ (km/s) $R_E$ (”) I
------------- ------------- --------------------- ------- ------- ------------------- ----------- -------- -- -- -- --
09:17:26.81 +31:09:25.1 5808 - 56325 - 0434 0.239 0.821 221.83 0.945 19.158
11:34:48.88 +54:11:13.1 6697 - 56419 - 0386 0.302 0.614 276.51 1.036 19.547
15:38:29.13 +41:17:54.3 6050 - 56089 - 0379 0.587 0.931 329.13 0.982 20.857
10:02:48.66 +20:48:52.0 5784 - 56029 - 0104 0.372 0.494 290.96 0.550 20.138
10:09:12.73 +01:31:42.7 4738 - 55650 - 0294 0.479 0.947 235.08 0.690 20.678
17:19:19.83 +27:53:16.4 5003 - 55715 - 0681 0.530 0.952 227.77 0.572 20.563
01:29:03.47 +03:59:01.7 4310 - 55508 - 0314 0.196 0.915 252.72 1.373 18.811
23:47:54.15 +20:03:08.1 6126 - 56269 - 0845 0.411 0.743 212.89 0.523 20.776
11:32:12.17 +06:07:12.7 4767 - 55946 - 0563 0.557 0.856 293.82 0.746 21.129
13:40:58.11 +18:37:52.4 5862 - 56045 - 0110 0.244 0.646 203.35 0.698 19.221
11:45:59.88 +10:22:41.1 5380 - 55980 - 0804 0.385 0.916 241.04 0.874 20.365
12:21:36.90 +16:52:39.3 5849 - 56033 - 0068 0.437 0.975 246.95 0.858 20.082
01:28:32.32 -02:35:00.0 4352 - 55533 - 0120 0.155 0.468 242.40 1.093 18.556
13:45:24.20 +07:05:46.6 4865 - 55713 - 0836 0.229 0.628 209.69 0.761 19.038
01:51:02.97 +04:20:13.1 4270 - 55511 - 0571 0.434 0.938 261.58 0.942 20.918
00:47:51.06 +31:04:32.1 6872 - 56540 - 0443 0.643 0.848 447.56 1.170 20.958
09:03:42.53 +40:32:46.2 4605 - 55971 - 0765 0.437 0.529 390.85 0.685 20.810
23:07:58.78 +33:13:19.0 6504 - 56540 - 0833 0.569 1.037 527.48 3.084 20.631
22:22:19.79 +21:56:58.4 5946 - 56101 - 0884 0.487 0.952 262.30 0.846 20.777
14:12:17.55 +46:56:30.7 6751 - 56368 - 0355 0.402 0.654 223.88 0.501 20.396
14:15:02.80 +54:42:37.5 6797 - 56426 - 0149 0.618 0.786 321.40 0.540 20.560
23:19:04.19 +19:32:42.5 6123 - 56217 - 0216 0.297 0.848 285.39 1.408 19.461
23:35:45.29 -01:15:17.5 4357 - 55829 - 0561 0.486 0.881 236.25 0.631 20.492
13:46:47.26 +04:10:19.5 4785 - 55659 - 0610 0.414 0.624 450.30 1.769 20.168
08:54:26.52 +07:27:01.7 4867 - 55924 - 0865 0.461 0.768 351.37 1.257 21.148
10:27:22.44 +46:06:47.5 6659 - 56607 - 0299 0.521 0.751 298.71 0.684 21.133
08:57:12.83 +19:43:52.1 5175 - 55955 - 0804 0.377 0.628 234.19 0.573 19.821
13:30:57.19 +47:39:54.2 6743 - 56385 - 0086 0.349 0.664 236.65 0.698 19.568
00:19:34.83 +32:45:19.7 7130 - 56568 - 0432 0.114 0.648 213.57 1.054 17.684
02:00:13.95 +03:53:20.1 4270 - 55511 - 0007 0.429 0.754 444.35 2.191 20.983
11:54:49.31 +04:05:41.8 4765 - 55674 - 0115 0.443 0.649 294.56 0.706 20.555
01:46:48.16 +22:00:09.6 5108 - 55888 - 0811 0.267 0.745 214.99 0.798 19.297
: A list of top-100 candidates from the selection algorithm. The first and second column are the Right Ascension (RA) and Declination (Dec) in J2000, respectively. The third column is the plate number (Plate), modified Julian date of the spectra observation (MJD), and fiber id of the target (FiberID). The fourth column, $z_{l}$, is the redshift of the lens galaxy, which is calculated by the BOSS pipeline. The fifth column, $z_{s}$, is the estimated redshift of the source galaxy, assuming the detected peak is the OII doublet. The sixth column is the velocity dispersion of the lens galaxy measured by BOSS pipeline using the width of emission lines. The seventh column, $R_E$ (”), refers to the estimated Einstein radius with SIS model in arc-seconds, assuming $H_0=70\ \rm km\ s^{-1}Mpc^{-1}$ and $\Omega_m=0.3$. The eighth column, $\rm I$, is the i-band magnitude of lens galaxy measured by SDSS photometry with extinction correction applied.[]{data-label="tab:appdx"}
RA (J2000) Dec (J2000) Plate-MJD-FiberID $z_l$ $z_s$ $\sigma_v$ (km/s) $R_E$ (”) I
------------- ------------- --------------------- ------- ------- ------------------- ----------- -------- -- -- -- --
09:33:14.20 +22:05:16.6 5789 - 56246 - 0362 0.350 0.695 216.82 0.615 20.065
14:27:09.14 +25:58:43.5 6015 - 56096 - 0927 0.356 0.756 279.67 1.086 20.907
12:00:51.72 +06:03:13.0 4847 - 55931 - 0040 0.334 0.738 330.80 1.579 19.541
14:09:49.98 +58:22:41.7 6804 - 56447 - 0957 0.360 0.744 285.14 1.100 19.297
07:29:51.26 +32:37:17.3 4444 - 55538 - 0596 0.342 0.839 270.38 1.140 19.394
00:46:17.03 +25:20:41.5 6286 - 56301 - 0265 0.428 0.829 260.19 0.841 20.216
23:53:33.02 -08:38:58.2 7166 - 56602 - 0784 0.167 0.657 209.51 0.904 18.061
00:13:10.41 +11:03:52.9 6113 - 56219 - 0841 0.578 0.811 312.92 0.693 20.873
22:10:54.47 +05:47:40.4 4427 - 56107 - 0478 0.177 0.778 181.41 0.700 18.946
01:00:02.69 +32:05:19.9 6593 - 56270 - 0403 0.487 0.875 241.41 0.651 20.169
13:22:18.78 +64:00:22.6 6822 - 56711 - 0571 0.482 0.786 247.31 0.599 20.625
23:47:36.70 +20:23:01.9 6126 - 56269 - 0834 0.450 0.588 304.53 0.558 19.960
02:05:56.30 -02:20:02.1 4347 - 55830 - 0156 0.283 0.462 256.85 0.687 19.056
13:07:22.99 +33:22:53.0 6488 - 56364 - 0927 0.589 1.022 382.08 1.510 21.068
23:22:25.23 +34:14:58.3 7139 - 56568 - 0687 0.292 0.613 272.90 1.043 19.265
15:50:22.42 +52:49:35.8 6715 - 56449 - 0938 0.442 0.734 241.26 0.594 20.787
00:27:32.70 +09:37:17.4 6195 - 56220 - 0345 0.463 0.864 209.84 0.520 20.571
01:39:28.96 +20:38:23.4 7242 - 56628 - 0051 0.597 0.879 370.07 1.076 21.022
01:25:59.26 +19:36:41.7 5135 - 55862 - 0850 0.101 0.585 226.37 1.191 17.946
12:38:13.00 +23:15:37.6 5983 - 56310 - 0025 0.480 0.852 246.12 0.668 20.341
12:38:46.67 +51:00:56.8 6674 - 56416 - 0262 0.388 0.766 236.44 0.716 19.914
10:51:50.40 +44:01:42.5 4689 - 55656 - 0573 0.435 1.008 325.06 1.533 20.280
12:34:18.66 +57:08:52.9 6832 - 56426 - 0591 0.432 0.799 241.03 0.685 20.253
09:16:28.33 +56:30:45.2 5725 - 56625 - 0595 0.434 0.895 238.81 0.752 20.396
08:50:21.70 +35:47:50.8 4602 - 55644 - 0603 0.475 0.664 446.82 1.445 20.697
23:28:48.60 +01:48:25.3 4284 - 55863 - 0107 0.437 0.855 259.83 0.845 20.846
23:38:28.85 +27:41:25.3 6516 - 56571 - 0543 0.549 0.849 264.13 0.611 20.779
01:57:48.01 -08:51:06.6 7184 - 56629 - 0855 0.501 1.035 227.24 0.667 20.534
09:18:07.92 +50:49:36.2 5730 - 56607 - 0351 0.581 0.913 353.94 1.117 20.230
13:06:40.98 +43:51:21.8 6621 - 56366 - 0263 0.482 0.720 266.59 0.595 20.713
16:39:56.53 +45:34:34.3 6027 - 56103 - 0388 0.422 0.840 324.38 1.347 20.938
13:36:35.91 +22:33:41.8 5998 - 56087 - 0336 0.269 0.834 197.44 0.709 19.288
10:20:52.84 +45:44:32.4 7385 - 56710 - 0198 0.198 0.524 653.74 7.308 20.281
14:43:54.17 +07:02:44.5 4858 - 55686 - 0601 0.514 0.997 209.95 0.533 20.427
15:09:41.98 +34:32:02.3 4717 - 55742 - 0121 0.390 0.769 647.44 5.359 19.902
00:35:34.70 +09:03:59.4 4540 - 55863 - 0069 0.214 0.805 215.77 0.932 19.288
02:52:34.60 -03:31:12.4 7054 - 56575 - 0701 0.412 0.787 231.24 0.657 20.420
23:48:17.74 +11:39:09.0 6150 - 56187 - 0089 0.472 0.775 236.82 0.556 20.354
10:16:37.24 +54:18:16.9 6696 - 56398 - 0487 0.487 0.863 289.27 0.921 20.452
13:17:19.45 +18:40:50.1 5867 - 56034 - 0780 0.291 0.967 226.74 0.957 19.267
00:27:25.59 +00:06:40.5 4220 - 55447 - 0737 0.218 0.535 206.72 0.691 21.118
10:11:55.93 +11:33:14.8 5331 - 55976 - 0910 0.387 0.876 296.83 1.278 20.189
21:58:05.89 +17:56:26.1 5020 - 55852 - 0899 0.628 0.854 338.48 0.735 20.716
: - continued[]{data-label="tab:appdx2"}
RA (J2000) Dec (J2000) Plate-MJD-FiberID $z_l$ $z_s$ $\sigma_v$ (km/s) $R_E$ (”) I
------------- ------------- --------------------- ------- ------- ------------------- ----------- -------- -- -- -- --
13:38:47.40 +05:06:01.8 4786 - 55651 - 0678 0.271 0.560 259.91 0.938 19.113
23:58:12.87 -09:18:25.0 7167 - 56604 - 0642 0.376 0.468 747.56 2.886 19.817
00:26:02.88 +09:44:17.2 6195 - 56220 - 0424 0.246 0.389 266.92 0.714 19.237
14:01:07.02 +41:03:44.0 6630 - 56358 - 0844 0.448 0.870 300.27 1.115 20.661
02:00:58.79 +04:35:46.1 4268 - 55483 - 0644 0.469 0.758 243.72 0.576 20.419
10:31:25.35 +32:42:14.2 6450 - 56331 - 0081 0.610 1.008 240.59 0.554 20.927
12:50:21.74 +19:05:05.4 5856 - 56090 - 0868 0.500 0.822 347.36 1.191 20.418
15:56:18.83 +12:46:51.6 4901 - 55711 - 0573 0.552 0.944 478.26 2.348 20.915
08:29:19.60 +54:11:57.4 5156 - 55925 - 0574 0.361 0.676 235.53 0.678 20.510
08:11:14.08 +22:43:19.7 4469 - 55863 - 0216 0.165 0.918 222.58 1.122 18.488
09:11:34.76 +25:17:08.9 5778 - 56328 - 0079 0.566 0.794 308.63 0.678 20.589
09:44:16.65 +06:27:43.5 4872 - 55944 - 0061 0.338 0.867 232.94 0.872 21.344
23:31:45.76 +29:21:48.7 6581 - 56540 - 0976 0.313 0.679 274.55 1.078 19.867
01:18:01.04 +08:53:17.4 4555 - 56189 - 0210 0.359 0.801 330.19 1.575 19.591
01:15:48.47 +05:09:48.8 4425 - 55864 - 0376 0.337 0.607 217.98 0.558 19.852
07:49:21.61 +27:20:01.0 4459 - 55533 - 0605 0.492 0.682 375.22 0.993 20.660
11:59:15.50 +55:09:07.2 6839 - 56425 - 0117 0.489 1.025 270.56 0.963 20.245
09:55:05.21 +30:32:30.6 5800 - 56279 - 0591 0.473 1.013 225.59 0.685 20.351
08:51:22.62 +11:09:29.2 5291 - 55947 - 0681 0.503 0.838 452.24 2.050 21.057
12:21:43.03 +40:43:42.5 6633 - 56369 - 0164 0.216 0.501 246.50 0.944 18.726
15:47:10.74 +56:52:52.0 6791 - 56429 - 0025 0.441 0.993 247.26 0.867 20.043
11:48:17.87 +26:06:28.9 6411 - 56331 - 0427 0.291 0.743 180.38 0.529 19.664
13:48:40.76 +39:40:45.4 4710 - 55707 - 0601 0.362 0.699 324.13 1.326 20.860
08:02:00.13 +55:17:23.8 5945 - 56213 - 0636 0.461 0.999 273.23 1.020 20.426
12:08:21.58 +29:33:32.1 6472 - 56362 - 0375 0.447 0.876 232.59 0.676 20.647
: - continued
\[lastpage\]
[^1]: Based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Sciences de l’Univers of the Centre National de la Recherche Scientifique of France, and the University of Hawaii. Based on spectroscopic data from the Baryon Oscillation SpectroscopicSurvey of the Sloan Digital Sky Survey III.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A correlation between two noise processes driving the thermally activated particles in a symmetric triple well potential, may cause a symmetry breaking and a difference in relative stability of the two side wells with respect to the middle one. This leads to an asymmetric localization of population and splitting of Kramers’ rate of escape from the middle well, ensuring a preferential distribution of the products in the course of a parallel reaction.'
author:
- 'Pulak Kumar Ghosh$^a$, Bidhan Chandra Bag$^b$ and Deb Shankar Ray$^a$[[^1]]{}'
title: 'Noise correlation-induced splitting of Kramers’ escape rate from a metastable state'
---
Introduction
============
The escape of particles from a metastable well has long been the focal theme of activated rate processes[@kram; @hangi; @ff22; @san; @db11] in chemical kinetics and condensed matter physics. When a stochastic system is thermodynamically made open by the action of external periodic or random forces significant changes in the dynamics take place which reflect the constructive role of noise in dynamical systems. The well-known examples include stochastic resonance[@ben; @luc; @mcn; @d41], resonant activation[@dor; @Bier; @van; @pk1], noise-induced transition[@sch; @hor; @fuli], ratchet and rectification of noise[@ajd; @mag1; @kul; @pkg2; @march12; @pcbcb] etc to mention a few. In the overwhelming majority of these cases the essential physics relies on double-well as a model potential. The present work concentrates on the motion of Brownian particles in the middle well of a symmetric triple well potential which diffuses symmetrically to the left and the right well. At a finite temperature and in absence of any bias force the particles are activated only by inherent thermal fluctuations resulting in equalization of population in the two side wells. Application of an additive noise can not lead to any change of time averaged relative population of the two wells. The situation is expected to remain unaltered even if, in addition, one introduces a multiplicative noise which makes the diffusion state dependent. The aim of the present work is to look for a scheme which leads to a symmetry breaking resulting in a preferential population distribution in one of the two side wells. In what follows we show that the presence of correlation between the applied additive and multiplicative noises may cause a change in relative stability of the side wells with respect to the middle one. This correlation induced interference of the two noises leads to a splitting of Kramers’ rate of escape from the metastable well.
The correlation between noise processes has been the subject of study in a number of issues. For example, it has been shown that correlation strongly influences the noise-induced phase transitions from unimodal to bimodal distribution [@fuli]. Fox[@fox1] has investigated the correlation between multi-component, Markovian and Gaussian stochastic processes. The effect of correlation between quantum noises[@singh] in laser modes and in the description of hydrodynamic modes[@fedc] are of interest in the related context. Our proposal in this work concentrates on altering the relative stability of the two side wells with respect to the middle one of a triple-well potential under the influence of correlation between the noises, where the underlying idea rests on controlling the pathways of a parallel reaction. A prototypical example may be set by considering a nucleophilic attack by $X^{-}$ (a halid ion of $HX$) at the carboxyl group of a ketone say $R_1(R_2)C=0$ which produces $D$ and $L$ forms of $R_1(R_2)C(OH)X$ - the two optical isomers. They are of same energy but differ in optical properties. The middle well represents the reactant state while the two side wells refer to the product states of the parallel reaction. To realize an experimental situation we introduce a light field of fluctuating intensity which polarizes the photosensitive carbonyl group rendering planarity of the polarized states and causing symmetric oscillation of the barrier height. If an electric field of fluctuating intensity and of common origin is now imposed, in addition, on the polarized system, then depending on the cross-correlation of the fluctuations of light and electric fields, the nucleophilic attack of an anion $X^-$ will be asymmetric. This is because the electric field rocks the side wells (similar to what one observes in stochastic resonance) implying the differential relative stability of the transition states of the complex comprising polarized system plus anion. Based on a Langevin and an associated Fokker-Planck description for the dynamics where the additive and multiplicative noise processes are independently Gaussian and $\delta-$correlated in character but the cross-correlation between them is exponential, we derive an analytical expression for the splitting of Kramers’ rate due to correlation induced asymmetry in the state dependent diffusion. This results in preferential distribution of reaction products of the parallel reaction due to differential relative stability of the two wells.
The model
=========
Consider an overdamped Brownian particle in a triple-well potential $V(x)$ kept in a thermal bath at temperature $T$ and subjected to two stochastic forces $\epsilon_1(t)$ and $\epsilon_2(t)$. The governing Langevin equation is given by $$\begin{aligned}
\label{1}
\gamma\dot{x}= -V'(x)\;+\;\epsilon_1(t)+\;x\epsilon_2(t)+\Gamma(t)\end{aligned}$$ where $V(x)=x^2(bx^2-c)^2$ symmetric triple-well potential (Fig.1); $b$ and $c$ are the parameters of the potential and $\gamma$ is the dissipation constant. Thermal fluctuation $\Gamma(t)$ of the bath is modeled by Gaussian, zero mean and delta correlated noise
$$\begin{aligned}
\langle \Gamma(t)\rangle&=&0\\\label{2a}
\langle
\Gamma(t)\Gamma(t')\rangle&=&2 D \delta(t-t')\label{2b}\end{aligned}$$
$D$ being the strength of thermal fluctuation and is given by $D=kT/\gamma$. Here the additive random force $\epsilon_1(t)$ rocks the potential wells sidewise randomly, whereas the multiplicative force $\epsilon_2(t)$ sets random fluctuation of the barrier height around $\Delta V_0( = {4c^3}/{27b})$, in a symmetric manner. The system as described by (\[1\]) is associated with both the thermal and non-thermal environments. To achieve our desired asymmetry in the dynamics of the particle in a symmetric triple-well potential, we apply an electric field and a radiation field simultaneously. The interaction with the radiation field is relatively stronger than that with the applied electric field. The multiplicative and additive noises in Eq.(\[1\]) correspond to fluctuating amplitude of the radiation field and the electric field, respectively. To keep the treatment on a general footing one may assume, $\epsilon_1$ and $\epsilon_2$ to be colored. This may cause a serious difficulty for an analytical approach. On the other hand to capture the essential physics we may assume $\epsilon_1(t)$ and $\epsilon_2(t)$ to be Gaussian white noises. This does not change the inherent feature of the proposed model problem. The characteristics of the noise processes can be summarized as follows
$$\begin{aligned}
\langle \epsilon_1(t)\rangle&=&\langle \epsilon_2(t)\rangle
=0\\\label{3a} \langle \epsilon_1(t)\epsilon_1(t')\rangle&=& 2\;
Q_1 \delta(t-t')\\\label{3b} \langle
\epsilon_2(t)\epsilon_2(t')\rangle&=& 2 \;Q_2
\delta(t-t')\label{3c}\end{aligned}$$
$Q_1$, $Q_2$ are the strength of $\epsilon_1(t)$ and $\epsilon_2(t)$. Now if the simultaneous action of fluctuating electric and radiation fields, is due to a common origin then the statistical properties of the noises are not expected to differ widely and may be correlated [@li; @mad; @jia1]. We characterize the correlation of $\zeta(t)$ and $\eta(t)$ as follows[@xia; @jia; @jia1; @mei; @mei1; @pb] $$\begin{aligned}
\langle \epsilon_1(t)\epsilon_2(t')\rangle &=& \langle
\epsilon_1(t')\epsilon_2(t)\rangle = \frac{\lambda\sqrt{Q_1
Q_2}}{\tau}\;\exp{\left[- \frac{(t-t')}{\tau}\right]}\\\label{3d}
\;\;&=&2\;\lambda\sqrt{Q_1 Q_2}\;\delta(t-t') \;\;\;\;\;\;\;\;
as \; \;\tau \rightarrow 0\nonumber\end{aligned}$$ $\lambda$ is strength of cross-correlation and $\tau$ is the cross-correlation time. By colored correlation between the white noises we mean that both the external fluctuations are affected by each other for certain ranges frequencies. Again if we assume that cross-correlation is a $\delta-$ function, then the gross feature of our model problem will remain unchanged. However for generality, we have assumed colored cross-correlation.
We now proceed with a probabilistic description corresponding to Langevin equation(\[1\]) with the prescriptions(2,3,4) for internal thermal noise and external forces, respectively. Following[@san; @san1; @str; @jai; @xia], the time evolution equation for the probability density is given by $$\begin{aligned}
\frac{\partial P(x,t)}{\partial t}=-\frac{\partial}{\partial
x}V'(x)P(x,t)-\frac{\partial}{\partial x}\langle \epsilon_1(t)
\;\delta(x(t)-x)\rangle &-&\langle x \epsilon_2(t)
\;\delta(x(t)-x)\rangle\nonumber\\
&+& D\frac{\partial ^2 P(x,t)}{\partial
x^2}\label{4}\end{aligned}$$ where $P(x, t)=\langle \delta(x(t)-x)\rangle$; the averages $\langle...\rangle$ in Eq.(\[4\]) can be calculated for Gaussian noise by the Novikov theorem[@nov1]. The resulting equation is the Fokker-Planck description as given by[@jai; @xia]. $$\begin{aligned}
\frac{\partial P(x,t)}{\partial t} =-&&\frac{\partial}{\partial x}
f(x) P(x,t) +Q_2\frac{\partial}{\partial x}x\frac{\partial}{\partial
x}x P(x,t)+\frac{\lambda\sqrt{Q_1 Q_2}}{1+8c^2\tau}
\;\frac{\partial}{\partial x}x\frac{\partial}{\partial x}
P(x,t)\nonumber
\\&&+ \frac{\lambda\sqrt{Q_1 Q_2}}{1+8c^2\tau}
\;\frac{\partial^2}{\partial x^2}x
P(x,t)\;+(Q_1+D)\frac{\partial^2}{\partial x^2}P(x,t) \label{5}\end{aligned}$$ The only constraint on $\tau$ is that, $$\begin{aligned}
1+8c^2\tau > 0\label{6}\end{aligned}$$ $c$ is a potential parameter and a real positive number. Thus practically there exists no restriction on $\tau$ in this case.
Stationary distribution and asymmetric localization
===================================================
We now return to the Fokker-Planck Eq.(\[5\]). We can recast it in the more simpler form as follows $$\begin{aligned}
\frac{\partial P(x,t)}{\partial t} =-\frac{\partial}{\partial
x}\left[ f(x) -\frac{1}{2}\frac{\partial}{\partial
x}D(x,\tau)-D(x,\tau)\frac{\partial }{\partial x}
\right]P(x,t)\label{8}\end{aligned}$$ where $D(x,\tau)$ is the effective diffusion constant $$\begin{aligned}
D(x,\tau) = Q_2 x^2+2\left(\frac{\lambda\sqrt{Q_1 Q_2}}{1+8c^2\tau}
\right) x +Q_1+D \label{9a}\end{aligned}$$ The effective diffusion constant is an asymmetric function of position, which means it is dissimilar in the three wells of the triple well potential. Around the minimum of the middle well diffusion is almost independent of space and has a constant value, while towards the right well diffusion strength quadratically increases with position and decreases towards the left well. So the implications for the simultaneous action of the two forces $\epsilon_1(t)$ and $\epsilon_2(t)$ is that, the diffusive nature of the Brownian motion in the three wells differs and therefore one can expect an asymmetric distribution of the particles in the two side wells and different transition rates from middle well to right and left well.
As $t\rightarrow \infty$, the system reaches the stationary state ($\frac{\partial P}{\partial t}=0$) with the current attaining a constant or zero value. As the process $x(t)$ is bound to the triple-well potential it is expected that in the stationary state there will have no net flow of particles, and hence one may assume a zero current stationary state. The solution of the Eq.(\[8\]) is given by $$\begin{aligned}
P(x)= D(x,\tau)^{-1/2}\exp{\left[\int^x dy
\frac{f(y)}{D(y,\tau)}\right]}\label{10}\end{aligned}$$ It is apparent from the expressions (\[9a\]) and (\[10\]) that as a result of interplay of two stochastic driving forces the distribution function is asymmetric in space. The asymmetry in the distribution function arises due to asymmetric diffusion of the particles. To illustrate the asymmetric localization of the particles we have plotted the distribution function as a function of position in Fig.2. The solid line of Fig.2 presents a symmetric distribution in absence the external driving forces. The dotted line presents the same plot in presence of additive and multiplicative noise but for no cross-correlation ($\lambda = 0$) . In this case the distribution function still remains symmetric but the population of the middle well is much higher than that of the side wells due to the fact that in presence of multiplicative noise the effective diffusion becomes space dependent, $D (x)=Q_2x^2+Q_1+D$. As a result the particles in the side wells diffuse to the middle more quickly as compared to the particles of the middle well which diffuse to the terminal wells at a relatively slower rate and the particles spend most of the time in the middle well. This implies that higher diffusion destabilizes the side wells compared to the middle well. When the two stochastic forces are correlated ($\lambda \neq 0$), interestingly, because of the spatial asymmetry in diffusion the Brownian particles are preferentially localized in the left well compared to the right. This has been presented by the dashed line in Fig.2.
To proceed further we require a quantifier which measures the asymmetry in localization in the two wells. To this end we choose the mean position of the particle as its measure. For a symmetric distribution mean position $\langle x\rangle=0$ and for the localization of the particles in left or right well, the value of mean position is negative or positive, respectively. To this end an expression for $\langle x \rangle$ from a direct steady state solution of Fokker-Planck equation (10) can be formally obtained. However, since this involves a complicated form of space dependent diffusion coefficient, the final expression, which, in principle, contains all the information regarding the interwell transitions, is a lumped expression and it is difficult to figure out the details of transition from one well to another. To have a closer look into this aspect we examine the variation of mean position $\langle x \rangle$ with temperature, $Q_1$, $Q_2$ and correlation time with the help of a discrete three-state model for the continuous triple well potential. Three states are denoted by $x_0$, $\pm x_m$ for the symmetric unperturbed system corresponding to three minima. The diffusional motion causes transitions between them and it is schematically presented by the following kinetic model $$\begin{aligned}
\;\;\;\;k_{L}\;\;\;\;\;\;\;\;\;\;\;&k_{M}^R& \nonumber\\
L\;\;\;\; \rightleftharpoons\;\;\; M \;\;\;&\rightleftharpoons& \;\;\;\;R \nonumber\\
\;\;\;\; k_{M}^L \;\;\;\;\;\;\;\;\;\;\;&k_{R}& \nonumber\end{aligned}$$ $k_L,\;k_R,\; k_M^R ,\; k_M^L $ denote the time averaged rate of transition from left to middle well, right to middle well, middle to right and middle to left well, respectively. The number of particles in the three states at time $t$ are denoted by $n_{L},\; n_R$ and $n_L$. The governing master equations for $n_{i}$ ($i=L,\; R,\; M$) read as $$\begin{aligned}
\frac{dn_L}{dt}&=&-k_L\;n_L+ k_M^L \;n_M\\\label{11}
\frac{dn_R}{dt}&=&-k_R\;n_R+ k_M^R \;n_M\\\label{12}
\frac{dn_M}{dt}&=&k_L\;n_L+k_R\;n_R -( k_M^L +k_M^R )\;n_M
\label{13}\end{aligned}$$ At the steady state ($\dot{n_L}=\dot{n_M}=\dot{n_R}=0$) the probability of finding the particles at the three wells $P_{i}$ ($i=L,\; R,\; M$) are $$\begin{aligned}
P_L= k_M^L k_R/P,\;\;\;\;\;\;\; P_R=k_M^R k_L/P,\;\;\;\;\; \;
P_M=k_L k_R/P \label{14}\end{aligned}$$ where $P=k_R k_M^L +k_R k_L+k_L k_M^R$. The expression for the mean position is then given by $$\begin{aligned}
\label{9}
\langle x \rangle &=&\int_{-\infty}^{+\infty}x P(x)\;dx=x_m
P_R+x_0P_M-x_mP_L\nonumber
\\
&=&\left(\frac{\sqrt{27\Delta V_0}}{2c}\right)\frac{ k_M^R
k_L-k_M^L k_R}{k_R k_M^L +k_R k_L+k_L k_M^R}\label{15}\end{aligned}$$ The above expression clearly expresses the dependence of mean position and probability on four rate constants. All the individual rate constants have a typical dependence on the system parameters, such as, temperature, intensity of the external noises and cross-correlation time. We therefore anticipate a distinct signature of asymmetric localization of the particles with the variation of system parameters. As revealed by Eq.(\[15\]) the mean position is directly proportional to ($k_M^R k_L-k_M^L k_R$), that is, the difference involving the product of right hand directed transition rates and the product of left hand directed transition rates. Our numerical illustration shows that the mean position $\langle x
\rangle \neq 0$ only if $\lambda\neq 0$. This means that the particles are asymmetrically localized in the triple-well potential only when there is a finite correlation between two external stochastic drives $\epsilon_1(t)$ and $\epsilon_2(t)$. This type of behaviour can be physically explained as follows: As both the external forces $\epsilon_1(t)$ and $\epsilon_2(t)$, act independently to rock the potential well randomly in an asymmetric manner and randomly modulate the barrier heights in a symmetric way, respectively, then the symmetry of the triple-well system remains intact as the individual action of $\epsilon_1(t)$ and $\epsilon_2(t)$ are not able to break the symmetry of the system. If there is some correlation between $\epsilon_1(t)$ and $\epsilon_2(t)$, which means that for a number of stochastic realizations, both forces have the same sign at a particular instant of time. A qualitative interpretation of the localization may be given as follows: as long as the force $\epsilon_2(t)$ causing symmetric fluctuation of the barrier height attains its lower value, the force $\epsilon_1(t)$ points to the right well so that the particle in the middle well move towards the right well very quickly. On the other hand as the tilting force points to the left $\epsilon_2(t)$ sets the barrier height at a larger value and consequently the particle in the middle well takes relatively larger time to speed up from middle to left well for the simultaneous action of the synchronized forces. The particles in the middle well therefore have a greater chance to cross the right-hand barrier and the particles in the right well have a greater chance to move to the middle well as the fluctuation of barrier height in the right well occurs with a higher amplitude. Therefore the relative stability of the two wells with respect to the middle differs and an asymmetric localization is observed.
Another important question on the asymmetric localization concerns in which well the particles will be preferentially localized, out of two side wells of the triple-well potential. Secondly, what will be the sign of $\langle x \rangle $. An answer to this question may be obtained as follows. As the diffusive motion of the particles in the right well is greater than that in the left well (as given by the Eq.(\[9\])), the particles will prefer to be localized in the left well, since higher diffusion makes the right well relatively less stable. So the sign of $\langle x \rangle $ will be always negative. The sign of $\langle x \rangle $ and hence asymmetric localization can be inverted by reversing the sign of $\epsilon_1(t)$ in the Eq.(\[1\]).
We now proceed to analyze the behavior of localization (mean position) with the variation of system parameters. The effect of temperature in asymmetric localization and in the product distribution of the parallel reaction at the steady state is intimately related to the manipulation of inherent condition rather than coherence in selecting and controlling the reaction pathways. Keeping in view of the Arrhenius temperature dependence of the individual rate constants, the variation of $\langle x \rangle$ with temperature is expected to show a bell-shaped curve. The departure of $\langle x \rangle$ from zero towards negative direction indicates the preferential distribution of the product in the left well. The variation of mean position $\langle x \rangle$ as a function of temperature for several values of strength of correlation for the input forces as shown in Fig.3 corroborates this assertion. Fig.3 also reveals that for fixed values of temperature and other parameter set the mean position increases with increase of the strength of cross-correlation. It clearly indicates that, for an appropriate correlation between additive and multiplicative noises, the particles will be preferentially localized more asymmetrically. In Fig.4 we show the mean position as a function of intensity of the multiplicative noise. With the increase of values of $Q_2$ the mean position gradually moves to a maximum negative value followed by a return to zero at high intensity ($Q_2$). This sort of behavior can be understood from the expression for the effective diffusion constant(\[9a\]). In this expression $Q_2$ appears as the symmetric contribution in the first term and as the asymmetric contribution (as $\sqrt{Q_2}$) in the second term, so that if one starts from the very low value of $Q_2$, the asymmetry in diffusion first increases then at a relatively high value it starts decreasing. In Fig.5 we present the variation of mean position as a function of intensity of the additive driving force. With increase of $Q_1$ the value of mean position starts departing from zero to reach finally a limiting value. Finally, in Fig.6 we have examined the effect of the cross-correlation time in the asymmetric localization by plotting the mean position as a function of $\tau$ for different values of the strength of cross-correlation. As the asymmetry in diffusion decreases with increase of $\tau$, the departure of the mean position decreases with increasing values of cross-correlation time.
Transition rate from middle well to side wells: Splitting of Kramers’ rate
==========================================================================
To what extent the correlation induced asymmetry in the effective diffusion coefficient is reflected in the kinetics of activated processes. An answer to this question lies in examining the transition rates of the particles from the middle well to the side wells. The differential behavior of the transition rates signifies the scope of controlling the path of a parallel reaction. We approach the problem by calculating the mean escape time. The expression for MPFT [@str] (mean first passage time) for a particle to reach the final point $\pm x_b$, starting from an initial point $x_0$ is given by
$$\begin{aligned}
T_R&=&\int^{+x_b}_{x_0} \frac{dx}{D(x,\tau)P(x)}\int_{-\infty}^x
P(y)dy\label{16a}
\\
T_L &=&\int^{-x_b}_{x_0} \frac{dx}{D(x,\tau)P(x)}\int_{-\infty}^x
P(y)dy\label{16b}\end{aligned}$$
respectively, where $T_R$ and $T_L$ denote the mean escape time of the particle from $x_0$ to $+x_b$ and $-x_b$ respectively ($+x_b$ and $-x_b$ are the coordinates of barrier tops toward right and left well, respectively). Putting the expressions for $P(x)$ and $D(x)$, the integrals in the above Eqs(\[16a\],\[16b\]) have been calculated using steepest-descent approximation to obtain the expressions for $T_i$ ($i=R, \;L$) in the usual way
$$\begin{aligned}
T_R&=&\frac{2\pi}{\sqrt{\omega_{R}
\omega_0}}\exp{\left[-\int_{x_0}^{+x_m}
dx\;\frac{f(x)}{D(x,\tau)}\right]}\label{17a}
\\
T_L&=&\frac{2\pi}{\sqrt{\omega_{L}
\omega_0}}\exp{\left[-\int_{x_0}^{-x_m}
dx\;\frac{f(x)}{D(x,\tau)}\right]}\label{17b}\end{aligned}$$
where $\omega_0,\;\omega_R,\; \omega_L$ are the frequencies corresponding to the potential minimum ($x_0$) and the barrier tops ($\pm x_b$), respectively. As $D(x,\tau)$ is an asymmetric function, the integrals $\int_{x_0}^{+x_b} dx\;\frac{f(x)}{D(x,\tau)}$ and $\int_{x_0}^{-x_b} dx\;\frac{f(x)}{D(x,\tau)}$ are not same. It is thus apparent from the above expression that the transition rate from the middle well to the terminal wells splits up due to the interplay of two correlated stochastic forces. The ratio of the transition rates ($k_M^R/k_M^L$) deviates from unity ($T_L/T_R=k_M^R/k_M^L\neq 1$) only when the two external drives are correlated ($\lambda\neq 0$). If $\lambda=0$, $D(x,\tau)$ is a symmetric function of $x$ and $\int_{x_0}^{+x_b}
dx\;\frac{f(x)}{D(x,\tau)}=\int_{x_0}^{-x_b}
dx\;\frac{f(x)}{D(x,\tau)}$. So the ratio of the transition rates become unity. In Fig.7 we present the variation of the ratio of the transition rates($k_M^R/k_M^L$) as a function of temperature for several values of the coupling strength. As revealed by Fig.7, the ratio of the transition rates at a very low temperature significantly differs from unity and tends to equalize in the high temperature limit. For a fixed temperature the ratio of the transition rates deviates more from unity for an increase of the coupling strength($\lambda$). In Fig.8(a,b) we plot the ratio of the transition rates as a function of the intensities ($Q_1, Q_2$) of the external noises. One observes that with increase of intensities of the external noises the magnitude of $k_M^R/k_M^L$ increases to maximum followed by a decrease. This resonance like behavior is observed due to the fact that with increase of intensities of the external noise the synchronization probability of the noise realizations from two noise processes( $\epsilon_1(t),
\;\epsilon_2(t) $) in a given time increases as a result of which the transition probability to the right increases. Further increase of noise intensities results in randomization of the system. In order to examine the influence of cross-correlation time of the nonthermal noises on the the ratio of transition rate we plot in Fig.9 the variation of $k_M^R/k_M^L$ as a function of $\tau$ for several values of $\lambda$. With increase of cross-correlation time the ratio of the transition rates monotonically decreases. This is again due to the asymmetry in the diffusion coefficient which decreases with increase of $\tau$.
conclusion
==========
We have considered the stochastic dynamics of the particles in a triple well potential driven simultaneously by two cross-correlated white noise processes. It has been shown that depending on the correlation between the two noise sources, one multiplicative and another additive, the relative stability of the two side wells with respect to the middle one may differ significantly. This originates from an asymmetry in diffusive motion of the interwell dynamics due to interference of the two noises. An offshoot of this symmetry breaking effect is the splitting of Kramers’ rate of escape from the middle well to the sides wells. In a wider context the kinetic scheme may serve as a technique for preferentially selecting a pathway for a parallel chemical reaction. Thanks are due to the Council of Scientific and industrial research, Govt. of India, for partial financial support.
[99]{} H. A. Kramers, Physica [**7**]{}, 284 (1940).
P. Hänggi, P. Talkner, and M. Borkovec Rev. Mod. Phys. [**62**]{}, 251 (1990).
S. Faetti, P. Grigolini and F. Marchesoni, Z. Phys. B [**47**]{}, 353 (1982).
J. M. Sancho, M. San Miguel, S. L. Katz and J. D. Gunton, Phys. Rev. A [**26**]{}, 1589 (1982).
D. Barik, D. Banerjee and D. S. Ray, *in Progress in Chemical Physics Research*, Vol- 1, Edited by A. N. Linke, (Nova Publishers, New York, ISBN: 1-59454-451-4, 2006).
R. Benzi, G. Parisi, A. Sutera, and A. Vulpiani, Tellus [**34**]{}, 10 (1982); R. Benzi, A. Sutera, and A. Vulpiani, J. Phys. A [**14**]{} L 453 (1981).
L. Gammaitoni, P. Hänggi, P. Jung and F.Marchesoni, Rev. Mod. Phys., [**70**]{}, 223(1998).
B. McNamara and K. Wiesenfeld, Phys. Rev. A, [**39**]{}, 4854 (1988); B. McNamara, K. Wiesenfeld and R. Roy, Phys. Rev. Lett. 60, 2628 (1988).
P. K. Ghosh, D. Barik and D. S. Ray, Phys. Lett. A [**342**]{} 12 (2005).
C. R. Doering and J. C. Gadoua, Phys. Rev. Lett. [**69**]{}, 2318 (1992).
M. Bier and R. D. Astumian, Phys. Rev. Lett. [**71**]{}, 1649 (1993); U. Zürcher and C. R. Doering, Phys. Rev. E [**47**]{}, 3862 (1993). C. Van den Broeck, Phys. Rev. E [**47**]{}, 4579 (1994).
P. K. Ghosh, D. Barik, B. C. Bag and D. S. Ray, J. Chem. Phys. [**123**]{}, 224104 (2005); G. Goswami, P.Majee, P. K. Ghosh and B. C. Bag, Physica A [**374**]{} 549 (2007).
A. Schenzle and H. Brand, Phys. Rev. A, [**20**]{}, 1628 (1979).
W. Horsthemke and R. Lefever, *Noise Induced Transition, Theory and Application in Physics, Chemistry and Biology*, (Springer, Berlin 1983).
A. Fulinski and T. Telejko, Phys. Lett. A [**152**]{} 11 (1991). A. Ajdari and J. Prost, Comptes Rendus Acad. Sci. Paris [**315**]{}, 1635 (1992); M. Büttiker, Z. Phys B [**68**]{}, 161 (1987).
M. O. Magnasco, Phys. Rev. Lett. [**71**]{}, 1477 (1993); P. Reiman, Phys. Reps. [**361**]{}, 57 (2002); F. Jülicher and J. Prost, Phys. Rev. Lett. [**75**]{}, 2618 (1995).
C. R. Doering, W. Horsthemke and J. Riordan, Phys. Rev. Lett. [**72**]{}, 2984 (1994); J. Kula, T. Czernik and J. Łuczka, Phys. Rev. Lett. [**80**]{}, 1377 (1998).
P. K. Ghosh, D. Barik and D. S. Ray, Phys. Rev. E [**71**]{}, 041107 (2005); P. K. Ghosh and D. S. Ray, Phys. Rev. E [**73** ]{}, 036103 (2006).
F. Marchesoni, Phys. Lett. A [**237**]{}, 126 (1998).
P. K. Ghosh and D. S. Ray, J. Stat. Mech. P03003 (2007); G. Goswami, P.Majee, P. K. Ghosh and B. C. Bag, Physica A [**375**]{} 429 (2007).
R. F. Fox, Phys. Lett. [**94A**]{}, 281 (1983).
S. Singh, Phys. Rep. [**108**]{}, 217 (1984).
I. Fedchenia, J. Stat. Phys. [**50**]{}, 1043 (1988).
M. Borromeo and F. Marchesoni, Europhys. Lett. [**68**]{}, 783 (2004); M. Borromeo and F. Marchesoni, Phys, Rev, E [**71**]{}, 031105 (2005); M. Borromeo and F. Marchesoni, Chaos [**15**]{}, 026110 (2005)
P. K. Ghosh, B. C. Bag and D. S. Ray; Phys. Rev. E [**75**]{}, 032101 (2007).
J. Masoliver, B. J. West and K. Lindenberg, Phys. Rev. A [**35**]{}, 3086 (1987).
R. L. Stratonovich, *Topics in the Theory of Random Noise*, Vol. 1 (Gordon and Breach, New York, 1963); K. Lindenberg and B. J. West, J. Stat. Phys. [**42**]{}, 201 (1986).
W. Da-jin, C. Li and K. Sheng-zhi Phys. Rev. E [**50**]{}, 2496 (1994); Y. Jia and J. Li, Phys. Rev. E [**53**]{}, 5786 (1996).
X. Luo and S. Zhu, Phys. Rev. E [**67**]{}, 021104 (2003).
E. A. Novikov, Zh. Eksp. Teor. Fiz [**47**]{}, 1919 (1964)\[Sov. Phys. JETP [**29**]{}, 1290 (1965)\]. J. H. Li and Z. Q. Huang, Phys. Rev. E [**57**]{}, 3917 (1998).
Ya Jia and Jia-rong Li, Phys. Rev. Lett. [**78**]{}, 994 (1997).
A. J. R. Madureira, P. Hängi and H. S. Wio, Phys. Lett. A [**217**]{} 248 (1996).
Ya Jia and Jia-rong Li, Phys. Rev. E [**53**]{}, 5786 (1996).
D. C. Mei, G. Z. Xie and L. Cao and D. J. Wu, Phys. Rev. E [**59**]{}, 3880 (1999).
D. C. Mei, C. Xie and L. Zhang and D. J. Wu, Phys. Rev. E [**68**]{}, 051102 (2003).
P. Majee and B. C. Bag, J. Phys. A:Math. Gen. [**37**]{}, 3352 (2004).
[**Figure Captions**]{}
Fig.1 A schematic illustration of a symmetric triple-well potential as an energy profile of the parallel reaction (specifically, optical isomerization reaction). Fig.2. Probability distribution function $P(x)$ vs position $x$ plot depicting the changes of distribution due to the addition of external noises for the parameter set: $\tau=0,\;T=0.44, \;b=0.1 $ and $c=1.0$. (i) Solid line presents the distribution function in absence of external stochastic forces, (ii) dotted line presents the same plot in presence of external additive and multiplicative noise in absence of cross-correlation, (iii) dashed line presents also same plot in presence of external additive and multiplicative noise and their cross-correlation.
Fig.3. Variation of mean position $ \langle x \rangle $ as a function of temperature $T$ for several values of coupling strength and for the parameter set: $\tau=0.01,\; Q_1=0.02, \;
Q_2=0.02,\;b=0.1 $ and $c=1.0$.
Fig.4. Mean position($\langle x \rangle$) vs $Q_2$ (strength of the multiplicative noise) plot for several values of coupling strength and for the parameter set: $\tau=0.01,\; Q_1=0.02, \;
T=0.34,\;b=0.1 $ and $c=1.0$.
Fig.5. Mean position($\langle x \rangle$) vs $Q_1$ (strength of the multiplicative noise) plot for several values of coupling strength and for the parameter set: $\tau=0.01,\; Q_2=0.02, \; T=0.34,\;b=0.1
$ and $c=1.0$.
Fig.6. Variation of $ \langle x \rangle $ as function of cross-correlation time $\tau$ for different values of coupling strength and for the parameter set: $T=0.34, Q_1=0.02,
Q_2=0.02,b=0.1 ,c=1.0$.
Fig.7. Ratio of transition rates ($k_M^R/k_M^L$) vs temperature plot for several values of coupling strength and for the parameter set: $\tau=1.0,\; Q_2=0.05, \; Q_1=0.05,\;b=0.1 $ and $c=1.0$.
Fig.8. (a) Variation of the ratio of the transition rate ($k_M^R/k_M^L$) as function of the strength of additive noise and for several values of coupling strength and for the parameter set: $T=0.05,\; \tau=1.0, \; Q_2=0.02,\;b=0.1 $ and $c=1.0$. (b) Variation of the ratio of the transition rate ($k_M^R/k_M^L$) as function of the strength of multiplicative noise for same parameter set as (a) but for $Q_1=0.05$.
Fig.9. Variation of the ratio of the transition rate ($k_M^R/k_M^L$) as function of cross-correlation time $\tau$ for different values of coupling strength and for the parameter set: $T=0.05,\; Q_1=0.05, \;
Q_2=0.05,\;b=0.1 $ and $c=1.0$.
{width="15.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
{width="20.2cm"}
[^1]: e-mail address: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Predicting the runtime complexity of a programming code is an arduous task. In fact, even for humans, it requires a subtle analysis and comprehensive knowledge of algorithms to predict time complexity with high fidelity, given any code. As per Turing’s Halting problem proof, estimating code complexity is mathematically impossible. Nevertheless, an approximate solution to such a task can help developers to get real-time feedback for the efficiency of their code. In this work, we model this problem as a machine learning task and check its feasibility with thorough analysis. Due to the lack of any open source dataset for this task, we propose our own annotated dataset *CoRCoD: Code Runtime Complexity Dataset*[^1], extracted from online judges. We establish baselines using two different approaches: feature engineering and code embeddings, to achieve state of the art results and compare their performances. Such solutions can be widely useful in potential applications like automatically grading coding assignments, IDE-integrated tools for static code analysis, and others.'
author:
- Jagriti Sikka
- Kushal Satya
- Yaman Kumar
- Shagun Uppal
- Rajiv Ratn Shah
- Roger Zimmermann
bibliography:
- 'references.bib'
title: Learning based Methods for Code Runtime Complexity Prediction
---
Introduction
============
Time Complexity computation is a crucial aspect in the study and design of well-structured and computationally efficient algorithms. It is a measure of the performance of a solution for a given problem. As a popular mistaken consideration, it is not the execution time of a code. Execution time depends upon a number of factors such as the operating system, hardware, processors etc. Since execution time is machine dependent, it is not used as a standard measure to analyze the efficiency of algorithms. Formally, *Time Complexity* quantifies the amount of time taken by an algorithm to process as a function of the input. For a given algorithm, we consider its worst case complexity, which reflects the maximum time required to process it, given an input. Time complexity is represented in **Big O** notation. $O(n)$ denotes the asymptotic upper bound of an algorithm as a function of the input size $n$. Typically, the complexity classes in Computer Science refer to P, NP classes of decision problems, however, for the entire length of this paper, complexity class refers to a category of time complexity. The commonly considered categories in computer science as well in our work are $O(1)$, $O(logn)$, $O(n)$ , $O(nlogn)$ and $O(n^2)$.
In this work, we try to predict the time complexity of a solution, given the code. This can have widespread applications, especially in the field of education. It can be used in automatic evaluation of code submissions on different online judges. It can also aid in static analyses, informing developers how optimized their code is, enabling more efficient development of industry level solutions.
A number of ways have been proposed to estimate code complexity. McCabe et al. [@mccabe] is one of the earliest works done in this area, which proposed **Cyclomatic Complexity**. It is defined as the quantitative measure of the number of linearly independent paths through a program’s source code computed using the control flow graph of the program. Cyclomatic Complexity of a code can be calculated through Equation \[Equation1\]. $$C = E - N + M
\label{Equation1}$$ where $E$ is the number of the edges, $N$ is the number of nodes and $M$ is the number of connected components.
Cyclomatic Complexity quantitatively measures a program’s logical strength based on existing decision paths in the source code. However, the number of independent paths does not represent how many times these paths were executed. Hence, it is not a robust measure for time complexity. Bentley et al. [@master] proposed the **Master Theorem**. For the generic divide and conquer problem which divides a problem of input size $n$ into $a$ subproblems each of size $\frac{n}{b}$ ,and combine the result in $f(n)$ operations, Master theorem expresses the recurrence relation as Equation \[masters theorem\]. $$T(n) = aT(\frac{n}{b}) + f(n)
\label{masters theorem}$$ However, the master theorem is limited to divide and conquer problems and has several constraints on the permissible values of $a$, $b$ and $f(n)$.
Mathematically speaking, it is impossible to find a universal function to compute the time complexity of all programs. Rice’s theorem and other works in this area [@haltingproof; @ricetheorem] have established that the runtime complexity for problems in category P is undecidable i.e. it is impossible to formulate a mathematical function to calculate the complexity of any code with polynomial order complexity. Therefore, we need a Machine Learning based solution which can learn the internal structure of the code effectively. Recent research in the areas of machine learning and deep learning for programming codes provide several potential approaches which can be extended to solve this problem [@code2vec; @graph2vec]. Also, several *“Big Code”* datasets have been made available publicly. The Public Git Archive is a dataset of a large collection of Github repositories [@public_git_archive_dataset], [@so_dataset] and [@StacQ] are datasets of Question-code pairs mined from Stack Overflow. However, to the best of our knowledge, at the time of writing this paper, there is no existing public dataset that, given the source code, gives runtime complexity of the source code. In our work, we have tried to address this problem by creating a Code Runtime Complexity Dataset *(CoRCoD)* consisting of 932 code files belonging to 5 different classes of complexities, namely *$O(1)$, $O(logn)$, $O(n)$, $O(nlogn)$* and *$O(n^2)$* (see Table \[table:classwise\_data\]).\
We aim to substantially explore and solve the problem of code runtime complexity prediction using machine learning with the following contributions:
- Releasing a novel annotated dataset of program codes with their runtime complexities.
- Proposing baselines of ML models with hand-engineered feature and study of how these features affect the computational efficiency of the codes.
- Proposing another baseline, the generation of code embeddings from Abstract Syntax Tree of source codes to perform classification.
Furthermore, we find that code embeddings have a comparable performance to hand-engineered features for classification using Support Vector Machines (SVMs). To the best of our knowledge, CoRCoD is the first public dataset for code runtime complexity, and this is the first work that uses Machine Learning for runtime complexity prediction.\
The rest of this paper is structured as follows. In Section 3, we talk about dataset curation and its key characteristics. We experiment using two different baselines on the dataset: classification using hand engineered features extracted from code and using graph based methods to extract the code embeddings via Abstract Syntax Tree of code. Section 4 explains the details and key findings of these two approaches. In Section 5, we enumerate the results of our model and data ablation experiments performed on these two baselines.
Related Work
============
In recent years, there has been extensive research in the deep learning community on programming codes. Most of the research has been focused on two buckets, either on predicting some structure/attribute in the program or generating code snippets that are syntactically and/or semantically correct. In the latter case, Sun et al. [@Sun_Zhu_Mou_Xiong_Li_Zhang_2019] made use of the Abstract Syntax Tree (AST) of a program, and generated code by predicting the grammar rules using a CNN.
Variable/Method name prediction is a widely attempted problem, wherein Allamanis et al. [@pmlr-v48-allamanis16] used a convolutional neural network with attention technique to predict method names, Alon et al. [@ASTPaths-code2vec] suggested the use of AST paths to be used as context for generating *code embeddings* and training classifiers on top of them. Yonai et al. [@MethodNamePred-CallGraphEmbedding] used call graphs to compute method embeddings and recommend names of existing methods with function similar to target function.
Another popular prediction problem is that of defect prediction, given a piece of code. Li et al. [@DefectPred] used Abstract Syntax Trees of programs in their CNN for feature generation which are then used for defect prediction. A major goal in all these approaches is to come up with a representation of the source program, which effectively captures the syntactic and semantic features of the program. Chen and Monperrus [@CodeEmbeddings-LiteratureStudy] performed a survey on word embedding techniques used on source codes. However, so far, there has been no such work for predicting time complexity of programs using code embeddings. We have established the same as one of our baselines using graph2vec [@graph2vec].
Srikant and Aggarwal[@AspiringMinds] extract hand-engineered features from Control Flow and Data Dependency graphs of programs such as number of nested loops, number of instances of *if* statements in a $loop$ etc. for automatic grading of programs. They then used the grading criteria, that correct test programs would have similar programming constructs/features as those in the correct hand-graded programs. We use the same idea of identifying key features as the other baseline, which are constructs that a human evaluator would look at, to compute complexity and use them to train the classification models. Though, unlike [@AspiringMinds], our features are problem independent. Moreover, their dataset is not publicly available.
Dataset
=======
To construct our dataset, we collected source codes of different problems from Codeforces[^2]. Codeforces is a platform that regularly hosts programming contests. The large availability of contests having a wide variety of problems both in terms of data structures and algorithms as well as runtime complexity, made Codeforces a viable choice for our dataset.
Complexity class Number of samples
------------------ -------------------
$O(n)$ 385
$O(n^2)$ 200
$O(nlogn)$ 150
$O(1)$ 143
$O(logn)$ 55
: Extracted features[]{data-label="table:features"}
[|c|c|]{}\
Number of methods & Number of breaks\
Number of switches & Number of loops\
Priority queue present & Sort present\
Hash map present & Hash set present\
Nested loop depth & Recursion present\
Number of variables & Number of ifs\
Number of statements & Number of jumps\
For the purpose of construction of our dataset, we collected Java source codes from Codeforces. We used the Codeforces API to retrieve problem and contest information, and further used web scraping to download the solution source codes.
In order to ensure correctness of evaluated runtime complexity, the source codes selected should be devoid of issues such as compilation errors and segmentation faults. To meet this criterion, we filtered the source codes on the basis of their verdict and only selected the codes having verdicts *Accepted* or *Time limit exceeded* (TLE). For codes having TLE verdict, we ensured accuracy of solutions by only selecting codes that successfully passed at least four Test Cases. This criterion also allowed us to include multiple solutions for a single problem, different solutions having different runtime complexities. These codes were then manually annotated by a group of five experts, hailing from programming background each with a bachelor’s degree in Computer Science or related field. Each code was analyzed and annotated by two experts, in order to minimize the potential for error. Only the order of complexity was recorded, for example, a solution having two variable inputs, $n$ and $m$, and having a runtime complexity of $O(n*m)$ is labeled as $n\_square$ ($O({n}^2)$).
Certain agreed upon rules were followed for the annotation process. The rationale relies on the underlying implementations of these data structures in Java. Following points list down the rules followed for annotation and the corresponding rationale:
- Sorting algorithm’s implementation in Java collections has worst case complexity $O(nlogn)$.
- Insertion/retrieval in HashSet and HashMap is annotated to be $O(1)$, given $ n $ elements.
- TreeSet and TreeMap are implemented as Red-Black trees and thus have $O(logn)$ complexity for insertion/retrieval.
\[Algorithm1\]
``` {.java breaklines="" frame="single"}
class noOfNestedLoops extends ASTVisitor {
int current = 0;
int max_depth = 0;
@Override
bool visit(WhileStatement node) {
current += 1;
max_depth = max(current, max_depth);
return true;
}
@Override
void endVisit(WhileStatement node){
current -= 1;
}
}
```
We removed few classes with insufficient data points, and ended up with 932 source codes, 5 complexity classes, corresponding annotation and extracted features. We selected nearly 400 problems from 170 contests, picking an average of 3 problems per contest. For 120 of these problems, we collected 4-5 different solutions, with different complexities.
Solution Approach
=================
The classification model is trained using two approaches: one, extracting hand-engineered features from code using static analysis and two, learning a generic representation of codes in the form of code embeddings.
Feature Engineering
-------------------
#### Feature Extraction.
We identified key coding constructs and extracted 14 features (refer Table \[table:features\]). We extracted these features from the Abstract Syntax Tree (AST) of source codes. AST is a tree representation of syntax rules of a programming language. ASTs are used by compilers to check codes for accuracy. We used Eclipse JDT for feature extraction. A generic representation of AST as parsed by ASTParser in JDT is shown in Figure \[fig:AST\].
![Extraction of features from code using AST Parser[]{data-label="fig:AST"}](ASTParser){height="8.5cm" width="9.5cm"}
An ASTParser object creates the AST, and the ASTVisitor object *“visits”* the nodes of the tree via *visit* and *endVisit* methods using Depth First Search. One of the features chosen was the maximum depth of nested loops. Code snippet in Listing \[Algorithm1\] depicts how the value of depth of Nested Loops was calculated using ASTVisitor provided by JDT. Other features were calculated in a similar manner.
We observed that our code samples often have unused code like methods or class implementations never invoked from the main function. Removing such unused code manually from each code sample is tedious. Instead, we used JDT plugins to identify the methods reachable from main function and used those methods for extracting listed features. The same technique was also used while creating the AST for the next baseline.
[0.5]{} ![Density plot for the different features[]{data-label="fig:density plot"}](images/NestedLoopDepthIm.png "fig:"){width="\textwidth"}
[0.5]{} ![Density plot for the different features[]{data-label="fig:density plot"}](images/SortIm.png "fig:"){width="\textwidth"}
Figure \[fig:density plot\] represents the density distribution of features across different classes. For nested loops, $n\_square$ has peak at depth $2$ as expected; similarly $n$ and $nlogn$ have peak at depth $1$ loop depth (see Figure \[fig:density plot\]()). For number of sort calls, $n$ and $n\_square$ have higher peaks at $0$, indicating peak in the absence of sort whereas $nlogn$ peaks both at $0$ and $1$ (see Figure \[fig:density plot\]()). Upon qualitative analysis, we found that a large number of solutions were performing binary search within a *for* loop, which explains the peak at $0$, and others were simply using sort, hence the peak at $1$. This confirms our intuition that sort calls and nested loops are important parameters in complexity computation.
Code Embeddings
---------------
The Abstract Syntax Tree of a program captures comprehensive information regarding a program’s structure, syntactic and semantic relationships between variables and methods. An effective method to incorporate this information is to compute code embeddings from the program’s AST. We use graph2vec , a neural embedding framework [@graph2vec], to compute embeddings. Graph2vec automatically generates task agnostic embeddings, and does not require a large corpus of data, making it apt for our problem.We used the graph2vec implementation from [@graph2vec_implementation] to compute code embeddings.
Graph2vec is analogous to doc2vec[@doc2vec] which predicts a document embedding given the sequence of words in it. The goal of graph2vec is, given a set of graphs $\mathbb{G} = \{G_1, G_2, ... G_n\}$, learn a $\delta$ dimensional embedding vector for each graph. Here, each graph $G$ is represented as $(N,E,\lambda)$ where $N$ are the nodes of the graph, E the edges and $\lambda$ represents a function $n \rightarrow l$ which assigns a unique label from alphabet $l$ to every node $n \in N$. To achieve the same, graph2vec extracts nonlinear substructures, more specifically, rooted subgraphs from each graph which are analogical to words in doc2vec. It uses skipgram model for learning graph embeddings which correspond to code embeddings in our scenario. The model works by considering a subgraph $s_j \in c(g_i)$ to be occurring in the context of graph $g_i$ and tries to maximize the log likelihood in Equation \[loglikelihood\]:
$$\sum_{j=1}^{D} log \; Pr({s_j}|{g_i})
\label{loglikelihood}$$
where $c(g_i)$ gives all subgraphs of a graph $g_i$ and $D$ is the total number of subgraphs in the entire graph corpus.
We extracted AST from all codes using the JDT plugins. Each node in AST has two attributes: a Node Type and an optional Node Value. For e.g., a MethodDeclaration Type node will have the declared function name as the node value. Graph2vec expects each node to have a single label. To get a single label, we followed two different representations:
1. Concatenating Node Type and Node Value.
2. Choosing selectively for each type of node whether to include node type or node value. For instance, every identifier node has a SimpleName node as its child. For all such nodes, only node value i.e. identifier name was considered as the label.
For both the AST representations, we used graph2vec to generate 1024-dimensional code embeddings. These embeddings are further used to train SVM based classification model and several experiments are performed as discussed in the next section.
Experiments and Results
=======================
Feature Engineering
-------------------
Deep Learning(*DL*) algorithms tend to improve their performance with the amount of data available unlike classical machine learning algorithms. With lesser amount of data and correctly hand engineered features, Machine Learning(*ML*) methods outperform many *DL* models. Moreover, the former are computationally less expensive as compared to the latter. Therefore, we choose traditional *ML* classification algorithms to verify the impact of various features present in programming codes on their runtime complexities. We also perform a similar analysis on a simple Multi level Perceptron*(MLP)* classifier and compare against others. Table \[table:score\_precision\_recall\] depicts the accuracy score, weighted precision and recall values for this classification task using 8 different algorithms, with the best accuracy score achieved using the ensemble approach of random forests.
Algorithm Accuracy Precision Recall
--------------------- ----------- ----------- -----------
K-means 54.76 54.34 53.95
Random forest **74.26** 70.85 73.19
Naive Bayes 57.75 60.35 58.06
k-Nearest 59.89 59.35 58.57
Logistic Regression 73.19 72.89 73.19
Decision Tree 73.79 71.86 71.12
MLP Classifier 63.10 59.13 58.28
**SVM** **72.96** **69.43** **70.58**
: Per feature accuracy score, averaged over different classification algorithms.[]{data-label="table:per feature"}
Feature Mean Accuracy
------------------------ ---------------
No. of ifs 44.35
No. of switches 44.38
No. of loops 51.33
No. of breaks 43.85
Priority Queue present 45.45
No. of sorts 52.40
Hash Set present 44.38
Hash Map present 43.85
Recursion present 42.38
Nested loop depth **66.31**
No. of Variables 42.78
No. of methods 42.19
No. of jumps 43.65
No. of statements 44.18
: Per feature accuracy score, averaged over different classification algorithms.[]{data-label="table:per feature"}
Further, as per Table \[table:per feature\] showing per-feature-analysis, we distinctly make out that for the collected dataset, the most prominent feature which solely gives maximum accuracy is nested loop depth, followed by number of sorts and loops. For all algorithms, accuracy peaks uptil the tenth feature, and decreases thereafter [@additional_results]. This comprehensibly depicts that for the five complexity classes considered here, the number of variables, number of methods, number of jumps and number of statements do not play a notable role as compared to the other features.
Another interesting feature that comes up with the qualitative study of these accuracy scores is that with these features, most of the true classification predictions come up with code samples of category $O{(1)}$, $O{(n)}$ and $O{(n}^2)$. This holds a subtle resonance with the inherent obvious instincts of a human classifier who is more likely to correctly classify a program for these complexities instantly as compared to $O{(logn)}$ or $O{(nlogn)}$. Tables \[table:accuracy\_3\_classes\] and \[table:accuracy\_3\_classes\_2\] demarcate the difference between accuracy scores considering data samples from classes $O{(1)}$, $O{(n)}$, $O{(n}^2)$ as compared to classes $O{(1)}$, $O{(logn)}$, $O{(nlogn})$. A clear increment in accuracy scores is noticed amongst all the algorithms considered for the classification task for the former set of 3 classes considered as compared to the latter.
Algorithm Accuracy Precision Recall
--------------------- ----------- ----------- --------
K-means 64.38 63.76 63.04
Random forest **82.61** 81.32 79.37
Naive Bayes 63.42 61.44 61.09
k-Nearest 65.56 67.54 66.23
Logistic Regression 81.58 81.70 80.98
Decision Tree 80.12 81.19 78.24
MLP Classifier 70.06 65.53 70.31
SVM 76.43 72.14 74.35
: Accuracy, Precision and Recall values for different classification algorithms considering samples from complexity classes $O{(1)}$, $O{(logn)}$ and $O{(nlogn})$[]{data-label="table:accuracy_3_classes_2"}
Algorithm Accuracy Precision Recall
--------------------- ---------- ----------- --------
K-means 52.31 53.23 50.04
Random forest 64.05 71.21 63.24
Naive Bayes 61.05 68.21 60.24
k-Nearest 63.56 64.45 64.07
Logistic Regression 75.46 71.24 70.98
Decision Tree 74.74 72.12 77.05
MLP Classifier 67.34 68.45 67.43
SVM 69.64 70.76 67.24
: Accuracy, Precision and Recall values for different classification algorithms considering samples from complexity classes $O{(1)}$, $O{(logn)}$ and $O{(nlogn})$[]{data-label="table:accuracy_3_classes_2"}
Code Embeddings
---------------
We extracted ASTs from source codes, computed 1024-dimensional code embeddings from ASTs using graph2vec and trained an SVM classifier on these embeddings. Results are tabulated in Tables \[table:graph2vec\_results\]. We note that the maximum accuracy obtained for SVM on code embeddings is comparable to that of SVM on statistical features. Also, code embeddings baseline has better precision and recall scores for both representations of AST.
Data Ablation Experiments
-------------------------
To get further insight into the learning framework, we perform following data ablation tests:
*Label Shuffling.* Training models with shuffled class labels can indicate whether the model is learning useful features pertaining to the task at hand. If the performance does not significantly decrease upon shuffling, it can imply that the model is hanging on to statistical cues that do not contain meaningful information w.r.t. the problem.
*Method/Variable Name Alteration.* Graph2vec uses node labels along with edge information to generate graph embeddings. Out of randomly selected 50 codes having correct prediction, if the predicted class labels before and after data ablation are different for a significant number of test samples, it would imply it would imply that the model relies on method/variable name tokens whereas it should only rely on the relationships between variables/methods.
*Replacing Input Variables with Constant Literals.* Program complexity is a function of input variables. Thus, to test the robustness of models, we replace the input variables with constant values making resultant complexity $O(1)$ for 50 randomly chosen codes, which earlier had non-constant complexity. A good model should have a higher percentage of codes with predicted complexity as $O(1)$.
*Removing Graph Substructures.* We randomly remove program elements such as *for*, *if* blocks with a probability of 0.1. The expectation is that the correctly predicted class labels should not change heavily as the complexity most likely does not change and hence should have a higher percentage of codes with same correct label before and after removing graph substructures. This would imply that the model is robust to changes in code that do not change the resultant complexity.
Following are our observations regarding data ablation results in Table \[table:graph2vec\_results2\]:
*Label Shuffling.* The drop in test performance is higher in graph2vec than that in the basic model indicating that graph2vec learns better features compared to simple statistical models.
*Method/Variable Name Alteration.* Table \[table:graph2vec\_results2\] shows that SVM correctly classifies most of the test samples’ embeddings upon altering method and variable names, implying that the embeddings generated do not rely heavily on the actual method/variable name tokens.
*Replacing Input Variables with Constant Literals.* We see a significant and unexpected dip in accuracy, highlighting one of the limitations of our model.
*Removing Graph Substructures.* Higher accuracy for code embeddings as compared to feature engineering implies that the model must be learning the types of nodes and their effect on complexity to at least some extent, as removing substructures does not change the predicted complexity class of a program significantly.
Limitations
===========
The most pertinent limitation of the dataset is its size which is fairly small compared to what is considered standard today. Another limitation of our work is low accuracy of the models. An important point to note is that although we established that using code embeddings is a better approach, still their accuracy does not beat feature engineering significantly. One possible solution is to increase dataset size so that generated code embeddings can better model the characteristics of programs that differentiate them into multiple complexity classes, when trained on larger number of codes. However, generating a larger dataset is a challenging task since annotation process is tedious and needs people with a sound knowledge of algorithms. Lastly, we observe that replacing variables with constant literals does not change the prediction to $O(1)$ which highlights the inability of graph2vec to identify the variable on which complexity depends.
Usefulness of the Dataset
=========================
Computational complexity is a quantification of computational efficiency. Computationally efficient programs better utilize resources and improve software performance. With rapid advancements, there is a growing demand for resources; at the same time, there is greater need for optimizing existing solutions. Thus, writing computationally efficient programs is an asset for both students and professionals. With this dataset, we aim to analyze attributes and capture relationships that best define the computational complexity of codes. We do so, not just by heuristically picking up evident features, but by investigating their role in the quality, structure and dynamics of the problem using *ML* paradigm. We also capture relationships between various programming constructs by generating code embeddings from Abstract Syntax Trees. This dataset can not only help automate the process of predicting complexities, but also, to help learners decide apt features for well-structured and efficient codes. It is essential for better programming ethics apart from a good algorithm, since a good algorithm might get restricted performance due to inefficient coding, which can add an additional ease and value to educational purposes. It can also be used to train models that can be further integrated with IDEs and assist professional developers in writing computationally efficient programs for fast performance software development.
Conclusion
==========
The dataset presented and the baseline models established should serve as guidelines for the future work in this area. The dataset presented is balanced and well-curated. Though both the baselines; Code Embeddings and Handcrafted features have comparable accuracy, we have established through data ablation tests that code embeddings learned from Abstract Syntax Tree of the code better capture relationships between different code constructs that are essential for predicting runtime complexity. Work can be done in future to increase the size of the dataset to verify our hypothesis that code embeddings will perform significantly better than hand crafted features. Moreover, we hope that the approaches discussed in this work, their usage becomes explicit for programmers and learners to bring into practice efficient and optimized codes.
[^1]: The complete dataset is available for use at <https://github.com/midas-research/corcod-dataset>.
[^2]: https://codeforces.com
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We derive a “Kompaneets equation” for neutrinos, which describes how the distribution function of neutrinos interacting with matter deviates from a Fermi-Dirac distribution with zero chemical potential. To this end, we expand the collision integral in the Boltzmann equation of neutrinos up to the second order in energy transfer between matter and neutrinos. The distortion of the neutrino distribution function changes the rate at which neutrinos heat matter, as the rate is proportional to the mean square energy of neutrinos, $E_\nu^2$. For electron-type neutrinos the enhancement in $E_\nu^2$ over its thermal value is given approximately by $E_\nu^2/E_{\nu,\rm thermal}^2=1+0.086(V/0.1)^2$ where $V$ is the bulk velocity of nucleons, while for the other neutrino species the enhancement is $(1+\delta_v)^3$, where $\delta_v=mV^2/3k_BT$ is the kinetic energy of nucleons divided by the thermal energy. This enhancement has a significant implication for supernova explosions, as it would aid neutrino-driven explosions.'
author:
- Yudai Suwa
- 'Hiroaki W. H. Tahara'
- Eiichiro Komatsu
date: 'Received April 10, 2019; Revised June 15, 2019; Accepted July 3, 2019; Published August 23, 2019'
title: 'Kompaneets equation for neutrinos: Application to neutrino heating in supernova explosions'
---
Introduction {#sec:intro}
============
The Boltzmann equation is ubiquitous in physics. In a system in which matter and radiation interact, a useful approximation can be obtained by expanding the collision integral in the Boltzmann equation up to second order in energy transfer between matter and radiation.
One example is the Kompaneets’s equation [@komp57], which describes how the distribution function of photons evolves via interaction with thermal electrons in the non-relativistic limit. A solution to the Kompaneets equation in the optically-thin limit is known as the thermal Sunyaev-Zel’dovich (tSZ) effect [@suny72], which describes a distortion of the black-body spectrum of the cosmic microwave background photons by inverse-Compton scattering off hot electrons in galaxy clusters. The tSZ effect has been routinely detected towards ${\mathcal O}(10^3)$ galaxy clusters [@plan15]. Another example is the Fermi acceleration [@ferm49]. The so-called second-order Fermi acceleration distorts the distribution of charged particles by stochastic acceleration due to time-dependent electromagnetic fields. Both examples can be formulated in the same form, namely a diffusion equation for the distribution function of photons or charged particles in momentum space.
In this paper, we apply the same approximation to the Boltzmann equation describing neutrinos interacting with matter. Specifically, we consider a system in which isotropic neutrinos interact with nucleons, and expand the collision integral in the Boltzmann equation up to second order in energy transfer between neutrinos and nucleons. We do not assume that the system is optically thin. By solving this Kompaneets-like equation for neutrinos, we obtain distortions of the neutrino distribution function from a Fermi-Dirac distribution with zero chemical potential as a function of the temperature and bulk velocity of nucleons. Our result has a significant implication for neutrino-driven supernova explosions, as the distortion of the distribution function changes the rate at which neutrinos heat nucleons.
The rest of this paper is organized as follows. In Section \[sec:kompaneets\], we derive the Kompaneets-like equation for neutrinos interacting with nucleons; the nucleon motion includes both thermal and bulk motion. We present solutions of this equation in Section \[sec:solution\], including the effect of opacity of electron-type neutrinos. We summarize our results and discuss their implications in Section \[sec:summary\]. In Appendix \[sec:matrix\] we review the matrix element of neutrino-nucleon scattering. In Appendix \[sec:rybicki\] we provide an alternative derivation of the main result of this paper following the argument of Ref. [@rybi79]. Throughout the paper we shall set the speed of light to be unity, $c=1$.
Kompaneets equation for neutrinos {#sec:kompaneets}
=================================
We follow derivations of the Kompaneets equation for photons interacting with electrons given in Refs. [@hu94; @dode95], and derive a similar equation for neutrinos interacting with nucleons. To this end, we expand the collision integral of the Boltzmann equation up to second order in energy transfer. We shall ignore the mass of neutrinos throughout this paper, as the typical neutrino temperature that we consider here (e.g. that in the supernova engine) is much greater than the current upper bound on the mass of neutrinos on the order of 1 eV.
The Boltzmann equation is $$\begin{aligned}
\frac{df(\bm{p},t)}{dt}=&
{\sum_{N}}
\frac1{2p}
\int\frac{d^3q}{(2\pi)^32E_N(\bm{q})}\int\frac{d^3q'}{(2\pi)^32E_N(\bm{q}')}
\int \frac{d^3p'
}{(2\pi)^32p'}
\delta^{(4)}(p+q-p'-q')|M_N|^2
\nonumber\\
&\times\{
g_N({\bm q}')f(\bm{p}',t)[1-f(\bm{p},t)]
-g_N({\bm q})f(\bm{p},t)[1-f(\bm{p}',t)]
\},\end{aligned}$$ where $N$ denotes nucleons (neutrons and protons), $d/dt$ is the Lagrangian time derivative along the trajectory of a phase-space volume element, $|M_N|^2$ is the spin-averaged matrix element of neutrino-nucleon scattering, and $f(\bm{p},t)$ and $g_N({\bm q})$ are the distribution functions of neutrinos and nucleons with three-momenta $\bm{p}$ and ${\bm q}$, respectively. The four-dimensional Dirac delta function $\delta^{(4)}(p+q-p'-q')$ ensures energy and momentum conservation.
We assume that nucleons are non-relativistic, i.e. $E_N(\bm{q})\equiv \sqrt{m_N^2+|\bm{q}|^2}\approx m_N+|\bm{q}|^2/(2m_N)$, with the distribution function given by $$g_N({\bm q})=n_N(2\pi m_N k_BT_N)^{-3/2}\exp\left[-\frac{({\bm q}-m_N {\bm v}_N)^2}{2m_Nk_BT_N}\right]\,,$$ where $n_N$, $m_N$, $T_N$ and $\bm{v}_N$ are the number density, mass, temperature and bulk velocity of nucleons, respectively, and $k_B$ is the Boltzmann constant.
Performing integration over $\bm{q}'$, we obtain $$\begin{aligned}
\frac{df}{dt}=&
{\sum_{N}}
\frac1{8\pi p}\int dp'p'\frac{d\Omega'}{4\pi}
\int\frac{d^3q}{(2\pi)^3}
\frac{|M_N|^2}{E_N(\bm{q})E_N(\bm{q}+\bm{p}-\bm{p}')}\delta[p+E_N(\bm{q})-p'-E(\bm{q}+\bm{p}-\bm{p}')]
\nonumber\\
&\times\{
g_N({\bm q}+\bm{p}-\bm{p}')f(\bm{p}',t)[1-f(\bm{p},t)]
-g_N({\bm q})f(\bm{p},t)[1-f(\bm{p}',t)]
\}.
\label{eq:boltzmann}\end{aligned}$$ The energy transfer is on the order of $E_N(\bm{q})-E_N(\bm{q}+\bm{p}-\bm{p}')\approx (\bm{p}'-\bm{p})\cdot\bm{q}/m_N$, which should be much smaller than $E_N(\bm{q})\approx |\bm{q}|^2/(2m_N)$.
The next step is to expand $|M_N|^2$, $E(\bm{q}+\bm{p}-\bm{p}')$, the delta function and the distribution functions up to the second order in energy transfer. In this paper we consider elastic neutrino-nucleon scattering, i.e. $\nu+N\to \nu+N$, and ignore other processes.
There are three types of the so-called “neutrinosphere,” which are determined by different micro processes [@raff01]:
- [**Number sphere**]{}: The optical depth of emission and absorption of neutrinos is of order unity. For electron-type neutrinos $\nu_\e$ and $\bar\nu_\e$, the processes of electron/positron capture and its inverse processes (i.e. $\nu_\e + \n \leftrightarrow \p+\e^-$ and $\bar\nu_\e + \p
\leftrightarrow \n+\e^+$) are important. For other two types of neutrinos $\nu_\mu, \nu_\tau,\bar\nu_\mu,\bar\nu_\tau$, the pair production/annihilation processes (i.e. $\nu\bar\nu\leftrightarrow\gamma\gamma$, $\nu\bar\nu\leftrightarrow
\e^+\e^-$, $NN\leftrightarrow NN\nu\bar\nu$) determine the opacity.
- [**Energy sphere**]{}: Inelastic scattering by electrons is important here. Electrons receive energy from neutrinos, as the electron rest mass energy (511 keV) is much smaller than the typical neutrino energy ($\sim 10$ MeV), which is determined by the matter temperature at the number sphere. Inside the energy sphere the neutrinos are thermalized due to energy transfer with electrons, which are tightly coupled with baryons.
- [**Transport sphere**]{}: Beyond the energy sphere, elastic scattering by nucleons and nuclei determines the opacity. As the rest mass energy of these particles is much larger than the typical neutrino energy, scattering can be treated as elastic.[^1]
For $\nu_\e$ and $\bar\nu_\e$ all the neutrinospheres are almost coincident, and thus the spectrum is almost thermal.[^2] For $\nu_\mu, \nu_\tau,\bar\nu_\mu$, and $\bar\nu_\tau$, on the other hand, the radii of the neutrinospheres are separated from each other [@raff01], and thus neutrinos may establish a non-thermal component, which would be produced between the energy and transport spheres. This is the region we consider in this paper.
With the matrix element for neutrino-nucleon scattering given in Appendix \[sec:matrix\], the right-hand side of Eq. (\[eq:boltzmann\]) becomes $$\begin{aligned}
&{\sum_{N}}\frac{G_F^2}{\pi} p
\int d p' p' \frac{d \Omega^{\prime}}{4 \pi} \int \frac{d^{3} \bm{q}}{(2 \pi)^{3}}~g_N(\bm{q})
\Big\{
\left[ ({F}_1^N)^2 + 3({F}_A^N)^2 \right]
+ \left[ ({F}_1^N)^2 - ({F}_A^N)^2 \right] \cos \theta
\nonumber\\
&\mp 4 \frac{p}{m_N} ({F}_1^N + {F}_2^N) {F}_A^N (1 - \cos \theta)
- \frac{p}{m_N} ({F}_1^N)^2 (1 - \cos^2 \theta)
- \frac{p}{m_N} ({F}_A^N)^2 (3 - 4 \cos \theta + \cos^2 \theta)
\nonumber\\
&-4 [({F}_1^N)^2 + ({F}_A^N)^2] \frac{{\hat p} \cdot {\bf q}}{m_N}
+ [({F}_1^N)^2 - ({F}_A^N)^2] (1 - \cos \theta)
\frac{{\hat p} \cdot {\bf q} - {\hat p'} \cdot {\bf q}}{m_N}
\Big\}
\nonumber\\
\times&\left\{\delta\left(p-p^{\prime}\right)+\frac{\left(\bm{p}-\bm{p}'\right) \cdot \bm{q}}{m_N} \frac{\partial \delta\left(p-p'\right)}{\partial p'}
+\frac{\left(\bm{p}-\bm{p}'\right)^{2}}{2 m_N} \frac{\partial \delta\left(p-p'\right)}{\partial p'}+\frac{1}{2}\left[\frac{\left(\bm{p}-\bm{p}' \right) \cdot \bm{q}}{m_N}\right]^{2} \frac{\partial^{2} \delta\left(p-p'\right)}{\partial p^{\prime 2}} \right\}
\nonumber\\
\times&
\left\{f^{(0)}\left(p^{\prime}\right)-f^{(0)}(p)-f^{(0)}\left(p^{\prime}\right)\left(1-f^{(0)}(p)\right) \frac{\left(\bm{p}-\bm{p}^{\prime}\right) \cdot\left(\bm{q}-m_N \bm{v}_N\right)}{m_Nk_BT_N}\right.
\nonumber\\
&\qquad \left. +f^{(0)}\left(p^{\prime}\right)\left(1-f^{(0)}(p)\right)\left(\frac{-\left(\bm{p}-\bm{p}^{\prime}\right)^{2}}{2 m_Nk_B T_N}+\frac{1}{2}\left(\frac{\left(\bm{p}-\bm{p}^{\prime}\right) \cdot\left(\bm{q}-m_N \bm{v}_N\right)}{m_N k_B T_N}\right)^{2}\right) \right\}\,,
\label{eq:expanded}\end{aligned}$$ where $\mp$ takes $-$ for neutrinos and $+$ for anti-neutrinos, and $f^{(0)}(p)$ is the zeroth-order distribution function of neutrinos, which is a Fermi-Dirac distribution with zero chemical potential. We shall ignore the feedback of distorted neutrino spectrum in this paper ($f^{(1)}$ and $f^{(2)}$ in Ref. [@dode95]). Here, $\hat{p}$ is a unit vector, $\cos\theta=\hat{p}\cdot\hat{p}'$, $G_F$ is the Fermi coupling constant and $F_1^N$, $F_2^N$ and $F_A^N$ are the form factors appearing in $|M_N|^2$ (see Appendix \[sec:matrix\]). The derivative of the Dirac delta function will be handled by integration by parts.
Performing integration over $\bm{q}$ yields $$\begin{aligned}
&\int \frac{d^3q}{(2\pi)^3}~g_N(\bm{q})=n_N\,,\quad
\int \frac{d^3q}{(2\pi)^3}~g_N(\bm{q})\bm{q}=n_Nm_N\bm{v}_N\,,\\
&\int \frac{d^3q}{(2\pi)^3}~g_N(\bm{q})q_iq_j=n_N\left(m_Nk_BT_N\delta_{ij}+m_N^2v_{Ni}v_{Nj}\right)\,.\end{aligned}$$
To perform integration over the solid angle of $\hat p'$, $\int d\Omega'$, we write $\bm{p}'\cdot\bm{v}_N=p'v_N\mu'$ and $\int d\Omega'=\int_0^{2\pi}d\phi'\int_{-1}^1 d\mu'$. Using the addition theorem of Legendre polynomials $P_\ell(x)$, we have $(2\pi)^{-1}\int d\phi'~P_\ell(\hat p\cdot\hat p')=P_\ell(\mu)P_\ell(\mu')$ [@dode95]. Therefore, integration over $\phi'$ yields the following substitutions: $\cos\theta\to \mu\mu'$ and $\cos^2\theta\to (1-\mu^2-\mu'{}^2+3\mu^2\mu'{}^2)/2$. Integrating over $\mu'$, we obtain $$\begin{aligned}
&{\sum_{N}}\frac{G_F^2n_N}{\pi m_N}
\Big\{-m_N v_N p^3\frac{2}{3}[({F}_1^N)^2 + 5({F}_A^N)^2] \mu \frac{\partial f^{(0)}}{\partial p}
\nonumber\\
&+ \frac{2}{3}[({F}_1^N)^2 + 5({F}_A^N)^2] \frac1{p^2}
\frac{\partial}{\partial p} \left[ p^6 \left(k_BT_N \frac{\partial f^{(0)}}{\partial p} + f^{(0)}(p)(1-f^{(0)}(p)) \right) \right]
\nonumber\\
&+ p^4v_N\frac{2}{3}[({F}_1^N)^2 + 7({F}_A^N)^2 \pm 8{F}_A^N({F}_1^N+{F}_2^N)] \mu \frac{\partial f^{(0)}}{\partial p}
\nonumber\\
&+ p^3 m_N v_N^2
\left( \frac{2}{3}[({F}_1^N)^2 + 5({F}_A^N)^2](1+3\mu^2)
\frac{\partial f^{(0)}}{\partial p}
\right.\nonumber\\
&\qquad\qquad\qquad +\left.\frac{p}{6}[({F}_1^N)^2(1+\mu^2) + ({F}_A^N)^2(3+11\mu^2) ]
\frac{\partial^2 f^{(0)}}{\partial p^2} \right)\Big\}\,.\end{aligned}$$
Finally, we average over the directions of the bulk velocities of the nucleons by performing $\int_{-1}^1 d\mu/2$, and obtain $$\begin{aligned}
\frac{df}{dt}=
\sum_N n_N\sigma_N\left(1-\frac13\delta_N\right)
\left\{
\frac{v_N^2}3\frac1{p^4}
\frac{\partial}{\partial p} \left[ p^6 \frac{\partial f^{(0)}}{\partial p}\right]
+
\frac1{m_Np^4}
\frac{\partial}{\partial p} \left[ p^6 \left(k_BT_N\frac{\partial f^{(0)}}{\partial p}
+f^{(0)}(1-f^{(0)})\right)
\right]
\right\}\,,
\label{eq:final}\end{aligned}$$ where $\sigma_N$ is the total scattering cross section given by $$\begin{aligned}
\sigma_N=\frac{G_F^2}{\pi}p^2\left[({F}_1^N)^2 + 3({F}_A^N)^2\right]=\left(\frac{p}{m_\e}\right)^2\times
\begin{cases}
5.35\times 10^{-45}~{\rm cm^2}~~(\mathrm{for}~N=\p)\\
6.45\times 10^{-45}~{\rm cm^2}~~(\mathrm{for}~N=\n)
\end{cases}
\label{eq:sigma_N}\end{aligned}$$ and the symbol [@burr06b] $$\begin{aligned}
\delta_N \equiv \frac{({F}_1^N)^2 - ({F}_A^N)^2}{({F}_1^N)^2 + 3({F}_A^N)^2}=
\begin{cases}
-0.33~~(\mathrm{for}~N=\p)\\
-0.11~~(\mathrm{for}~N=\n)
\end{cases}
\label{eq:delta_N}\end{aligned}$$ captures angular dependence of scattering in the rest frame of nucleons, $d\sigma_N/d\Omega=(\sigma_N/4\pi)(1+\delta_N\cos\theta)$ —see the first line of Eq. (\[eq:expanded\]).
Eq. (\[eq:final\]) is the main result of this paper. Comparing this to the result for photon-electron scattering (see Eq. (15) of [@hu94]), $$\begin{aligned}
\frac{df}{dt}=n_\e\sigma_T
\left\{
\frac{v^2_\e}3\frac1{p^2}
\frac{\partial}{\partial p} \left[ p^4 \frac{\partial f^{(0)}}{\partial p}\right]
+
\frac1{m_\e p^2}
\frac{\partial}{\partial p} \left[ p^4 \left(k_BT_\e\frac{\partial f^{(0)}}{\partial p}
+f^{(0)}(1+f^{(0)})\right)
\right]
\right\}\,,
\label{eq:photon}\end{aligned}$$ we find two differences. First, the power of $p$ is different by two because the neutrino-nucleon cross section is proportional to $p^2$ whereas the Thomson scattering cross section $\sigma_T$ is independent of the photon momenta. Second, we have $1-f^{(0)}$ in the last term in Eq. (\[eq:final\]) instead of $1+f^{(0)}$ in Eq. (\[eq:photon\]), because of Fermi statistics.
Solutions {#sec:solution}
=========
Chemical potential distortion
-----------------------------
From now on we shall assume that protons and neutrons share the same temperature $T$ and bulk velocity $V$. We shall use the approximation that the masses of protons and neutrons are the same and denote them as $m$, i.e., $m_N\to m$, and drop the superscript $(0)$ on $f$.
Let us define dimensionless variables $x\equiv p/(k_BT)$ and $$\begin{aligned}
\nonumber
dy&\equiv \sum_N\sigma_N(x=1) \left(1-\frac13\delta_N\right)n_N\frac{k_BT}{m}dt\\
&=\left<n \sigma(x=1)\right>\frac{k_BT}{m_N}dt,\end{aligned}$$ where $$\begin{aligned}
\left<n \sigma\right>=\sum_Nn_N\sigma_N\left(1-\frac{1}{3}\delta_N\right)
=n_\n\left(1-\frac{1}{3}\delta_\n\right)\sigma_\n+n_\p\left(1-\frac{1}{3}\delta_\p\right)\sigma_\p.
\label{eq:nsigma}\end{aligned}$$ We then write Eq. (\[eq:final\]) as $$\frac{df}{dy}=\frac1{x^2}\frac{\partial}{\partial x}\left\{x^6\left[f-f^2+(1+\delta_v)f'\right] \right\}\,,
\label{eq:kompaneets}$$ where the prime denotes derivative with respect to $x$, and $\delta_v\equiv mV^2/(3k_BT)$ is the ratio of nucleon’s bulk kinetic energy to thermal energy.
An equilibrium solution, $df/dy=0$, is given by $$f(p)=\frac{1}{e^{(p-\mu_\nu)/[k_BT(1+\delta_v)]}+1}\,,
\label{eq:spectrum}$$ where $\mu_\nu$ is a chemical potential. The bulk velocity effectively enlarges the temperature by a factor of $1+\delta_v$.
As the number of neutrinos is conserved in our setup, we can obtain the emergent spectrum after stochastic scattering by bulk fluid motion as follows. For an initially thermal spectrum (i.e. $\mu_\nu=0$), we get the same number density of neutrinos if we have $\mu_\nu=-0.341$, $-4.34$, and $-80.3k_BT$ for $\delta_v=0.1$, $1$, and $10$, respectively. Therefore, scattering distorts the neutrino thermal spectrum by giving a non-zero chemical potential, which is well known as “$\mu$-distortion” in the cosmic microwave background research [@zeld69].
With the spectrum of Eq. (\[eq:spectrum\]), we can calculate the neutrino-annihilation rate which is one of the important heating processes in supernova explosions. The energy deposition rate via neutrino annihilation ($\nu+\bar\nu\to \mathrm{e}^++\mathrm{e}^-$) is given by [@good87; @seti06] $$\dot E_{\nu\bar\nu}=\mathcal{C}F_{3,\nu}F_{3,\bar\nu}
\left(\frac{{\big<{\epsilon_\nu^2}\big>}{\big<{\epsilon_{\bar\nu}}\big>}+{\big<{\epsilon_{\bar\nu}^2}\big>}{\big<{\epsilon_{\nu}}\big>}}{{\big<{\epsilon_{\nu}}\big>}{\big<{\epsilon_{\bar\nu}}\big>}}\right)\,,$$ where $F_{i,\nu}\equiv\int p^idpf$, ${\big<{\epsilon_\nu}\big>}=F_{3,\nu}/F_{2,\nu}$ and ${\big<{\epsilon_\nu^2}\big>}=F_{4,\nu}/F_{2,\nu}$. The subscript $\bar\nu$ indicates the corresponding quantities for anti-neutrinos. The overall factor $\mathcal{C}$ includes the weak interaction coefficients and information on the angular distribution of the neutrinos. To calculate this factor we need to determine the geometry of the neutrino-emitting source. Since this factor is expected not to change significantly by including the neutrino acceleration process, we concentrate on the effect of the spectral change here. For simplicity, we assume that the spectra of $\nu$ and $\bar\nu$ are identical. Then, we obtain $$\dot E_{\nu\bar\nu}\propto \frac{F_{3,\nu}^2{\big<{\epsilon_\nu^2}\big>}}{{\big<{\epsilon_\nu}\big>}}
\propto {\big<{\epsilon_\nu}\big>}{\big<{\epsilon_\nu^2}\big>}\,.$$ Here we use $F_{3,\nu}\propto {\big<{\epsilon_\nu}\big>}$, as $F_{2,\nu}$ does not change by scattering processes alone. Roughly speaking, ${\big<{\epsilon_\nu}\big>}=(1+\delta_v){\big<{\epsilon_\nu}\big>}_\mathrm{thermal}$ and ${\big<{\epsilon_\nu^2}\big>}=(1+\delta_v)^2{\big<{\epsilon_\nu^2}\big>}_\mathrm{thermal}$, where ${\big<{\epsilon_\nu}\big>}_\mathrm{thermal}$ and ${\big<{\epsilon_\nu^2}\big>}_\mathrm{thermal}$ are mean energy and mean square energy based on a purely thermal distribution function. Therefore, we find $$\dot E_{\nu\bar\nu}\propto (1+\delta_v)^3\,.$$
Including opacity {#sec:extended_boltzmann}
-----------------
The calculation of the energy deposition rate given above includes only scattering; thus, it can describe only $\mu$- and $\tau$-type neutrinos. However, there are also terms related to emission and absorption. Absorption is important for electron-type neutrinos and leads to thermalization of the spectrum of $\nu_e$ and $\bar\nu_e$. In the following we include all relevant terms.
The revised transfer equation in which both ${\mathcal O}(v)$ and ${\mathcal
O}(v^2)$ terms are taken into account is written as follows (see also Refs. [@psal97; @tita97] for photon cases): $$\begin{aligned}
\frac{df}{dt}&=
\nabla\cdot\left(\frac{1}{3\kappa^\mathrm{t}}\nabla f\right)
+\frac{1}{3}\left(\nabla\cdot{\bm V}\right)p\frac{\partial f}{\partial p}+\kappa^\mathrm{a}(f^{\mathrm{eq}}-f)\nonumber\\
&+
\frac1{p^2}\frac{\partial}{\partial p}
\left\{\frac{\kappa^\mathrm{sc}}{m}p^4\left[f-f^2+\left(k_BT+\frac{mV^2}{3}\right)\frac{\partial f}{\partial p}\right]\right\}\,,
\label{eq:transfer}\end{aligned}$$ where $\kappa^\mathrm{a}$ and $\kappa^\mathrm{sc}$ are the opacities (inverse of the mean free path) for absorption and scattering, respectively, $\kappa^\mathrm{t}\equiv\kappa^\mathrm{a}+\kappa^\mathrm{sc}$, and $f^\mathrm{eq}$ is the neutrino distribution function in thermal equilibrium. When we omit the second line on the right-hand side of this equation, we get exactly the same as the equation for neutrino transfer based on diffusion approximation, which is solved in a number of numerical simulations; see Eq. (A27) of Ref. [@brue85] for a flux-limited diffusion approximation and Eqs. (5) and (6) of Ref. [@lieb09] for an isotropic diffusion source approximation. Both of them were originally derived from the Boltzmann equation including velocity-dependent terms up to ${\mathcal O}(v)$ and are commonly used in neutrino-radiation hydrodynamics simulations.
The effect of the second term proportional to $\nabla\cdot{\bm
V}$ on the right-hand side is analogous to the bulk Comptonization process in photon cases and the first-order Fermi acceleration for charged particles. This term also modifies the neutrino spectrum when there is a compressional flow, such as a shock or an accretion flow onto a black hole. In core-collapse supernovae, a neutron star is formed and compression is almost negligible in the optically thick region for neutrinos so that this first-order acceleration does not work at all [@suwa13a].
Numerical solution {#sec:numerical_solution}
------------------
Assuming one-zone (i.e. $\nabla f=0$) and incompressible bulk flow (i.e. $\nabla\cdot{\bm V}=0$), Eq. (\[eq:transfer\]) reads $$\begin{aligned}
\frac{df}{dt}=\kappa^\mathrm{a}(f^{\mathrm{eq}}-f)+
\frac{\partial}{\partial p}
\left\{\frac{\kappa^\mathrm{sc}}{m}p^4\left[f-f^2+\left(k_BT+\frac{mV^2}{3}\right)\frac{\partial f}{\partial p}\right]\right\}\,.
\label{eq:small_bulk_vel}\end{aligned}$$ The scattering opacity is given from Eqs. , and as $$\begin{aligned}
\nonumber
\kappa^\mathrm{sc}=\left<\sigma n\right>
&=\left[\sigma_\n\left(1-\frac{1}{3}\delta_\n\right)Y_\n+\sigma_\p\left(1-\frac{1}{3}\delta_\p\right)Y_\p\right]\frac{\rho}{m}\\
&= 4.0\times 10^{-10}~\mathrm{cm}^{-1}\left(\frac{p}{m_\mathrm{e}}\right)^2\left(\frac{\rho}{10^{11}~\mathrm{g~cm}^{-3}}\right)\left(1-0.11Y_\p\right),\end{aligned}$$ where $Y_\n=n_\n/n_\mathrm{b}$ and $Y_\p=n_\p/n_\mathrm{b}$ are the number fractions of free neutrons and protons with $n_\mathrm{b}$ being the number density of baryons, and $m=1.67\times 10^{-24}$ g. Since all nuclei are photodisintegrated into neutrons and protons, $Y_\n+Y_\p=1$ here. The absorption opacity ($\nu_\e+\n\to \p+\e^-$ for $\nu_\e$ and $\bar\nu_\e+\p\to \n+\e^+$ for $\bar\nu_\e$) is given by [@burr06b] $$\begin{aligned}
\kappa^\mathrm{a}&=\frac{3g_A^2+1}{4}\sigma_0\left(\frac{p}{m_\mathrm{e}}\right)^2\frac{\rho}{m}\times
\begin{cases}
Y_\n~~(\mathrm{for}~\nu_\mathrm{e})\\
Y_\p~~(\mathrm{for}~\bar\nu_\mathrm{e})
\end{cases}\nonumber\\
&= 7.5\times 10^{-10}~\mathrm{cm}^{-1}\left(\frac{p}{m_\mathrm{e}}\right)^2\left(\frac{\rho}{10^{11}~\mathrm{g~cm}^{-3}}\right)\left(\frac{Y}{0.5}\right)\,,
\label{eq:kappa^a}\end{aligned}$$ where $g_A=1.2723$ is the axial-vector coupling constant, and $\sigma_0=1.705\times 10^{-44}$ cm$^2$ is a reference neutrino cross section. The small corrections due to the mass difference between a neutron and a proton, and to weak magnetism and recoil are neglected here. For characteristic values $Y\equiv
Y_\n=Y_\p=0.5$ is used in Eq. (\[eq:kappa\^a\]).
For constant temperature and chemical potential, $V^2=0$, and $f^\mathrm{eq}_\nu=(e^{(\epsilon_\nu-\mu_\nu)/k_BT}+1)^{-1}$, we have a steady-state solution ($\partial f/\partial t=0$) of Eq. (\[eq:small\_bulk\_vel\]) given by $f=f^\mathrm{eq}_\nu$. This solution is the same as the steady-state solution of the equation without the second term on the right-hand side in Eq. (\[eq:small\_bulk\_vel\]).
Writing the Boltzmann equation in a dimensionless form, we have $$\begin{aligned}
\frac{\partial f}{\partial y}=
\frac{1}{x^2}\frac{\partial}{\partial x}
\left\{
x^6
\left[
f-f^2+(1+\delta_v)f'
\right]
\right\}
+\Theta_T^{-1}(f^\mathrm{eq}-f)\,,
\label{eq:nondimensional}\end{aligned}$$ where $\Theta_T\equiv
(\kappa^\mathrm{sc}/\kappa^\mathrm{a})
(k_BT/m)$.
![[*(Top)*]{} Numerical solutions for $k_BT=6$ MeV, $\delta_v=1$, and $Y_\p=0.1$ (red lines). The initial spectrum and an analytic solution without absorption are shown by the blue- and green-dashed lines, respectively. For numerical solutions, two different values of $y$ are used ($y=1$ and $10^{-5}$ for the thick and thin lines, respectively). [ *(Bottom)*]{} Difference between numerical solutions at $y=1$ and the initial spectrum. The low-energy part ($x\lesssim 7$) is reduced, while the high-energy part is increased due to up-scattering.[]{data-label="fig:spec"}](temp_spec.eps){width="100.00000%"}
Figure \[fig:spec\] shows numerical solutions of Eq. (\[eq:nondimensional\]) as red solid lines. For this calculation we use $k_BT=6$ MeV, $\delta_v=1$, and $Y_\p=0.1$. Since we use $\kappa^\mathrm{a}\propto Y_\p$, the solutions shown in this figure are valid for $\bar\nu_e$. The initial condition at $y=0$ is given by $(e^x+1)^{-1}$, which is shown by the blue-dashed line in this figure. The green-dashed line is an analytic solution of Eq. (\[eq:spectrum\]), which is realized when the absorption is not taken into account. We find that, for higher energy ($x\gtrsim 10$), the numerical solutions deviate from the thermal equilibrium (blue-dashed line), and the spectra become slightly non-thermal toward the analytic solution (green-dashed line). This is because the ratio of the two terms on the right-hand side of Eq. (\[eq:small\_bulk\_vel\]) is $\propto p/m$, such that the absorption term dominates in the low-energy regime. We find that the numerical solutions are very similar for $y\gtrsim 0.001$.
Implications for neutrino heating rate
--------------------------------------
Since the neutrino interaction rate is a strong function of neutrino energy, the neutrino heating rate is affected when the neutrino spectrum changes. The neutrino heating rate is given as [@jank01] $$Q_\nu^+\propto E_\nu^2L_\nu\,,
\label{eq:qnu}$$ where $$\begin{aligned}
&E_\nu^2\propto \frac{\int^\infty_0 dp~p^5 f}{\int^\infty_0 dp~p^3 f}\,,
\label{eq:ebra}\\
&L_\nu\propto \int^\infty_0 dp~p^3 f\,.
\label{eq:Lnu}\end{aligned}$$ With the numerical solutions for the neutrino spectrum (Figure \[fig:spec\]), we find the following results: $$\begin{aligned}
\frac{\int_0^\infty f(x)x^5dx}{\int_0^\infty f^\mathrm{eq}(x)x^5dx}
&=1+0.017~\delta_v \left(\frac{k_BT}{1~\mathrm{MeV}}\right)=1+0.053\left(\frac{V}{0.1}\right)^2\,,
\label{eq:f5}\\
\frac{\int_0^\infty f(x)x^3dx}{\int_0^\infty f^\mathrm{eq}(x)x^3dx}
&=1-0.010~\delta_v \left(\frac{k_BT}{1~\mathrm{MeV}}\right)=1-0.031\left(\frac{V}{0.1}\right)^2\,.
\label{eq:f3}\end{aligned}$$ They are valid for $\delta_v\lesssim 1$. For larger $\delta_v$, non-linear terms in $\delta_v$ contribute. Combining them gives an enhancement in the square mean energy over its thermal value as $$\frac{E_\nu^2}{E_{\nu,\mathrm{thermal}}^2}=1+0.086\left(\frac{V}{0.1}\right)^2\,.$$ Thus, the neutrino heating rate $Q_\nu^+$ is amplified inside the gain region, where the neutrino heating overwhelms neutrino cooling between the shock and the neutrinospheres —see Eqs. and . The neutrino cooling rate is reduced inside the protoneutron star —see Eqs. and . Both of them make shock revival of a supernova easier. The actual velocity would be $V\sim$ 0.02–0.08 for the turbulent motion. To give an accurate value of the correction, we need hydrodynamical simulations.
Summary and discussion {#sec:summary}
======================
We have derived a Kompaneets-like equation for neutrinos by expanding the collision integral of the Boltzmann equation up to the second order in energy transfer, or ${\mathcal O}(v^2)$, including thermal and bulk velocities of nucleons. We also included absorption and emission of neutrinos, to arrive at the full neutrino transport equation given in Eq. (\[eq:transfer\]). The dimensionless form of the equation suitable for numerical calculations is given in Eq. (\[eq:nondimensional\]), and the numerical solutions are presented in Figure \[fig:spec\].
We find that the distortion of the neutrino spectrum due to interaction with nucleons leads to a larger neutrino heating rate in the gain region and a smaller neutrino cooling rate in the protoneutron star, which provides a better condition for supernova explosions than solutions without the effects we find in this paper.
The formulation given in this paper is a natural extension of the transfer equation that is solved in some neutrino-radiation hydrodynamics simulations for core-collapse supernovae, in which velocity-dependent terms in the collision integral are included only up to ${\mathcal O}(v)$. However, this effect would be implicitly included in numerical simulations if the transfer equation is solved in a comoving frame and the Lorentz transformation is properly performed between the comoving and laboratory frames. Note that the current equation is derived in the non-relativistic limit, i.e. $k_BT\ll m$ and $V\ll 1$. Relativistic corrections are known to amplify the spectral distortion of the SZ effect (e.g. see Ref. [@noza00]), which could even enhance the neutrino heating rates.
The numerical solutions presented in Section \[sec:numerical\_solution\] give the emergent spectrum of the electron-type neutrino ($\nu_e$) and anti-neutrinos ($\bar\nu_e$), as we include the absorption and emission term (the last term on the right-hand side in Eq. \[eq:nondimensional\]), which is relevant only for charged-current reactions. For other heavier leptonic flavors, i.e. muon-type and tauon-type, this term does not appear in the transfer equation so that their spectrum would be like Eq. (\[eq:spectrum\]). The neutrino annihilation rate of these heavier leptonic neutrinos [@good87; @jank91] can also be amplified by a factor $(1+\delta_v)^3$, which would make supernova explosion easier as well.
Our finding motivates self-consistent hydrodynamics simulations including the spectral distortion of neutrinos, which will be needed to calculate the quantitative impact on supernova explosions. Our solutions would also be useful in testing the numerical code handling the capability of the terms of $\mathcal{O}(v^2)$.
Acknowledgments {#acknowledgments .unnumbered}
===============
E. K. thanks Jens Chluba for useful discussions. Y. S. and H. W. H. T. thank the Max Planck Institute for Astrophysics, where this work was initiated and completed, for their hospitality. Y. S. was supported in part by Japan Society for the Promotion of Science (JSPS) postdoctoral fellowships for research abroad. This study was supported in part by the Grant-in-Aid for Scientific Research (Nos. 15H05896, 16H00869, 16K17665, 17H02864, 18H04586, and 18H05437). H. W. H. T. was supported in part by the Advanced Leading Graduate Course for Photon Science (ALPS).
Neutrino-nucleon scattering {#sec:matrix}
===========================
Matrix element
--------------
In this appendix, we follow Refs. [@smith71; @leitner05] to write an expression for the spin-averaged matrix element of neutrino-nucleon scattering. It is given by $$\begin{aligned}
|M_N|^2 = 16 G_F^2 m_N^4
\left[A \mp \frac{s-u}{m_N^{2}} B+\frac{(s-u)^{2}}{m_N^{4}} C\right]\,,\end{aligned}$$ where $\mp$ takes $-$ for neutrinos and $+$ for anti-neutrinos, and $$\begin{aligned}
s&=(p+q)^2= m_N^2 + 2 p \cdot q \,,\quad
u=(p'-q)^2= m_N^2 - 2 p' \cdot q\,,\quad
\tau=\frac{Q^{2}}{4 m_N^{2}}\,,
\\
A&=4\tau\left[(1+\tau)\left({F}_{A}^{N}\right)^{2}-(1-\tau)\left({F}_{1}^{N}\right)^{2}+\tau(1-\tau)\left({F}_{2}^{N}\right)^{2}+4 \tau {F}_{1}^{N} {F}_{2}^{N}\right]\,,
\\
B&=4\tau {F}_{A}^{N}\left({F}_{1}^{N}+{F}_{2}^{N}\right)\,,\quad
C=\frac{1}{4}\left[\left({F}_{A}^{N}\right)^{2}+\left({F}_{1}^{N}\right)^{2}+\tau\left({F}_{2}^{N}\right)^{2}\right]\,,\end{aligned}$$ where all momenta are four-vectors, and ${F}_1^N$, ${F}_2^N$ and ${F}_A^N$ are the so-called neutral-current Dirac, Pauli and axial form factors, respectively, which depend on $Q^2\equiv -(p-p')^2$.
When nucleons are non-relativistic, we have $$\begin{aligned}
p \cdot q &= p \; m_N - {\bf p} \cdot {\bf q}
= p_{c.m.} m_N + p_{c.m.}^2\,,
\\
p' \cdot q &= p' m_N - {\bf p'} \cdot {\bf q}
= p_{c.m.} m_N + p_{c.m.}^2 \cos{\theta_{c.m.}}\,,
\\
p \cdot p' &= p p' (1 - {\hat p} \cdot {\hat p'} )
= p_{c.m.}^2 (1-\cos{\theta_{c.m.}})\,,\end{aligned}$$ where the momenta on the left-hand side are four-vectors, whereas those in the middle and on the right-hand side are three-vectors and their magnitudes. $p_{c.m.}$ and $\theta_{c.m.}$ are the energy and scattering angle in the center-of-mass frame. Eliminating $p_{c.m.}$ and $\theta_{c.m.}$, we obtain $$\begin{aligned}
p'= p \frac{m_N-{\hat p} \cdot {\bf q}}{m_N-{\hat p'} \cdot {\bf q} + p (1 - {\hat p} \cdot {\hat p'} )}\,. \end{aligned}$$ Using this relationship to expand the matrix element in powers of $p/m_N$ and ${\bf q}/m_N$ up to their linear order, we obtain $$\begin{aligned}
|M_N|^2 &= 16 G_F^2 m_N^2 p^2
\left\{
\left[ ({F}_1^N)^2 + 3({F}_A^N)^2 \right]
+ \left[ ({F}_1^N)^2 - ({F}_A^N)^2 \right] \cos \theta
\right. \nonumber\\
&\mp 4 \frac{p}{m_N} ({F}_1^N + {F}_2^N) {F}_A^N (1 - \cos \theta)
- \frac{p}{m_N} ({F}_1^N)^2 (1 - \cos^2 \theta)
- \frac{p}{m_N} ({F}_A^N)^2 (3 - 4 \cos \theta + \cos^2 \theta)
\nonumber\\
&\left. -4 [({F}_1^N)^2 + ({F}_A^N)^2] {\hat p} \cdot \frac{\bf q}{m_N}
+ [({F}_1^N)^2 - ({F}_A^N)^2] (1 - \cos \theta)
\frac{{\hat p} \cdot {\bf q} - {\hat p'} \cdot {\bf q}}{m_N}
\right\}\,,\end{aligned}$$ where ${\hat p} \cdot {\hat p'} \equiv \cos \theta$.
Total cross section
-------------------
In the center-of-mass frame, the differential cross section is given by $$\begin{aligned}
\left. \frac{d\sigma_N}{d\Omega} \right|_\text{c.m.}
=
\frac{1}{64\pi^2 s} |M_N|^{2}\,.\end{aligned}$$ Integrating this over angles, we obtain the total cross section in the limit of non-relativistic nucleons ($s\approx m_N^2$) as $$\begin{aligned}
\sigma_N
=
\frac{G_F^2}{\pi} p^2 \left[ ({F}_1^N)^2 + 3({F}_A^N)^2 \right]\,.\end{aligned}$$
For the energy scale we are interested in, the only relevant form factors are the ones with $Q^2=0$. Thus, $$\begin{aligned}
2 {F}_{1}^{\p}(0)&
=1-4 \sin ^{2} \theta_{W}\,,\quad
2 {F}_{1}^{\n}(0)=-1\,,\quad 2 {F}_{A}^{\p}(0)=g_A\,,\quad
2 {F}_{A}^{\n}(0)=-g_A\,,
\\
2 {F}_{2}^{\p}(0)
&=\left(1-4 \sin ^{2} \theta_{W}\right)\left( \mu_\p - 1 \right)-\mu_\n\,,\quad
2 {F}_{2}^{\n}(0)=\left(1-4 \sin ^{2} \theta_{W}\right)\mu_\n-\left(\mu_\p-1\right)\,.\end{aligned}$$ Here, $g_A=1.2723$ is the axial-vector coupling constant, $\sin^2\theta_W=0.23122$ is the weak angle, $G_F=1.1663787\times 10^{-5}~{\rm GeV}^{-2}$ is the Fermi coupling constant, and $\mu_\p$ and $\mu_\n$ are the magnetic moments of protons and neutrons, respectively. As ${F}_{2}^{N}$ does not contribute to the angle-averaged collision integral in Eq. , we do not need to evaluate them here. We thus find that $$\begin{aligned}
\sigma_\p
&=
\frac{G_F^2}{4\pi} p^2 \left[ (1-4\sin^2\theta_W)^2 + 3g_A^2 \right]=
5.35\times 10^{-45}~{\rm cm^2}\cdot \left( \frac{p}{m_\e} \right)^2\,,
\\
\sigma_\n
&=
\frac{G_F^2}{4\pi} p^2 \left[ 1 + 3g_A^2 \right]=
6.45\times 10^{-45}~{\rm cm^2} \cdot \left( \frac{p}{m_\e} \right)^2\,.\end{aligned}$$
Alternative derivation à la Rybicki and Lightman {#sec:rybicki}
================================================
In this appendix, we provide an alternative derivation of the main result of this paper, Eq. (\[eq:final\]), following the argument given in Section 7.6 of Ref. [@rybi79]. For ease of comparison, let us change our notation of momenta to follow closely that of Ref. [@rybi79]: $\bm{q}\to \bm{p}$, $\bm{q}'\to\bm{p}_1$, $\bm{p}\to \omega\hat{n}$, and $\bm{p}'\to \omega_1\hat{n}_1$.
The Boltzmann equation is $$\begin{aligned}
{\frac{d f(\omega,\hat n,t)}{cd t}}=
&
{\sum_{N}}\int d^3p\int d\Omega~\frac{d\sigma_N}{d\Omega}&\nonumber\\
&\times\{
g_N({\bm p_1})f(\omega_1,\hat n_1,t)[1-f(\omega,\hat n,t)]-g_N({\bm p})f(\omega,\hat n,t)[1-f(\omega_1,\hat n_1,t)]
\}\,,
\label{eq:boltzmann_eq}\end{aligned}$$ where $d\sigma_N/d\Omega=(\sigma_N/4\pi)(1+\delta_N\cos\theta)$ (with $N=\p$ and $\n$) is the differential cross section of scattering by free nucleons, $g_N({\bm p})$ the distribution function of nucleons with three-momenta ${\bm p}$, $\omega$ and $\omega_1$ the neutrino energies, $\hat n$ and $\hat n_1$ the unit vectors of the neutrino three-momenta, and $\cos\theta=\hat n\cdot\hat n_1$.
Thermal nucleons
----------------
Since the typical energy transfer is small compared to a nucleon’s kinetic energy, we define $$\Delta\equiv \frac{\omega_1-\omega}{k_BT}\ll 1\,,
\label{eq:delta}$$ and expand the right-hand side of Eq. (\[eq:boltzmann\_eq\]) up to the second order in $\Delta$.
In a frame in which the initial nucleon velocity is $\boldsymbol\beta={\bm p}/E$ with $E=|{\bm p}|^2/(2m_N)$, we have $$\begin{aligned}
\nonumber
\frac{\omega_1}{\omega}
&=\frac{1-\boldsymbol\beta\cdot \hat n}{1-\boldsymbol\beta\cdot \hat
n_1+\omega(1-\hat n\cdot\hat n_1)
{\sqrt{1-\beta^2}/m_N}}\\
&= 1+\frac{{\bm p}\cdot(\hat n_1-\hat n)}{m_N}+\mathcal{O}\left(\left(\frac{k_BT}{m_N}\right)^2\right)\,.
\label{eq:e1/e}\end{aligned}$$ Note that $\beta\equiv|\boldsymbol\beta|\approx
k_BT/m_N\ll 1$. Comparing Eqs. (\[eq:e1/e\]) and (\[eq:delta\]), we find $$\Delta=x(\hat n_1-\hat n)\cdot\frac{{\bm p}}{m_N},$$ where $x\equiv \omega/(k_BT)$.
Expanding the neutrino distribution function with $\omega_1$ up to second order in $\Delta$, we obtain $$f(\omega_1,\hat n_1)\approx f(\omega,\hat n)
+\Delta f'(\omega,\hat n) +\frac{\Delta^2}2 f''(\omega,\hat n)\,,$$ where the primes denote derivatives with respect to $\omega$. We also expand the distribution function of non-relativistic nucleons up to second order in $\Delta=(E-E_1)/(k_BT)$ as $g_N(E_1)\approx\left(1+\Delta+\Delta^2/2\right)g_N(E)$ with $g_N(E)=n_N(2\pi m_N k_BT)^{-3/2}e^{-E/k_BT}$. Here, $n_N$ is the number density of nucleons. Then Eq. (\[eq:boltzmann\_eq\]) becomes $$\begin{aligned}
{\frac{d f}{d t}}&=
\left[f'+f(1-f)\right]\sum_N\int d^3p\int d\Omega\frac{d\sigma_N}{d\Omega}g_N\Delta\nonumber\\
&+\left[\frac{1}{2}f''+\left(f'+\frac{f}{2}\right)(1-f)\right]\sum_N\int d^3p\int d\Omega\frac{d\sigma_N}{d\Omega}g_N\Delta^2\,.
\label{eq:dfdt}\end{aligned}$$ To check the calculation, note that a Fermi-Dirac distribution $f_\mathrm{FD}(x)=1/(e^{x+\eta}+1)$ is a steady-state solution to Eq. (\[eq:dfdt\]). Here, $\eta$ is an integration constant corresponding to a chemical potential term.
Let us evaluate the integral in the second term of Eq. (\[eq:dfdt\]). Writing $(\hat n_1-\hat n)\cdot{\bm p}=|\hat n_1-\hat n|p\cos\xi$ and $\hat
n\cdot\hat n_1=\cos\theta$, we have $$\Delta^2=2x^2\frac{p^2}{m_N^2}\cos^2\xi (1-\cos\theta)\,.$$ Using $d^3p=p^2 dp d(\cos\xi) d\varphi$ and $d\Omega=d(\cos\theta)
d\phi$, the integral evaluates to $$\begin{aligned}
\int d^3p\int d\Omega\frac{d\sigma_N}{d\Omega}g_N\Delta^2
&=n_Nx^2\frac{k_BT}{m_N}\sigma_N\left(2-\frac{2}{3}\delta_N\right)\,.\end{aligned}$$
The integral in the first term of Eq. (\[eq:dfdt\]) is more complicated. Therefore, we follow Ref. [@rybi79] and use the conservation law to calculate it. Since scattering does not change the number of neutrinos, $\int dx x^2\partial f/\partial t=0$. This equation is satisfied by the continuity equation, $${\frac{d f}{d t}}=
-\frac{1}{x^2}\frac{\partial}{\partial x}[x^2 j(x)]\,,
\label{eq:cons}$$ where $j$ is the flux of neutrinos in momentum space. Next, we write the unknown integral as $$\int d^3p\int d\Omega\frac{d\sigma_N}{d\Omega}g_N\Delta=n_N\sigma_N\frac{k_BT}{m_N}I\,.$$ The Boltzmann equation then becomes $$\begin{aligned}
{\frac{d f}{d t}}=&
\sum_N n_N\sigma_N\frac{k_BT}{m_N}\left[
\alpha_N x^2f''+2\alpha_N x^2f'(1-f)+f'I
+f(1-f)(\alpha_N x^2+I)
\right]\,,
\label{eq:dfdt2}\end{aligned}$$ where $\alpha_N\equiv 1-\delta_N/3$. Requiring that this equation be equal to Eq. (\[eq:cons\]), we write an [*ansatz*]{}, $$j(x)=g(x)\left[f'+h(f,x)\right]\,,$$ with two unknown functions $g$ and $h$, so that $j'$ has a term in $f''$ but no higher derivatives. We then determine the function $h$ from the equilibrium form, $f_\mathrm{FD}$, which has the following relation; $f_\mathrm{FD}'+f_\mathrm{FD}(1-f_\mathrm{FD})=0$. If $j$ vanishes in equilibrium then $h=f(1+f)$. We find that $$\begin{aligned}
-\frac{1}{x^2}\frac{\partial}{\partial x}(x^2j)=-
\left\{
gf''+f'
\left[
g'+g\left(1-2f+\frac{2}{x}\right)\right]
+f(1-f)\left(g'+\frac{2g}{x}\right)
\label{eq:deriv_j}
\right\}\,.\end{aligned}$$ Comparing the $f''$ terms in Eqs. (\[eq:dfdt2\]) and (\[eq:deriv\_j\]), we obtain $$\begin{aligned}
g=-\sum_Nn_N\sigma_N\frac{k_BT}{m_N}\alpha_N x^2\,.\end{aligned}$$ Comparing the $f(1-f)$ terms gives $$\begin{aligned}
\sum_Nn_N\sigma_N \frac{k_BT}{m_N}(\alpha_N x^2+I)
=-\left(g'+\frac{2g}{x}\right)
=\sum_Nn_N\sigma_N\frac{k_BT}{m_N}\alpha_N(4x+2x)\,,\end{aligned}$$ which yields $I=\alpha_N(6-x)x$. Note that this makes $f'$ terms consistent as well.
The final result is the following equation: $${\frac{d f}{d t}}=
\sum_N
\alpha_N\frac{k_BT}{m_N}\frac{1}{x^2}\frac{\partial}{\partial x}
\left[n_N\sigma_N(x)x^4(f-f^2+f')\right]\,.
\label{eq:kompaneets_dimensional}$$ By changing the variable from $t$ to $y$ via $dy\equiv
\sum_N\alpha_N\frac{k_BT}{m_N}n_N\sigma_N(x=1)dt$, we obtain $${\frac{d f}{dy}}=
\frac{1}{x^2}\frac{\partial}{\partial x}
\left[
x^6
\left(
f-f^2+f'
\right)
\right],
\label{eq:kompaneets-wo-v}$$ which agrees with Eq. (\[eq:kompaneets\]) without the bulk velocity effect $\delta_v$.
Including bulk motion {#sec:correction_bulk}
---------------------
In the above derivation, we have used the Maxwell distribution for nucleons. Next, let us include the bulk motion. The distribution function of nucleons including bulk motion is given by $g_N({\bm p})=n_N(2\pi m_N k_BT)^{-3/2}e^{-({\bm p}-m_N {\bm V})^2/2m_Nk_BT}$. Then, the bulk velocity of a fluid, ${\bm V}$, is given by $${\bm V}=\left<\frac{{\bm p}}{m_N}\right>\equiv \frac{\int d^3p \left(\displaystyle\frac{{\bm p}}{m_N}\right)g_N({\bm p})}{\int d^3p~ g_N({\bm p})}\,,$$ where $\langle\dots \rangle$ denotes the average over the momentum distribution of nucleons. The second moment of ${\bm p}/m_N$ is $$\left<\frac{p^2}{m_N^2}\right>={\big<{v^2}\big>}+V^2\,,$$ where ${\bm v}\equiv{\bm p}/m_N-{\bm V}$ is the velocity of a nucleon in the rest frame of a fluid element. Assuming that a nucleon’s momentum distribution in the rest frame of a fluid element is a Maxwell distribution, we have ${\big<{v^2}\big>}=3k_BT/m_N$. Including the bulk velocity in Eq. (\[eq:kompaneets-wo-v\]), we obtain $$\frac{d f}{d y}=
\frac{1}{x^2}\frac{\partial}{\partial x}
\left\{
x^6
\left[
f-f^2+(1+\delta_v)f'
\right]
\right\}\,,
\label{eq:kompaneets_v2}$$ which agrees with Eq. (\[eq:kompaneets\]).
[10]{}
A. S. Kompaneets, Sov. Phys. JETP, [**4**]{}(5), 730–737 (1957).
R. A. Sunyaev and [Ya]{}. B. Zeldovich, Comments Astrophys. Space Phys., [**4**]{}, 173–178 (1972).
P. A. R. Ade et al., Astron. Astrophys., [**594**]{}, A27 (2016), [[arXiv:1502.01598]{}]{}.
E. Fermi, Phys. Rev., [**75**]{}, 1169–1174 (1949).
G. B. [Rybicki]{} and A. P. [Lightman]{}, (1979).
W. Hu, D. Scott, and J. Silk, Phys. Rev., [**D49**]{}, 648–670 (1994), [[arXiv:astro-ph/9305038]{}]{}.
S. Dodelson and J. M. Jubas, Astrophys. J., [**439**]{}, 503–516 (1995), [[arXiv:astro-ph/9308019]{}]{}.
G. G. Raffelt, Astrophys. J., [**561**]{}, 890–914 (2001), [[arXiv:astro-ph/0105250]{}]{}.
H.-Th. Janka and W. Hillebrandt, Astron. Astrophys., [**224**]{}, 49–56 (1989).
I. Tamborra, B. M[ü]{}ller, L. H[ü]{}depohl, et al. Phys. Rev. [**D86**]{}, 125031 (2012).
A. Burrows, S. Reddy, and T. A. Thompson, Nucl. Phys., [**A777**]{}, 356–394 (2006), [[arXiv:astro-ph/0404432]{}]{}.
. B. Zeldovich and R. A. Sunyaev, Astrophys. Space Sci., [**4**]{}, 301–316 (1969).
J. [Goodman]{}, A. [Dar]{}, and S. [Nussinov]{}, Astrophys. J. Lett., [ **314**]{}, L7 (1987).
S. Setiawan, M. Ruffert, and H. Th. Janka, Astron. Astrophys., [ **458**]{}, 553 (2006), [[arXiv:astro-ph/0509300]{}]{}.
D. Psaltis and F. K. Lamb, Astrophys. J., [**488**]{}, 881 (1997), [[arXiv:astro-ph/9706017]{}]{}.
L. Titarchuk, A. Mastichiadis, and N. D. Kylafis, Astrophys. J., [**487**]{}, 834 (1997), [[arXiv:astro-ph/9702092]{}]{}.
S. W. Bruenn, Astrophys. J. Suppl., [**58**]{}, 771–841 (1985).
M. Liebendoerfer, S. C. Whitehouse, and T. Fischer, Astrophys. J., [**698**]{}, 1174–1190 (2009), [[arXiv:0711.2929]{}]{}.
Y. Suwa, Mon. Not. Roy. Astron. Soc., [**428**]{}, 2443 (2013), [[arXiv:1210.4162]{}]{}.
H.-Th. Janka, Astron. Astrophys., [**368**]{}, 527 (2000), [[arXiv:astro-ph/0008432]{}]{}.
S. Nozawa, N. Itoh, Y. Kawana, and Y. Kohyama, Astrophys. J., [**536**]{}, 31–35 (2000), [[arXiv:astro-ph/9912008]{}]{}.
H. Th. [Janka]{}, Astron. Astrophys., [**244**]{}, 378–382 (1991).
C. H. Llewellyn Smith, Phys. Rept., [**3**]{}, 261–379 (1972).
T.-J. Leitner, [Diplomarbeit, Justus-Liebig-Universität Giessen]{} (2005).
[^1]: Note that Ref. [@raff01] investigated how the recoil term affects the neutrino spectrum.
[^2]: More precisely, because of the energy dependencies of neutrino-matter interactions high-energy neutrinos decouple further outside at lower matter temperature so that the spectrum becomes pinched from the pure Fermi-Dirac spectrum due to the negative temperature gradient [@jank89; @tamb12].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Since the launch of the Fermi Gamma-ray Space Telescope on 11 June 2008, significant detections of high energy emission have been reported only in six Gamma-ray Bursts (GRBs) until now. In this work we show that the lack of detection of a GeV spectrum excess in almost all GRBs, though somewhat surprisingly, can be well understood within the standard internal shock model and several alternatives like the photosphere-internal shock (gradual magnetic dissipation) model and the magnetized internal shock model. The delay of the arrival of the $>100$ MeV photons from some Fermi bursts can be interpreted too. We then show that with the polarimetry of prompt emission these models may be distinguishable. In the magnetized internal shock model, high linear polarization level should be typical. In the standard internal shock model, high linear polarization level is still possible but much less frequent. In the photosphere-internal shock model, the linear polarization degree is expected to be roughly anti-correlated with the weight of the photosphere/thermal component, which may be a unique signature of this kind of model. We also briefly discuss the implication of the current Fermi GRB data on the detection prospect of the prompt PeV neutrinos. The influences of the intrinsic proton spectrum and the enhancement of the neutrino number at some specific energies, due to the cooling of pions (muons), are outlined.'
author:
- |
Yi-Zhong Fan$^{1,2}$ [^1]\
$^{1}$[Niels Bohr International Academy, Niels Bohr Institute, University of Copenhagen, Blegdamsvej 17, DK-2100 Copenhagen, Denmark]{}\
$^{2}$[Purple Mountain Observatory, Chinese Academy of Sciences, Nanjing 210008, China]{}
title: 'Interpretation and implication of the non-detection of GeV spectrum excess by Fermi $\gamma$-ray Space Telescope in most GRBs'
---
\[firstpage\]
gamma rays: bursts $-$ polarization $-$ radiation mechanism: nonthermal $-$ acceleration of particles $-$ elementary particles: neutrino
Introduction
============
Gamma-ray Bursts (GRBs) are the most extreme explosion discovered so far in the universe. With the discovery of the afterglows and then the measurement of the redshifts in 1997 [see @wijers00 for a review], the cosmological origin of GRBs has been firmly established. The modeling of the late ($t>10^{4}$ s) afterglow data favors the external forward shock model [see @piran99; @mesz02; @zm04 for reviews]. The radiation mechanisms employed in the modeling are synchrotron radiation and synchrotron self-Compton (SSC) scattering. In the early time the prolonged activity of the central engine plays an important role in producing afterglow emission too, particularly in X-ray band [e.g., @Katz98; @FW05; @Nousek06; @Zhang06]. The radiation mechanisms, remaining unclear, are assumed to be the same as those of the prompt soft gamma-ray emission. It is expected that in the Fermi era the origin of the prompt emission can be better understood. This is because the Large Area Telescope (LAT) and the Gamma-ray Burst Monitor (GBM) onboard Fermi satellite (http://fermi.gsfc.nasa.gov/) can measure the spectrum in a very wide energy band (from 8 keV to more than 300 GeV), with which some models may be well distinguished. For example, in the standard internal shock model the SSC radiation can give rise to a distinct GeV excess while in the magnetized outflow model no GeV excess is expected.
Motivated by the detection of some $>100$ MeV photons from quite a few GRBs by the Compton Gamma Ray Observatory satellite in 1991$-$2000 [e.g., @Hurley94; @fm95; @Gonz03], the prompt high energy emission has been extensively investigated and most calculations are within the framework of the standard internal shocks [e.g., @Pilla98; @pw04; @gz07; @Bosnjak09 cf. Giannios 2007]. The detection prospect for LAT seems very promising [see @fp08 for a recent review].
Since the launch of Fermi satellite on 11 June 2008, significant detections of prompt high energy emission from GRBs have been only reported in GRB 080825C [@Bouvier08], GRB 080916C [@Abdo09], GRB 081024B [@Omodei08], GRB 090323 [@Ohno09b], GRB 090328 [@Cutini09], possibly and GRB 090217 [@Ohno09] until now (5 May 2009). Though the detection of 3 prompt photons above $10$ GeV from GRB 080916C at redshift $z\sim 4.5$ [@Abdo09; @Greiner09] is amazing and may imply a very high initial Lorentz factor of the outflow $\Gamma_{\rm i}>1800$ and an efficient acceleration of particles to very high energy [@Zou09], the non-detection of a significant $>100$ MeV emission from most GRBs may be a better clue of the underlying physics. A delay in the onset of the $>100$ MeV emission with respect to the soft gamma-rays, as detected in GRB 080825C, GRB 080916C, GRB 081024B and GRB 090323, may be the other clue of the GRB physics [@Abdo09; @Ohno09b]. In this work we focus on these two novel observational features.
This work is structured as the following. In section 2 we discuss the constraint of the current Fermi GRB data on the standard internal shock model and several alternatives. In section 3 we look for distinguished signals in linear polarization of the prompt emission. In section 4 we briefly discuss the implication of the current Fermi results on the detection prospect of PeV neutrinos from GRBs. We summarize our results in section 5.
Interpreting the lack of GeV excess in most GRBs and the delay of the arrival of the $>100$ MeV photons
=======================================================================================================
In the leading [*internal shock model*]{} for the prompt emission [@npp92; @px94; @rm94; @dm98], the ultra-relativistic outflows are highly variable. The faster shells ejected at late times catch up with the slower ones ejected earlier and then power energetic forward/reverse shocks at a radius $R_{\rm int} \sim 5\times
10^{13}~(\Gamma_{\rm i}/300)^2 (\delta t_{\rm v}/{\rm 10~ms})$ cm, where $\Gamma_{\rm i}$ is the initial Lorentz factor of the outflow and $\delta t_{\rm v}$ is the intrinsic variability timescale. Part of the shock energy has been used to accelerate electrons and part has been given to the magnetic field. If the outflow is magnetized [@usov92; @dt92; @lb03; @gs05], we call the shocks generated in the collisions within the outflow the [*magnetized internal shocks*]{} [@Spruit01; @fan04]. The synchrotron radiation of the shock-accelerated electrons may peak in soft gamma-ray band and then account for the observed prompt emission. This model has been widely accepted for the following good reasons: (1) For an ultra-relativistic outflow moving with an initial Lorentz factor $\Gamma_{\rm i}$, the velocity (in units of $c$) is $\beta_{\rm
i}=\sqrt{1-1/\Gamma_{\rm i}^2}$. A small velocity dispersion $\delta
\beta_{\rm i}\sim \beta_{\rm i}/(2\Gamma_{\rm i}^2)$ will yield a very different Lorentz factor. As a result, internal shocks within the GRB outflow seem inevitable. (2) In the numerical simulation of the collapsar launching relativistic outflow, people found highly variable energy deposition in the polar regions in a timescale as short as $\sim 50$ ms [@mw99]. (3) This model can naturally account for the variability that is well detected in prompt gamma-ray emission [@kps97]. On the other hand, this model usually predicts a fast cooling spectrum $F_\nu \propto \nu^{-1/2}$ in the X-ray band. However, the data analysis finds a typical X-ray spectrum $F_\nu \propto \nu^{0}$ [@preece00; @band93]. Such a divergence between the model and the observation data is the so-called “the low energy spectral index crisis" [@gc99]. Another potential disadvantage of the internal shock model is its low efficiency of converting the kinetic energy of the outflow into prompt emission [e.g., @Kumar99]. Among the various solutions put forward, a plausible scenario is the [*photosphere-internal shock model*]{} [e.g., @rm05; @peer06; @tmr07]. The idea is that the thermal emission leaking from the photosphere is the dominant component of the prompt sub-MeV photons [@tho94; @mr00; @Ryde05; @ioka07]. The nonthermal high energy emission is likely the external inverse Compton (EIC) radiation of the internal shock-accelerated electrons cooled by the thermal photons from the photosphere [@rm05; @peer05; @peer06; @tmr07]. If the electrons are accelerated by gradual magnetic energy dissipation rather than by internal shocks, it is called the [*photosphere-gradual magnetic dissipation model*]{} [@gian07]. There is an increasing interest in these two kinds of models since: (1) In the spectrum analysis people did find evidences for a thermal emission component in dozens of bright GRBs [@Ryde05; @Ryde06; @RF09; @McGl09]. (2) The emission from the photosphere can naturally account for the temporal behaviors of the temperature and flux of these thermal radiation [@peer08]. (3) The overall spectrum of the prompt emission can be reasonably interpreted [e.g., @peer06; @gian07]. (4) The GRB efficiency can be much higher than that of the internal shock model [see @RF09 and the references therein].
In this section we test these four models with the current Fermi GRB data. It is somewhat surprisingly to see that none of these models have been ruled out.
Explaining the lack of GeV spectrum excess in most GRBs {#sec:GeV-excess}
-------------------------------------------------------
### The standard internal shock model
In this model, the outflow is baryonic and the thermal emission during the initial acceleration of the outflow is ignorable. The prompt emission is powered by energetic internal shocks. There are three basic assumptions. (i) $\epsilon_{\rm e}$, $\epsilon_{\rm B}$, $\epsilon_{\rm p}$ fractions of shock energy have been given to electrons, magnetic field and protons, respectively (note that $\epsilon_{\rm e}+\epsilon_{\rm B}+\epsilon_{\rm
p}=1$). (ii) The energy distribution of the shock-accelerated electrons is a single power-law. (iii) The prompt soft gamma-ray emission is attributed to the synchrotron radiation of the shocked electrons.
For internal shocks generating at $R_{\rm int}$, the typical random Lorentz factor of the electrons can be estimated as (see section 4.1.1 of @fp08 for details) $$\gamma'_{\rm e, m} \sim 760~(1+Y_{\rm ssc})^{1/4}L_{\rm
syn,52}^{-1/4} R_{\rm int,13}^{1/2}(1+z)^{1/2}(\varepsilon_{\rm
p}/300~{\rm keV})^{1/2}, \label{eq:gamma_em}$$ where $\varepsilon_{\rm p}=h\nu_{\rm m}$ is the observed peak energy of the synchrotron-radiation spectrum ($\nu F_\nu$), $h$ is the Planck’s constant, $L_{\rm syn}$ is the synchrotron-radiation luminosity of the internal shock emission, $Y_{\rm ssc} \sim {[-1+\sqrt{1+4\epsilon_{\rm
e}/(1+g^{2})\epsilon_{\rm B}}]}/2$ is the regular SSC parameter [@se01; @fp08; @Tsvi09], and $g\sim \gamma'_{\rm e,m}
\varepsilon_{\rm p}/\Gamma_{\rm i} m_{\rm e}c^2$. The SSC in the extreme Klein-Nishina regime ($g\gg 1$) is very inefficient. If that happens the non-detection of GeV spectrum excess in most Fermi GRBs can be naturally explained. With the typical parameters adopted in eq.(\[eq:gamma\_em\]) we have $g\sim 1$, for which the SSC may still be important (i.e., $Y_{\rm ssc}\geq 1$). In this work the convenience $Q_{\rm x}=Q/10^{\rm x}$ has been adopted in cgs units except for some specific notations.
The SSC radiation will peak at $$\begin{aligned}
h\nu_{\rm m, ssc} &\sim & {2{\gamma'}_{\rm e, m}^2 \varepsilon_{\rm
p} \over 1+2g} \sim 220 {\rm GeV}~(1+2g)^{-1}(1+Y_{\rm
ssc})^{1/2}\nonumber\\
&& L_{\rm syn,52}^{-1/2} R_{\rm int,13}(1+z)(\varepsilon_{\rm
p}/300~{\rm keV})^{2}. \label{eq:SSC_flare}\end{aligned}$$ Taking into account the energy loss of the electrons via inverse Compton scattering on prompt soft gamma-rays, the cooling Lorentz factor can be roughly estimated as $$\gamma'_{\rm e,c} \sim 0.03 L_{\rm syn,52}^{-1}R_{\rm int,
13}\Gamma_{2.5}^3.$$ In reality $\gamma'_{\rm e,c}$ is always larger than 1. The derived $\gamma'_{\rm e,c}<1$ just means that the electrons have lost almost all their energies and are sub-relativistic.
Prompt high energy photons above the cut-off frequency $\nu_{\rm
cut}$ will produce pairs by interacting with softer photons and will not escape from the fireball. Following @ls01 and @fp08, we have $$\begin{aligned}
h\nu_{\rm cut} &\approx & 2~{\rm GeV}~(1+z)^{-1}(\varepsilon_{\rm
p}/300~{\rm
keV})^{(2-p)/p}L_{\rm syn,52}^{-2/p}\nonumber\\
&& \delta t_{\rm
v,-2}^{\rm 2/p}\Gamma_{\rm i,2.5}^{(2p+8)/p}. \label{eq:nu_cut}\end{aligned}$$
The SSC radiation spectra can be approximated by $F_{\nu_{\rm ssc}}
\propto \nu^{-1/2}$ for $\nu_{\rm m}<\nu<\nu_{\rm m,ssc}$, and $F_{\nu_{\rm ssc}} \propto \nu^{-p/2}$ ($F_{\nu_{\rm ssc}} \propto
\nu^{-p}$) for $\nu>\nu_{\rm m,ssc}$ and $g \leq 1$ ($g\gg 1$). The energy ratio of the SSC radiation emitted below $\nu_{\rm cut}$ to the synchrotron radiation in the energy range $\nu_{\rm
m}<\nu<\max\{\nu_{\rm M},~\nu_{\rm cut}\}$ can be estimated as $$\begin{aligned}
{\cal R} &\sim & {Y_{\rm ssc} \nu_{\rm m}^{(2-p)/2} \int^{\nu_{\rm
cut}}_{\nu_{\rm m}}\nu^{-1/2}d\nu \over \nu_{\rm m,ssc}^{1/2}
\int^{\max\{\nu_{\rm
M},~\nu_{\rm cut}\}}_{\nu_{\rm m}}\nu^{-p/2}d\nu } \nonumber\\
&\approx & (p-2)(\nu_{\rm cut}/\nu_{\rm m,ssc})^{1/2}Y_{\rm ssc},\end{aligned}$$ where $\nu_{\rm M}\approx 30~\Gamma_{\rm i}(1+z)^{-1}~$ MeV is the maximal synchrotron radiation frequency of the shocked electrons [@cw96].
For $Y_{\rm ssc} \sim 1$, $p\sim 2.5$ (corresponding to the typical $\gamma-$ray spectrum $F_\nu \propto \nu^{-1.25}$ for $h\nu>\varepsilon_{\rm p}$) and $\nu_{\rm cut}\ll \nu_{\rm m,ssc}$, we have ${\cal R}\ll 1$. Therefore there is no GeV excess in the spectrum, in agreement with the data. In other words, the non-detection of a significant high energy component is due to a too large $h\nu_{\rm m,ssc}\sim$ TeV and a relative low $h\nu_{\rm
cut}\sim$ GeV (see Fig.\[fig:Schematic-1\] for a schematic plot). The other possibility is that $(1+g^{2})\varepsilon_{\rm
B}>4\epsilon_{\rm e}$, for which $Y_{\rm ssc}\sim {\cal O}(1)$, i.e., the SSC radiation is unimportant and can be ignored.
![A possible interpretation of the non-detection of the SSC component by Fermi satellite in the internal shock model. The high energy photons with an energy $>h\nu_{\rm cut}$ have been absorbed by the soft gamma-rays and energetic $e^{\pm}$ pairs are formed. The pairs will lost their energy through synchrotron radiation and/or inverse Compton scattering and then produce soft gamma-rays.[]{data-label="fig:Schematic-1"}](Fig1.eps)
### The photosphere$-$internal shock model
@tho94 proposed the first photosphere model for the prompt gamma-ray emission, in which the nonthermal X-ray and gamma-ray emission are attributed to the Compton upscattering of the thermal emission by the mildly relativistic Alfvén turbulence. The spectra of GRBs can be nicely reproduced [see also @peer06; @gian07]. It is, however, difficult to explain the energy dependence of the width of the gamma-ray pulse [@feni95; @norr96]. As shown in @tmr07, such a puzzle may be solved in the photosphere$-$ internal shocks model, in which the sub-MeV emission is dominated by the thermal emission of the fireball and the nonthermal tail is the EIC radiation of the electrons accelerated in internal shocks at a radius $R_{\rm int}\sim 10^{14}$ cm [e.g., @mr00; @rm05; @peer06]. In this model the shock accelerated-electrons take a power-law energy distribution $dn/d\gamma'_{\rm e} \propto {\gamma'}_{\rm e}^{-p}$ for $\gamma'_{\rm e}\geq \gamma'_{\rm e,m} \sim 1$. Such an initial distribution can not keep if the electrons cool down rapidly. As shown in @SPN98, in the presence of steady injection of electrons, the energy distribution can be approximated by $N_{\gamma_{\rm e}} \propto {\gamma'}_{\rm e}^{-(p+1)}$ for $\gamma'_{\rm e}>\max\{\gamma'_{\rm e,c}, ~\gamma'_{\rm e,m}\}$ and $N_{\gamma_{\rm e}} \propto {\gamma'}_{\rm e}^{-2}$ for $\gamma'_{\rm e,c}< \gamma'_{\rm e}<\gamma'_{\rm e,m}$. Under what conditions can shock acceleration generate a particle distribution with $\gamma'_{\rm e,m} \sim 1$, with a significant fraction of the outflow energy deposited in the nonthermal particles? There could be two ways. One is to assume that electron/positron pair creation in the outflow is so significant that the resulting pairs are much more than the electrons associated with the protons. As a result, the fraction of shock energy given to each electron/positron will be much smaller than that in the case of a pair-free outflow and a $\gamma'_{\rm e,m} \sim 1$ is achievable [@tmr07]. The other way is to assume that the particle heating is continuous. In the internal shock scenario, this could happen if an outflow shell consists of many sub-shells and the weak interaction between these sub-shells may be able to produce multiple shocks that can accelerate electrons continually with a very small $\gamma'_{\rm
e,m}$.
Since both $\gamma'_{\rm e,m}$ and $\gamma'_{\rm e,c}$ are $\sim 1$, the EIC spectrum should be $F_\nu \propto \nu^{-p/2}$ in the MeV-TeV energy range and there is no GeV excess for $p\sim 2.5$, consistent with the Fermi data.
### The magnetized internal shock model {#sec:mag}
If the unsteady GRB outflow carries a moderate/small fraction of magnetic field, the collision between the fast and slow parts will generate strong internal shocks and then produce energetic soft gamma-ray emission. As usual, the ratio between the magnetic energy density and the particle energy density is denoted as $\sigma$. In the ideal MHD limit, for $\sigma\gg 1$ just a very small fraction of the upstream energy can be converted into the downstream thermal energy. Therefore the GRB efficiency is very low[^2]. That’s why people concentrate on the internal shocks with a magnetization $\sigma\leq 1$ [@Spruit01; @fan04].
For the magnetized internal shocks with $\sigma\geq 0.1$, no significant high energy emission is expected since: (a) The SSC emission of the internal shocks is weak. Therefore there is no distinct GeV excess in the spectrum. (b) The synchrotron-radiation spectrum may be very soft, which renders the detection of GeV photons from GRBs more difficult. The reason is the following. For an isotropic diffusion and a relativistic shock, the electron energy distribution index can be estimated by [@KW05] $$p \sim (3\beta_{\rm u}-2\beta_{\rm u}\beta_{\rm d}^2+\beta_{\rm
d}^3)/(\beta_{\rm u}-\beta_{\rm d})-2. \label{eq:KW05}$$ However, in the presence of a large scale coherent magnetic field, the diffusion is highly anisotropic rather than isotropic [@Morl07a]. There are thus corrections to eq.(\[eq:KW05\]). But as long as the scattering is not very forward- or backward-peaked, these corrections are small [@Keshet06]. Taking into account the anisotropic correction, @Morl07 found a spectrum steep to $p\sim 3$ for $\sigma \sim 0.05$. In the ion-electron shock simulation, the acceleration of particles at the un-magnetized shock front is a lot more efficient than that with a $\sigma\sim 0.1$ [@Spit06]. Motivated by these two possible evidences, we adopt eq.(\[eq:KW05\]) to estimate the spectral slope of accelerated particles at the magnetized shock fronts. The validity of our approach can be tested by the advanced numerical simulations in the future.
In the case of $\sigma=0$, for an ultra-relativistic shock, $\beta_{\rm u} \rightarrow 1$ and $\beta_{\rm d} \rightarrow 1/3$, we have $p \rightarrow 2.22$. But for an ultra-relativistic magnetized shock, $\beta_{\rm u} \rightarrow 1$ and [e.g., @fan04] $$\beta_{\rm d} \approx {1\over 6}(1+\chi+\sqrt{1+14\chi+\chi^2}),
\label{eq:fan04}$$ where $\chi\equiv \sigma/(1+\sigma)$, please note that $\sigma$ is measured in the upstream. Note in this work we just discuss the ideal MHD limit, i.e., there is no magnetic energy dissipation at the shock front. For $0<\sigma\ll 1$, we have $p\sim
(4.22-2\sigma)/(1-2\sigma)-2>2.22$. For $\sigma\gg1$, we have $\beta_{\rm d} \rightarrow 1-1/2\sigma$ and $p \sim 4\sigma-1 \gg
2.22$. Correspondingly, the electron spectrum is very soft or even thermal-like. Adopting $\beta_{\rm u}\sim 1$ and substituting $\sigma \sim (1,~0.5,~0.1,~0.01)$ into eqs.(\[eq:KW05\]-\[eq:fan04\]), we have $p\sim
(6.6,~4.5,~2.7,~2.3)$. The (very) soft high energy spectra of some GRBs [e.g., @preece00 see also our Fig.\[fig:beta\] for the Fermi GRBs] may be interpreted in this way.
(0,180) (0,0)
If the weak prompt high energy emission of GRBs is indeed attributed to the magnetization of the outflow, one can expect that the smaller the $p$, the stronger the high energy emission. The ongoing analysis of the LAT data will test such a correlation.
### The photosphere$-$gradual magnetic dissipation model
@gian07 calculated the emission of a Poynting-flux-dominated GRB outflow with gradual magnetic energy dissipation (reconnection). In his scenario, the energy of the radiating electrons is determined by heating and cooling balance. The mildly relativistic electrons stay thermal throughout the dissipation region because of Coulomb collisions (Thomson thick part of the flow) and exchange of synchrotron photons (Thomson thin part). Rather similar to @tho94, the resulting spectrum naturally explains the observed sub-MeV break of the GRB emission and the spectral slopes. In this scenario, different from the magnetized internal shock model, the higher the initial $\sigma$, the harder the spectrum (see the Fig.2 of @gian07 for illustration). For an initial $\sigma\leq 40$ (corresponding to the baryon loading $L/\dot{M}c^2
\sim \sigma^{3/2} \leq 250$, where $\dot{M}$ is the mass loading rate), the resulting $>10$ MeV spectrum is very soft [see also @Drenkhahn02; @DS02], accounting for the failed detection of the GeV spectrum excess in most GRBs.
Interpreting the delay of the arrival of the $>100$ MeV photons {#sec:GeV-delay}
---------------------------------------------------------------
In both the collapsar and the compact star merger models for GRBs [see @piran99; @mesz02; @zm04 for reviews], the early outflow may suffer more serious baryon pollution and thus have a smaller $\Gamma_{\rm i}$ than the late ejecta [@ZhangW04]. This may explain the delay of the arrival of the $>100$ MeV emission since as long as $$\Gamma_{\rm i}\leq \Gamma_{\rm i,c}=180~(1+z)^{p \over
2p+8}({h\nu_{\rm cut} \over 100~{\rm MeV}})^{p\over
2p+8}({\varepsilon_{\rm p}\over 300~{\rm keV}})^{p-2 \over
2p+8}L_{\rm syn,52}^{1 \over 4+p}\delta t_{\rm v,-2}^{-{1\over
4+p}},$$ [*the $>100$ MeV photons can not escape from the emitting region freely and thus can not be detected.*]{}
In the photosphere-gradual magnetic dissipation model, a small $\Gamma_{\rm i}$ implies a low initial magnetization of the outflow, for which the high energy spectrum can be very soft [@DS02; @gian07]. In the magnetized internal shock model, the delay of the onset of the LAT observation indicates a larger magnetization of the early internal shocks if $\Gamma_{\rm i}>\Gamma_{\rm i,c}$.
In the collapsar scenario, before the breakout, the initial outflow is choked by the envelope material of the massive star (Zhang et al. 2004). The ultra-relativistic reverse shock may be able to smooth out the velocity/energy-density dispersion of the initial ejecta. So the internal shocks generated within the early/breakout outflow may be too weak to produce a significant non-thermal radiation component. The early emission is then dominated by the thermal component from the photosphere and may last a few seconds (provided that the chocked material has a width comparable to that of the envelope of the progenitor). The outflow launched after the breakout of the early ejecta can escape from the progenitor freely and the consequent internal shocks can be strong enough to produce energetic non-thermal radiation. The photosphere-internal shock model therefore might be able to naturally account for the delay in the onset of the LAT observation.
In summary, before and after the onset of the $>100$ MeV emission, it seems the physical properties of the outflow have changed.
The linear polarization signal of the prompt $\gamma-$ray emission
==================================================================
As discussed in section 2, the failed detection of the GeV spectrum excess in most GRBs can be understood in either the standard internal shock model or several alternatives. Therefore we need independent probes to distinguish between these scenarios. Our current purpose is to see whether the polarimetry in gamma-ray band can achieve such a goal. In this section, we firstly investigate the linear polarization property of the photosphere-internal shock model ([*the results may apply to the photosphere-gradual magnetic dissipation model as well*]{}) since it has not been reported by others yet[^3]. We then briefly discuss the linear polarization signals expected in the magnetized internal shock model and in the standard internal shock model since they have been extensively discussed in the literature [e.g., @Lyu03; @Granot03; @Waxman03; @NPW03; @FXW08; @Toma09].
Linear polarization signal of the photosphere-internal shock model {#sec:Lin-thermal}
------------------------------------------------------------------
![The coordinates in the comoving frame of the emitting region. The polarization vector $\hat{\Pi}'$ is along the ${\rm
X'}-$direction for a positive $P$ given in eq.(\[eq:AA81-b\]) otherwise it would be along $\hat{\rm X}'\times \hat{k}'_{\rm o}$. []{data-label="fig:coordinate-2"}](Fig3.eps)
In this work we assume an uniform outflow. At any point in the outflow there is a preferred direction, the radial direction, in which the fluid moves. We choose the ${\rm Z}'$-direction of the fluid local frame coordinate to be in that direction. The ${\rm
Y}'-$direction is chosen to be within the place containing the line of sight (i.e., the scattered photon $k'_{\rm o}$) and the ${\rm
Z}'-$axis (see Fig.\[fig:coordinate-2\]). In this frame, the incident photons ($k'$) are along the ${\rm Z}'-$direction.
As usual, we assume that the electrons are isotropic in the comoving frame of the emitting region[^4]. The incident photons are the thermal emission from the photosphere and are unpolarized. The energy of the incident and the scattered photons are $h\nu'_{\rm se}$ and $h\nu'$, respectively. As shown in @AA81 and @Bere82: (I) In the cloud of the isotropic electrons of energies $\gamma'_{\rm e}m_{\rm e}c^2$, the spectrum of photons, upscattered at the angle $\theta'$ relative to the direction of the seed photon beam, can be approximated by $$\begin{aligned}
{d N_\gamma \over dt d\nu'd\Omega'}& \approx &{3\sigma_T c \over
16\pi {\gamma'}_e^2}{n_{\nu'_{\rm se}} d\nu'_{\rm se} \over
\nu'_{\rm se}} {x\over 2 \beta'_{\rm e}Q} (4A_0^2-4A_0+B_0),
\label{eq:AA81-a}\end{aligned}$$ where $\beta'_{\rm e}=(1-1/{\gamma'}_{\rm e}^2)^{-1/2}$ and $$\begin{aligned}
A_0 &=& {1\over 2\delta {\gamma'}_{\rm e}^2 x}({1\over {\cal
T}_1}-{1\over {\cal
T}_2}), \nonumber\\
B_0 &=& {{\cal T}_2\over {\cal T}_1}+{{\cal T}_1\over {\cal
T}_2},\nonumber\\
{\cal T}_1 &=& (1-\cos \theta'){1+x+x\delta \cos \theta'-\delta x^2 \over Q^2},\nonumber\\
{\cal T}_2 &=& (1-\cos \theta'){1+x-x\delta \cos \theta'+\delta \over Q^2},\nonumber\\
Q &\equiv & \sqrt{1+x^{2}-2x\cos \theta'},\nonumber\\
\delta &\equiv & h \nu'_{\rm se}/({\gamma'}_{\rm e} m_{\rm e}
c^{2}),~~x\equiv \nu'/\nu'_{\rm se}.\end{aligned}$$ Please note that $x$ ranges from $x_{\rm m}$ to $x_{_{\rm M}}$ that are given by $$\begin{aligned}
x_{\rm m,_{\rm M}}=1+{\{{\cal A} \mp {\gamma'}_{\rm e}{\beta'}_{\rm
e}\sqrt{{\gamma'}_{\rm e}^2(1+\delta)^{2}(1-\cos
\theta')^{2}+\sin^{2}\theta'}\}\over 1+2\delta {\gamma'}_{\rm e}^2
(1-\cos \theta')+{\gamma'}_{\rm e}^2\delta^2(1-\cos \theta')^{2}},\end{aligned}$$ where ${\cal A}\equiv (1-\cos \theta'){\gamma'}_{\rm
e}^2[{\beta'}_{\rm e}^2-\delta -\delta^2 (1-\cos \theta')]$.
\(II) The polarization degree is $$P \approx {4(A_0-A_0^2) \over B_0-4A_0+4A_0^2}. \label{eq:AA81-b}$$
For the relativistic electrons (i.e., $\beta_{\rm e}\rightarrow 1$), eq.(\[eq:AA81-a\]) and eq.(\[eq:AA81-b\]) take the simplified forms $$\begin{aligned}
{d N_\gamma \over dt d\nu' d\Omega'} &\approx & {3\sigma_T c \over
16\pi {\gamma'}_e^2}{n_{\nu'_{\rm se}} d\nu'_{\rm se} \over
\nu'_{\rm se}} [1+ {\xi^2 \over 2(1-{\xi})}-{2\xi \over b_\theta
(1-\xi)}\nonumber\\
&+&{2\xi^2 \over b_\theta^2 (1-\xi)^2}], \label{eq:AA81-a1}\end{aligned}$$ $$P\approx {{2\xi \over b_\theta (1-\xi)}-{2\xi^2 \over b_\theta^2
(1-\xi)^2}\over 1+ {\xi^2 \over 2(1-{\xi})}-{2\xi \over b_\theta
(1-\xi)}+{2\xi^2 \over b_\theta^2 (1-\xi)^2}}, \label{eq:AA81-b1}$$ where $\xi \equiv h\nu'/({\gamma'}_{\rm e} m_{\rm e} c^2)$, $b_\theta=2(1-\cos \theta'){\gamma'}_{\rm e} h\nu'_{\rm se}/(m_{\rm
e} c^2)$, and $h\nu'_{\rm se}\ll h\nu' \leq {\gamma'}_{\rm e} m_{\rm
e} c^2 b_\theta /(1+b_\theta)$.
Since the emitting region is moving relativistically, the angle $\theta'$ (in Fig.\[fig:coordinate-2\]) corresponding to the line of sight (L.o.S) is given by $$\cos \theta'=(\cos \theta-\beta_{\rm i})/(1-\beta_{\rm i} \cos
\theta),$$ where $\theta$ is the angle between the line of sight and the emitting point (measured in the observer’s frame).
The azimuthal angle $\phi$ varying from $0$ to $2\pi$ is defined in Fig.\[fig:Cartoon\]. The polar angle $\theta$ ranges from $0$ to $\theta_{\rm v}+\theta_{\rm j}$. The angle between the vector $(\sin
\theta \cos \phi,~\sin \theta \sin \phi,~\cos \theta)$ and the central axis of the ejecta (C.A. in Fig.\[fig:Cartoon\]) is denoted as $\Theta$ and is given by $$\cos \Theta =-\sin \theta_{\rm v}\sin \theta \sin \phi+\cos
\theta_{\rm v}\cos \theta.$$ Please [*bear in mind*]{} that in the following radiation calculation, the flux is set to be zero if $\cos \Theta<\cos
\theta_{\rm j}$ because these points $(\theta,\phi)$ are outside of the cone of the ejecta.
![Sketch of the geometrical set-up used to compute the polarization signal. We take the L.o.S as the $z-$axis. The ${\rm
y}$-axis ($x-$axis) is within (perpendicular to) the plane containing the line of sight and central axis of the ejecta.[]{data-label="fig:Cartoon"}](Fig4.eps)
The EIC radiation flux in the observer frame is $$F_{\nu,\rm EIC}\propto
%={(1+z)\over 4 \pi D_L^2}
\int{ {\cal D}^3 h\nu'{d
N_\gamma \over dt d\nu'd\Omega'} N_{\gamma_{\rm e}}d{\gamma'}_{\rm
e} d\Omega},$$ where $\nu={\cal D}\nu'/(1+z)$, ${\cal D}=[\Gamma_{\rm
i}(1-\beta_{\rm i} \cos \theta)]^{-1}$ is the Doppler factor, and $\Omega$ is the solid angle satisfying $d\Omega=\sin \theta d\theta
d\phi$.
The polarized radiation flux is $$Q_{\nu,\rm EIC}\propto
%={(1+z)\over 4 \pi D_L^2}
\int{ {\cal D}^3 h\nu' P
\cos 2 \phi {d N_\gamma \over dt d\nu'd\Omega'} N_{\gamma_{\rm
e}}d{\gamma'}_{\rm e} d\Omega}.$$ The polarization degree of the EIC emission is $$P_{\rm \nu, EIC}=|Q_{\rm \nu, EIC}|/F_{\nu,\rm EIC}.$$
One can see that a non-zero net polarization is expected as long as $\theta_{\rm v}>0$. In the numerical example, we assume that the seed photons have a thermal spectrum (as suggested in the photosphere model) $$n_{\nu'_{\rm se}}\propto {(h\nu'_{\rm se})^{2} \over e^{h\nu'_{\rm
se}/kT'}-1},$$ where $kT'\approx kT/2\Gamma_{\rm i}$ is the temperature (measured in the rest frame of the emitting region) of the thermal emission. In the calculation we take $\Gamma_{\rm i} \sim 300$ and $kT \sim
100$ keV. The electron distribution is taken as $N_{\gamma_{\rm e}}
\propto {\gamma'}_{\rm e}^{-(1+p)} \propto {\gamma'}_{\rm e}^{-3.5}$ for ${\gamma'}_{\rm e}>2$, otherwise $N_{\gamma_{\rm e}} =0$. For comparison purpose we also consider the case of $N_{\gamma_{\rm e}}
\propto {\gamma'}_{\rm e}^{-2}$. The numerical results are presented in Fig.\[fig:Pol-1\]. One can see that the polarization degrees expected in these two representative cases are only slightly different. We also find that a moderate linear polarization level ($P_{\rm \nu,EIC}>10\%$) is achievable only for $\theta_{\rm
v}\gtrsim \theta_{\rm j}+1/(3\Gamma_{\rm i})$.
(0,200) (0,0)
(0,200) (0,0)
Currently the prompt emission consists of a thermal and a non-thermal components. The thermal component with the flux $F_{\rm
\nu,th}$ is expected to be unpolarized while the nonthermal EIC component may have a high linear polarization level. The observed polarization degree $$P_{\rm \nu,obs}={|Q_{\rm \nu,EIC}|\over F_{\rm \nu,EIC}+F_{\rm
\nu,th}}$$ should be strongly frequency-dependent. Roughly speaking, the linear polarization degree is anti-correlated with the weight of the thermal component. With an energy $\sim kT$, the emission is dominated by the thermal component and $P_{\rm kT,obs}$ is low. For $h\nu \gg kT$, the emission is dominated by the EIC component and $P_{\nu,\rm obs} \sim P_{\nu,\rm EIC}$, as illustrated in Fig.\[fig:Pol-2\]. This unique behavior[^5] can help us to distinguish it from other models.
The probability of detecting a moderate/high linear polarization degree (${\cal R}_{\rm pol}$), however, is not high (Please note that we do not take into account the weak events for which a reliable polarimetry is impossible). On the one hand, a high linear polarization level is achievable only for $\theta_{\rm v} \geq \theta_{\rm j}+1/3\Gamma_{\rm i}$. On the other hand, $\theta_{\rm v}-\theta_{\rm j}\lesssim 1/\Gamma_{\rm i}$ is needed otherwise the burst will be too weak to perform the gamma-ray polarimetry. For $\Gamma_{\rm i}\theta_{\rm j}\gg 1$, we have $${\cal R}_{\rm pol} \sim 4 /(3\Gamma_{\rm i} \theta_{\rm j}) \approx
5\% ~\Gamma_{\rm i,2.5}^{-1}\theta_{\rm j,-1}^{-1}. \label{eq:R_pol}$$
During the revision of this work, @McGl09 reported their analysis on the spectrum and the polarization properties of GRB 061122. They found out that the spectrum was better fitted by the superposition of a thermal and a non-thermal components and the photons in the “thermal" emission dominated energy range had a (much) lower polarization level than those in the higher energy band. These two characters are in agreement with the photosphere-internal shock model (or the photosphere-gradual magnetic dissipation model).
Linear polarization level expected in the standard internal shock model
-----------------------------------------------------------------------
In the standard internal shock model, the polarization of the synchrotron radiation depends on both the [*poorly known*]{} configuration of magnetic field generated in internal shocks and the geometry of the visible emitting region. Assuming a random magnetic field that remains planar in the plane of the shock, @Waxman03 and @NPW03 showed that a high linear polarization level can be obtained when a narrow jet is observed from the edge, like in the photosphere-internal shock model. For the jets on-axis ($\theta_{\rm v}\leq \theta_{\rm j}$), the linear polarization degree is low (see also Gruzinov 1999 and Toma et al. 2009). The detection probability of a moderate/high linear polarization degree can also be estimated by eq.(\[eq:R\_pol\]).
High linear polarization degree expected in the magnetized internal shock model
-------------------------------------------------------------------------------
For the magnetized internal shock model, the prompt soft $\gamma-$ray emission is attributed to the synchrotron radiation of the electrons in ordered magnetic field and a high linear polarization level is expected [@Lyu03; @Granot03]. The physical reason is the following. The magnetic fields from the central engine are likely frozen in the expanding shells. The toroidal magnetic field component decreases as $R^{-1}$, while the poloidal magnetic field component decreases as $R^{-2}$. At the radius of the “internal” energy dissipation (or the reverse shock emission), the frozen-in field is dominated by the toroidal component. For an ultra-relativistic outflow, due to the relativistic beaming effect, only the radiation from a very narrow cone (with the half-opening angle $\leq 1/\Gamma_{\rm i}$) around the line of sight can be detected. As long as the line of sight is off the symmetric axis of the toroidal magnetic field, the orientation of the viewed magnetic field is nearly the same within the field of view. The synchrotron emission from such an ordered magnetic field therefore has a preferred polarization orientation (i.e. perpendicular to the direction of the toroidal field and the line of sight). Consequently, the linear polarization of the synchrotron emission of each electrons could not be effectively averaged out and the net emission should be highly polarized. The detection prospect of a high linear polarization degree is very promising (i.e., ${\cal R}_{\rm pol}\sim 100\%$). The above argument applies to the reverse shock emission as well if the outflow is magnetized [@fan04a].\
[l|c|c|c|c|c]{} model & unique polarization property & ${\cal R}_{\rm pol}$\
standard internal shocks & & $\lesssim$ 10%\
photosphere-internal shocks$^\dagger$ & strongly frequency-dependent$^\ddagger$ & $\lesssim$ 10%\
magnetized internal shocks & & $\sim$ 100%\
\
\
\[tab:stat\]
As summarized in Tab.\[tab:stat\], for the magnetized internal shock model, a high linear polarization level should be typical, while for two other models a moderate/high linear polarization degree is still possible but much less frequent. So the statistical analysis of the GRB polarimetry results may be able to distinguish the magnetized internal shock model from the others [see aslo @Toma09]. In the photosphere-internal shock model the polarization degree is expected to be strongly frequency-dependent. Such a remarkable behavior, if detected, labels its physical origin.
Indeed there were some claims of the detection of high linear polarization degree in the soft $\gamma-$ray emission of GRB 021206 [@CB03 however see Rutledge & Fox 2004], GRB 930131, GRB 960924 [@Willis05], GRB 041219A [@McGl07; @Gotz09], and GRB 061122 [@McGl09]. These results are consistent with each other as the errors are very large. The situation is inconclusive and additional data is needed to test these results. Measuring polarization is of growing interest in high energy astronomy. New technologies are being invented, and several polarimeter projects are proposed, such as, in the gamma-ray band, there are the Advanced Compton Telescope Mission [@Boggs06], POET [@Hill08] and others [see @Toma09 for a summary]. So in the next decade reliable polarimetry of GRBs in gamma-ray band may be realized and we can impose tight constraint on the models. At present, the most reliable polarimetry is in UV/optical band [e.g., @Covi99; @Wijers99]. The optical polarimetry of the prompt emission and the reverse shock emission require a quick response of the telescope to the GRB alert. This is very challenging. Mundell et al. (2007) reported the optical polarization of the afterglow, at 203 sec after the initial burst of $\gamma-$rays from GRB 060418, using a ring polarimeter on the robotic Liverpool Telescope. Their robust ($90\%$ confidence level) upper limit on the percentage of polarization, less than $8\%$, coincides with the fireball deceleration time at the onset of the afterglow. Such a null detection is, however, not a surprise because for this particular burst the reverse shock emission is too weak to outshine the unpolarized forward shock emission [@JF07]. Quite recently, the robotic Liverpool Telescope performed the polarimetry measurement of the reverse shock emission of GRB 090102 (Kobayashi 2009, private communication). Following @fan02, @zhang03 and @KP03, it is straightforward to show that the reverse shock of GRB 090102 is magnetized. Consequently the optical flash is expected to be highly polarized. If confirmed in the ongoing data analysis, the magnetized outflow model for some GRBs will be favored.
Implication on the detection prospect of PeV neutrino emission
==============================================================
The site of the prompt $\gamma-$ray emission may be an ideal place accelerating protons to ultra-high energy [@Vietri95; @Waxman95]. These energetic protons can produce high-energy neutrinos via photomeson interaction, mainly through $\Delta$-resonance [@WB97]. The resulting neutrinos have a typical energy $E_{\rm
\nu,obs} \sim 5\times 10^{14}~{\rm eV}~\Gamma_{2.5}^2
[(1+z)^2\epsilon_{\gamma, \rm obs}/1~{\rm MeV}]^{-1}$. Significant detections are expected if GRBs are the main source of ultra-high energy cosmic rays [@WB97]. The underlying assumption is that the proton spectrum is not significantly softer than $dN/dE\propto
E^{-2}$. The current Fermi observations do not provide an observational evidence for such a flat particle spectrum. Below we discuss the detection prospect of PeV neutrinos implicated by the non-detection of high energy emission from most GRBs. In the magnetized outflow model, the acceleration of a significant part of protons to energies $\geq 10^{16}$ eV is highly questionable because of the resulting soft proton spectrum. In the photosphere-internal shock model, $\gamma'_{\rm e,m}\sim 1$ is needed [@tmr07]. The efficiency of accelerating protons to very high energy depends on the mechanism of the particle heating. For example, in the case of multiple internal shocks, each pair of internal shocks are expected to be very weak since $\Gamma_{\rm sh}-1 \sim 0.04 (\gamma'_{\rm
e,m}/5)(\epsilon_{\rm e}/0.2)^{-1}[3(p-2)/(p-1)]^{-1}$, where $\Gamma_{\rm sh}$ is the Lorentz factor representing the strength of the shock ($\Gamma_{\rm sh}\sim 1$ for Newtonian shocks). So the acceleration of the protons to ultra-high energy is less efficient than the standard internal shocks. This is particularly the case if the acceleration is mainly via second-order Fermi process, in which the acceleration of particles depends on the shock velocity sensitively.
Even in the standard internal shock model, the generation of $10^{20}$ eV protons and the production of PeV-EeV neutrinos may be not as promising as that claimed in most literature adopting a proton spectrum $dN/dE \propto E^{-2}$. Such a flat spectrum is predicted for [*Newtonian*]{} shocks and has been confirmed by the supernova remnant observations. But the typical MeV spectrum $F_\nu \propto \nu^{-1.25}$ [@preece00] of GRBs suggests $dN/dE\propto E^{-2.5}$, supposing the accelerated protons and electrons have the same spectrum. The non-detection of $>100$ MeV photon emission from most GRBs implies soft electron (possibly) and proton spectra. Given a proton spectrum $dN/dE\propto E^{-2.22}$ that is predicted in the relativistic shock acceleration model (the First-order Fermi mechanism), the kinetic energy of the ejecta needs to be $\sim 100$ times of the $\gamma-$ray radiation energy if GRBs are indeed the main source of the observed $\sim 10^{20}$ eV cosmic rays [see @Dermer08 and the reference therein]. In other words, the GRB efficiency should be as low as $\sim 1\%$. If correct, the number of protons at $E\sim 10^{16}$ eV will be quite a few times what assumed in @Guetta04. Correspondingly the PeV neutrino flux will be higher. However, the current afterglow modeling usually yields a typical GRB efficiency $\sim 10\%$ [e.g., @FP06; @Zhang07] or larger [e.g., @PK01; @Granot06]. Below we discuss a new possibility–The proton spectrum is curved. In the “low-energy" part, the spectrum may be steepened significantly by the leakage of the very high energy cosmic rays from the ejecta [see @Hillas05 and the references therein]. The “high-energy" spectrum part may be a lot flatter. For example, in the numerical simulation of cosmic rays accelerated in some supernova remnants, a spectrum $dN/dE
\propto E^{-1.7}$ at the high energy part is obtained [e.g., @Volk02; @Bere03]. If holding for GRBs as well and GRBs are the main source of the $10^{20}$ eV cosmic rays, the PeV neutrino spectrum will be harder than that predicted in @Guetta04. For instance, the neutron spectra $\varepsilon_{\rm \nu}^2dN/d\varepsilon_{\nu} \propto
\varepsilon_{\nu}^0$ and $\varepsilon_{\rm
\nu}^2dN/d\varepsilon_{\nu} \propto \varepsilon_{\nu}^{-2}$ in their Fig.3 will be hardened by a factor of $\varepsilon_{\nu}^{0.3}$. But the total flux may be just $\sim 10\%$ times that predicted in @Guetta04 because in this scenario the protons are not as many as that suggested in a flat spectrum $dN/dE\propto E^{-2}$ for $E\ll 10^{20}$ eV.
There is a process, ignored in some previous works, that can enhance the detection prospect a little bit. After the pions (muons) are generated, the high energy pions (muons) will lose energy via synchrotron radiation before decaying, thus reducing the energy of the decay neutrinos (e.g. Guetta et al. 2004). As a result, above $\varepsilon_{\nu_{\mu}}^{\rm c} \sim{10^{17} \over
1+z}\epsilon_{\rm e}^{1/2}\epsilon_{\rm
B}^{-1/2}L_{\gamma,52}^{-1/2}\Gamma_{\rm i, 2.5}^{4}\delta t_{\rm
v,-2}~{\rm eV}$ ($\varepsilon_{\bar{\nu}_{\mu},\nu_{\rm e}}^{\rm
c} \sim \varepsilon_{\nu_{\mu}}^{\rm c}/10$), the slope of the corresponding neutrino spectrum steepens by 2, where $L_{\gamma}$ is the luminosity of the $\gamma-$ray emission [@Guetta04]. However we do not suggest a smooth spectral transition around $\varepsilon_{\nu_{\mu}}^{\rm c}$ or $\varepsilon_{\bar{\nu}_{\mu},\nu_{\rm e}}^{\rm c}$ because the cooling of pions (muons) will cause a pile of particles at these energies (see also Murase & Nagataki 2006). A simple estimate suggests that the number of neutrinos in the energy range $(0.5,~1)\varepsilon_{\nu_{\mu}}^{\rm c}$ should be enhanced by a factor of $\sim 3$. A schematic plot of the muon neutrino spectrum in the standard internal shock model is shown in Fig.\[fig:Schematic\].
![The schematic plot of the PeV muon neutrino spectrum in the standard internal shock model.[]{data-label="fig:Schematic"}](Fig7.eps)
Conclusion
==========
In the pre-Fermi era, it is widely expected that significant GeV emission will be detected in a good fraction of bright GRBs if they are powered by un-magnetized internal shocks [e.g., @Pilla98; @pw04; @gz07; @fp08]. The detection of a distinct excess at GeV-TeV energies, the SSC radiation component of such shocks, will be a crucial evidence for the standard fireball model. The non-detection of the GeV spectrum excess in almost all Fermi bursts [@Abdo09] is a surprise but does not impose a tight constraint on the models. For example, in the standard internal shock model, the non-detection can be attributed to a too large $h\nu_{\rm m,ssc} \sim $ TeV and a relative low $h\nu_{\rm cut}
\sim$ GeV. Some alternatives, such as the photosphere-internal shock model, the magnetized internal shock model and the photosphere-gradual magnetic dissipation model, can be in agreement with the data, too (see Tab.\[tab:sum\] for a summary). We attribute the delay in the onset of LAT detection in quite a few Fermi bursts to the unfavorable condition for GeV emission of the early outflow (see section \[sec:GeV-delay\] for details).
model the physical reason
----------------------------- -------------------------------------------------------- -- -- --
standard internal shocks $h(\nu_{\rm m,ssc},~\nu_{\rm cut})\sim (\rm TeV,~GeV)$
photosphere-internal shocks very small $\gamma'_{\rm e,m}$ and $\gamma'_{\rm e,c}$
magnetized internal shocks soft synchrotron spectrum and weak SSC
photosphere-gradual
magnetic dissipation very small $\gamma'_{\rm e,m}$ and $\gamma'_{\rm e,c}$
: The physical reasons for the lack of GeV spectrum excess in most GRBs.
\[tab:sum\]
With the polarimetry of GRBs people can potentially distinguish between some prompt emission models (see Tab.\[tab:stat\] for a summary; see also Toma et al. 2009). We show in section \[sec:Lin-thermal\] that in the photosphere-internal shock model the linear polarization degree is roughly anti-correlated with the weight of the thermal component and will be highly frequency-dependent. Such a unique behavior, if detected, labels its physical origin. However, a moderate/high linear polarization level is expected only when the line of sight is outside of the cone of the ejecta (i.e., $\theta_{\rm v}>\theta_{\rm j}$). In addition, $\theta_{\rm
v}-\theta_{\rm j}\lesssim 1/\Gamma_{\rm i}$ is needed otherwise the burst will be too weak to perform the gamma-ray polarimetry. Consequently the detection prospect is not very promising.
In this work we have also briefly discussed the detection prospect of prompt PeV neutrinos from GRBs. The roles of the intrinsic spectrum of the protons and the cooling of pions (muons) have been outlined. The latter always increases the neutrino numbers at the energies $\varepsilon_{\nu_{\mu}}^{\rm c}$ or $\varepsilon_{\bar{\nu}_{\mu},\nu_{\rm e}}^{\rm c}$ by a factor of 3. The former, however, is uncertain. If the protons have an intrinsic spectrum $dN/dE\propto E^{-2.22}$ and have a total energy about tens times that emitted in gamma-rays, the detection prospect would be as good as, or even better than that presented in @Guetta04. If the proton spectrum traces that of the electrons, i.e., typically $dN/dE\propto E^{-2.5}$, the detection prospect would be discouraging.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous referee for very helpful suggestions/comments and Drs. X. F. Wu, K. Toma, and Y. C. Zou for communication. This work was supported in part by the Danish National Science Foundation, Chinese Academy of Sciences, National basic research program of China (grant 2009CB824800), and the National Natural Science Foundation of China (grant 10673034).
Abdo A. A., et al., 2009, Science, 323, 1688 Aharonian F. A., Atoyan A. M., 1981, Ap&SS, 79, 321 Band D. et al., 1993, ApJ, 413, 281 Berezhko E. G., Pühlhofer G., Vöolk H. J., 2003, A&A, 400, 971 Berestetskii V. B., Lifshitz E. M., Pitaevski L. P., 1982, Quantum Electrodynamics, Oxford OX3 OBW, England Boggs S. E., et al., 2006, New A. Rev., 50, 604 Bosnjak Z., Daigne F., Dubus G., 2009, A&A (arXiv:0811.2956) Bouvier A., 2008, GCN Circ. 8183 Cheng K. S., Wei D. M., 1996, MNRAS, 283, L133 Coburn W., Boggs S. E., 2003, Nature, 423, 415 Covino S. et al., 1999, A&A, 348, L1 Cutini S., Vasileiou V., Chiang J., 2009, GCN Circ. 9077 Daigne F., Mochkovitch R., 1998, MNRAS, 296, 275 Dermer C. D. 2008, arXiv:0711.2804 Drenkhahn G., 2002, A&A, 387, 714 Drenkhahn G., Spruit H. C., 2002, A&A, 391, 1141 Duncan R. C., Thompson C., 1992, ApJ, 392, L9 Fan Y. Z., Dai Z. G., Huang Y. F., Lu T., 2002, ChJAA, 2, 449 Fan Y. Z., Piran T., 2006, MNRAS, 369, 197 Fan Y. Z., Piran T., 2008, Front. Phys. Chin., 3, 306 Fan Y. Z., Wei D. M., 2005, MNRAS, 364, L42 Fan Y. Z., Wei D. M., Wang C. F., 2004a, A&A, 424, 477 Fan Y. Z., Wei D. M., Zhang B. 2004b, MNRAS, 354, 1031 Fan Y. Z., Xu D., Wei D. M., 2008, MNRAS, 387, 92 Fan Y. Z., Zhang B., Proga D., 2005, ApJ, 635, L129 Fenimore E. E., in ’t Zand J. J. M., Norris J. P., Bonnell J. T., Nemiroff R. J. 1995, ApJ, 448, L101 Fishman G. J., Meegan C. A., 1995, ARA&A, 33, 415 Ghisellini G., Celotti A. 1999, ApJ, 511, L93 Giannios D., 2007, A&A, 480, 305 Giannios D., Spruit H. C., 2005, A&A, 430, 1 Gotz D., Laurent P., Lebrun F., Daigne F., Bosnjak Z., 2009, ApJ, 695, L208 Granot J., 2003, ApJ, 596, L17 Granot J., Königl A., Piran T. 2006, MNRAS, 370, 1946 Greiner J., et al., 2009, A&A, submitted (arXiv:0902.0761) Gruzinov A., 1999, 525, L29 González M. M., Dingus B. L., Kaneko Y., Preeze R. D., Dermer C. D., Briggs M. S., 2003, Nature, 424, 749 Guetta, D., Hooper, D., Aliarez-Muniz, J., Halzen, F., Reuveni, E. 2004, Astropart. Phys., 20, 429 Gupta N., Zhang B., 2007, MNRAS, 380, 78 Hillas A. M., 2005, J. Phys. G: Nucl. Part. Phys., 31, R95 Hill J. E., et al., 2008, AIPC, 1065, 331 Hurley K., Dingus B. L., Mukherjee R., et al., 1994, Nature, 372, 652 Ioka K., Murase K., Toma K., Nagataki S., Nakamura T., 2007, ApJ, 670, L77 Jin Z. P., Fan Y. Z., 2007, MNRAS, 378, 1043 Katz J. I., Piran T., Sari R. 1998, Phys. Rev. Lett., 80, 1580 Keshet U., 2006, Phys. Rev. Lett. 97, 221104 Keshet U., Waxman E., 2005, Phys. Rev. Lett., 94, 111102 Klebesadel R. W., Strong I. B., and Olson R. A., 1973, ApJ, 182, L85 Kobayashi S., Piran T., Sari R., 1997, ApJ, 490, 92 Kumar P., 1999, ApJ, 523, L113 Kumar P., Panaitescu A., 2003, MNRAS, 346, 905 Lithwick Y., Sari R., 2001. ApJ, 555, 540 Lyutikov M., Blandford R., 2003 (arXiv:astro-ph/0312347) Lyutikov M., Pariev V. I., Blandford R. D. 2003, ApJ, 597, 998 MacFadyen A. I., Woosley S. E. 1999, ApJ, 524, 262 McGlynn S. et al., 2007, A&A, 466, 895 McGlynn S. et al., 2009, A&A, in press (arXiv0903.5218) Mészáros P., 2002, ARA&A, 40, 137 Mészáros P., Rees M. J., 2000, ApJ, 530, 292 Morlino G., Blasi P., Vietri M., 2007a, ApJ, 658, 1069 Morlino G., Blasi P., Vietri M., 2007b, ApJ, 662, 980 Mundell C. G., et al., 2007, Science, 315, 1822 Murase K., Nagataki S., 2006, Phys. Rev. D., 73, 063002 Nakar E., Piran T., Waxman E., 2003, JCAP, 0310, 005 Narayan R., Paczyński B., Piran T., 1992, ApJ, 395, L83 Norris J. P., Nemiroff R. J., Bonnell J. T., Scargle J. D., Kouveliotou C., Paciesas W. S., Meegan C. A., Fishman G. J. 1996, ApJ, 459, 393 Nousek, J. A., et al., 2006, ApJ, 642, 389 Ohno M., McEnery J., Pelassa V., 2009a, GCN Circ. 8903 Ohno M., Cutini S., McEnery J., Chiang J., Koerding E., 2009b, GCN Circ. 9021 Omodei N., 2008, GCN Circ. 8407 Paczyński B., Xu G. H., 1994, ApJ, 427, 708 Panaitescu A., Kumar P., 2002, ApJ, 571, 779 Pe’er A., 2008, ApJ, 682, 463 Pe’er A., Mészáros P., Rees M. J. 2005, ApJ, 635, 476 Pe’er A., Mészáros P., Rees M. J. 2006, ApJ, 642, 995 Pe’er A., Waxman E., 2004, ApJ, 613, 448 Pilla R. P., Loeb A., 1998, ApJ, 494, L167 Piran T., 1999, Phys. Rep., 314, 575 Piran T., Sari R., Zou Y. C., 2009, MNRAS, 393, 1107 Preece R. D., Briggs M. S., Mallozzi, R. S., Pendleton G. N., Paciesas W. S., Band D. L. 2000, ApJS, 126, 19 Rees M. J., Mészáros P., 1994, ApJ, 430, L93 Rees M. J., Mészáros P., 2005, ApJ, 628, 847 Rutledge R. E., Fox D. B., 2004, MNRAS, 350, 1288 Ryde F., 2005, ApJ, 625, L95 Ryde F., Björnsson C., Kaneko Y., Mészáros P., Preece R., Battelino M., 2006, ApJ, 652, 1400 Ryde F., Pe’er A., 2009, ApJ submitted (arXiv:0811.4135) Sari R., Esin A. A., 2001, ApJ, 548, 787 Sari R., Piran T., Narayan R., 1998, ApJ, 497, L17 Sironi L., Spitkovsky, A., 2009, ApJ, submitted (arXiv:0901.2578) Spruit H. C., Daigne F., Drenkhahn G., 2001, A&A, 369, 694 Thompson C., 1994, MNRAS, 270, 480 Thompson C., Mészáros P., Rees M. J., 2007, ApJ, 666, 1012 Toma K., et al., 2009, ApJ in press (arXiv:0812.2483) Usov V. V., 1992, Nature, 270, 480 Usov V. V., 1994, MNRAS, 267, 1035 van Paradijs J, Kouveliotou C., Wijers R. A. M. J., 2000, ARA&A, 38, 379 Vietri M., 1995, ApJ, 453, 883 Völk H. J. et al., 2002, A&A, 396, 649 Waxman E., 1995, Phys. Rev. Lett., 75, 386 Waxman E., 2003, Nature, 423, 388 Waxman E., Bahcall J. N., 1997, Phys. Rev. Lett., 78, 2292 Wijers R. A. M. J., et al., 1999, ApJ, 523, L33 Willis D. R. et al., 2005, A&A, 439, 245 Zhang B., et al. 2006, ApJ, 642, 354 Zhang B., et al., 2007, ApJ, 655, 989 Zhang B., Kobayashi S., Mészáros P. 2003, ApJ, 595, 950 Zhang B., Mészáros P., 2002, ApJ, 581, 1236 Zhang B., Mészáros P. 2004, Int. J. Mod. Phys. A, 19, 2385 Zhang W. Q., Woosley S. E., MacFadyen A. I., 2004, ApJ, 586, 356 Zou Y. C., Fan Y. Z., Piran T., 2009, MNRAS, in press (arXiv:0811.2997)
[^1]: Email: [email protected]
[^2]: For the Poynting-flux dominated outflow (i.e., $\sigma\gg 1$), @usov94 proposed that at the radius $r_{_{\rm MHD}}\sim
6\times10^{15}L_{52}^{1/2} \sigma_{2}^{-1} {t_{v,m}}_{-3}
\Gamma_{\rm i,2.5}^{-1}~{\rm cm}$, the MHD condition breaks down and large scale electromagnetic waves are generated, where $L$ is the total luminosity of the outflow and $t_{v,m}$ is the minimum variability timescale of the central engine [see also @zm02]. The particles are accelerated and the synchrotron radiation of the ultra-relativistic electrons peaks at $\nu_{\rm
m}\sim 5\times10^{19}\sigma_{2}^{3}(\epsilon_{\rm
e}/0.2)^2[3(p-2)/(p-1)]^{2} \Gamma_{2.5} {t_{v,m}}_{-3}~{\rm Hz}$ [@fzp05], provided that a significant part of the magnetic energy has been converted to the thermal energy of the particles. In this scenario, the SSC radiation is expected to be very weak because in the rest frame of the electrons having $\gamma'_{\rm e,m} \sim
10^{4}(\epsilon_{\rm e}/0.2)[3(p-2)/(p-1)]\sigma_{2}$ [@fzp05], the soft gamma-rays have an energy $\gamma'_{\rm
e,m}h\nu_{\rm m}/\Gamma_{\rm i}\gg m_{\rm e}c^2$, i.e., the SSC is in the extreme Klein-Nishina regime and is very inefficient. The non-detection of high energy emission from most GRBs is, of course, consistent with this model.
[^3]: In the final stage of preparing the manuscript, the author was informed by K. Toma and X. F. Wu that the POET (Polarimeters for Energetic Transients) group discussed the polarization property of the photosphere-internal shock model independently. The preliminary results were included in their white paper.
[^4]: In the current scenario, the electrons are heated by the internal shocks and an isotropic distribution in the rest frame of the emitting region may be a good approximation. Such a distribution, however, will be modified by the cooling of the electrons via EIC scattering on the thermal photons from the inner part of the outflow. Detailed numerical simulation is needed to see such a modification and then its influence on the inverse Compton radiation, which is beyond the scope of this work.
[^5]: For the synchrotron radiation of the electrons moving in a random magnetic field (the standard internal shock model) or in an ordered magnetic filed (e.g., the magnetized internal shock model), before and after the peak of the spectrum, the polarization degree changes because the polarization properties depend on the profile of the spectrum. However, such a dependence is weak, as shown in @Granot03.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'S.McBreen'
- 'F.Quilligan'
- 'B.McBreen'
- 'L.Hanlon'
- 'D.Watson'
bibliography:
- 'katmonic.bib'
- 'Dh291.bib'
date: 'Received / Accepted '
nocite:
- '[@macfad:1999]'
- '[@ruffjan:1999]'
title: 'Temporal properties of the short gamma-ray bursts'
---
Introduction
============
It has been recognised that GRBs may occur in two sub-classes based on spectral hardness and duration with $T_{90}$$>$$2\,$s and $T_{90}$$<$$2\,$s [@kmf:1993; @dbt:1995; @paciesas:2001]. The bimodal distribution can be fit by two Gaussian distributions in the logarithmic durations [@mhlm:1994]. There is significant evidence for a third subgroup as part of the long duration GRBs [@mfb:1998; @horv:1998] but this has been questioned because of a possible BATSE selection effect [@hhp:2000]. It has been suggested [@cmo:1999] that the small group of GRBs with T$_{90} <$ 0.1 s form an additional category. The short GRBs have a higher value of $\langle V/V_{\rm max}\rangle$ [@kc:1996], a much smaller value of the spectral lag [@nsb:2000], a pulse shape that depends on position in the burst [@gupta:2000] and a smaller space density than long GRBs [@schmidt:2001].
A variety of statistical methods have been applied to the temporal properties of GRBs with $T_{90}$$>$$2\,$s. It is important to compare the temporal profiles of the long and short GRBs to determine the similarities and differences between the two classes in an objective way. A detailed objective analysis has been performed on the temporal profiles of a large sample of 319 bright GRBs with $T_{90}$$>$$2\,$s [@quillig:2001; @hmq:1998]. The properties of the pulses in GRBs and the time intervals between them were found to be consistent with lognormal distributions. These results can be used as templates for comparison with a similar analysis of GRBs with $T_{90}$$<$$2\,$s.
The analysis method is presented in section 2 and the results in section 3. In section 4 the results are discussed and compared with the sample of long GRBs.
Property Median $\mu$ $\sigma$ Width ($\pm50\%$) KS ($\mathrm{\tau}_{\rm \sigma}$ $\geq$ 3,$\mathrm{\tau}_{\rm i}$ $\geq$ 20%) KS ($\mathrm{\tau}_{\rm \sigma}$ $\geq$ 8, $\mathrm{\tau}_{\rm i}$ $\geq$ 80%)
----------------- --------------------- --------- ---------- ------------------------ ------------------------------------------------------------------------------- -------------------------------------------------------------------------------- --
Rise Time 0.035 $-$3.31 0.94 0.012-0.11 0.48 0.16
Fall Time 0.056 $-$2.89 0.98 0.017-0.176 0.07 0.19
FWHM 0.045 $-$3.17 1.01 0.013-0.138 0.73 0.48
Area 1.5$\times 10^{5}$ 11.7 1.24 28-520($\times10^{3}$) 0.44 0.15
Pulse Amplitude 1.02$\times 10^{4}$ 9.22 0.83 3.8-27($\times10^{3}$) 0.90 0.72
Asymmetry Ratio 0.65 $-$0.42 0.91 0.19-2.22 0.78 0.30
$\Delta$T 0.095 $-$2.24 0.87 0.038-0.3 0.83 0.96
Data Preparation
================
The dataset was taken from the BATSE current catalog. The time tagged event data at 5ms resolution was used. The data from the four energy channels were combined into a single channel to maximise the signal to noise ratio. A subset of the current catalogue was selected with GRB durations less than two seconds ($T_{90}<2$s). These GRBs were ordered according to their peak flux in 256ms ($P_{\rm 256 ms}$) and the first 100 bursts without data gaps and with 5ms data available for the complete burst where selected . All GRBs in the sample had $P_{\rm 256
ms}$ $>$ 1.6 photons cm$^{-2}$s$^{-1}$.
Background subtraction
----------------------
The first step in the data preparation involved selecting the appropriate background for subtraction. The start and end times for each burst were identified . A background section was then identified which was usually after the main section of the GRB. If 5ms data was available for the background level estimation, it was used, otherwise a 64ms section was selected. The median value of the data points in the background section was subtracted from all other values in the active section of the GRB.
Denoising and pulse selection
-----------------------------
A median filter was used to denoise the GRBs. This denoises a signal by finding the median of each bin. The best denoised signal was found by varying the bin size and the value of $\sigma$. The wavelet method [@quillig:2001] was not used because the short durations of the GRBs produced a smaller number of data points than that required for our wavelet method.
The pulse selection method was described elsewhere [@quillig:2001]. The pulses selected had a threshold of 5 $\sigma$ ($\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5) and isolated from adjacent pulses by at least 50% ($\mathrm{\tau}_{\rm i}$ $\geq$ 50%). A value of $\mathrm{\tau}_{\rm i}$ $\geq$ 50% implies that the two minima on either side of the pulse maximum must be at or below half the maximum value. A total of 313 pulses were selected with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 and 181 of these had $\mathrm{\tau}_{\rm i}$ $\geq$ 50%.
Results
=======
Distributions of $t_{\rm r}$, $t_{\rm f}$, FWHM, pulse area, pulse amplitude and $\Delta$T
------------------------------------------------------------------------------------------
The distribution of the number of pulses with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 is given in Fig. 1. The median value of the number of pulses per GRB is 2.5 and is smaller than the value of 6 for long GRBs [@quillig:2001]. The distributions of rise time $t_{\rm r}$, fall time $t_{\rm f}$ and full width at half maximum (FWHM) for the isolated pulses are presented in Fig. 2 along with the lognormal fits to the data. The distributions of pulse amplitudes and areas are presented in Fig. 3 for the isolated pulses. The distribution of time intervals between pulses with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 is given in Fig. 4.
\[b\][Number of Pulses]{} \[b\][Number of GRBs]{} {width="0.85\columnwidth"} \[fig:grbs\_pulses\]
The median values and the parameters of the lognormal fits are presented in Table 1. The distribution of the pulse asymmetry ratios, $t_{\rm r}$/$t_{\rm f}$ , is not presented but the values are listed in Table 1. The pulse fall time is generally longer than the rise time. The widths or ranges of the lognormal distributions at the $\pm50\%$ are are also given in Table 1. The Kolmogorov-Smirnov (KS) test was used to indicate the effects of varying $\mathrm{\tau}_{\rm \sigma}$ and $\tau_{\rm i}$. The pulse property distributions at $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 and $\tau_{\rm i}$ $\geq$ 50% were compared with the distributions selected over large range of $\mathrm{\tau}_{\rm \sigma}$ and $\tau_{\rm i}$. The significance level probabilities from the KS test for the comparison are presented in Table 1 for $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 3, $\tau_{\rm i}$ $\geq$ 20% and $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 8, $\tau_{\rm i}$ $\geq$ 80% and all have acceptable values. It was found that when $\tau_{\rm i}$ increased the pulses become less contaminated by overlaps from adjacent pulses.
\[b\][$t_{\rm r}$ (sec)]{} \[b\][frequency]{} ![ The distribution of $t_{\rm r}$, $t_{\rm f}$, and FWHM ([a-c]{}) for pulses with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 and $\mathrm{\tau}_{\rm i}$ $\geq$ 50%. The dashed lines are lognormals that are consistent with the data. ](Dh291_f2.eps "fig:"){width="0.85\columnwidth"}
\[b\][$t_{\rm f}$ (sec)]{} \[b\][frequency]{} ![ The distribution of $t_{\rm r}$, $t_{\rm f}$, and FWHM ([a-c]{}) for pulses with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 and $\mathrm{\tau}_{\rm i}$ $\geq$ 50%. The dashed lines are lognormals that are consistent with the data. ](Dh291_f3.eps "fig:"){width="0.85\columnwidth"}
\[b\][FWHM (sec)]{} \[b\][frequency]{} ![ The distribution of $t_{\rm r}$, $t_{\rm f}$, and FWHM ([a-c]{}) for pulses with $\mathrm{\tau}_{\rm \sigma}$ $\geq$ 5 and $\mathrm{\tau}_{\rm i}$ $\geq$ 50%. The dashed lines are lognormals that are consistent with the data. ](Dh291_f4.eps "fig:"){width="0.85\columnwidth"}
Correlation Analysis
--------------------
The correlation coefficients between a range of GRB parameters and pulse parameters were obtained using the Spearman Rank Order correlation coefficents $\rho$ with associated probabilities and the values are listed in Table 2. The correlation coefficients for the time intervals between pulses are listed in Table 3 and they are strongly correlated with each other.
Properties $\rho$ Probability
--------------------------------- --------- ------------------------ --
No. of Pulses vs. $T_{90}$ 0.18 0.073
No. of Pulses vs. Total Fluence 0.38 $9.8 \times 10^{-5}$
$T_{90}$ vs Total Fluence 0.21 0.037
Rise Time vs. Fall Time 0.54 $2.7 \times 10^{-15}$
Rise Time vs. FWHM 0.79 $1.3 \times 10^{-40}$
Rise Time vs. Area 0.59 $3.3 \times 10^{-10}$
Rise Time vs. Pulse Amplitude 0.013 0.90
Fall Time vs. FWHM 0.75 $2.6 \times 10^{-33}$
Fall Time vs. Area 0.64 $2.0 \times 10^{-12}$
Fall Time vs. Pulse Amplitude 0.078 0.45
FWHM vs. Area 0.60 $6.89 \times 10^{-11}$
FWHM vs. Pulse Amplitude $-$0.05 0.64
Area vs. Pulse Amplitude 0.65 $1.1 \times 10^{-12}$
: Spearman Rank Order correlation coefficients $\rho$ and associated probabilities between a range of burst (first 3 rows) and pulse parameters (second 10 rows).
Total Number $\rho$ Probability
-------------- -------------- -------------------------------------------- -- -- --
140 0.42/0.30 $1.7 \times 10^{-12}$/$3.0 \times 10^{-3}$
84 0.48/0.32 $4.9 \times 10^{-6}$/$3.2 \times 10^{-3}$
213/107 0.39/0.24 $5 \times 10^{-9}$/$1.1 \times 10^{-2}$
140/84 0.13/$-$0.09 0.13/0.42
: The Spearman correlation coefficients $\rho$ and probabilities for $\Delta$Ts (first 2 rows) and PAs (second 2 rows). The first and third rows refer to adjacent $\Delta$Ts and PAs while the second and fourth rows refer to $\Delta$Ts or PAs separated by one time interval or pulse. The first / second entry for the $\Delta$Ts gives $\rho$ when the $\Delta$Ts are unnormalised / normalised by $T_{90}$. The first / second entry for the PAs gives $\rho$ when the PAs are unnormalised / normalised by the maximum peak in the burst.
\[t\][Area (counts)]{} \[b\][frequency]{} {width="0.85\columnwidth"}
\[t\][Pulse Amplitude (counts)]{} \[b\][frequency]{} {width="0.85\columnwidth"}
\[t\][$\Delta$T (sec)]{} \[b\][frequency]{}
{width="0.85\columnwidth"} \[fig:time\_int\]
Discussion
==========
The major result is that pulses in GRBs with $T_{90}$$<$$2\,$s have very similar distributions of $t_{\rm r}$, $t_{\rm f}$, FWHM, pulse amplitude, pulse area and $\Delta$T. The frequency distributions are broad and all are compatible with lognormal distributions, and at the 50% level span an order of magnitude in the particular pulse property (Table 1). The same pulse properties in GRBs with $T_{90}$$>$$2\,$s are also well described by lognormal distributions [@quillig:2001]. The results of the correlation analysis on the GRBs and the properties of the pulses (Table 2) are generally in good agreement between long and short GRBS. The main difference is the pulse amplitude because it is not anticorrelated with $t_{\rm r}$, $t_{\rm f}$ and FWHM. In GRBs with $T_{90}$$>$$2\,$s the values of $\Delta$T are not random but consistent with a lognormal distribution with a Pareto-Levy tail for a small number of long time intervals in excess of 15s. It was also found that the values of $\Delta$T were correlated over most of the GRBs and up to more than 20 pulses in the GRBs with large numbers of pulses. In short GRBs there is a strong correlation between the values of $\Delta$T and also between the pulse amplitudes (Table 3) but the latter are better correlated in GRBs with $T_{90}$$>$$2\,$s. There are remarkable similarities between the statistical properties of the two classes of GRBs. The clear conclusion is that the same emission mechanism accounts for the two types of GRBs. This conclusion is in agreement with a very different analysis of the temporal structure of short GRBs [@np:2001]. The external shock model [@derm:1999] has serious difficulties in accounting for GRBs with $T_{90}$$<$$2\,$s and with the non-random distribution of correlated time intervals between pulses. The results presented here provide considerable support for the internal shock model [@reemes:1994]. The internal shock model can account for the results obtained on long and short GRBs provided the cause of the pulses and the correlated values of $\Delta$T can be attributed to the central engine. It has been shown that the time intervals between glitches in pulsars and outbursts in SGRs are lognormally distributed [@quillig:2001]. The similar distributions of $\Delta$T in short and long GRBs are indirect evidence for rotation powered systems with super-strong magnectic fields.
There is strong evidence that long GRBs originate in star forming galaxies providing evidence that they are linked to massive stars and supernovae [@mes:2001]. A variety of models, such as collapsars and hypernovae have been proposed as progenitors (e.g. MacFadyen & Woosley 1999). Models of these systems imply that jets can be generated to produce GRBs by an internal shock model. These models cannot account for GRBs with $T_{90}$ less than about 5s. The shorter GRBs have been attributed to neutron star (NS) - NS mergers or NS - black hole mergers (e.g. Ruffert & Janka 1999). A common ingredient of most GRB models is the formation of a black hole with a temporary torus that accretes and powers the relativistic jets and the GRBs. The overall energetics of the various progenitor models differ by at most an order of magnitude. The GRBs appear to have the same emission mechanism and possibly different progenitors for long and short bursts.
Conclusions
===========
A sample of bright GRBs with $T_{90}$$<$$2\,$s have been denoised and analysed by an automatic pulse selection algorithm. The results show that the distribution of the properties of isolated pulses and time intervals between all pulses are compatible with lognormal distributions. The same mechanism seems to be responsible for both long and short GRBs and is attributed to the internal shock model.
|
{
"pile_set_name": "ArXiv"
}
|
=-5mm
¶
[**Does the “Soft Pomeron” Cope with the HERA Data?**]{}
V.A.Petrov[^1], A.V.Prokudin[^2]\
The effect of acceleration of the cross sections growth with c.m.s. energy $\sqrt{s}$ in presence of a hard scale ( a current virtuality, a high mass in the final state etc) revealed at HERA \[1\] raises a question on validity ( or sufficiency) of the Regge-pole model that has a good reputation in the description of cross sections of hadron-hadron scattering \[2\]. This phenomenon can be described by a simple power-like phenomenological parametrization:
\_\^[\^\*p]{}\~s\^, 0.2 where the exponent $\lambda$ not only exceeds its “soft” counterpart $\Delta=\alpha_{\Bbb
{P}}(0)-1\simeq 0.08$ in the Regge-pole method but also depends on the current virtuality $Q^2$. In the framework of the complex J-plane (in the $t$-channel) this means that there are some singularities located to the right of $Re J=\alpha_{\P}$ and which depend on outer kinematic parameters beyond $t$ so that these poles are not of Regge type.
If such poles exist then one should explore their nature and physical meaning. Besides one should find out and verify restrictions due to general principles (unitarity in the first rate).
The effect discovered at HERA caused a vivid response of theoreticians. Among numerous phenomenological theories \[3\] there exists a particular method based on well known school \[4\] of finding solutions of some specific approximations to Bethe-Salpeter equation for gluonic amplitudes. In fact one has to know how QCD works at large distances (“confinement region”) in order to work out a full analysis, but as is well known there is still no theory answering this question. If one adds some assumptions about behaviour in this region to perturbative analysis then one of the practical results is prediction of Regge poles with high intercepts $(\alpha(0)-1>0.4)$. An evident inconsistency of these poles with observed behaviour of hadron-hadron cross sections was explained by smallness of corresponding residues \[4\]. However a “pomeron” of high intercept (so called “hard pomeron”) could be relevant for description of the cross sections rapid growth revealed at HERA. It is sufficient to assume that in presence of additional “hard scale” (photon’s virtuality) corresponding residues are not small.
Taking into account the fact that the “hard pomeron” status as a QCD solution is based on additional hypotheses and its intercept can be restricted by unitarity \[5\] it could be extremely worth trying to analyse the possibility of using the “soft pomeron”.
An alike problem was addressed in paper \[6\]. However the solution is not satisfactory because of a great number of parameters and presence of additional factors of non-Regge type energy dependence.
In this paper we consider that amplitude for vector meson production in the process
-1cm to 6cm [to 6cm]{}
$$\gamma^*+p\to V+p\;\;\;\;$$ is $$T_{\gamma^*p\to Vp}(s,t,Q^2)=C_{\Bbb{P}} (t,Q^2)
\left (i + \cot\frac{\pi\alpha_{\Bbb{P}}(t)}{2}\right)
\left (\frac{s}{Q^2+m^2_V}\right )^{\alpha_{\Bbb{P}}(t)}+$$ C\_(t, Q\^2) (i + ) ()\^[\_(t)]{}. Here $Q^2$ is the photon’s virtuality, $s=W^2=(p+q)^2, \;\;\Bbb{P}$ stands for the pomeron, $\Bbb{R}$ for a secondary reggeon, Regge trajectories have the following form in a linear approximation $$\alpha_{\Bbb{P,R}} (t)=\alpha_{\Bbb{P,R}}(0)+\alpha'_{\Bbb{P,R}}(0)t,$$ $$\alpha_{\Bbb{P}}(0)-1\equiv \epsilon=0.071\pm 0.018,\;\;\;
\alpha_{\Bbb{R}}(0)-1\equiv -\mu=0.46\pm 0.25,$$ $$\alpha'_{\Bbb{P}}(0)=0.25,\;\;\;
\alpha'_{\Bbb{P}}(0)=1.00.$$ The values $\alpha_{\Bbb{P,R}}(0)$ are taken from the papers \[2,7\], where a good description of hadron-hadron processes is achieved by means of these Regge poles. The $Q^2$ dependence of residues $C_{\Bbb{P,R}}$ is not fixed in the Regge approximation and we assume the $t$ dependence in the form $\exp \left [\frac{1}{4}R^2_{\Bbb{P,R}}(Q^2)t\right ]$ where $R_{\Bbb{P,R}}$ are pomeron and reggeon “radii”.
Fig.1,2,3 show the results of the description of cross sections for the processes \^\*+p && \^0+p\
\^\*+p && +p\
\^\*+p && J/+p by means of formula (2).
0.2cm to 7cm [to 7cm]{} -7cm 8.5cm to 7cm [to 7cm]{}
0.2cm to 7cm [to 7cm]{} -7cm 8.5cm to 7cm [to 7cm]{}
-1cm
0.2cm to 7cm [to 7cm]{} -7cm 8.5cm to 7cm [to 7cm]{}
As is noticed in the paper \[8\], unitarity corrections contribute to the cross section less than $10\%$, so with present experimental accuracy we decided to consider the Born term only. It is seen that the Regge method is relevant enough for the processes (3) description and hence there is no need for taking into account new poles (probably of non-Regge type) besides well known $\Bbb{P}$ and $\Bbb{R}$.
Now let us consider the total cross section for the DIS process $$\gamma^*+p\to X.$$ The result of using formula similar to (2) is shown in fig.4. It is obvious that in this case poles $\P$ and $\R$ are insufficient. Since
-1cm
0.2cm to 7cm [to 7cm]{}
-7cm 8.5cm to 7cm [to 7cm]{}
$$\sigma^{\gamma^*p}_{\mbox\small{{tot}}}=\sum_{V}\sigma_{\gamma^*p\to Vp}+
\sigma^{\gamma^*p}_{\mbox\small{{inel}}},$$ the insufficiency of “usual” Regge poles stems from behaviour of $\sigma^{\gamma^*p}_{\mbox\small{{inel}}}$ which is connected with processes of multiple production.
Does it mean that for successful description of the total cross-sections we need poles with a high intercept? Not at all. If one adds “secondary” Regge poles with $\alpha (0)<1$ then one will have a good description as is seen in fig.4.They have the following intercepts:
$$\alpha_{{\Bbb{R}}_{1}}(0)-1=-0.08;$$ $$\alpha_{{\Bbb{R}}_{2}}(0)-1=-0.1.$$
The problem of interpretation of these new poles, which have intercepts lying close to each other and to 1, rises immediately. One should also include these poles into the fit of exclusive cross-sections.
Thus, we have demonstrated that Regge poles describing hadron-hadron processes do cope with a rapid growth of exclusive vector meson production cross sections at HERA and we explain this growth as a “threshold effect”. Meanwhile it is not possible to do the same for the total cross sections unless one adds new poles, which are not “hard” however.
The authors would like to express their acknowledgment to A.De Roeck kindly granted them with experimental data and to J.Field for useful discussions.
[99]{} ,(ZEUS Collab), ICHEP96, pa02-028;\
[*S.Aid et al.*]{} (H1 Collab.)// Nucl.Phys. 1996. V.B468. P.3;\
[*M. Derrick et al.*]{} (ZEUS Collab.)// Phys.Lett. 1995. V.B356. P.601;\
[*M. Derrick et al.*]{} (ZEUS Collab.)// Phys.Lett. 1996. V.B380. P.220;\
[*S.Aid et al.*]{} (H1 Collab.)// Nucl.Phys. 1996. V.B472. P.3;\
[*M.R.Adams et al*]{}.(E655 Collab.) MPI-PhE/97-03 (submitted to Zeitshrift für Physik C). . // Phys.Lett. 1992. V.B296. P.227. Deep inelastic scattering and related phenomena DIS96. Editors G.D’Agostini and A.Nigro. World Scientific. // Sov.Phys.JETP. 1986. V.63. P.904. . Proc. of QCD-96, Editor S.Narrison 1996, Montpellier, France 4-12 July, p.160. . Frontiers in Strong Interactions, Eds. P.Chiapetta, M.Haguenauer and J.Tr\^ an Thanh V\^ an 1996, Editions Frontièrs. P.203. . Phys.Rev. 1996. D54. P.192. Is the “Soft Pomeron” Valid for the Description of the Data from HERA? hep-ph/9706257. Submitted to Yad.Fiz.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
UMN-TH-3219/13\
[**Gauge-flation confronted with Planck** ]{}\
[ [**Ryo Namba , Emanuela Dimastrogiovanni , Marco Peloso** ]{} ]{}\
School of Physics and Astronomy, University of Minnesota, Minneapolis, 55455, USA\
[**Abstract**]{}
> Gauge-flation is a recently proposed model in which inflation is driven solely by a non-Abelian gauge field thanks to a specific higher order derivative operator. The nature of the operator is such that it does not introduce ghosts. We compute the cosmological scalar and tensor perturbations for this model, improving over an existing computation. We then confront these results with the Planck data. The model is characterized by the quantity $\gamma \equiv \frac{g^2 Q^2}{H^2}$ (where $g$ is the gauge coupling constant, $Q$ the vector vev, and $H$ the Hubble rate). For $\gamma < 2$, the scalar perturbations show a strong tachyonic instability. In the stable region, the scalar power spectrum $n_s$ is too low at small $\gamma$, while the tensor-to-scalar ratio $r$ is too high at large $\gamma$. No value of $\gamma$ leads to acceptable values for $n_s$ and $r$, and so the model is ruled out by the CMB data. The same behavior with $\gamma$ was obtained in Chromo-natural inflation, a model in which inflation is driven by a pseudo-scalar coupled to a non-Abelian gauge field. When the pseudo-scalar can be integrated out, one recovers the model of Gauge-flation plus corrections. It was shown that this identification is very accurate at the background level, but differences emerged in the literature concerning the perturbations of the two models. On the contrary, our results show that the analogy between the two models continues to be accurate also at the perturbative level.
Introduction {#sec:intro}
============
The paradigm of inflation offers a successful theoretical framework for the physics of the Early Universe [@Linde:2005ht]. In most of its realizations, inflation is driven by scalar fields, a feature which can easily result in an isotropic expansion. Quantum fluctuations generated from scalar fields during or at the end of inflation can account both for the cosmological fluctuations in the Cosmic Microwave Background (CMB) Radiation and for the formation of Large Scale Structures. A large number of inflationary models with scalar fields provide results in excellent agreement with observations [@Ade:2013uln]. Nevertheless, models of inflation where vector fields play a major role have been receiving quite some attention, inspired both by theoretical and observationally-related motivations (see e.g. [@review1; @review2; @review3; @review4; @review5] for some recent reviews).
In this paper, we study one of these models, namely the recently proposed Gauge-flation model [@Maleknejad:2011jw]. Gauge-flation is a non-Abelian gauge theory minimally coupled to gravity. At the background level, it is characterized by an effective scalar degree of freedom in the form of the vacuum expectation value (vev) of the spatial component of the gauge field. The action for the model reads [@Maleknejad:2011jw] $$\label{oneeq}
S=\int d^{4}x \sqrt{-g}\left[\frac{M_{p}^{2}}{2}R-\frac{1}{4}F_{\mu\nu}^{a}F^{a,\mu\nu}+\frac{\kappa}{96}\left(F_{\mu\nu}^{a}\tilde{F}^{a,\mu\nu}\right)^{2}\right],$$ where $R$ is the Ricci scalar, $F_{\mu\nu}^{a} \equiv \partial_\mu A_\nu^a - \partial_\nu A_\mu^a - g \epsilon^{abc} A_\mu^b A_\nu^c$ is the field-strength tensor of a SU(2) gauge field with coupling constant $g$, and $\tilde{F}^{a,\mu\nu} \equiv \frac{\epsilon^{\mu\nu\alpha\beta}}{2 \sqrt{-g}} F_{\alpha\beta}^a$ is its dual ($\epsilon^{\mu\nu\alpha\beta}$ is totally anti-symmetric, and $\epsilon^{0123}=1$).
Thanks to its anti-symmetric structure, the last term in (\[oneeq\]) does not introduce more than two time derivatives in the equations of motion, and therefore, for positive $\kappa$, the model is ghost free. [^1] Moreover, the gauge symmetry is not explicitly broken in this model, and there are no problems associated with the longitudinal vector polarizations, which instead destabilize several models of vector fields in cosmology [@hcp].
The model admits an isotropic background solution, with $$\label{gF}
\langle A_{i}^{a} \rangle={\hat \phi } \left( t \right) \, \delta_{i}^{a}.$$ Further studies of the model were carried out for example in [@Maleknejad:2011jr], where the cosmic no-hair theorem is tested within Gauge-flation with a Bianchi I background, in [@Ghalee:2012gg], where the background solutions of Gauge-flation were explored, and in [@Noorbala:2012fh], where it is shown that the theory can be embedded in the gravi-leptogenesis scenario of [@Alexander:2004us]. Recently, it was also shown [@Adshead:2012qe; @SheikhJabbari:2012qf] that Gauge-flation shares some background trajectories with Chromo-natural inflation [@Adshead:2012kp], another recent model which assigns to non-Abelian gauge fields a crucial role in the dynamics of inflation and in the generation of primordial fluctuations (see [@review4] for a comprehensive review). [^2]
In Chromo-natural inflation an SU(2) gauge field is coupled to an axion $\varphi$ which plays the role of the inflaton: $$\label{cn}
S=\int d^{4}x\sqrt{-g}\left[\frac{M_{p}^{2}}{2}R-\frac{1}{4}F_{\mu\nu}^{a}F^{a,\mu\nu}-\frac{1}{2}\left(\partial\varphi\right)^{2}-V(\varphi)+\frac{\lambda}{4f}\varphi F_{\mu\nu}^{a}\tilde{F}^{a,\mu\nu}\right].$$ It turns out that, if sufficiently close to the bottom of its potential, the $\varphi$ field can be integrated out, leading precisely to the action in Eq. (\[oneeq\]) plus corrections [@SheikhJabbari:2012qf].
The main goal of Chromo-natural model, and Gauge-flation as well, is to solve an issue that affects most scalar field models of inflation, i.e. the difficulty in ensuring a quasi-flat potential for the inflaton. The condition of quasi-flatness for the potential is important in order to allow enough e-foldings of expansion and to obtain a nearly scale invariant spectrum of perturbations. This motivated the so-called *natural inflation* model [@Freese:1990rb; @Adams:1992bn], where the inflaton enjoys a (broken) axionic shift symmetry, $\varphi\rightarrow\varphi+const.$, that protects its potential from large quantum corrections. The simplest implementation of this idea, namely a single field $\varphi$ evolving in a potential $V \propto \cos \frac{\varphi}{f}$, agrees with observations only if the axion decay constant $f$ is of the order of or greater than the Planck mass [@Savage:2006tr]. It has been debated in the literature whether a trans-Planckian breaking may be compatible with gravity, in the case in which the symmetry is global [@f-Mp-global]. The shift symmetry may emerge from a gauge symmetry, as typically in string theory. However, also in this case a trans-Planckian $f$ is regarded as problematic, since all known controlled string theory constructions are characterized by $f<M_{p}$ [@Banks:2003sx; @Svrcek:2006yi].
A scale $f < M_p$ can be compatible with inflation through a number of mechanisms, several of which also lead to an interesting phenomenology (see [@review5] for a recent review). One can, for instance, consider more than one axion [@sol1; @sol2], require nontrivial compactifications in string theory [@sol3], couple the axion to a $4$-form [@sol4], modify the axion kinetic term [@sol5], or slow-down the axion through particle production [@sol6a; @sol6b]. In particular, in the mechanism of [@sol6a] the dissipation occurs through the production of a U(1) field coupled to the inflaton $\varphi$ through the interaction $\frac{\varphi}{f} F {\tilde F}$ between the inflaton and an Abelian gauge field. Ref. [@Adshead:2012kp] showed that this coupling can also affect the background evolution (prior to any particle production consideration) if the U(1) field is replaced by a SU(2) field with a nonvanishing vacuum expectation value (vev). Due to the interaction with the vev of the vector multiplet, the inflaton can be in slow roll even if its potential would otherwise (i.e. in absence of this interaction) be too steep to give inflation. As shown in [@Adshead:2012qe; @SheikhJabbari:2012qf], Gauge-flation appears as an analogous (“dual”) version of this mechanism. The Gauge-flation formulation is particularly suggestive since it is characterized by the vector field only. To our knowledge, this is the first and only existing stable model in which inflation is driven by a vector field alone (previously introduced models with dynamical vector fields during inflation either have dynamically relevant scalar fields [@review4], or ghosts [@hcp]).
The perturbation analysis for Chromo-natural inflation was first carried out in [@Dimastrogiovanni:2012st] in the limit of heavy vector field (see below), in which the gauge field can be integrated out, leading to an effective single-scalar field $P[(\partial\varphi)^{2},\varphi]$ model of inflation with a non-canonical kinetic term. A full analysis at linear order in perturbation theory was later presented in [@Dimastrogiovanni:2012ew], where the model was found to be highly unstable in the sub-horizon regime for: $$\frac{g^2 \, Q^2}{H^2} < 2 \;\;\;,\;\;\; {\rm instability \; in \; Chromo\mbox{-}natural \; inflation} \;\;.
\label{cn-instab}$$ In this relation $Q$ is the SU(2) vev, related to (\[gF\]) by $Q = {\hat \phi } / a$ (where $a$ is the scale factor of the universe), while $H$ is the Hubble rate. In Chromo-natural inflation, the quantity $m_g \equiv \sqrt{2} \, g \, Q$ coincides with the mass of the vector field fluctuations in the $m_g \gg H$ regime [@Dimastrogiovanni:2012st], which is the regime where the computation of [@Dimastrogiovanni:2012st] applies.
Ref. [@Dimastrogiovanni:2012ew], also noted that the model can lead to a large production of the vacuum gravity wave mode, in excess of the standard Lyth bound [@Lyth:1996im] $r > 16 \epsilon$ (see eqs. (\[epsilon\]) and (\[def-r\]) for the definition of $\epsilon$ and $r$, respectively), which does not apply in this context, since it holds for a free inflaton and unsourced tensor modes. The stability result (\[cn-instab\]) was later obtained independently by [@Adshead:2013qp], that also presented a complete study of the gravitational waves produced in the model (correcting an error in the original version of [@Dimastrogiovanni:2012ew]), showing that they are chiral. Ref. [@Dimastrogiovanni:2012ew] also presented the scalar and tensor power spectrum of the Chromo-natural inflation for some illustrative choices of the parameters. None of the examples presented in [@Dimastrogiovanni:2012ew] leads to an acceptable phenomenology. In particular, these example showed that the spectrum of the scalar perturbations is too red at small $m_g/H$, while the tensor modes are too high at large $m_g/H$. Based on this, ref. [@Dimastrogiovanni:2012ew] argued that the simultaneous requirements of sufficiently flat scalar spectrum and of sufficiently small tensor mode could pose significant bounds on the model, and potentially rule it out as a model of sub-Planckian $f$. This was later confirmed by [@Adshead:2013nka] through an exhaustive parameter scan in the model.
A study of the perturbations of Gauge-flation can instead be found in [@Maleknejad:2011jw]. Ref. [@Maleknejad:2011jw] concluded that all parameter choices lead to a stable solution (contrary to what happens in Chromo-natural inflation, see eq. (\[cn-instab\])). Moreover, the scalar spectrum obtained in [@Maleknejad:2011jw] is significantly bluer than the one found for Chromo-natural inflation; compare for instance Figure 8 of the last work in [@Maleknejad:2011jw] with Figure 12 of [@Adshead:2013nka]. This different behavior is somewhat puzzling, given the analogy that the two models present at the background level. Specifically, ref. [@SheikhJabbari:2012qf] shows that, integrating out the axion $\varphi$ from the action (\[cn\]) of Chromo-natural inflation, one obtains the action (\[oneeq\]) of Gauge-flation, plus corrections. [^3] The procedure is conceptually simple, but it is not straightforward to compute the corrections explicitly. The integration of $\varphi$ is done in [@SheikhJabbari:2012qf] at the level of the one loop effective action; slow roll approximations are used, and in computing the functional determinant the laplacian has been replaced by its flat space value, under the assumption that the main contribution comes from sub-horizon momenta [@SheikhJabbari:2012qf]. One may wonder about the impact of the corrections that would emerge by going beyond these approximations. For instance, the departure of the perturbations from scale invariance is controlled by the full evolution (and not only the sub-horizon regime) and by the slow roll parameters. Barring mistakes, the difference emerged in the literature between the perturbations in the two models is a clear indication that a non-negligible difference between the two models must appear when $\varphi$ is accurately integrated out.
It is interesting to understand how precisely this difference emerges, as this may provide some general indication on how accurate one needs to be in integrating out fields from a model of inflation, and still obtain an accurate effective description. This was the main motivation for the present study. As a byproduct, we also deemed useful to update the phenomenological limits on Gauge-flation in the light of the new Planck results [@Ade:2013zuv], particularly if the analogy with Chromo-natural inflation turns out to be accurate also at the perturbative level. To understand this, we computed the cosmological perturbations of Gauge-flation, employing an analogous procedure to the one that we used to study the perturbations of Chromo-natural inflation [@Dimastrogiovanni:2012ew]. Disregarding the vector perturbations (as we realized that considering them would not impact our conclusions) the model has two physical scalar perturbations, two physical left-handed tensor perturbations, and two physical right-handed perturbations. These three groups are decoupled from each other at the linearized level, and each of them is described by two coupled second order differential equations. The equations (particularly, those in the scalar sector), are extremely involved and we could not solve them analytically. However, they can be integrated numerically without any particular difficulty. After imposing a given duration of inflation (we study both the cases in which there are $N=50$ or $N=60$ e-folds between the moment when the largest CMB modes exit the horizon and the end of inflation), and that the scalar perturbations have the observed amplitude, Gauge-flation is characterized by a single free parameter, that can be chosen to be the same combination $\gamma \equiv g^2 \, Q^2 / H^2$ that also plays a relevant role for the perturbations of Chromo-natural inflation. We obtained phenomenological results for Gauge-flation as a function of this parameter. Our results differ from those of [@Maleknejad:2011jw], and agree with those emerged from the several analyses of Chromo-natural inflation. In particular, we find that the stability and departure from scale invariance of the scalar modes, and the level of the tensor modes of Gauge-flation scale with $\gamma$ analogously to what found in Chromo-natural inflation. We find that also Gauge-flation is ruled out by the CMB data.
The paper is organized as follows. In Section \[sec:model\] we introduce the model and review the background analysis, mostly summarizing the analogous computation of [@Maleknejad:2011jw]. In Section \[sec:linpert\] we present our formalism for the perturbations. Specifically, we discuss at the formal level how we fix the gauge freedom, integrate out the non-dynamical modes, and quantize the early time action, to obtain the initial condition for the perturbations. The explicit computations for the model are presented in Section \[sec:tensor\] for the tensor perturbations, and in Section \[sec:scalar\] for the scalar modes. In Section \[sec:pheno\] we study the phenomenological implications of these results. In Section \[sec:comparison\] we compare our computation with that of [@Maleknejad:2011jw]. Finally, in the concluding Section \[sec:conclusions\] we summarize our findings, and compare them with those in Chromo-natural inflation.
The model, the background solution, and the slow roll approximation {#sec:model}
===================================================================
The “matter” Lagrangian of Gauge-flation, see eq. (\[oneeq\]), consists of a standard Yang-Mills term and of a gauge-invariant contribution of the form $(F\tilde{F})^{2}$ ($\tilde{F}$ being the dual field-strength tensor). Due to the SU(2) gauge symmetry, the theory admits a rotationally invariant vev for the SU(2) multiplet, specified by eq. ([\[gF\]]{}), which is compatible with an isotropic background. We choose an FRW background metric [^4] $$\label{s2one}
ds^{2}=-dt^{2}+a^{2}(t)\delta_{ij}dx^{i}dx^{j}.$$
The energy-momentum tensor of the system, $$T_{\mu \nu} = F^{a}_{\mu \alpha} F^a_{\nu \beta} g^{\alpha \beta} - g_{\mu \nu} \left[ \frac{1}{4} F^2 + \frac{\kappa}{96} \left( F {\tilde F} \right)^2 \right] \;\;,
\label{Tdd}$$ once evaluated on the background, assumes the standard form $$\label{s2two}
T_{\;\; \nu}^{\mu}=diag\left(-\rho,P,P,P\right) \;\;.$$ In this expression, the energy density and the pressure can be written as the sum of two contributions, the first one arising from the Yang-Mills terms and the second one from the $(F\tilde{F})^{2}$ interaction $$\label{s2three}
\rho=\rho_{YM}+\rho_{\kappa},\quad\quad\quad P=\frac{1}{3}\rho_{YM}-\rho_{\kappa} \;\;,$$ where $$\label{s2four}
\rho_{YM}\equiv\frac{3}{2}\left(\frac{\dot{{\hat \phi } }^{2}}{a^{2}}+g^{2}\frac{{\hat \phi } ^{4}}{a^{4}}\right),\quad\quad\quad \rho_{\kappa}\equiv\frac{3}{2}\kappa g^{2}\frac{\dot{{\hat \phi } }^{2}{\hat \phi } ^{4}}{a^{6}} \;\;.$$ We can immediately see that the standard Yang-Mills contribution leads to the equation of state of radiation, while the second contribution to an effective equation of state of a cosmological constant. The $(F\tilde{F})^{2}$ term thus turns out to be a convenient choice for building a model of inflation. Despite being higher order, this term has only two time derivatives, due to its anti-symmetric structure, and the theory can be stable for some choice of parameters (as our study below shows). However, when this term dominates over the Yang-Mills one, it is natural to wonder whether higher order operators (which may arise from loops) can be safely suppressed. The authors of [@Maleknejad:2011jr] claim that this is the case within the parameter space of the theory.
The background evolution of Gauge-flation has been exhaustively studied in the literature, starting from the original proposal [@Maleknejad:2011jw]. We summarize it here for completeness, and to discuss the freedom in the choice of parameters and initial conditions in the theory.
We can immediately verify that slow roll inflation requires (the standard FRW equations for $H$ and $\dot{H}$ apply) $$\epsilon \equiv - \frac{\dot{H}}{H^2} = \frac{2 \rho_{YM}}{\rho_{YM} + \rho_\kappa} \ll 1 \;\; \Leftrightarrow \;\; \rho_{\kappa}\gg \rho_{YM} \;\;,
\label{epsilon}$$ confirming that the energy associated to the higher order term in (\[oneeq\]) must dominate over the one from the conventional Yang-Mills term. Besides the slow roll parameter $\epsilon$, it is useful to introduce the two additional slow-roll parameters: $$\label{s2seven}
\eta\equiv -\frac{\ddot{H}}{2H\dot{H}}=\epsilon-\frac{\dot{\epsilon}}{2\epsilon H} \;\;\;,\;\;\;
\delta\equiv -\frac{\dot{Q}}{HQ} \;\;.$$ The parameter $\eta$ is related to the time variation of $\epsilon$, so that $\eta \ll 1$ imposes that $\epsilon$ varies very slowly during inflation; the other slow-roll parameter is defined from the time variation of the physical field $$\label{s2eight}
Q \equiv \frac{{\hat \phi } }{a} \;\;,$$ so $ \delta \ll 1$ ensures that the gauge inflaton $Q$ sustains inflation long enough. The condition $\epsilon\ll 1$ in turn translates into $\kappa g^{2}Q^{4}\gg 1$ and $\kappa H^{2}Q^{2}\left(1-\delta\right)^{2}\simeq\kappa H^{2}Q^{2}\gg 1$. The slow-roll parameters satisfy the relations $$\label{sl2}
\epsilon\simeq\frac{Q^{2}}{M_{p}^{2}}\left(1+\gamma\right),\quad\quad\eta\simeq\frac{\epsilon}{1+\gamma}\simeq\frac{Q^{2}}{M_{p}^{2}},\quad\quad\delta\simeq\frac{\gamma}{6\left(1+\gamma\right)}\epsilon^{2},$$ where $\gamma $ is defined in (\[def-gamma\]). The relations (\[sl2\]) have been derived in [@Maleknejad:2011jw] and for brevity we do not repeat the derivation here. We note that, for the slow-roll conditions to be satisfied, $Q\ll M_{p}$ is needed, so that, according to standard terminology, this is a “small-field” inflationary model. However, as we show below, contrary to what happens in standard small-field models, Gauge-flation can lead to too large an amount of gravitational waves for some choice of parameters.
Following the notation of [@Maleknejad:2011jw], in eq. (\[sl2\]) we have introduced the quantity $$\gamma \equiv \frac{g^{2}Q^{2}}{H^{2}} \;\;,
\label{def-gamma}$$ which we see corresponds to the same combination (\[cn-instab\]) bounded by the stability of the scalar modes in Chromo-natural inflation. One of the goals of the current work is to study whether a bound analogous to (\[cn-instab\]) is also present in Gauge-flation.
By combining the definition (\[def-gamma\]) with eqs. (\[epsilon\]) and (\[s2seven\]) we can obtain the exact relation $$\kappa = \frac{1}{H^2 \gamma Q^2} \; \frac{\left( 1 - \delta \right)^2 + \gamma}{\left( 1 - \delta \right)^2} \, \frac{2-\epsilon}{\epsilon} \;\;,
\label{relation-kappa}$$ which will be used in the following Sections to eliminate the parameter $\kappa$ from the explicit form of the equations of the perturbations that we solve. Another useful exact relation that we will use in the study of the perturbations is $$M_p = Q \sqrt{\frac{\left( 1 - \delta \right)^2 + \gamma}{\epsilon}} \;\;.
\label{relation-Mp}$$ To prove this relation, we rewrite (\[epsilon\]) as $\epsilon = \frac{2 \rho_{YM}}{3 H^2 M_p^2}$, we express $\rho_{YM}$ through eqs. (\[s2four\]), (\[s2eight\]), and, finally, we trade $\dot{Q}$ for $\delta$ - using (\[s2seven\]) - and $g^2$ for $\gamma$ - using (\[def-gamma\]).
Differentiating (\[def-gamma\]), and using the definition of the slow roll parameters, we obtain that $$\dot{\gamma}= 2 \gamma \left(\epsilon-\delta \right) H \;\;,$$ so that $\gamma$ is a slowly rolling quantity. More in general, from the slow roll conditions, one can show [@Maleknejad:2011jw] that the quantities $\gamma H^2 \propto Q^2 \propto \epsilon / \left( 1 + \gamma \right)$ are constant at leading order in slow roll. Therefore, $$\label{s2fourteen}
\frac{\epsilon}{\epsilon_{\rm in}}\simeq\frac{1+\gamma}{1+\gamma_{\rm in}}
\;\;,\;\;
\frac{H^{2}}{H_{\rm in}^{2}}\simeq \frac{\gamma_{\rm in}}{\gamma} \;\;,$$ where the subscript “in” indicates some initial time, which we take to correspond to $N$ e-folds before the end of inflation.
The number of e-folds can be related to the initial values of $\gamma$ and $\epsilon$ by [@Maleknejad:2011jw] $$\label{nN}
N\simeq \frac{1+\gamma_{\rm in}}{2\epsilon_{\rm in}}\ln\left[\frac{1+\gamma_{\rm in}}{\gamma_{\rm in}}\right].$$ up to subleading terms in slow roll.
We introduce the dimensionless quantities $${\tilde Q} \equiv \frac{Q}{M_p} \;\;,\;\; {\tilde t} \equiv \frac{t}{\sqrt{\kappa} M_p} \;\;,\;\;
{\tilde g} \equiv \sqrt{\kappa} M_p^2 g \;\;.
\label{tilde-combo}$$ and we denote ${\tilde \partial} \equiv \frac{\partial }{ \partial {\tilde t} } $. In terms of these quantities, the only nontrivial SU(2) background equation and the background “ii-Einstein” equation read, respectively $$\begin{aligned}
& &
{\tilde \partial}^2 {\tilde Q} + 2 \frac{ {\tilde \partial} a}{a} {\tilde \partial} {\tilde Q} + \frac{ {\tilde \partial}^2 a}{a} {\tilde Q} + \frac{ 2 {\tilde g}^2 }{ 1 + {\tilde g}^2 {\tilde Q}^4 }
\left[ 1 + \left( {\tilde \partial} {\tilde Q} + \frac{ {\tilde \partial} a}{a} {\tilde Q} \right)^2 \right] {\tilde Q}^3
+ \frac{ {\tilde \partial} a }{a} \frac{ 1 - 3 {\tilde g}^2 {\tilde Q}^4 }{ 1 + {\tilde g}^2 {\tilde Q}^4 }
\left( {\tilde \partial} {\tilde Q} + \frac{ {\tilde \partial} a}{a} {\tilde Q} \right) = 0 \;\; , \nonumber\\
& &
\frac{{\tilde \partial}^2 a}{a} - \left( \frac{{\tilde \partial} a}{a} \right)^2 = - \left[ \left( {\tilde \partial} {\tilde Q} + \frac{ {\tilde \partial} a}{a} {\tilde Q} \right)^2 + {\tilde g}^2 {\tilde Q}^4 \right] \;\; ,
\label{eom-Q-ii}\end{aligned}$$ while the background "00-Einstein” equation reads $$\left( \frac{{\tilde \partial} a}{a} \right)^2 = \frac{1}{2} \left[ \left( {\tilde \partial} {\tilde Q} + \frac{ {\tilde \partial} a}{a} {\tilde Q} \right)^2 \left( 1 + {\tilde g}^2 {\tilde Q}^4 \right) + {\tilde g}^2 {\tilde Q}^4 \right] \;\;.
\label{eom-00}$$ These three equations are related to each other by a nontrivial Bianchi identity. A set of independent equations is provided by either one of (\[eom-Q-ii\]) and by (\[eom-00\]). Equivalently, one can solve the two equations (\[eom-Q-ii\]), but then eq. (\[eom-00\]) has to be imposed as an initial condition. Whatever the choice, the background evolution is completely determined by the initial values ${\tilde Q}_{\rm in}$ and ${\tilde \partial } {\tilde Q}_{\rm in}$ (for a flat universe, the normalization of the scale factor is arbitrary, and we set $a_{\rm in} = 1 $).
We now show that all quantities needed to specify a background evolution can be given in terms of $N$ and of $\gamma_{\rm in}$. Eliminating $\epsilon$ through (\[sl2\]), eq. (\[nN\]) becomes a relation in terms of $N, \gamma_{\rm in}$ and ${\tilde Q}_{\rm in}$, that we solve numerically to obtain ${\tilde Q}_{\rm in}$ in terms of the other two quantities: $${\tilde Q}_{\rm in} \simeq \left[ \frac{1}{2 N}\ln\left(\frac{1+\gamma_{\rm in}}{\gamma_{\rm in}}\right) \right]^{1/2} \;\;.
\label{getQin}$$ Next, consider eq. (\[def-gamma\]), and combine eqs. (\[s2seven\]) and (\[sl2\]). These relations, written in terms of dimensionless quantities, and evaluated at the initial time, give $$\begin{aligned}
&& {\tilde g} = \frac{ \gamma_{\rm in}^{1/2} {\tilde \partial } a_{\rm in} }{ {\tilde Q}_{\rm in} } \;\;, \nonumber\\
&& {\tilde \partial} {\tilde Q}_{\rm in} = - \delta_{\rm in} {\tilde Q}_{\rm in} {\tilde \partial a}_{\rm in} \simeq - \frac{\gamma_{\rm in} \left( 1 + \gamma_{\rm in} \right) {\tilde Q}_{\rm in}^5}{6} \,
{\tilde \partial a}_{\rm in} \;\; ,
\label{getgQp}\end{aligned}$$ These two equations, plus eq. (\[eom-00\]) are three relations in terms of three unknown quantities ${\tilde g} ,\,
{\tilde \partial } a_{\rm in} ,\, $ and $ {\tilde \partial} {\tilde Q}_{\rm in} $. In this way we obtain all the quantities needed for the background evolution.
To summarize: Gauge-flation is characterized by the two parameters $\kappa$ and $g$. A given inflationary evolution is specified by the number of e-folds $N$, the values of the inflaton ($Q$) and its time derivative at some initial time. One of the parameters ($\kappa$, in the current case) can be rescaled out of the background evolution by reabsorbing it in the units of time (eq. (\[tilde-combo\]), in the current case). We will determine this parameter by imposing that the scalar perturbations have the measured amplitude (this is completely analogous to what happens to the parameter $m$ in massive chaotic inflation, $V = \frac{1}{2} m^2 \varphi^2$). For any choice of the remaining parameter ${\tilde g}$, the initial value of the inflaton is one-to-one related to the number of e-folds of inflation. The initial derivative of the inflaton is instead obtained by imposing that the evolution is in the slow roll inflationary attractor (in our case, the inflationary attractor is characterized by the last of (\[getgQp\])).
In our analysis, we “traded” the remaining parameter ${\tilde g}$ for $\gamma_{\rm in}$ by imposing the first of (\[getgQp\]). The quantity $\gamma$ is constant at leading order in slow roll , so its initial value $\gamma_{\rm in}$ is a good quantity to characterize a given evolution. The choice of presenting our results in terms of $\gamma_{\rm in}$ rather than ${\tilde g}$ is dictated by the fact that $\gamma$ played an important role in the related model of Chromo-natural inflation (see eq. (\[cn-instab\])). However, we could have equivalently used ${\tilde g}$, and the left panel of Figure \[g-eps\] shows how these two quantities are related for $N=50$ and $N=60$. In the right panel of the Figure we show instead how the initial value of the slow roll parameter $\epsilon$ is related to $\gamma_{\rm in}$.
![Left panel: Relation between the dimensionless parameter ${\tilde g} \equiv \sqrt{\kappa} M_p^2 g $, and the initial value of $\gamma $, for $N=50$ and $N=60$ e-folds of inflation. The quantities $N$ and $\gamma_{\rm in}$ completely characterize the model and the background evolution. Right panel: relation between the initial value of the slow roll parameter $\epsilon$ and of $\gamma$. \[g-eps\]](g-vs-gamma "fig:"){width="45.00000%"} ![Left panel: Relation between the dimensionless parameter ${\tilde g} \equiv \sqrt{\kappa} M_p^2 g $, and the initial value of $\gamma $, for $N=50$ and $N=60$ e-folds of inflation. The quantities $N$ and $\gamma_{\rm in}$ completely characterize the model and the background evolution. Right panel: relation between the initial value of the slow roll parameter $\epsilon$ and of $\gamma$. \[g-eps\]](eps-vs-gamma "fig:"){width="45.00000%"}
In Figure \[fig:Q-eps-bck\], we show both the inflationary evolution of $Q(t)$ (left panel) and of the slow roll parameter $\epsilon$ (right panel), for $60$ e-folds of inflation and for two values of the parameter $\gamma$. Solid lines in the Figure give the numerical exact evolution, while dashed lines the analytic slow-roll approximation. The solid and dashed lines are almost superimposed to each other, signaling the great accuracy of the slow roll approximation.
![ Inflationary evolution of $Q$ (left panel) and of $\epsilon$ (right panel) as a function of the number of e-folds, for two different values of $\gamma_{60}$ (the value $\gamma_{\rm in}$ at $60$ e-folds before the end of inflation). Solid lines in the Figure give the numerical exact evolution, while dashed lines the analytic slow-roll approximation. The solid and dashed lines are almost superimposed to each other, signaling the great accuracy of the slow roll approximation. []{data-label="fig:Q-eps-bck"}](Q_SR "fig:"){width="45.00000%"} ![ Inflationary evolution of $Q$ (left panel) and of $\epsilon$ (right panel) as a function of the number of e-folds, for two different values of $\gamma_{60}$ (the value $\gamma_{\rm in}$ at $60$ e-folds before the end of inflation). Solid lines in the Figure give the numerical exact evolution, while dashed lines the analytic slow-roll approximation. The solid and dashed lines are almost superimposed to each other, signaling the great accuracy of the slow roll approximation. []{data-label="fig:Q-eps-bck"}](epsilon-SR "fig:"){width="45.00000%"}
Linear perturbations {#sec:linpert}
====================
In this Section, we study at a formal level the most general perturbations of the background solution presented in Section \[sec:model\]. In Subsection \[subsesec:gauge\] we split the perturbations in three groups that are decoupled from each other at the linearized level (and that can be therefore studied independently), and we fix the gauge freedoms associated with general coordinate transformations and with the SU(2) group. In Subsection \[subsec:formal-deco\] we present our formalism to solve the linearized equations for these perturbations.
General decomposition and gauge choice {#subsesec:gauge}
--------------------------------------
This discussion closely follows Subsection III A of [@Dimastrogiovanni:2012ew], where the reader is referred to for details. There are 22 perturbations in the model, $12$ of which in the SU(2) vector field, and $10$ in the metric. We decompose: $$\begin{aligned}
A_0^a & = & a \left( Y_a + \partial_a Y \right) \;\;, \nonumber\\
A_i^a & = & a \left[ \left( Q + \delta Q \right) \delta_{ai} + \partial_i \left( M_a + \partial_a M \right)
+ \epsilon_{iab} \left( U_b + \partial_b U \right) + t_{ia} \right] \;\;, \nonumber\\
g_{00} & = & - a^2 \left( 1 - 2 \phi \right) \;\;, \nonumber\\
g_{0i} & = & a^2 \left( B_i + \partial_i B \right) \;\;, \nonumber\\
g_{ij} & = & a^2 \left[ \left( 1 + 2 \psi \right) \delta_{ij} + 2 \partial_i \partial_j E + \partial_i E_j + \partial_j E_i + h_{ij} \right]
\;\;,\end{aligned}$$ where $a$ denotes both the SU(2) index and the scale factor (as we believe that this does not cause ambiguity), while $i=1,2,3$ ranges over the spatial coordinates. The modes $t_{ia}$ and $h_{ij}$ are transverse and traceless ($\partial_i h_{ij} = \partial_i t_{ia} = \partial_a t_{ia} = t_{ii} = h_{ii} = 0$) and we denote them as “tensor modes”. The modes $Y_a , M_a, U_a, B_i , E_i$ are transverse ($\partial_i Y_i = \dots = \partial_i E_i = 0$), and we denote them as “vector modes”. We denote the remaining modes as “scalar modes”. These three groups of modes are separate from each other at the linearized level, and can be studied independently [@Dimastrogiovanni:2012ew]. [^5]
As shown in [@Dimastrogiovanni:2012ew], we can always choose the gauge $\psi=E=E_i=U=U_i=0$. This gauge choice completely fixes the freedom associated to general coordinate and SU(2) transformations. We Fourier transform each perturbation according to $$\delta \left( t , \vec{x} \right) = \int \frac{d^3 k}{\left( 2 \pi \right)^{3/2} } \, {\rm e}^{i \vec{k} \cdot \vec{x}} \, \delta \left( t , \vec{k} \right) \;\;.
\label{FT}$$
We are only interested in studying the linearized theory of the perturbations. This amounts in expanding the equations of motion of the system up to first order in the perturbations, or in expanding the action of the system in quadratic order in the perturbation (this is equivalent, since the linearized equations for the perturbations are obtained by extremizing the quadratic action for the perturbations). In both these computations, the modes are decoupled from each other (since the perturbations are coupled to each other only at the nonlinear level). Therefore, without loss of generality in the linearized computation, we can choose the momentum $\vec{k}$ of the modes to be oriented along one given direction, which we choose to be the $3$rd one [^6]
We end up with the following perturbations: $$\begin{aligned}
{\rm scalar}: & & \delta A_\mu^1 = a \left( 0 , \delta Q , 0 , 0 \right) \;\;,\;\;
\delta A_\mu^2 = a \left( 0 , 0 , \delta Q , 0 \right) \;\;,\;\;
\delta A_\mu^3 = a \left( \partial_z Y , 0 , 0 , \delta Q + \partial_z^2 M \right) \;\;, \nonumber\\
& & \delta g_{00} = a^2 2 \phi \;\;,\;\; \delta g_{03} = a^2 \partial_z B \;\;, \nonumber\\
{\rm vector}: & & \delta A_\mu^i = a \left( Y_i , 0 , 0 , \partial_z M_i \right) \;\;,\;\;
\delta g_{0i} = a^2 B_i \;\;,\;\; i=1,2 \nonumber\\
{\rm tensor}: & & \delta A_\mu^1 = a \left( 0 , t_+ , t_\times , 0 \right) \;\;,\;\;
\delta A_\mu^2 = a \left( 0 , t_\times , - t_+ , 0 \right) \;\;,\;\;
\delta g_{11} = - \delta g_{22} = a^2 h_+ \;\;,\;\; \delta g_{12} = a^2 h_\times \;\;.
\label{modes}\end{aligned}$$
Namely, after $7$ perturbations are removed by the gauge fixing, we end up with $5$ modes in the scalar sector, $6$ modes in the vector sector, and $4$ modes in the tensor sector. Not all these modes correspond to physically propagating independent degrees of freedom. Due to the structure of the kinetic terms, the modes originating from $\delta g_{0\mu}$ and $\delta A_0^a$ enter in the quadratic action of the perturbations without time derivatives (up to boundary terms) and are commonly denoted as “non-dynamical” perturbations. [^7] Extremizing the quadratic action of the perturbations with respect to these modes provide equations that are algebraic in the non-dynamical modes. These equations are called “constraint equations”. We solve these equations by expressing the non-dynamical modes in terms of the dynamical modes and their first time derivatives.
Therefore, the non-dynamical modes do not introduce additional degrees of freedom in the initial condition, but are uniquely determined in terms of the dynamical modes. We thus see that Gauge-flation is characterized by $22$ (the starting number of modes) minus $7$ (gauge-fixing) minus $7$ (the amount of non-dynamical modes) equal $8$ physically propagating degrees of freedom. In our choice (\[modes\]), these are the modes $\delta Q$ and $M$ in the scalar sector, the modes $M_1$ and $M_2$ (in the vector sector), and all the $4$ modes in the tensor sector.
In the remainder of this work we disregard the vector sector, since it is possible to show that the model is ruled out by observations from the study of the scalar and tensor sectors alone.
Formal expressions for the the linearized equations, and formal solutions {#subsec:formal-deco}
-------------------------------------------------------------------------
In this Subsection we study the quadratic action for the perturbations and the corresponding linearized equations of motion at the formal level. Particular care is taken in distinguishing the role of the non-dynamical vs the dynamical modes. The discussion follows and summarizes the analogous one presented in Appendix A of [@Dimastrogiovanni:2012ew].
Once expanded at second order in the perturbations, the action of the system splits in three separate parts $$S_{\rm quadratic} = S_{\rm scalar} + S_{\rm vector} + S_{\rm tensor} \;\;,$$ all of which are Hermitian. The tensor action only contains dynamical modes, while the other two actions contain both dynamical and non-dynamical modes (see the final part of the previous Subsection).
In complete generality, a Hermitian and quadratic action for a set of dynamical modes $\left\{ X_i \right\}$ and a set of non-dynamical modes $ \left\{ N_i \right\}$ is of the form (in matrix notation) $$S = \int d \tau d^3 k \left[ X^{'\dagger} A X' + \left( X^{'\dagger} B X + {\rm h.c.} \right)
+ X^\dagger C X + \left( N^\dagger D X' + {\rm h.c.} \right) + \left( N^\dagger E X + {\rm h.c.} \right) + N^\dagger F N \right]
\;\;,
\label{S2-formal}$$ where $A,C,F$ are Hermitian, and (up to an integration by parts) $B$ is anti-Hermitian. These matrices are function of background quantities, and therefore they are time dependent. Extremizing (\[S2-formal\]) with respect to the non-dynamical and the dynamical fields, we find, respectively, $$\begin{aligned}
&& D X' + E X + F N = 0 \;\;, \nonumber\\
&& \left( A X' + B X + D^\dagger N \right)' - B^\dagger X' - C X - E^\dagger N = 0 \;\;.
\label{eom1-formal}\end{aligned}$$ These equations are nothing but the linearized equations for the perturbations, and can equivalently be obtained by expanding the exact SU(2) and Einstein equations of the model at first order in the perturbations.
The first line in (\[eom1-formal\]) is the constraint equations of the system, which are solved by $$N = - F^{-1} \left( D X' + E X \right) \;\;.
\label{constraint-sol}$$ We can insert this solution back into the second line of (\[eom1-formal\]), so that the remaining linearized equations (that we still have to solve) are expressed in terms of dynamical variables only: $$\begin{aligned}
&&
\left( A - D^\dagger F^{-1} D \right) X'' + \left[ \left( A-D^\dagger F^{-1} D \right)' +
\left( B - D^\dagger F^{-1} E - {\rm h.c.} \right) \right] X' \nonumber\\
& & \quad\quad\quad\quad \quad\quad\quad\quad \quad\quad\quad\quad \quad\quad\quad\quad +
\left[ \left( B - D^\dagger F^{-1} E \right)' - C + E^\dagger F^{-1} E \right] X = 0 \;\;.
\label{eom2-formal}\end{aligned}$$ We can also insert the solution (\[constraint-sol\]) back into the starting action (\[S2-formal\]), and obtain a quadratic action for the dynamical variables only $$S = \int d \tau d^3 k \left[ X^{'\dagger} \left( A - D^\dagger F^{-1} D \right) X' +
\left( X^{'\dagger} \left( B - D^\dagger F^{-1} E \right) X + {\rm h.c.} \right)
+ X^\dagger \left( C - E^\dagger F^{-1} E \right) X \right] \;\;.
\label{S22-formal}$$ Obviously, extremizing the quadratic action (\[S22-formal\]) precisely produces the equations of motion (\[eom2-formal\]).
To summarize, we have explicitly outlined at the formal level the procedure that is commonly denoted as “integrating out the non-dynamical variables”. We have obtained a quadratic action in terms of the dynamical variables only. All the information necessary to solve the system of these variables is contained in this action, and only in it. As we have seen, this action provides the linearized equations (\[eom2-formal\]) in terms of the dynamical variables only; moreover, as we discuss in Subsection \[subsec:PS\], this action is the starting point of the quantization of the perturbations, which will allow us to set the initial conditions for the modes. Therefore, the action (\[S22-formal\]) completely determines the Cauchy problem that uniquely determines the solutions for the dynamical modes.
In our computations below we explicitly derive the action (\[S22-formal\]) for the dynamical variables of Gauge-flation. We stress that, once the explicit form of (\[S22-formal\]) has been obtained, it contains all the information needed to derive the final solutions for the dynamical modes, and that the constraint equations can no longer be used for this purpose (we have just used them to eliminate the non-dynamical modes, and using them again would reintroduce the non-dynamical modes in the system that we are solving). In particular, we can no longer use the constraint equations to “learn” anything new on the dynamical variables beyond what the action (\[S22-formal\]) indicates. Only after the solution for the dynamical variables have been obtained from (\[S22-formal\]), the constraint equations can be used in the form (\[constraint-sol\]) to determine the explicit solution for the non-dynamical modes in terms of the explicit solutions for the dynamical modes that we have obtained from (\[S22-formal\]).
The procedure outlined in this Subsection is commonly used in the case of scalar field inflation. For instance, in appendix A of [@Dimastrogiovanni:2012ew] we explicitly worked out the example of a single scalar field, and showed how the results of [@Mukhanov:1990me] are reproduced. The fact that some of the variables in the problem at hand originated from a vector multiplet does not impact this discussion in any way, since the study presented here starts from the action (\[S2-formal\]), which can always been written in this form for any system of perturbations.
Quantization and power sectra {#subsec:PS}
-----------------------------
The kinetic matrix $A-D^\dagger F^{-1} D$ in (\[S22-formal\]) is Hermitian, and can be diagonalized through $$X_i = {\cal M}_{ij} \Delta_j \;\;.
\label{XtoD}$$ The “mix term” in the resulting action in terms of $\Delta$ can be simplified by removing a boundary term, and we can write the action in the form $$S = \frac{1}{2} \int d \tau d^3 k \left[ \Delta^{' \dagger} T \Delta' + \Delta^{' \dagger} K \Delta - \Delta^\dagger K \Delta' - \Delta^\dagger \Omega^2 \Delta \right] \;\;,\;\; T = 1 \;\;,
\label{S23-formal}$$ where the Hermitianity of the action implies that $K$ is anti-Hermitian, and $\Omega^2$ is Hermitian. For the model we are investigating, the matrices $K$ and $\Omega^2$ actually turn out to be real, and therefore we assume that this is the case also in the present discussion.
We choose to perform the quantization in terms of the fields $\Delta_i$: $$\Delta_i \equiv {\cal D}_{ij} a_j + {\cal D}_{ij}^* a_j^\dagger \;\;\;\;\;\;,\;\;\;\;\;\;
\left[ a_i \left( \vec{k} \right) , a_j ^\dagger \left( \vec{p} \right) \right] = \delta^{(3)} \left( \vec{k} - \vec{p} \right) \delta_{ij} \;\;.
\label{D-deco}$$ To perform the quantization, besides these relations, we need to impose the equal time commutation relations (ETCR) between $\Delta_i$ and its conjugate momentum in (\[S23-formal\]): $$\left[ \Delta_i \left( t , \vec{x} \right) , \Pi_j \left( t , \vec{y} \right) \right] = i \delta_{ij} \delta^{(3) } \left( \vec{x} - \vec{y} \right)
\;\;\;\;\;\;,\;\;\;\;\;\;
\Pi_i \equiv \frac{\partial L}{\partial \Delta_i'} \;\;,
\label{conjugate}$$ where, with an abuse of notation, in this relation (and only in this relation) the fields are in real space. Decomposing $\Pi$ in terms of the same annihilation / creation operators as in (\[D-deco\]), we have $$\Pi_i = \pi_{ij} a_j + \pi_{ij}^* a_j^\dagger \;\;\;,\;\;\; \pi_{ij} = {\cal D}_{ij}' + K_{il} {\cal D}_{lj} \;\;.$$
The relations (\[D-deco\]) and (\[conjugate\]) can be simultaneously imposed only if the condition, $$\left[ {\cal D} \pi^\dagger - {\cal D}^* \pi^T \right]_{ij} = i \, \delta_{ij} \;\;,
\label{wronskian}$$ is satisfied. Notice that we can impose (\[wronskian\]) as an initial condition; however, consistency of the computation requires that this condition holds at all times. This is the case if the initial conditions also satisfy $$\pi \pi^\dagger - \pi^* \pi^T = {\cal D} {\cal D}^\dagger - {\cal D}^* {\cal D}^T = 0 \;\;,
\label{wronskian2}$$ namely if the products $\pi \pi^\dagger $ and ${\cal D} {\cal D}^\dagger$ are real. Indeed, the three conditions in (\[wronskian\]) and (\[wronskian2\]) are preserved by the equations of motion if they all hold initially: to verify this, one can write the equations of motion following from (\[S23-formal\]) in terms of ${\cal D}$ and $\pi$. Using these equations, one can then show that the first derivatives of (\[wronskian\]) and (\[wronskian2\]) vanish if these relations hold. Therefore, if these relations hold initially, they are preserved by the evolution, and continue to hold at all times.
The condition (\[wronskian\]) generalizes to the multi-field case the standard Wronskian condition imposed by the single field quantization. The conditions (\[wronskian2\]) instead vanish identically in the single field case, and have no counterpart in this case. We will collectively denote the three conditions (\[wronskian\]) and (\[wronskian2\]) as Wronskian conditions in the remainder of this work.
In the explicit computations below, we show that, in the initial sub-horizon regime, the modes can be chosen according to the positive frequency initial adiabatic vacuum prescription, and satisfying the conditions (\[wronskian\]) and (\[wronskian2\]). Starting from these initial conditions, we then compute the time evolution for the mode functions and obtain the solution at late time, when the mode is outside the horizon. The solutions provide the late time correlators of any observable of our interest. Specifically, the observables we are interested in can be written as linear combinations of the dynamical variables and their time derivatives: $${\cal O } \left( t , \vec{x} \right) = \int \frac{d^3 k}{\left( 2 \pi \right)^{3/2} } \, {\rm e}^{i \vec{k} \cdot \vec{x}} \,
{\cal O } \left( t , \vec{k} \right) \;\;,\;\;
{\cal O } \left( t , \vec{k} \right) = c_i \left( t ,\, \vec{k} \right) X_i \left( t ,\, \vec{k} \right) + d_i \left( t ,\, \vec{k} \right) X_i' \left( t ,\, \vec{k} \right)
\;\;.
\label{observable}$$ where the reality of ${\cal O}$ enforces that ${\cal O }^* \left( \vec{k} \right) = {\cal O} \left( - \vec{k} \right)$. We can therefore decompose also ${\cal O}$ in terms of annihilation / creation operators $$\begin{aligned}
{\cal O } & = & {\cal O}_i \left( \vec{k} \right) a_i \left( \vec{k} \right) + {\cal O}^*_i \left( - \vec{k} \right) a_i \left( - \vec{k} \right)^\dagger \;\;, \nonumber\\
{\cal O}_i & = & c_j {\cal M}_{jl} {\cal D}_{li} + d_j \left( {\cal M}_{jl} {\cal D}_{li} \right)' \;\;.
\label{cal-O}\end{aligned}$$ and we obtain the correlator $$\langle {\cal O } \left( t , \vec{x} \right) {\cal O } \left( t , \vec{y} \right) \rangle = \int \frac{d^3 k}{\left( 2 \pi \right)^3}
{\rm e}^{i \vec{k} \cdot \left( \vec{x} - \vec{y} \right)} \, \sum_i \left\vert {\cal O}_i \left( \vec{k} \right) \right\vert^2$$
Assuming statistical isotropy, $ \left\vert {\cal O}_i \left( \vec{k} \right) \right\vert^2 = \left\vert {\cal O}_i \left( k \right) \right\vert^2$, and the correlator can be written in terms of the isotropic power spectrum $$\begin{aligned}
& &
\langle {\cal O } \left( t , \vec{x} \right) {\cal O } \left( t , \vec{y} \right) \rangle = \int \frac{d k}{k} \, \frac{\sin \left( k \, r \right)}{k \, r} \, P_{{\cal O}{\cal O}} \left( k \right) \;\;\;,\;\;\; r \equiv \vert \vec{x} - \vec{y} \vert \;\;, \nonumber\\
& &
P_{{\cal O}{\cal O}} \left( k \right) = \frac{k^3}{2 \pi^2} \sum_i \vert {\cal O}_i \vert^2
\label{PS-OO}\end{aligned}$$
We conclude this Subsection by noting a symmetry of the above solutions. Specifically, if we replace $${\cal D} \rightarrow {\cal D} \, U \;\;,
\label{phase-arbit}$$ where $U$ is a constant and unitary matrix, and we perform the same transformation on ${\cal D}'$ and ${\cal \pi}$, the conditions (\[wronskian\]) and (\[wronskian2\]), as well as the final power spectrum (\[PS-OO\]) are unchanged. The matrix $U$ is arbitrary and unphysical; this generalizes to the $N$-field case the usual phase arbitrarily of the mode function in the single field case. We use this arbitrariness in Subsection \[subsec:initial\] to simplify the initial conditions.
Tensor modes {#sec:tensor}
============
The tensor sector of Gauge-flation is given in eq. (\[modes\]). It is actually convenient to work in terms of the left-handed and right-handed canonical modes, related to the fields given in (\[modes\]) through $$h_+ = \frac{h_L + h_R}{\sqrt{2}} \;\;,\;\;
h_\times = \frac{h_L - h_R}{i \sqrt{2}} \;\;,\;\;
t_+ = \frac{t_L + t_R}{\sqrt{2}} \;\;,\;\;
t_\times = \frac{t_L-t_R}{ i \sqrt{2}} \;\;.
\label{LR-def}$$
We insert these modes in the action of the model, and expand to quadratic order. The action splits in two decoupled parts, $$S_{\rm tensor } = S_L + S_R \;\;,$$ so that the doublet $\left\{ h_L ,\, t_L \right\}$ is the counterpart of $\left\{ X \right\}_i$ - defined above eq. (\[S2-formal\]) - in the left-handed sector (and analogously for the right-handed sector). We recall that there are no non-dynamical tensor modes.
The modes in (\[LR-def\]) are not canonically normalized. Canonical normalization is achieved through $$h_L = \frac{\sqrt{2}}{M_p a} \, H_L \;\; , \;\;
t_L = \frac{1}{\sqrt{2} a} \, T_L
\label{canonical-tensor}$$ (and analogously in the right-handed sector). This diagonalization is the explicit expression for (\[XtoD\]), and leads to an action of the form $$S_L = \frac{1}{2} \int d \tau d^3 k \left[ \Delta_L^{'\dagger} \Delta_L'
+ \Delta_L^{'\dagger} K_L \Delta_L
- \Delta_L^\dagger K_L \Delta_L'
- \Delta_L^\dagger \Omega_L^2 \Delta_L \right] \;\;,\;\;
\Delta_L = \left( \begin{array}{c} H_L \\ T_L \end{array} \right) \;\;,
\label{tensor-S}$$ and identically for the right-handed sector. The matrix $K_L$ is anti-symmetric, with $$K_{L,12} = \frac{1}{M_p} \left( Q' + \frac{a'}{a} Q \right) \;\;,
\label{KL}$$ while the $\Omega_L^2$ matrix is symmetric, with $$\begin{aligned}
\Omega_{L,11}^2 & = & k^2 - 2 \frac{a'^2}{a^2} + \frac{3 g^2 a^2 Q^4}{M_p^2} - \frac{\left( a Q \right)^{'2}}{M_p^2 a^2} \;\;, \nonumber\\
\Omega_{L,12}^2 & = & k \frac{2 g a Q^2}{M_p} + \frac{\left( a Q \right)'}{a M_p} \frac{a'}{a} - \frac{2 \kappa g^2 Q^3}{M_p a^2} \frac{g^2 a^4 Q^4 + a^{'2} Q^2 - a^2 Q^{'2}}{1+\kappa g^2 Q^4} \;\;, \nonumber\\
\Omega_{L,22}^2 & = & k^2 - 2 k g a Q \left[ 1 + \kappa \frac{g^2 a^4 Q^4 + a^{' 2} Q^2 - a^2 Q^{' 2} }{a^4 \left( 1 + \kappa g^2 Q^4 \right) } \right] + \frac{2 \kappa g^2 Q^2}{ a^2} \frac{g^2 a^4 Q^4 + a^{'2} Q^2 - a^2 Q^{'2}}{1+\kappa g^2 Q^4} \;\;.
\label{omegaL}\end{aligned}$$ The matrices $K_R$ and $\Omega_R^2$ are obtained from $K_L$ and $\Omega_L^2$, respectively, by replacing $k \rightarrow - k$. This causes a difference between the two helicities, signaling a violation of parity invariance in the tensor sector, analogously to what happens for the model of Chromo-natural inflation [@Adshead:2013qp].
To gain an analytical understanding of the solutions in the tensor sector, we first of all rewrite the above matrix elements eliminating as many parameters as possible (using some background relations), and then we approximate them in slow roll. Specifically, we first of all perform the following substitutions on each matrix element: $$\begin{aligned}
&& Q' \rightarrow - a Q H \delta \;\;,\;\; a' \rightarrow a^2 H \;\;,\;\; k \rightarrow p \, a \;\;,\;\;
\kappa \rightarrow \frac{1}{H^2 \gamma Q^2} \; \frac{\left( 1 - \delta \right)^2 + \gamma}{\left( 1 - \delta \right)^2} \, \frac{2-\epsilon}{\epsilon} \;\;, \nonumber\\
&& M_p \rightarrow Q \sqrt{\frac{\left( 1 - \delta \right)^2 + \gamma}{\epsilon}}
\;\;,\;\; g \rightarrow \sqrt{\gamma} \frac{H}{Q} \;\;.
\label{subs-before-slowroll}
\end{aligned}$$ Even if they employ slow roll parameters, all these substitutions are exact. The first substitution follows from the definition of $\delta$ in eq. (\[s2seven\]), the second and third substitutions are standard relations between comoving/conformal and physical quantities ($p$ is the physical momentum of a mode), while the last three substitutions follow directly from eqs. (\[relation-kappa\]), (\[relation-Mp\]), and (\[def-gamma\]), respectively. From these substitutions, we obtain: $$\begin{aligned}
\frac{ K_{L,12} }{a} & = & H \frac{\left( 1 - \delta \right) \sqrt{\epsilon}}{\sqrt{\gamma+\left( 1 - \delta \right)^2} } \;\;, \nonumber\\
\frac{ \Omega_{L,11}^2 }{a^2} & = & p^2 - H^2 \frac{\left( 1 - \delta \right)^2 \left( 2 + \epsilon \right) + \gamma \left( 2 - 3 \epsilon \right)}{\gamma + \left( 1 - \delta \right)^2} \;\; , \nonumber\\
\frac{ \Omega_{L,12}^2 }{a^2} & = & \frac{ H p \, 2 \sqrt{\gamma \epsilon}}{\sqrt{\gamma + \left( 1 - \delta \right)^2} } -
\frac{ H^2 \sqrt{\epsilon}}{\sqrt{\gamma+\left( 1 - \delta \right)^2} } \frac{2 \gamma^2 \left( 2 - \epsilon \right) + 3 \gamma \left( 1 - \delta \right) \left( 2 - \epsilon \right) + 2 \left( 1 - \delta \right)^3 \left[ 1 + \delta \left( 2 - \epsilon \right) - \epsilon \right]}{2 \left( 1 - \delta \right)^2 + \gamma \left( 2 - \epsilon \right) } \;\; , \nonumber\\
\frac{ \Omega_{L,22}^2 }{a^2} & = & p^2 - H p \; \frac{2 }{\sqrt{\gamma}} \frac{2 \gamma^2 \left( 2 - \epsilon \right) + \left( 1 - \delta \right)^3 \left( 1 + \delta \right) \left( 2 - \epsilon \right) + 2 \gamma \left( 1 - \delta \right) \left( 3 - \delta - \epsilon \right)}{2 \left( 1 - \delta \right)^3 + \gamma \left( 2 - \epsilon \right)} \nonumber\\
& & \quad \quad
+ H^2 \; 2 \left( 2 - \epsilon \right) \frac{\gamma^2 + 2 \gamma \left( 1 - \delta \right) + \left( 1 - \delta \right)^3 \left( 1 + \delta \right)}{2 \left( 1 - \delta \right)^2 + \gamma \left( 2 - \epsilon \right) } \;\;.
\label{tensor-substituted}\end{aligned}$$ We stress that these expressions are completely equivalent to the corresponding ones in eqs. (\[KL\]) and (\[omegaL\]) and that no slow roll approximation has yet been done. However, the expressions in (\[tensor-substituted\]) are much more transparent, as the only dimensional parameters are the physical momentum $p$ and he physical Hubble rate $H$, while any explicit reference to $Q$ has been eliminated. Besides $H$ and $p$, the expressions (\[tensor-substituted\]) are given in terms of three dimensionless parameters: the two slow-roll parameters $\epsilon$ and $\delta$, and the parameter $\gamma$.
All expressions in (\[tensor-substituted\]) have a power law dependence on $p$ and $H$ of the type $K_{12} \propto c_1 \, H \;,\; \Omega_{ij}^2 = c_2 p^2 + c_3 p H + c_4 H^2$. Therefore these expressions immediately indicate which terms dominate in the sub-horizon and in the super-horizon regime. All the $c_i$ coefficients are slowly evolving, and can be immediately expanded in slow roll. To perform the expansion, we $${\rm Substitute \;\; } \delta \rightarrow \frac{\gamma}{6 \left( 1 + \gamma \right)} \epsilon^2 \;\;\;\; {\rm and \; expand \; in \;\; } \epsilon \ll 1 \;\;,
\label{subs-slowroll}$$ where the first expression is the slow roll solution for $\delta$, see eq. (\[sl2\]). The expansion leads to $$\begin{aligned}
\frac{ K_{L,12} }{a} & \simeq & H \frac{\sqrt{\epsilon}}{\sqrt{1+\gamma}} \;\;, \nonumber\\
\frac{ \Omega_{L,11}^2 }{a^2} & \simeq & p^2 - 2 H^2 \;\; , \nonumber\\
\frac{ \Omega_{L,12}^2 }{a^2} & \simeq & \frac{2 \sqrt{\gamma \epsilon}}{\sqrt{1+\gamma}} H p - \frac{\left( 1 + 2 \gamma \right) \sqrt{\epsilon} }{\sqrt{1+\gamma}} H^2 \;\; , \nonumber\\
\frac{ \Omega_{L,22}^2 }{a^2} & \simeq & p^2 - \frac{ 2 \left( 1 + 2 \gamma \right)}{\sqrt{\gamma}} \, p \, H + 2 \left( 1 + \gamma \right) \, H^2 \;\;.
\label{tensor-slow}\end{aligned}$$ We stress that, as we do not disregard any of the $c_i$ coefficients, but we simply expand them in slow roll, the expressions in (\[tensor-slow\]) provide a very accurate approximation to the exact expressions (\[tensor-substituted\]) at all times (namely, for any value of $p/H$). All of them agree with the exact expressions at all times, with an ${\rm O } \left( \epsilon \right)$ accuracy or better.
From the expressions (\[tensor-slow\]) we can gain an intuitive understanding of the tensor sector. We first of all note that, if it was not for the coupling with the $t_{L/R}$ modes from the gauge field, the tensor mode helicities would obey the standard relation $h_{L/R}'' + \left( k^2 - 2 H^2 a^2 \right) h_{L/R} \simeq 0$. However, for some choice of parameters the mode $t_L$ exhibits a large tachyonic growth for some time close to horizon crossing, and it can then source a large growth of $h_L$. The same does not occur in the right-handed sector. The situation is completely analogous to what takes place in Chromo-natural inflation [@Adshead:2013qp].
To understand why $t_L$ can grow for some choice of parameters, and why the same does not take place for $t_R$, we can disregard their coupling to the $h_{L/R}$ modes. This is an accurate approximation, since the coupling affects $t_{L/R}$ through the $K_{12}$ and $\Omega_{12}^2$ terms, that are slow roll suppressed with respect to the $\Omega_{22}^2$ term at all times. We see that $$\Omega_{L,22}^2 < 0 \;\;\;\; {\rm for } \;\;\; r_* - \Delta r < \frac{p}{H} < r_* + \Delta r \;\; , \;\;
{\rm with } \;\;\;\; r_* \equiv \frac{1+2 \gamma}{\sqrt{\gamma}} \;\;\; , \;\;\;
\Delta r \equiv \frac{\sqrt{1+2\gamma+2\gamma^2}}{\sqrt{\gamma} } \;\; .$$ Therefore, for any choice of $\gamma$, each mode of $t_L$ experiences a tachyonic instability in a neighborhood of $p/H = r_*$ (therefore, the tachyonic growth is absent at sufficiently small and sufficiently large scale, but always takes place close to horizon crossing). We recall that $ \Omega_{R,22}^2 $ is related to $ \Omega_{R,22}^2 $ by $p \rightarrow - p$, so that $ \Omega_{R,22}^2 $ is instead positive at all times, and there is no tachyonic growth for the right-handed modes. In the left-handed sector, for any fixed value of $H$, the most negative value of $\Omega_{L,22}^2$ takes place at $$\Omega_{L,22}^2 \Big\vert_{p=\frac{1+2 \gamma}{\sqrt{\gamma}} H } \simeq - H^2 \; \frac{1+2\gamma+2\gamma^2}{\gamma} \;\;.$$
Both the minimum value of $\Omega_{L,22}^2 $ and the duration of the tachyonic phase increase both at large and small $\gamma$. However, as we shall see in the next Section, the region $\gamma < 2$ is excluded as the scalar perturbations are unstable there. For $\gamma > 2$, we see that the instability grows with $\gamma$, and so we should expect that too large values of $\gamma$ are excluded because the growth of $t_L$ will source too large a growth of $h_L$. In Section \[sec:pheno\] we show that this is indeed the case, and that the Planck constraint $r \lta 0.11$ [@Ade:2013zuv] forces $\gamma \lta 5$ (the precise value depending on the number of e-folds of inflation).
The actions $S_{L/R} \rightarrow \frac{1}{2} \int d \tau d^3 k \left[ \vert \Delta_{L/R,i}' \vert^2 - k^2 \vert \Delta_{L/R,i} \vert^2 \right]$ at asymptotically early times. This gives the initial conditions $$\sqrt{2 k } {\cal D}_{L,{\rm in}} = \mathbb{I} \;\; , \;\;
\sqrt{2 k } {\cal D}_{L,{\rm in}}' = - i k \, \mathbb{I} \;\; ,$$ where $ \mathbb{I} $ is the identity operator, $ \mathbb{I}_{ij} = \delta_{ij} $.
Starting from these initial conditions, we numerically integrate the equations following from (\[tensor-S\]) and (\[D-deco\]): $${\tilde \partial }^2 \left[ \sqrt{2 k} {\cal D}_L \right] +
\left[ 2 \frac{\sqrt{\kappa } M_p K_L}{a} + \frac{{\tilde \partial } a}{a} \right] {\tilde \partial } \left[ \sqrt{2 k} {\cal D}_L \right]
+ \left[ \frac{ {\tilde \partial } \left( \sqrt{\kappa} M_p K_L \right) }{a} + \frac{\kappa M_p^2 \Omega_L^2}{a^2} \right] \left[ \sqrt{2 k} {\cal D}_L \right] = 0 \;\;.
\label{eom-DL}$$ (and identically for the right-handed sector) and obtain the power $$P_{L/R} = \frac{k^3}{2 \pi^2} \frac{2}{M_p^2 a^2} \frac{1}{2 k} \sum_{i=1}^2 \left\vert \sqrt{2 k} {\cal D}_{L/R,1i} \right\vert^2 = \frac{1}{\kappa M_p^4} \; \frac{ {\tilde p}^2 }{2 \pi^2} \sum_{i=1}^2 \left\vert \sqrt{2 k} {\cal D}_{L/R,1i} \right\vert^2 \;\;,
\label{PLR}$$ where in the second expression we defined the dimensionless physical momentum $${\tilde p} \equiv \sqrt{\kappa} M_p p \;\; .
\label{p-tilde}$$
Eq. (\[eom-DL\]) has been written in terms of the dimensionless quantities that we use in our numerical integration: namely, the quantities defined in (\[tilde-combo\]) and ${\tilde p}$. We recall that ${\tilde \partial}$ denotes derivative with respect to the dimensionless physical time ${\tilde t}$, related to the derivative with respect to conformal time $\tau$ by $\frac{\partial }{\partial \tau } = \frac{a}{\sqrt{\kappa} M_p } \, {\tilde \partial} $. In this way, all the matrix elements appearing in (\[eom-DL\]) can be expressed solely in terms of dimensionless quantities, and are “ready” for the numerical integration. For instance, $$\frac{\sqrt{\kappa } M_p K_{L,12}}{a} = {\tilde \partial } {\tilde Q } + \frac{{\tilde \partial } a}{a} {\tilde Q} \;\; ,$$ and analogously for all the other matrix elements. We recall that $ \frac{{\tilde \partial } a}{a} $ is the Hubble rate in rescaled physical time.
In Figure \[fig:PLPR\], we show the time evolution of the power $P_L$ and $P_R$ for a mode that leaves the horizon at $60$ e-folds before the end of inflation. In the left panel and right panel we show the evolution for $\gamma_{\rm in} = 3$ and $\gamma_{\rm in} = 10$, respectively. In both cases, the initial values of the two powers coincide at early times (since the early time actions in the left-handed and right-handed sectors are identical), but then $P_L > P_R$ at late times, signaling the breaking of parity in the tensor sector, and the tachyonic growth in the left-handed sector that we have also observed analytically. As discussed above, the final value of $P_L$ grows with growing $\gamma$.
![ Growth of the power and freeze out for the left-handed and right-handed gravity waves, for two different values of $\gamma_{60} = 3$ (left panel) and for $\gamma_{60} = 10$ (right panel), where $\gamma_{60}$ is the value $\gamma_{\rm in}$ at $60$ e-folds before the end of inflation. Both panel show the evolution of the power for a mode that leaves the horizon $60$ e-folds before the end of inflation. The $x-$axis is the ratio between the Hubble horizon and the wavelength of the mode, and therefore it is a measure of time. Horizon crossing occurs at $a H / k = 1$.[]{data-label="fig:PLPR"}](tenpower-small "fig:"){width="45.00000%"} ![ Growth of the power and freeze out for the left-handed and right-handed gravity waves, for two different values of $\gamma_{60} = 3$ (left panel) and for $\gamma_{60} = 10$ (right panel), where $\gamma_{60}$ is the value $\gamma_{\rm in}$ at $60$ e-folds before the end of inflation. Both panel show the evolution of the power for a mode that leaves the horizon $60$ e-folds before the end of inflation. The $x-$axis is the ratio between the Hubble horizon and the wavelength of the mode, and therefore it is a measure of time. Horizon crossing occurs at $a H / k = 1$.[]{data-label="fig:PLPR"}](tenpower-large "fig:"){width="45.00000%"}
Scalar modes {#sec:scalar}
============
In this Section we study the scalar system of perturbations of Gauge-flation in the gauge (\[modes\]). The discussion is divided in several parts. In Subsection \[subsec:S2scal-explicit\] we provide the explicit form of the quadratic action of the perturbations. In Subsection \[subsec:initial\] we provide the initial conditions imposed by this action and by the conditions (\[wronskian\]) and (\[wronskian2\]). In Subsection \[subsec:scalarsol\] we study the numerical evolution of the equations of motion following from this action, and we present some examples.
The quadratic action {#subsec:S2scal-explicit}
--------------------
Introducing the scalar modes in (\[modes\]) into the action of Gauge-flation, and expanding up to second order, we obtain $$\begin{aligned}
S_{\rm scalar} & = & \int d \tau d^3 k \, {\cal L}_{\rm scalar} \;\;, \nonumber\\
{\cal L}_{\rm scalar} & = &
\frac{a^2 \left( 1 + g^2 \kappa Q^4 \right)}{6} \left\vert 3 \delta Q' - k^2 M' \right\vert^2 + \frac{k^4 a^2}{3} \left\vert M' \right\vert^2 - \frac{a^2}{3} \Bigg\{ \frac{k^2}{3} + \frac{g^2 \kappa Q^4\left( 1 + g^2 \kappa Q^4 \right) }{2 a^2 M_p^2 } \left( a Q \right)^{' 2} \nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
+ \frac{1}{1+g^2 \kappa Q^4} \left[ g^2 Q^2 \left( 3 - g^2 \kappa Q^4 \right) \left( a^2 + \kappa Q^2 \frac{a'^2}{a^2} + \kappa Q^{'2} \right) - 8 g^2 \kappa Q^4 \frac{a'^2}{a^2} \right] \Bigg\} \left\vert 3 \delta Q - k^2 M \right\vert^2
\nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
- \frac{k^4 a^2}{3} \left( \delta Q^* M + {\rm h.c.} \right)
+ \frac{a^2 k^4}{3} \left[ \frac{k^2}{3} - \frac{g^2 \kappa Q^4}{a^2 M_p^2 } \left( a Q \right)^{' 2}
- \frac{2 g^2 \kappa Q^2}{1+g^2 \kappa Q^4} \left( g^2 a^2 Q^4 + \frac{a^{'2}}{a^2} Q^2 - Q^{'2} \right) \right] \left\vert M \right\vert^2 \nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
+ \Bigg\{ \frac{a^2 k^2}{6} \Bigg[ \left( 1 + g^2 \kappa Q^4 \right) \left( 3 \delta Q^{'*} - k^2 M^{'*} \right) - 2 k^2 M^{'*} + \frac{1}{a} \left[ \left( 1 + g^2 \kappa Q^4 \right) a' + 2 g^2 \kappa Q^3 \left( a Q \right)' \right] \left( 3 \delta Q^* - k^2 M^* \right)
\nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
- 2 k^2 \frac{a'}{a} M^* \Bigg] Y + {\rm h. c. } \Bigg\} + k^2 a^2 \left[ g^2 a^2 Q^2 + \frac{k^2}{6} \left( 3 + g^2 \kappa Q^4 \right) \right] \left\vert Y \right\vert^2
+ \frac{a}{2} \left( 1 + g^2 \kappa Q^4 \right) \left( a Q \right)' \left[ \phi^* \left( 3 \delta Q' - k^2 M' \right) + {\rm h. c. } \right]
\nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
+ \left[ \frac{a' \left( 1 + g^2 \kappa Q^4 \right)}{2} \left( a Q \right)' + g^2 \kappa Q^3 \left( a Q \right)^{' 2} + g^2 a^4 Q^3 \right] \left[ \phi^* \left( 3 \delta Q - k^2 M \right) + {\rm h. c. } \right]
+ \frac{a k^2 \left( 1 + g^2 \kappa Q^4 \right) }{2} \left( a Q \right)' \left( Y^* \phi + {\rm h.c.} \right) \nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
- a k^2 \left( a Q \right)' \left[ B^* \delta Q + {\rm h. c. } \right] - g^2 k^2 a^4 Q^3 \left( Y^* B + {\rm h.c.} \right)
+ \frac{3}{4} \left[ 3 \left( 1 + g^2 \kappa Q^4 \right) \left( a Q \right)^{'2} + g^2 a^4 Q^4 - 6 M_p^2 a^{'2} \right] \left\vert \phi \right\vert^2 \nonumber\\
& & \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\! \!\!\!\!
+ \frac{k^2}{4} \left[ g^2 a^4 Q^4 + 6 M_p^2 a^{'2} - 3 \left( 1 + g^2 \kappa Q^4 \right) \left( a Q \right)^{'2} \right] \left\vert B \right\vert^2 - k^2 M_p^2 a a' \left( B^* \phi + {\rm h. c. } \right) \;\;.
\label{s2sca-explicit}\end{aligned}$$ which is indeed of the form (\[S2-formal\]) in terms of the dynamical modes $X_i = \left\{ \delta Q , M \right\}_i$ and of the non-dynamical modes $N_i = \left\{ Y , \phi , B \right\}_i$. We integrated out the non-dynamical modes as outlined in Subsection \[subsec:formal-deco\], leading to an action of the form (\[S22-formal\]) in terms of the dynamical modes only.
We transformed the two dynamical fields according to eq. (\[XtoD\]), with $$X_i = {\cal M}_{ij} \Delta_j \;\;\;,\;\;\;
{\cal M} = \left( \begin{array}{cc}
- \frac{1}{\sqrt{6} a } \sqrt{ 1 + \frac{2}{1+\kappa g^2 Q^4}} & 0 \\
- \sqrt{\frac{3}{2}} \frac{1}{a k^2} \sqrt{1-\frac{2}{3+\kappa g^2 Q^4}} &
- \frac{1}{a k^2} \sqrt{\frac{k^2}{2 a^2 g^2 Q^2 } + \frac{3}{3+\kappa g^2 Q^4}}
\end{array} \right) \;\;.
\label{XtoD-explicit}$$ The resulting action is of the form $$S = \frac{1}{2} \int d \tau d^3 k \left[ \Delta^{' \dagger} T \Delta' + \Delta^{' \dagger} K \Delta - \Delta^\dagger K \Delta' - \Delta^\dagger \Omega^2 \Delta \right] \;\;,\;\;
\label{S24-formal}$$ which is nearly of the form (\[S23-formal\]), with the difference that the matrix $T$ is not the identity in this expression.
Formally speaking, it is straightforward to set the matrix $T$ equal to the identity for an action of the type (\[S24-formal\]), through a second field redefinition, and to include this second redefinition in the transformation matrix ${\cal M}$. However, the explicit form of the matrix ${\cal M}$ would be extremely involved for the problem at hand. Fortunately, this additional step is not necessary. Indeed, the action (\[S24-formal\]) is needed for two purposes. The first purpose is to obtain the evolution equations for the mode functions. Performing the decomposition (\[D-deco\]), the equations of motion following from (\[S24-formal\]) are $$\begin{aligned}
&& {\cal D}'' + \alpha \, {\cal D}' + \beta \, {\cal D} = 0 \;\;, \nonumber\\
&&
\alpha \equiv T^{-1} \left( T' + 2 K \right) \;\;,\;\;
\beta \equiv T^{-1} \left( K' + \Omega^2 \right) \;\;,
\label{eom-formal}\end{aligned}$$ which can be integrated for any invertible $T$ (we verified analytically that $T$ is indeed invertible). The second purpose is to impose the initial conditions. Fortunately (as we explicitly show below) the matrix $T \simeq 1$ with extremely good accuracy in the early time / sub-horizon regime. We can therefore simply disregard the departure of $T$ from the identity matrix in the early time regime, and perform all the steps outlined in Subsection \[subsec:PS\]. The initial conditions obtained in this way are extremely accurate. Finally, the expressions (\[cal-O\]) - (\[PS-OO\]) for the observables do not require that $T = 1$. Therefore, we do not need to explicitly perform the transformation that sets $T$ to unity.
The procedure to obtain the matrices $T,K,\Omega^2$ in (\[S24-formal\]) is a straightforward algebraic procedure, that we have outlined in Subsection \[subsec:formal-deco\]. The action (\[s2sca-explicit\]) is of the form (\[S2-formal\]), and by comparing these two expressions we can immediately read off the explicit expressions of the matrices $A, \dots , F$. We therefore also have the explicit expression for the action in terms of the dynamical modes, see eq. (\[S22-formal\]). It is conceptually straightforward to perform the transformation (\[XtoD-explicit\]) in this action, and obtain the explicit expressions for $T,K,\Omega^2$. This leads to an expression for the matrix elements that is the scalar-sector counterpart of the expressions (\[KL\]) and (\[omegaL\]) that we had obtained for the tensor sector. However, these expressions (that we obtained using Mathematica) are extremely much longer and more involved than those in the tensor sector, and their full form is not particularly illuminating. For this reason we do not report them here. All entries in these matrices are formally of the type [^8]
$$\frac{\sum_i c_i p^{\alpha_i} }{\sum_j d_j p^{\alpha_j} }
\times \sqrt{ \frac{\sum_m {\tilde c}_m p^{\alpha_m} }{\sum_n {\tilde d}_n p^{\alpha_n } } } \;\;,
\label{TKOM-formal}$$
where the sums are finite sums, and where the coefficients $c_i, d_j, {\tilde c}_m , {\tilde d}_n$ are slowly evolving functions of time.
We expanded each of these coefficients in slow roll, performing the same two steps (\[subs-before-slowroll\]) and (\[subs-slowroll\]) that we performed in the tensor sector. This leads to the expressions (\[slowT\]), (\[slowK\]), (\[slowOm\]) that we report in Appendix \[app-scalaraction\], and that are the scalar-sector counterpart of eqs. (\[tensor-slow\]). We stress that the expressions given in the Appendix retain all the coefficients of (\[TKOM-formal\]), and approximate each of them in slow roll. Therefore, these expressions are very accurate at all times (namely for all values of $p/H$). These are the expressions that we used to integrate the equations (\[eom-formal\]) numerically.
We now discuss the asymptotic limits of the matrices $T,K,\Omega^2$ in the early time / sub-horizon and late-time / super-horizon regime. At early times, these evaluate to $$\begin{aligned}
p \gg H :
&& T_{11} \simeq 1 + \frac{3 \epsilon^2 \gamma}{\left( 1 + \gamma \right)^2} \frac{H^2}{p^2} \;\;,\;\;
T_{12} \simeq - \frac{\sqrt{3 \gamma} \epsilon}{1+\gamma} \frac{H}{p} \;\;,\;\;
T_{22} \simeq 1 + \frac{3 \epsilon \gamma}{ 1 + \gamma } \frac{H^2}{p^2} \;\;, \nonumber\\
&& \frac{K_{12}}{a} \simeq - \frac{p}{\sqrt{3 \gamma}} \;\;, \nonumber\\
&& \frac{\Omega_{11}^2}{a^2} \simeq \frac{p^2}{3} \;\;,\;\;
\frac{\Omega_{12}^2}{a^2} \simeq \frac{2 \left( 1 + \gamma \right)}{\sqrt{3 \gamma}} p H \;\;,\;\;
\frac{\Omega_{22}^2}{a^2} \simeq \left( 1 - \frac{2}{\gamma} \right) p^2 \;\;,
\label{earlyTKOm}
\end{aligned}$$ where we recall that $T$ and $\Omega^2$ are symmetric, while $K$ is anti-symmetric. These expressions have two properties worth noting. Firstly, the asymptotic form of $\Omega^2$ shows that the scalar perturbations have a strong instability in the sub-horizon regime for $\gamma < 2$. We discuss this more in details in Subsections \[subsec:initial\] and \[subsec:scalarsol\]. Secondly, we see that the matrix $T$ indeed coincides with the identity in this regime, up to terms suppressed both by $H/p$ and by slow roll. The initial conditions are given by setting the modes in the initial adiabatic vacuum. This requires for instance that $\dot{\Omega} \ll \Omega^2$. We see from (\[earlyTKOm\]) that $\frac{\dot{\Omega} }{\Omega^2 } = {\rm O } \left( \frac{H}{p} \right)$, while $T- \mathbb{I} =
{\rm O } \left( \epsilon \, \frac{H}{p} \right)$. Therefore it is completely safe to approximate $T$ with the identity at the initial time.
However, at horizon crossing and after it, $T$ departs from the identity. More in general, at late times we have $$\begin{aligned}
p \ll H :
&& T_{11} \simeq 1 + \frac{2}{\gamma} \;\;,\;\;
T_{12} \simeq - \frac{2 \sqrt{1+\gamma}}{\gamma \sqrt{\epsilon}} \;\;,\;\;
T_{22} \simeq 1 + 2 \frac{1+\gamma}{\gamma \epsilon} \;\;, \nonumber\\
&& \frac{K_{12}}{a} \simeq \frac{2 \epsilon^{3/2} H}{3 \sqrt{1+\gamma}} \;\;, \nonumber\\
&& \frac{\Omega_{11}^2}{a^2} \simeq \frac{2 \left( - 2 + \gamma + \gamma^2 \right)}{\gamma} H^2 \;\;,\;\;
\frac{\Omega_{12}^2}{a^2} \simeq \frac{4 \sqrt{1+\gamma}}{\gamma \sqrt{\epsilon}} H^2 \;\;,\;\;
\frac{\Omega_{22}^2}{a^2} \simeq - \frac{4 \left( 1 + \gamma \right)}{\gamma \epsilon} H^2 \;\;.
\label{lateTKOm}
\end{aligned}$$ The asymptotic limits (\[earlyTKOm\]) and (\[lateTKOm\]) are significantly simpler than the expressions (\[slowT\]), (\[slowK\]), (\[slowOm\]). However, they are not accurate at horizon crossing, and for this reason the expressions of Appendix \[app-scalaraction\] (which are valid at all times) are used in the numerical integration of (\[eom-formal\]).
We conclude this Subsection with a note on the effect of the metric perturbations on the scalar system. To study their impact, we (i) artificially removed them from (\[s2sca-explicit\]), (ii) integrated out the only remaining non-dynamical mode $Y$, and (iii) performed the transformation (\[XtoD-explicit\]). The resulting action is of the form (\[S24-formal\]) with $T= \mathbb{I} $ (this is true exactly, and not just in the slow roll or early/late time approximation). Therefore the transformation (\[XtoD-explicit\]) is the one that sets the kinetic matrix to unity in absence of metric perturbations. As it is always the case in slow roll inflation, the effect of metric perturbations is negligible in the sub-horizon regime, and indeed we could verify that the action obtained by the steps (i)-(ii)-(iii) coincides with the asymptotic early times expressions (\[earlyTKOm\]) to leading order in slow roll.
Initial adiabatic solution {#subsec:initial}
--------------------------
In this Subsection we first of all obtain an analytic approximate but very accurate solution for the equation (\[eom-formal\]) in the early time regime. This solution is characterized by some integration constants that we fix by imposing that the solution is initially in the positive frequency adiabatic vacuum, with the conditions (\[wronskian\]) and (\[wronskian2\]) respected. This provides the initial conditions for the perturbations, that we use in Subsection \[subsec:scalarsol\] when we numerically integrate the equation (\[eom-formal\]) at all times.
From the asymptotic early time expressions (\[earlyTKOm\]) we obtain $$\begin{aligned}
\frac{k}{a} \gg H :
&& \alpha =
\left( \begin{array}{cc}
0 & - \frac{2 k}{\sqrt{3 \gamma}} \\
\frac{2 k}{\sqrt{3 \gamma}} & 0
\end{array} \right) + {\rm O } \left( a H \right) \;\;,\;\;
\beta = \left( \begin{array}{cc}
\frac{k^2}{3} & 0 \\
0 & \frac{\gamma-2}{\gamma} k^2 \end{array} \right) + {\rm O } \left( a H k \right) \;\;,
\label{alpha-beta-early}
\end{aligned}$$ for the two matrices that characterize the evolution equation (\[eom-formal\]).
If we keep only the terms explicitly given in (\[alpha-beta-early\]), we can guess that eq. (\[eom-formal\]) should admit a solution characterized by $ \frac{\cal D'}{\cal D} = {\rm O } \left( k \right) \;, \frac{\cal D''}{\cal D} = {\rm O } \left( k^2 \right) $. Indeed, such a solution is $$\begin{aligned}
{\cal D}^{\rm early}_{1j} & = & c_{1j} {\rm e}^{-i k \tau} + c_{2j} {\rm e}^{i k \tau} + c_{3j} {\rm e}^{-i \frac{\sqrt{\gamma-2} k \tau}{\sqrt{3 \gamma}}} + c_{4j} {\rm e}^{i \frac{\sqrt{\gamma-2} k \tau}{\sqrt{3 \gamma}}} \;\;, \nonumber\\
{\cal D}^{\rm early}_{2j} & = & - \frac{i \sqrt{\gamma}}{\sqrt{3}} \left[ c_{1j} {\rm e}^{-i k \tau} - c_{2j} {\rm e}^{i k \tau} \right]
+ \frac{i}{\sqrt{\gamma-2}} \left[ c_{3j} {\rm e}^{-i \frac{\sqrt{\gamma-2} k \tau}{\sqrt{3 \gamma}}} - c_{4j} {\rm e}^{i \frac{\sqrt{\gamma-2} k \tau}{\sqrt{3 \gamma}}} \right] \;\;, \;\; j=1,2 \;\;,
\label{early-sol-1}\end{aligned}$$ where $c_{ij}$ are integration constants. It is straightforward to verify that (\[early-sol-1\]) indeed satisfies (\[eom-formal\]) if only the terms explicitly given in (\[alpha-beta-early\]) are retained. It is also straightforward to verify that, once the early time solution is inserted in (\[eom-formal\]), the subdominant terms in (\[alpha-beta-early\]) contribute to (\[eom-formal\]) with terms that are suppressed by at least ${\rm O } \left( \frac{a H}{k} \right) \ll 1$ factors with respect to the contributions that we include. This confirms that (\[early-sol-1\]) solves the evolution equations in the early time / sub-horizon regime. Moreover, we can see that the system (\[eom-formal\]) should admit $8$ complex integration constants. They are precisely given by the coefficients $c_{ij}$ in (\[early-sol-1\]). Therefore, the expression (\[early-sol-1\]) is the most general early time solution of eq. (\[eom-formal\]).
As we anticipated in Subsection \[subsec:S2scal-explicit\], one linear combination of the modes $\propto c_{3j}, c_{4j}$ is exponentially growing in the solution (\[early-sol-1\]) for $\gamma < 2$, signaling a tachyonic instability. In general, a tachyonic instability is not an intrinsic pathology of a model, but “simply” an instability of a given background solution. If the instability is mild enough, it does not lead to any problem (as it is the case for the tachyonic instability in the left-handed tensor sector at sufficiently small $\gamma_{\rm in}$, see Section \[sec:tensor\]). However, the present instability at $\gamma < 2$ is extremely strong. It takes place all throughout the sub-horizon regime, with a rate $\propto \frac{k}{a H}$, which is arbitrarily large the deeper inside the horizon a mode is. At any given moment, all modes in the deep sub-horizon regime are highly unstable.
In principle, the instability would be absent if we could set to zero the coefficient of the linear combination that is exponentially growing in (\[early-sol-1\]) at $\gamma < 2$. Doing so, however, we verified by direct inspection that the first Wronskian condition (\[wronskian\]) cannot be satisfied for any $\gamma < 2$ (the different matrix components in that relation give incompatible requirements on the integration constants). Therefore, already from this condition alone, we learn that it is not possible to avoid the instability for $\gamma < 2$. Condition (\[wronskian\]) is obtained from the initial quantization. One may be tempted to simply postulate that quantization is impossible in that regime, and so set that coefficient to zero by hand. This would however be a completely unjustified and acausal choice, as there is no causal reason why all sub-horizon modes should have at some given initial moment $t_{\rm in}$ a classical initial condition that avoids the tachyonic growth taking place at $t > t_{\rm in}$. It would be equivalent to stating that the model $V= \frac{1}{2} m^2 \varphi^2$, with a large and negative $m^2$, is stable because, classically, the choice $\delta \varphi = 0$ can be made. Thus, we conclude that the background solution obtained in Section \[sec:model\] is invalid for $\gamma < 2$.
We note that an instability inside the horizon takes place for some choice of parameters takes place also in the related model of Chromo-Natural inflation in [@Dimastrogiovanni:2012ew]. In that case the instability occurs for a finite range of $\frac{k}{a H}$ inside the horizon. We see that the instability in Gauge-flation is even stronger. However, both instabilities have the effect of invalidating the range of parameters in which they take place.
Let us now study the stable $\gamma > 2$ region. As standard in inflation, we impose that the negative frequency terms in (\[early-sol-1\]) are initially vanishing, $c_{2j} = c_{4j} = 0$. This corresponds to no quanta present at the initial time. [^9] After imposing this, from a direct inspection we obtain that the conditions (\[wronskian\]) and (\[wronskian2\]) are satisfied [^10] if and only if $$\begin{aligned}
\vert c_{11} \vert^2 + \vert c_{12} \vert^2 = \frac{3}{2 k \left( 1 + \gamma \right)} \;\;,\;\;
\vert c_{31} \vert^2 + \vert c_{32} \vert^2 = \frac{\sqrt{3 \gamma} \sqrt{\gamma-2}}{2 k \left( 1 + \gamma \right)} \;\;,\;\;
c_{11} c_{31}^* + c_{12} c_{32}^* = 0 \;\;.
\label{sol-combo}\end{aligned}$$
These relations do not uniquely specify the remaining coefficients $c_{1i}$ and $c_{3i}$. This however was expected: as we discussed in Subsection \[subsec:PS\], if the matrix ${\cal D}$ satisfies the Wronskian conditions, then the product ${\cal D} \times U$, where $U$ is unitary, also satisfies them. Under a unitary transformation, the coefficients $c_{1i}$ and $c_{3i}$ transform as $$c_{1j} \rightarrow c_{1k} \, U_{kj} \;\;,\;\; c_{3j} \rightarrow c_{3k} \, U_{kj} \;\;,
\label{U2}$$ and it is immediate to observe that the combinations in (\[sol-combo\]) are left invariant by these transformations. As we discussed in Subsection \[subsec:PS\], the matrix $U$ is arbitrary, but unphysical, and it generalizes to the $N$ fields case the well known phase arbitrarity of the wave function of the single field quantization. There are $8$ real parameters in the $c_{1j},c_{3j}$ coefficients, and eqs. (\[sol-combo\]) are $4$ real constraints. The space of allowed solutions is therefore a $4-$dimensional real space, which is the same dimension of the U(2) arbitrarity in (\[U2\]). Therefore, we simply need to provide one solution of (\[sol-combo\]), with the understanding that the other solutions will be related to it by the arbitrary and unphysical transformation (\[U2\]). An immediate solution of (\[sol-combo\]) is provided by $$c_{11} = \sqrt{\frac{3}{1+\gamma}} \, \frac{1}{\sqrt{2 k}} \;\;,\;\; c_{12} = 0 \;\;,\;\;
c_{31} = 0 \;\;,\;\; c_{32} = \frac{\left( 3 \gamma \right)^{1/4} \left( \gamma - 2 \right)^{1/4}}{\sqrt{1+\gamma}} \, \frac{1}{\sqrt{2 k}} \;\;.$$
Inserting this solution in (\[early-sol-1\]), and setting $\tau = 0$ as our initial time (changing the value of the initial time in (\[early-sol-1\]) also corresponds to an unphysical U(2) transformation of ${\cal D}$), we obtain the initial conditions for ${\cal D}$ and ${\cal D}'$: $$\sqrt{2 k} \, {\cal D}^{\rm in} =
\left( \begin{array}{cc}
\sqrt{\frac{3}{1+\gamma}} & \frac{\left( 3 \gamma \right)^{1/4} \left( \gamma - 2 \right)^{1/4}}{\sqrt{1+\gamma}} \\
- \frac{i\sqrt{\gamma}}{\sqrt{1+\gamma}} & \frac{i\left( 3 \gamma \right)^{1/4} }{ \left( \gamma - 2 \right)^{1/4} \sqrt{1+\gamma}}
\end{array} \right) \;\;,\;\;
\sqrt{2 k} {\cal D}^{'{\rm in} } = - i \, k \left( \begin{array}{cc}
\sqrt{\frac{3}{1+\gamma}} & \frac{ \left( \gamma - 2 \right)^{3/4}}{ \left( 3 \gamma \right)^{1/4} \sqrt{1+\gamma}} \\
- \frac{i\sqrt{\gamma}}{\sqrt{1+\gamma}} & \frac{i \left( \gamma - 2 \right)^{1/4}}{ \left( 3 \gamma \right)^{1/4} \sqrt{1+\gamma}}
\end{array} \right) \;\;.
\label{Din}$$
These quantities are the initial conditions for the numerical integration of the evolution equation (\[eom-formal\]) the we perform in the next Subsection.
Solution at all times, and power of $\zeta$ {#subsec:scalarsol}
--------------------------------------------
We solve for the variables $\sqrt{2 k} {\cal D}_{ij}$. Starting from the initial conditions (\[Din\]), we numerically integrate the equations of motion following what is formally written in (\[eom-formal\]), where the explicit expressions for the matrix elements are given in Appendix \[app-scalaraction\]. We recall that these expressions are accurate in slow roll at all times (namely for all values of the ratio $p/H$). We integrate these equations after writing them in physical time, and in rescaled dimensionless quantities (\[tilde-combo\]). The equations are solely written in terms of these rescaled quantities and of the rescaled physical momenta (\[p-tilde\]).
From the solutions, we compute the curvature perturbation on uniform density hypersurfaces, $\zeta$, that, in the spatially flat gauge that we are using, is given by $$\zeta \equiv - \frac{H}{\dot{\rho}} \delta \rho \;\; .
\label{def-zeta}$$
Combining the expressions (\[Pzeta\]) and (\[coeffzeta-super\]), we obtain the super-horizon relation between the power of $\zeta$ and the $\sqrt{2 k} {\cal D}_{ij}$ solutions: $$\begin{aligned}
\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\! P_\zeta \simeq \frac{1}{\kappa M_p^4} \, \frac{\left( 1 + \gamma \right)^2 {\tilde p}^2}{6 \pi^2 \gamma^2 \epsilon^4 } \sum_i \!\!\!\! & & \!\!\!\! \Bigg\vert - \frac{\sqrt{\epsilon}}{\sqrt{1+\gamma}} \left( 2 \sqrt{2 k } {\cal D}_{1i} + \frac{{\tilde \partial } \left[ \sqrt{2 k } {\cal D}_{1i} \right] }{ { \tilde \partial } a / a } \right) \nonumber\\
& & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad + 2 \sqrt{2 k} {\cal D}_{2i} + \frac{{\tilde \partial } \left[ \sqrt{2 k } {\cal D}_{2i} \right] }{ {\tilde \partial } a / a } \Bigg\vert^2 \;\;,\;\; p \ll H \;\;.
\label{PzetaD}\end{aligned}$$ We verified numerically that this expression is accurate only $ \sim 10 $ e-folds after horizon crossing, due to the fact that the super horizon approximated results (\[coeffzeta-super\]) have been used. In the left panel of Figure \[fig:Pzeta-kappa\] we show the time evolution of the power for a mode that leaves the horizon $60$ e-folds before the end of inflation, for two different choices of the parameter $\gamma$ evaluated at $60$ e-folds before the end of inflation. In this Figure and in the ones presented in the next Section, the expressions (\[coeffzeta-all\]), rather than (\[coeffzeta-super\]), have been used. The expressions (\[coeffzeta-all\]) are valid at all times. The expressions (\[coeffzeta-super\]) are however significantly simpler, and accurate from $ \sim 10 $ e-folds after horizon crossing onwards.
In the left panel of Figure \[fig:Pzeta-kappa\], and in Figures \[fig:PLPR\], the power is given in units of $\frac{1}{\kappa M_p^4}$. The coefficient $\kappa$ can be finally obtained by imposing the power spectrum normalization $P_\zeta = 2.2 \cdot 10^{-9}$ [@Ade:2013zuv] for the large scale modes leaving the horizon either $N=60$ or $N=50$ e-folds before the end of inflation (according to the value specified in each plot shown). In the right panel of Figure \[fig:Pzeta-kappa\] we show the resulting value of $\kappa^{1/4} M_p$ (this is a dimensionless quantity) as a function of $\gamma_{\rm in}$.
![ Left panel: time evolution of the power of $\zeta$ for a mode that leaves the horizon $60$ e-folds before the end of inflation, and for two different values of the variable $\gamma$ evaluated at that time (namely when $a H / k = 1$ in the Figure). Right panel: value of the dimensionless quantity $\kappa^{1/4} M_p$ as a function of the initial value of $\gamma$, for $N=50$ and $N=60$ e-folds of inflation. This value is obtained by imposing the observed normalization $P_\zeta = 2.2 \cdot 10^{-9}$ [@Ade:2013zuv] for the large scale modes leaving the horizon either $N=60$ or $N=50$ e-folds before the end of inflation. []{data-label="fig:Pzeta-kappa"}](scapower "fig:"){width="45.00000%"} ![ Left panel: time evolution of the power of $\zeta$ for a mode that leaves the horizon $60$ e-folds before the end of inflation, and for two different values of the variable $\gamma$ evaluated at that time (namely when $a H / k = 1$ in the Figure). Right panel: value of the dimensionless quantity $\kappa^{1/4} M_p$ as a function of the initial value of $\gamma$, for $N=50$ and $N=60$ e-folds of inflation. This value is obtained by imposing the observed normalization $P_\zeta = 2.2 \cdot 10^{-9}$ [@Ade:2013zuv] for the large scale modes leaving the horizon either $N=60$ or $N=50$ e-folds before the end of inflation. []{data-label="fig:Pzeta-kappa"}](kappa-COBE "fig:"){width="45.00000%"}
Phenomenology {#sec:pheno}
=============
Let us now discuss the phenomenological implications of the comparison between the results obtained in the previous two Sections for the tensor and scalar modes and the Planck results [@Ade:2013zuv]. Figure \[fig:Pzeta\] shows the power spectrum $P_\zeta$ for $N=60$ e-folds of inflation and for two different values of $\gamma$ at that moment.
For each value of $k$ that has been used in the Figure, we show the value of $P_\zeta$ obtained at $3$ e-folds after the horizon crossing; in each case we verified that by that time the power had already saturated to its final freeze out value (this can also be seen in the example shown in Figures \[fig:PLPR\] and \[fig:Pzeta-kappa\]; we recall that $3$ efolds corresponds to $a H / k \simeq 20$). We show the spectrum for a limited range of momenta, namely from $k=k_{60}$ (defined as the mode that leaves the horizon precisely $60$ e-folds of inflation) to $k = 10^3 k_{60}$. However, given the mild (and smooth) scale dependence of $P_\zeta$ (typical of slow-roll inflation), the range of momenta shown is enough to determine the value of the spectral tilt $n_s$, defined [^11] as $P_\zeta \propto k^{n_s-1}$. We see that the spectrum is more red at the smaller value of $\gamma_{\rm in}$ shown.
![ Power spectrum of $\zeta$ after freeze out. $k_{60}$ corresponds to the comoving momentum of a mode leaving the horizon $60$ e-folds before the end of inflation. We show the spectrum obtained for two different values of $\gamma$ at that moment. []{data-label="fig:Pzeta"}](scalarPS){width="45.00000%"}
Figure \[fig:ns-r\] is the main result of this work. The left panel shows the spectral index $n_s$ obtained for $N=50$ and $N=60$ e-folds of inflation, as a function of the initial value of $\gamma$ (evaluated at $50$ or $60$ e-folds before the end of inflation, respectively). Also the $95 \%$ CL Planck limits [@Ade:2013zuv] are shown: $0. 9457 < n_s < 0.9749$. The comparison excludes at $95 \%$ CL all values $\gamma_{\rm in} \lta 13.5$ for $N=50$ and all values $\gamma_{\rm in} \lta 9.3$ for $N=60$. [^12] The right panel of the Figure shows the value of the tensor-to-scalar ratio $$r \equiv \frac{ P_L + P_R }{ P_\zeta } \;\; ,
\label{def-r}$$ as a function of the initial value of $\gamma$. The $95 \%$ CL Planck limit [@Ade:2013zuv] $r < 0.11$, rules out all values $\gamma_{\rm in} \gta 4.8$ for $N=50$ and $\gamma_{\rm in} \gta 5$ for $N=60$ [^13]
![ Spectral index $n_s$ and tensor-to-scalar ratio $r$ obtained in Gauge-flation for two different values of the number of e-folds $N$ and of the initial value of $\gamma$ (we recall that these two parameters completely characterize the model and the background evolution). Also shown are the $95 \%$ CL Planck limits [@Ade:2013zuv] $0. 9457 < n_s < 0.9749$ and $r < 0.11$. No value of these parameters is compatible with these limits, so that Gauge-flation is phenomenologically ruled out. []{data-label="fig:ns-r"}](ns-vs-gamma "fig:"){width="45.00000%"} ![ Spectral index $n_s$ and tensor-to-scalar ratio $r$ obtained in Gauge-flation for two different values of the number of e-folds $N$ and of the initial value of $\gamma$ (we recall that these two parameters completely characterize the model and the background evolution). Also shown are the $95 \%$ CL Planck limits [@Ade:2013zuv] $0. 9457 < n_s < 0.9749$ and $r < 0.11$. No value of these parameters is compatible with these limits, so that Gauge-flation is phenomenologically ruled out. []{data-label="fig:ns-r"}](r-vs-gamma "fig:"){width="45.00000%"}
We see that no value of the parameters leads to both a sufficiently flat spectrum, and sufficiently small tensor modes. We therefore conclude that Gauge-flation is ruled out by the CMB results. We employed the Planck limits given in the abstract of [@Ade:2013zuv], that correspond to a specific choice of priors. We can however see that the interval of $\gamma_{\rm in}$ allowed by the $n_s$ bound is rather far from the interval allowed by the $r$ bound, so that a small change of the allowed intervals (due to the different priors studied in [@Ade:2013zuv]) does not impact our conclusion.
Comparison with previous results {#sec:comparison}
================================
Refs. [@Maleknejad:2011jw] also performed a study of cosmological perturbations in Gauge-flation. To be concrete, all our remarks in this Section refer to the most recent version (namely, v$5$ posted on the archive) of the second paper in [@Maleknejad:2011jw]. Ref. [@Maleknejad:2011jw] reaches different conclusions from ours in the scalar sector. This crucially affects the phenomenological study of the model. Their computation is in terms of gauge invariant variables, and it is rather different from the procedure that we have outlined in Section \[sec:linpert\]. Therefore, a direct comparison between our and their procedure is not straightforward. We refrain from commenting on their algebra. However, we can point out one crucial assumption and one crucial simplification done in [@Maleknejad:2011jw], which we believe invalidate their conclusions, and contribute to the disagreement between our and their results.
A crucial assumption of [@Maleknejad:2011jw], which we believe to be incorrect is in the choice of initial conditions for $\gamma < 2 $. We have seen that $\gamma < 2$ leads to a strong tachyonic instability. Ref. [@Maleknejad:2011jw] essentially removes the instability by setting to zero the coefficient of the exponentially large mode. Specifically, ref. [@Maleknejad:2011jw] works in terms of the two modes ${\cal Q}$ and $\Phi$ (loosely speaking, they are the gauge invariant counterparts of our modes $\delta Q$ and $\phi$). They decompose these modes into ${\cal Q} = {\cal Q}_1+ {\cal Q}_2$ and $\Phi = \Phi_1+ \Phi_2$, and each ${\cal Q}_i$ and $\Phi_i$ obeys a decoupled equation (their eqs. (V.68) and (V.69)) at asymptotically early times. The mode ${\cal Q}_1$ is related to $\Phi_1$ (and analogously for ${\cal Q}_2$ and $\Phi_2$) as explained in footnote $7$, so that this decomposition is in reality a rotation that diagonalizes the early times equations of motion. The eigenfrequencies coincide with those in our eq. (\[early-sol-1\]), and the eigenfrequencies of the modes ${\cal Q}_2$ and $\Phi_2$ are tachyonic for $\gamma < 2$.
Eq. (V.88) of [@Maleknejad:2011jw] is the solution for the mode ${\cal Q}_2$, and eq. (V.82) is the solution for $\Phi_2$. It is claimed in [@Maleknejad:2011jw] that these are good solutions at all times, (we disagree on this, as we discuss below), but let us here only discuss the early time expansion of these solutions. The first line of (V.88) is the solution for $\gamma <2$, while the second line is the solution for $\gamma > 2$. Ref. [@Maleknejad:2011jw] performs $2$ steps to obtain the initial conditions in these solutions.
Step 1: the coefficients $q_2$ and ${\tilde q}_2$ in the first line of (V.88) are related to those in the second line, by stating that they are chosen “such that in both cases, ${\cal Q}_2$ has the same super horizon value.”
Step 2: The coefficient ${\tilde q}_2$ is set to zero based on the fact that, in the second line , this is the coefficient of the negative frequency mode.
Due to the identification in step 1, the choice done in step $2$ also sets ${\tilde q}_2$ to zero in the first line of (V.88). This coefficients multiplies the exponentially large mode at early times, and so this choice effectively removes the instability.
Step 2 is correct. For $\gamma > 2$, there is no tachyonic instability, and the standard quantization indeed sets to zero the negative frequency modes (we are speaking of negative frequency, not negative squared frequency). This is analogous to setting $c_{2j} = c_{4j}= 0$ in our eq. (\[early-sol-1\]), and it is a standard choice of the vacuum in the sub-horizon regime. On the other hand, step 1 cannot be accepted. First of all, this step is mathematically correct only for ${\tilde q}_2 = 0$. If this is not the case, the first and second line of (V.88) do not coincide in the super-horizon regime (this is not clarified in [@Maleknejad:2011jw]). Moreover step 1 assumes that the solution (V.88) is valid at all times (as we discuss below, we disagree on this). These are however mathematical issues which are not the core of the problem. The real problem with step 1 is that there is no physical reason why the initial conditions (in the sub-horizon regime) for one choice of the parameters should be given in such a way that they lead to a solution that, in the super-horizon regime, coincides with the solution obtained for another choice of parameters. Namely, according to [@Maleknejad:2011jw], if the background solution of Gauge-flation has $\gamma_{\rm in} = 1$, then the sub-horizon value for the scalar perturbations should be given such that the solution, once evolved to much later times, should coincide with the scalar perturbations obtained for a background solution with $\gamma_{\rm in} = 3$. For $N=60$ e-folds of inflation, the choice $\gamma_{\rm in} = 1$ corresponds to roughly $Q_{\rm in} \simeq 0.076 M_p$ and $\sqrt{\kappa} g \simeq 3,200 / M_p^2$, while the choice $\gamma_{\rm in} = 3$ corresponds to roughly $Q_{\rm in} \simeq 0.049 M_p$ and $\sqrt{\kappa} g \simeq 12,000 / M_p^2$. Namely, the two cases are characterized by completely different values of the parameters and of the background initial conditions, and there is no physical motivation for taking information from one case to set the initial conditions for the perturbations in another case. It is as unmotivated as for example having the model $V=\frac{1}{2} m^2 \varphi^2$, and setting the initial perturbations for $\delta \varphi$ in the case $\varphi_{\rm in} = 1, m^2 = -2$, based on the fact that they should agree at late times with the perturbations for $\varphi_{\rm in} = 3, m^2 = +4$.
This is the reason why ref. [@Maleknejad:2011jw] concludes that the $\gamma < 2$ region is stable. We have shown that this choice is inconsistent with a proper quantization, and physically unmotivated.
Also in the stable $\gamma >2$ region our results in the scalar sector disagree with those of [@Maleknejad:2011jw]. Ref. [@Maleknejad:2011jw] extends the decomposition ${\cal Q} = {\cal Q}_1 + {\cal Q}_2$ and $\Phi = \Phi_1 + \Phi_2$ also to the late time regime. This decomposition must be supplemented by two relations between these variables, not to artificially increase the number of degrees of freedom. These relations are given by footnote $7$ for the early time regime only. However, ref. [@Maleknejad:2011jw] continues to use these modes for all the evolution, without providing a clear and explicit definition of how the ${\cal Q}_i$ and $\Phi_i$ modes are defined after the asymptotic early time regime. Maybe, one could argue that the solutions themselves encode how the modes are related to each other, and, although a clear definition of these quantities (prior the solutions are given) would be useful, perhaps this is just an issue of presentation of the results. Let us therefore disregard this issue, and rather let us discuss more in details the results given in [@Maleknejad:2011jw] for $\gamma > 2$. These are given by their eqs. (V.81)-(V.82)-(V.87)-(V.88), and it is claimed in [@Maleknejad:2011jw] that these results are the solutions for ${\cal Q}_i$ and $\Phi_i$ at all times. The problem with this statement is that the equations solved by (V.81)-(V.82)-(V.87)-(V.88) are not the actual linearized equations for the perturbations of Gauge-flation. Specifically (V.81)-(V.82)-(V.87)-(V.88) are the solutions of equations (V.77)-(V.78)-(V.83)-(V.84). This set of equations is constructed so to coincide at early times with the early time equations derived in [@Maleknejad:2011jw] - these are equations (V.68) and (V.69) - and so to coincide at late times with the late time equations derived in [@Maleknejad:2011jw] - these are equations (V.75) and (V.76). Therefore, already in the derivation of ref. [@Maleknejad:2011jw], eqs. (V.77)-(V.78)-(V.83)-(V.84) are not presented as the real equations for the perturbations, but as equations that coincide with the real ones only in the deep sub-horizon and super-horizon regimes. In eqs. (V.77)-(V.78)-(V.83)-(V.84) the variables ${\cal Q}_i$ and $\Phi_i$ are fully decoupled. While - according to the derivation of [@Maleknejad:2011jw] - these variables are decoupled in the deep sub- and super-horizon regimes, this is certainly not the case at horizon crossing. Therefore, assuming that eqs. (V.77)-(V.78)-(V.83)-(V.84) are valid at all times trivializes all the dynamics, particularly in the most crucial times when the mixing may be expected to be mostly relevant (or, at the very least, where no $p \gg H$ nor $p \ll H$ approximation can be made). Ref. [@Maleknejad:2011jw] does not study the accuracy of (V.77)-(V.78)-(V.83)-(V.84), neither by comparing their solutions with the ones that can be obtained by numerically solving the real system of equations, nor by performing an analytic comparison of this set of equations with the real ones.
We stress that all entries in the equations that we have solved numerically coincide (at leading order in slow roll) with the exact ones at all times (for all values of $p/H$). Our results disagree with those of [@Maleknejad:2011jw]. This, together with the lack of a proper justification of eqs. (V.77)-(V.78)-(V.83)-(V.84), leads us to the conclusion that this set of equations does not correctly characterize the dynamics of the scalar perturbations of the model.
Conclusions {#sec:conclusions}
===========
Gauge-flation has the interesting peculiarity that it is a model of inflation driven by a vector field without ghost instabilities. It has been shown that the model is related to a specific limit of Chromo-natural inflation [@SheikhJabbari:2012qf]. This identification has been shown to be accurate at the background level [@Adshead:2012qe; @SheikhJabbari:2012qf]. However, as we discussed in the Introduction, it may be possible that integrating out the axion of Chromo-natural inflation leads to appreciable differences between the perturbations in the two models. This possibility appears to be corroborated by the existing literature. The perturbations of Chromo-natural inflation have been studied by a number of recent works [@Dimastrogiovanni:2012st; @Dimastrogiovanni:2012ew; @Adshead:2013qp; @Adshead:2013nka]. Their results agree with each other when they overlap: the model is stable only if the vector field is sufficiently heavy, which is encoded by the condition (\[cn-instab\]). In the stable region, there is no value of parameters leading to sufficiently flat scalar spectrum and sufficiently small tensor modes. Perturbations in Gauge-flation were studied in [@Maleknejad:2011jw], which did not report any unstable region, and which obtained a spectrum of scalar perturbations significantly bluer than those in Chromo-natural inflation.
Our results for the perturbations in Gauge-flation disagree with those of [@Maleknejad:2011jw], and confirm that the identification [@SheikhJabbari:2012qf] between the two models is accurate also at the perturbative level. Firstly, we have seen that the scalar perturbations of Gauge-flation are highly unstable when $\gamma \equiv \frac{g^2 Q^2}{H^2} < 2$. This condition coincides with the result obtained for Chromo-natural inflation [@Dimastrogiovanni:2012ew], see eq. (\[cn-instab\]). The instability was discussed in [@Dimastrogiovanni:2012ew] for the Chromo-natural formulation of the model: it is first of all not an instability of the model, but “only” of the inflationary background solution. A background instability typically manifests itself through the presence of tachyonic modes, and this is also the case for these two models. The instability can be most easily explained in the Chromo-natural formulation [@Dimastrogiovanni:2012ew]: for any given value of $H$, the axion potential is too steep to drive inflation by itself. Chromo-natural inflation realizes a slow-roll axion evolution by coupling it with a gauge field. The mechanism requires that the gauge field is non-Abelian ($g\neq 0$) and it has a vev ($Q \neq 0$), and so - a posteriori - it is not surprising that, for any given $H$, the mechanism cannot work at arbitrarily small $g \, Q$.
Secondly, let us compare the results for the perturbations in the two models in the stable $\gamma > 2$ region. For Gauge-flation, as shown in Fig. 6, both the tilt of the scalar power spectrum ($n_{s}$) and the tensor-to-scalar ratio ($r$) are growing functions of $\gamma_{\rm in}$. We find that the $95\%$ CL Planck limits [@Ade:2013uln] on the spectral index ($0.9457 < n_{s} < 0.9749$) requires at $95\%$ CL $\gamma_{\rm in} \geq 13.5$ for $N = 50$ and $\gamma_{\rm in} \geq 9.3$ for $N = 60$. The $95\%$ CL Planck limit on the tensor-to-scalar ratio $r < 0.11$, requires instead $\gamma_{\rm in}\leq 4.8$ for $N = 50$ and $\gamma_{\rm in}\leq 5$ for $N = 60$. For Chromo-natural inflation, we refer to the parameter scan performed in ref. [@Adshead:2013nka]. As shown in their Figure 12, the Planck bounds on the scalar spectral tilt translate into the limits $6.5 \lta \gamma_{\rm in} \lta 16$ for a number of e-folds between $50$ and $60$. Ref. [@Adshead:2013nka] does not provide a direct plot (or equation) of $r$ as a function of $\gamma_{\rm in}$. However, we can approximately deduce a bound by comparing their Figures 12 and 13. From Figure 13, we see that $r$ is compatible with the Planck bound only when $n_s \lta 0.935$. From Figure 12 we see that this limit on $n_s$ translates into $\gamma_{\rm in} \lta 5$. We recall that the statement of equivalence between the two models is that Gauge-flation should correspond to a specific limit of Chromo-natural inflation. Therefore there is a large fraction of parameters in Chromo-natural inflation that do not correspond to Gauge-flation. Figures $12$ and $13$ of [@Adshead:2013nka] scan also over these parameters. So, if we could restrict the results in the two Figures to only those regions of parameters compatible with Gauge-flation, we may find more tighter bounds for $\gamma_{\rm in}$. Therefore our result $\gamma_{\rm in} \gta {\rm O } \left( 10 \right)$ from the spectral index is compatible with that obtained in Chromo-natural inflation, while the result obtained in [@Maleknejad:2011jw] is significantly bluer (see their Figure 8). Also the limits from $r$ obtained here are in good agreement with that of Chromo-natural inflation.
To conclude, we have applied the formalism that we have developed in our study [@Dimastrogiovanni:2012ew] of Chromo-natural inflation to the model of Gauge-flation. Gauge-flation was found to be equivalent to Chromo-natural inflation plus corrections [@SheikhJabbari:2012qf], and the corrections were shown to be negligible at the background level [@Adshead:2012qe; @SheikhJabbari:2012qf]. The situation was less clear at the perturbative level. Our study has shown that the analogy persists also at this level. Both models are unstable at $\gamma < 2$, and have too red a scalar spectrum at small $\gamma_{\rm in}$. In both cases, raising $\gamma_{\rm in} $ to a level that is compatible with the Planck bounds [@Ade:2013uln] on $n_s$ results in too large a gravity wave signal. Therefore, identically to what happens for Chromo-natural inflation, also Gauge-flation is ruled out by the CMB data.
[**Acknowledgements**]{}
We thank A. Maleknejad and M. M. Sheikh-Jabbari for correspondence on their work [@Maleknejad:2011jw] and on our Section \[sec:comparison\]. This work is supported in part by DOE grant DE-FG02-94ER-40823 at the University of Minnesota. MP would like to thank the University of Padova, INFN, Sezione di Padova, and the Cosmology Group at the Department of Theoretical Physics of the University of Geneva for their friendly hospitality and for partial support during his sabbatical leave.
Scalar action in slow roll approximation {#app-scalaraction}
========================================
Eq. (\[s2sca-explicit\]) is the action of the scalar perturbations of Gauge-flation, in the gauge discussed in Subsection \[subsesec:gauge\]. After integrating out the non-dynamical modes and performing the transformation (\[XtoD-explicit\]), we obtain an action of the form (\[S24-formal\]), where each element in the matrices $T,K,\Omega^2$ is of the form (\[TKOM-formal\]).
We compute the slow roll approximation of each coefficient entering in (\[TKOM-formal\]), by performing the two steps outlined in eqs. (\[subs-before-slowroll\]) and (\[subs-slowroll\]). We obtain $$\begin{aligned}
T_{11} & \simeq & 1 + \frac{6 \epsilon^2 \gamma H^2}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\;, \nonumber\\
T_{12} = T_{21} & \simeq & - 2 \sqrt{3} \epsilon \frac{\sqrt{\gamma \left( 1 + \gamma \right)} \sqrt{ \left( 1 + \gamma \right) p^2 + 3 \epsilon \gamma H^2} H}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\;, \nonumber\\
T_{22} & \simeq & 1 + \frac{6 \epsilon \gamma \left( 1 + \gamma \right) H^2}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\;,
\label{slowT}\end{aligned}$$ for the first matrix, $$\begin{aligned}
K_{11} = K_{22} & = & 0 \;\;, \nonumber\\
\frac{K_{12}}{a} = - \frac{K_{21}}{a} & \simeq & - \frac{2 \left( 1 + \gamma \right)^3 p^4 + 3 \epsilon \gamma \left( 1 + \gamma \right)^2 p^2 H^2 - 6 \epsilon^4 \gamma^3 H^4}{ \sqrt{3 \gamma \left( 1 + \gamma \right)} \sqrt{\left( 1 + \gamma \right) p^2 + 3 \epsilon \gamma H^2} \left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2 \right]} \;\;,
\label{slowK}\end{aligned}$$ for the second matrix, and $$\begin{aligned}
\frac{\Omega_{11}^2}{a^2} & \simeq &
\frac{p^2}{3} + 2 \left( 1 + \gamma \right) H^2 - \epsilon^2 \gamma H^2 \frac{4 \gamma \left( 1 + \gamma \right)^2 p^4-24 \left( 1 + \gamma \right)^2 p^2 H^2 + 36 \epsilon^2 \gamma^2 H^4}{\left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2 \right]^2} \;\;,
\nonumber\\
\frac{\Omega_{12}^2}{a^2} = \frac{\Omega_{21}^2}{a^2} & \simeq &
\frac{8\sqrt{1+\gamma}H}{\sqrt{3 \gamma} \sqrt{\left( 1 + \gamma \right) p^2 + 3 \epsilon \gamma H^2} \left\{ 4 \left( 1 + \gamma \right)^5 p^6 + 3 \epsilon \gamma \left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2 \right]^2 H^2 \right\}} \times \Bigg[ \left( 1 + \gamma \right)^6 p^8
\nonumber\\
& &
\!\!\!\!\!\! \!\!\!\!\!\!
+ 3 \epsilon \gamma \left( 1 + \gamma \right)^4 \left( 1 + 2 \gamma \right) p^6 H^2 + \frac{9}{4} \epsilon^2 \gamma^2 \left( 1 + \gamma \right)^3 \left( 4 \gamma - 5 \right) p^4 H^4 - 27 \epsilon^3 \gamma^3 \left( 1 + \gamma \right)^2 p^2 H^6 + \frac{81}{2} \epsilon^5 \gamma^5 H^8 \Bigg] \;\;,
\nonumber\\
\frac{\Omega_{22}^2}{a^2} \!\! \!\! & \simeq & \!\! \!\!
\frac{\gamma-2}{\gamma} p^2 + \epsilon H^2 \left( 1 + \gamma \right) \frac{ \left( \gamma - 6 \right) \left( 1 + \gamma \right)^3 p^6 + 9 \gamma \left( 1 + \gamma \right)^3 p^4 H^2 + 18 \epsilon \gamma^2 \left( 1 + \gamma \right)^2 p^2 H^4 - 27 \epsilon^3 \gamma^4 H^6}{ \left( 1 + \gamma \right)^5 p^6 + \frac{ 3 }{ 4 } \epsilon \gamma \left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2 \right]^2 H^2} \;\;, \nonumber\\
\label{slowOm}\end{aligned}$$ for the third matrix.
Expression for $\zeta$ in terms of ${\cal D}_{ij}$ {#app-zetaD}
==================================================
We are interested in the curvature perturbation on uniform density hypersurfaces $\zeta$, that, in the spatially flat gauge that we are using is given by $$\zeta \equiv - \frac{H}{\dot{\rho}} \delta \rho = \frac{\delta \rho}{6 \left[ \left( \dot{Q} + H Q \right)^2 + g^2 Q^4 \right] }
\simeq \frac{\delta \rho}{6 H^2 Q^2 \left( 1 + \gamma \right) } \;\;,$$ where the background equations have been used in the first equality, while the final expression is given at leading order in slow roll.
The variation of the energy is obtained by perturbing to linear order the definition $\rho \equiv - T^0_0$, and it reads $$\begin{aligned}
\delta \rho & = & 3 \frac{\left( a Q \right)'}{a^3} \left( 1 + \kappa g^2 Q^4 \right) \delta Q' - k^2 \frac{\left( a Q \right)'}{a^3} \left( 1 + \kappa g^2 Q^4 \right) M'
\nonumber\\
&& \!\!\!\!\!\!\!\!\!\!
+ 3 \left[ 2 \, g^2 Q^3 + \left( 1 + 3 \, \kappa g^2 Q^4 \right) \frac{a'^2}{a^4} Q + \left( 1 + 5 \, \kappa g^2 Q^4 \right) \frac{a' Q'}{a^3} +2 \, \kappa g^2 Q^3 \frac{Q'^2}{a^2} \right] \delta Q
\nonumber\\
&& \!\!\!\!\!\!\!\!\!\!
- k^2 \left[ 2 \, g^2 Q^3 + \left( 1 + 3 \, \kappa g^2 Q^4 \right) \frac{a'^2}{a^4} Q + \left( 1 + 5 \, \kappa g^2 Q^4 \right) \frac{a' Q'}{a^3} +2 \, \kappa g^2 Q^3 \frac{Q'^2}{a^2} \right] M
\nonumber\\
&& \!\!\!\!\!\!\!\!\!\!
+ k^2 \, \frac{\left( a Q \right)'}{a^3} \left( 1 + \kappa g^2 Q^4 \right) Y + 3 \, \frac{\left( a Q \right)'^2}{a^4} \left( 1 + \kappa g^2 Q^4 \right) \Phi \; .\end{aligned}$$ We express the non-dynamical variables $Y$ and $\Phi$ in terms of the dynamical variables. [^14] In this way, we express $\delta \rho$ as a linear combination of $\delta Q$, $M$, and their first derivative. We expand to leading order in slow roll the coefficients of this expansion, and we obtain $$\begin{aligned}
\delta \rho & \simeq &
\frac{4 \left( 1 + \gamma \right)^{3/2} H M_p \sqrt{\epsilon}}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \Bigg\{ 3 H \left[ \left( 3 + \gamma \right) p^2 + 9 \gamma H^2 \right] \delta Q
- a^2 p^2 H \left[ \left( 2 + \gamma \right) p^2 + 9 \gamma H^2 \right] M \nonumber\\
&& \quad\quad\quad\quad \quad\quad\quad\quad \quad\quad\quad\quad
+ \frac{3}{a} \left[ p^2 + 3 \gamma H^2 \right] \delta Q' - 3 a \gamma p^2 H^2 M' \Bigg\} \;\;. \end{aligned}$$
The variable $\zeta$ is our observable in the scalar sector, and we are interested in its power spectrum. When comparing with (\[observable\]), we have $X_1 = \delta Q $ and $X_2 = M$, and therefore $$\zeta = c_1 \, \delta Q + c_2 M + d_1 \delta Q' + d_2 M' \;\;,$$ with $$\begin{aligned}
& & c_1 \simeq
\frac{2 \left( 1 + \gamma \right)^{3/2} }{ M_p \sqrt{\epsilon} } \;
\frac{\left( 3 + \gamma \right) p^2 + 9 \gamma H^2}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\;\; , \;\;\;
c_2 \simeq
- 2 a^2 p^2 \frac{ \left( 1 + \gamma \right)^{3/2} }{ 3 M_p \sqrt{\epsilon} } \; \frac{\left( 2 + \gamma \right) p^2 + 9 \gamma H^2}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\; , \nonumber\\
& &
d_1 \simeq \frac{2 \left( 1 + \gamma \right)^{3/2}}{a H M_p \sqrt{\epsilon}} \; \frac{ p^2 + 3 \gamma H^2}{2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\;\; , \;\;\;
d_2 \simeq - \frac{2 a H \gamma \left( 1 + \gamma \right)^{3/2}}{M_p \sqrt{\epsilon}} \; \frac{p^2}{ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \epsilon^2 \gamma^2 H^2} \;\; . \end{aligned}$$
Using eqs. (\[cal-O\]) and (\[PS-OO\]) we can see that the correlator of $\zeta$ is related to the matrix elements ${\cal D}$ by $$P_{\zeta } = \frac{k^2}{4 \pi^2} \sum_{i} \left\vert \left( c_j {\cal M}_{jl} + d_j {\cal M}_{jl}' \right) \sqrt{2 k} {\cal D}_{li} + d_j {\cal M}_{jl} \sqrt{2 k} {\cal D}_{li}' \right\vert^2 \;\;.
\label{Pzeta}$$ which is the scalar-sector analogous of the expression (\[PLR\]).
We have $$\begin{aligned}
& & \!\!\!\!\!\!\!\!
c_j {\cal M}_{jl} + d_j {\cal M}_{jl}' =
\frac{\sqrt{2/3} \sqrt{1+\gamma}}{\left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \gamma^2 \epsilon^2 H^2 \right] a M_p \sqrt{\epsilon} }
\nonumber\\ & &
\quad\quad\quad\quad \times
\left\{ - \left[ \left( 2 + \gamma \right) \epsilon p^2 + 6 \epsilon \gamma H^2 \right] ,\,
\frac{\sqrt{1+\gamma}}{\sqrt{3 \gamma}} \frac{\left( 1 + \gamma \right) \left( 2 + \gamma \right) p^4 + 3 \gamma \left( 1 + \gamma \right) H^2 p^2 + 18 \gamma^2 \epsilon H^4}{H \sqrt{\left( 1 + \gamma \right) p^2 + 3 \gamma \epsilon H^2} } \right\}_l \;\;, \nonumber\\
& & \!\!\!\!\!\!\!\!
d_j {\cal M}_{jl} =
\frac{\sqrt{2/3} \sqrt{1+\gamma}}{\left[ 2 \left( 1 + \gamma \right)^2 p^2 + 3 \gamma^2 \epsilon^2 H^2 \right] H a^2 M_p \sqrt{\epsilon} }
\left\{ - \left[ \left( 1 + \gamma \right) p^2 + 3 \epsilon \gamma H^2 \right] ,\,
\sqrt{3 \gamma \left( 1 + \gamma \right) } H \sqrt{ \left( 1 + \gamma \right) p^2 + 3 \gamma \epsilon H^2 } \right\}_l \;\;.
\nonumber\\
\label{coeffzeta-all}\end{aligned}$$
In the super-horizon regime, these expressions approximate to $$\begin{aligned}
c_j {\cal M}_{jl} + d_j {\cal M}_{jl}' & \simeq & \frac{2 \sqrt{2} \left( 1+\gamma \right)}{\sqrt{3} \, \gamma} \frac{1}{ \epsilon^2 a M_p} \left\{ - \frac{\sqrt{\epsilon}}{\sqrt{1+\gamma} } , \, 1\right\}_l \;\; , \nonumber\\
d_j {\cal M}_{jl} & \simeq & \frac{1}{H a} \frac{ \sqrt{2} \left( 1+\gamma \right)}{\sqrt{3} \, \gamma} \frac{1}{ \epsilon^2 a M_p } \left\{ - \frac{\sqrt{\epsilon}}{\sqrt{1+\gamma}} , \, 1 \right\}_l \;\; , \;\; p \ll H \;\; .
\label{coeffzeta-super}
\end{aligned}$$ These are the expressions that we use in the main text in eq. (\[PzetaD\]). We verified numerically that these expressions provide the correct result for $\zeta$ starting from $\sim 10$ e-folds after horizon crossing. The freeze out of the power however occurs soon after horizon crossing, as can be observed by using the expressions (\[coeffzeta-all\]), which are valid at all times (these are the expressions used in the left panel of Figure \[fig:Pzeta-kappa\]).
[99]{}
A. D. Linde, Contemp. Concepts Phys. [**5**]{}, 1 (1990) \[hep-th/0503203\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], arXiv:1303.5082 \[astro-ph.CO\].
E. Dimastrogiovanni, N. Bartolo, S. Matarrese and A. Riotto, Adv. Astron. [**2010**]{}, 752670 (2010) \[arXiv:1001.4049 \[astro-ph.CO\]\].
K. Dimopoulos, Int. J. Mod. Phys. D [**21**]{}, 1250023 (2012) \[Erratum-ibid. D [**21**]{}, 1292003 (2012)\] \[arXiv:1107.2779 \[hep-ph\]\]. J. Soda, Class. Quant. Grav. [**29**]{}, 083001 (2012) \[arXiv:1201.6434 \[hep-th\]\]. A. Maleknejad, M. M. Sheikh-Jabbari and J. Soda, arXiv:1212.2921 \[hep-th\]. E. Pajer and M. Peloso, arXiv:1305.3557 \[hep-th\].
A. Maleknejad and M. M. Sheikh-Jabbari, arXiv:1102.1513 \[hep-ph\]. A. Maleknejad and M. M. Sheikh-Jabbari, Phys. Rev. D [**84**]{}, 043515 (2011) \[arXiv:1102.1932 \[hep-ph\]\].
E. Elizalde, A. J. Lopez-Revelles, S. D. Odintsov and S. Y. .Vernov, arXiv:1201.4302 \[hep-th\].
B. Himmetoglu, C. R. Contaldi and M. Peloso, Phys. Rev. Lett. [**102**]{}, 111301 (2009) \[arXiv:0809.2779 \[astro-ph\]\]; B. Himmetoglu, C. R. Contaldi and M. Peloso, Phys. Rev. D [**79**]{}, 063517 (2009) \[arXiv:0812.1231 \[astro-ph\]\]; B. Himmetoglu, C. R. Contaldi and M. Peloso, Phys. Rev. D [**80**]{}, 123530 (2009) \[arXiv:0909.3524 \[astro-ph.CO\]\].
A. Maleknejad, M. M. Sheikh-Jabbari and J. Soda, JCAP [**1201**]{}, 016 (2012) \[arXiv:1109.5573 \[hep-th\]\]. A. Ghalee, Phys. Lett. B [**717**]{}, 307 (2012) \[arXiv:1206.1650 \[gr-qc\]\]. M. Noorbala and M. M. Sheikh-Jabbari, arXiv:1208.2807 \[hep-ph\]. S. H. -S. Alexander, M. E. Peskin and M. M. Sheikh-Jabbari, Phys. Rev. Lett. [**96**]{}, 081301 (2006) \[hep-th/0403069\].
P. Adshead and M. Wyman, Phys. Rev. D [**86**]{}, 043530 (2012) \[arXiv:1203.2264 \[hep-th\]\]. M. M. Sheikh-Jabbari, Phys. Lett. B [**717**]{}, 6 (2012) \[arXiv:1203.2265 \[hep-th\]\].
P. Adshead and M. Wyman, Phys. Rev. Lett. [**108**]{}, 261302 (2012) \[arXiv:1202.2366 \[hep-th\]\]. E. Martinec, P. Adshead and M. Wyman, arXiv:1206.2889 \[hep-th\]. K. Freese, J. A. Frieman and A. V. Olinto, Phys. Rev. Lett. [**65**]{}, 3233 (1990). F. C. Adams, J. R. Bond, K. Freese, J. A. Frieman and A. V. Olinto, Phys. Rev. D [**47**]{}, 426 (1993) \[hep-ph/9207245\]. C. Savage, K. Freese and W. H. Kinney, Phys. Rev. D [**74**]{}, 123511 (2006) \[hep-ph/0609144\]. S. Ghigna, M. Lusignoli and M. Roncadelli, Phys. Lett. B [**283**]{} (1992) 278; R. Holman, S. D. H. Hsu, T. W. Kephart, E. W. Kolb, R. Watkins and L. M. Widrow, Phys. Lett. B [**282**]{} (1992) 132 \[hep-ph/9203206\]; M. Kamionkowski and J. March-Russell, Phys. Lett. B [**282**]{} (1992) 137 \[hep-th/9202003\]; S. B. Giddings and A. Strominger, Nucl. Phys. B [**306**]{} (1988) 890; S. B. Giddings and A. Strominger, “String Wormholes,” Phys. Lett. B [**230**]{} (1989) 46; R. Kallosh, A. D. Linde, D. A. Linde and L. Susskind, Phys. Rev. D [**52**]{}, 912 (1995) \[hep-th/9502069\].
T. Banks, M. Dine, P. J. Fox and E. Gorbatov, JCAP [**0306**]{}, 001 (2003) \[hep-th/0303252\]. P. Svrcek and E. Witten, “Axions in String Theory,” JHEP [**0606**]{} (2006) 051 \[hep-th/0605206\].
J. E. Kim, H. P. Nilles and M. Peloso, JCAP [**0501**]{}, 005 (2005) \[hep-ph/0409138\]; M. Berg, E. Pajer and S. Sjors, Phys. Rev. D [**81**]{}, 103535 (2010) \[arXiv:0912.1341 \[hep-th\]\].
S. Dimopoulos, S. Kachru, J. McGreevy and J. G. Wacker, JCAP [**0808**]{}, 003 (2008) \[hep-th/0507205\]. R. Kallosh, Lect. Notes Phys. [**738**]{}, 119 (2008) \[hep-th/0702059 \[HEP-TH\]\]; R. Kallosh, N. Sivanandam and M. Soroush, Phys. Rev. D [**77**]{}, 043501 (2008) \[arXiv:0710.3429 \[hep-th\]\]. E. Silverstein and A. Westphal, Phys. Rev. D [**78**]{}, 106003 (2008) \[arXiv:0803.3085 \[hep-th\]\]; L. McAllister, E. Silverstein and A. Westphal, Phys. Rev. D [**82**]{}, 046003 (2010) \[arXiv:0808.0706 \[hep-th\]\]; N. Kaloper and L. Sorbo, Phys. Rev. Lett. [**102**]{}, 121301 (2009) \[arXiv:0811.1989 \[hep-th\]\]. C. Germani and A. Kehagias, Phys. Rev. Lett. [**106**]{}, 161302 (2011) \[arXiv:1012.0853 \[hep-ph\]\]; J. Ohashi and S. Tsujikawa, JCAP [**1210**]{}, 035 (2012) \[arXiv:1207.4879 \[gr-qc\]\]; D. Maity, arXiv:1209.6554 \[hep-th\].
M. M. Anber and L. Sorbo, Phys. Rev. D [**81**]{}, 043534 (2010) \[arXiv:0908.4089 \[hep-th\]\]. S. Mohanty and A. Nautiyal, Phys. Rev. D [**78**]{}, 123515 (2008) \[arXiv:0807.0317 \[hep-ph\]\]; L. Visinelli, JCAP [**1109**]{}, 013 (2011) \[arXiv:1107.3523 \[astro-ph.CO\]\].
E. Dimastrogiovanni, M. Fasiello and A. J. Tolley, JCAP [**1302**]{}, 046 (2013) \[arXiv:1211.1396 \[hep-th\]\].
E. Dimastrogiovanni and M. Peloso, Phys. Rev. D [**87**]{}, 103501 (2013) \[arXiv:1212.5184 \[astro-ph.CO\]\].
D. H. Lyth, Phys. Rev. Lett. [**78**]{}, 1861 (1997) \[hep-ph/9606387\]. P. Adshead, E. Martinec and M. Wyman, arXiv:1301.2598 \[hep-th\]. P. Adshead, E. Martinec and M. Wyman, arXiv:1305.2930 \[hep-th\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], arXiv:1303.5076 \[astro-ph.CO\].
V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, Phys. Rept. [**215**]{} (1992) 203.
H. P. Nilles, M. Peloso and L. Sorbo, JHEP [**0104**]{}, 004 (2001) \[hep-th/0103202\].
[^1]: Ref. [@Elizalde:2012yk] studied at the background level a model characterized by the last term in (\[oneeq\]) plus a more generic kinetic structure ${\cal L } \supset {\cal F} \left[ - F^2 \right]$, where ${\cal F}$ is an arbitrary function.
[^2]: See [@Martinec:2012bv] for an extension of Chromo-natural inflation.
[^3]: To be precise, integrating out $\varphi$ is feasible only when $\varphi$ is close to the minimum of its potential; so, when the two models are compared to each other, one is actually identifying Gauge-flation with a specific limit of Chromo-natural inflation. For brevity of exposition, we refer to the two models as being “analogous”, or “dual”, but we stress that Chromo-natural inflation has actually a larger parameter space than Gauge-flation [@Adshead:2012qe; @SheikhJabbari:2012qf].
[^4]: Notation: in this work, dot denotes derivative with respect to physical time $t$, while prime denotes derivative with respect to conformal time $\tau$, related to the physical time by $d t = a \, d \tau$. Greek indices span all coordinates, while latin indices span the spatial coordinates.
[^5]: The terms “tensor/vector/scalar" are somewhat misnomers for some of the modes, since they refer to the transformation properties of the corresponding modes in $\delta g$ under spatial rotations, but not to the transformation properties of the modes in $\delta A$ (for instance, given that SU(2) indices have been used in that decomposition, the mode $M_a$ is not a vector under a spatial rotation).
[^6]: Equivalently, starting from a mode with generic $\vec{k}$, we can always rotate the axis such that $k_x=k_y=0$ and simultaneously perform a SU(2) global transformation such that the background relation $\langle A_\mu^a \rangle \propto \delta_\mu^a$ is preserved. This also proves that we can set $k=k_z$ with no loss of generality at the linearized level [@Dimastrogiovanni:2012ew].
[^7]: This justifies our gauge choice [@Dimastrogiovanni:2012ew], since this choice preserves all the non-dynamical $\delta g_{0\mu}$ and $\delta A_0^a$ modes. The number of physically independent propagating degrees of freedom is gauge independent, and in other gauges the non-dynamical degrees of freedom appear as (in general, nontrivial) linear combinations of the modes preserved in that gauge.
[^8]: The formal expression (\[TKOM-formal\]) has been written in a way that applies to all matrix elements, but not all the sums appearing in (\[TKOM-formal\]) are nontrivial for all the matrix elements. For instance, there is no square root in $\Omega_{11}^2$. Hence, $\Omega_{11}^2$ is also formally of the the type (\[TKOM-formal\]), but the sums inside the square root only contain the monomial $1$. See Appendix \[app-scalaraction\] for the explicit forms of the various entries.
[^9]: This is well known for the single field case. See [@Nilles:2001fg] for the multiple field case.
[^10]: More precisely, we impose these conditions at $\tau = 0$; as we discussed in Subsection \[subsec:PS\], if all the conditions (\[wronskian\]) and (\[wronskian2\]) are satisfied at some time, then they are satisfied at all times.
[^11]: To obtain $n_s$, we first compute $P_\zeta$ for a dense grid of values of $k$; we then interpolate $ \log P_\zeta$ as a function of $\log \frac{k}{k_{60}}$, and we differentiate this function. We verified that the grid is sufficiently dense, and that removing some of its points does not change the value of $n_s$.
[^12]: This confirms the observation made based on Figure \[fig:Pzeta\] that the spectrum is redder at smaller $\gamma_{\rm in}$; we also see that, for fixed $\gamma_{\rm in}$, the spectrum is redder the smaller $N$ is (this is common in slow-roll inflation, since the departure from scale invariance is due to the slow roll parameters, that typically increase during inflation).
[^13]: As we showed in Section \[sec:tensor\], the growth of $r$ with $\gamma_{\rm in}$ is due to the fact that the tachyonic instability in the left-handed tensor sector becomes stronger with increasing $\gamma$.
[^14]: Specifically, we impose eqs. (\[constraint-sol\]), namely the solutions of the constraint equations. The explicit form of the matrices $D$, $E$, $F$ that appear in (\[constraint-sol\]) are obtained as explained in the paragraph before eq. (\[TKOM-formal\]).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '**Abstract:** [Holographic dark energy (HDE), presents a dynamical view of dark energy which is consistent with the observational data and has a solid theoretical background. Its definition follows from the entropy-area relation $S(A)$, where $S$ and $A$ are entropy and area respectively. In the framework of loop quantum gravity, a modified definition of HDE called “entropy-corrected holographic dark energy” (ECHDE) has been proposed recently to explain dark energy with the help of quantum corrections to the entropy-area relation. Using this new definition, we establish a correspondence between modified variable Chaplygin gas, new modified Chaplygin gas and the viscous generalized Chaplygin gas with the entropy corrected holographic dark energy and reconstruct the corresponding scalar potentials which describe the dynamics of the scalar field. ]{}'
author:
- 'M. Umar Farooq'
- 'Muneer A. Rashid'
- Mubasher Jamil
title: 'Interacting entropy-corrected holographic Chaplygin gas model'
---
**Introduction**
================
Numerous cosmological observations of type Ia supernova, cosmic microwave background anisotropies measured with WMAP and large scale structure, suggest that our universe is undergoing in an accelerated expansion possibly due to the presence of dark energy which possesses negative pressure [1]{}. In the astrophysics community, the nature of such mysterious form of energy is still a matter of debate. In cosmology, several candidates responsible for this expansion have been proposed namely, Chaplygin gas, modified gravity, scalar tensor theory, tachyon and braneworld model, to a name a few [@2]. Despite the fact that the cosmological constant is the most obvious candidate that offers a solution to the dark energy problem, yet it has several drawbacks like fine-tuning and coincidence problems.
In order to soften the cosmic coincidence and fine tuning problems, models of dark energy interacting with dark matter have been proposed [@3]. It is generally asked that if the universe evolves from the earlier quintessence ($\omega >-1$) to late accelerating universe with phantom regime ($\omega <-1$), then why $\omega =-1$ crossing occurs at the present time, where $\omega={p}/{\rho }$. The equation of state (EoS) parameter $\omega $ changes at different cosmic epoch which supports the evolving dark energy [@4]. In the cosmological model explaining dark energy, the ratio of energy densities of dark matter and dark energy $\mathit{r}_{m}$ is of order unity, while observations suggest that in pure dark energy model, this ratio must decrease. It may implies the transfer of energy occurs between these two entities to keep a subtle balance at the current time. The interaction between these entities involves a coupling constant that determines the strength of this interaction. Besides Einstein gravity, this study of interaction can also be extended to $f(R)$, Brans-Dicke, braneworld, Horava-Lifshitz and Gauss-Bonnet gravities [@5].
In recent times, considerable interest has been stimulated to explain the observed dark energy (dominant force in cosmos) with the help of holographic dark energy model [@6]. According to holographic principle, the number of degrees of freedom in a bounded physical system should be finite and has relationship with the area of its boundary rather than with its volume [7]{}. It is commonly believed that the holographic principle is a fundamental principle of quantum gravity which is used to explain the events involving high energy scale. It is motivated from an observation that in quantum field theory, the ultra-violet cut-off $\Lambda $ could be related to the infrared cut-off $L$ due to the limit set by forming a black hole i.e. the quantum zero-point energy of a system with size $L$ should not exceed the mass of a black hole with the same size, i.e. $L^{3}\Lambda ^{3}\leq (M_{p}L)^{3/2}$ [@8]$.$ This last expression can be re-written as $L^{3}\rho
_{\Lambda }\leq LM_{p}^{2},$ where $\rho _{\Lambda }\backsim \Lambda
^{4}$ is the energy density corresponding to the zero point energy and cut-off $\Lambda .$ Now the last inequality takes the form $\rho
_{\Lambda }\leq M_{p}^{2}L^{-2}$ or $\rho _{\Lambda
}=3n^{2}M_{p}^{2}L^{-2}.$ Here $3n^{2}$ is a constant and attached for convenience. A sufficient literature is available on the study of interaction between the holographic dark energy and the matter [@9]. We are interested in studying the dynamics of entropy corrected holographic dark energy (ECHDE) when it interacts with some exotic type fluids. The entropy corrected holographic dark energy is given by [@10]$$\rho _{\Lambda }=3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4},$$where the first term on right hand side is the usual holographic dark energy while the other two appeared are mainly due to quantum corrections in the loop quantum gravity (LQG). The correction term play crucial role in early universe when L is small. When L gets large, ECHDE reduces to HDE. The expression for the corrected entropy of the black hole is given by $S=\frac{A}{4G}+\gamma \ln
(\frac{A}{4G})+\beta $ is mainly arise due to the thermal equilibrium and quantum fluctuation [@11]. The parameters $n^{2},$ $\gamma $ and $\beta $ constant of order unity. If we choose $\gamma $ and $\beta $ to be zero we arrive at the usual holographic dark energy model. The parameter $n$ can be a function of time [@12], but in our discussion it is purely a constant quantity.
The plane of the paper is as follows: In section 2, the model of Friedmann-Robertson-Walker (FRW) universe and the relevant equations are presented. In section 3, we discuss a correspondence between ECHDE and different Chaplygin gas forms including the modified variable Chaplygin gas (MVCG) and new modified Chaplygin gas (NMCG) and the viscous generalized Chaplygin gas (VGCG). In each case, we reconstruct the potentials and the dynamics of the scalar field which describe the entropy corrected holographic Chaplygin cosmology. The final section is devoted to the conclusion.
The model
=========
We start by assuming the background spacetime to be spatially homogeneous and isotropic FRW spacetime given by $$ds^{2}=-dt^{2}+a^{2}(t)\left[ \frac{dr^{2}}{1-kr^{2}}+r^{2}(d\theta
^{2}+\sin ^{2}\theta d\varphi ^{2})\right] .$$Here $a(t)$ is the dimensionless scale factor which is an arbitrary function of time and $k$ is defined to be the curvature parameter which has dimensions of *length*$^{-2}$ and its different values describe the spatial geometry. For instance, for $k=-1,0,1,$ the above metric (2) represents the spatially open, flat and closed FRW spacetimes respectively. The Friedmann equation is given by$$H^{2}+\frac{k}{a^{2}}=\frac{1}{3m_{p}^{2}}[\rho _{\Lambda }+\rho _{m}],$$where $m_{p}^{2}=(8\pi G)^{-1}$ is modified Planck mass. Here $H=\dot{a}/a$ is the Hubble parameter while $\rho _{\Lambda }$ and $\rho _{m}$ are the energy densities of dark energy and matter respectively. In dimensionless form, Eq. (3) can be written as$$1+\Omega _{k}=\Omega _{\Lambda }+\Omega _{m}.$$The dimensionless density parameters corresponding to matter, dark energy and curvature are$$\Omega _{m}=\frac{\rho _{m}}{\rho _{cr}}=\frac{\rho _{m}}{3H^{2}m_{p}^{2}},%
\text{ }\Omega _{\Lambda }=\frac{\rho _{\Lambda }}{\rho _{cr}}=\frac{\rho
_{\Lambda }}{3H^{2}m_{p}^{2}},\text{ }\Omega _{k}=\frac{k}{(aH)^{2}}.$$Here $\rho _{cr}=3H^{2}m_{p}^{2}$ is the critical density. The energy conservation equations for dark energy and dark matter are$$\begin{aligned}
\dot{\rho}_{\Lambda }+3H(\rho _{\Lambda }+p_{\Lambda }) &=&-Q, \\
\dot{\rho}_{m}+3H\rho _{m} &=&Q.\end{aligned}$$Here overdot represents the differentiation with respect to cosmic co-moving time $t$. Eq. (6) and (7) show that if there is an interaction between dark energy and dark matter, the energy conservation for dark energy and matter would not hold independently but for the total interacting system. In explicit form, we have $p_{\Lambda }=\omega _{\Lambda }\rho _{\Lambda }$ and $p_{m}=0.$ During the energy transfer, local energy conservation will not hold in general but for the whole interacting system. Naturally if two species are present in dominant form, it is obvious that they will interact. If the quantity $Q$ is positive, it shows the transfer of energy from dark matter to dark energy and vice versa in the case when $Q$ is negative. The importance of interacting dark energy and dark matter model also emerges since it is the best fit for the data we obtain from the physical observations for instance SN Ia and cosmic microwave background [@1].
Defining the effective equation of state for dark energy and dark matter as [@13]$$\omega _{\Lambda }^{\text{eff}}=\omega _{\Lambda }+\frac{\Gamma }{3H},\text{
}\omega _{m}^{\text{eff}}=-\frac{1}{r_{m}}\frac{\Gamma }{3H},$$where the term $\Gamma =Q/\rho _{\Lambda }$ represents the decay rate. After employing Eq. (8) in (6) and (7), we obtain the following pair of continuity equations$$\begin{aligned}
\dot{\rho}_{\Lambda }+3H(1+\omega _{\Lambda }^{\text{eff}})\rho _{\Lambda }
&=&0 \\
\dot{\rho}_{m}+3H(1+\omega _{m}^{\text{eff}})\rho _{m} &=&0.\end{aligned}$$If we take $L$ as the Hubble scale $H^{-1}$ i.e. $L=H^{-1}$ at the present epoch $H=H_{0}\backsim 10^{-33}eV,$ then the energy density $\rho _{\Lambda }
$ is comparable with the observed dark energy density $\backsim
10^{-10}eV^{4}.$ The second option for the infra-red cut-off is the particle horizon. Hsu [@14] showed that under this scenario the resulting EoS becomes zero and does not lead to an accelerated universe. So to get an accelerated expansion of the observable universe, Li [@15] proposed that the IR cut-off $L$ should be taken as the future event horizon and he defined it as$$L=a(t)r(t),$$here $r(t)$ is related to the future event horizon of the observable universe. Using the FRW metric, we can obtain [@15] $$L=a(t)\frac{\text{sinn}(\sqrt{|k|}y)}{\sqrt{|k|}},\ \ \ y=\frac{R_{h}}{a(t)},$$where $R_{h}$ is the size of the future event horizon defined as $$R_{h}=a(t)\int\limits_{t}^{\infty }\frac{dt^{\prime }}{a(t^{\prime })}%
=a(t)\int\limits_{0}^{r_{1}}\frac{dr}{\sqrt{1-kr^{2}}}.$$The last integral has the explicit form as $$\int\limits_{0}^{r_{1}}\frac{dr}{\sqrt{1-kr^{2}}}=\frac{1}{\sqrt{|k|}}\text{%
sinn}^{-1}(\sqrt{|k|}r_{1})=\begin{cases}\text{sin}^{-1}(r_1) , & \,
\,k=+1,\\ r_1, & \, \, k=0,\\ \text{sinh}^{-1}(r_1), & \, \,k=-1.\\
\end{cases}$$Using the definition of $\rho _{\Lambda }$ and $\rho _{cr},$ we obtain the following relation $$HL=\sqrt{\frac{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}%
}{3m_{p}^{2}\Omega _{\Lambda }}}.$$Differentiate eq. (12) with respect to time $t$ and using Eq. (16), we have$$\dot{L}=\sqrt{\frac{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}{3m_{p}^{2}\Omega _{\Lambda
}}}-\text{cosn}(\sqrt{|k|}y),$$where$$\text{cosn}(\sqrt{|k|}y)=\begin{cases} \text{cos}y , \, \,k=+1,\\ 1, \, \,
k=0,\\ \text{cosh}y, \, \,k=-1,\\ \end{cases}$$After taking derivative of (1) with respect to $t,$ we get$$\begin{aligned}
\dot{\rho}_{\Lambda } &=&(2\gamma L^{-5}-4\gamma L^{-5}\ln
(m_{p}^{2}L^{2})-4\beta L^{-5}-6n^{2}m_{p}^{2}L^{-3}) \nonumber \\
&&\times \left[ \sqrt{\frac{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}{3m_{p}^{2}\Omega _{\Lambda }}}-\text{cosn}(%
\sqrt{|k|}y)\right] .\end{aligned}$$Using Eq. (18) in (6), we arrive at $$\begin{aligned}
w_{\Lambda } &=&-1-\Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\Big) \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] -\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}.\end{aligned}$$Now making use of Eq. (19) in (8) yields$$\begin{aligned}
w_{\Lambda }^{\text{eff}} &=&-1-\Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\Big) \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] .\end{aligned}$$
Correspondence between ECHDE and Chaplygin gas variants
=======================================================
Since there are several candidates of dark energy, so it is essential to develop a correspondence and the relationships between them. Kamenshchik et al [@16] studied a homogenous model based on a single fluid obeying the EoS $p=-\frac{A_{0}}{\rho }$ called the Chaplygin gas, where $p$ and $\rho $ represent the pressure and energy density of the fluid and $A_{0}$ is some positive constant. Possessing many physically interesting features, several authors have used it to model the accelerated expansion of universe [@17]. But it does not satisfactorily address the problems like structure formation and cosmological perturbation power spectrum [@18]. Subsequently, this equation was modified to the form $p=-\frac{A_{0}}{\rho ^{\alpha }}$ called generalized Chaplygin gas (GCG) to construct viable cosmological models. Two free parameters involved in it: one is $A_{0}$ and the other $0\leq \alpha \leq 1.$ The GCG fluid behaves like dust for small size of the universe while it acts as cosmological constant when universe gets sufficiently large. The GCG equation has been further modified to $p=B\rho -\frac{A_{0}}{%
\rho ^{\alpha }}$ with $0\leq \alpha \leq 1,$ which is called modified Chaplygin gas (MCG) [@19] and it involves three parameters. An interesting feature connected with MCG equation of state is that it shows radiations era in the early universe. At the late time it behaves as cosmological constant which can be fitted to a $\Lambda $CDM model. Late on, Guo and Jhang [@20] first proposed a model $p=B\rho -\frac{A_{0}}{\rho ^{\alpha }}$ by taking $A_{0}$ as a function of the cosmological scale factor $a(t)$ i.e. $A_{0}=A_{0}(a(t)),$ which is known as modified variable Chaplygin gas (MVCG) [@21]. This assumption seems to be reasonable since $A_{0}(a(t))$ is related to scalar potential if we interpret Chaplygin gas via Born-Infeld scalar field.
Modified variable Chaplygin gas and ECHDE
-----------------------------------------
Suppose we have two species i.e. dark matter and dark energy. The later is specified by the MVCG which is given by $$p=B\rho _{\Lambda }-\frac{B_{0}a^{-\delta }}{\rho _{\Lambda
}^{\alpha }}.$$The evolution of the energy density of MVCG is $$\rho _{\Lambda }=\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha +1)(B+1)-\delta ]}%
\frac{1}{a^{\delta }}-\frac{C}{a^{3(\alpha +1)(B+1)}}\Big]^{\frac{1}{\alpha
+1}},$$where $B_{0}$ and $C$ are some constants.
We now reconstruct expressions of the potential and the dynamics of the scalar field in the presence of ECHDE. So for this, consider a time dependent scalar field $\phi (t)$ with potential $V(\phi ),$ which are directly related with the energy density and pressure of MVCG as$$\begin{aligned}
\rho _{\Lambda } &=&\frac{1}{2}\dot{\phi}^{2}+V(\phi ), \\
p_{\Lambda } &=&\frac{1}{2}\dot{\phi}^{2}-V(\phi ).\end{aligned}$$Since the kinetic terms are positive hence it means that MVCG is of quintessence type. Adding Eqs. (23) and (24), we get the kinetic term$$\begin{aligned}
\dot{\phi}^{2} &=&(1+B)\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha
+1)(B+1)-\delta ]}\frac{1}{a^{\delta }}-\frac{C}{a^{3(\alpha +1)(B+1)}}\Big]%
^{\frac{1}{\alpha +1}} \nonumber \\
&&-\frac{B_{0}a^{-\delta }}{\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha
+1)(B+1)-\delta ]}\frac{1}{a^{\delta }}-\frac{C}{a^{3(\alpha +1)(B+1)}}\Big]%
^{\frac{\alpha }{\alpha +1}}}.\end{aligned}$$Subtraction of Eqs. (23) and (24) yields the potential term$$\begin{aligned}
2V(\phi ) &=&(1-B)\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha +1)(B+1)-\delta ]}%
\frac{1}{a^{n}}-\frac{C}{a^{3(\alpha +1)(B+1)}}\Big]^{\frac{1}{\alpha +1}}
\nonumber \\
&&+\frac{B_{0}a^{-\delta }}{\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha
+1)(B+1)-\delta ]}\frac{1}{a^{\delta }}-\frac{C}{a^{3(\alpha +1)(B+1)}}\Big]%
^{\frac{\alpha }{\alpha +1}}}.\end{aligned}$$To see the correspondence between the ECHDE and MVCG energy density, we use Eqs. (1) and (22) to get $$C=a^{3(\alpha +1)(B+1)}\Big[\frac{3(\alpha +1)B_{0}}{[3(\alpha
+1)(B+1)-\delta ]}\frac{1}{a^{\delta }}-(3n^{2}m_{p}^{2}L^{-2}+\gamma
L^{-4}\ln (m_{p}^{2}L^{2})+\beta L^{-4})^{\alpha +1}\Big].$$Writing Eq. (21) in an alternate form$$B_{0}=a^{\delta }(B-w_{\Lambda })\rho _{\Lambda }^{\alpha +1}.$$Therefore in view of Eq. (19), the above equation (28) gives the value of the parameter $B_{0}$ as$$\begin{aligned}
B_{0} &=&a^{\delta }(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4})^{\alpha +1}\Big[1+B \nonumber \\
&&+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left( 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right) +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big].\end{aligned}$$Now to determine the other parameter $C$ substitute the value of $B_{0}$ in Eq. (27), we have$$\begin{aligned}
C &=&a^{3(\alpha +1)(A+1)}(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4})^{\alpha +1}\Big(\frac{3(\alpha +1)}{[3(\alpha
+1)(B+1)-\delta ]} \nonumber \\
&&\times (B+1+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})}) \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] -1\Big)\end{aligned}$$Now we can re-write the kinetic energy and scalar potential terms as $$\begin{aligned}
\dot{\phi}^{2} &=&-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4}) \nonumber \\
&&\Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big),\end{aligned}$$and$$\begin{aligned}
2V(\phi ) &=&(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4}) \nonumber \\
&&\times \Big(2+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big)\end{aligned}$$Using the relation $x=\ln a,$ we get $\dot{\phi}=\phi ^{\prime }H,$ where $%
^{\prime }$ represents the derivative with respect to e-folding time parameter $\ln a.$ After putting the value of $\dot{\phi}$ and applying integration, we obtain$$\begin{aligned}
\phi (a)-\phi (a_{0}) &=&\frac{1}{H}\int_{0}^{\ln
a}[-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta L^{-4})
\nonumber \\
&&\times \Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] \Big)+\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}%
]^{1/2}d\ln a, \nonumber \\
&&\end{aligned}$$where $a_{0}$ is the present value of the scale factor.
New modified Chaplygin gas model and ECHDE
------------------------------------------
The model that represents the dark energy is now the new modified Chaplygin gas (NMCG) given by [@22]$$p_{\Lambda }=B\rho _{\Lambda }-\frac{K(a)}{\rho _{\Lambda }^{\alpha }},\text{
\ \ where }B>0\text{ and }0\leq \alpha \leq 1.$$Here $K(a)$ is a function of scale factor of the universe. Taking $K(a)$ in the form $K(a)=-\omega _{\Lambda }A_{1}a^{-3(w_{\Lambda }+1)(\alpha +1)}$ as introduced by the authors in [@23], we get$$p_{\Lambda }=B\rho _{\Lambda }+\frac{w_{\Lambda }A_{1}}{\rho _{\Lambda
}^{\alpha }}a^{-3(w_{\Lambda }+1)(\alpha +1)}.$$The energy density of the NMCG can be expressed as$$\rho _{\Lambda }=\Big[\frac{w_{\Lambda }}{w_{\Lambda }-B}A_{1}a^{-3(w_{%
\Lambda }+1)(\alpha +1)}+B_{1}a^{-3(B+1)(\alpha +1)}\Big]^{\frac{1}{\alpha +1%
}},$$where $B_{1}$ is a constant of integration. Following the previous section we establish the correspondence between the ECHDE and NMCG energy density. Comparing Eqs. (36) and (1), we get$$\begin{aligned}
B_{1} &=&a^{3(B+1)(\alpha +1)}\Big((3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4})^{\alpha +1} \nonumber \\
&&-\frac{1+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})}\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega
_{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}%
\text{cosn}(\sqrt{|k|}y)\right] }{1+B+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\times \left[ 1-\sqrt{\frac{%
3m_{p}^{2}\Omega _{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(\sqrt{|k|}y)\right] } \nonumber
\\
&&\times \Big(A_{1}a^{3(\alpha +1)\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\times \left[ 1-\sqrt{\frac{%
3m_{p}^{2}\Omega _{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(\sqrt{|k|}y)\right] }\Big)\Big).\end{aligned}$$With the use of $p_{\Lambda }=\omega _{\Lambda }\rho _{\Lambda },$ Eq. (35) gives an expression for $A_{1}$ i.e.$$\begin{aligned}
A_{1} &=&(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4})^{\alpha +1} \nonumber \\
&&\times a^{-3(\alpha +1)\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\times \Big(1-\sqrt{\frac{%
3m_{p}^{2}\Omega _{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(\sqrt{|k|}y)\Big)} \nonumber \\
&&\times \left[ \frac{B+1+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})}\times \left[ 1-\sqrt{\frac{%
3m_{p}^{2}\Omega _{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(\sqrt{|k|}y)\right] }{1+\frac{%
2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})}\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega
_{\Lambda }}{3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}%
\text{cosn}(\sqrt{|k|}y)\right] }\right] . \nonumber \\
&&\end{aligned}$$So we have found the expressions for the constants $A_{1}$ and $B_{1}.$ Employing Eqs. (23) and (24), the kinetic and potential terms are found to be$$\begin{aligned}
\dot{\phi}^{2} &=&(B+1)\Big[\frac{w_{\Lambda }}{w_{\Lambda }-B}%
A_{1}a^{-3(w_{\Lambda }+1)(\alpha +1)}+B_{1}a^{-3(B+1)(\alpha +1)}\Big]^{%
\frac{1}{\alpha +1}} \nonumber \\
&&+\frac{A_{1}w_{\Lambda }a^{-3(w_{\Lambda }+1)(\alpha +1)}}{[\frac{%
w_{\Lambda }}{w_{\Lambda }-B}A_{1}a^{-3(w_{\Lambda }+1)(\alpha
+1)}+B_{1}a^{-3(B+1)(\alpha +1)}]^{\frac{\alpha }{\alpha +1}}},\end{aligned}$$and $$\begin{aligned}
2V(\phi ) &=&(1-B)[\frac{w_{\Lambda }}{w_{\Lambda }-B}A_{1}a^{-3(w_{\Lambda
}+1)(\alpha +1)}+B_{1}a^{-3(B+1)(\alpha +1)}]^{\frac{1}{\alpha +1}}
\nonumber \\
&&-\frac{A_{1}w_{\Lambda }a^{-3(w_{\Lambda }+1)(\alpha +1)}}{[\frac{%
w_{\Lambda }}{w_{\Lambda }-B}A_{1}a^{-3(w_{\Lambda }+1)(\alpha
+1)}+B_{1}a^{-3(B+1)(\alpha +1)}]^{\frac{\alpha }{\alpha +1}}},\end{aligned}$$where $A_{1}$ and $B_{1}$ are given in (38) and (37) respectively. In all Eqs. (35), (36), (39) and (40) the value of $ w_{\Lambda }$ is given by Eq. (19). From Eqs. (23) and (24), the kinetic energy term is re-written to be$$\begin{aligned}
\dot{\phi}^{2} &=&-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4}) \nonumber \\
&&\times \Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big),\end{aligned}$$while the potential energy term has the form$$\begin{aligned}
2V(\phi ) &=&(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4}) \nonumber \\
&&\times \Big(2+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big).
\nonumber \\
&&\end{aligned}$$Following the same steps as done for the MVCG, the kinetic term is easily transformable to the following form$$\begin{aligned}
\phi (a)-\phi (a_{0}) &=&\frac{1}{H}\int_{0}^{\ln
a}[-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta L^{-4})
\nonumber \\
&&\times \frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}%
]^{1/2}d\ln a. \nonumber \\
&&\end{aligned}$$
Viscous generalized Chaplygin gas model and ECHDE
-------------------------------------------------
If we assume that the dark energy as non-viscous as taken in section A and B, it must results in the occurrence of a cosmic singularity (Big Rip) in the far future. This singularity can be alleviated by introducing the quantum corrections due to the conformal anomaly while the other option is to consider the bulk viscosity $\xi $ of the cosmic fluid [@24]. The theory of bulk viscosity was initially investigated by Eckart and later on pursued by Landau and Lifshitz [@25]. The important feature of dark energy with bulk viscosity is that it shows the accelerated expansion of phantom type in the later epoch and softens the coincidence problem, age problem and phantom crossing [@26]. The effective pressure containing the isotropic pressure and viscous stress is given by the equation of state$$p_{\text{eff}}=p_{\Lambda }+\Pi ,$$where $p_{\Lambda }=\frac{\chi }{\rho _{\Lambda }^{\alpha }},$ $\chi >0.$ Notice that first term on the right hand side mimics the GCG and the parameter $\alpha $ varies as $0<\alpha \leq 1.$ If $\alpha =1$ it represents the Chaplygin gas model. On the other hand if $\alpha <0$, it corresponds to a polytropic gas. The bulk viscous fluid is represented by $%
\Pi =-\xi (\rho _{\Lambda })u_{;\mu }^{\mu }$ where $u^{\mu }$ is the four-velocity vector of the viscous fluid and $\xi >0$ to get the positive entropy production in conformity with second law of thermodynamics [@27]. We choose $\xi (\rho _{\Lambda })=\nu \rho _{\Lambda }^{1/2}$, with $\nu $ as constant. The energy conservation equation yields the energy density of VGCG as$$\rho _{\Lambda }=[\frac{Da^{-3(\alpha +1)(1-\nu \gamma _{1})}-\chi }{(1-\nu
\gamma _{1})}]^{\frac{1}{\alpha +1}},$$ [@28]. Here $\gamma _{1}=m_{p}^{-1}\sqrt{1-r_{m}},$ where $r_{m}=\frac{%
\rho _{m}}{\rho _{\Lambda }}=\frac{\Omega _{m}}{\Omega _{\Lambda }}$ and $D$ some constant of integration. The effective equation of state (45) says that the universe is expanding in accelerated manner. It has been shown that this phenomenon of universe can be studied with the help of dynamical evolving scalar fields usually called inflation. So it was introduced to construct models dealing with minimally coupled scalar field. Hence we construct the dynamical scalar field $\phi $ with potential $V(\phi ),$ related to the energy density and pressure of viscous dark energy model as$$\begin{aligned}
\rho _{\Lambda } &=&[\frac{Da^{-3(\alpha +1)(1-\nu \gamma _{1})}-\chi }{%
(1-\nu \gamma _{1})}]^{\frac{1}{\alpha +1}}, \nonumber \\
p_{\Lambda } &=&\chi \lbrack \frac{(1-\nu \gamma _{1})}{Da^{-3(\alpha
+1)(1-\nu \gamma _{1})}-\chi }]^{\frac{\alpha }{\alpha +1}}-3\nu H[\frac{%
Da^{-3(\alpha +1)(1-\nu \gamma _{1})}-\chi }{(1-\nu \gamma _{1})}]^{1/2}.\end{aligned}$$To determine the expressions for the unknown $D$ and $\chi $ we take the following steps. The effective equation of state for the interacting VGCG is$$w_{\Lambda }^{\text{eff}}=\frac{\chi }{\rho _{\Lambda }^{\alpha +1}}%
-3\upsilon H\rho _{\Lambda }^{-\frac{1}{2}}+\frac{b^{2}(1+\Omega _{k})}{%
\Omega _{\Lambda }}.$$After substituting the value of $w_{\Lambda }^{\text{eff}}$ in (47) and making simplification, we obtain$$\begin{aligned}
\chi &=&(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4})^{\alpha +1} \nonumber \\
&&\times \Big(-1-\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] \nonumber \\
&&+3\upsilon H(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4})^{-1/2}-\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big).
\nonumber \\
&&\end{aligned}$$Inserting the value of $\chi $ in (45), we get$$\begin{aligned}
D &=&a^{3(\alpha +1)(1-\nu \gamma _{1})}(3n^{2}m_{p}^{2}L^{-2}+\gamma
L^{-4}\ln (m_{p}^{2}L^{2})+\beta L^{-4})^{\alpha +1} \nonumber \\
&&\times \lbrack -\nu \gamma _{1}-\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln
(m_{p}^{2}L^{2})-4\beta L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma
L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] \nonumber \\
&&+3\upsilon H(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4})^{-\frac{1}{2}}-\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}].
\nonumber \\
&&\end{aligned}$$Now we can re-write the scalar potential and kinetic energy terms as following$$\begin{aligned}
2V(\phi ) &=&(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta
L^{-4}) \nonumber \\
&&\times (2+\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}).
\nonumber \\
&&\end{aligned}$$and$$\begin{aligned}
\dot{\phi}^{2} &=&-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln
(m_{p}^{2}L^{2})+\beta L^{-4}) \nonumber \\
&&\times \Big(\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}\Big).\end{aligned}$$Again using the equation $\dot{\phi}=\phi ^{\prime }H,$ we can write$$\begin{aligned}
\phi (a)-\phi (a_{0}) &=&\frac{1}{H}\int_{0}^{\ln
a}[-(3n^{2}m_{p}^{2}L^{-2}+\gamma L^{-4}\ln (m_{p}^{2}L^{2})+\beta L^{-4})
\nonumber \\
&&\times (\frac{2\gamma L^{-2}-4\gamma L^{-2}\ln (m_{p}^{2}L^{2})-4\beta
L^{-2}-6n^{2}m_{p}^{2}}{3(3n^{2}m_{p}^{2}+\gamma L^{-2}\ln
(m_{p}^{2}L^{2})+\beta L^{-2})} \nonumber \\
&&\times \left[ 1-\sqrt{\frac{3m_{p}^{2}\Omega _{\Lambda }}{%
3n^{2}m_{p}^{2}+\gamma L^{-2}\ln (m_{p}^{2}L^{2})+\beta L^{-2}}}\text{cosn}(%
\sqrt{|k|}y)\right] +\frac{b^{2}(1+\Omega _{k})}{\Omega _{\Lambda }}%
)]^{1/2}d\ln a. \nonumber \\
&&\end{aligned}$$It is interesting to note that the above potential and kinetic energy expressions for the interacting ECHDE with the viscous generalized Chaplygin gas coincide with the non-viscous case.
Conclusion
==========
Enormous literature dealing with the subject of dark energy is available but the holographic dark energy is considered to be the most promising candidate of dark energy. In this paper we have constructed a correspondence between the interacting ECHDE and the Chaplygin gas variants. Several candidates of dark energy have been suggested to describe cosmic acceleration but Chaplygin gas has emerged as a unification of dark energy and dark matter. It’s cosmic evolution is similar to initial dust like matter while it behaves as a cosmological constant at a later epoch. In the present work, we have investigated a model of dark energy in the presence of entropy corrections to holographic dark energy. In this context, a link between the ECHDE and various models of Chaplygin gas has been established. We have found the kinetic and potential energies corresponding to each model and also reconstructed the potentials.
[7]{} A. G.Riess et al. \[Supernova Search Team Collaboration\], Astron. J. **116** (1998)1009; S. perlmutter et al., \[Supernova Cosmology Project Collaboration\], Astrophys. J. **517** (1999) 565; W. J. Percival et al., \[The 2d FGRS Collaboration\], Mon .Not. Roy. Astron. Soc. **327** (2001) 1297; P. Astier et al., Astron., Astrophs. **447** (2006) 31; A. G. Riess et al., \[Supernova Search Team Collaboration\], Astrophys. J. **607** (2004) 665; P. de Bernardis et al., Nature (London) **404** (2000) 955; R. A Knop et al., Astrophys. J. **598** (2003) 102; J. L. Tonry et al., Astrophys. J. **594** (2003) 1; M. V. Jhon Astrophys. J. **614** (2004) 1; D. N. Spergel et al., Astrophys. J. Suppl. **170** (2007) 377; G. Hinshaw Astrophys. J. **170** (2007) 288; M. Colless et al., Mon. Not. R. Astro. Soc. **328** (2001) 1039; M. Tegmark et al., Phys Rev. D **69** (2004) 103501; V. Springel et al., Nature (London) **440** (2006) 1137.
P. J. E. Peebles and B. Ratra, Rev. Mod. Phys. **75** (2003) 559; E. J. Copeland et al., Int. J. Mod. Phys. D **15** (2006) 1753; R. R. Caldwell et al., Phys. Rev. Lett. **80**, (1998) 1582; A. R. Liddle et al., Phys. Rev. D **59** (1999) 023509; T. padmanabhan **380** (2003) 235; A. Sen, JHEP **0207**, (2002) 065; N.N. Weinberg et al., Phys. Rev. Lett. **91**, (2003) 071301 ; B. Feng et al., Phys. Lett B **607** (2005) 35; Z. K. Guo et al., Phys. Lett. B **608** (2005) 177; H. Wei et al., Class. Quantum Grav. **22** (2005) 3189; S. M. Carrol. The Cosmological Constant, Living Rev. Rel. **4** (2001) 1; V Sahni et al., Int. J. Mod. Phys. D **9** (2000) 373.
H. Zhang et al., Phys.Lett. B **678** (2009) 331; G. Caldera-Cabral et al.,JCAP **07** (2009) 027; M. Jamil et al., Eur. Phys. J. C **64** (2009) 97; M. Jamil et al., Eur. Phys. J. C **60** (2009) 141; M. Jamil et al., Eur. Phys. J. C **58** (2008) 111.
Sadjadi H. M. et al., Phys. Rev. D **74** (2006) 103007.
N. J. Poplawski Phys. Rev. D **74** (2006) 084032; A. Sheykhi Phys. Lett. B **681** (2009) 205; K. Y. Kim et al., Mod. Phys. Lett. A **22** (2007) 2631 M. R. Setare arXiv: 0909.0456; M. R. Setare Chin. Phys. Lett. **26** (2009) 029501.
K. Enqvist, S et al., JCAP **2** (2005) 004; X. Zhang. Int. J. Mod. Phys. D **14** (2005) 1597; D. Pavon et al., hep-th/0511053; ; P. F. Gonzalez-Diaz, Phys. Rev. D **27** (1983) 3042.
G’t Hooft, gr-qc/9310026; L. Susskind, J. Math. Phys. **36** (1995) 6377.
J. D. Bekenstein, Phys. Rev. D **7** (1973) 2333; S. W. Hawking Comm. Math. Phys. **43** (1975) 199; S. W. Hawking Phys. Rev. D **13** (1976) 191; J. D. bekenstein, Phys. Rev. D **23** (1981) 287; A. G. Cohen et al., Phys. Rev. Lett. **82** (1999) 4971.
M. R. Setare Eur. Phys. J. C **50** (2007) 991; M. R. Setare Eur. Phys. J. C **52** (2007) 689; K. Karami et al., arXiv: 0912.1536.
H. Wei, Commun. Theor. Phys. **52** (2009) 743; M. Jamil and M.U. Farooq, JCAP 1003 (2010) 001.
A. Ghosh and P. Mitra, Phys. Rev. D 71 (2005) 027502; A. Ashtekar et al, Phys. Rev. Lett. 80 (1998) 904; C. Rovelli, Phys. Rev. Lett. **77** (1996) 3288; K. A. Miessner, Class. Quant. Grav. **21** (2004) 5245.
L. Xu JCAP **09** (2009) 016.
H. Kim et al., Phys. Lett. B **632** (2006) 605; H. M. Sadjadi et al., Phys. Rev. D **74** (2006) 103007.
S. D. H. Hsu, Phys. Lett. B **594** (2004) 13.
M. Li, Phys. Lett. B **603** (2004) 1.
A. Kamenshchik et al., Phys. Lett. B **511** (2001) 265.
V. Gorini et al., preprint 0403062 \[gr-qc\]; A. Y. Kamenshchik et al., Phys. Lett. B **511** (2001) 265.
Sandvik et al., Phys. Rev. D **69** (2004) 123524; R. Bean R et al., Phys. Rev. D **68** (2003) 023515.
H. B. Benaoum, preprint hep-th/0205140 (2002); ;U. Debnath et al., Class.Qunatum Grav. **21** (2004) 5609; V. Gorini et al., Phys. Rev. D **67** (2003) 063509; U. Alam et al., Mon. Not. Roy. Astrn. Soc. **344** (2003) 1057; A. Dev et al., Phys Rev D **67** (2003) 023515. V. Sahni et al JETP Lett. **77** (2003) 201; M. C. Bento, et al., Phys. Rev. D **66** (2002) 043507.
Z. K. Guo and Y. Z. Zhang, Phys. Lett. B **645** (2007) 326.
U. Debnath, Astrophys. Space Sci., **312** (2007) 295; M. Jamil et al., Eur. J. C **61** (2009) 471.
Surajit Chattopadhyay, Ujjal Debnath, arXiv.: 0805.007v \[gr-qc\].
X. Zhang et al., JCAP **01** (2006) 003.
I. Brevik, S. D. Odintsov, Phys. Lett. B **455** (1999) 104; S. Nojiri, S.D. Odintsov Phys. Lett. B **562** (2003) 147.
C. Eckart, Phys. Rev. **58** (1940) 919; L. D. Landau and E. M. Lifschitz, Fluid Mechanics (Butterworth Heineman, Oxford (1987))
C. Feng et al., Phys Lett. B **680** (2009) 355; J. Chen, Y. Wang. :arXiv: 0904.2808v2 \[gr,qc\]; I. Brevik, Int. J. Mod. Phys. D **15**, (2006) 767.
W. Zimdahl, D. Pavon, Int. J. Mod. Phys. D **15** (2006) 767.
M. Jamil, M. A. Rashid, Eur. Phys. J. C **56** (2008) 429.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We present a detailed analysis of the class of regression decision tree algorithms which employ a regulized piecewise-linear node-splitting criterion and have regularized linear models at the leaves. From a theoretic standpoint, based on Rademacher complexity framework, we present new high-probability upper bounds for the generalization error for the proposed classes of regularized regression decision tree algorithms, including LASSO-type, and $\ell_{2}$ regularization for linear models at the leaves. Theoretical result are further extended by considering a general type of variable selection procedure. Furthermore, in our work we demonstrate that the analyzed class of regression trees is not only numerically stable but can furthermore be made tractable via an algorithmic implementation, presented herein, as well as with the help of modern GPU technology. Empirically, we present results on multiple datasets which highlight the strengths and potential pitfalls, of the proposed tree algorithms compared to baselines which grow trees based on piecewise constant models.
**Keywords:** Decision trees, Piecewise-Linear Regression, Rademacher complexity, regularization.
author:
- 'Leonidas Lefakis, Oleksandr Zadorozhnyi, Gilles Blanchard'
bibliography:
- 'lrt.bib'
title: 'Efficient Regularized Piecewise-Linear Regression Trees '
---
Introduction {#sec:intro}
============
Decision trees and random forests remain to be very popular machine learning tools because of their general applicability and considerable empirical evidence with regards to their performance on diverse tasks. Both theoretical aspects and practical applications of such algorithms are active fields of study.
At the same time in recent years another family of predictors, namely those falling under the umbrella of “deep-learning”, have met with success. Despite recent efforts to attain a better understanding of the properties of this model class, there remains much which is not well understood (from a theoretical perspective). Nonetheless, it is clear that the increase in performance of such algorithms is due, at least in part, to their ability to leverage cutting edge GPU technology to efficiently build complex models using huge datasets.
Consequently, an important and largely open question is whether other families of predictors can exploit this technology to increase their performance. In the case of decision trees, recent work [@Kont15] has in fact shown that decision trees and deep learning are not mutually exclusive and can be combined to create powerful predictors. Here, however, we are interested in investigating enhancements of the decision trees algorithms, and in particular regression trees without resorting to building hybrid deep systems.
In particular, we note that an integral component of every tree growing algorithm, to wit the criterion used for defining the optimal split at each node, has remained relatively simple in nature and is typically based on the squared error of a piecewise constant model. The motivation behind this simplicity has invariably been linked to the tractability and numerical stability of the underlying algorithm. In the following we propose a piecewise linear splitting criterion for building a regression tree and show how modern GPU technology makes such complex criteria both tractable and numerically stable. We furthermore present a detailed theoretical analysis that provides insights into the generalization error of this model class, obtaining high probability upper bounds on the generalization error. Finally, we present empirical evidence showing that proposed algorithm consistently outperforms many benchmark algorithms which employ the aforementioned piecewise constant model criterion.
Related work {#sec:rel_work}
============
Least squares regression trees are introduced in the seminal work by Breiman et al. [@breiman1984classification]. As with the large majority of subsequent tree building algorithms, the proposed CART trees proceed in a top-down greedy manner. Nodes are split based on local criteria which in the case of regression trees takes the form of minimizing the variances of the target values in the two resulting subtrees. This can be interpreted as minimizing the squared error of a piecewise constant model.
The M5 algorithm proposed by Quinlan [@quinlan1992learning] allows construction of linear models at the leaves of the resulting tree. The splitting of samples at the internal nodes is chosen so as to minimize the standard deviations on the two sub-populations. Effectively the algorithm builds a regression tree based on a piecewise constant model and exchanges these constant models for more powerful linear models at the leaves. As noted in [@landwehr2003logit] due to the similarity in splitting criteria, the CART and M5 algorithms result in similar tree structures. This similarity not withstanding, M5, and its “rational reconstruction” M5$'$ [@wang1997inducing] have been shown to outperform CART in practice [@wang1997inducing; @vogel07scalable].
Similar to the M5, the HTL algorithm [@torgo1997functional], first builds a CART tree and then replaces the models at the nodes. The main difference in this work from the M5 algorithm is that HTL allows for non-linear regressors at the leaves. Yet another approach, SECRET [@dobra2002secret], constructs an artificial classification problem, and shows that there is merit in building the tree structure using a classification tree algorithm and then finally assigning linear regression models to each of the leaves.
One characteristic of these and other approaches, is that they avoid using the final model (usually linear regression) as a criterion when optimizing the splits of the internal nodes. M5 for instance, optimizes the split for piecewise constant models in the leaves and only inserts the linear models in the leaves post-hoc.
However, it would be more desirable to build the tree using a split criterion that takes into account the models actually employed in the leaves: $$\begin{aligned}
\label{eq:lrt}
\sum_{\bx_i \in A} \paren[2]{ y_i- \hat{f}_A(\bx_i;w_A) }^2
+ \sum_{\bx_j \in A^c} \paren[2]{ y_j- \hat{f}_{A^c}(\bx_j;w_{A^c}) }^2,\end{aligned}$$ where $A,A^c$ are the two split subsets and $\hat{f}_A(\cdot;w_A)$ is a linear regressor built on the corresponding subset. The discrepancy of using different models for optimizing a split and predicting at the leaves, was first pointed out by Karalic [@karalic1992employing], at the time however, solving the full optimization problem that results from using a piecewise linear model as part of the splitting criterion was neither tractable nor numerically stable. This intractability and instability was asserted many times in subsequent years [@dobra2002secret; @Potts2005; @Nata], and in the view of solving this optimization problem, multiple efforts were made to efficiently approximate the above split criterion. Two such approaches [@GUIDE; @SUPPORT] use the residuals of a linear regressor at each node to perform a kind of clustering, while another approach [@vogel07scalable] replaces linear regression with forward selection regression at each node in order to reduce the dimensionality of the models.
We mention here, for the sake of completeness, that though greedy top-down growing of trees is the most common approach, there exist other approaches, such as so-called “soft” trees [@chipman2010] which predefine the tree architecture and then optimize, in a global manner, the split functions. Other non-greedy strategies for learning the split criteria have also been studied in the literature, as for example via multi-linear programming [@benett1994global] and structured prediction [@norouzi2015efficient].
We focus however on the specific, sometimes called “hard”, family of tree algorithms. We show in the following that despite the increased computational cost of using the criterion in Equation , modern GPU technology ensures that this approach is both tractable and numerically stable. More importantly we provide a comprehensive theoretical analysis of this model class, by separating the “split” space $\mathbf{\mathcal{X}^{'}}$ from the “regression” space ${\mathcal{X}}$ we show that the generalization capabilities of the model class are linked to the dimensionality of the latter as well to the regularization constraints, which is induced by some norm on ${\mathcal{X}}$ (in particular we consider $\ell_{1}-$norm and LASSO, which is well-known for inducing the sparcity).
The paper is organized as follows: in Section \[sec:prob\_setup\_new\] we introduce the notation, present the general framework of Piecewise-Linear Regression Tree (PLRT) algorithm and provide necessary formalism to describe the model class. Section \[sec:theor\_analysis\] contains main theoretical contribution of the paper, namely the upper bound for the Rademacher complexity of the underlying regression trees class and the high probability inequality which controls the deviations of the generalization error. Futhermore, in this Section we provide the important corollaries which relate the different types of penalization (i.e. those which induce the sparsity) on the leaves to the performance of the risk bounds. Section \[sec:ctn\] is devoted to the question of numerical tractability and stability, while computing the regressor estimates on GPU, whereas Section \[sec:emp\_eval\] reports the empirical results of wide range of experiments, showing practical advantage of using GPU combined with superiority of models returned by PLRT algorithm (in comparison to well-known decision trees baselines). Finally, we conclude in Section \[sec:conc\], by highlighting the possible extensions. All the proofs can be found in the Appendix.
Problem setup {#sec:prob_setup_new}
=============
Regression algorithm {#subsec: regression_alg}
--------------------
Let $\mathcal{X} = \mathbb{R}^{d}$, equipped with euclidean norm $\norm{\cdot}_{2}$ and $\mathcal{Y} \subset \mathbb{R}$ be some (closed) interval of the real line. Denote through $\mathbf{S}=\{\bx_{i},y_{i}\}_{i=1}^{n}$ the i.i.d. sample of of size $n$ from some unknown distribution $\mathbb{P}_{{X, Y}}$ over the space $\mathcal{X}\times \mathcal{Y}$. For some input $x \in \mathcal{X}$ we also consider its feature representation in some feature space $\mathcal{X}^{'}$ and denote it as $\psi(x)$. In this work we consider the case where $\mathcal{X}^{'} = \mathbb{R}^{D}$ and denote for an arbitrary point $\bx_{i}$ its representation $\psi(\bx_{i}) := \psi_{i} = \paren{\psi_{i}^{1},\ldots,\psi_{i}^{D}}$ with $d \ll D$, where as usual through $\psi_{i}^{k}$ we denote the $k$-th coordinate of the feature representation $\psi_{i}$. Therefore, in our work we consider the *extended sample* $\mathbb{S} =\{\bx_{i},y_{i},\psi_{i}\}_{i=1}^{n}$. Lastly, we assume (for the simplicity of the theoretical analysis) that the distribution of $\norm{X}_{2}$ has bounded support in the interval $[0,K]$. We will use the notation $[D]$ for the integer interval $\set{1,\ldots,D}$.
We investigate both the theoretical properties and empirical performance of a general form piecewise linear regressor with regularization constraints built via the the following regression tree algorithm. The proposed algorithm proceeds in a top-down greedy fashion as is typically done in building a decision tree. At each node in the tree, a split is chosen in the $\mathcal{X}^{'}$ space so as to minimize the empirical least square errors of linear predictors in the $\mathcal{X}$ space after splitting.
Let ${\mathcal{J}}$ be a subset of indices of the training dataset, corresponding to instances present at a given tree node. For all $(i,k) \in [D]\times {\mathcal{J}}$ denote $A_{i,k}=\{ j \in {\mathcal{J}}: \psi_{j}^{i} \geq \psi_{k}^{i} \}$ and $A^{c}_{i,k}= {\mathcal{J}}\setminus A_{i,k}$ corresponding to the subsets obtained after a split according to the $i$-th feature coordinate and threshold $\psi^i_{k}$. Define the matrix $\bX_{A_{i,k}}$ of dimensions $(|A_{i,k}|,d)$ whose lines are given by $(\bx^t_{l}, l \in A_{i,k})$, and similarly $\bX_{A^{c}_{i,k}}$ of dimensions $(|A^c_{i,k}|, d)$. Define label vectors $Y_{A_{i,k}}=\paren[1]{y_{l}: l \in A_{i,k}}^t$ and $Y_{A^{c}_{i,k}}=\paren[1]{y_{l}: l \in A^{c}_{i,k}}^t$ of dimension $|{A_{i,k}}|$, $|{A^{c}_{i,k}}|$ respectively.
For every $(i,k) \in [D] \times {\mathcal{J}}$ consider the optimal cumulative penalized loss of linear predictors after splitting: $$\begin{gathered}
\label{eq:L_diff}
L^\lambda_{i,k}= \min_{w_{i,k} \in \mathbb{R}^{d}} \paren{ \norm[1]{\bX_{A^{}_{i,k}}{w}^{}_{i,k}-Y_{A^{}_{i,k}}}_{P}^{2} + \lambda \| w_{i,k} - w_0\|^2_Q} \\ + \min_{w^c_{i,k} \in \mathbb{R}^{d}} \paren{\norm[1] {\bX_{A^{c}_{i,k}}{w}^{c}_{i,k}-Y_{A^{c}_{i,k}}}_{P}^{2} + \lambda \| {w}^{c}_{i,k} - {w}_0\|^2_Q},\end{gathered}$$ where $\norm{x}^{2}_{P}:= x^{\top}Px$ and $w_{0}$ may be any vector, though in the following we set $w_0$ to be the linear regression vector of the parent node. In particular, if for a given leaf $s$ we have computed an optimal regularized least squares solution $w_s^*$; when further splitting this node into two (children) leaves we optimize problems of the form $ \min_{w_{i,k} \in \mathbb{R}^{d}} \paren{ \norm[1]{\bX_{A^{}_{i,k}}{w}^{}_{i,k}-Y_{A^{}_{i,k}}}_{P}^{2} + \lambda \| w_{i,k} - w_s^*\|^2_Q}$ and in the similar form for $A_{i,k}^{c}$. Thus, the solution at lower nodes are regularized by the propagation of the solutions from higher (parential) nodes. As will be shown in the empirical evaluation, such a regularization is crucial to the stability of the proposed algorithm. This was in fact a key insight of early experiments, without strong regularization of this form the proposed trees were not able to generalize well. We note that $L_{i,k}$ can be analytically computed by standard linear algebra formulas, see Section \[sec:ctn\] for a more detailed discussion of numerical aspects.
As the optimal empirical splitting rule, we choose a pair $(i^{\star},k^{\star})$ minimizing the above: $$(i^{\star},k^{\star}) = \argmin_{(i,k) \in [D] \times {\mathcal{J}}} L^\lambda_{i,k}.$$ For the pair $(i^{\star},k^{\star})$, we split the data accordingly and get the subsets $\paren{\bx_\ell,y_\ell}_{\ell \in {A_{i^{\star},k^{\star}}}}$ and $\paren{\bx_\ell,y_\ell}_{\ell \in {A^c_{i^{\star},k^{\star}}}}$ for the two children of the node. Now, provided that none of the stopping criteria has been reached we apply the splitting algorithm recursively to these two subnodes. If the stopping criteria is reached, then the linear regression with constraints is performed on the leaves.
Regression tree formalism
-------------------------
For the learning-theoretic study of the algorithm, we need to define formally the set of possible decision functions that can be output by the learning algorithm. This set must be data-independent for classical learning-theoretic arguments to apply.
We assume that the total number of leaves is fixed and equal to $\ell$. Let $\mathcal{T}_{\ell}$ be the set of all binary (unlabeled) trees with $\ell$ leaves. For each tree $T \in \mathcal{T}_{\ell}$, denote by $T^{\circ}$ its interior nodes, and by $\partial T$ its leaves. Consider some interior node $s \in T^{\circ}$ and some splitting rule according to the algorithmic scheme in Section \[subsec: regression\_alg\]. Notice that a pair $(i_{s},t_{s}) \in [D]\times \mathbb{R}$ fully parametrizes the splitting criteria in the form $\mathbbm{1}\paren{{\psi^{i_{s}}(\bx) \geq t_{s}}}$ for some instance $\bx$. Any split corresponding to a set $A_{i,k}$ which can be obtained by the partition procedure of the PLRT algorithm can be represented as a pair $(i_s=i,t_s=\psi^{i_s}_k)$ for some $k\in [n]$. Finally, for each leaf $ L_{i} \in \partial T$, $i \in \{1,\ldots,\ell\}$, the local prediction function at leaf $L_i$ is a linear predictor $\bx \mapsto \inner{f_i,\bx}$, which we parametrize by the vector $f_i$, under the constraint $f_{i} \in B$; we will consider the constrained classes of the form $ B := \{ \norm{f} \leq W\} \subset \mbr^d$, where $\norm{\cdot}$ is some norm in $\mathbb{R}^{d}$, and track the influence of norm constraints (for some specific choice of norm) on the complexity terms.
Using the aforementioned notation, we can describe class of regression decision trees as follows: $$\begin{aligned}
\label{func_class001}
\mathcal{F} := \{f : f = \paren{ T, (i_{s},t_{s})_{s \in T^{\circ}},(f_{k})_{k \in \partial T}}, T \in \mathcal{T}_{\ell}, \paren{i_s,t_s} \in [D]\times \mathbb{R}, f_{k} \in B \}.\end{aligned}$$ The main aim of the next section is to obtain error bounds for statistical performance of the functional class by means of the deviation of its generalization error. The main technical tool is the concept of Rademacher complexity.
Theoretical analysis {#sec:theor_analysis}
=====================
Preliminaries and aim.
----------------------
In the following analysis, we obtain high probability upper bounds on the deviation of the *statistical risk* $L(f) = E_{P_{X,Y}}[\ell(f(\bx),y)]$ of the model $f \in\mathcal{F}$ from *empirical risk* $\hat{L}_{n}(f) = \frac{1}{n}\sum_{i=1}^{n}\ell(f({\bx_{i}}),y_{i})$ uniformly over the model class $\mathcal{F}$. We define the *generalization error* of the class $\mathcal{F}$, as $Z := \sup\limits_{f \in \mathcal{F}} \big(L(f) - \hat{L}_{n}(f) \big)$. The main result of this section is a high probability upper bound on the deviations of $Z$ under the different assumptions on the regularization constraints in $B$.
The expectation and deviation of the generalization error is the typical measure to control the statistical performance of the underlying model class and it was widely studied in [@bartlett2002rademacher],[@koltchinskii2001rademacher]. In these works the framework of Rademacher complexity is used as the complexity measure for structured regularized risk estimation methods, including kernel methods.
Let now $\bm{\sigma}:=(\sigma_{i})_{i=1}^{n}$ be a $n$-vector of i.i.d. random variables independent of $\mathbb{S}$, uniformly distributed over $\{-1,+1\}$ (Rademacher random variables). For a given loss function $l$, such that $|l(\cdot,\cdot)| \leq C$ and for all $f$ in (an arbitrary) model class $\mathcal{G}$, with probability at least $1-\delta$ it holds (see for example [@bartlett2002rademacher]):
$$L(f) \leq L_{n}(f)+2\mathfrak{R}_{n}(l \circ \mathcal{\mathcal{G}}) + C\sqrt{\frac{\log \delta^{-1}}{2n}},
\label{eq:gen_error}$$
where $$\begin{aligned}
\mathfrak{R}_{n}(l \circ \mathcal{F}) := E \left[\sup\limits_{f \in \mathcal{F}} \frac{1}{n} \sum\limits_{i=1}^{n}\sigma_{i}l(f(\bx_{i}),y_{i})) \right],\end{aligned}$$ is the *Rademacher complexity* of the model class $\mathcal{F}$, the last expectation being taken under the product distribution of $(\mbs,\bm{\sigma})$, where $l \circ \mathcal{G}:= \{ (\bx,y) \mapsto l(f(\bx),y) | f \in {\mathcal{G}} \}$ is the loss function class associated to $\cG$. If we treat the sample $\mathbb{S}$ as fixed, we introduce the *empirical Rademacher complexity*
$$\hat{\mathfrak{R}}_{\mbs}(l \circ \mathcal{F}) := E_{\bm{\sigma}}\left[ \sup\limits_{f \in \mathcal{F}} \frac{1}{n}\sum\limits_{i=1}^{n}\sigma_{i}l(f(\bx_{i}),y_{i}))\right],
\label{eq:empirical_rademacher_f}$$
so that $\mathfrak{R}_{n}(l \circ \mathcal{F}) = E_{\mathbb{S} \sim P^{\otimes n}_{x,y}}\left[\hat{\mathfrak{R}}_{\mbs}(l \circ \mathcal{F})\right] $. It is also known (see for example [@bartlett2005local]) that if the loss-function $\ell$ is $L$-Lipschitz in its second argument,i.e. $\forall y \in \mathbf{Y}$ $ |l(a,y)-l(b,y)| \leq L|a-b| \; \forall a,b \in \mathbb{R}$ then: $$\hat{\mathfrak{R}}_{\mbs}(l \circ \mathcal{F}) \leq L \hat{\mathfrak{R}}_{\mbs}(\mathcal{{F}}),
\label{eq:ledoux}$$ and for we get for all $f$ in the class $\mathcal{F}$: $$\label{eq:gen_bound}
L(f) \leq L_{n}(f)+2L \mathfrak{R}_{n} ( \mathcal{F}) + C\sqrt{\frac{\log\delta ^{-1}}{2n}}.$$ Now, for the PRLT learning algorithm we consider the squared loss function $l(y',y)=(y-y')^2$. Recall that from the previous assumptions we have $\norm{\bx}_{2} \leq K$ and $|y| \leq R$ almost surely with respect to ${P}_{X,Y}$.
Assume also that for all $\bx$ with $\norm{\bx}_{2}\leq K$ and $f\in \cF$ it holds that $\abs{f(\bx)} \leq F$ (for linear predictors, the constant $F$ depends on the norm-constraints in the set $B$ and will be specified later). Then, we have easily for the squared-loss function: $$\begin{aligned}
|l(\cdot,\cdot)| \leq (R+F)^2 := L_{R,F,B}.\end{aligned}$$ Furthermore, we can control $\mathfrak{R}_{n}(\mathcal{F})$ by means of its empirical version $\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F})$, since we observe that changing one point in $\mathbb{S}$ changes $\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F})$ by at most $\frac{F}{n}$, so that by McDiarmid’s inequality, with probability at least $1-\delta/2$ it holds: $$\begin{aligned}
\mathfrak{R}_{n}\paren{\mathcal{F}} \leq \hat{\mathfrak{R}_{\mathbb{S}}}\paren{\mathcal{F}} + F\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}.\end{aligned}$$
This argument, together with the aforementioned reasoning, implies that in order to obtain high-probability upper bounds for the generalization error $Z$ of model class $\mathcal{F}$, it is sufficient to obtain bounds on its empirical Rademacher complexity based on the sample $\mathbb{S}$. More precisely, with probability at least $1-\delta/2$ it holds for all $f \in \mathcal{F}$ that: $$\begin{aligned}
\label{eq:risk_ineq}
L(f) \leq L_{n}(f) + 2L\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} + 2LF\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}} + L_{R,F,B}\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}.\end{aligned}$$ In the next part we concentrate mainly on the bounds on the (empirical) Rademacher complexity of the model class $\mathcal{{F}}$ of PLRT, but also provide bounds on the true Rademacher complexity.
(Empirical) Rademacher complexity and generalization error deviation bound for the class of PLRT
------------------------------------------------------------------------------------------------
Recalling the formal definition of the predictor class in , we observe that for a fixed a element (tree) $T \in \mathcal{T}_{l}$, and some partition generated by the family $(i_{s},t_{s})_{s \in T^{\circ}} \in S$, we obtain a submodel class $\mathcal{{F}}_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}$ which is a product of the decision models over the leaves, and such that each $\bx_{i} \in \mathbb{S}$ belongs to exactly one leaf $L_{j}$. Thus, the whole model class $\mathcal{F}$ can be represented as follows: $$\begin{aligned}
\label{eq:class_str}
\mathcal{F} = \bigcup_{\substack{T \in \cT_l\\(i_{s},t_{s})_{s} \in ([D] \times \mbr)^{\cT^\circ}}}\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}} ,\end{aligned}$$ where we formally write $\mathcal{F}_{T,(i_{s},t_{s})_{s\in \cT^\circ}} = \{f: f = (f_{1},\ldots,f_{\ell}): \forall x \in \mathcal{X}, f(x) = \sum_{j} \inner{f_{j},x}\mathbbm{1}(x \in L_{j}), f_{j} \in \partial T \} $ for each tree $T$ and split family $\paren{i_{s},t_{s}}_{s \in T^{\circ}}$. To study the statistical performance of the classes with the union-type structure we make use the of Lemma 2 from [@maurer2014inequality] and its corollary, which we provide below for completeness.
\[thm:sup\_bound\] Let $N \geq 4$ be some natural number and $A_{1},\ldots,A_{N} \subset \mathbb{R}^{n}$ some subsets, such that for a given $A \subset \mathbb{R}^{n}$ we have $A =\bigcup_{i=1}^{N} A_{i}$. Consider $\bm{\sigma} = \paren{\sigma_1,\ldots,\sigma_n}$ to be a vector of i.i.d. Rademacher variables (i.e. uniformly distributed over $\set{-1,1}^n$). Then we have: $$\begin{aligned}
\ee{}{\sup_{z \in A} \inner{\bm{\sigma},z}} \leq \max_{i=1}^{N}\ee{}{\sup_{z \in A_{i}}\inner{\bm{\sigma},z}} + 4 \sup_{z \in A}\norm{z}\sqrt{\log{N}}.
\end{aligned}$$
Assume we have a finite family of functional classes $\mathcal{F},\mathcal{F}_{1},\ldots,\mathcal{F}_{N}$, such that $\mathcal{F} = \bigcup_{i=1,\ldots,N} \mathcal{F}_{i}$ and a sample $\mathbb{S}$ as before. Denote $A_{j} \subset \mathbb{R}^{n}, A_{j} = \{\paren{f(\bx_{1}),\ldots,f(\bx_{n})}: f \in \mathcal{F}_{j} \}$, $j \in \{1,\ldots, N\}$ (i.e the vector image of the evaluation of function $f \in \mathcal{{F}}_{j}$ on the sample $\mbs$). From the previous result, we have the next corollary (see also [@maurer2014inequality]): Now, assume we have a finite family of functional classes $\mathcal{F},\mathcal{F}_{1},\ldots,\mathcal{F}_{N}$, such that $\mathcal{F} = \bigcup_{i=1}^{N} \mathcal{F}_{i}$ and a sample $\mathbb{S}$ as before. Denote $A_{j} \subset \mathbb{R}^{n}, A_{j} = \{\paren{f(\bx_{1}),\ldots,f(\bx_{n})}: f \in \mathcal{F}_{j} \}$, $j \in \{1,\ldots, N\}$ (i.e the vector image of the evaluation of function $f \in \mathcal{{F}}_{j}$ on the sample $\mbs$). From the previous result, we have the next corollary (see also [@maurer2014inequality]):
\[cor:rad\_union\] For the empirical Rademacher complexity of the class $\mathcal{F}= \bigcup_{i=1,\ldots,N} \mathcal{F}_{i}$ based on the sample $\mathbb{S}$, we have: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} \leq \max_{m=1}^{N}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{m}) + 4\mathcal{M}\sqrt{\frac{\log{N}}{n}},
\end{aligned}$$ where $\mathcal{M} = \sqrt{\sup_{f \in \mathcal{F}}\frac{1}{n}\sum_{i=1}^{n}f^{2}(\bx_{i})}$.
For the analysis of the class of regression decision trees algorithms , this means that we can in principle first analyze the Rademacher complexity of the predictor class for any fixed tree structure and splits, and then pay as a price the $\log$-cardinality of the union appearing in . One issue is that (because of the real-valued thresholds $t_s$) this union is not finite; however using a classical argument we can reduce it to a finite union when considering the empirical Rademacher complexity.
First bound we obtain by a simple upper bounding of the number of the possible splits ($((n-1)D)^{\ell-1}$) in the tree and the number of different trees with $\ell$ leaves (which is the $\ell-1$ Catalan’s number). Futhermore, by using simple inequalities $(\frac{k}{e})^k<k!<e(\frac{k}{2})^k$ we obtain $\frac{1}{\ell} \binom{2(\ell-1)}{\ell-1} \leq \frac{e^{\ell}}{\ell}$ and the second claim follows by simple calculations.
\[lem: help\_lem01\] $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}} \paren{\mathcal{F}}
= \hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{\substack{T \in \cT_l\\(i_s,t_s)_{s} \in ([D] \times [n])^{T^\circ}}} \mathcal{F}_{T,(i_{s},\psi_{k_s}^{i_s})_{s\in T^{\circ}}}}
&\leq \max_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}) + 4\mathcal{M}\sqrt{\frac{\ell\log{enD}}{n}}, \\
\end{aligned}$$ where $\mathcal{M} = \sqrt{\sup\limits_{f \in \cF}\frac{1}{n}\sum_{i=1}^{n}f^{2}(\bx_{i})}$.
With the notation as in Section \[sec:prob\_setup\_new\] we obtain:
\[thm:main\_theorem01\] With probability at least $1-\delta/2$ it holds uniformly over all $f \in \mathcal{{F}}$: $$\label{eq:Main_inequality01}
L(f) - L_{n}(f) \leq 4\paren{R+F} \paren[3]{\max_{T,(i_{s},t_{s})_{s\in T^\circ}}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^\circ}})+ 4\mathcal{M}\sqrt{\frac{\ell \log(enD)}{n}}
+ (R+2F) \sqrt{\frac{\log{ \paren{\frac{2}{\delta}}}}{2n}}}
$$ where we recall that $\abs{f(\bx)} \leq F$ uniformly over all $\bx$ and $\abs{y} \leq R$.
In the next section we specify the constraints induced by the set $B$, and give explicit bounds for both Rademacher complexity and generalization error for $\ell_{2}$ and $\ell_{1}$ norm constraints.
Different penalty constraints and corresponding bounds
------------------------------------------------------
In what follows we denote $\Sigma = \ee{}{\bx \bx^{\top}}$ the covariance matrix of the random vector $\bx$, and $\hat{\Sigma} = \frac{1}{n} \sum_{i=1}^n \bx_i\bx_i^t$ its empirical counterpart.
### Euclidean-norm penalty.
Let $B = \{ \norm{f}_{2} \leq W \}$. Recall that our goal is now to control $\max_{T,(i_{s},t_{s})_{s \in T^{\circ}}} \hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}})$ for decision trees of total number of leaves $\ell$. The following result holds true.
\[eq:l2\_bound\] For a fixed tree structure $T$ with $\ell$ leaves and the splits $(i_{s},t_{s})_{s \in T^{\circ}}$ we have the following upper bounds: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} &\leq \frac{W\sqrt{2\ell} \sqrt{\sum_{j}\norm{\bx_{j}}_{2}^{2}}}{n} = W\sqrt{\frac{2 \ell \tr{\hat{\Sigma}}}{n}};\\
{\mathfrak{R}}_{n}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} &\leq W\sqrt{\frac{{2\ell \tr{{\Sigma}}}}{n}}.
\end{aligned}$$
Notice that the upper bound for the empirical Rademacher complexity is data-dependent, but does not depend on the structure of the decision tree with fixed splits.
\[lem:rademacher\_ell2\] By linearity of expectation and taking into account that leaves $(L_{j})_{j=1}^{\ell}$ are disjoint sets we obtain: $$\begin{aligned}
\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\epsilon_{i}f(\bx_{i})} &= \sum_{j=1}^{\ell}\ee{}{\sup_{f_{j} \in B} \inner{f_{j},\sum_{i: \bx_{i} \in L_{j}}^{}\epsilon_{i}\bx_{i}}} \\
& \leq W \sum_{j=1}^{\ell} \ee{}{\norm[2]{\sum_{i: \bx_{i} \in L_{j}}^{}\epsilon_{i}\bx_{i}}_{2}} \\
& \leq W\sum_{j=1}^{\ell}\sqrt{\ee{}{\sum_{i,k: \bx_{i},\bx_{k} \in L_{j}}^{}\epsilon_{i}\epsilon_{k}\inner{\bx_{i},\bx_{k}}}},
\end{aligned}$$ where we used Jensen’s inequality in the last line. Since $(\epsilon_{i})_{i=1}^{n}$ are independent Rademacher variables, by taking expectation, we obtain: $$\begin{aligned}
\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\epsilon_{i}f(\bx_{i})} \leq W \sum_{j=1}^{\ell}\sqrt{\sum_{i: \bx_{i} \in L_{j}}\norm{\bx_{i}}^{2}} \leq W \sqrt{\ell} \sqrt{\sum_{i=1}^{n}\norm{\bx_{i}}^{2}},
\end{aligned}$$ where in the last inequality we just used the Jensen’s inequality for the function $t \mapsto \sqrt{t}$. The claim of the Lemma now follows from the definition of Rademacher Complexity.
Plugging in the result of Lemma \[eq:l2\_bound\] into Lemma \[lem: help\_lem01\] and consequently plugging the implication of the latter into the general result of Theorem \[thm:main\_theorem01\] we deduce the following general result.
\[eq:rad\_comp\_ell\_2\] Let the function class $\mathcal{F}$ be as given in Equation with $B=\{\norm{f}_{2} \leq W\}$. Assume $\norm{X}_{2} \leq K$ and $\abs{Y} \leq R$, $\mathbb{P}_{X}$ and $\mathbb{P}_{Y}$-almost surely. Then the following upper bounds for the empirical and true Rademacher complexities of the class $\mathcal{F}$ hold: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} & \leq
W \sqrt{\frac{\ell}{n}} \paren{\sqrt{2\tr{\hat{\Sigma}}} + 4\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log(enD)}};
\\
{\mathfrak{R}}_{n}\paren{\mathcal{F}} &\leq W \sqrt{\frac{\ell}{n}} \paren{\sqrt{2\tr{\Sigma}} + 4 \sqrt{\ee[1]{}{\norm[1]{\hat{\Sigma}}_{op}}} \sqrt{\log(enD)} }.
$$ where we denote the operator norm of a matrix as $\norm{\cdot}_{op}$. Furthermore, with probability at least $1-\delta$ it holds: $$\begin{aligned}
L\paren{f} - L_{n}\paren{f} \leq 4\paren{R+F} \paren{ W\sqrt{\frac{\ell}{n}}\paren{\sqrt{2\tr \hat{\Sigma}} + 4 \sqrt{\norm[2]{\hat{\Sigma}}_{op} }\sqrt{\log\paren{enD}} } + \paren{R+2F}\sqrt{\frac{\log\paren{\frac{2}{\delta}}}{2n}}},
\end{aligned}$$ or similarly in data-independent form (when using bound on Rademacher complexity instead of its data-dependent proxy) we have: $$\begin{aligned}
L\paren{f} - L_{n}\paren{f} \leq 4\paren{R+F} \paren{ W \sqrt{\frac{\ell}{n}} \paren{\sqrt{2\tr{\Sigma}} + 4 \sqrt{\ee[1]{}{\norm[1]{\hat{\Sigma}}_{op}}} \sqrt{\log(enD)} } + \paren{R+F}\sqrt{\frac{\log\paren{\frac{2}{\delta}}}{2n}}}
\end{aligned}$$
Furthermore, with probability at least $1-\delta/2$, we have for all $f \in \mathcal{{F}}$: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1} \paren[3]{W\sqrt{\frac{\ell}{n}} \paren{\sqrt{2\tr{\hat{\Sigma}}} + 4 \sqrt{{\norm[1]{\hat{\Sigma}}_{op}}} \sqrt{\log(enD)} } + WK\sqrt{\frac{\log(\frac{2}{\delta})}{2n}}} + C\sqrt{\frac{\log({\frac{2}{\delta}})}{2n}}.
\end{aligned}$$
Alternatively, using the bound on the true Rademacher complexity we obtain with probability at least $1-\delta$: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1}W \sqrt{\frac{\ell}{n}} \paren{ \sqrt{2\tr{\Sigma}} + 4
\sqrt{\ee[1]{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\log{enD}}}
+C\sqrt{\frac{\log({\frac{1}{\delta}})}{2n}},
\end{aligned}$$ where $C_{1} = 4 \paren{R+WK}$, $C:=\paren{R+WK}^{2}$.
\[rem:gen\_error\] Consider the simple case $\Sigma = \mathbbm{I}_{d}$. Then, in the last bound, the first term scales as $\tr{\Sigma} = d$, but the second term scales like $\e[1]{\norm[1]{\hat{\Sigma}}_{op}}$, which is, due to classical matrix concentration results [@Tropp:15], up to a $\log(d)$ factor the same as $\norm{\Sigma}_{op} = 1$. This implies: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1} W \sqrt{\frac{\ell}{n}} \paren{ {\sqrt{2d} + 8C_{3}
\sqrt{\log{d}}\sqrt{\log{enD}}}} +C\sqrt{\frac{\log({\frac{1}{\delta}})}{2n}},\end{aligned}$$ where $C_{3}$ is some universal numerical constant and $C=(R+WK)^{2}$, $C_{1}=4(R+WK)$. In this case, the first term contributes to the bound exponentially more ($d$ against $\log d$) comparing to the second. Furthermore, if the eigenvalues of $\Sigma$ decay geometrically fast, i.e. $\lambda_{k} \leq \lambda\rho^{k-1}$, with $0 < \rho <1$ for $k \in [d]$, we have $\tr{\Sigma} \leq \frac{\lambda}{1-\rho}$ and $\norm{\Sigma}_{op} = \lambda$ and the bound transforms to: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1}W \sqrt{\frac{\ell}{{n}}} \paren{
{\sqrt{\frac{2\lambda}{1-\rho}} + 4C_{3}\sqrt{\lambda}\sqrt{\log{d}}\sqrt{\log{enD}}}} +C\sqrt{\frac{\log({\frac{1}{\delta}})}{2n}}.\end{aligned}$$ Here the balance between the first and second term depends on the trade-off between the log of dimensionality $d$ and the spectral decay parameter $\rho$.
### Lasso-type constraints. {#sec:lasso_type_constraints}
Consider now $B = \{f: \norm{f}_{1} \leq W \}$, where $\norm{\cdot}_{1}$ is the $\ell_{1}-$norm in the space $\mathbb{R}^{d}$. For a fixed tree $T$ and split family $\paren{i_{s},t_{s}}_{s \in T^{\circ}}$, the empirical Rademacher complexity $\hat{\mathfrak{R}}(\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}})$ can be upper bounded as follows:
\[lem:rad\_lasso\_con\] $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}) &\leq \frac{W\sqrt{2\ell\sum_{i=1}^{n}\norm{\bx_{i}}_{\infty}^{2}}}{{n}}\paren{1+4\sqrt{\log(d)}}.
\end{aligned}$$
Repeating the similar arguments as in the case with $\ell_{2}-$penalty we obtain: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} = \frac{2}{n}\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\epsilon_{i}f(\bx_{i})} &= \sum_{j=1}^{\ell}\frac{2}{n}\ee{}{\sup_{f_{j} \in B} \inner{f_{j},\sum_{i: \bx_{i} \in L_{j}}^{}\epsilon_{i}\bx_{i}}} \\
& = \sum_{j=1}^{\ell}\frac{r_{j}}{n}\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\hat{\mathcal{F}}_{j}},
\end{aligned}$$ where $r_{j}=\abs{L_{j}}$ and $\hat{\mathcal{F}}_{j}$ is the empirical Rademacher complexity of class of functions, such that $\norm{f}_{1} \leq 1$ computed on the $r_{j}$ samples. Using Proposition 3 from [@maurer2012structured] with $\mathcal{P}$ being the set of orthogonal projectors on the coordinates, we obtain (recalling that $\norm{x}_{2} \leq X$, which implies that $\norm{x}_{\infty} \leq X$): $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\hat{\mathcal{F}}_{j}} \leq \frac{W\sqrt{2\sum_{i: \bx_{i} \in L_{j}} \norm{\bx_{i}}^{2}_{\infty}}}{r_{j}} \paren{1+4\sqrt{\log{d}}},
\end{aligned}$$ which implies $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} \leq \frac{2}{n}W\paren{1+4\sqrt{\log{d}}}\sum_{j=1}^{\ell}\sqrt{2\sum_{i: \bx_{i} \in L_{j}} \norm{\bx_{i}}^{2}_{\infty}} \leq \frac{2^{\frac{3}{2}}\ell^{\frac{1}{2}}W\sqrt{\sum_{i}\norm{\bx_{i}}^{2}_{\infty}}}{{n}}\paren{1+4\sqrt{\log{d}}},
\end{aligned}$$ where in the last inequality we just used Jensen’s inequality in the similar fashion as for $\ell_{2}-$norm constaints. The Lemma is therefore proved.
In the similar way, as we used above, plugging in the result of Lemma \[lem:rad\_lasso\_con\] into \[lem: help\_lem01\] we get the following result.
\[prop:Lasso\_rad\_gen\] Let the function class $\mathcal{F}$ be as given by Equation with $B=\{\norm{f}_{1} \leq W\}$. Then the following upper bounds for the empirical and true Rademacher complexity of the class $\mathcal{F}$ hold: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} &\leq \frac{W\sqrt{\ell}}{\sqrt{n}} \paren{\sqrt{\frac{2}{n}\sum_{j}\norm{\bx_{j}}_{\infty}^{2}}\paren{1 + 4 \sqrt{\log{d}}} + 4\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{{\log{enD}}}}; \\ {\mathfrak{R}}_{n}\paren{\mathcal{F}} &\leq
\frac{W\sqrt{\ell}}{\sqrt{n}} \paren{\sqrt{2\ee{}{\norm{\bx_{j}}_{\infty}^{2}}}\paren{1 + 4 \sqrt{\log{d}}} + 4\sqrt{ \ee[1]{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{{\log{enD}}}}.
\end{aligned}$$ Furthermore, in the similar vein to the $\ell_{2}$ regularization case, bounds on the generalization error is obtained by simply applying Theorem \[thm:main\_theorem01\] and plugging-in the upper bounds on the empirical Rademacher complexity. with probability at least $1-\delta/2$ we have for all $f \in \mathcal{{F}}$: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1} \paren{ \frac{W\sqrt{\ell}}{\sqrt{n}} \paren{\sqrt{\frac{2}{n}\sum_{j=1}^{n}\norm{\bx_{j}}_{\infty}^{2}}\paren{1 + 4 \sqrt{\log{d}}} + 4\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{{\log{enD}}}} + WK\sqrt{\frac{\log(\frac{2}{\delta})}{2n}}} \\
& + C\sqrt{\frac{\log(\frac{2}{\delta})}{2n}}.
\end{aligned}$$ Or an alternative (distribution dependent bound) in the form: $$\begin{aligned}
L(f) \leq L_{n}(f) & + C_{1}\paren{ \frac{W\sqrt{\ell}}{\sqrt{n}} \paren{2^{\frac{1}{2}}\sqrt{\ee{}{\norm{\bx}_{\infty}^{2}}}\paren{1 + 4 \sqrt{\log{d}}} + 4\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{{\log{enD}}}}} + C\sqrt{\frac{\log(\frac{2}{\delta})}{2n}},
\end{aligned}$$ where $C_{1} =4(R+WK)$, $C = (R+WK)^{2}$.
Discussion of the results
-------------------------
\[rem:gen\_error\_lasso\] Notice that the second complexity term in the bound on the Rademacher complexity in Proposition \[prop:Lasso\_rad\_gen\] is the same as in the analogue result with the $\ell_{2}$ constraints. For the first complexity term, it holds $\frac{1}{n}\sum_{j=1}^{n}\norm{\bx_{j}}^{2}_{\infty} \leq \frac{1}{n}\sum_{j=1}^{n}\norm{\bx_{j}}^{2}_{2} = \tr{\hat{\Sigma}} $ and correspondingly $\ee{}{\norm{\bx}_{\infty}^{2}} \leq \tr{\Sigma}$. On the other hand, we have the additional factor of order $\sqrt{\log{d}}$ due to LASSO-type of constraints. Thus, whether LASSO-type regularization gives better bounds for generalization error in comparison to $\ell_{2}$ penalty depends on the magnitude of the ratio $r := \tr{(\Sigma)}/{\ee{}{{\norm{\bx}}_{\infty}^{2}}}$ (or its empirical counterpart $\hat{r} := \tr{(\hat{\Sigma})}/({\frac{1}{n}\sum_{j}\norm{\bx_{j}}_{\infty}^{2}})$). Namely, the LASSO-type penalty approach gives a better bound if $r$ or $\hat{r}$ are of order $\log{d}$, where $d$ is the dimensionality of the regression space. In the simple case $\Sigma=\mathbbm{I}_d$ and assuming the coordinates have subgaussian distribution, we will in fact have $\tr{\Sigma}=d$ while $\ee{}{\norm{\bx}_{\infty}^{2}} \lesssim \log(d)$.
Theoretical analysis of decision trees ( in particular regression trees) is certanly not new topic and has been considered in various settings before. For the case of $1$-dim classification, bounds for generalization error with $0/1$ loss is given in [@Mansur:99]. Furthermore, for multi-class classification the analysis of statistical performance (in terms of bound for generalization error) of a brode class of composite decision trees and forests is given in [@Salvo:16]. Although it is formally incorrect to compare these results to our setting, our bounds have the same order of magnitude in terms of the sample size and number of leaves as in [@Mansur:99]. Furthermore, the error bounds of scale linearly with the number of leaves, whereas we have sublinear scale ($\sqrt{\ell}$) while having the same order of magnitude in the number of training samples ($1/\sqrt{n}$). In the regression setting the recent work [@Lauer:17] author considers the theoretical analysis of the switching regression framework (which includes the framework of regression decision trees). However, this framework does not consider the regularization of the models on the partitioned space. In such setting, using similar toolbox of the Rademacher complexity for statistical error conctrol, in Equation $(16)$ author obtains risk upper bound which scales linearly with the number of partition cells (which corresponds to the number of leaves in regression trees scenario and comparably worse to our setting). Also, author provides only the majorant term in the upper bound, whereas (as we stressed out before) in our paper we investigate the influence of the smaller order terms (in terms of scaling in logs of dimension and of the norm of covariance matrix). Through usage of more advanced chaining argument, in the section 3.B.2 author shows an upper bound on the Rademacher complexity in the case of linear regression which matches (in the dominant term) the worse case of our bound of Proposition \[eq:rad\_comp\_ell\_2\] (when empirical covariance matrix is close to identity).
Notice that our analysis directly imply the bounds for the well-known particular (and much simpler) case of algorithm where the binary regression tree structure are combined with *constant* prediction models on the leaves. This algorithm is know as CART (see original work [@Breiman:84],also in [@Donoho:97]). Using the formalism we considered in our setting, this type of model class can be described as follows: $$\begin{aligned}
\label{func_class_pc}
\mathcal{F}_{c} := \{\overline{f} : \overline{f} = \paren{ T, (i_{s},t_{s})_{s \in T^{\circ}},(\overline{f}_{k})_{k \in \partial T}}, T \in \mathcal{T}_{\ell}, \paren{i_s,t_s} \in [D]\times \mathbb{R}, \overline{f}_{k}= \frac{1}{\abs{\partial T_{k}}}\sum_{i}^{}X_{i}\mbi_{ X_{i} \in \partial T_{k}} \},
\end{aligned}$$ whereas splits are done by minimizing emprical $\ell_{2}-$error. In the work [@Nedelec:03] the statistical performance of pruned CART decision trees (where prunning is perfomed from the maximal tree) of type \[func\_class\_pc\] was analyzed in different setting to ours. Namely, in [@Nedelec:03] the oracle-type inequalities for a fixed tree-structure in the frameworks of gaussian and bounded regression have been established. We notice, however, that our results actually applies to much broader setting which as the specific case includes CART. We point out that despite the estimation error will be the same, our PLRT methodology will have smaller empirical error than CART, since instead of greedy error minimization, it employs exact optimization. Since generalization error is bounded by empirical error “plus” estimation error, the second term is bounded by the same quantity for both methods but the first term is smaller for our method, the latter also a better bound on generalization error.
Combining PLRT with variable selection procedure {#subsec:variable_select}
------------------------------------------------
In this section we extend the setting by considering the application of some general feature selection rule (data-dependent, and considered as a black box) before applying the PLRT algorithm, and study its influence on the theoretical generalization performance.
Let $s \in \mathbb{N}$ be the target number of features, which a learner aims to select. Variable selection for the PLRT learning algorithm can be implemented in different ways, from which we consider the following. At step $0$ the learner can perform selection of $s$ coordinates from the $d-$dimensional vector and build a training sample $\mathbb{S}^{v} = \{\bx_{i}^{v}, \psi_{i},y_{i}\}_{i=1}^{n}$, where $\bx_{i}^{v}$ stands for vector $\bx_{i}$ which has only selected $s$ features given by an index vector $v$, and apply the learning procedure to the sample $\mathbb{S}^{v}$ instead of the original $\mathbb{S}$. Notice that in this case the dimensionality of the feature representation remains the same. In Appendix \[sec:appendix\] we also consider the alternate situation of interest when the feature selection is performed separately at each leaf of the tree (i.e. the the set of selected features can change from leaf to leaf).
Previous theoretical arguments of the deviations control for Rademacher complexity can be extended to the class of decision trees which includes variable selection procedure in the following way. For a given number $s$ define the set $\overline{S} = \{a \in \{0,1\}^{d}: \sum a_{i} =s \}$ and for vector $f \in \mathbb{R}^{d}$ denote through $f^{a}$ its $s-$dimensional restriction to the coordinates $i \in \{1,\ldots, d\}$, for which $a_{i} =1$. Define also for $\bw \in \mathbb{R}^{d}$, ${s_{0}}(\bw) = \sharp \{i: \bw_{i} \neq 0 \}$. Given some constraint set $B$ and a number $s \in \mathbb{N}$ we define $B^{s} = \{\bw: \bw \in B, s_{0}\paren{\bw}=s\}$.
Now consider the extension of the class of PLRT with variable selection (PLRTVS) procedure which can be formally presented as the follows: $$\begin{aligned}
\label{func_class02}
\mathcal{F}_{s} := \{f : f = \paren{ T, (i_{k},t_{k})_{k \in T^{\circ}},(f_{m})_{m \in \partial T}}, T \in \mathcal{T}_{\ell}, k\in T^{\circ} \paren{i_k,t_{k}} \in [D]\times\mathbb{R}, f_{m} \in B^{s} \}.\end{aligned}$$ We remark that the correction $B^{s}$ instead of $B$ captures all possible choices of the model $f$ on the leaves according to the constraint set $B$ with *arbitrary* $s$ coordinates (which were chosen by a variable selection procedure).
Using a similar argument, for a fixed tree $T \in \mathcal{T}_{l}$, some splits generated by sequence $(i_{k},t_{k})_{k \in T^{\circ}} $ and some set of variables $\overline{s} \in \overline{S}$ we have a model which is a union of the all possible models on the leaves and such that each $\bx_{i} \in \mathbb{S}$ belongs to exactly one leaf $L_{j}$ and is restricted to the coordinates indicated by $\overline{s}$ (and thus of dimension $s$). The model class $\mathcal{F}$ can be represented as follows: $$\begin{aligned}
\label{eq:union_select}
\mathcal{F}_{s} = \bigcup_{T,(i_{k},t_{k}),\overline{s}} \mathcal{F}_{T,(i_k,t_k)_{k \in T^{\circ}},\overline{s} \in \overline{S}},\end{aligned}$$ where $\mathcal{F}_{T,(i_{s},t_{s})_{s \in T^{\circ}},\overline{s}} = \{f^{\overline{s}}: f^{\overline{s}} = (f^{\overline{s}}_{1},\ldots,f^{\overline{s}}_{\ell}): \forall x \in \mathcal{X}, f^{\overline{s}}(x) = \sum_{j} \inner{f^{\overline{s}}_{j},x}\mathbbm{1}(x \in L_{j}), f^{\overline{s}}_{j} \in \partial T \}$.
Now, in order to study the Rademacher complexity of the class $\mathcal{{F}}_{s}$ we can apply the same scheme as before. Once more, because of the real-valued thresholds $t_k$ this union is not finite; however using a classical argument same as in the proof of Lemma \[lem: help\_lem01\] we reduce it to finite union.
\[lem:card02\]
$$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}} \paren{\mathcal{F}_{s}}
= \hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{T,(i_{k},t_{k}),\overline{s}} \mathcal{F}_{T,(i_k,t_k)_{k \in T^{\circ}},\overline{s} \in \overline{S}}}
&\leq \max_{T,i_{s},t_{s},\overline{s}}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,i_{s},t_{s},\overline{s}}) + 4\mathcal{M}\paren{\sqrt{\frac{\ell\log{enD} + s \log\paren{\frac{de}{s}}}{n}} }, \\
\end{aligned}$$
where $\mathcal{M} = \sqrt{\sup\limits_{f \in \cF}\frac{1}{n}\sum_{i=1}^{n}f^{2}(\bx_{i})}$.
Proof is based on the observation that the number of possible choices of $s$ coordinates from $d$ is $\binom{d}{s}$; the other part of the proof replies the argument made in the \[lem:card\] .
For the bound on the binomial coefficient we need use that $s! \geq \paren{\frac{s}{e}}^{s}$, from which we can derive that: $$\begin{aligned}
\binom{d}{s} = \frac{\prod_{j=0}^{s-1}\paren{d-j}}{s!} \leq \frac{d^{s}}{s!} \leq \paren{\frac{de}{s}}^{s}\end{aligned}$$
We demonstrate the influence of the variable selection on the statistical performance for LASSO-type penalization constraints on leaves. In a very similar vein, we provide data-dependent and distribution-dependent bounds for the empirical Rademacher complexity as well as the generalization error high probability upper bound. Using the analogues of Lemma \[lem: help\_lem01\] and Theorem \[thm:main\_theorem01\] for class $\mathcal{{F}}_{s}$ and LASSO-type regularization on $s$ selected variables, we get:
\[prop:Lasso\_Varsel\] Let decision class $\mathcal{F}_{s}$ be as given in Equation with $B=\{\norm{f}_{1} \leq W\}$. Then the following upper bound for the Rademacher complexity of the class $\mathcal{F}_{s}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{s}) &\leq
\frac{\sqrt{\ell}W}{\sqrt{n}} \paren[4]{ \sqrt{\frac{2}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}}} + \frac{16\sqrt{s}W}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}} \\
{\mathfrak{R}}_{n}(\mathcal{F}_{s}) &\leq
\frac{\sqrt{\ell}W}{\sqrt{n}} \paren[3]{\sqrt{2\ee{}{\norm{\bx}_{\infty}^{2}}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}}} + \frac{16\sqrt{s}W}{\sqrt{n}}\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\log\paren{\frac{de}{s}}}
\end{aligned}$$
Also, with probability at least $1-\delta/2$ we obtain that for all $f \in \mathcal{{F}}_{s}$ it holds: $$\begin{aligned}
L(f) &\leq L_{n}(f) + \frac{C_{1}\sqrt{\ell}W }{\sqrt{n}} \paren[4]{\sqrt{\frac{2}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} +4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}}} \\ &+\frac{8C_{1}\sqrt{s}W}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}} + C_{1}WK \sqrt{\frac{\log(\frac{2}{\delta})}{2n}} + C\sqrt{\frac{\log{\paren{\frac{2}{\delta}}}}{2n}},
\end{aligned}$$ where constants $C_{1}=4(R+WK)$, $C=(R+WK)^{2}$ are as before.
Discussion: generalization properties of variable selection procedure.
----------------------------------------------------------------------
Firstly, we observe that using the uniform boundedness assumption for the norm of random variables ${\bx_{i}}$ by $K$ we can obtain the data independent bounds for Rademacher complexity (and for the high-probability deviations of generalization error) of the underlying model classes, which are specified by corresponding norm. Futher control of the Rademacher complexity (and consequently for the generalization error ) is enabled through assumptions on the distributions of the input variable $\bx$.
Firstly, from the results of the Propositions \[eq:rad\_comp\_ell\_2\], \[prop:Lasso\_rad\_gen\], \[prop:Lasso\_Varsel\], we observe that the convergence rates for generalization error are optimal up to the $\log(n)$ factor. Also all bounds on the Rademacher complexities scale linearly with the norm constraint on the prediction vector $f$.
As it was already mentioned before, both Proposition \[eq:rad\_comp\_ell\_2\] and Proposition \[prop:Lasso\_rad\_gen\] imply high probability deviation bounds (both empirical and distribution dependent) for the statistical risk $L(f)$ and can be used for confidence intervals construction.
Analysing the case of variable selection procedure at root with the LASSO-penalized regressors at leaves, we notice that it reduces the impact of the dimensionality of the underlying linear model class (from $\log{d}$ to $\log{s}$ in the complexity term), adding however an additional factor of $\sqrt{{s\log{\paren{\frac{de}{s}}}}/{n}}$ (where the dependence on the dimension of the regression space is only in $\log(d)$ term) for the generality of the selection rule (which in our framework can be arbitrary).
It is worth to notice that our analysis can be naturally extended to the case in which at each node the covariates are projected on the lower-dimensional linear manifold (which may still depend on the internal nodes, in which the split is performed). In this case our analysis, as well as the discussions of Remarks \[rem:gen\_error\] and \[rem:gen\_error\_lasso\], hold with $d$ replaced by $m$, thus leading to sharper bounds for this particular subclass of PLRTs. We notice, that in this general case (where splits are data-dependent) in order to control the complexity of the functional class (i.e. apply the Rademacher-type analysis in the same vein as in Proposition \[eq:rad\_comp\_ell\_2\] or Proposition \[prop:Lasso\_rad\_gen\] ), we need to have an upper bound on the projected dimension $m$.
The feature selection procedure in the regression space $\mathcal{X}$ can be also performed both in the internal nodes for finding the penalized least squares solution and additionally on the leaves to build the final regressors. More precisely, at each internal node $1,\ldots, \ell-1$ we select $s$ features among $d$ from the avaliable dataset $\bX$, find the best split by building the (penalized) cumulative least squares loss, based on selected $s$ features. After the tree has been build, at each leaf ($1,\ldots,\ell$) select $s$ (possibly another) features from $d$ and compute the regression solutions. Thus, feature selection procedure is performed $2\ell-1$ times. Formally, class of decision allows the union representation as follows: $$\begin{aligned}
\label{eq:class_str_all}
\mathcal{F}_{\ell,sel} = \bigcup_{\substack{T \in \cT_l\\(i_{k},t_{k})_{k} \in ([D] \times \mbr)^{\cT^\circ} \\ \paren{\overline{s}}_{p} \in \overline{S}^{\otimes 2\ell-1}}}\mathcal{F}_{T,\paren{(i_{k},t_{k})_{k\in T^{\circ}}}, \paren{\overline{s}}_{p}}.
\end{aligned}$$ The aforementioned reasoning in Section \[subsec:variable\_select\] (with the only feature selection procedure at the root) can be straightforwardly extended to the case when we use feature selection at each internal node and we obtain the following analogues of Lemma \[lem:card02\] and Proposition \[prop:Lasso\_Varsel\].
\[lem:card03\] For the log-cardinality of the union of the sets in it holds: $$\begin{aligned}
\log(\abs{\mathcal{{F}}_{\ell,sel}} ) \leq \ell\paren{ \log({enD}) + 2s\log{\paren{\frac{de}{s}}}}.
\end{aligned}$$
\[prop:Lasso\_Varsel\_all\] Let function class $\mathcal{F}_{\ell,sel}$ be as given in the equation with $B=\{\norm{f}_{1} \leq W\}$. Then the following upper bound for the Rademacher compelxity of the class $\mathcal{F}_{\ell,sel}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{\ell,sel}) \leq
\frac{\sqrt{\ell}W}{\sqrt{n}} \paren[4]{2^{\frac{1}{2}} \sqrt{\frac{1}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}} + \frac{8\sqrt{s}}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}}} .
\end{aligned}$$ Also, with probability at least $1-\delta/2$ we obtain that for all $f \in \mathcal{{F}}_{\ell,sel}$ it holds: $$\begin{aligned}
L(f) &\leq L_{n}(f) + C_{1} \paren[3]{ \frac{\sqrt{\ell}W}{\sqrt{n}} \paren[2]{2^{\frac{1}{2}} \sqrt{\frac{1}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}}} }\\ &+\frac{8C_{1}\sqrt{s\ell}W}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}} + C_{1}WK \sqrt{\frac{\log(\frac{2}{\delta})}{2n}} + C\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}
\end{aligned}$$
Tractability Analysis {#sec:ctn}
=====================
As noted in the Introduction, the tractability and stability issues[^1] through using piecewise linear regressors for split optimization have been long been called into question (see for example [@dobra2002secret; @Potts2005; @Nata]). In the following we demonstrate that the full optimization can be efficiently solved.
Computational Complexity
-------------------------
The optimization problem to be solved for every candidate partitioning $\left(A,A^c\right)$ has the form $$\left\| \bX_A w - Y_A \right\|^2_P + \lambda \left\| w-w_0\right\|^2_Q,$$ which has an explicit solution $$w_A= (\bX_A^T P \bX_A + \lambda Q)^{-1} (\bX_A^T P Y_A + \lambda Q w_0).$$ As firstly noted by [@torgo2002comp], an efficient algorithm before scanning the thresholds $\psi^{k}_i$, will first sort them and then sequentially move one position in the sorted list at each iteration. Thus, once it has evaluated partition $A$ it will move on to evaluate the next partition, $A \cup i$ which implies that $$w_{A \cup i}
= \left( \bX^T_A P \bX_A + \bx_i ^T P \bx_i + \lambda Q \right)^{-1} \left( \bX_A^T P Y_A +\bx_i^T P y_i + \lambda Q w_0 \right).
\label{eq:k}$$ The first factor in the above product (eq. \[eq:k\]) can be expressed via the Sherman-Morrison formula. Therefore, the complexity of computing the above inverse is $O(d^2)$, given the inverse $\left( \bX^T_A P \bX_A + \lambda Q \right)^{-1}$. Thus, given the sorting of the $n$ samples along the $D$ dimensions, which itself has a complexity $O(Dn\log{n})$ , the complexity of finding the optimal split at each node is $O(nDd^2)$. It depends quadratically on underlying data dimension $d$ and only linearly on $D$ (dimension of the feature space). The complexity of the proposed method is $O(Dn\log{n} + Dnd^2)$ whereas the complexity of analogue model, which uses piecewise constant splits is $O(Dn\log{n} + Dn)$ .
Speeding up
-----------
Recall that $L^\lambda_{i,k}$ is defined as in . For each candidate split $i,k$ in order to evaluate $L^\lambda_{i,k}$ we need to compute the two losses $$l^\lambda_{i,k} = \min_{w_{i,k} \in \mbr^{d}} \paren{ \norm[1]{\bX_{A^{}_{i,k}}{w}^{}_{i,k}-Y_{A^{}_{i,k}}}_{P}^{2} + \lambda \| w_{i,k} - w_0\|^2_Q}$$ $$r^\lambda_{i,k}= \min_{w^c_{i,k} \in \mbr^{d}} \paren{\norm[1] {\bX_{A^{c}_{i,k}}{w}^{c}_{i,k}-Y_{A^{c}_{i,k}}}_{P}^{2} + \lambda \| {w}^{c}_{i,k} - {w}_0\|^2_Q}.$$ Assume that we have $N$ samples, which should be divided into two sets by a splitting procedure. We can significantly speedup the part of the Algorithm which performs tree construction by noticing that $\forall m>k, l^\lambda_{i,m} \geq l^\lambda_{i,k}$ and $\forall m<k, r^\lambda_{i,m} \geq r^\lambda_{i,k}$. For all $ i>1$ the algorithm possess some candidate pair $i_T,k_T$ for which $L^\lambda_{i_T,k_T}$ is the smallest encountered so far. Therefore the algorithm can avoid unneccesary computations by employing the following strategy: at each iteration for given $j \in [D]$ and $k \in [N]$ instead of computing $l^\lambda_{j,k}$ and $ r^\lambda_{j,k} $, the algorithm calculates $l^\lambda_{j,k}$ and $ r^\lambda_{j,N-k} $. If for some $k$, $l^\lambda_{j,k} + r^\lambda_{j,N-k} \geq L^\lambda_{i_T,k_T}$ then $\forall m, k \leq m \leq N-k$ it also holds that $l^\lambda_{j,m} + r^\lambda_{j,m} \geq L^\lambda_{i_T,k_T}$ and those computations can be avoided. We note that this speedup does not reduce the computational complexity of the algorithm, however it is an exact algorithm and as seen in Fig. \[fig:trac2\] can lead to significantly lower training times in practice.
We can significantly speedup the tree-constructing algorithm by noting that $\forall j>i, l^\lambda_{j,k} \geq l^\lambda_{i,k}$ and $\forall j<i, r^\lambda_{j,k} \geq r^\lambda_{i,k}$. Given that $\forall k>1$ the algorithm will have some candidate pair $i_T,k_T$ for which $L^\lambda_{i_T,k_T}$ is the smallest encountered so far the algorithm can avoid many computations by employing the following strategy: at each iteration instead of computing $l^\lambda_{j,k}$ and $ r^\lambda_{j,k} $, the algorithm calculates $l^\lambda_{j,k}$ and $ r^\lambda_{j,K-k} $. If for some $k$, $l^\lambda_{j,k} + r^\lambda_{j,K-k} \geq L^\lambda_{i_T,k_T}$ then $\forall m, k \leq m \leq K-k$ it also holds that $l^\lambda_{j,m} + r^\lambda_{j,m} \geq L^\lambda_{i_T,k_T}$ and those computations can be avoided. We note that this speedup does not reduce the computational complexity of the algorithm, however it is an exact algorithm and as seen in Fig. \[fig:trac2\] can lead to significantly lower training times in practice.
We also illustrate the results for two non-exact versions of the algorithm which use approximations to non-calculated losses to quickly identify non-promising splits, dimensions of the data and cut them from the computations. In particular we consider the next approaches:
1. $\forall m, k \leq m \leq N-k$ and for fixed $j \in [D]$ we approximate $$L^\lambda_{i,m} \approx l^\lambda_{j,k} + r^\lambda_{j,N-k} + \left(N-2k\right) \left(\min(l^\lambda_{j,k}-\lambda \| {w}_{j,k} - {w}_0\|^2_Q,r^\lambda_{j,N-k}-\lambda \| {w}_{j,N-k} - {w}_0\|^2_Q)\right).$$ If for a given $k$ , $L^\lambda_{i,m} \geq L^\lambda_{i_T,k_T}$,then as with the exact algorithm we forgo calculating $l^\lambda_{j,m},r^\lambda_{j,K-m}, \forall m, k \leq m \leq N-k$.
2. In a second approach where we consider that for all $m, k \leq m \leq N-k$ the following approximation: $$L^\lambda_{i,m} \approx l^\lambda_{j,k} + r^\lambda_{j,N-k} + \left(N-2k\right) \left(\max(l^\lambda_{j,k}-\lambda \| {w}_{j,k} - {w}_0\|^2_Q,r^\lambda_{j,N-k}-\lambda \| {w^c}_{j,N-k} - {w}_0\|^2_Q)\right).$$ As before, if for given $k$ , $L^\lambda_{i,m} \geq L^\lambda_{i_T,k_T}$, we proceed calculating $l^\lambda_{j,m},r^\lambda_{j,N-m}, \forall m, k \leq m \leq N-k$.
Training Times in Practice
--------------------------
As mentioned above, the complexity of calculating the optimal split, once the thresholds are sorted, is $O(nDd^2)$ which can involve a significant number of computations. In order to empirically evaluate the tractability of the proposed algorithm, we present in Fig. \[fig:trac2\] the time which required on a Tesla-K80 GPU to build a full regression tree of depth $10$ for various datasets. As can be seen, a careful algorithmic implementation, in conjunction with the use of GPU technology makes the models tractable despite their considerable complexity.Furthermore, by employing the proposed approximation, the speedup of the Algorithm can be significant and in some cases provide a gain up to an order of magnitude. In the appendix we present an empirical analysis of the effects of the proposed approximations on the accuracy of the trained models and demonstrate that the effects of the approximation can be negligible in many experiments.
Dataset KDD04 Forest KinH KinM MNISTOE MNISTBS CT Slice Air Energy
------------------------------ ------- -------- ------ ------ --------- --------- ---------- ----- --------
CART/M5 10 5 0.5 0.5 160 124 75 0.5 2.5
PLRT (no speedup) 6202 1454 550 532 60286 59583 14660 221 2247
PLRT (exact speedup) 1803 1276 481 458 29870 33171 5928 84 1878
PLRT (approx. speedup (min)) 1014 840 276 236 17379 16393 4590 62 1087
PLRT (approx. speedup (max)) 659 444 148 154 9240 8831 3222 55 421
Empirical Evaluation {#sec:emp_eval}
====================
In this section we present an empirical evaluation of the generalization capabilities of the proposed PLRT algorithm. In particular, we show results on a number of different regression tasks comparing different regression tree algorithms, among them
- PLRT trees that have been build using a $l_2$ regularized linear regression criterion to split the internal nodes and which similarly at the leaves employ $l_2$ regularization.
- PLRT trees that have been build using a $l_2$ regularized linear regression criterion to split the internal nodes but which employ $l_1$ regularization at the leaves.
- M5 trees that are built using a simple piecewise constant model but which employ $l_2$ regularized linear regression at the leaves. We slightly adapt the algorithm to allow it to use all variables to compute the linear regressors at its leaves.
- In the few cases where CART trees are competitive we also provide their results. As has been noted CART trees typically empirically underperform when compared to M5 [@wang1997inducing; @vogel07scalable].
Summary statistics on the datasets is provided in the Table \[tbl:datasets\], we refer reader to Appendix for a detailed presentation of the various tasks. We also notice that in order to distinguish between the regularization parameter for the $l_2$ and $l_1$ penalties we use $\gamma$ in case of $\ell_{2}$ norm and $\lambda$ for LASSO-type. For example in the plots for PLRT $\gamma=1.0$,$\lambda=0.1$ denotes a tree built by solving $\left\| \bX_A w - Y_A \right\|^2_P + \gamma \left\| w-w_0\right\|^2_Q$, while using LASSO in the leaves $ \left\| \bX_A w - Y_A \right\|^2 + \lambda \left\| w\right\|$. We note that for the experiments presented here we set $P,Q$ to be the identity matrices of appropriate dimension.
-- -- -- -- -- -- -- --
-- -- -- -- -- -- -- --
: Datasets description[]{data-label="tbl:datasets"}
- [**KDD-cup 2004**]{} which was part of the KDD-cup 2004 competition data. In particular we use the data from the physics competition which comprises data on 78 properties of 150.000 particles. As in [@vogel07scalable], we use the value of the 24th column as our target value.
- [**Forest**]{}, here the task is to predict the forest cover type from $54$ cartographic variables. As in [@Ronan], we create a regression task by assigning a label of $+1$ to samples corresponding to cover type $2$ and $-1$ to the rest. We use $50.000$ samples for training and another $50.000$ for testing.
- [**CT-Slice**]{} comprises $53500$ CT images, from $74$ different patients, represented by two histograms, one for bone structure and one for air inclusions. The dimensionality of the data is $383$. The target value of the task is the relative location on of the image on the axial axis. We assign images from two thirds of the patients to the training datasets, while the remaining third we keep as a test set.
- [**MNIST**]{} consists of $60.000$ training samples and $10.000$ test samples of dimensionality $784$, representing handwritten digits. In order to create a regression task, we assign a label of $+1$ to all odd digits and $-1$ to the even digits. Following [@Ronan] we create a second regression task on the MNIST dataset by assigning a label of $+1$ to the digits $0$ to $4$ and $-1$ to digits $5$ through $9$.
- [**Kinematic**]{} which comes from a realistic simulation of the dynamics of an 8 link robot arm. The task is to predict the distance of the arm’s end-effector from some target. We use two versions of this dataset, one with medium noise and another with high noise.
\
\
\
\
\
\
\
\
\
As can be seen in Fig. \[fig:RES\], the proposed algorithm PLRT consistently outperforms the M5 baseline. This is due to the fact that tree structure of PLRT has been optimized in a greedy fashion and provides at the end complex regression model in the leaves; M5 on the other hand optimizes the tree structure based on a different model than ultimately employed. This allows PLRT to obtain similar or better generalization at smaller tree depths. Though we restrict the results shown here to $\gamma \in \{10.0,1.0\}$, we provide a more detailed analysis of the effects of the $\gamma$ parameter in the Appendix.
The proposed method also outperforms the simpler CART method on all datasets; in fact in most cases CART significantly underperforms and its results are supressed in order to keep the plot compact. As highlighted by the results on the Kinematic and KDD2004 datasets, PLRT can be prone to overfitting for small sized datasets at larger depths. Nonetheless, in all $3$ tasks involving these datasets PLRT achieves better empirical mean-squared errors when compared to M5. We also note that using $\ell_{1}-$penalization at the leaves does not lead to any added benefit in the empirical performance. This is due to the discrepancy between the usage of the regularization at the nodes ($l_2$ distance from weight vector of parent node $\| w-w_0\|$) and at the leaves ($l_1$ regularization $| w |$). A more detailed empirical evaluation of the effects of the $\lambda$ parameter can be found in the Appendix. In two cases (datasets KDD2004 and Air), PLRT + LASSO performs significantly worse than PLRT and M5 and we do not show these results in the plots.
Conclusion {#sec:conc}
==========
In this paper we provided a broad analysis of the class of regression decision trees algorithms with regularized linear prediction models on the leaves. From a theoretic perspective, using the celebrated toolbox of Rademacher complexities, we obtained new high probability upper bounds for the generalization error of the regularized PLRT-algorithms with $\ell_{2}$, $\ell_{1}$ penalty constraints (also including the extension to variable selection setting in the root). The resulting algorithms are based on the minimization of squared error splitting criteria. We illustrated that the proposed PLRT algorithm ( together with its speedups) are numerically stable and tractable and can be efficiently implemented with the help of GPU technology. The empirical study on the diversified types of datasets reveals that the resulting algorithm outperforms the well-known piecewise constant models when using the mean-squared error metrics with $\ell_{2}$ regularized solutions. From an empirical perspective, we investigate various regularization setups and reach a key insight regarding the type of regularization needed to avoid overfitting in practice.
Acknowledgments {#acknowledgments .unnumbered}
===============
[OZ would like to acknowledge the full support of the Deutsche Forschungsgemeinschaft (DFG) SFB 1294. ]{}
Appendix. Proofs of the main results {#sec:appendix}
=====================================
Generalization error bounds of the class of Piecewise Linear Regression tree models
-----------------------------------------------------------------------------------
We remind the formal definition used for the class of piecewise-linear regression trees $\mathcal{F}$. $$\begin{aligned}
\label{eq:regr_trees}
\mathcal{F} := \{f : f = \paren{ T, (i_{s},t_{s})_{s \in T^{\circ}},(f_{k})_{k \in \partial T}}, T \in \mathcal{T}_{\ell}, \paren{i_{s},t_s}_{s\in T^{\circ}} \in [D]\times \mathbb{R}, f_{k} \in B \},\end{aligned}$$ where $\mathcal{T}_{\ell}$ is the set of all trees, $T^{\circ}$ is the set of all internal nodes and $\partial T$ is the set of leaves. Through $\abs{A}$ we denote the cardinality of (finite) set $A$. Also we recall the following representation $$\begin{aligned}
\label{eq:class_str01}
\mathcal{F} = \bigcup_{\substack{T \in \cT_l\\(i_{s},t_{s})_{s \in T^{\circ}} \in ([D] \times \mbr)^{ \abs{T^\circ}}}}\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}} ,\end{aligned}$$ of the model class as the union of the smaller sub-classes, where we formally write $\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^\circ}} = \{f: f = (f_{1},\ldots,f_{\ell}): \forall x \in \mathcal{X}, f(x) = \sum_{j} \inner{f_{j},x}\mathbbm{1}(x \in L_{j}), f_{j} \in \partial T \} $ for each tree $T$ and split family $\paren{i_{s},t_{s}}_{s \in T^{\circ}}$. We present the proofs of the main results of the paper next.
Observe, that for a fixed coordinate $i \in [D]$ there are $(n-1)$ different splits into two sets that can be obtained by the procedure, implemented in the algorithm. Indeed, if we consider $\psi^{i}$ sorted in decreasing order (w.r.t. samples), then any threshold $t \in [\max_{j} \{\psi_{j}^{i}: \psi_{j}^{i} < \psi_{k}^{i} \}, \psi_{k}^{i}]$ will give exact the same split as $t= \psi_{k}^{i}$. Thus, going iteratively over the (sorted) sample we obtain $n-1$ possibilities for choice of coordinate $k$ (excluding the split, when one subset is empty). The total number of possibilities at one node is therefore $(n-1)D$ and the total number of possible partitions in the fixed tree $\paren{(n-1)D}^{\ell-1}$ which means, that the infinite union $\mathcal{{F}}$ can be restricted to the finite, when for example choosing the threshold in each step $t_{s} = \psi^{i}_{k}$. Thus, we have representation that $\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} = \hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{\substack{T \in \cT_l\\(i_s,t_s)_{s} \in ([D] \times [n])^{\abs{T^\circ}}}} \mathcal{F}_{T,(i_{s},\psi_{k_s}^{i_s})_{s\in T^{\circ}}}}$. Also, the number of different trees with $\ell$ leaves is the $\ell-1$ Catalan’s number which is $\frac{1}{l} \binom{2(l-1)}{l-1}$. Futhermore, by using simple inequalities $(\frac{k}{e})^k<k!<e(\frac{k}{2})^k$ we obtain $\frac{1}{\ell} \binom{2(\ell-1)}{\ell-1} \leq \frac{e^{\ell}}{\ell}$ and thus the log-cardinality of the union is bounded by $\log{ \frac{e^{\ell}}{\ell}\paren{(n-1)D}^{\ell-1}} \leq \ell \log{enD}$. Applying Corollary \[cor:rad\_union\] to $\hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{\substack{T \in \cT_l\\(i_s,t_s)_{s \in T^{\circ}} \in ([D] \times [n])^{\abs{T^\circ}}}} \mathcal{F}_{T,(i_{s},\psi_{k_s}^{i_s})_{s\in T^{\circ}}}}$, and plugging the bound on the log-cardinality of the union in \[eq:class\_str01\] we obtain the claim of the Lemma.
Next we present the self-contained (alternative) proof of the Rademacher complexity upper bound for the case of $\ell_{2}$ regularization constraints for the class $\mathcal{{F}}$.
\[lem:Comp\_rad\_average\] Let functional class $\mathcal{F}$ as defined in and $B = \{\norm{f}_{2}\leq W\}$. Assume also that $\norm{X} \leq K$ $\mathbb{P}_{X}-$ almost surely and $\abs{Y} \leq R$ $\mathbb{P}_{Y}$- almost surely. Denote $\Sigma = \ee{}{\bx \bx^{\top}}$ the covariance matrix of random vector $\bx$ and $\hat{\Sigma}$ is the empirical analogue. Let also we have the sample $\mathbb{S}=\{\bx_{i},\psi_{i},y_{i}\}_{i=1}^{n}$ of size $n$. In this case the following bounds on empirical and true Rademacher complexities of class $\mathcal{F}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{S}(\mathcal{F}) &\leq \sqrt{2}W\sqrt{\frac{l\log(enD)}{n}\tr{\hat{\Sigma}}} \\
\mathfrak{R}_{n}\paren{\mathcal{F}} &\leq \sqrt{2}W \sqrt{\frac{l\log(enD)\tr{\Sigma}}{n}}
\end{aligned}$$
In this proof we use the notation $||\cdot||$ to denote the standard euclidean norm $\norm{\cdot}_{2}$. We enumerate leaves from $1$ to $l$ and denote through $L_{j}$ the leaf with the number $j$ and associate $(f_{1}, \ldots, f_{l})$ with the tuple of linear regression function attached to the leaves.
For the fixed number of leaves $l$, the empirical Rademacher complexity for $\mathcal{F}$ can be written as follows: $$\begin{aligned}
\hat{\mathfrak{R}}_{S}(\mathcal{F})&= E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sup\limits_{(f)_{k \in \partial T} }\frac{1}{n}\sum \limits_{i=1}^{n}\sigma_{i}\langle f,\bx_{i} \rangle \right] = E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sup\limits_{f_1,...,f_l}\frac{1}{n}\sum\limits_{j=1}^{l}\sum\limits_{i:\phi_{i} \in L_{j}}\sigma_{i}\langle f_{j},\bx_{i}\rangle\right] \\
& =\frac{1}{n} E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sum\limits_{j=1}^{l} \sup\limits_{f_{j} \in B}\sum\limits_{i: \bx_{i} \in L_{j}}\sigma_{i}\langle f_{j},\bx_{i}\rangle\right]\\
& =
\frac{1}{n} E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sum\limits_{j=1}^{l} \sup\limits_{f_{j} \in B}\langle \ f_{j} ,\sum\limits_{i: \bx_{i} \in L_{j}}\sigma_{i}\bx_{i}\rangle\right].
\end{aligned}$$ Using the Cauchy-Schwartz inequality in the last equality for the scalar product of $f_{j}$ with $\sum\limits_{i: \bx_{i} \in L_{j}}\sigma_{i}\bx_{i}$ and taking the supremum over $f_{j} \in B$ we obtain: $$\label{eq:through_S}
\frac{1}{n} E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sum\limits_{j=1}^{l} \sup\limits_{f_{j}\in B}\sum\limits_{i: \bx_{i} \in L_{j}}\sigma_{i}\langle f_{j},\bx_{i}\rangle\right] \leq \frac{W}{n} E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sum\limits_{j=1}^{l}||S_{j}||\right],$$ in which we denote $S_{j} :=\sum\limits_{i: \bx_{i} \in L_{j}}\sigma_{i}\bx_{i}$.
To upper bound we use the **softmax inequality** . This gives us: $$\label{eq:softmax}
E_{\sigma}\left[\sup\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} \sum\limits_{j=1}^{l}||S_{j}||\right] \leq E_{\sigma}\left[\frac{1}{\lambda}\log \sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}\exp(\lambda \sum\limits_{j=1}^{l}||S_{j}||)\right],$$ where the last sum is taken over all possible partitions $i_{s},t_{s}$ and tree structures $T$. Using Jensen’s inequality and the linearity of expectation for we obtain: $$\begin{aligned}
\label{eq:prepare}
E_{\sigma}\left[\frac{1}{\lambda}\log \sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}\exp(\lambda \sum\limits_{j=1}^{l}||S_{j}||)\right] &\leq \frac{1}{\lambda} \log \left[\sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}E_{\sigma}\left[\exp(\lambda \sum\limits_{j=1}^{l}||S_{j}||)\right]\right].
\end{aligned}$$ $S_{j}$ are mutually independent random variables (as variables in $\sigma$) as the leaves contain disjoint sets of data-points, which gives: $$\begin{aligned}
\label{eq:Prod_represent}
E_{\sigma} \left[\exp(\lambda \sum\limits_{j=1}^{l}||S_{j}||)\right] = \prod\limits
_{j=1}^{l}E_{\sigma}\left[\exp(\lambda||S_{j}||)\right].
\end{aligned}$$ Using for each $j \in \{1,...l\}$ the inequality $e^{tx} \leq 2\cosh(tx), t>0 $ and exponential inequalities from Theorem 3 in [@pinelis1986remark] we get: $$\begin{aligned}
\begin{aligned}
\label{eq:trans}
E_{\sigma}\left[\exp(\lambda||S_{j}||)\right] & \leq 2E_{\sigma}\left[\cosh(\lambda ||S_{j}||)\right] \leq 2 \prod\limits_{i: \bx_{i} \in L_{j}}E_{\sigma}\left[\exp(\lambda||\sigma_{i}\bx_{i}||)-\lambda||\sigma_{i}\bx_{i}||\right] \\
& = 2\prod\limits_{i: \bx_{i} \in L_{j}} \left[ \exp(\lambda||\phi_{i}||)-\lambda||\phi_{i}|| \right] \leq 2 \exp \bigg(\frac{\lambda^{2}}{2}\sum\limits_{i: \bx_{i} \in L_{j}}\norm{\bx_{i}}^{2}\bigg),
\end{aligned}
\end{aligned}$$ where in the last point we used the inequality $\exp(x) - x \leq \exp(\frac{x^{2}}{2}) $ for $x>0$. Inserting the results of $\eqref{eq:trans}$ and $\eqref{eq:Prod_represent}$ into $\eqref{eq:prepare}$ we get: $$\begin{aligned}
\begin{aligned}
\label{eq:pre_log_last}
\frac{1}{\lambda} \log &\left[\sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} E_{\sigma}\left[\exp(\lambda \sum\limits_{j=1}^{l}||S_{j}||)\right]\right] \\
& \leq \frac{1}{\lambda} \log \Bigg(\sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} 2^{l} \exp \bigg(\frac{\lambda^{2}}{2}\sum\limits_{i=1}^{n}\norm{\bx_{i}}^{2}\bigg)\Bigg).
\end{aligned}
\end{aligned}$$ We denote, $\hat{S} := \sum_{i=1}^{n}\norm{\bx_{i}}^{2}$. By a simple upper bounding of the number of the possible splits ($((n+1)D)^{l-1}$) in the tree and the number of different trees (which is the $l-1$ Catalan’s number) we obtain:
$$\begin{aligned}
\label{eq:catalan_numb}
\sum\limits_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}} 2^{l} \exp \bigg(\frac{\lambda^{2}}{2}\sum\limits_{i=1}^{n}\norm{\bx_{i}}^{2}\bigg) \leq \frac{1}{l} \binom{2(l-1)}{l-1} (nD)^{l-1} 2^{l} \exp\bigg(\frac{\lambda^{2}}{2}\hat{S}\bigg).
\end{aligned}$$
By using $(\frac{k}{e})^k<k!<e(\frac{k}{2})^k$ we obtain $\frac{1}{l} \binom{2(l-1)}{l-1} \leq \frac{e^{l}}{l}$. Plugging this into and then plugging into and into we obtain:
$$\begin{aligned}
\begin{aligned}
\hat{\mathfrak{R}_{S}}(\mathcal{F}) & \leq \frac{W}{n\lambda}\log \left[ \frac{(2e(n+1)D)^{l}}{l(n+1)D}\exp\bigg(\frac{\lambda^{2}}{2}\hat{S}\bigg) \right] = \frac{W}{n}\bigg(\frac{C}{\lambda}+ \frac{\lambda}{2}\hat{S}\bigg) := \frac{W}{n}\psi(\lambda).
\end{aligned}
\end{aligned}$$
where $C = \log \big(\frac{(2e(n+1)D)^{l}}{l(n+1)D} \big)$. After minimizing the function $\psi(\lambda)$ in $\lambda$ we obtain: $\lambda_{min} = \sqrt{\frac{2C}{\hat{S}}}$ which implies $ \hat{\mathfrak{R}_{S}}(\mathcal{F}) \leq W\psi(\lambda_{min}) = \frac{W}{n}\sqrt{2\hat{S}C} $. Upper bounding $C$ by $\log((2enD)^{l})$ we obtain: $$\begin{aligned}
\hat{\mathfrak{R}}_{S}(\mathcal{F}) &\leq \sqrt{2}W\sqrt{\frac{\ell\log(2enD)}{n^{2}}\sum\limits_{i=1}^{n}\norm{\bx_{i}}^{2}} = 2^{\frac{1}{2}}W\sqrt{\frac{\ell \log(2enD)}{n}\frac{1}{n}\sum_{i=1}^{n}\tr{\bx_{i} \bx_{i}^{\top}}} \\
& = 2^{\frac{1}{2}}W\sqrt{\frac{\ell \log(2enD)}{n}\tr{\hat{\Sigma}}},
\end{aligned}$$ which proves the first part of the lemma. For the second claim we just have by using Jensen’s inequality and the exchanging linearity between trace and expectation: $$\begin{aligned}
\mathfrak{R}_{n}\paren{\mathcal{F}} & \leq \sqrt{2}W\sqrt{\frac{l\log(2enD)}{n^{2}}\sum\limits_{i=1}^{n}\ee{}{\norm{\bx_{i}}^{2}}} = \sqrt{2}W\sqrt{\frac{l\log(2enD)}{n^{2}}n \ee{}{\tr{\bx \bx^{\top}}}} \\
& =\sqrt{2}W\sqrt{\frac{l\log(2enD)}{n}{\tr{\Sigma}}}.
\end{aligned}$$
This bound can be improved using the general result of [@maurer2014inequality] in the sense of scaling in front of the $\sqrt{\ell}\log{enD}$ term. We consider firstly the bound on one specific tree (with fixed structure $T$ and partition $(i_{s},t_{s})_{s \in T^{\circ}}$) given by Lemma \[eq:l2\_bound\].
By linearity of expectation and taking into account, that leaves $(L_{j})_{j=1}^{\ell}$ are disjoint sets we obtain: $$\begin{aligned}
\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\sigma_{i}f(\bx_{i})} &= \sum_{j=1}^{\ell}\ee{}{\sup_{f_{j} \in B} \inner{f_{j},\sum_{i: \bx_{i} \in L_{j}}^{}\sigma_{i}\bx_{i}}} \\
& \leq W \sum_{j=1}^{\ell} \ee{}{\norm[2]{\sum_{i: \bx_{i} \in L_{j}}^{}\sigma_{i}\bx_{i}}_{2}} \\
& \leq W\sum_{j=1}^{\ell}\sqrt{\ee{}{\sum_{i,k: \bx_{i},\bx_{k} \in L_{j}}^{}\sigma_{i}\sigma_{k}\inner{\bx_{i},\bx_{k}}}},
\end{aligned}$$ where we used Jensen’s inequality in the last line. Since $(\sigma_{i})_{i=1}^{n}$ are independent Rademacher variables, by taking expectation, we obtain: $$\begin{aligned}
\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\sigma_{i}f(\bx_{i})} \leq W \sum_{j=1}^{\ell}\sqrt{\sum_{i: \bx_{i} \in L_{j}}\norm{\bx_{i}}^{2}} \leq W \sqrt{\ell} \sqrt{\sum_{i=1}^{n}\norm{\bx_{i}}^{2}},
\end{aligned}$$ where in the last inequality we used Jensen’s inequality for the function $t \mapsto \sqrt{t}$. The claim of the Lemma now follows from the definition of Rademacher complexity of the fixed class $\mathcal{F}_{T, (i_{s},t_{s})_{s \in T^{\circ}}}$ and the fact that $\sum_{i=1}^{n}\norm{\bx_{i}}^{2} = n \sum_{i=1}^{n}\tr{\frac{1}{n}\bx_{i} \bx_{i}^{\top}} = n \tr{\hat{\Sigma}}$, where $\hat{\Sigma} := \frac{1}{n}\sum_{i=1}^{n}\bx_{i} \bx_{i}^{\top}$ is the empirical covariance matrix.
First of all, we notice that for linear prediction vector $f$ and the constant $\mathcal{M}$ from Lemma \[lem: help\_lem01\] we obtain: $$\begin{aligned}
\mathcal{M} = \sqrt{\sup\limits_{f \in \mathcal{F},f\in B}\frac{1}{n}\sum_{i=1}^{n}\inner{f,\bx_{i}}^{2}}&= W\sqrt {\sup_{f \in \mathcal{F},f\in B}\inner{\frac{f}{W},\frac{1}{n}\sum_{i=1}^{n}\inner[1]{\frac{f}{W},\bx_{i}}\bx_{i}}} \\
& = W \sqrt{\sup_{f \in \mathcal{F},\norm{f}_{2} \leq 1}\inner{f,\frac{1}{n}\sum_{i=1}^{n}\inner[1]{f,\bx_{i}}\bx_{i}} } \\
& = W\sqrt{\norm{\hat{\Sigma}}_{op}},
\end{aligned}$$ where as before we denote $\hat{\Sigma} = \frac{1}{n}\sum_{i=1}^{n}\bx_{i}\bx_{i}^{\top}$ to be the empirical covariance matrix estimated from sample $\mathbb{S}$. Considering the fact that the (empirical) Rademacher complexity of one fixed (sub) class $\mathcal{F}_{T,\paren{i_{s},t_{s}}_{s \in T^{\circ}}}$ does not depend on the structure and partition (splitting family) of the tree, by applying Lemma \[lem: help\_lem01\] together with the claim of Lemma \[eq:l2\_bound\] for Rademacher complexity of $\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_s,t_{s})_{s \in T^{\circ}}}}$ we deduce: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} \leq \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\tr{\hat{\Sigma}}}}{\sqrt{n}} + 4W\sqrt{\norm{\hat{\Sigma}}_{op}}\sqrt{\ell\frac{\log{enD}}{n}}.
\end{aligned}$$ For the true Rademacher complexity, we simply take the expectation of $\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}}$, use Jensen’s inequality and exchange trace and expectation (due to their linearity) in the first summand; thus we get: $$\begin{aligned}
\label{eq:rad_bound_true02}
\begin{aligned}
\mathfrak{R}_{n}\paren{\mathcal{F}} & \leq \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\ee{}{\tr{\hat{\Sigma}}}}}{\sqrt{n}} + 4W\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\ell\frac{\log{2enD}}{n}} \\
&= \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\tr{\Sigma}}}{\sqrt{n}} + 4W\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\ell\frac{\log{2enD}}{n}}.
\end{aligned}
\end{aligned}$$ For the expectation of the operator norm of empirical covariance matrix by using triangle inequality and upper bound for the expectation of the difference between in operator norm between empirical covariance matrix and its population analogue( [@Tropp:15]) we deduce: $$\begin{aligned}
\ee{}{\norm{\hat{\Sigma}}_{op}} & \leq \ee{}{\norm{\hat{\Sigma} - \Sigma }_{op}} + \norm{\Sigma}_{op} \\
& \leq \sqrt{\frac{2K\norm{\Sigma}_{op}\log(2d)}{2n}} + \frac{2K \log(2p)}{n} + \norm{\Sigma}.
\end{aligned}$$ Inserting the last bound into the inequality we obtain the result for Rademacher complexity. Furthermore, since $f(\bx) \leq \norm{f}_{2}\norm{\bx}_{2} \leq WK$, putting this together the upper bound for the empirical Rademacher complexity and the bounds for Lipschitz constant, constants $F$ and $L_{R,B,F}$ into the Theorem \[thm:main\_theorem01\] we obtain with probability at least $1-\delta/2$: $$\begin{aligned}
L(f) \leq L_{n}(f) & + 4\paren{R+WK} \paren{\frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\tr{\hat{\Sigma}}}}{\sqrt{n}} + 4W\sqrt{\norm{\hat{\Sigma}}_{op}}\sqrt{\ell\frac{\log{2enD}}{n}}}\\
& + 4\paren{R+WK}WK \sqrt{\frac{\log(\frac{2}{\delta})}{2n}} + \paren{R+WK}^{2}\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}.
\end{aligned}$$ Finally, to obtain the alternative upper bound for the generalization error we just use the distribution dependent result for Rademacher complexity and get: $$\begin{aligned}
L(f) \leq L_{n}(f) & + 4\paren{R+WK}\frac{W\sqrt{\ell}}{\sqrt{n}} \paren{{2^{\frac{1}{2}}\sqrt{\tr{\Sigma}}} + 4\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{{\log{2enD}}}}
+ \paren{R+WK}^{2}\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}.
\end{aligned}$$
Repeating the similar arguments as in the case with $\ell_{2}-$penalty we obtain: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} = \frac{2}{n}\ee{}{\sup_{f = (f_{1},\ldots,f_{\ell})} \sum_{i=1}^{n}\sigma_{i}f(\bx_{i})} &= \sum_{j=1}^{\ell}\frac{2}{n}\ee{}{\sup_{f_{j} \in B} \inner{f_{j},\sum_{i: \bx_{i} \in L_{j}}^{}\sigma_{i}\bx_{i}}} \\
& = \sum_{j=1}^{\ell}\frac{r_{j}}{n}\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\hat{\mathcal{F}}_{j}},
\end{aligned}$$ where $r_{j}=\abs{L_{j}}$ and $\hat{\mathcal{F}}_{j}$ is the empirical Rademacher complexity of class of functions, such that $\norm{f}_{1} \leq W$ computed on the $r_{j}$ samples. Applying Proposition 3 from [@maurer2012structured] with $\mathcal{P}$ being the set of orthogonal projectors on the coordinates and $\norm{f}_{\mathcal{P}} = \norm{f}_{1}$, after rescaling the function $f$ by $W$ we obtain: : $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\hat{\mathcal{F}}_{j}} \leq \frac{W\sqrt{2\sum_{i: \bx_{i} \in L_{j}} \norm{\bx_{i}}^{2}_{\infty}}}{r_{j}} \paren{1+4\sqrt{\log{d}}},
\end{aligned}$$ which implies that $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{T,(i_{s},t_{s})_{s\in T^{\circ}}}} \leq \frac{1}{n}W\paren{1+4\sqrt{\log{d}}}\sum_{j=1}^{\ell}\sqrt{2\sum_{i: \bx_{i} \in L_{j}} \norm{\bx_{i}}^{2}_{\infty}} \leq \frac{2^{\frac{1}{2}}\ell^{\frac{1}{2}}W\sqrt{\sum_{i}\norm{\bx_{i}}^{2}_{\infty}}}{{n}}\paren{1+4\sqrt{\log{d}}},
\end{aligned}$$ where in the last inequality we used Jensen’s inequality in the similar fashion as for $\ell_{2}-$norm constraints for the function $t \mapsto \sqrt{t}$. The statement is therefore proved.
The general scheme of the proof remains to be the same as in the proof of Proposition \[eq:rad\_comp\_ell\_2\]. Namely, since we have that for any $f \in \mathbb{R}^{d}$ $\norm{f}_{2} \leq \norm{f}_{1}$ then, for the constant $\mathcal{M}$ we obtain: $$\begin{aligned}
\mathcal{M} = W\sqrt {\sup_{f \in \mathcal{F},f\in B}\inner{\frac{f}{W},\frac{1}{n}\sum_{i=1}^{n}\inner[1]{\frac{f}{W},\bx_{i}}\bx_{i}}} = W \sqrt{\sup_{f \in \mathcal{F},\norm{f}_{1} \leq 1}\inner{f,\frac{1}{n}\sum_{i=1}^{n}\inner[1]{f,\bx_{i}}\bx_{i}} } \leq W\sqrt{\norm{\hat{\Sigma}}_{op}}.
\end{aligned}$$ Thus, using Lemma \[lem: help\_lem01\] with simply $\mathcal{M} \leq W \sqrt{\norm{{\hat{\Sigma}}}_{op}}$ and bound on the log-cardinality of the union, together with Lemma \[lem:rad\_lasso\_con\] results into: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}} \leq
\frac{2^{\frac{3}{2}}\ell^{\frac{1}{2}}W\sqrt{\sum_{i=1}^{n}\norm{\bx_{i}}^{2}_{\infty}}}{{n}}\paren{1+4\sqrt{\log{d}}} + 8W\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\ell\frac{\log{2enD}}{n}}.
\end{aligned}$$ For the true Rademacher complexity in the similar vein to the case of $\ell_{2}$ penalty we get: $$\begin{aligned}
\label{eq:rad_bound_true_lasso}
\begin{aligned}
\mathfrak{R}_{n}\paren{\mathcal{F}} & \leq \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\sum_{j}\ee{}{\norm{\bx_{j}}_{\infty}^{2}}}}{{n}} + 8W\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\ell\frac{\log{2enD}}{n}} \\
& = \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\ee{}{\max_{k}\paren[1]{\bx^{k}}^{2}}}}{\sqrt{n}}\paren{1+4\sqrt{\log d}} + 8W\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}\sqrt{\ell\frac{\log{2enD}}{n}},
\end{aligned}
\end{aligned}$$ which proves the bound for Rademacher complexity. In order to control $\sqrt{\ee{}{\norm[1]{\hat{\Sigma}}_{op}}}$ we can apply the same argument as in the proof of Proposition \[eq:rad\_comp\_ell\_2\] and get: $$\begin{aligned}
\ee{}{\norm{\hat{\Sigma}}_{op}} \leq \sqrt{\frac{2K\norm{\Sigma}_{op}\log(2d)}{2n}} + \frac{2K \log(2p)}{n} + \norm{\Sigma}.
\end{aligned}$$ In order to control $\ee{}{\max_{k}\paren[1]{\bx^{k}}^{2}}$ we use the assumption that $\bx$ is subgaussian in each direction. Define $z_{k} = {\paren[1]{\bx^{k}}^{2}}$ for $k \in \{1,\ldots,d\}$ and notice that $z_{k} \geq 0$. Thus we have for any $\delta > 0$: $$\begin{aligned}
\ee{}{\max_{k} z_{k}} &\leq \max_{k} \ee{}{z_{k}} + \delta + \int_{\delta + \max_{k} \ee{}{z_{k}}}^{\infty}\probb{}{\max_{k} {z_{k}} > s}ds \\
& = \max_{k} \ee{}{z_{k}} + \delta + \int_{\delta }^{\infty}\probb{}{\max_{k} {z_{k}} > + \max_{k} \ee{}{z_{k}} + s}ds \\
& \leq \max_{k} \ee{}{z_{k}} + \delta + \sum_{k=1}^{d}\int_{\delta}^{+\infty}\probb{}{z_{k} \geq \ee{}{z_{k}} +s }ds. \\
\end{aligned}$$ Now, since variable $z_{k}-\ee{}{z_{k}}$ is centered subexponential, there exist $c_{1}>2,c_{2} >0$ such that for all $s>0$ it holds that $\probb{}{z_{k}-\ee{}{z_{k}} > s} \leq c_{1}\exp(-c_{2}s)$. After integration this gives the following upper bound: $$\begin{aligned}
\ee{}{\max_{k} z_{k}} \leq \max_{k}\ee{}{z_{k}} + \delta + \frac{dc_{1}}{c_{2}}\exp\paren{-c_{2}\delta}.
\end{aligned}$$ Minimizing the right hand side in $\delta$ we get $\delta^{\star} := c_{2}^{-1}\log(dc_{1})$ which implies: $$\begin{aligned}
\ee{}{\max_{k} z_{k}} \leq \max_{k} \ee{}{z_{k}} + C\log(dc_{1}),
\end{aligned}$$ where $C = c_{2} + c_{2}^{-1}$. Now since $\ee{}{z_{k}} = \ee{}{\bx_{k}^{2}} \leq \sigma^{2}$, we finally have: $$\begin{aligned}
\ee{}{\max_{k} z_{k}} \leq \sigma^{2} + C\log(dc_{1}).
\end{aligned}$$ Putting this inequality into the upper bound for Rademacher complexity gives: $$\begin{aligned}
\mathfrak{R}_{n}\paren{\mathcal{F}} \leq \frac{2^{\frac{1}{2}}\sqrt{\ell}W\sqrt{\sigma^{2} + C\log(dc_{1})}}{\sqrt{n}}\paren{1+4\sqrt{\log d}} + 8W\sqrt{ \paren{\frac{2K\norm{\Sigma}_{op}\log(2d)}{2n}}^{1/2} + \frac{2K \log(2p)}{n} + \norm{\Sigma} }\sqrt{\ell\frac{\log{2enD}}{n}}.
\end{aligned}$$ The result for the generalization error follows immediately from Theorem \[thm:main\_theorem01\].
Proofs for variable selection results
-------------------------------------
In this part we provide the results and sketch the proofs for the control of Rademacher complexities and generalization error upper bounds for the extension of PLRT algorithm to the variable selection criteria. and the following upper bound is true:
\[lem:card02\] $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}} \paren{\mathcal{F}_{s}}
&= \hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{T,(i_{k},t_{k}),\overline{s}} \mathcal{F}_{T,(i_k,t_k)_{k \in T^{\circ}},\overline{s} \in \overline{S}}}
\\ &\leq \max_{T,i_{s},t_{s},\overline{s}}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,i_{s},t_{s},\overline{s}}) + \\& + 4\mathcal{M}\paren{\sqrt{\frac{\ell\log{enD} + s \log\paren{\frac{de}{s}}}{n}} }, \\
\end{aligned}$$ where $\mathcal{M} = \sqrt{\sup\limits_{f \in \cF}\frac{1}{n}\sum_{i=1}^{n}f^{2}(\bx_{i})}$.
Repeating the argument of the proof of Lemma \[lem: help\_lem01\], for finite $n$, we can restrict the model class $\mathcal{F}_{s}$ to a finite union, encounting only splits with $t_{s} = \psi^{i}_{k}$ for all $i \in [D]$, $k \in [n]$. Notice, the number of possible choices of $s$ coordinates from $d$ is $\binom{d}{s}$ we derive, that the log-cardinality of the (finite) union is upper bounded by $\ell \log{enD} + \log{\binom{d}{s}}$ Now using the inequality $s! \geq \paren{\frac{s}{e}}^{s}$, from which we can derive that: $$\begin{aligned}
\binom{d}{s} = \frac{\prod_{j=0}^{s-1}\paren{d-j}}{s!} \leq \frac{d^{s}}{s!} \leq \paren{\frac{de}{s}}^{s}.
\end{aligned}$$ Then, the result is follows, when taking the logarithm and summing up the terms and proceeding directly as in Lemma \[lem: help\_lem01\].
Multiple feature selection. {#S-mfs}
---------------------------
The feature selection procedure in the regression space $\mathcal{X}$ can be also performed both in the internal nodes for finding the penalized least squares solution *and* additionally on the leaves to build the final regressors. More precisely, at each internal node $1,\ldots, \ell-1$ we select $s$ features among $d$ from the available dataset $\bX$, find the best split by building the (penalized) cumulative least squares loss, based on selected $s$ features. After the tree has been build, at each leaf ($1,\ldots,\ell$) select $s$ (possibly different) features from $d$ and compute the regression solutions. Thus, feature selection procedure is performed $2\ell-1$ times. Formally, class of decision allows the union representation as follows: $$\begin{aligned}
\label{eq:class_str_all}
\mathcal{F}_{\ell,sel} = \bigcup_{\substack{T \in \cT_l\\(i_{k},t_{k})_{k} \in ([D] \times \mbr)^{\cT^\circ} \\ \paren{\overline{s}}_{p} \in \overline{S}^{\otimes 2\ell-1}}}\mathcal{F}_{T,\paren{(i_{k},t_{k})_{k\in T^{\circ}}}, \paren{\overline{s}}_{p}}.\end{aligned}$$ The reasoning in Section \[subsec:variable\_select\] (with the only feature selection procedure at the root) can be straightforwardly extended to the case when we use feature selection at each internal node and we obtain the following analogues of Lemma \[lem:card02\] and Proposition \[prop:Lasso\_Varsel\]. Firstly we formulate the quite analogues version of Lemma \[lem:card02\] with multiple splits.
\[lem:card03\] $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}} \paren{\mathcal{F}_{\ell,sel}}
= \hat{\mathfrak{R}}_{\mathbb{S}} \paren[4]{\bigcup_{\substack{T \in \cT_l\\(i_{k},t_{k})_{k} \in ([D] \times \mbr)^{\cT^\circ} \\ \paren{\overline{s}}_{p} \in \overline{S}^{\otimes 2\ell-1}}}\mathcal{F}_{T,\paren{(i_{k},t_{k})_{k\in T^{\circ}}}, \paren{\overline{s}}_{p}}}
&\leq \max_{T,i_{s},t_{s},\overline{s}}\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{T,i_{s},t_{s},\paren{\overline{s}}_{p}}) \\
+ 4\mathcal{M}\paren{\sqrt{\frac{\ell\log{enD} + 2s\ell \log\paren{\frac{de}{s}}}{n}} },
\end{aligned}$$ where $\mathcal{M} = \sqrt{\sup\limits_{f \in \cF}\frac{1}{n}\sum_{i=1}^{n}f^{2}(\bx_{i})}$.
The proof is done in the very same way as in the Lemma \[lem:card02\], noticing that now we choose $s$ variables out of $d$ *at each node* which increases the factor of the union’s cardinality to $\binom{d}{s}^{2\ell-1}$. Rest of the proof follows the same argument as before.
\[prop:Lasso\_Varsel\_all\] Let function class $\mathcal{F}_{\ell,sel}$ be as given in the equation with $B=\{\norm{f}_{1} \leq W\}$. Then the following upper bound for the Rademacher complexity of the class $\mathcal{F}_{\ell,sel}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{\ell,sel}) \leq
\frac{\sqrt{\ell}W}{\sqrt{n}} \paren[4]{2^{\frac{1}{2}} \sqrt{\frac{1}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}} + \frac{8\sqrt{s}}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}}} .
\end{aligned}$$ Also, with probability at least $1-\delta/2$ we obtain that for all $f \in \mathcal{{F}}_{\ell,sel}$ it holds: $$\begin{aligned}
L(f) - L_{n}(f)&\leq + C_{1} \paren[3]{ \frac{\sqrt{\ell}W}{\sqrt{n}} \paren[2]{2^{\frac{1}{2}} \sqrt{\frac{1}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}} + 4 \sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{neD}}} }\\ &+\frac{8C_{1}\sqrt{s\ell}W}{\sqrt{n}}\sqrt{\norm[1]{\hat{\Sigma}}_{op}}\sqrt{\log\paren{\frac{de}{s}}} + C_{1}WK \sqrt{\frac{\log(\frac{2}{\delta})}{2n}} + C\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}}
\end{aligned}$$
Proof of both Propositions follows the same scheme as the Proposition \[prop:Lasso\_rad\_gen\]. Notice, that after selecting $s$ features the dimensionality of regression vector reduces to $s$, and thus the impact of the LASSO regularization in the complexity term will change from $\log{d}$ to $\log{s}$. Furthermore, for the Proposition \[prop:Lasso\_Varsel\] we use bound from Lemma 1 combined with the bound on the log-cardinality from Lemma \[lem:card02\] and general bound for a fixed class from Lemma \[lem:rad\_lasso\_con\]. In the same vein, to obtain the bounds from Proposition \[prop:Lasso\_Varsel\] we use the result of Lemma 1 combined with the bound on the log-cardinality from Lemma \[lem:card03\] and general bound for a fixed class from Lemma \[lem:rad\_lasso\_con\].
Case of clipped loss function
-----------------------------
In this part we additionally investigate the influence of the scaling constants of the model class regularization constraints, under the assumption that the square loss function can be restricted to some bounded domain, when using some apriori knowledge independent of the constraints upper bounds on the target function. We give the definition (following [@steinwart2008support], chapter 2) of the clipped loss function below.
A loss function $\ell\paren{y,t}: (y,t) \mapsto \mathbb{R}$ can be clipped at the value $H>0$ if for all $\paren{y,t}$ we have: $$\begin{aligned}
\ell\paren{y,\widetilde{t}} \leq \ell\paren{y,t},
\end{aligned}$$ where $\widetilde{t}$ denotes the clipped value of $t$ at $H$, i.e. :
$$\begin{aligned}
\widetilde{t} = \begin{cases}
-H & t < -H \\
t & t \in [-H,H] \\
H & t > H,
\end{cases}
\end{aligned}$$
or equivalently $\widetilde{t} = \max\{-H,\min\{t,H\}\}$.
Now let us consider the squared loss, i.e. $\ell(y,t) = \paren{y-t}^{2}$ and for the apriori given $M>0$ define the loss function $\widetilde{\ell}(y,t) = \min\paren{M,\paren{y-t}^{2}}$. Assuming, that the output variable $y$ has bounded support on $[-R,R]$ one can readily check, that the loss function $\ell(y,t)$ can be clipped at value $R+\sqrt{M}$ and that the loss function $\widetilde{\ell}(y,t)$ is its clipped version. Moreover, doing straightforward calculations, one can obtain, that the Lipschitz constant of $\tilde{\ell}\paren{y,\cdot}$ is bounded by $4R+2\sqrt{M}$.
Let firstly $\mathcal{F}$ be some model class, such that $\mathcal{F} = \cup_{i=1}^{N} \mathcal{F}_{i}$ and $\mathcal{F}_{i}$ be the arbitrary functional classes of real-valued functions with domain in $\mathcal{X}$. Applying Theorem \[thm:sup\_bound\] with the sets $A_{j}= \{\paren{\widetilde{\ell} \circ f(\bx_{1}),\ldots,\widetilde{\ell} \circ f(\bx_{n})}: f \in \mathcal{F}_{j} \}$, $j \in \{1,\ldots, N\}$ we obtain the following result for the Rademacher complexity $\hat{\mathfrak{R}}_{\mathbb{S}}(\widetilde{\ell} \circ \mathcal{F})$ of the image of the class $\mathcal{F}$ under the $\ell(\cdot,\cdot)$ map:
\[lem:img\_comp\] $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\widetilde{\ell} \circ \mathcal{F}) \leq \max_{m=1}^{N} \hat{\mathfrak{R}}_{\mathbb{S}}\paren{\widetilde{\ell} \circ \mathcal{F}_{m}} + 4 \mathcal{M} \sqrt{\frac{\log N}{n}},
\end{aligned}$$ where we have $\mathcal{M} := \sqrt{\sup_{f \in \cup_{m=1}\mathcal{F}_{m}}\frac{1}{n}\sum_{i=1}^{n} \paren[2]{\widetilde{\ell} \circ f(\bx_{i})}^{2}}$
**Remark** Notice, that from the definition of the function $\widetilde{\ell}\paren{y,t}$ it follows that $\mathcal{M} \leq M$. Notice that in Lemma \[lem:img\_comp\] a constant $\mathcal{M}$ depends only on the apriori bound $M$ and on the contrary to that from the Propositions \[eq:rad\_comp\_ell\_2\] or \[prop:Lasso\_rad\_gen\] where it scales linearly with the norm constraint of the prediction function $f$.
We demonstrate this more precisely on the example of $\ell_{1}$-type regularization constraints below. Firstly, through contraction principle with Lipschitz constant $L:= 4R+2\sqrt{M}$ of function $\widetilde{\ell}(y,t)$ we deduce from Lemma \[lem:img\_comp\] that it holds: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\widetilde{\ell} \circ \mathcal{F}) \leq L\max_{m=1}^{N} \hat{\mathfrak{R}}_{\mathbb{S}}\paren{\mathcal{F}_{m}} + 4 M \sqrt{\frac{\log N}{n}}.\end{aligned}$$ Thus, considering class $\mathcal{{F}}$ from Equation with the norm constraints $B =\{f: \norm{f}_{1} \leq W \}$ and using the upper bound of the Lemma \[lem: help\_lem01\] and upper bound on the empirical Rademacher complexity of one tree with fixed structure and partition (Lemma \[lem:rad\_lasso\_con\] ) we obtain.
\[prop:Lasso\_rad\_\_clipped\] Let the function class $\mathcal{F}$ be as given in the equation with $B=\{\norm{f}_{1} \leq W\}$. Let also the underlying loss function be the clipped loss $\widetilde{\ell}$ of the squared loss, clipped at the point $R+\sqrt{M}$. Then the following (data-dependent) upper bounds for both the empirical and true Rademacher complexities of the class $\mathcal{F}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}\paren{\widetilde{\ell} \circ \mathcal{F}} &\leq \frac{\sqrt{\ell}}{\sqrt{n}} \paren{2^{\frac{1}{2}}WL\sqrt{\frac{1}{n}\sum_{j}\norm{\bx_{j}}_{\infty}^{2}}\paren{1 + 4 \sqrt{\log{d}}} +4M \sqrt{\log\paren{enD}}} \\
{\mathfrak{R}}_{n}\paren{\widetilde{\ell} \circ \mathcal{F}} &\leq \frac{\sqrt{\ell}}{\sqrt{n}} \paren{ 2^{\frac{1}{2}}WL\sqrt{\ee{}{\norm{\bx}_{\infty}^{2}}}\paren{1 + 4 \sqrt{\log{d}}} + 4M \sqrt{\log\paren{enD}}}
\end{aligned}$$ Furthermore, with probability at least $1-\delta/2$ w.r.t. sample $\mathbb{S}$ we have for all $f \in \mathcal{{F}}$: $$\begin{aligned}
L(f) - L_{n}(f)\leq & + \frac{2\sqrt{\ell}}{\sqrt{n}} \paren{ 2^{\frac{1}{2}}WL\sqrt{\ee{}{\norm{\bx}_{\infty}^{2}}}\paren{1 + 4 \sqrt{\log{d}}} + 4M \sqrt{\log\paren{enD}}} + C\sqrt{\frac{\log(\frac{2}{\delta})}{2n}}.
\end{aligned}$$ where $C = (R+WK)^{2}$.
**Remark.** We observe that in the previous bound for Rademacher complexity only the first term scales linearly with the norm constraint, but the second term scales only with the constant $M$ (which apriori does not depend on the norm constraint $W$).
The effect of the clippable loss can be extended to the setting of feature selection procedure, when it is performed at each internal node and in the leaves. Recall that in this case the underlying model class can be written as: $$\begin{aligned}
\label{eq:class_str_all_sup}
\mathcal{F}_{\ell,sel} = \bigcup_{\substack{T \in \cT_l\\(i_{k},t_{k})_{k} \in ([D] \times \mbr)^{\cT^\circ} \\ \paren{\overline{s}}_{p} \in \overline{S}^{\otimes 2\ell-1}}}\mathcal{F}_{T,\paren{(i_{k},t_{k})_{k\in T^{\circ}}}, \paren{\overline{s}}_{p}}.\end{aligned}$$ The following result for the clipped square-loss function holds true.
\[prop:Lasso\_Varsel\_all\_clipped\] Let function class $\mathcal{F}_{\ell,sel}$ be the class of decision trees with regularization constraints $B=\{\norm{f}_{1} \leq W\}$ described by Equation \[eq:class\_str\_all\_sup\]. Then the following upper bound for the Rademacher complexity of the class $\mathcal{F}_{\ell,sel}$ is true: $$\begin{aligned}
\hat{\mathfrak{R}}_{\mathbb{S}}(\mathcal{F}_{s}) \leq
\frac{\sqrt{\ell}WL}{\sqrt{n}} \paren[4]{2^{\frac{1}{2}} \sqrt{\frac{1}{n} \sum_{i=1}^{n}\norm{\bx_{j}}^{2}_{\infty}}\paren{1 + 4\sqrt{\log(s)}}} + \frac{4M\sqrt{\ell}}{\sqrt{n}} \sqrt{\log\paren{neD}} + \frac{8\sqrt{s}}{\sqrt{n}}M\sqrt{\log\paren{\frac{de}{s}}} .
\end{aligned}$$ Also, with probability at least $1-\delta$ we obtain that for all $f \in \mathcal{{F}}_{\ell,sel}$ it holds: $$\begin{aligned}
L(f) - L_{n}(f)&\leq + \frac{2\sqrt{\ell}WL}{\sqrt{n}} \paren[4]{2^{\frac{1}{2}} \sqrt{\ee{}{\norm{\bx}^{2}_{\infty}}}\paren{1 + 4\sqrt{\log(s)}}} + \frac{8M\sqrt{\ell}}{\sqrt{n}} \sqrt{\log\paren{neD}} + \frac{16\sqrt{s\ell}}{\sqrt{n}}M\sqrt{\log\paren{\frac{de}{s}}} \\ &+ M\sqrt{\frac{\log{\frac{2}{\delta}}}{2n}},
\end{aligned}$$ where $L= 4R + 2\sqrt{M}$ is the Lipschitz constant as before.
**Remark.** Notice that the three terms in the last bound on the generalization error in the right hand side have the same scaling in terms of $\sqrt{\ell}/\sqrt{n}$, but only the first term scales linearly with respect to norm constraints $W$. Also, for $n$ large and $d<<D$, second term dominates the bound, but we can balance (and thus make the bound sharper) between first and third terms by choosing the number of selected variables depending on the norm constraints. More precisely, let $A:= \frac{2\sqrt{\ell}WL}{\sqrt{n}} \paren[4]{2^{\frac{1}{2}} \sqrt{\ee{}{\norm{\bx}^{2}_{\infty}}}\paren{1 + 4\sqrt{\log(s)}}}$ and $B := \frac{16\sqrt{s\ell}}{\sqrt{n}}M\sqrt{\log\paren{\frac{de}{s}}}$. Solving equation $A=B$ in $s$ for fixed norm constraint $W$ we have $$\begin{aligned}
\label{eq:s_optimal_scale}
s^{\star} = \Psi^{-1}\paren{\frac{WL\sqrt{\ee{}{\norm{\bx}_{\infty}^{2}}}}{2^{\frac{5}{2}}}},\end{aligned}$$ where $\Psi\paren{s} = \frac{\sqrt{s\log{\frac{de}{s}}}}{1+4\sqrt{\log{s}}}$ and $\Psi^{-1}$ is its inverse. This provides us the choice of number $s$ of variables to select, depending on the input dimension $d$, constant $L$ and norm constraints $W$ that balances the statistical bound for generalization error from the Proposition \[prop:Lasso\_Varsel\_all\_clipped\]. Notice that we can also obtain purely data-dependent selection rule, when substituting $\ee{}{\norm{\bx}}$ with its empirical counterpart $\frac{1}{n}\sum_{i=1}^{n}\norm{\bx_{i}}^{2}_{\infty}$.
Empirical Evaluations
=====================
Datasets
--------
- [**KDD-cup 2004**]{} this data set which was part of the KDD-cup 2004 competition data. In particular we use the data from the physics competition which comprises data on 78 properties of 150.000 particles. As in [@vogel07scalable], we use the value of the 24th column as our target value.
- [**Forest**]{}, here the task is to predict the forest cover type from $54$ cartographic variables. As in [@Ronan], we create a regression task by assigning a label of $+1$ to samples corresponding to cover type $2$ and $-1$ to the rest. As in [@Ronan], we use $25.000$ samples for training and another $10.000$ for testing.
- [**CT-Slice**]{} comprises $53500$ CT images, from $74$ different patients, represented by two histograms, one for bone structure and one for air inclusions. The dimensionality of the data is $383$. The target value of the task is the relative location on of the image on the axial axis. We assign images from two thirds of the patients to the training datasets, while the remaining third we keep as a test set.
- [**MNIST**]{} consists of $60.000$ training samples and $10.000$ test samples of dimensionality $784$, representing handwritten digits. In order to create a regression task, we assign a label of $+1$ to all odd digits and $-1$ to the even digits. Following [@Ronan] we create a second regression task on the MNIST dataset by assigning a label of $+1$ to the digits $0$ to $4$ and $-1$ to digits $5$ through $9$.
- [**Kinematic**]{} which comes from a realistic simulation of the dynamics of an 8 link robot arm. The task is to predict the distance of the arm’s end-effector from some target. We use two versions of this dataset, one with medium noise and another with high noise. In both cases the dimensionality of the data is 32.
- [**Energy**]{} The regression task here is to predict the appliances energy consumption (in Wh) of a household given the 28 measurements of a wireless sensor network of the temperature and humidity of various rooms in the household. The dataset’s github repository provides a train/test split comprising 14,803 training samples and 4932 testing samples.
- [**Air**]{} The dataset contains 9358 instances of hourly averaged responses from an array of 5 metal oxide chemical sensors embedded in an Air Quality Chemical Multisensor Device. We randomly subsampled 6000 samples for training and retained the rest for testing.
Effects of Speedups
-------------------
We present here an empirical analysis of the effects on the prediction error of the two approximation speedups proposed in the paper. Recall that we considered two settings for the approximation.
1. $\forall m, k \leq m \leq N-k$ we approximate $$L^\lambda_{i,m} \approx l^\lambda_{j,k} + r^\lambda_{j,N-k} + \left(N-2*k\right) \left(\min(l^\lambda_{j,k}-\lambda \| {w}_{j,k} - {w}_0\|^2_Q,r^\lambda_{j,N-k}-\lambda \| {w^c}_{j,N-k} - {w^c}_0\|^2_Q)\right).$$ If for a given $k$ , $L^\lambda_{i,m} \geq L^\lambda_{i_T,k_T}$, we forgo calculating $l^\lambda_{j,m},r^\lambda_{j,N-m}, \forall m, k \leq m \leq N-k$.
2. $\forall m, k \leq m \leq N-k$ we approximate $$L^\lambda_{i,m} \approx l^\lambda_{j,k} + r^\lambda_{j,N-k} + \left(N-2*k\right) \left(\max(l^\lambda_{j,k}-\lambda \| {w}_{j,k} - {w}_0\|^2_Q,r^\lambda_{j,N-k}-\lambda \| {w^c}_{j,N-k} - {w^c}_0\|^2_Q)\right).$$ If for a given $i$ , $L^\lambda_{i,m} \geq L^\lambda_{i_T,k_T}$, we forego calculating $l^\lambda_{j,m},r^\lambda_{j,N-m}, \forall m, k \leq m \leq N-k$.
We set regularization parameter $\gamma=1$ throughout these experiments and plot the loss for various tree depths on the various datasets for the 3 type of algorithmic procedures (exact, approximate (1.), approximate (2.)).
\
\
\
\
\
\
\
\
\
As can be seen in most cases the difference in accuracy is negligible. Only in two of the Datasets (Energy, KDD2004) the approximation algorithms lead to poorer performance. Given the speedups they offer in computation time, we suggest that there is considerable merit in employing these algorithms despite the fact that they are not exact.
Empirical effects of the choice of parameter $\lambda$ ($\ell_{1}-$penalization)
--------------------------------------------------------------------------------
We present here an empirical analysis of the effect on generalization of using LASSO in the leaves of a PLRT tree built with $l_2$-regularization in the nodes. In particular we set $\gamma=1.0$ constantly throughout all these experiments meaning that all the trees were built by optimizing $\left\| \bX_A w - Y_A \right\|^2 + \left\| w-w_0\right\|^2$, and evaluate the performance of the trees for various values of the LASSO regularization parameter $\lambda$.
\
\
\
\
\
\
\
\
\
As can be seen, LASSO does not lead to improved performance on the datasets presented here. This may be related to the difference, when using LASSO, in regularization criteria in the nodes and the leaves. The tree itself is constructed by using an $l_2$-regularization forcing the optimization algorithms in the prospective leaves to find weight vector solution close (in an $l_2$ sense) to the optimal weight vector already calculated in the node. This form of regularization played a crucial role in the development of the presented PLRT algorithm. As can be seen in the next section, strong regularization, i.e. a large $\gamma$ value, is needed for the stability of the algorithm.
Effects of parameter $\gamma$ ($l_2$ regularization)
----------------------------------------------------
We present here an empirical evaluation of the effect of the $\gamma$ on generalization, for the various datasets. As can be seen the proposed algorithm is prone to overfitting for small values of $\gamma$. In fact for almost all datasets the algorithm overfits even for moderately deep trees. We surmise that strong regularization $\gamma\geq 1.0$ is crucial to the performance of the algorithm. By propagating the weight vectors of the higher nodes, through the $l_2$ regularization, the space of viable weight vector solutions is constrained and overfitting is avoided. As noted this was a key insight in developing the proposed solution.
\
\
\
\
\
\
\
\
\
Numerical Stability
===================
As noted, one concern that can be raised regarding the proposed method is its numerical stability. the proposed method involves calculating, for each threshold, the inverse of a matrix of the form $\bX^T P \bX + \lambda Q$, where $\bX$ is a $N \times d$ matrix, meaning $\bX^T P \bX + \lambda Q$ is the result of $N$ rank-one updates to $Q$.
We consider in the following that $P,Q$ are identity matrices and investigate the numerical stability of calculating $\Lambda = \left( \bX^T \bX + I \right)^{-1}$ as $N$ rank-updates using Sherman-Morrison $(\Lambda_{R1})$ and calculating the inverse from scratch each time via the Cholesky decomposition $(\Lambda_{CH})$. In particular we consider the relative error of the Frobenius norm $\frac{\|\Lambda_{CH} - \Lambda_{R1}\|_F}{\|\Lambda_{CH}\|_F}$, where $\| \Lambda \|_F = \sum\limits_{i,j} \Lambda^2_{ij}$. For randomly generated data $\bX$, we plot in Figure \[fig:stab\](A) the error for various values of $N$ and $d=2048$. As can be seen the relative error in the Frobenius norm remains very small (of the order of magnitude $10^{-4}$) even for large values of $N$.
We furthermore investigate the numerical stability of the calculated vector $w = \left( \bX^T \bX + I \right)^{-1}\left( \bX^T Y\right)$, as this is ultimately what is used to calculate the optimal split. In Figure \[fig:stab\](B) we plot the angle, in degrees, between the vectors $w_{CH}$ and $w_{R1}$ calculated using Cholesky decompositions and $N$ rank-one updates respectively, as before we plot the angle for various values of $N$ and for $d=2048$. As can be seen there is some inaccuracy in a certain bandwidth of values of $N$ of the order $O(d)$ but even in this case the angle between the two vectors is very small $\left(0.03^{\circ}\right)$. Plotting, in Fig. \[fig:stab\](C) the conditioning number of the matrix $\kappa\left( \Lambda_{CH}\right) = \| \Lambda_{CH}^{-1} \| \| \Lambda_{CH} \|$ against $N$ reveals the source of this, relatively small, numerical inaccuracy. The experiments were run using the cuBLAS library on the same type of GPU as used in the experiments (Tesla K80) .
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![[\[fig:stab\]]{} Numerical stability of calculating $\Lambda = \left( \bX^T \bX + I \right)^{-1}$ and $w = \left( \bX^T \bX + I \right)^{-1}\left( \bX^T Y\right)$ via $N$ rank one updates to $I$ compared to calculating the same quantities via the Cholesky Decomposition: A) The relative error in the Frobenius norm, B) the angle, in degrees, between the vectors $w_{CH}$ and $w_{R1}$, and C) the conditioning number of $\kappa\left( \Lambda_{CH}\right) = \| \Lambda_{CH}^{-1} \| \| \Lambda_{CH} \|$, for $d=2048$ and various values of $N$](pics/froberr.png "fig:"){width="33.00000%"} ![[\[fig:stab\]]{} Numerical stability of calculating $\Lambda = \left( \bX^T \bX + I \right)^{-1}$ and $w = \left( \bX^T \bX + I \right)^{-1}\left( \bX^T Y\right)$ via $N$ rank one updates to $I$ compared to calculating the same quantities via the Cholesky Decomposition: A) The relative error in the Frobenius norm, B) the angle, in degrees, between the vectors $w_{CH}$ and $w_{R1}$, and C) the conditioning number of $\kappa\left( \Lambda_{CH}\right) = \| \Lambda_{CH}^{-1} \| \| \Lambda_{CH} \|$, for $d=2048$ and various values of $N$](pics/Angles2.png "fig:"){width="33.00000%"} ![[\[fig:stab\]]{} Numerical stability of calculating $\Lambda = \left( \bX^T \bX + I \right)^{-1}$ and $w = \left( \bX^T \bX + I \right)^{-1}\left( \bX^T Y\right)$ via $N$ rank one updates to $I$ compared to calculating the same quantities via the Cholesky Decomposition: A) The relative error in the Frobenius norm, B) the angle, in degrees, between the vectors $w_{CH}$ and $w_{R1}$, and C) the conditioning number of $\kappa\left( \Lambda_{CH}\right) = \| \Lambda_{CH}^{-1} \| \| \Lambda_{CH} \|$, for $d=2048$ and various values of $N$](pics/Cond.png "fig:"){width="33.00000%"}
(A) (B) (C)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[^1]: We contend that the empirical evaluations presented in Section \[sec:emp\_eval\] highlight that stability is not an issue. We illustrate in Appendix the evaluation of the numerical stability of the underlying algebraic computations
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We give an asymptotic expression for the expected number of spanning trees in a random graph with a given degree sequence $\dvec=(d_1,\ldots, d_n)$, provided that the number of edges is at least $n + \nfrac{1}{2} \dmax^4$, where $\dmax$ is the maximum degree. A key part of our argument involves establishing a concentration result for a certain family of functions over random trees with given degrees, using Pr[" u]{}fer codes.'
author:
- |
Catherine Greenhill\
School of Mathematics and Statistics\
UNSW Australia\
Sydney NSW 2052, Australia\
[email protected]\
- |
Mikhail Isaev\
Research School of Computer Science\
Australian National University\
Canberra, ACT 2601, Australia\
Moscow Institute of Physics and Technology\
Dolgoprudny, 141700, Russia\
[email protected]
- |
Matthew Kwan\
Department of Mathematics\
ETH Z[" u]{}rich\
Z[" u]{}rich, 8092, Switzerland\
[email protected]\
- |
Brendan D. McKay\
Research School of Computer Science\
Australian National University\
Canberra, ACT 2601, Australia\
[email protected]
date: 20 February 2017
title: |
The average number of spanning trees in\
sparse graphs with given degrees[^1]
---
Introduction {#s:intro}
============
The number of spanning trees $\tau(G)$ in a graph $G$ (also called the *complexity* of $G$) is an important graph parameter that has connections to a wide range of topics, including the study of electrical networks, algebraic graph theory, statistical physics and number theory (see for example [@Alo90; @MW99; @Sha87; @Wu77]). These connections are largely related to the *matrix tree theorem*, which says that $\tau(G)$ is equal to any cofactor of the Laplacian matrix of $G$.
There is a large body of existing work concerning the approximate value of $\tau (G)$ for graphs with given degree sequences, and random graphs with given degree sequences, especially in the regular case. Let $\dvec=(d_1,\ldots,d_n)$ be a vector of positive integers with even sum, and let $\Gamma_{\dvec}$ denote the set of all graphs on the vertex set $\{1,2,\ldots, n\}$ with degree sequence $\dvec$. If every entry of $\dvec$ equals $d$ then we write $\Gamma_{n,d}$ for the set of all $d$-regular graphs on $\{ 1,2, \ldots, n\}$. Let $\calG_{\dvec}$ be the random graph with degree sequence $\dvec$, chosen uniformly at random from $\Gamma_{\dvec}$, and let $\calG_{n,d}$ be the random $d$-regular graph on vertex set $\{ 1,2,\ldots, n\}$, chosen uniformly at random from $\Gamma_{n,d}$. Unless otherwise stated, all asymptotics in this paper hold as $n\to\infty$, possibly along some infinite subsequence of $\mathbb{N}$.
The number of spanning trees in a graph is strongly controlled by its degree sequence. Let $$d = \frac 1n\sum_j d_j, \qquad
\hat d = \biggl(\,\prod_{j=1}^n d_j\biggr)^{\!1/n}$$ denote the arithmetic and geometric means of the degree sequence $\dvec$. The best uniform upper bound for regular graphs is due to McKay [@McK83], who proved that when $d\geq 3$, $$\tau(G)=O(1) \biggl( \frac{(d-1)^{d-1}}{(d^2-2d)^{d/2-1}}\biggr)^{\!n}
\frac{\log n}{nd\log d}$$ for all $G\in \Gamma_{n,d}$. This was proved sharp within a constant by Chung and Yau [@chung]. Kostochka [@Kos95] proved that $$\(\hat d(1-\varepsilon)\)^n \leq \tau(G)\leq \frac{\hat{d}^n}{n-1}$$ for any connected $G\in \Gamma_{\dvec}$, where $\varepsilon =\varepsilon(\delta)>0$ tends to zero as $\delta=\min_j d_j\to \infty$. This lower bound extended a result of Alon [@Alo90] on $\tau(G)$ in the case of $d$-regular graphs.
To discuss random graphs, define the random variables $$\tau_\dvec=\tau(\calG_{\dvec}) \,\, \text{ and } \,\,
\tau_{n,d} = \tau(\calG_{n,d}).$$ That is, $\tau_{\dvec}$ is the number of spanning trees in $\calG_{\dvec}$, and $\tau_{n,d}$ is the number of spanning trees in $\calG_{n,d}$. McKay [@McK81] proved that for fixed $d$, $$\tau_{n,d}^{1/n}\to \frac{(d-1)^{d-1}}{(d^2-2d)^{d/2-1}}$$ with probability 1. An alternative proof in a much more general framework was given by Lyons in [@Lyo05 Example 3.16].
McKay [@subgraphs] gave the expected value $\E \tau_\dvec$ to within a constant factor, in the case that $d_j = O(1)$ for all $j$ and the average degree is at least $2+\eps$, for some $\eps > 0$. Specifically, McKay proved that under these conditions, the expected number of spanning trees is $$\label{McKay}
\E \tau_\dvec=\Theta(1) \;\frac 1n \biggl( \frac{\hat d\,(d-1)^{d-1}}
{d^{d/2} (d-2)^{d/2-1}}
\biggr)^{\!n}.$$ Greenhill, Kwan and Wind [@GKW] recently found the asymptotic value of this $\Theta(1)$ factor, for random $d$-regular graphs with $3\le d = O(1)$. Specifically, they proved that the $\Theta(1)$ in (\[McKay\]) is asymptotic to the constant $$\label{GKW-old}
\frac{(d-1)^{1/2}}{(d-2)^{3/2}}\,
\exp\biggl( \frac{6d^2-14d+7}{4(d-1)^2} \biggr).$$ (This is about $e^{3/2}/d$ for large $d$.) They also gave the asymptotic distribution of the number of spanning trees in a random cubic graph.
In this paper, we obtain an asymptotic expression for $\E \tau_\dvec$ for a wider range of sparse degree sequences $\dvec$ than in any of the above random graph results.
Our main result is the following.
\[main-sparse\] Let $\dvec = \dvec(n)=(d_1,\ldots, d_n)$ be a vector of positive integers with even sum, for every $n$ in some infinite subsequence of $\mathbb{N}$. Define $$\begin{aligned}
d_\mx &= \max_j d_j,\qquad d = \dfrac{1}{n}\, \sum_{j=1}^n d_j, \qquad
\hat d = \biggl(\,\prod_{j=1}^n d_j\biggr)^{\!1/n}, \qquad
R = \dfrac{1}{n}\, \sum_{j=1}^n (d_j - d)^2\end{aligned}$$ and let $$\Hd =
\frac{(d-1)^{1/2}}{(d-2)^{3/2}\, n}\,
\biggl( \frac{\hat{d}\, (d-1)^{d-1}}{d^{d/2}\, (d-2)^{d/2-1}}\biggr)^{\!n} .$$ Suppose that $d_\mx^4 \leq (d-2) n$. Then the sequence $\dvec$ is graphical for sufficiently large $n$, and the expected number of spanning trees in $\mathcal{G}_{\dvec}$ is given by $$\E\tau_{\dvec}
= \Hd\,
\exp\biggl( \frac{6d^2-14d+7}{4(d-1)^2}
+ \frac{R}{2(d-1)^3}
+ \frac{(2d^2-4d+1)R^2}{4(d-1)^4\, d^2}
+ O\biggl( \frac{d_\mx^4}{(d-2)n} + \eta\biggr)\biggr),$$ where $$\begin{aligned}
\eta &= \min\biggl\{ \frac{d_\mx^4}{(d-2)^2n},\,
\frac{d_\mx^3\log n}{(d-2)n},\,
d_\mx(d-2)\biggr\} \\
&= O\biggl( \frac{d_\mx^4}{(d-2)n} + \frac{(\log n)^{5/2}}{n^{1/2}}\biggr).\end{aligned}$$
Some remarks about this result are given below.
- Due to the Erd[ő]{}s-Gallai Theorem, under the conditions of Theorem \[main-sparse\] the sequence $\dvec$ is always graphical (without any requirement for $n$ to be large). Since this fact is not required for our asymptotic formula, we omit the proof.
- Since $\dmax \geq 1$, the condition $\dmax^4\leq (d-2) n$ implies that $d > 2$.
- Other than the relative error term, the expression given by Theorem \[main-sparse\] matches in the regular case, showing that the formula obtained in for regular graphs with constant $d\geq 3$ also holds for $d$-regular graphs with slowly growing $d$ (in particular, it holds when $d=o(n^{1/3})$).
- Under our assumptions, the relative error term may not be vanishing, though it is always bounded. Let $m = \nfrac{1}{2}\sum_{j=1}^n d_j$ be the number of edges in any graph in $\Gamma_{\dvec}$. The condition $\dmax^4 \leq (d-2)n$ is equivalent to the condition that $m\geq n + \nfrac{1}{2} \dmax^4$, or in other words, that there are at least $\nfrac{1}{2}\dmax^4 + 1$ more edges in any graph in $\Gamma_{\dvec}$ than in a tree on $n$ vertices. For example, when $\dmax = 3$, our result holds with a bounded error if the number of edges exceeds $n-1$ by at least 42.
In particular, we have the following corollary when $d$ is close to 2.
Suppose that $d=2+2x/n$ where $\nfrac{1}{2}\dmax^4 \leq x\leq n^{1/2}$. (This corresponds to graphs with $n+x$ edges.) Then $$\begin{aligned}
\E \tau_{\dvec} &= \frac1n\left(\frac{e}{2}\right)^x\, \left(\frac{n}{2x}\right)^{3/2 + x}
\, \biggl(\frac{\hat{d}}{2}\,\biggr)^n\,
\exp\left(\frac{(6+R)(2+R)}{16}
+ \frac{3x^2}{2n} +
O\left(\frac{\dmax^4}{x} + \frac{x^3}{n^2}\right)\right).\end{aligned}$$ \[close-to-2\]
We estimate the various terms in Theorem \[main-sparse\]. First note that $$\left(d-1\right)^{1/2}=\left(1+2x/n\right)^{1/2}=e^{O\left(x/n\right)}$$ and $$\frac{1}{(d-2)^{3/2}}\biggl(\frac{\hat{d}}{(d-2)^{d/2-1}}\biggr)^{n}=\hat{d}^{n}\left(\frac{n}{2x}\right)^{3/2+x}.$$ Next, a series expansion yields $$\begin{aligned}
\log\biggl(\frac{(d-1)^{d-1}}{d^{d/2}}\biggr) & =\left(1+2x/n\right)\log\left(1+\frac{2x}{n}\right)-\left(1+x/n\right)\log\left(2+\frac{2x}{n}\right)\\
& =-\log2+\left(1-\log2\right)\frac{x}{n}+\frac{3}{2}\left(\frac{x}{n}\right)^{2}+O\left(\left(\frac{x}{n}\right)^{3}\right)\end{aligned}$$ so we have $$\biggl(\frac{(d-1)^{d-1}}{d^{d/2}}\biggr)^{n}=2^{-n}\left(\frac{e}{2}\right)^{x}\exp\left(\frac{3x^{2}}{2n}+O\left(\frac{x^{3}}{n^{2}}\right)\right).$$ Then, we can compute $$\begin{aligned}
\frac{6d^{2}-14d+7}{4\left(d-1\right)^{2}} & =\frac{3}{4}+O\left(\frac{x}{n}\right),\\
\frac{1}{2\left(d-1\right)^{3}} & =\frac{1}{2}+O\left(\frac{x}{n}\right),\\
\frac{\left(2d^{2}-4d+1\right)}{4\left(d-1\right)^{4}d^{2}} &= \frac{1}{16}+O\left(\frac{x}{n}\right).\end{aligned}$$ So, noting that $R\le\dmax^{2}$, we have $$\begin{aligned}
\frac{6d^{2}-14d+7}{4\left(d-1\right)^{2}}+\frac{R}{2\left(d-1\right)^{3}}+\frac{\left(2d^{2}-4d+1\right)R^{2}}{4\left(d-1\right)^{4}d^{2}}
& =\frac{\left(6+R\right)\left(2+R\right)}{16}+O\left(\frac{\dmax^4\, x}{n}\right).
\label{medium}\end{aligned}$$ Finally, the error term from Theorem \[main-sparse\] is at most $$O\left(\frac{\dmax^4}{(d-2)n} + (d-2)\dmax\right) = O\left(\frac{\dmax^4}{x}
\right).$$ Since this error term dominates the error from (\[medium\]) under our assumptions, the result follows.
From Corollary \[close-to-2\] we see that when the average degree is close to but above 2, and the geometric mean $\hat{d}$ is strictly greater than 2, then $\E \tau_{\dvec}$ tends to infinity. This can be true even for degree sequences where the probability of connectivity tends to zero, even when Corollary \[close-to-2\] does not apply. For example, consider the degree sequence $\dvec$ with $n/2$ vertices of degree 5 and $n/2$ vertices of degree 1 (restricted to even $n$). Here $d = 3$ and $\hat{d} = \sqrt{5} > 2$. From Theorem \[main-sparse\] it follows that the expected number of spanning trees in $\calG_{\dvec}$ is $$\Theta(1/n)\, \left(\frac{80}{27}\right)^{n/2}$$ which tends to infinity as $n\to \infty$. However, the probability that $\calG_{\dvec}$ is connected tends to zero. To see this, we work in the configuration model [@bollobas]. For ease of notation, write $n=2t$ and let $S$ be the set of configurations with $t$ cells containing 5 points and $t$ cells containing 1 point. If a configuration in $S$ gives rise to a connected graph then every point in a cell of size 1 is paired with a point from a cell of size 5. There are at most $$\label{example-numerator}
\frac{(5t)_t \, (4t)!}{2^{2t}\, (2t)!}$$ such configurations, and the probability that a random configuration in $S$ is simple, conditioned on connectedness, is at most 1. The total number of simple configurations in $S$ is $$\label{example-denominator}
\Theta(1) \, \frac{(6t)!}{2^{3t}\, (3t)!}$$ where the $\Theta(1)$ factor is the probability that a random configuration in $S$ is simple: this tends to a constant bounded away from zero, by [@symm Theorem 4.6]. Dividing by gives the upper bound $$\Theta(1)\, \left( \frac{5^5}{2^7\, 3^3}\right)^{n/2} = o(1)$$ on the probability that a random element of $\calG_{\dvec}$ is connected.
The case of dense irregular degree sequences will be treated in a separate paper.
Outline of our approach {#s:outline}
------------------------
Let $(a)_k$ denote the falling factorial $a(a-1)\cdots (a-k+1)$. We say that a sequence $\xvec=(x_1,\ldots, x_n)$ of positive integers is a *tree degree sequence* if the entries of $\xvec$ sum to $2n-2$. We say that a tree degree sequence $\xvec$ is a *suitable* degree sequence if $1\leq x_j\leq d_j$ for all $j\in \{1,2,\ldots, n\}$. (The intended meaning is that $\xvec$ is suitable as a degree sequence for a spanning tree of a graph with degree sequence $\dvec$.)
For a suitable degree sequence $\xvec$, let $\calT_{\xvec}$ be the set of all trees with degree sequence $\xvec = (x_1,\ldots, x_n)$ and $\calT$ be the set of all trees with vertex set $\{ 1,2,\ldots, n\}$. It is well-known that $$\label{alltrees}
\card{\calT_\xvec}= \binom{n-2}{x_1-1,\ldots,x_n-1}.$$ (See for example [@moon Theorem 3.1].) Let $\tau_{\dvec}(\xvec)$ denote the number of spanning trees of $\calG_{\dvec}$ with degree sequence $\xvec$, and denote by $P(\dvec,T)$ the probability that the random graph $\mathcal{G}_{\dvec}$ has $T$ as a subgraph, for all $T\in \calT$. Then the expected number of spanning trees with degree sequence $\xvec$ in $\mathcal{G}_{\dvec}$ can be written as $$\label{nd}
\E \tau_\dvec(\xvec) = \sum_{T\in \calT_\xvec} P(\dvec,T)$$ and furthermore, the expected number of spanning trees (of any degree sequence) in $\calG_{\dvec}$ is $$\E \tau_\dvec = \sum_{\xvec} \E \tau_\dvec(\xvec)$$ where the sum is over all suitable degree sequences $\xvec$.
We will estimate the summand in (\[nd\]) using a theorem by McKay [@symm Theorem 4.6], which we will restate below, including some necessary terminology, and with some minor rewording for consistency.
*[@symm Theorem 4.6]* \[mckay\] Let $\gvec = (g_1,\ldots, g_n)$ be a sequence of non-negative integers with even sum $2m$, and let $\gmax = \max\{ g_1,\ldots, g_n\}$. Let $X$ be a simple graph on the vertex set $\{1,2,\ldots, n\}$ with degree sequence $\xvec = (x_1,\ldots, x_n)$, where $\xmax = \max\{ x_1,\ldots, x_n\}$. Suppose that $\gmax\geq 1$ and $\hat{\Delta} \leq \epsilon_1 m$, where $\epsilon_1 < 2/3$ and $\hat{\Delta} = 2 + \gmax(\nfrac{3}{2}\gmax + \xmax + 1)$. Define $$\lambda = \frac{1}{4m}\, \sum_{j=1}^n (g_i)_2 \quad \text{ and } \quad
\mu = \frac{1}{2m}\, \sum_{ij\in X} g_ig_j.$$ Let $N(\gvec,X)$ denote the number of simple graphs with degree sequence $\gvec$ and no edge in common with $X$. Then $$N(\gvec,X) = \frac{(2m)!}{m!\, 2^{m}\, \prod_{j=1}^n g_i!}\, \exp(-\lambda - \lambda^2 - \mu
+ O(\hat{\Delta}^2/m))$$ uniformly as $n\to\infty$.
Given a suitable degree sequence $\xvec$ and a tree $T\in\mathcal{T}_{\xvec}$, define the parameters $$\begin{aligned}
\lambda_0 &= \frac{1}{2dn}\, \sum_{j=1}^n \, (d_j)_2, \\
\lambda(\xvec) &= \frac{1}{2(d-2)n+4} \, \sum_{j=1}^n \, (d_j-x_j)_2, \\
\mu(T) &= \frac{1}{(d-2)n+2}\, \sum_{\{i,j\}\in E(T)} (d_i-x_i)(d_j-x_j)\end{aligned}$$ Using Theorem \[mckay\], we may prove the following.
\[prob\] Suppose that $\xvec$ is a suitable degree sequence and let $T\in\mathcal{T}_{\xvec}$. With notation as above, provided $\dmax^4\leq (d-2)n$, $$\begin{aligned}
P(\dvec,T)
&= \frac{ (dn/2)_{n-1}\, 2^{n-1}}{ (dn)_{2n-2} }
\prod_{j=1}^n (d_j)_{x_j} \exp\left( \lambda_0 + \lambda_0^2 - \lambda(\xvec) - \lambda(\xvec)^2 - \mu(T)
+ O\!\left(\frac{\dmax^4}{(d-2)n}\right) \right). \end{aligned}$$
There is a bijection between the set of graphs with degree sequence $\dvec$ which contain $T$, and those with degree sequence $\dvec - \xvec$ which contain no edges of $T$. Therefore, we can write $$P(\dvec,T) = \frac{N(\dvec-\xvec,T)}{N(\dvec,\emptyset)}
\label{PdT}$$ and Theorem \[mckay\] to estimate the numerator and denominator.
First, consider $N(\dvec-\xvec, T)$. Let $$\hat{\Delta} = 2 + g_{\mx}\left(\nfrac{3}{2} g_{\mx} + x_{\mx} + 1\right)$$ where $g_{\mx} = \max_{j=1,\ldots, n} (d_j-x_j)$. We require that $\hat{\Delta}^2 \leq \eps_1 m$, where $\eps_1 < \nfrac{2}{3}$ is a constant and $m = \nfrac{1}{2} ((d-2)n+2)$ is the number of edges in a graph with degree sequence $\dvec-\xvec$. By assumption, $$m = \nfrac{1}{2}\, ((d-2)n+2) \geq 1 + \nfrac{1}{2}\dmax^4.$$ Since $g_{\mx}, x_{\mx}\leq d_{\mx}$ and $\dmax \geq 3$ (which follows since $d > 2$), we have $$\frac{\hat{\Delta}}{m} \leq
\frac{2 + \dmax(\nfrac{5}{2}\dmax + 1)}{1 + \nfrac{1}{2} \, \dmax^4} \leq \frac{55}{83}$$ which is strictly less than $\nfrac{2}{3}$. Observe also that $\hat{\Delta} = O(\dmax^2)$. Hence Theorem \[mckay\] applies and says that $$\begin{aligned}
N(\dvec-\xvec,T) &= \frac{((d-2)n+2)!}{((d-2)n/2 + 1)! 2^{(d-2)n/2+1}\, \prod_{j=1}^n (d_j-x_j)!}\\
& \hspace*{4cm}
\times \exp\left( -\lambda(\xvec) - \lambda(\xvec)^2 - \mu(T) + O\Big(\dmax^4/((d-2)n)\Big)\right).\end{aligned}$$ Similarly, we obtain $$\begin{aligned}
N(\dvec,\emptyset) &= \frac{(dn)!}{(dn/2)! 2^{dn/2}\, \prod_{j=1}^n d_j!}
\, \exp\left( -\lambda_0 - \lambda_0^2 + O\Big(\dmax^4/((d-2)n)\Big)\right),
\label{Nd}\end{aligned}$$ noting that the value of the $\hat{\Delta}$ is smaller than in the previous application of Theorem \[mckay\], while the parameter $m$ is larger. Substituting these expressions into (\[PdT\]) completes the proof.
Observe that the only term in the argument of the exponential in Lemma \[prob\] which depends on the structure of $T$ (rather than just the degree sequence of $T$) is $\mu(T)$. For any suitable degree sequence $\xvec$ and any tree $T\in\calT_{\xvec}$, define $$f(\xvec)=\lambda_0+\lambda_0^2 - \lambda(\xvec)-\lambda(\xvec)^2
\label{fg-def}$$ and let $$\label{beta-def}
\eg(\xvec) = \frac{1}{|\calT_\xvec|}\, \sum_{T\in\calT_{\xvec}}\, e^{- \mu(T)}$$ be the average value of $e^{-\mu(T)}$ over all $T\in\calT_\xvec$.
Combining (\[alltrees\]), (\[nd\]) and Lemma \[prob\], for any suitable degree sequence $\xvec$ we have $$\begin{aligned}
\E \tau_\dvec(\xvec) &= e^{O(\dmax^4/((d-2)n))}
\frac{ (dn/2)_{n-1} \,2^{n-1}\, \prod_{j=1}^n d_j}
{ (dn)_{2n-2} }\,
\sum_{T\in\calT_\xvec}\, \biggl(\, \prod_{j=1}^n (d_j-1)_{x_j-1}
e^{f(\xvec) - \mu(T)} \biggr) \nonumber \\
&= e^{O(\dmax^4/((d-2)n))}
\frac{ (dn/2)_{n-1} \,2^{n-1}\, \hat{d}^n}
{ (dn)_{2n-2} } \, (n-2)!\,
\biggl(\, \prod_{j=1}^n \binom{d_j-1}{x_j-1} \biggr)
e^{f(\xvec)}\, \eg(\xvec). \label{one-x}\end{aligned}$$ Now define $$\label{mubar-def}
\barg(\xvec) = \frac{1}{|\calT_{\xvec}|}\, \sum_{T\in\calT_{\xvec}} \mu(T),$$ the average value of $\mu(T)$ over $\calT_{\xvec}$. By proving that $\eg(\xvec)$ is close to $e^{-\barg(\xvec)}$ for each suitable degree sequence $\xvec$, and evaluating $\barg(\xvec)$, we will establish the following.
Suppose that the conditions of Theorem \[main-sparse\] hold and that $\xvec$ is a suitable degree sequence. Then with $\eta$, $R$ and $\Hd$ as defined as in Theorem \[main-sparse\], $$\begin{aligned}
\E \tau_{\dvec}(\xvec)
= \Hd\, \binom{(d-1)n}{n-2}^{\!-1}
&\biggl(\,\prod_{j=1}^n \binom{d_j-1}{x_j-1} \biggr)
\,\exp\biggl(
\frac{(R+d^2)^2}{4d^2} - \dfrac{1}{4} - \lambda(\xvec) - \lambda(\xvec)^2 \\
&{\kern5em}- \frac{1}{n}\sum_{j=1}^n (x_j-1)(d_j-x_j)
+ O\left(\frac{\dmax^4}{(d-2) n} + \eta \right)\biggr).
\end{aligned}$$ \[just-one-x\]
The structure of the rest of the paper is as follows. In Section \[s:tree-averaging\] we generalise the function $\mu$ and prove a concentration result for trees with given degrees. This proof will involve a martingale concentration result of McDiarmid [@McDiarmidChapter] which we discuss in Section \[s:concentration\]. The results of Section \[s:tree-averaging\] are applied in Section \[s:completing-calculations\] to prove that the average of $e^{-\mu(T)}$ over $T\in\calT_{\xvec}$ is close to $\exp(-\barg(\xvec))$, and hence to prove Theorem \[just-one-x\], for any suitable degree sequence $\xvec$. Finally, Theorem \[main-sparse\] is proved in Section \[s:final-approx\].
Before we begin, note that we use the following conventions in our summation notation: $\sum_{i\neq j}$ will always denote a sum over all ordered pairs $(i,j)$ with $i\neq j$ (over some appropriate range which will be clear from the context, usually $i,j = 1,\ldots, n$). On the other hand, if $i$ is fixed and we wish to sum over all $j\neq i$ (for example, over all $j\in\{ 1,2,\ldots, n\}\setminus i$) then we will write $\sum_{j: j\neq i}$.
Concentration results {#s:concentration}
=====================
Let $\mathcal{P}=(\varOmega,\calF,\mathbb{P})$ be a finite probability space. A sequence $\calF_0,\ldots,\calF_n$ of $\sigma$-subfields of $\calF$ is a *filter* if $\calF_0\subseteq\cdots\subseteq\calF_n$. A sequence $Y_0,\ldots,Y_n$ of random variables on $\mathcal{P}$ is a *martingale with respect to $\calF_0,\ldots,\calF_n$* if
- $Y_j$ is $\calF_j$-measurable and has finite expectation, for $j=0,\ldots, n$;
- $\E(Y_j \st \calF_{j-1}) = Y_{j-1}$ for $j=1,\ldots, n$.
An important example of a martingale is made by the so-called *Doob martingale process*. Suppose $X_1,X_2,\ldots,X_n$ are random variables on $\mathcal{P}$ and $f(X_1,X_2,\ldots,X_n)$ is a random variable on $\mathcal{P}$ of bounded expectation. Let $\sigma(X_1,\ldots, X_j)$ denote the $\sigma$-field generated by the random variables $X_1,\ldots, X_j$. Define the martingale $\{Y_j\}$ with respect to the filter $\{\calF_j\}$, where for each $j$, $\calF_j=\sigma(X_1,\ldots,X_j)$ and $Y_j=\E(f(X_1,X_2,\ldots,X_n)\st \calF_j)$. In particular, $\calF_0=\{\emptyset,\varOmega\}$ and $Y_0=\E f(X_1,X_2,\ldots,X_n)$.
In this section we state some concentration results for martingales. See McDiarmid [@McDiarmidChapter] for further background and for any definitions not given here. Following McDiarmid [@McDiarmidChapter], for $j=1,\ldots, n$ we define the *conditional range* of $Y_j$ as $$\operatorname{ran}(Y_j\mid \calF_{j-1}) =
\operatorname{ess\, sup}(Y_j\mid \calF_{j-1}) +
\operatorname{ess\, sup}(- Y_j\mid \calF_{j-1}).
\label{con-range}$$ Here “essential supremum” may be replaced by “supremum”, as in [@McDiarmidChapter], if the probability distribution is positive over $\varOmega$.
Our main tool is the following result from McDiarmid [@McDiarmidChapter]. The tail bound on the probability is given by [@McDiarmidChapter Theorem 3.14]. The upper estimate on the moment generating function $\E(e^{h Y_n})$ is an intermediate step of McDiarmid’s proof, see [@McDiarmidChapter Section 3.5]. The lower bound on $K$ is due to Jensen’s inequality.
*([@McDiarmidChapter])*Suppose that $\mathcal{P}=(\varOmega,\calF,\mathbb{P})$ is a finite probability space. Let $Y_0,Y_1,\ldots, Y_n$ be a martingale on $\mathcal{P}$ with respect to a filter $\calF_0,\calF_1\ldots,\calF_n$, where $\calF_0 = \{\emptyset,\varOmega\}$, such that $$\sum_{j=1}^n \left( \operatorname{ran}(Y_j\mid \calF_{j-1})\right)^2
\leq \hat{r}^2\quad a.s.$$ for some real $\hat{r}$. Then $$\E e^{Y_n} = e^{Y_0 + K}$$ where $0\leq K\leq \nfrac{1}{8}\, \hat{r}^2$. Furthermore, for any real $t > 0$, $$\Pr(|Y_n - Y_0| \geq t) \leq 2\, \exp( -2t^2/\hat{r}^2).$$ \[martingales\]
As a corollary, we obtain a concentration result for functions of sets of a given size.
\[subsets\] Let $\binom{[N]}{r}$ be the set of $r$-subsets of $\{1,\ldots,N\}$ and let $h:\binom{[N]}{r}\to \Reals$ be given. Let $C$ be a uniformly random element of $\binom{[N]}{r}$. Suppose that there exists $\alpha\geq 0$ such that $$\abs{h(A)-h(A')} \le \alpha$$ for any $A, A' \in \binom{[N]}{r}$ with $|A \cap A'| = r-1$. Then $$\E e^{h(C)} = \exp\left(\E h(C)+ K\right)
\label{henry}$$ where $K$ is a real constant such that $0 \leq K \leq \tfrac18 \min\{r,N-r\} \alpha^2 $. Furthermore, for any real $t>0$, $${\rm Pr}( |h(C) - \E h(C)| \geq t) \leq 2 \exp\left( -\frac{2t^2}{\min\{r,N-r\} \alpha^2}\right).$$
Let $S_N$ denote the set of permutations of $\{1,\ldots, N\}$ and $\tau = (\tau_1,\ldots, \tau_N)$ be a uniform random element of $S_N$. Note that the set $\{\tau_{1}, \ldots, \tau_r\}$ is a uniformly random element of $\binom{[N]}{r}$. Define $\tilde{h}:S_N \to \mathbb{R}$ by $\tilde{h}(\omega) = h(\{\omega_{1}, \ldots, \omega_r\})$ for all $\omega\in S_N$. Then $$|\tilde{h}(\rho) - \tilde{h}(\rho')| \leq \alpha$$ for all permutations $\rho$, $\rho'\in S_N$ such that $\rho^{-1}\rho'$ is a transposition. Given $\omega = (\omega_1,\ldots,\omega_N) \in S_N$, for $k=0,\ldots, N$ let $$\tilde{h}_k(\omega) =
\E\left(\tilde{h}(\tau) \st \tau_j=\omega_j \, \text{ for } \, j=1,\ldots k
\right).$$ Clearly, $\tilde{h}_0(\tau), \ldots, \tilde{h}_N(\tau)$ forms a martingale: it is the result of the Doob martingale process for $\tilde{h}(\tau)$. It follows from Frieze and Pittel [@FP Lemma 11] that $$\operatorname{ran}\(\tilde{h}_k(\tau) \mid \sigma(\tau_1,\ldots, \tau_{k-1}) \) \leq \alpha.$$ Moreover, for any $\omega \in S_N$ and $k\in\{ r,\ldots, N\}$, we have $$\E\(\tilde{h}(\tau) \st \tau_j=\omega_j, \ 1\leq j\leq k\) = h(\{\omega_{1}, \ldots, \omega_r\}).$$ Therefore $\operatorname{ran}\(\tilde{h}_k(\tau) \mid \sigma(\tau_1,\ldots, \tau_{k-1}) \) = 0$ for all $k > r$. Applying Theorem \[martingales\] to the martingale $\tilde{h}_0(\tau),\ldots, \tilde{h}_N(\tau)$, we conclude that (\[henry\]) holds with $0\leq K \leq \tfrac18 r \alpha^2$, and that $${\rm Pr}( |h(C) - \E h(C)| \geq t) \leq
2 \exp\left( -\frac{2 t^2}{r \alpha^2}\right).$$ If $r\leq N-r$ then we are finished. Otherwise we repeat the above argument using the bijection between subsets and their complements.
Trees with given degrees {#s:tree-averaging}
========================
In this section we consider sums of the form $$\label{Fdef}
F(T) = \sum_{\{ j,k\} \in E(T)}\phi(j)\, \phi(k)$$ for a given function $\phi:\{ 1,2,\ldots, n\}\rightarrow [\a, \b]\subset \mathbb{R}$.
Let $\bar{F}(\xvec)$ be the average value of $F$ over all trees with a given degree sequence $\xvec$: $$\bar{F}(\xvec) = \frac{1}{|\calT_{\xvec}|}\, \sum_{T\in\calT_{\xvec}} F(T).$$ The goal of this section is to prove the following theorem, showing that the average of $e^{\xi F(T)}$ over $\calT_{\xvec}$ is close to $e^{\xi \bar{F}(\xvec)}$, for $\xi\in \{ -1,1\}$. We will measure this distance using the seminorm of $\phi$ given by $$\label{phi-norm}
\| \phi\|_m = \min_{c\in \mathbb{R}}\, \sum_{j=1}^n |\phi(j) - c |.$$ Here the minimising value of $c$ is any median of $\{ \phi(j) : j = 1,\ldots, n\}$.
Let $F$ satisfy *(\[Fdef\])*. Then for any tree degree sequence $\xvec$ and for $\xi\in\{-1,1\}$, $$\frac{1}{|\calT_{\xvec}|}\, \sum_{T\in\calT_{\xvec}}\, e^{\xi F(T)} =\exp\left(\xi \bar{F}(\xvec)+ K\right)$$ for some real constant $K$ which satisfies $0\leq K\leq \nfrac{1}{8} L_\phi$, where $$L_\phi =
(\b - \a)^3\, \min\big\{ (\b- \a) n,\,\, \| \phi\|_m\, (\ln n + 2)\big\}.$$ Furthermore, if $\widehat{T}$ is a uniformly random element of $\calT_{\xvec}$ then for any real constant $t>0$, $$\Pr( |F(\widehat{T}) - \bar{F}(\xvec)| > t)
\leq 2\,\exp\left( - 2 t^2/L_\phi \right).$$ \[tree-general\]
First we give some explicit formulae which we will need later.
\[tree-prob\] Let $\xvec$ be a tree degree sequence and consider the set $\cal{T}_\xvec$ of all trees with degree sequence $\xvec$.
- Let $S$ be a disconnected forest with vertex set $\{1,\ldots,n\}$ and degree sequence $(s_1,\ldots,s_n)$, where $s_j\le x_j$ for $j = 1,\ldots, n$. Let $S_1,\ldots,S_r$ be the components of $S$. Then the probability that a uniform random tree in $\calT_\xvec$ contains $S$ is $$\frac{ \prod_{i=1}^r \sum_{j\in V(S_i)} (x_j-s_j)}{(n-2)_{n-r}}
\prod_{j=1}^n \, (x_j-1)_{s_j-1},$$ where $(x_j-1)_{s_j-1}=x_j^{-1}$ if $s_j=0$. In particular, for distinct $j,k\in \{1,2,\ldots, n\}$, the fraction of trees in $\calT_\xvec$ in which vertices $j,k$ are adjacent is $$\frac{x_j+x_k-2}{n-2}.$$
- The average value of $F$ over $\calT_{\xvec}$ is $$\bar F(\xvec) = \frac{1}{n-2}\, \left(\sum_{k=1}^n\phi(k)\right)\left(\sum_{j=1}^n (x_j-1)\, \phi(j)\right) -
\frac{1}{n-2}\, \left(\sum_{j=1}^n (x_j-1)\, \phi(j)^2\right).$$
Define $\xvec' = (x'_1,\ldots,x'_r)$, where $x'_i = \sum_{j\in V(S_i)} (x_j - s_j)$ for $i=1,\ldots, r$. If $x'_i=0$ for any $i$ then $T$ cannot contain $S$, as $S$ is disconnected. Hence the result holds trivially in that case, and for the remainder of the proof we may assume that all entries of $\xvec'$ are positive. Next, observe that the entries of $\xvec'$ sum to $2(r-1)$, and hence $\xvec'$ is a tree degree sequence.
Each tree in $\calT_\xvec$ that contains $S$ can be formed uniquely by the following process:
1. Take any tree $T'$ on the vertex set $\{1,\ldots,r\}$ with degree sequence $\xvec'$.
2. For $i=1,\ldots, r$, replace vertex $i$ of $T'$ by $S_i$ and distribute the edges of $T'$ that were incident with $i$ amongst the vertices of $S_i$, so that each vertex $j\in V(S_i)$ has degree $x_j$ in the resulting tree.
By (\[alltrees\]), the number of choices for $T'$ in Step 1 is $\binom{r-2}{x'_1-1,\ldots,x'_r-1}$, while the number of ways to distribute edges in Step 2 is $$\prod_{i=1}^r\, \frac{x'_i!}{\prod_{j\in V(S_i)} (x_j-s_j)!}.$$ The first statement of (i) is proven by multiplying these expressions together, dividing by and simplifying. Then taking $S$ to be the edge $jk$ together with $n-2$ trivial components completes the proof of (i).
Now using linearity of expectation, (\[Fdef\]), and part (i), we calculate that $$\begin{aligned}
(n-2)\, \bar F(\xvec) &= \sum_{j < k} \,(x_j+x_k-2)\, \phi(j)\phi(k) \label{fred0}\\
&= \sum_{j\neq k} (x_j-1) \, \phi(j)\phi(k) \label{fred}\\
&= \sum_{j=1}^n (x_j-1)\, \phi(j)\, \left(\left(\sum_{k=1}^n \phi(k)\right) - \phi(j)\right) \nonumber \\
&= \left(\left(\sum_{k=1}^n \phi(k)\right)\, \sum_{j=1}^n (x_j-1)\phi(j) \right)
- \left(\sum_{j=1}^n (x_j-1)\,\phi(j)^2\right),\nonumber
\end{aligned}$$ establishing (ii).
We complete this section with the proof of Theorem \[tree-general\], which involves the process used to construct the *Pr[" u]{}fer code* of a labelled tree. The Pr[" u]{}fer code of a tree $T\in\calT$ is a sequence $\boldsymbol{b} = (b_1,\ldots, b_{n-2})\in \{1,2,\ldots, n\}^{n-2}$. Given $T$, find the unique neighbour $b_1$ of the lowest-labelled leaf $a_1$. Then $b_1$ becomes the first entry in the Pr[" u]{}fer code for $T$. We find the next entry recursively by considering the tree $T - a_1$ with the first leaf deleted. The process stops when a single edge remains: this edge is determined by the degree sequence and does not need to be recorded in the code $\boldsymbol{b}$. We will refer to this process as the *Pr[" u]{}fer process* with input $T$. See Figure \[f:prufer\] for an example.
(0,1) – (4,1); (1,1) – (1,2); (2,1) – (2,2); (0,1) circle (0.1); (1,1) circle (0.1); (1,2) circle (0.1); (2,1) circle (0.1); (2,2) circle (0.1); (3,1) circle (0.1); (4,1) circle (0.1); at (0,0.9) [3]{}; at (1,0.9) [2]{}; at (0.9,2) [6]{}; at (2,0.9) [7]{}; at (2.1,2) [4]{}; at (3,0.9) [1]{}; at (4,0.9) [5]{}; at (6,1) [$\Rightarrow$]{}; at (8,1) [$(2,7,1,7,2)$]{};
The correspondence between trees and Pr[" u]{}fer codes is a bijection: see for example Moon [@moon pp. 5-6]. This provides a proof of Cayley’s formula and of (\[alltrees\]).
The following useful property of the Pr[" u]{}fer process may be proved by induction on $j$.
Let $\xvec$ be a tree degree sequence and let $T\in\calT_{\xvec}$. Suppose that the Pr[" u]{}fer process with input $T$ produces the Pr[" u]{}fer code $\bvec$ and the sequence $(a_1,\ldots, a_{n-2})$ of “leaves”. For any $j = 1,\ldots, n-2$, the initial sequence $(a_1,\ldots, a_j)$ is uniquely determined by $\xvec$ and $(b_1,\ldots, b_{j-1})$. \[determined\]
When there is more than one tree under consideration we will write $a_j(T)$, $b_j(T)$ for the vertices identified at step $j$ of the Pr[" u]{}fer process for the tree $T$. To prove Theorem \[tree-general\] we work with a martingale defined using the Pr[" u]{}fer code of a tree. A martingale construction based on the Pr[" u]{}fer code was given by Cooper, McGrae and Zito [@CMZ], for all labelled trees. Our martingale is restricted to trees with a given degree sequence and we study a function for which it is more difficult to bound the conditional ranges.
Suppose that $T_1$ and $T_2$ are trees on $\{ 1,2,\ldots, n\}$ with the same degree sequence. For $j=0,\ldots, n-3$, say that $T_1$ and $T_2$ are *$j$-equivalent*, and write $T_1 \simj{j} T_2$, if $b_i(T_1) = b_i(T_2)$ for $i=1,2,\ldots, j$. By Lemma \[determined\], if $T_1\simj{j} T_2$ then $a_i(T_1) = a_i(T_2)$ for $i=1,\ldots, j$. The $j$-equivalence relation induces a partition of $\calT_{\xvec}$ into equivalence classes $C_{j,1},\ldots, C_{j,r_j}$, say, with $\cup_{\ell=1}^{r_j} C_{j,\ell} = \calT_{\xvec}$.
Let $Y_{j,\ell}$ equal the average of $F(T)$ over $T\in C_{j,\ell}$, and define the function $Y_j$ on $\calT_{\xvec}$ by $Y_j(T) = Y_{j,\ell}$ if $T\in C_{j,\ell}$. Finally, define the random variable $Y_j = Y_j(\widehat{T})$, where $\widehat{T}$ is a uniformly random element of $\calT_{\xvec}$. Then $Y_0$ is the constant function which takes the value $\E F(\widehat{T})$ everywhere, and $Y_{n-3} = F(\widehat{T})$, since each equivalence class $C_{n-3,\ell}$ is a set of size 1. Observe that $Y_0,\ldots, Y_{n-3}$ is a martingale with respect to the filter $\calF_0, \ldots, \calF_{n-3}$, where, for each $j$, $\calF_j$ is generated by the sets $C_{j,1},\ldots, C_{j,r_j}$. In fact, this is the Doob martingale process for the function $F(\widehat{T})$ of the random variables $b_1(\widehat{T}),\ldots,b_{n-3}(\widehat{T})$, which determine $\widehat{T}$ uniquely.
To apply Theorem \[martingales\] we must calculate a value for $\hat{r}^2$. Suppose that $T_1$ and $T_2$ are $(j-1)$-equivalent, where $T_1, T_2\in\calT_{\xvec}$ and $j\in \{1, \ldots, n-3\}$. Then $a_j(T_1) = a_j(T_2)$, again by Lemma \[determined\]. For ease of notation, write $a_i$ instead of $a_i(T_1)$ (or $a_i(T_2)$) for $i=1,\ldots, j$, and write $b_i$ instead of $b_i(T_1)$ (or $b_i(T_2)$) for $i=1,\ldots, j-1$.
For $s=1,2$ let $T_s'$ be the tree (with $n-j$ vertices) obtained by deleting the vertices $a_1,\ldots, a_j$ from $T_s$. Both $T'_1$ and $T'_2$ have vertex set $$V_j = \{ 1,2, \ldots, n\} \setminus \{ a_1,\ldots, a_{j}\}.$$ If $b_j(T_1) = b_j(T_2)$ then $T'_1$ and $T'_2$ have the same degree sequence, since (in this case) precisely the same edges have been deleted from $T_1$ and $T_2$. In this case, $Y_j(T_1)= Y_j(T_2)$.
Otherwise, the degree sequences of $T'_1$, $T'_2$ differ only for the two vertices $b_j(T_1)$ and $b_j(T_2)$. Specifically, vertex $b_j(T_1)$ has degree in $T_1'$ which is equal to its degree in $T_2'$ minus 1, while vertex $b_j(T_2)$ has degree in $T_1'$ which is equal to its degree in $T_2'$ plus 1. Hence $T'_1$ and $T'_2$ have the same degree on all vertices in the set $$U_j(T_1,T_2) = V_j \setminus \{ b_j(T_1),\, b_j(T_2)\}.$$ For $s=1,2$, let $\yvec_s$ be the degree sequence of $T_s'$ (on the vertex set $V_j$) and let $\calT'_s$ denote the set of all trees on the vertex set $V_j$ with degree sequence $\yvec_s$. Observe that $\yvec_s$ and $\calT'_s$ depend only on $(b_1,\ldots, b_{j-1}, b_j(T_s))$ and $\xvec$. By relabelling the equivalence classes if necessary, we may assume that $T_s\in C_{j,s}$ for $s=1,2$. The map $\varphi:C_{j,s} \rightarrow \calT'_s$ which sends a tree $T\in C_{j,s}$ to $T\setminus \{ a_1,\ldots, a_j\}$ is a bijection. To see this, observe that the inverse map $\varphi^{-1}$ takes a tree in $\calT'_s$, adds the vertices $a_1,\ldots, a_j$ and the edges $$\{ \{ a_1,b_1\},\ldots, \{ a_{j-1}, b_{j-1}\},\,\, \{ a_j, b_j(T_s)\}\}$$ giving a tree in $C_{j,s}$. Therefore, for $s=1,2$, $$\frac{1}{|C_{j,s}|}\, \sum_{T\in C_{j,s}} F(T\setminus \{ a_1,\ldots, a_j\})
= \frac{1}{|\calT'_s|}\, \sum_{T'\in \calT'_s} F(T').$$ Combining this with (\[Fdef\]) and the definition of $Y_{j,s}$, we see that for $s=1,2$, $$\begin{aligned}
Y_{j,s} &= \frac{1}{|C_{j,s}|}\,\sum_{T\in C_{j,s}}
\, \sum_{\{k,\ell\}\in E(T)} \phi(k)\phi(\ell)\\
&= \left( \sum_{i=1}^{j-1} \phi(a_i)\phi(b_i)\right)
+ \phi(a_j)\phi(b_j(T_s)) +
\frac{1}{|C_{j,s}|}\, \sum_{T\in C_{j,s}} F(T\setminus \{ a_1,\ldots, a_j\})\\
&= \left( \sum_{i=1}^{j-1} \phi(a_i)\phi(b_i)\right)
+ \phi(a_j)\phi(b_j(T_s)) +
\frac{1}{|\calT'_s|}\, \sum_{T'\in \calT'_s} F(T').\end{aligned}$$ Applying Lemma \[tree-prob\](ii) gives $$\begin{aligned}
Y_{j,1} - Y_{j,2}
&= \left(\phi(b_j(T_1)) - \phi(b_j(T_2))\right)\,
\biggl( \phi(a_j) - \frac{1}{n-j-2}\sum_{\ell \in U_j(T_1,T_2)} \phi(\ell) \biggr)\\
&= \frac{\phi(b_j(T_1)) - \phi(b_j(T_2))}{n-j-2}\, \sum_{\ell\in U_j(T_1,T_2)}
\(\phi(a_j) - \phi(\ell)\).
\end{aligned}$$ (Note that if $T_1\simj{j} T_2$ then $b_j(T_1)=b_j(T_2)$ and the above equality also holds.)
Recall the definition of $\|\phi\|_m$ from (\[phi-norm\]), and let $c\in\mathbb{R}$ be the minimising value in this definition. By the triangle inequality, $$\begin{aligned}
\frac{1}{n-j-2}\, \biggl|\sum_{\ell\in U_j(T_1,T_2)} (\phi(a_j) - \phi(\ell)) \biggr|
&\leq |\phi(a_j) - c| + \frac{1}{n-j-2}\, \biggl|\sum_{\ell\in U_j(T_1,T_2)} (c - \phi(\ell))\biggr| \nonumber \\
&\leq |\phi(a_j) - c| + \frac{\| \phi\|_m }{n-j-2}
\label{intermediate}
\end{aligned}$$ since $U_j(T_1,T_2)$ has $n-j-2$ elements and $j\leq n-3$. Therefore, for any equivalence class $C_{j-1,\ell}$, we have $$\begin{aligned}
& \biggl(\, \sup_{T'\in C_{j-1,\ell}} Y_j(T') + \sup_{T'\in C_{j-1,\ell}} (-Y_j(T'))\biggr)^2 \nonumber\\
&\hspace*{3cm} \leq \frac{(\b-\a)^2}{(n-j-2)^2}\,
\sup_{T_1,T_2\in C_{j-1,\ell}} \biggl(\, \sum_{\ell\in U_j(T_1,T_2)} (\phi(a_j) - \phi(\ell))\biggr)^{\!2}\nonumber\\
&\hspace*{3cm} \leq (\b-\a)^3\, \min\left\{ \b-\a,\, | \phi(a_j) -c| + \frac{\| \phi\|_m}{n-j-2}\right\}.\label{OK}
\end{aligned}$$ (Here we take the minimum of two possible upper bounds: the first arises from taking the worst case summand for both factors in the line above, while the second arises by applying (\[intermediate\]) to one of the factors.)
Now let $C_{j-1}(\widehat{T})$ denote the random set which is the equivalence class with respect to $\simj{j-1}$ which contains $\widehat{T}$. It follows from (\[OK\]) that $$\begin{aligned}
\operatorname{ran}(Y_j\mid \calF_{j-1})^2 &=
\biggl(\,\sup_{T\in C_{j-1}(\widehat{T})} Y_j(T)
+ \sup_{T\in C_{j-1}(\widehat{T})} (-Y_j(T)) \biggr)^2
\\
&\leq (\b-\a)^3\, \min\left\{ \b-\a,\, | \phi(a_j(\widehat{T})) -c| + \frac{\| \phi\|_m}{n-j-2}\right\}.\end{aligned}$$ Using the definition of $c$, the standard upper bound on the harmonic series and the fact that each vertex is chosen as $a_j(\widehat{T})$ at most once during the Pr[" u]{}fer process, we get that $$\sum_{j=1}^{n-3} \operatorname{ran}(Y_j\mid \calF_{j-1})^2
\leq (\b - \a)^3 \, \min\left\{ (\b - \a) n,\,\, \| \phi\|_m\ (\ln n + 2)\right\}.$$ Observe that the left hand side does not change if $F$ is replaced by $-F$ (and hence, the same bound is obtained whether $\xi = 1$ or $\xi = -1$). Since $\E(e^{Y_{n-3}}) =
\E(e^{F(\widehat{T})})$ and $Y_0 = \E(F(\widehat{T}))$, applying Theorem \[martingales\] completes the proof.
Proof of Theorem \[just-one-x\] {#s:completing-calculations}
===============================
First we note the following corollary of Theorem \[tree-general\]. Recall the definition of $\eg(\xvec)$ and $\barg(\xvec)$ from (\[beta-def\]), (\[mubar-def\]), respectively.
Under the conditions of Theorem \[main-sparse\], $$\eg(\xvec) =\exp\left(-\barg(\xvec)+O\!\left(\min\left\{\frac{\dmax^4}{(d-2)^2n},\,
\frac{\dmax^3\, \ln n}{(d-2)n}\right\} \right)\right).$$ \[Eeg\]
Set $$\phi(j) = \frac{d_j-x_j}{\sqrt{(d-2)n+2}}$$ for $j\in \{ 1,2,\ldots, n\}$, and let $\xi=-1$. We can take $\a=0$ and $\b=\dmax/\sqrt{(d-2)n+2}$. Next, we bound $$\| \phi\|_m \leq \sum_{j=1}^n \frac{d_j-x_j}{\sqrt{(d-2)n+2}} = \sqrt{(d-2)n + 2}.$$ Finally, observe that $$\frac{\dmax^3 (\| \phi\|_m + \b)(\ln n + 2)}{((d-2)n+2)^{3/2}}
= O\left(\frac{\dmax^3\, \ln n}{(d-2)n} + \frac{\dmax^4\ln n}{((d-2)n)^{2}}\right)
= O\left(\frac{\dmax^3\,\ln n}{(d-2)n}\right).$$ Now the result follows from Theorem \[tree-general\].
We can now prove Theorem \[just-one-x\], giving an asymptotic expression for the expected number $\E \tau_\dvec(\xvec)$ of spanning trees in $\calG_{\dvec}$ with degree sequence $\xvec$.
Firstly note that, by , $$\begin{aligned}
\barg(\xvec)
&= \frac{1}{(n-2)((d-2)n+2)}\,
\sum_{j\ne k} \,(x_j-1)(d_j-x_j)(d_k-x_k)
\label{g-useful}\\
&= \frac{1}{n}\, \sum_{j=1}^n (x_j-1)(d_j-x_j)
+ O\left(\frac{\dmax^2}{(d-2)n}\right). \label{barg-expression}\end{aligned}$$ We rewrite (\[one-x\]) as $$\begin{aligned}
&\E \tau_{\dvec}(\xvec) \nonumber\\*
&\quad = e^{O(\dmax^4/((d-2)n))}\,
\frac{(dn/2)_{n-1}\, 2^{n-1}\, \hat{d}^n}{(dn)_{n}}\,
\binom{(d-1)n}{n-2}^{\!\!-1}\, \left(\prod_{j=1}^n \binom{d_j-1}{x_j-1} \right) \,
e^{f(\xvec)}\, \beta(\xvec)
\label{new}\end{aligned}$$ with $f(\xvec)$ as defined in (\[fg-def\]). Applying Stirling’s approximation gives $$\frac{(dn/2)_{n-1}\, 2^{n-1}\, \hat{d}^n}{(dn)_{n}} = \Hd\,
\left(1 + O\left(\frac{1}{(d-2)n}\right)\right)$$ where $\Hd$ is defined in the statement of Theorem \[main-sparse\]. Combining Lemma \[Eeg\] and (\[barg-expression\]) gives $$\begin{aligned}
\beta(\xvec)
&=
\exp\left(-\barg(\xvec) + O\left(\min\left\{\frac{\dmax^4}{(d-2)^2 n},\,
\frac{\dmax^3\, \ln n}{(d-2)n}\right\} \right)\right)\nonumber \\[0.5ex]
&= \exp\left(-
\frac{1}{n}\, \sum_{j=1}^n (x_j-1)(d_j-x_j) \right. \nonumber \\[-0.6ex]
& \left. \hspace*{3cm} {}
+ O\left(\frac{\dmax^2}{(d-2)n} + \min\left\{\frac{\dmax^4}{(d-2)^2 n},\,
\frac{\dmax^3\, \ln n}{(d-2)n}\right\} \right)\right).
\label{beta-approx}\end{aligned}$$ In some cases, when $d-2$ is small, we can obtain a smaller error bound by a different argument. Observe that $$\label{beta-bounds}
e^{- \barg(\xvec)} \leq \beta(\xvec)\leq 1,$$ using Jensen’s inequality for the lower bound. It follows from (\[barg-expression\]) that $$\barg(\xvec) = O\left(\frac{\dmax^2}{(d-2)n} + (d-2)\dmax\right).$$ Hence we can replace the upper bound on $\beta(\xvec)$ in (\[beta-bounds\]) by $e^{-\barg(\xvec)}$ if we include an error term of this magnitude, leading to $$\begin{aligned}
\beta(\xvec)
&= \exp\left(-\barg(\xvec)
+ O\left( \frac{\dmax^2}{(d-2)n} + (d-2)\dmax\right)\right) \nonumber \\
&= \exp\biggl(- \frac{1}{n}\, \sum_{j=1}^n (x_j-1)(d_j-x_j)
+ O\biggl( \frac{\dmax^2}{(d-2)n} + (d-2)\dmax\biggr) \biggr)\label{beta-alternative}\end{aligned}$$ using (\[barg-expression\]). We may choose to use either this expression or (\[beta-approx\]), whichever gives the smaller bound. Finally, observe that $$\label{lambda0}
\lambda_0 + \lambda_0^2 = \frac{(R+d^2)^2}{4d^2} - \dfrac{1}{4}.$$ Combining this with (\[fg-def\]), (\[new\]), (\[beta-approx\]) and (\[beta-alternative\]) completes the proof.
Proof of Theorem \[main-sparse\] {#s:final-approx}
================================
In this section we prove Theorem \[main-sparse\] by summing the expression from Theorem \[just-one-x\] over all suitable degree sequences $\xvec$. Given a suitable degree sequence $\xvec$, define $$g(\xvec) = f(\xvec) - \barg(\xvec) =
\frac{(R+d^2)^2}{4d^2} - \dfrac{1}{4} -\lambda(\xvec) - \lambda(\xvec)^2 - \barg(\xvec),
\label{g-def}$$ using (\[lambda0\]). By (\[barg-expression\]) and Theorem \[just-one-x\] we have $$\begin{aligned}
\E \tau_\dvec &=
\Hd\, \sum_{\xvec}
\binom{(d-1)n}{n-2}^{\!\!-1}\, \left(\,\prod_{j=1}^n \binom{d_j-1}{x_j-1}\right)
\nonumber \\ & \hspace*{4cm} {} \times
\exp\left( g(\xvec)
+ O\left( \frac{\dmax^4}{(d-2)n} + \eta \right)\right)\,
\label{sum-over-x}\end{aligned}$$ where the sum is over all suitable degree sequences $\xvec$. We now interpret this sum as an expected value of a function of a nonuniform distribution on suitable degree sequences.
\[useful-later\] Fix a partition $A_1,\ldots, A_n$ of $\{ 1,2,\ldots, (d-1)n\}$ such that $\card{A_j} = d_j-1$ for $j=1,\ldots, n$, and let $B$ be a uniformly random subset of $\{ 1,2,\ldots, (d-1)n\}$ of size $n-2$. Define the random vector $\X = \X(B) = (X_1,\ldots, X_n)$ by $X_j = |A_j\cap B| + 1$. Then $$\E \tau_\dvec = \Hd\,
\exp\left( O\left( \frac{d_\mx^4}{(d-2)n} + \eta\right)\right)\,
\E\left(e^{g(\X)}\right).$$
Let $\xvec$ be a suitable degree sequence. Since the sets $A_j$ are disjoint, there are $\prod_{j=1}^n \binom{d_j-1}{x_j-1}$ ways to choose a subset of $\{ 1,\ldots, (d-1)n\}$ with precisely $x_j-1$ elements in $A_j$, for $j=1,\ldots, n$. It follows that $$\Pr(\X = \xvec) = \binom{(d-1)n}{n-2}^{\!\!-1}\, \prod_{j=1}^n \binom{d_j-1}{x_j-1}.$$ Substituting this into (\[sum-over-x\]) completes the proof.
Next, we prove that $\E\(e^{g(\X)}\)$ can be approximated by $e^{\E g(\X)}$ by applying Corollary \[subsets\]. We say that two suitable degree sequences $\xvec$ and $\xvec'$ are *adjacent* if $\xvec$ and $\xvec'$ differ in precisely two entries, say in entries $j$ and $k$, such that $x'_j = x_j + 1$ and $x_k' = x_k-1$. Adjacent degree sequences correspond to subsets $A, A'$ of $\{ 1,2,\ldots, (d-1)n\}$ of size $n-2$ which have $n-3$ elements in common. In order to apply Corollary \[subsets\] to $g$ we must bound the amount by which $g(\xvec)$ can differ from $g(\xvec')$ when $\xvec$ and $\xvec'$ are adjacent.
\[f+g\] Suppose that $\xvec$, $\xvec'$ are two suitable degree sequences which are adjacent. Then $$\left| g(\xvec) - g(\xvec')\right| =
O\left(\frac{\dmax^2}{(d-2) n}\right).$$
Recall the definition of $g$ in (\[g-def\]). Firstly, observe that $$\lambda(\xvec')^2 - \lambda(\xvec)^2 =
\left( \lambda(\xvec') - \lambda(\xvec)\right)\left(\lambda(\xvec') + \lambda(\xvec)\right)
= O(\dmax)\, \left(\lambda(\xvec') - \lambda(\xvec)\right)$$ since for any suitable $\xvec$ we have $$\lambda(\xvec) = O\left(\frac{\dmax}{(d-2)n}\right)\, \sum_{j=1}^n (d_j-x_j) = O(\dmax).$$ Next we calculate that $$\begin{aligned}
|\lambda(\xvec') - \lambda(\xvec)| = \frac{| (d_k-x_k) -(d_j-x_j-1)|}{(d-2)n+2}
= O\left(\frac{\dmax}{(d-2)n}\right).
\end{aligned}$$ Therefore $$\left|\lambda(\xvec) + \lambda(\xvec)^2 - \( \lambda(\xvec') - \lambda(\xvec')^2\)\right|
= O\left(\frac{\dmax^2}{(d-2)n}\right).$$ Now we consider $\barg$. Suppose that $\yvec$ is a vector which disagrees with $\xvec$ in precisely one position, say $y_i=x_i+ \zeta$ where $\zeta\in\{-1,1\}$. Then using (\[g-useful\]) (most conveniently in the form in ), $$\begin{aligned}
| \barg(\yvec) -\barg(\xvec)| &\leq \frac{1}{(n-2)\, ((d-2)n+2)}\, \sum_{j:j\ne i}
(d_j-x_j) \, \left| (d_i - x_i) - (x_i + x_j - 2) + \zeta\right| \\
&= O\left(\frac{\dmax}{n}\right) = O\left(\frac{\dmax^2}{(d-2)n}\right). \end{aligned}$$ Applying this twice gives a bound of the same magnitude on $|\barg(\xvec') - \barg(\xvec)|$, completing the proof.
Now we apply Corollary \[subsets\] to prove the following.
Under the conditions of Theorem \[main-sparse\], \[eEfg\] $$\E\left(e^{g(\X)}\right)=\exp\left(\E g(\X)
+O\!\left(\frac{\dmax^4}{(d-2) n}\right)\right).$$
We will apply Corollary \[subsets\] to $h(B)=g(\X(B))$, where the random set $B$ is defined in Lemma \[useful-later\]. We set $N=(d-1)n$ and $r=n-2$. Lemma \[f+g\] says that $h$ changes by at most $\alpha=O(\dmax^2/((d-2)n))$ if two entries of the vector change by 1 (one increasing and one decreasing). The value of the error term given by Corollary \[subsets\] also depends on $\min\{r,N-r\} = \min\{n-2,(d-2)n+2\}$. We consider two cases.
If $(n-2) \leq (d-2)n+2$ then Corollary \[subsets\] gives $$\begin{aligned}
\E(e^{g(\X)}) &=\exp\left(\E g(\X) +O\!\left(\frac{\dmax^4 (n-2)}{(d-2)^2 n^2}\right)\right)
\\*
&= \exp\left(\E g(\X) +O\!\left(\frac{\dmax^4}{(d-2) n}\right)\right).\end{aligned}$$ (The second equality follows since in this case $d-2 \geq 1 - \dfrac{4}{n}\geq \dfrac{1}{2}$.)
Otherwise it holds that $(d-2)n+2 < n-2$, and here Corollary \[subsets\] says that $$\begin{aligned}
\E(e^{g(\X)}) &=\exp\left(\E g(\X) +O\!\left(\frac{\dmax^4((d-2)n+2)}{(d-2)^2 n^2}\right)\right)
\\
&= \exp\left(\E g(\X) +O\!\left(\frac{\dmax^4}{(d-2) n}\right)\right),\end{aligned}$$ as required.
To approximate $\E g(\X)$, we need to be able to compute joint moments of the form $\E\big((X_j-1)_s\, (X_k-1)_t \big) $, where $\X=\X(B) = (X_1,\ldots, X_n)$. The random vector $$\X - (1,1,\ldots, 1) = (X_1-1,\ldots, X_n-1)$$ has a multivariate hypergeometric distribution, and from this it follows that the entries $X_1,\ldots, X_n$ of $\X=\X(B)$ satisfy $$\label{joint}
\E((X_i-1)_s(X_j-1)_t)=(d_i-1)_s (d_j-1)_t \, \frac{(n-2)_{s+t} }{ ((d-1)n)_{s+t}}$$ for $i\neq j$. See for example [@JKB Equation (39.6)].
We now find an asymptotic expression for $\E(g(\X))$.
\[g-expression\] Under the conditions of Theorem \[main-sparse\], $$\E(g(\X)) =
\frac{6d^2-14d+7}{4(d-1)^2}
+ \frac{R}{2(d-1)^3}
+ \frac{(2d^2-4d+1)R^2}{4(d-1)^4\, d^2}
+ O\biggl(\frac{d_\mx^3}{dn}\biggr).$$
First we estimate $\E\barg(\X)$ using (\[g-useful\]). By (\[joint\]) we have $$\begin{aligned}
&\frac{\E( (X_j-1)(d_j-X_j)(d_k-X_k) )}{(n-2)((d-2)n+2)} \\[-1ex]
&\qquad = \frac{1}{(n-2)((d-2)n+2)}\, \Big( (d_j-2)(d_k-1)\E(X_j-1)
\\ & \hspace*{12em} {} - (d_j-2)\E((X_j-1)(X_k-1)) \\
& \hspace*{12em} {} - (d_k-1)\E((X_j-1)_2)
+ \E( (X_j-1)_2 (X_k-1))\Big) \\
&\qquad = \frac{(d_j-1)_2 (d_k-1)}{(n-2)((d-2)n+2)}\, \bigg( \frac{n-2}{(d-1)n}
- \frac{2(n-2)_2}{((d-1)n)_2}
+ \frac{(n-2)_3}{((d-1)n)_3} \bigg) \\
&\qquad = \frac{(d_j-1)_2\, (d_k-1)\, ((d-2)n + 1)}{((d-1)n)_3}\\
&\qquad = (d_j-1)_2\, (d_k-1)\, \left(\frac{d-2}{(d-1)^3 n^2} + O\left(\frac{1}{d^3n^3}\right)\right).\end{aligned}$$ Now $$\begin{aligned}
\sum_{j\neq k} (d_j-1)_2 (d_k-1)
&= \sum_{j=1}^n (d_j-1)_2 \bigl( (d-1)n - (d_j-1)\bigr)\\*
&= (d-1) (R + (d-1)_2) n^2 + O(\dmax^2 dn).
\end{aligned}$$ Hence the expected value of $\barg(\X)$ is given by $$\begin{aligned}
\E \barg(\X) &= \left(\frac{d-2}{(d-1)^3 n^2} + O\left(\frac{1}{d^3n^3}\right)\right)\,
\sum_{j\neq k} (d_j-1)_2 \, (d_k-1) \nonumber \\
&= \left(\frac{d-2}{(d-1)^3 n^2} + O\left(\frac{1}{d^3n^3}\right)\right)\,
\left( (d-1)(R + (d-1)_2)\, n^2 + O(\dmax^2 dn)\right) \nonumber \\
&= \frac{(d-2)(R + (d-1)_2)}{(d-1)^2} + O\left(\frac{\dmax^2}{dn}\right).\label{emu}\end{aligned}$$
Next, recall that $$\lambda(\X ) =\frac1{2((d-2)n+2)}\, \sum_{j=1}^n (d_j-X_j)_2.$$ Applying (\[joint\]) shows that $$\begin{aligned}
& \frac{\E( (d_j - X_j)_2 )}{2((d-2)n+2)}\\
&\qquad = \frac{1}{2((d-2)n+2)}\,\left( (d_j-1)_2 - 2(d_j-2)\E(X_j-1 ) + \E((X_j-1)_2 )\right)\\
&\qquad = \frac{(d_j-1)_2}{2((d-2)n+2)}\,\left( 1 - \frac{2(n-2)}{(d-1)n} + \frac{(n-2)_2}{((d-1)n)_2}\right)\\
&\qquad = \frac{(d_j-1)_2\, ((d-2)n+1)}{2 ((d-1)n)_2} \\
& \qquad = (d_j-1)_2\,
\left(\frac{(d-2)}{2 (d-1)^2 n} + O\left(\frac{1}{d^2n^2}\right)\right). \end{aligned}$$ Therefore $$\begin{aligned}
\E( \lambda(\X) ) &=
\sum_{j=1}^n (d_j-1)_2\,
\left(\frac{(d-2)}{2 (d-1)^2 n} + O\left(\frac{1}{d^2n^2}\right)\right) \nonumber\\
&= \frac{(d-2)(R + (d-1)_2)}{2 (d-1)^2 } + O\left(\frac{\dmax}{dn}\right).
\label{f1}\end{aligned}$$ The same approach works for $\E(\lambda(\X)^2)$ but the details are a little messier. Observe that $$\begin{aligned}
&\lambda(\X)^2 \nonumber \\*[-1ex]
&= \frac{1}{4((d-2)n + 2)^2}\,
\left(\left(\sum_{j\neq k} (d_j-X_j)_2 (d_k-X_k)_2 \right)
+ \sum_{j=1}^n (d_j-X_j)^2 (d_j-X_j-1)^2\right).
\label{expand}\end{aligned}$$ Applying (\[joint\]) to the off-diagonal summands gives $$\begin{aligned}
&\frac{\E( (d_j-X_j)_2 (d_k-X_k)_2 )}{4((d-2)n+2)^2}\, \\
&\quad= \frac{1}{4((d-2)n+2)^2}\, \Big(
(d_j-1)_2(d_k-1)_2 - 2(d_j-1)_2 (d_k-2) \E(X_k-1)\\ & \qquad \quad {} - 2(d_j-2)(d_k-1)_2 \E(X_j-1) + (d_j-1)_2 \E((X_k-1)_2 ) \\ & \qquad \quad {}+ 4(d_j-2)(d_k-2) \E((X_j-1)(X_k-1))
+ (d_k-1)_2 \E((X_j-1)_2 ) \\ & \qquad \quad {}
- 2(d_j-2)\, \E( (X_j-1)(X_k-1)_2 ) - 2(d_k-2)\, \E((X_j-1)_2(X_k-1)) \\
& \hspace*{95mm} {} + \E((X_j-1)_2(X_k-1)_2) \Big)\\
&\quad = \frac{(d_j-1)_2 (d_k-1)_2}{4((d-2)n+2)^2}\,
\left( 1 - \frac{4(n-2)}{(d-1)n} + \frac{6(n-2)_2}{((d-1)n)_2} - \frac{4(n-2)_3}{((d-1)n)_3}
+ \frac{(n-2)_4}{((d-1)n)_4}\right)\\
&\quad =
\frac{(d_j-1)_2 (d_k-1)_2 \,((d-2)n+1)_3}{4((d-1)n)_4 ((d-2)n+2)}\\
&\quad = (d_j-1)_2 (d_k-1)_2 \, \left( \frac{(d-2)^2\, }{4(d-1)^4 n^2} + O\left(\frac{1}{d^3n^3}\right)\right).
\end{aligned}$$ Next, calculate $$\sum_{j\neq k} (d_j-1)_2 (d_k-1)_2 = (R + (d-1)_2)^2 n^2 +
O(\dmax^3 dn).$$ Therefore the contribution to $\lambda(\X )^2$ from the off-diagonal summands is $$\begin{aligned}
& \left(\frac{(d-2)^2}{4(d-1)^4 n^2} + O\left(\frac{1}{d^3n^3}\right)\right)\,
\sum_{j \neq k} (d_j-1)_2 (d_k-1)_2\\
&= \left(\frac{(d-2)^2}{4(d-1)^4 n^2 } + O\left(\frac{1}{d^3 n^3}\right)\right)\,
\Bigl( (R + (d-1)_2)^2 n^2 + O\left(\dmax^3 d n\right)\Bigr)\\
&= \frac{(d-2)^2 (R+(d-1)_2)^2}{4(d-1)^4} + O\left(\frac{\dmax^3}{dn}\right).\end{aligned}$$ The contribution to $\E(\lambda(\X)^2 )$ from the diagonal terms of (\[expand\]) (that is, the second summation in (\[expand\])) is $$\begin{aligned}
\frac{1}{4((d-2)n + 2)^2} \, \sum_{j=1}^n \E( (d_j-X_j)^2 (d_j-X_j-1)^2 )
&= O\left(\frac{\dmax^2}{(d-2) n}\right)\, \E(\lambda(\X))\\
&= O\left(\frac{\dmax^3}{d n}\right), \end{aligned}$$ using (\[f1\]). Therefore $$\E(\lambda(\X)^2 ) = \frac{(d-2)^2\, (R+ (d-1)_2)^2}{4(d-1)^4}
+ O\left(\frac{\dmax^3}{dn}\right).
\label{f2}$$ The result follows by combining (\[lambda0\]), (\[emu\]), (\[f1\]) and (\[f2\]), after some rearranging.
Now we may easily prove our main theorem.
The number of graphs with degree sequence $\dvec$ is positive when $n$ is sufficiently large, by (\[Nd\]). That is, $\dvec$ is graphical for sufficiently large $n$. The claimed asymptotic expression for $\E \tau_{\dvec}$ then follows immediately from Lemmas \[useful-later\], \[eEfg\] and \[g-expression\]. We also briefly justify the bound $$\begin{aligned}
\eta &= \min\biggl\{ \frac{d_\mx^4}{(d-2)^2n},\,
\frac{d_\mx^3\log n}{(d-2)n},\,
d_\mx(d-2)\biggr\} \\
&= O\biggl( \frac{d_\mx^4}{(d-2)n} + \frac{(\log n)^{5/2}}{n^{1/2}}\biggr).\end{aligned}$$ Note that $(d_\mx^3\log n)/((d-2)n)\le d_\mx^4/((d-2)n)$ if $\dmax\ge \log n$. When $\dmax\leq \log n$, take the geometric mean of $(d_\mx^3\log n)/((d-2)n)$ and $d_\mx(d-2)$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We would like to thank the referees for their helpful comments.
[99]{}
N. Alon, The number of spanning trees in regular graphs, *Random Structures Algorithms* **1** (1990), no. 2, 175–182.
B. Bollob[á]{}s, A probabilistic proof of an asymptotic formula for the number of labelled regular graphs, *Europ. J. Combin.* [**1**]{} (1980), 311–316.
F. Chung and S.-T. Yau, Coverings, heat kernels and spanning trees, *Electron. J. Combin.* [**6**]{} (1999), \#R12.
C. Cooper, A.R.A. McGrae and M. Zito, Martingales on trees and the empire chromatic number of random trees, in *Fundamentals of Computation Theory 2009* (M. Kuty[ł]{}owski, M. G[ȩ]{}bala and W. Charatonik, eds.), LNCS vol. 5699, Springer, Heidelberg, 2009, pp. 74–83.
A. Frieze and B. Pittel, Perfect matchings in random graphs with prescribed minimal degree, in *Mathematics and Computer Science III: Algorithms, Trees, Combinatorics and Probabilities* (M. Drmota, P. Flajolet, D. Gardy and B. Gittenberger, eds.), Birkh[ä]{}user, Basel, 2004, pp. 95–132.
C. Greenhill, M. Kwan and D. Wind, On the number of spanning trees in random regular graphs, *Electron. J. Combin.* [**21**]{} (2014) \#P1.45.
N.L. Johnson, S. Kotz and N. Balakrishnan, *Discrete Multivariate Distributions*, Wiley, New York, 1997.
A.V. Kostochka, The number of spanning trees in graphs with a given degree sequence, *Random Structures Algorithms* **6** (1995), no. 2-3, 269–274.
R. Lyons, Asymptotic enumeration of spanning trees, *Combin. Probab. Comput.* **14** (2005), no. 4, 491–522.
B.D. McKay, Subgraphs of random graphs with specified degrees, *Congr. Numer.*, [**33**]{} (1981) 213–223.
B.D. McKay, Spanning trees in random regular graphs, in *Proceedings of the Third Carribean Conference on Combinatorics and Computing* (C. Cadogan, ed.), Department of Mathematics, University of the West Indies, 1981, pp. 139–143.
B. D. McKay, Spanning trees in regular graphs, *European J. Combin.* **4** (1983), no. 2, 149–160.
B.D. McKay, Asymptotics for symmetric 0-1 matrices with prescribed row sums, *Ars Combin.*, [**19A**]{} (1985) 15–26.
C. McDiarmid, Concentration, in *Probabilistic Methods for Algorithmic Discrete Mathematics* (M. Habib, C. McDiarmid, J. Ramirez-Alfonsin, B. Reed, eds.), Springer, Berlin, 1998, pages 195–248.
C. Merino and D.J. Welsh, Forests, colorings and acyclic orientations of the square lattice, *Ann. Combin.* **3** (1999), no. 2-4, 417–429.
J.W. Moon, *Counting Labelled Trees*, Canadian Mathematical Monographs, Vol. 1, Canadian Mathematical Congress, Montreal, 1970.
L.W. Shapiro, An electrical lemma, *Math. Mag.* **60** (1987), no. 1, 36–38.
F. Wu, Number of spanning trees on a lattice, *J. Phys. A* **10** (1977), no. 6, L113.
[^1]: Research supported by the Australian Research Council, Discovery Project DP140101519.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the finite part of the adjoint $L$-function (including contributions from all non-archimedean places, including ramified places) is holomorphic in ${{\mathrm{Re}}}(s) \ge 1/2$ for a cuspidal automorphic representation of ${{\mathrm{GL}}}_3$ over a number field. This improves the main result of [@H18]. We obtain more general results for twisted adjoint $L$-functions of both ${{\mathrm{GL}}}_3$ and quasisplit unitary groups. For unitary groups, we explicate the relationship between poles of twisted adjoint $L$-functions, endoscopy, and the structure of the stable base change lifting.'
address:
- 'Department of Mathematics, State University of New York at Buffalo, Buffalo, NY, USA'
- 'Department of Mathematics and Statistics, University of Calgary, Canada'
author:
- Joseph Hundley
- Qing Zhang
title: 'Adjoint $L$-functions for ${{\mathrm{GL}}}(3)$ and ${{\mathrm {U}}}_{2,1}$'
---
Introduction
============
Motivation and Background
-------------------------
Let $F$ be a number field and ${{\mathbb {A}}}$ be its ring of adeles. Let $H$ be a connected, reductive $F$-group, isomorphic to ${{\mathrm{GL}}}_3$ over the algebraic closure ${\overline}F$ of $F.$ The $L$-group $^LH$ of $H$ is then the semi-direct product of ${{\mathrm{GL}}}_3({\mathbb{C}})$ with a Galois or Weil group, depending on which form of the $L$-group we consider (cf.[@BorelCorvallis]). Differentiating the action of $^LH$ on its own identity component by conjugation, we obtain an action on ${\mathfrak}{gl}_3({\mathbb{C}}).$ The subspace ${\mathfrak}{sl}_3({\mathbb{C}})$ is preserved. We denote the action of $^LH$ on ${\mathfrak}{sl}_3({\mathbb{C}})$ by ${{\mathrm{Ad}}}.$
In this paper we study the twisted adjoint $L$-function $L(s, \pi, {{\mathrm{Ad}}}\times \chi)$ where $\chi$ is a Hecke character. This work is motivated by the following simple conjecture regarding the untwisted $L$-function in the split case.
\[conj: adjoint GL3 is entire\] Let $\pi$ be an irreducible cuspidal automorphic representation of ${{\mathrm{GL}}}_n({\mathbb{A}})$, where ${\mathbb{A}}$ is the adele ring of a global field. Let ${{\mathrm{Ad}}}$ denote the adjoint representation of ${{\mathrm{GL}}}_n({{\mathbb {C}}})$ on ${\mathfrak}{sl}_n({\mathbb{C}}).$ Then the global Langlands $L$-function $L(s, \pi, {{\mathrm{Ad}}})$ is entire.
In the case $n=2,$ Conjecture \[conj: adjoint GL3 is entire\] was proved by Gelbart-Jacquet [@GelbartJacquet], generalizing a method of Shimura [@Shimura]. A different proof was given in [@JZ]. Flicker [@Flicker] proved that under a certain assumption about ratios of Hecke $L$ series (see “$({\operatorname}{Ass}; E, \omega),$” [@Flicker p. 233]), Conjecture \[conj: adjoint GL3 is entire\] holds if $\pi\cong \otimes_v \pi_v$ with at least one $\pi_v$ supercuspidal. More precisely, he proved that for such $\pi,$ the twisted adjoint $L$-function $L(s, \pi, {{\mathrm{Ad}}}\times \chi)$ is entire unless $\chi$ is nontrivial and $\pi \cong \pi \otimes \chi.$ This proof was based on a trace formula. Our approach is based on an integral representation, which was pioneered in [@G], and a method of ruling out poles which was pioneered in [@GJ].
Note that the action of ${{\mathrm{GL}}}_n$ on the space of $n\times n$ matrices may be regarded as the tensor product of the standard representation with its own dual. It decomposes as the direct sum of our adjoint representation, and the one dimensional span of the identity matrix equipped with trivial action. It follows that $$\label{Ad = RS/zeta}
L( s, \pi, {{\mathrm{Ad}}}\times \chi) = \frac{L( s, \pi \times {\widetilde}\pi\times \chi )}{L(s, \chi)},$$ where ${\widetilde}\pi$ is the contragredient of $\pi.$ From this perspective, Conjecture \[conj: adjoint GL3 is entire\] may be viewed as saying that $L(s, \pi \times {\widetilde}\pi\times \chi)$ should be “evenly divisible” by $L(s, \chi).$ More concretely, it says that $L(s, \pi \times {\widetilde}\pi)$ must vanish at each zero of $\zeta(s),$ to at least the same order. (Cf. [@Flicker p. 234].)
As explained in [@JngR], Conjecture \[conj: adjoint GL3 is entire\] is expected to be related to the conjecture that $\zeta_K(s)/\zeta_F(s)$ is entire for a field extension $K/F$. As explained in the introduction of [@BumpGinzburg], since $\zeta_K(s)/\zeta_F(s)$ is a product of Artin $L$-functions, the latter conjecture is a consequence of the Artin conjecture. This relationship is one display in Flicker’s conditional result. For more details about this relationship, see [@JngR].
In the case when the global field is ${\mathbb{Q}},$ Conjecture \[conj: adjoint GL3 is entire\] may also be viewed from the point of view of the Selberg class. We recall that the Selberg class is a class of meromorphic functions ${\mathbb{C}}\to {\mathbb{C}}$ introduced by Selberg in [@Selberg]. For $\pi$ a cuspidal automorphic representation of ${{\mathrm{GL}}}_n({\mathbb{A}}),$ the finite $L$-function $L_f (s, \pi \times {\widetilde}\pi)$ will be an element of the Selberg class, unless $\pi$ is a counterexample to the Ramanujan conjecture. It then would follow from the conjectures in [@Selberg] that $L_f(s, \pi \times {\widetilde}\pi)$ must be the product of the Riemann zeta function and another element of the Selberg class (cf. remark ii at the bottom of p. 370 of [@Selberg]). In other words it would follow that $L_f(s, \pi, {{\mathrm{Ad}}})$ is itself an element of the Selberg class. But then it must be entire since elements of the Selberg class have no poles except possibly at one.
Main results
------------
For $n>2$ the adjoint $L$-function of ${{\mathrm{GL}}}_n$ is not accessible via the Langlands-Shahidi method. Some information can be obtained through the relationship with the better-understood Rankin-Selberg $L$-function. In particular, one can get a global functional equation of the adjoint $L$-function $L(s,\pi,{{\mathrm{Ad}}}\times \chi)$ by the global functional equations of $L(s,\pi\times {\widetilde}\pi\times \chi)$ and $L(s,\chi)$. To obtain further information of the adjoint $L$-function in the case $n=3$, our main tool here is an integral representation pioneered in [@G] together with a method of ruling out poles pioneered in [@GJ].
Ginzburg’s local zeta integral is based on an embedding of ${{\mathrm{SL}}}_3$ into a split exceptional group of type $G_2.$ The argument was adapted in [@H12] to apply to special quasisplit unitary group ${{\mathrm{SU}}}_{2,1}$ (which depends on a fixed quadratic extension of $F$), which also embed into $G_2.$ Let $H$ be one of these groups, regarded as a subgroup of $G_2.$ Then Ginzburg’s global zeta integral is given by $$Z(\varphi, f_s)=\int_{H(F) {\backslash}H({\mathbb{A}})} \varphi(g) E(g,f_s) \, dg,$$ where $\varphi$ is a cuspidal automorphic form on ${{\mathrm{GL}}}_3({{\mathbb {A}}})$ or ${{\mathrm {U}}}_{2,1}({{\mathbb {A}}})$ in a fixed irreducible cuspidal automorphic representation $\pi$ and $E(g, f_s)$ is an Eisenstein series on $G_2$. See [@G; @H18] or $\S$\[sec: split main theorem\] for more details. In the split case (when $H={{\mathrm{SL}}}_3$), among other things, Ginzburg proved that the above local zeta integral is Eulerian $$Z(\varphi,f_s)=\prod_v Z^*(W_v,f_{s,v}),$$ with $$Z^*(W_v,f_{s,v})=L(3s,\chi_v)L(6s-2,\chi_v^2)L(9s-3,\chi_v^3)\int_{N_2(F_v)\setminus H(F_v)}W_v(g)f_s(w_\beta g)dg,$$ where $\chi=\otimes_v\chi_v$ is a character of $F^\times\setminus {{\mathbb {A}}}^\times$ which is part of the datum of the Eisenstein series, $W_v$ is the Whittaker function associated with local component of $\varphi$, $N_2$ is certain subgroup of $H={{\mathrm{SL}}}_3$, and $w_\beta$ is the Weyl element associated with the long simple root $\beta$ of $G_2$. Ginzburg also showed that at unramified places, the local zeta integral $Z^*(W_v,f_{s,v})$ represents the adjoint $L$-function $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$.
It is worth noting that both $Z(\varphi, f_s)$ depends only on the restriction of $\varphi$ to ${{\mathrm{SL}}}_3,$ and hence that one could, in principle begin with an automorphic representation of ${{\mathrm{SL}}}_3({\mathbb{A}})$ instead of ${{\mathrm{GL}}}_3({\mathbb{A}}).$ However, the unfolding argument in [@G] results in a particular Whittaker integral of $\varphi.$ Thus, one must restrict attention to representations which are globally generic with respect to this particular character. Now, each cuspidal automorphic representation of ${{\mathrm{SL}}}_3({\mathbb{A}})$ may be obtained as one of the components in the restriction of some cuspidal automorphic representation of ${{\mathrm{GL}}}_3({\mathbb{A}})$ (unique up to twist). As explained in [@BumpGinzburg p.120], the phenomenon that the integral $Z(\varphi,f_s)$ only depends on its restriction to ${{\mathrm{SL}}}_3({\mathbb{A}})$ is a reflection of the fact that the the adjoint representation of ${{\mathrm{GL}}}_3({{\mathbb {C}}})$ factors through ${{\mathrm {P}}}{{\mathrm{GL}}}_3({{\mathbb {C}}})$, the $L$-group of ${{\mathrm{SL}}}_3$. Similarly, the local zeta integral $Z^*(W_v,f_{s,v})$ also only depends on $ W_v|_{{{\mathrm{SL}}}_3(F_v)}$.
To obtain the holomorphy of $L(s,\pi,{{\mathrm{Ad}}}\times \chi)$, on one hand, one needs to analyze the properties of the global integral $Z(\varphi,f_s)$; on the other hand, one needs to analyze the local zeta integral $Z^*(W_v,f_{s,v})$ at every ramified place $v$. Because of the functional equation, it suffices to rule out poles in the half plane ${{\mathrm{Re}}}(s) \ge \frac 12.$ After the pioneering work of Ginzburg [@G] and Ginzburg-Jiang [@GJ], some progress in this direction was obtained in [@H18]. The goal of our first main result in this paper is to extend [@H18 Theorem 6.1], which treats the partial $L$-function $L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)=\prod_{v\notin S}L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ attached to a finite set $S=S(\pi,\chi)$ of places of $F$. Here $S(\pi,\chi)$ contains the set of all infinite places $S_\infty$ and all of the ramified places of $\pi$ and $\chi$. The main result in [@H18] states that for $\chi =1$ or $\chi$ unitary and $\pi \not \cong \pi \times \chi,$ the function $L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)$ has no poles on the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2}.$ However, during the referee process for this paper, a gap in the proof of the main theorem in [@H18] was pointed out. In this paper, we first explain how to close this gap, and then extend [@H18 Theorem 6.1] to the finite part $L$-function $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi)=L^{S_\infty}(s, \pi, {{\mathrm{Ad}}}\times \chi)$:
\[thm: main for split case\] Let $\chi$ be a unitary Hecke character of $F^\times\setminus {{\mathbb {A}}}^\times$ and $\pi$ be an irreducible cuspidal automorphic representation of ${{\mathrm{GL}}}_3({{\mathbb {A}}})$, then the finite part of the adjoint $L$-function $L_f(s,\pi, {{\mathrm{Ad}}}\times \chi)$ is holomorphic on the region ${{\mathrm{Re}}}(s)\ge 1/2$, except for a simple pole at ${{\mathrm{Re}}}(s)=1$ when $\chi$ is non-trivial and $\pi \cong \pi \otimes \chi$ $($which forces $\chi$ to be cubic$)$.
As mentioned above, to prove Theorem \[thm: main for split case\], we need to analyze Ginzburg’s local zeta integral at places $v\in S-S_\infty$, and compare the local zeta integral with the corresponding local L-function. More precisely, we need to show that for each place $v\in S-S_\infty$, there exists a “test" Whittaker function $W_v$ of $\pi_v$ and a “test" section $f_{s,v}\in I(s,\chi_v)$ such that the local zeta integral $Z^*(W_v,f_{s,v})$ “detects" all poles of $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ on the region ${{\mathrm{Re}}}(s)\ge 1/2$, i.e., the quotient $Z^*(W_v,f_{s,v})/L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ has no zeros on the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2}$. The main obstacle here is that, at this time, we can not rule out the existence of a place $v\in S-S_\infty$ such that $\chi_v$ is unramified $\pi_v$ is both ramified and non-tempered. In this case, $\pi_v$ is of the form ${{\mathrm{Ind}}}_{B_3}^{{{\mathrm{GL}}}_3(F_v)}(||^{\alpha}\otimes \mu_2\otimes ||^{-\alpha})$, where $B_3$ is the upper triangular Borel subgroup of ${{\mathrm{GL}}}_3(F_v)$, $\alpha$ is a real number with $0<\alpha<1/2 $, and $\mu_2$ is a ramified character of $F_v^\times$. It turns out that the required “test" Whittaker function comes from a certain new form of $\pi_v$. More precisely, let $\psi_v$ be an unramified additive character, and let $W_c\in {{\mathcal {W}}}(\pi_v,\psi_v)$ be the Whittaker function associated with a new vector of $\pi_v$ with respect to the group $$K_c=\begin{pmatrix}{{\mathfrak{o}}}& {{\mathfrak{o}}}& {{\mathfrak{p}}}^{-c} \\ {{\mathfrak{p}}}^c & 1+{{\mathfrak{p}}}^c & {{\mathfrak{o}}}\\ {{\mathfrak{p}}}^c & {{\mathfrak{p}}}^c &{{\mathfrak{o}}}\end{pmatrix}^\times,$$ where ${{\mathfrak{o}}}$ is the ring of integers of $F_v$, ${{\mathfrak{p}}}$ is the maximal ideal of ${{\mathfrak{o}}}$ and $c$ is the conductor of $\pi_v$ in the sense of [@JPSS81]. Here a new vector of $\pi_v$ with respect to $K_c$ is a nonzero vector in the space $\pi_v^{K_c}$. It is known that $W_c$ is unique (up to scalar). The main local ingredient to attack Theorem \[thm: main for split case\] is the following Whittaker function formula for $W_c$:
\[thm: whittaker function formula\] We have $W_c(1)\ne 0$ and $$W_c({{\mathrm{diag}}}(\varpi^m,1,\varpi^{-m}))=\frac{q^{-2m}}{t_1-t_1^{-1}}(t_1^{2m+1}-t_1^{-(2m+1)})W_c(1),$$ where $\varpi$ is a fixed uniformizer of $F_v$, $q$ is the number of the residue field of $F_v$ and $t_1=|\varpi|^\alpha$.
In fact, Theorem \[thm1.3\] is slightly general than the above Theorem \[thm: whittaker function formula\]. After Theorem \[thm: whittaker function formula\], a careful choice of test section $f_{s,v}\in I(s,\chi_v)$ (See §\[computation of the local zeta integral in a special case\] for the details) will give the desired property, i.e., $Z^*(W_c,f_{s,v})/L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ has no zeros in the region ${{\mathrm{Re}}}(s)\ge 1/2$.
[After [@H18], Buttcane and Zhou [@BuZh Theorem 2.4] were able to show that the adjoint $L$-function of a Maass form for ${{\mathrm{SL}}}_3({\mathbb{Z}})$ must be entire. Their argument relies on a simple comparison between where the Gamma factor can have poles and where the Riemann zeta function can have zeros. Our result gives an extension of theirs to Maass forms for congruence subgroups of ${{\mathrm{SL}}}_3({\mathbb{Z}}).$]{}
[Theorem \[thm: main for split case\] shows that $L_f(s,\pi,{{\mathrm{Ad}}})$ has no poles in the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2}$. To prove Conjecture \[conj: adjoint GL3 is entire\] in the case $n=3$, one still needs to analyze Ginzburg’s local zeta integral $Z(W_v,f_{s,v})$ for any archimedean place $v$. In this direction, recently F. Tian is able to show the meromorphic continuation, local functional equation at archimedean places in [@Ti18]. Moreover, the local gamma factors for principal series representations at archimedean places is explicitly computed in [@Tianthesis; @Ti18c]. ]{}
The second part of this paper is devoted to the nonsplit case, i.e., when $H={{\mathrm{SU}}}_{2,1}$ is a quasisplit unitary group attached to a quadratic extension $E$ of $F$, and $\pi$ is a globally generic, irreducible cuspidal automorphic representation of ${{\mathrm {U}}}_{2,1}({{\mathbb {A}}}).$ In this case, we have not performed a careful analysis on local zeta integrals at ramified primes, because there is still a lot to say about the simpler question of when the partial $L$ function attached to all finite unramified places will have a pole. As mentioned, the adaptation of Ginzburg’s integral representation to apply in nonsplit case was carried out in [@H12]. The adaptation of Ginzburg-Jiang’s method of ruling out poles was carried out in [@H18], with interesting results.
The key to Ginzburg and Jiang’s result is a “first term identity” relating residues of Eisenstein series attached to different parabolics. Using such an identity, they showed that if the global integral from [@G] has a double pole, then a global integral involving an Eisenstein series on the other maximal parabolic subgroup of $G_2$ must be nonzero. They then showed that this global integral unfolds to a period integral which vanishes on every cuspidal representation of ${{\mathrm{GL}}}_3({{\mathbb {A}}}).$ When this argument is adapted to the nonsplit case the resulting period, which is on ${{\mathrm{SU}}}_{1,1},$ does [*not*]{} vanish on every cuspidal representation of ${{\mathrm {U}}}_{2,1}({{\mathbb {A}}}).$ Rather, it fails to vanish, by earlier work of Gelbart, Rogawski, and Soudry [@GeRoSo2], precisely on the image of an endoscopic lift constructed in [@RogawskiBook]. By carefully combining this information with information about the relationship between Rogawski’s liftings and the stable base change lift of Kim and Krishnamurthy [@KimKrishOdd; @KimKrishEven], as well as results about the image of that lift obtained by Ginzburg, Rallis and Soudry by the method of functorial descent [@GRSBook], we obtain the following main result.
Take $\pi$ a globally generic irreducible cuspidal automorphic representation of the quasisplit unitary group ${{\mathrm {U}}}_{2,1}({\mathbb{A}})$ attached to a quadratic extension $E$ of $F.$ Let $\chi_{E/F}$ be the quadratic character attached to $E/F$ by class field theory. Then $L^S(s, \pi, {{\mathrm{Ad}}})$ is holomorphic and nonvanishing at $s=1,$ while $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi_{E/F})$ can have at most a double pole. More precisely, we have three sets of equivalent conditions.
1. The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi_{E/F})$ is holomorphic and nonvanishing at $s=1$
2. the stable base change lift, ${{\operatorname}{sbc}}(\pi),$ of $\pi$ is cuspidal
3. $\pi$ is not endoscopic
4. $\pi$ has a nonzero ${{\mathrm{SU}}}_{1,1}$ period
2. The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi_{E/F})$ has a simple pole at $s=1$
2. ${{\operatorname}{sbc}}(\pi)$ is the isobaric sum of a character and a cuspidal representation of ${{\mathrm{GL}}}_2({\mathbb{A}}_E)$
3. $\pi$ is an endoscopic lift from ${{\mathrm {U}}}_{1,1}({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}}),$ but not from ${{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}).$
3. The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi_{E/F})$ has a double pole at $s=1$
2. ${{\operatorname}{sbc}}(\pi)$ is the isobaric sum of three characters of ${\mathbb{A}}_E^\times,$
3. $\pi$ is an endoscopic lift from ${{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}).$
The twist by $\chi_{E/F}$ is of special importance for at least two reasons. First, in the nonsplit case the local zeta integral $Z^*(W_v, f_{v,s})$ is equal to $L(s, \pi_v, {{\mathrm{Ad}}}\times \chi_{E/F,v}\chi_v)$ as opposed to $L(s, \pi_v, {{\mathrm{Ad}}}\times \chi_v).$ (See [@H12].) Second, the analogue of in the nonsplit case relates $L(s, \pi , {{\mathrm{Ad}}}\times \chi)$ to the Asai $L$-function $L(s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}})$ of the stable base change lift ${{\operatorname}{sbc}}(\pi)$ ([@KimKrishOdd; @KimKrishEven]) of $\pi.$ Recall that ${{\operatorname}{sbc}}(\pi)$ is an automorphic representation of ${{\mathrm{GL}}}_3({\mathbb{A}}_E).$ The twist by $\chi_{E/F}$ is of particular importance because for a cuspidal representation $\Pi$ of ${{\mathrm{GL}}}_n({\mathbb{A}}_E),$ the $L(s, \Pi, {{\operatorname}{Asai}}) L(s, \Pi, {{\operatorname}{Asai}}\times \chi_{E/F})$ is equal to the Rankin-Selberg convolution $L$-function of $\Pi$ with the automorphic representation of ${{\mathrm{GL}}}_3({\mathbb{A}}_E)$ obtained by composing $\Pi$ with the nontrivial element of ${{\mathrm{Gal}}}(E/F).$ (cf [@GRSBook p.317]). We also obtain results which characterize when an adjoint $L$ function twisted by a character other than $\chi_{E/F}$ can have a pole. See theorems \[thm:other poles; type3\], \[thm:other poles: type 2+1\] and proposition \[prop: other poles, type 1+1+1\].
Because it relies on an embedding of ${{\mathrm{SL}}}_3$ into $G_2$, Ginzburg’s integral representation does not generalize to ${{\mathrm{GL}}}_n$ for general $n.$ However, it was shown in [@BumpGinzburg] that a certain global integral in the split exceptional group $F_4$ represents the adjoint $L$-function of ${{\mathrm{GL}}}_4,$ and some evidence was presented in [@GH3] that a certain global integral in the exceptional group $E_8$ represents the adjoint $L$-function for ${{\mathrm{GL}}}_5.$ In both cases, it appears possible to adapt the construction to apply to quasisplit unitary groups as well. It is expected similar results could be done for adjoint representations of ${{\mathrm{GL}}}_4$ and ${{\mathrm{GL}}}_5$ by studying these integral representations.
Acknowledgements
----------------
The first named author was supported by NSA grants H98230-15-1-0234 and H98230-16-1-0125. The second named author was supported by NSFC grant 11801577 and a fellowship from the Pacific Institute for the Mathematical Sciences (PIMS). We thank Jack Buttcane, Jim Cogdell, Clifton Cunningham, Sol Friedberg, Paul Garrett, Dorian Goldfeld, Dihua Jiang, Muthukrishnan Krishnamurthy, Xiaoqing Li, Baiying Liu, David Loeffler, Stephen Miller, Michitaka Miyauchi, Ralf Schmidt, Richard Taylor, and Fan Zhou for stimulating questions, helpful suggestions, and useful discussions. We thank Fangyang Tian for sending us his unpublished preprints [@Tianthesis; @Ti18c] and for his comments on our Lemma \[lemnonvanishing\] of a previous draft. We are very grateful to the anonymous referees for their careful reading and many valuable suggestions.
Local zeta integral for the adjoint representation of Lg and Lg
===============================================================
In this section, we review Ginzburg’s local zeta integral [@G] for the adjoint representation of ${{\mathrm{GL}}}_3$ and prove the local functional equation for these local zeta integrals. Ginzburg’s construction was extended to the ${{\mathrm {U}}}_{2,1}$ case in [@H12]. Let $F$ be a local field in this section. If $F$ is a non-archimedean local field, let ${{\mathfrak{o}}}$ be the ring of integers of $F$ and let $q$ be the number of the residue field of $F$. We also fix a uniformizer $\varpi$ of $F$ when $F$ is non-archimedean.
The local zeta integrals defined by Ginzburg involve the unique split exceptional group of type $G_2$. We denote this group $G_2$ and realize it as a set of $8 \times 8$ matrices as in [@H12]. Let $B=TU$ be the upper triangular Borel subgroup of $G_2$ under the realization in [@H12] with torus $T$ and maximal unipotent subgroup $U$.
The group $G_2$ has two simple roots $\alpha,\beta$, where $\alpha$ is the short root and $\beta$ is the long root. Then the set of positive roots of $G_2$ is ${\left\{{\alpha,\beta,\alpha+\beta,2\alpha+\beta,3\alpha+\beta,3\alpha+2\beta}\right\}}$. We denote the set of all roots by $\Phi$ and the set of positive roots by $\Phi^+.$ For a root $\delta$, we denote $U_\delta$ the corresponding root space of $\delta$ and ${{\bf {x}}}_\delta: {{\mathbb {G}}}_a{\rightarrow}U_\delta$ a choice of isomorphism. We let $X_\delta = d{{\bf {x}}}_\delta(1).$ (Here, $d{{\bf {x}}}_\delta$ is the differential.) We assume that $X_\alpha = E_{12}+E_{34}+E_{35}-E_{46}-E_{56}-E_{78},$ $X_{\beta}=E_{23}-E_{67},$ and that the family $\{ {{\bf {x}}}_\delta: \delta \in \Phi\}$ is chosen as in [@G], where $E_{ij}\in \textrm{Mat}_{8\times 8}(F)$ is the matrix such that its $i$-th row and $j$-th column is 1 and zero elsewhere. (See also [@Ree]). For a root $\delta$, denote $w_\delta(t)={{\bf {x}}}_{\delta}(t){{\bf {x}}}_{-\delta}(-t^{-1}){{\bf {x}}}_\delta(t)$ and $w_\delta=w_\delta(1)$. Let $h_\delta(t)=w_\delta(t)w_\delta^{-1}$. For $a,b\in F^\times$, we denote $h(a,b)=h_\alpha(ab)h_\beta(a^2b)$. Then $T(F)={\left\{{h(a,b): a,b\in F^\times}\right\}},$ and $U=\prod_{\delta>0}U_\delta.$
We briefly recall some facts about $G_2$ and the particular realization of it given in [@H12]. For more details see [@H12] and references therein. The group $G_2$ can be realized as the fixed points of an order three automorphism of ${\operatorname}{Spin}_8.$ The embedding into eight dimensional matrices in [@H12] is obtained by embedding into ${\operatorname}{Spin}_8$ and then projecting to ${{\mathrm{SO}}}_8.$ This projection is actually injective on $G_2$. Thus a symmetric bilinear form is preserved. For the realization in [@H12] it is the form attached to the matrix $J$ with ones on the diagonal running from top right to lower left, and zeros elsewhere. The standard representation of ${{\mathrm{SO}}}_8$ does not restrict to an irreducible representation of $G_2.$ It decomposes as a one dimensional space on which $G_2$ acts trivially and a copy of the seven dimensional “standard” representation of $G_2.$ In [@H12], the invariant one dimensional space is spanned by $v_0:={}^t\!{\begin{bmatrix}}0&0&0&1&-1&0&0&0{\end{bmatrix}}.$ The seven dimensional “standard” representation of $G_2$ supports an invariant quadratic form $Q.$ (In [@H12], it’s obtained by restricting the form induced by $J$ to the orthogonal complement of $v_0$.) The vectors satisfying $Q(v)=c$ form a single $G_2(F)$-orbit for each $c \in F^\times.$ The stabilizer of such a vector is isomorphic to ${{\mathrm{SL}}}_3,$ either over $F,$ or over a quadratic extension depending on $c.$ For example, the stabilizer $H_\rho$ of $v_\rho:={}^t\! {\begin{bmatrix}}0&0&1&0&0&\rho&0&0{\end{bmatrix}}$ is isomorphic to ${{\mathrm{SL}}}_3$ over the smallest extension of $F$ in which $\rho$ is a square. Indeed, if $\rho = \tau^2$ then $h_\beta(\frac 1{2\tau}){{\bf {x}}}_\alpha(\frac{1}{2\tau}){{\bf {x}}}_{-\alpha}(-\tau).v_\rho = {}^t\!{\begin{bmatrix}}0&0&0&-\tau &-\tau&0&0&0
{\end{bmatrix}}.$ (In fact, ${{\bf {x}}}_\alpha(\frac{1}{2\tau}){{\bf {x}}}_{-\alpha}(-\tau).v_\rho = {}^t\!{\begin{bmatrix}}0&0&0&-\tau &-\tau&0&0&0{\end{bmatrix}},$ and then acting by $h_\beta(\frac1{2\tau})$ has no effect. The reason for including $h_\beta(\frac1{2\tau})$ will be clear in a moment.) The stabilizer of this vector in ${\mathfrak}g_2$ is the image of the embedding $i:{\mathfrak}{sl}_3\to {\mathfrak}g_2$ given by $$i{\begin{pmatrix}}a&b&c\\d&e&f\\ g&h&-a-e {\end{pmatrix}}=
{\begin{pmatrix}}a+e &&&&&-f&-c&\\
&a&b&&&&&c\\
&d&e&&&&&f\\
&&&0&&&&\\
&&&&0&&&\\
-h&&&&&-e&-b&\\
-g&&&&&-d&-a&\\
&g&h&&&&&-a-e\\
{\end{pmatrix}}.$$ A general element of ${\mathfrak}h_\rho$ is given by $${\begin{pmatrix}}T_1&a&-\rho e&d&d&e&f&0\\
\rho a&T_1&-\rho d&\rho e&\rho e&d&0&-f\\
h&l&0&a&a&0&-d&-e\\
-\rho l&-h&\rho a&0&0&-a&-\rho e&-d\\
-\rho l&-h&\rho a&0&0&-a&-\rho e&-d\\
-\rho h&-\rho l&0&-\rho a&-\rho a&0&\rho d&\rho e\\
k&0&\rho l&h&h&-l&-T_1&-a\\
0&-k&\rho h&\rho l&\rho l&-h&-\rho a&-T_1{\end{pmatrix}}.$$ Conjugating by $h_\beta(\frac1{2\tau}){{\bf {x}}}_\alpha(\frac{1}{2\tau}){{\bf {x}}}_{-\alpha}(-\tau)$ yields the image under the injection $i$ of the matrix $${\begin{pmatrix}}-a \tau + T_{1} & e \tau - d & -\frac{f}{2 \, \tau } \\
2 \, l \rho - 2 \, h \tau & 2 \, a \tau & -e \tau - d \\
-2 \, k \tau & 2 \, l \rho + 2 \, h \tau & -a \tau - T_{1}
{\end{pmatrix}},
\label{cm}$$(this is where $h_\beta(\frac1{2\tau})$ is needed) which is a general element of ${{\mathrm{SL}}}_3(F)$ if $\tau \in F.$ If $\tau \notin F,$ then is a general element of $$\left\{X \in {\mathfrak}{sl}_3(F(\tau)): X{\left(\begin{smallmatrix}}&&-1\\&1&\\-1&& {\end{smallmatrix}\right)}+ {\left(\begin{smallmatrix}}&&-1\\&1&\\-1&& {\end{smallmatrix}\right)}{}^t{\overline}{X} = 0\right \},$$ where $\overline{\phantom{F}}$ denotes the action of the nontrivial element of ${{\mathrm{Gal}}}(F(\tau)/F).$ This is the Lie algebra of a quasisplit special unitary group attached to the matrix ${\left(\begin{smallmatrix}}&&-1\\&1&\\-1&& {\end{smallmatrix}\right)}.$ In [@H12], the group $\rm U_{2,1}$ is defined (relative to a choice of quadratic extension) using the matrix ${\left(\begin{smallmatrix}}&&1\\&1&\\1&& {\end{smallmatrix}\right)},$ i.e., $${{\mathrm {U}}}_{2,1}={\left\{{g\in {{\mathrm{GL}}}_3(F(\tau)): {}^t\!\bar g{\left(\begin{smallmatrix}}&&1\\&1&\\1&& {\end{smallmatrix}\right)}g={\left(\begin{smallmatrix}}&&1\\&1&\\1&& {\end{smallmatrix}\right)}. }\right\}}$$ For consistency with [@H12] we compose conjugation by ${{\mathrm{diag}}}(1,1,-1)$ in ${{\mathrm{GL}}}_3$ with $i$ followed conjugation by $(h_\beta(\frac1{2\tau}){{\bf {x}}}_\alpha(\frac{1}{2\tau}){{\bf {x}}}_{-\alpha}(-\tau))^{-1}$ in $G_2$ to obtain an injection ${\rm SU}_{2,1} {\hookrightarrow}G_2$ in the case $\tau \notin F.$ In the case $\tau \in F$ we obtain an injection ${{\mathrm{SL}}}_3 {\hookrightarrow}G_2,$ which is slightly different from the one used in [@G].
Let $P=MU^\alpha$ be the parabolic subgroup of $G_2$ such that $U_\alpha$ is contained in the Levi $M\cong {{\mathrm{GL}}}_2$ of $P$. Here $U^\alpha$ is the unipotent radical of $P$, which is the product of the root subgroups of $\beta,\alpha+\beta,2\alpha+\beta,3\alpha+\beta,3\alpha+2\beta$. Then $B_\rho:=P \cap H_\rho$ is a Borel subgroup of $H_\rho.$ We denote its unipotent radical $U_\rho$ and its maximal torus $T_\rho.$
Induced representations
-----------------------
Let $N_{2,\rho}= U_\rho \cap w_\beta P w_\beta^{-1}.$ Then the image of $N_{2,\rho}(F)$ in ${{\mathrm{GL}}}_3(F(\tau))$ is $${\left\{{\begin{pmatrix}
1&r\tau&t\tau +\frac{r^2\rho}2\\
&1&r\tau \\&&1\end{pmatrix},r,t\in F}\right\}}.$$ In the case $\tau \in F$, this simplifies to $ {\left\{{\begin{pmatrix}
1&r&t\\
&1&r\\&&1\end{pmatrix},r,t\in F}\right\}}.$ This is a little different from [@G], where $N_2$ is $ {\left\{{\begin{pmatrix}
1&r&t\\&1&-r\\&&1\end{pmatrix},r,t\in F}\right\}}.$ The reason for the difference is the extra conjugation by ${{\mathrm{diag}}}(1,1,-1).$
The Levi subgroup $M$ is generated by elements ${{\bf {x}}}_\alpha(r), {{\bf {x}}}_{-\alpha}(r), h(a,b)$. Note that $h_\alpha(r)=h(1/r,r^2).$ We consider the isomorphism $M{\rightarrow}{{\mathrm{GL}}}_2(F)$ defined by $${{\bf {x}}}_\alpha(r)\mapsto \begin{pmatrix}1& r\\ &1 \end{pmatrix},$$ $$h(a,b)\mapsto \begin{pmatrix}ab &\\ & a \end{pmatrix}.$$ For $m\in M$, we define $\det(m)$ using the isomorphism $M\cong {{\mathrm{GL}}}_2$.
Let $\delta_P$ be the modulus character of $P$. One can check that $\delta_P(m)=|\det(m)|^3$ for $m\in M$. Let $\chi$ be a character of $F^\times$, we consider the normalized induced representation \[defn of I(s,chi)\] $$I(s,\chi)={{\mathrm{Ind}}}_{P}^{G_2}(\chi_{s-1/2}),$$ where $\chi_s$ is the character of $M$ defined by $\chi(\det(m))\delta_P^s(m)$. Note that $$\chi_s(h(a,b))=\chi(a^2b)|a^2b|^{3s}.$$
Let $\tilde w=w_\beta w_\alpha w_\beta w_\alpha w_\beta$. Then $\tilde w$ represents the unique Weyl element such that $\tilde w(\alpha)>0$ but $\tilde w(U^\alpha)$ is in the opposite of $U$. Consider the standard intertwining operator $$M_{\tilde w}: I(s,\chi){\rightarrow}I(1-s,\chi^{-1})$$ which is defined by $$M_{\tilde w}(f_s)(g)=\int_{U^\alpha}f_s( \tilde wug)du.$$ By the general theory of intertwining operators, $M_{\tilde w}$ is absolutely convergent for ${{\mathrm{Re}}}(s)\gg0$ and can be meromorphically continued to all $s\in {{\mathbb {C}}}$.
An exact sequence for induced representations of Lg
---------------------------------------------------
For each $\rho \in F^\times,$ we have the double coset decomposition $G_2=Pw_\beta H_\rho\cup P H_\rho$, see [@H12 Lemma 3]. The subgroup $T_0:= {\left\{{h(a,1): a\in {{\mathrm{GL}}}_1}\right\}}$ of $T$ is contained in $H_\rho$ for all $\rho$ and is the maximal $F$-split subtorus of $T_\rho$ when $\rho$ is not a square. We also have $$H_\rho\cap w_\beta P w_\beta^{-1}=N_{2,\rho}\cdot T_0,$$ for all $\rho.$ See [@H12 pp. 198-199]. Moreover we have the relation $$\label{eq2.1}
w_\beta h(a,1)=h(1,a)w_\beta.$$ On the other hand, recall that $H_\rho\cap P$ is the Borel subgroup $B_\rho$ of $H_\rho$. Mackey’s theory gives us an exact sequence of $H_\rho$-modules: $$\label{eq2.2}0{\rightarrow}{{\mathrm{ind}}}_{N_{2,\rho}\cdot T_0}^{H_\rho}(\chi'_s){\rightarrow}I(s,\chi){\rightarrow}\textrm{n-}{{\mathrm{Ind}}}_{B_\rho}^{H_\rho}(\chi_{s}){\rightarrow}0,$$ where ${{\mathrm{ind}}}$ means compact induction, $\textrm{n-}{{\mathrm{Ind}}}$ means non-normalized induction, $\chi'_s$ is the character on $N_{2,\rho}\cdot T_0$ defined by $$\chi_s'(nh(a,1))=\chi_s(h(1,a))=\chi(a)|a|^{3s},\qquad (n \in N_{2, \rho}, \ a \in {{\mathrm{GL}}}_1)$$ and $\chi_s$ is viewed as character on $B_\rho$ by restriction, since $B_\rho \subset P.$ Here the embedding $${{\mathrm{ind}}}_{N_{2,\rho} \cdot T_0}^{H_\rho}(\chi'_s) {\hookrightarrow}I(s,\chi)$$ is defined as follows. For $f_s\in {{\mathrm{ind}}}_{N_{2,\rho} \cdot T_0}^{H_\rho}(\chi'_s)$, then the corresponding $\tilde f_s\in I(s,\chi)$ is defined by \[def of tilde f s\] $$\tilde f_s( mu w_\beta g)=\chi_s(m) f_s(g),$$ where $m\in M, u\in U^\alpha, g\in H_\rho$, and $$\tilde f_s(h)=0, \textrm{ if } h\notin Pw_\beta H_\rho.$$ By , $\tilde f_s$ is well-defined.
The local zeta integrals
------------------------
For $\rho$ in $F^\times$ fix $\tau$ with $\tau^2=\rho,$ and let ${\widetilde}H_\rho$ be ${{\mathrm{GL}}}_3$ when $\rho$ is a square and the quasi-split unitary group ${\rm U}_{2,1}$ otherwise. Let $\pi$ be an irreducible generic representation of ${\widetilde}H_\rho(F)$ and $\chi$ be a quasi-character of $F^\times$. Let $\psi$ be a nontrivial additive character of $F$.
The maximal unipotent subgroup $U_\rho$ is given by $$\left\{\begin{pmatrix} 1&x+y\tau&\frac{(x^2-y^2\rho)}2+w\tau\\&1&-x+y\tau\\ &&1\end{pmatrix}x,y,z\in F\right\}.$$ If $\tau\in F$, this is simply a parametrization of the standard maximal unipotent of ${{\mathrm{SL}}}_3$. Let $\psi_\rho:U_\rho(F) \to {\mathbb{C}}^\times$ be the character given by $$\label{psirho}
\psi_\rho\begin{pmatrix} 1&x+y\tau&\frac{(x^2-y^2\rho)}2+w\tau\\&1&-x+y\tau\\ &&1\end{pmatrix}
= \psi(x).$$ Note that $\psi_\rho|_{N_{2,\rho}}=1$. Given $W\in {{\mathcal {W}}}(\pi,\psi),f_s\in I(s,\chi)$, the local zeta integral of [@G],[@H12] is $$Z(W,f_s)=\int_{N_{2,\rho}\setminus H_\rho(F)}W(g)f_s(w_\beta g)dg.$$ Observe that if $\tilde f_s\in {{\mathrm{Ind}}}_P^{G_2}(\chi_s)$ is given in terms of $f_s\in {{\mathrm{ind}}}_{N_{2,\rho} \cdot T_0}^{H_\rho}(\chi_s)$ as in section \[def of tilde f s\] then $$Z(W,\tilde f_s)=\int_{N_{2,\rho}\setminus H_\rho(F)}W(g)f_s( g)dg.$$
Theorem 5.1 of [@H18] states that for each $s_0$, there exists a choice of data $W,f_s$ (depending on $s_0$) such that $Z(W,f_s)$ is holomorphic and nonvanishing at $s_0$. However, there is a gap in the proof of this theorem in the archimedean case, because it is never shown that $f_{s_0} \mapsto Z(W, f_{s_0})$ is a continuous function of $f_{s_0} \in I(s_0, \chi),$ when $s_0$ is outside the domain of convergence. In the split case, it is shown in [@Tianthesis; @Ti18] that $Z$ is, in fact, continuous in both of its arguments. The technique, which was pioneered in [@Soudry-Archimedean], should extend to the non-split case as well. In fact, the non-split case is easier, because the rank of the maximal split torus is only one. Here, we content ourselves with sketching how to close the gap in [@H18].
For any fixed $s_0 \in {\mathbb{C}}$ and $W$ in the Whittaker model of some irreducible unitary representation of ${{\mathrm{SU}}}_{2,1}({\mathbb{R}}),$ the function $f_{s_0} \mapsto Z(W, f_{s_0})$ is a continuous function $I(s_0, \chi) \to {\mathbb{C}}.$
Factoring the Haar measure on ${{\mathrm{SU}}}_{2,1}({\mathbb{R}})$ using the Iwasawa decomposition, we write the local zeta integral as $$\int_K \int_0^\infty\int_{-\infty}^\infty
f_s(w_\beta {{\bf {x}}}_\beta(r)) W(d(t)) \psi(tr) t^{3s-3}
\, dr\, dt\, dk,$$ where $d(t)= {\left(\begin{smallmatrix}}t&\\&1&\\&&t^{-1}{\end{smallmatrix}\right)}\in {{\mathrm{SU}}}_{2,1}({\mathbb{R}}).$ Arguing as in [@Soudry-Archimedean Lemma 1, p.197], or [@Ti18 Lemma 4.1], we show that continuity of $Z$ follows from continuity of the inner integral $$\int_0^\infty W(t) t^{3s-3} J f_s(t)\, dt,
\qquad J f_s(t) : = \int_{-\infty}^\infty f_s(w_\beta {{\bf {x}}}_\beta(r)) \psi(tr)\, dr.$$ Arguing as in [@JS-ExteriorSquare], we may write $$W(t) = \sum_{\xi \in X} \varphi_\xi(t) \xi(t)$$ where $X$ is a finite set of finite functions on the maximal split torus, and, for each $\xi \in X,$ $\varphi_\xi$ is a Schwartz function on ${\mathbb{R}}.$ Recall that a finite function on the multiplicative group of positive reals is of the form $t\mapsto t^u (\log t)^n$ for some complex number $u$ and non-negative integer $n.$ Hence it suffices to prove continuity of the mapping $$\label{mapping}
f \mapsto \int_0^{\infty} Jf_s(t) \varphi(t) t^u (\log t)^n \, dt,$$ for any complex number $u,$ nonnegative integer $n,$ and Schwartz function $\varphi.$ Since $Jf_s$ is just the Jacquet integral of an embedded ${{\mathrm{SL}}}_2$ we have an asymptotic expansion as in [@Wallach-RRG2 15.2]. This can be formulated as follows. Let $(z_n)_{n=1}^\infty$ be the sequence of complex numbers obtained by numbering the elements of $\{ 3s_0+2k: k = 0,1,2, \dots\} \cup \{ 1-3s_0+2k: k = 0,1,2 \dots\}$ in increasing order of real part. Then there is a sequence $(a_k)_{k=0}^\infty$ of continuous linear functionals $I(s_0, \chi)\to {\mathbb{C}}$ and a sequence $(A_k)_{k=1}^\infty$ of continuous functions $I(s_0, \chi) \to (0, \infty)$ such that for any non-negative integer $N$, $$\left|
Jf_{s_0}(t) - \sum_{k=0}^N a_k(f_{s_0}) t^{z_k}
\right|\le A_{N+1}(f_{s_0}) t^{z_{N+1}},~ \forall t < 1.$$ Let $Z(\varphi, u, n)$ be defined by $$Z(\varphi, u, n)=\int_0^\infty \varphi(t) t^u (\log t)^n \, dt.$$ for $u > 0$ and by meromorphic continuation elsewhere. (Notice that $Z(\varphi u, 0)$ is a standard Tate zeta factor, while $Z(\varphi, u, n)$ is its $n$th derivative.) Then $$\int_0^{\infty} Jf_s(t) \varphi(t) t^u (\log t)^n \, dt
- \sum_{k=0}^n a_k(f_{s_0})
Z( \varphi, u+z_k, n)
+E(f_{s_0}),$$ where $$|E(f_{s_0})| \le A_{N+1}(f_{s_0})
\int_0^\infty |\varphi(t)| t^{{{\mathrm{Re}}}(u+z_{N+1})} ( \log t)^n \, dt,$$ provided we choose $N$ sufficiently large to ensure that this last integral is convergent. Since each of the functions $a_k, k=0, \dots, N$ and $A_{N+1}$ tends to zero with $f_{s_0},$ it is now clear that is continuous.
If $F$ is non-archimedean, we can prove a stronger form of the non-vanishing result.
\[lemnonvanishing\] Let $F$ be non-archimedean. The local zeta integral $Z(W,f_s)$ is absolutely convergent for ${{\mathrm{Re}}}(s)\gg 0$ and can be meromorphically continued to a rational function of $q^s$. Moreover, there exist choices of data $W,f_s$ such that $Z(W,f_s)$ is a nonzero constant.
The meromorphic continuation of the local zeta integral $Z(W,f_s)$ at the archimedean places in the split case is proved in [@Ti18].
The first statement follows from the Bruhat decomposition and the asymptotic behavior of $W$ on the torus element. We next consider the “moreover" part. To simplify the notation, we only deal with the split case, i.e., when $H_\rho\cong {{\mathrm{SL}}}_3$. Let $K_3^m=(1+\textrm{Mat}_{3\times 3}({{\mathfrak{p}}}^m))\cap {{\mathrm{SL}}}_3(F)$ be the standard level $m$ congruence subgroup of ${{\mathrm{SL}}}_3$. Recall that $B_3$ is the upper triangular Borel subgroup of ${{\mathrm{GL}}}_3(F)$.
In this case $T_\rho \cong (F^\times)^2.$ For each $a \in F^\times,$ the group $T_\rho$ contains the element $h(a,1),$ which is identified with ${{\mathrm{diag}}}(a,1,a^{-1})$ in ${{\mathrm{SL}}}_3(F).$ Note that $h(a,b)\in T$ is identified with ${{\mathrm{diag}}}(a,b,a^{-1}b^{-1})\in {{\mathrm{SL}}}_3(F)$ Consider the following function on ${{\mathrm{SL}}}_3(F)$: $$f_{s}^m(g)=\left\{\begin{array}{lll}0,& \textrm{ if } g\notin (B_3\cap {{\mathrm{SL}}}_3(F))K_3^m\\ \chi_s'(nh(a,1))\phi_1(b)\phi_2(x), & \textrm{ if } g=nh(a,1)n(x)h(1,b) k, n\in N_2, k\in K_3^m, \end{array}\right.$$ where $$n(x)=\begin{pmatrix}1& x& 0\\ &1&0\\ &&1 \end{pmatrix}\in {{\mathrm{SL}}}_3(F), \phi_1\in {{\mathcal {S}}}(F^\times),\phi_2\in {{\mathcal {S}}}(F).$$ To make $f_s^m$ well-defined, we need to require that $\phi_1,\chi_s'$ are constant on $K_3^m\cap T_0$, and $\phi_2$ is invariant under the translation of $\varpi^m{{\mathfrak{o}}}_F$. We have $f_s^m\in {{\mathrm{ind}}}_{N_{2,\rho}\cdot T_0}^{H_\rho}\chi_s'$. We now compute $Z(W,\tilde f_s^m).$ We assume $m$ is large enough such that $W(gk)=W(g)$ for $k\in K_3^m$. Then factoring the Haar measure on $H_\rho$ using the Iwasawa decomposition yields $$\begin{aligned}
Z(W, \tilde f_s^m)&=c_1\int_{(F^*)^2\times F} W(h(a,1) n(x) h(1,b))f_s^m( h(a,1)n(x)h(1,b) )\, dx\,\frac{ d^*a\, d^*b}{|a^3b^2|}\\
&=c_1\int_{(F^*)^2\times F} W(n(ax) h(a,b)) \chi_s(h(1,a))\phi_1(b)\phi_2(x)\, dx\,\frac{ d^*a\, d^*b}{|a^3b^2|}\\
&=c_1\int_{(F^*)^2} W(h(a,b)) \check \phi_2(a)\phi_1(b)\chi_s(h(1,a))\, \,\frac{ d^*a\, d^*b}{|a^3b^2|},\end{aligned}$$ where $c_1={{\mathrm{Vol}}}(B_\rho({{\mathfrak{o}}})) K_3^m$ and $\check \phi_2(a)=\int_F \psi(ax)\phi_2(x)dx$.
Assume $m$ is sufficiently large that $a\mapsto \chi_s(h(1,a))$ is trivial on $1+{{\mathfrak{p}}}^m,$ and choose $\phi_1, \phi_2$ such that $\phi_1=\check \phi_2$ is the characteristic function of $1+{{\mathfrak{p}}}^m$, we get $Z(W, \tilde f_s^m) =c_1{{\mathrm{vol}}}(1+{{\mathfrak{p}}}^m)^2W(1).$ Clearly, this is constant and $W$ may be chosen so that it is nonzero. This concludes the proof.
### Dependence on $\psi$
In this section we discuss the dependence of the local Ginzburg zeta integral on the choice of additive character $\psi.$ Let ${\widehat}F$ be the Pontriagin dual of $F$ . Then $\widehat F$ is isomorphic to $F,$ but not canonically: indeed if $\psi$ is any fixed nontrivial element of ${\widehat}F,$ then every other element of ${\widehat}F$ is of the form $\psi^a(x):=\psi(ax)$ for some $a \in F.$ The formula gives an injection $\widehat F {\hookrightarrow}\widehat{U_\rho(F)}.$ If $\pi$ is an irreducible $\psi_\rho$-generic representation of ${\widetilde}H_\rho(F),$ write ${{\mathcal {W}}}(\pi,\psi_{\rho})$ for the Whittaker model of $\pi.$
Fix $a, \rho \in F^\times.$ Then $$(\psi^a)_\rho(u) = \psi_\rho(h(a,1)uh(a,1)^{-1}), \qquad \forall u \in U_\rho(F).$$
Direct computation.
If $\pi$ is $\psi_\rho$ generic then it is $(\psi^a)_\rho$-generic for every $a,$ and left translation by $h(a,1)$ is an isomorphism ${{\mathcal {W}}}(\pi,\psi_{\rho}) \to {{\mathcal {W}}}(\pi,(\psi^a)_{\rho}).$
\[prop:dependence on psi\] If $f_s \in I(s, \chi), \ W \in {{\mathcal {W}}}(\pi, \psi_\rho),$ and $a \in F^\times,$ then there exists $W' \in {{\mathcal {W}}}(\pi, (\psi^a)_\rho)$ such that $Z(W', f_s) = \chi_s(h(1,a))^{-1}Z(W, f_s).$
Indeed $W'(g) := W(h(a,1)g)$ is an element of ${{\mathcal {W}}}(\pi, (\psi^a)_\rho).$ A change of variable in the integral defining $Z(W, f_s),$ together with the identity $w_\beta h(a,1) = h(1,a) w_\beta$ shows that $Z(W', f_s) = \chi_s(h(1,a))^{-1}Z(W, f_s).$
The local functional equation
-----------------------------
In this subsection, we assume that $F$ is a non-archimedean local field. Let $\pi$ be an irreducible generic representation of ${\widetilde}H_\rho$. It is known that $\pi|_{H_\rho}$ has finite length, see [@GeK].
\[prelfe\] Except for a finite number of $q^{-s}$, we have $${{\mathrm{Hom}}}_{H_\rho}( \textrm{n-}{{\mathrm{Ind}}}_{B_\rho}^{H_\rho}(\chi_{s}),\pi)=0.$$
By the Frobenius reciprocity law, we have $${{\mathrm{Hom}}}_{H_\rho}(\textrm{n-}{{\mathrm{Ind}}}_{B_\rho}^{H_\rho}(\chi_{s}),\pi)={{\mathrm{Hom}}}_{T_\rho}( \chi_s,\pi_{U_\rho}).$$ Since $U_\rho$ is the maximal unipotent subgroup of the upper triangular Borel subgroup of both $H_\rho$ and ${\widetilde}H_\rho$, we have $\dim\pi_{U_\rho}<\infty$. The assertion follows.
\[prop: multiplicity one\] Excluding a finite number of $q^{-s}$, we have $$\dim {{\mathrm{Hom}}}_{H_\rho}
(I(s,\chi),\pi )\le 1.$$
By the exact sequence (\[eq2.2\]) and Lemma \[prelfe\], it suffices to show that $$\dim{{\mathrm{Hom}}}_{H_\rho}({{\mathrm{ind}}}_{N_{2,\rho}\cdot T_0}^{H_\rho}(\chi'_s), \pi)\le 1$$ except for a finite number of $q^{-s}$.
By Frobenius reciprocity law [@BZ Proposition 2.29], we have $$\begin{aligned}
&{{\mathrm{Hom}}}_{H_\rho}({{\mathrm{ind}}}_{N_{2,\rho}\cdot T_0}^{H_\rho}(\chi'_s), \pi)\\
=&{{\mathrm{Hom}}}_{N_{2,\rho}\cdot T_0}(\chi'_s,\pi)\\
=&{{\mathrm{Hom}}}_{T_0}(\chi'_s, \pi_{N_{2,\rho}}),\end{aligned}$$ where $\pi_{N_{2,\rho}}$ is the Jacquet module. The Jacquet module $\pi_{N_{2,\rho}}$ can be viewed as a representation of $T_0\cdot U_\rho$. Since $N_{2,\rho}$ acts trivially on $\pi_{N_{2,\rho}}$, we know that $\pi_{N_{2,\rho}}$ can be viewed a representation of $T_0\ltimes U_\rho/N_{2,\rho}\cong {{\mathrm{GL}}}_1\ltimes F$, where the action of ${{\mathrm{GL}}}_1\cong F^\times$ on $ F$ is given by multiplication. As a representation of $F\cong U_\rho/N_{2,\rho}$, $\pi_{N_{2,\rho}}$ is smooth. Denote $\sigma=\pi_{N_{2,\rho}}$ and $V_\sigma$ the space of $\sigma$. Thus we have ${{\mathcal {S}}}(F). V_\sigma=V_\sigma$. From the isomorphism induced by the Fourier transform ${{\mathcal {S}}}(\hat F)\cong {{\mathcal {S}}}(F)$, we get $V_\sigma={{\mathcal {S}}}(\hat F).V_\sigma$, i.e., $\sigma$ is smooth as a ${{\mathcal {S}}}(\hat F)$-module. Thus by [@BZ Proposition 1.14] there exists a unique sheaf ${{\mathcal {V}}}$ on ${{\mathcal {S}}}(\hat F)$ such that ${{\mathcal {V}}}_c\cong V_\sigma$, where ${{\mathcal {V}}}_c$ denotes the compact support sections in ${{\mathcal {V}}}$.
The action of $F^\times$ on $\hat F$ has two orbits. Let $\psi$ be a nontrivial additive character of $F$, and let $O={\left\{{\psi_a, a\in F^{\times}}\right\}}$. Then $O$ is the open orbit of the action of $F^\times$ on $\hat F$, and its complement is the trivial character on $F$. We have the usual short exact sequence $$0{\rightarrow}{{\mathcal {V}}}_c(O){\rightarrow}{{\mathcal {V}}}_c {\rightarrow}{{\mathcal {V}}}_c(0){\rightarrow}0,$$ where $0$ denotes the zero character. Consider the element $\psi$ in $O$, which may be identified with the character $\psi_\rho$ of $U_\rho$ given in . The stalk of the sheaf ${{\mathcal {V}}}$ at the point $\psi$ is given by $$(\pi_{N_{2,\rho}})_{U_\rho/N_{2,\rho},\psi}=\pi_{U_\rho,\psi}\cong {{\mathbb {C}}}_\psi,$$ since $\pi$ is an irreducible representation of ${\widetilde}H_\rho$ and $\psi$ is a generic character of the maximal unipotent subgroup $U_\rho$ of a Borel subgroup in ${\widetilde}H_\rho$. The stabilizer of ${{\mathcal {V}}}$ of the point $\psi$ in ${{\mathrm{GL}}}_1$ is ${\left\{{1}\right\}}$. Thus by [@BZ Proposition 2.23], we get $${{\mathcal {V}}}_c(O)={{\mathrm{ind}}}_{1}^{{{\mathrm{GL}}}_1}({{\mathbb {C}}}_\psi).$$ Similarly, we have ${{\mathcal {V}}}_c(0)=\pi_{U_\rho}$ which has finite dimension. Thus we get the short exact sequence $$0{\rightarrow}{{\mathrm{ind}}}_{1}^{{{\mathrm{GL}}}_1}({{\mathbb {C}}}_\psi){\rightarrow}\pi_{N_{2,\rho}}{\rightarrow}\pi_{U_\rho}{\rightarrow}0.$$ Since $\pi_{U_\rho}$ has finite dimension, after excluding a finite number of $q^s$, we have $${{\mathrm{Hom}}}_{T_0}(\chi'_s, \pi_{N_{2,\rho}}) ={{\mathrm{Hom}}}_{{{\mathrm{GL}}}_1}(\chi'_s, {{\mathrm{ind}}}_{1}^{{{\mathrm{GL}}}_1}({{\mathbb {C}}}_\psi))={{\mathrm{Hom}}}(\chi'_s, {{\mathbb {C}}}_\psi).$$ Since $\chi'_s$ and ${{\mathbb {C}}}_\psi$ have dimension 1, we get $${{\mathrm{Hom}}}(\chi'_s, {{\mathbb {C}}}_\psi)$$ has dimension 1.
\[corlfe\] There exists a rational function $\gamma(s,\pi, \chi,\psi)$ of $q^s$ such that $$Z(W, M_{\tilde w}(f_s))=\gamma(s,\pi,\chi,\psi)Z(W,f_s),$$ for all $W\in {{\mathcal {W}}}(\pi,\psi), f_s\in I(s,\chi)$.
Since both $(W,f_s)\mapsto Z(W,f_s)$ and $(W,f_s)\mapsto Z(W,M_{\tilde w}(f_s))$ define a bilinear form on ${{\mathrm{Hom}}}_{H_\rho}(I(s,\chi)\otimes \pi,1)$. By Proposition \[prop: multiplicity one\], such bilinear form is unique up to a scalar. Thus there is factor $\gamma(s,\pi,\chi,\psi)$ such that $Z(W,M_{{\widetilde}w}(f_s))=\gamma(s,\pi,\chi,\psi)Z(W,f_s)$ for all $W\in {{\mathcal {W}}}(\pi,\psi),f_s\in I(s,\chi)$. Let $ W\in {{\mathcal {W}}}(\pi,\psi),\tilde f_s^m\in I(s,\chi)$ be as in the proof of Lemma \[lemnonvanishing\]. Then $Z(W,\tilde f_s^m)$ is a non-zero constant, say $c$, by Lemma \[lemnonvanishing\]. Thus we have $\gamma(s,\pi,\chi,\psi)=c^{-1}Z(W,M_{\tilde w}(\tilde f_s^m))$, which is a rational function of $q^s$ by Lemma \[lemnonvanishing\] again.
Based on the relationship between the $L$-functions, it is reasonable to define the local gamma factors.
\[defn: local gamma factors\] When $\rho$ is a square $($so $H_\rho \cong {{\mathrm{SL}}}_3)$, let the twisted local adjoint gamma factor be given by $$\gamma(s,\pi,{{\mathrm{Ad}}}\times \chi,\psi)=\frac{\gamma(s,(\chi\pi)\times \tilde \pi,\psi)}{\gamma(s,\chi,\psi)},$$ where $\gamma(s,\chi\pi\otimes \tilde \pi,\psi)$ is the local Rankin-Selberg gamma factor defined by Jacquet-Piatetski-Shapiro-Shalika [@JPSS83] and $\gamma(s,\chi,\psi)$ is the local gamma factor of Tate. When $\rho$ is a nonsquare $($so $H_\rho$ is a unitary group$)$, let the twisted local ${{\mathrm{Ad}}}'$ gamma factor be given by $${\gamma}(s, \pi, {{\mathrm{Ad}}}' \times \chi,\psi) = \frac{{\gamma}(s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}\times \chi,\psi)}{{\gamma}(s, \chi, \psi)},$$ where the denominator is the Tate gamma factor and the numerator is defined by the Langlands-Shahidi method.
Here ${{\mathrm{Ad}}}'$ is defined as in [@H18]. Following [@H18] we define ${{\mathrm{Ad}}}'$ to be ${{\mathrm{Ad}}}$ in the split case. Then the unramified computations from [@H12] show that in both the split and nonsplit cases, $$\label{gamma factor formula}
{\gamma}(s, \pi, \chi, \psi)
=\frac{{\gamma}(3s-1, \pi,{{\mathrm{Ad}}}' \times \chi, \psi)}
{{\gamma}(3s-2, \chi, \psi), {\gamma}(6s-3, \chi^2, \psi){\gamma}(9s-5, \chi^3, \psi)},$$ where the individual gamma factors on the right hand side are defined as in [@JPSS83] or the Langlands-Shahidi method. (For ${{\mathrm{GL}}}_1$ factors in the denominator, either of these other definitions reduces to the one in Tate’s thesis.) It is natural to expect that the local gamma factors in Corollary \[corlfe\] is essentially the same as the local gamma factors defined in Definition \[defn: local gamma factors\], i.e., should be true for all irreducible generic representation $\pi$ of ${\widetilde}H_\rho$.
[Over archimedean local field, the local functional equation of Ginzburg’s local zeta integral is proved in [@Tianthesis; @Ti18; @Ti18c] recently. Moreover, it is verified in [@Tianthesis; @Ti18c] that the local gamma factors for principal series representation of ${{\mathrm{GL}}}_3({{\mathbb {R}}})$ obtained from the local functional equation satisfies .]{}
A Whittaker function formula for Lg
===================================
In this section, we develop a Whittaker function formula for certain ramified induced representation of ${{\mathrm{GL}}}_3$ over a $p$-adic field, see Theorem \[thm1.3\]. This Whittaker function formula will be used to compute the Ginzburg’s local zeta integral in a special case, see Proposition \[prop: special case\], which is the main ingredient in the proof of one of our main theorem, Theorem \[thm4.2\].
In this section, let $F$ be a non-archimedean local field, ${{\mathfrak{o}}}$ the ring of integers of $F$, ${{\mathfrak{p}}}$ the maximal ideal of ${{\mathfrak{o}}}$, and $\varpi$ a fixed generator of ${{\mathfrak{p}}}$. Let $q=|{{\mathfrak{o}}}/{{\mathfrak{p}}}|$. \[notion on F\]
Certain subgroups of Lg
-----------------------
Let $ B_3=T_3U_3$ be the upper triangular Borel subgroup of ${{\mathrm{GL}}}_3(F)$ with diagonal torus $T_3$ and upper triangular unipotent subgroup $U_3$. Let $K={{\mathrm{GL}}}_3({{\mathfrak{o}}})$. Denote $$w=\begin{pmatrix}&&1\\ &1&\\ 1&& \end{pmatrix},$$ and $${{\mathfrak{u}}}(x,y,z)=\begin{pmatrix}1&x&z\\ &1&y\\ &&1 \end{pmatrix}, \hat {{\mathfrak{u}}}(x,y,z)=\begin{pmatrix}1&&\\ x&1&\\z&y&1 \end{pmatrix}, \textrm{ for }x,y,z\in F.$$ For a nonnegative integer $n\ge 0$, we consider the subgroup $$K_n'=\begin{pmatrix} {{\mathfrak{o}}}&{{\mathfrak{o}}}&{{\mathfrak{o}}}\\ {{\mathfrak{o}}}& {{\mathfrak{o}}}&{{\mathfrak{o}}}\\ {{\mathfrak{p}}}^n&{{\mathfrak{p}}}^n&1+{{\mathfrak{p}}}^n\end{pmatrix}^\times$$ of ${{\mathrm{GL}}}_3(F)$. Here and in the following, for a subset $A\subset \textrm{Mat}_{3\times 3}(F)$ which is closed under multiplication, $A^\times$ is used to denote the subset $A^\times:={\left\{{a\in A: a^{-1} \textrm{ exists and }a^{-1}\in A}\right\}}$ of $A$. It’s clear that $A^\times$ is a group, whenever it is nonempty.
Given an irreducible smooth generic complex representation $(\pi,V)$ of ${{\mathrm{GL}}}_3(F)$, we consider the $K_n'$-fixed subspace $V'(n)=V^{K'_n}$ of $V$. Denote $c=c_\pi=\min\{n| V'(n)\ne 0\}$, which is called the conductor of $\pi$. By [@JPSS81], the space $V'(c)$ has dimension 1. A nonzero vector of $V'(c)$ is called a new form or new vector of $\pi$. Let $\psi$ be an unramified additive character of $F$. It is known that the epsilon factor $\epsilon(s,\pi,\psi)$ of $\pi$ has the form $C q^{-c_\pi s}$, where $C\in {{\mathbb {C}}}^\times$, see [@JPSS81 $\S$5]. It is worth to note that there was an error in [@JPSS81], which was fixed in [@J12] and [@Ma13].
We consider a variant of the above notions. Denote $$\epsilon_n=\begin{pmatrix}1&&\\ &&1\\&\varpi^n & \end{pmatrix}\in {{\mathrm{GL}}}_3(F),$$ and $K_n=\epsilon_n K_n'\epsilon_n^{-1}$. In matrix form, we have $$K_n=\begin{pmatrix}{{\mathfrak{o}}}& {{\mathfrak{o}}}& {{\mathfrak{p}}}^{-n} \\ {{\mathfrak{p}}}^n & 1+{{\mathfrak{p}}}^n & {{\mathfrak{o}}}\\ {{\mathfrak{p}}}^n & {{\mathfrak{p}}}^n &{{\mathfrak{o}}}\end{pmatrix}^\times.$$ Given an irreducible smooth generic complex representation $(\pi,V)$ of ${{\mathrm{GL}}}_3(F)$, denote $V(n)=V^{K_n}$, the subspace of $V$ which is fixed by $K_n$. Then $V(n)=\pi(\epsilon_n)V'(n)$. In particular we have $\dim V(c_\pi)=1.$
[Let $E/F$ be an unramified quadratic extension of $p$-adic fields and ${{\mathrm {U}}}_{2,1}$ be the unitary group with $3$ variables associated with $E/F$ realized by the matrix $w\in {{\mathrm{GL}}}_3$. Let ${{\mathfrak{o}}}_E$ be the ring of integers of $E$ and ${{\mathfrak{p}}}_E$ be the maximal ideal of ${{\mathfrak{o}}}_E$. In [@M13], Miyauchi developed a theory of local new forms for ${{\mathrm {U}}}_{2,1}$ using the group $${\begin{pmatrix}}{{\mathfrak{o}}}_E& {{\mathfrak{o}}}_E& {{\mathfrak{p}}}_E^{-n}\\ {{\mathfrak{p}}}_E^n & 1+{{\mathfrak{p}}}_E^n &{{\mathfrak{o}}}_E\\ {{\mathfrak{p}}}_E^n & {{\mathfrak{p}}}_E^n & {{\mathfrak{o}}}_E {\end{pmatrix}}^\times \cap {{\mathrm {U}}}_{2,1}.$$ The group $K_n$ we choose is inspired by the above group considered by Miyauchi.]{}
\[lem1.1\] Let $n\ge 1$.
1. Let $x,y\in F$. If $\hat {{\mathfrak{u}}}(x,0,0)\in B_3 K_n$, then $x\in {{\mathfrak{p}}}^n$. Similarly, if $\hat {{\mathfrak{u}}}(0,y,0)\in B_3\cdot K_n$, then $y\in {{\mathfrak{p}}}^n$.
2. Given $r\in F$, the element $w'(r):=\begin{pmatrix}&&1\\ 1&&\\r &1& \end{pmatrix}$ is not in $B_3\cdot K_n$.
\(1) If $\hat {{\mathfrak{u}}}(x,0,0)\in B_3 K_n$, then there exists an element $b\in B_3$ such that $b\hat {{\mathfrak{u}}}(x,0,0)\in K_n$. Write $$b=\begin{pmatrix}a_1&x_1&z_1\\ &a_2&y_1\\ &&a_3 \end{pmatrix}.$$ Then $$b\hat {{\mathfrak{u}}}(x,0,0)=\begin{pmatrix}a_1+x_1x& x_1&z_1\\ a_2x& a_2& y_1\\ &&a_3 \end{pmatrix}.$$ From the condition $b\hat {{\mathfrak{u}}}(x,0,0)\in K_n$, we get $a_2\in 1+{{\mathfrak{p}}}^n, a_2x\in {{\mathfrak{p}}}^n$. Thus $x\in {{\mathfrak{p}}}^n$. This proves the first statement. The second one can be proved similarly.
\(2) If $w'(r)\in B_3K_n$, there exists $b_3\in B_3$ as above such that $k=bw'(r)\in K_n$. After expanding the matrix of $k$, we get $a_3\in {{\mathfrak{p}}}^n.$ From the condition $k^{-1}=(w'(r))^{-1}b^{-1}\in K_n$, we can get $a_3^{-1}\in {{\mathfrak{o}}}$. This is a contradiction.
[Given $z\in F,$ $\hat {{\mathfrak{u}}}(0,0,z)\in B_3K_n$ does not imply $z\in {{\mathfrak{p}}}^n$.]{}
Induced representation and Whittaker functional {#notation I(mu)}
-----------------------------------------------
Let $\mu=(\mu_1,\mu_2,\mu_3)$ be a triple of quasi-characters of $F^\times$. We consider the normalized induced representation $$I(\mu)={{\mathrm{Ind}}}_{B_3}^{{{\mathrm{GL}}}_3}(\mu_1\otimes \mu_2\otimes \mu_3).$$ We assume that the representation $I(\mu)$ is irreducible. By [@BZ77 Theorem 4.2], the irreducibility of $I(\mu)$ is equivalent to that $\mu_i\mu_j^{-1}\ne |~|^{\pm}$, $1\le i,j\le 3$. It is known that the representation $I(\mu)$ is generic.
An element $f\in I(\mu)$ is a function $f:{{\mathrm{GL}}}_3(F){\rightarrow}{{\mathbb {C}}}$ such that $$f({{\mathrm{diag}}}(a_1,a_2,a_3)ug)=\mu_1(a_1)\mu_2(a_2)\mu_3(a_3)|a_1/a_3| f(g),$$ for all $ {{\mathrm{diag}}}(a_1,a_2,a_3)\in T_3, u\in U_3, g\in {{\mathrm{GL}}}_3(F).$
Let $U^1\subset U^2\subset\dots $ be a sequence of open compact subgroups of $U_3$ such that $\bigcup_{k=1}^\infty U^k=U_3$. For $f\in I(\mu)$, the sequence of integrals $$\int_{U^k} f(wu)\psi^{-1}(u)du, k\ge 1$$ is stable and its limit is independent on the choice of the sequence ${\left\{{U^k,k\ge 1}\right\}}$, see [@Sh Proposition 3.2] and [@CS Corollary 1.8]. Denote $$\int_{U_3}^{st}f(wu)\psi^{-1}(u)du:=\lim_{k{\rightarrow}\infty}\int_{U^k} f(wu)\psi^{-1}(u)du,$$ where “$st$" stands for stable integral. The linear map $f\mapsto \int_{U_3}^{st}f(wu)\psi^{-1}(u)du$ is a Whittaker functional on $I(\mu)$.
In the rest of this section, we assume that $\mu_1,\mu_3$ are unramified and $\mu_2$ is ramified with conductor $c\ge 1$. Let $t_i=\mu_i(\varpi),i=1,3$. By [@GoJ Theorem 3.4, page 36], we have $\epsilon(s,I(\mu),\psi)=\prod_{i=1}^3 \epsilon(s,\mu_i,\psi)$. Since $\mu_1,\mu_3$ are unramified, and $\mu_2$ has conductor $c$, we have $\epsilon(s,I(\mu),\psi)=Cq^{-cs}$ for some $C\in {{\mathbb {C}}}^\times$. Thus $c$ is also the conductor of the representation $I(\mu)$. Consequently, we have $\dim I(\mu)^{K_c}=1$.
We consider the following function $f$ on ${{\mathrm{GL}}}_3(F)$. We require that ${\mathrm{supp}}(f)\subset B_3K_c$ and $$f({{\mathrm{diag}}}(a_1,a_2,a_3)u k)=\mu_1(a_1)\mu_2(a_2)\mu_3(a_3)|a_1/a_3|, \forall a_1,a_2,a_3\in F^\times, u\in U_3, k\in K_c.$$ The function $f$ is well-defined and right $K_c$-invariant. Thus $f$ is a new form of $I(\mu)$, i.e., $f\in I(\mu)^{K_c}$.
In the following, we fix a nontrivial unramified character $\psi$ of $F$, and we consider the Whittaker function $W_f$ associated with the new form $f$: $$\label{eq1.1}W_f(g)=\int_{U_3}^{st} f(wug)\psi^{-1}(u)du.$$
\[lem: root killing\] Let $a_i\in F^\times$ with $a_i\in \varpi^{n_i}{{\mathfrak{o}}}^\times$ for $i=1,2,3$. If $W_f({{\mathrm{diag}}}(a_1,a_2,a_3))\ne 0$, then $n_1\ge n_2\ge n_3$.
For any $x\in {{\mathfrak{o}}}$, we have $${{\mathrm{diag}}}(a_1,a_2,a_3){{\mathfrak{u}}}(x,0,0)={{\mathfrak{u}}}(a_1a_2^{-1}x,0,0){{\mathrm{diag}}}(a_1,a_2,a_3).$$ Since ${{\mathfrak{u}}}(x,0,0)\in K_c$ and $W_f$ is right $K_c$-invariant, we have $$W_f({{\mathrm{diag}}}(a_1,a_2,a_3))=\psi(a_1a_2^{-1}x)W_f({{\mathrm{diag}}}(a_1,a_2,a_3)), \forall x\in {{\mathfrak{o}}}.$$ If $W_f({{\mathrm{diag}}}(a_1,a_2,a_3))\ne 0$, we must have $\psi(a_1a_2^{-1}x)=1$ for all $x\in {{\mathfrak{o}}}$. Since $\psi$ is assumed to be unramified, we have $a_1a_2^{-1}\in {{\mathfrak{o}}}$. This shows that $n_1\ge n_2$. Similarly, we can show that $n_2\ge n_3$.
\[thm1.3\]
1. We have $W_f(1)\ne 0$.
2. Let $a\in \varpi^m {{\mathfrak{o}}}^\times, b\in \varpi^{-n}{{\mathfrak{o}}}^\times$ with $m,n\ge 0$. We have $$W_f({{\mathrm{diag}}}(a,1,b))=\frac{q^{-(m+n)}}{t_1-t_3}(t_1t_3)^{-n}(t_1^{m+n+1}-t_3^{m+n+1})W_f(1).$$
[We consider the embedding of ${{\mathrm{GL}}}_2$ into ${{\mathrm{GL}}}_3$ by $${\begin{pmatrix}}a&b\\ c&d {\end{pmatrix}}\mapsto {\begin{pmatrix}}a&&b\\ &1&\\ c&&d {\end{pmatrix}}.$$ Let $\sigma$ be the unramified representation ${{\mathrm{Ind}}}_{B_2}^{{{\mathrm{GL}}}_2(F)}(\mu_1\otimes \mu_3)$ of ${{\mathrm{GL}}}_2(F)$, where $B_2$ is the upper triangular subgroup of ${{\mathrm{GL}}}_2(F)$. Let $W^\sigma$ be the unramified Whittaker function of $\sigma$ normalized by $W^\sigma(1)=1$. Then by Shintani’s formula for ${{\mathrm{GL}}}_2$ [@Shin], the above theorem says that $$W_f({{\mathrm{diag}}}(a,1,a^{-1}))=W^\sigma({{\mathrm{diag}}}(a,a^{-1}))W_f(1).$$ This property is what we need in the proof of the holomorphy of the finite part of the adjoint L-function for ${{\mathrm{GL}}}_3$ on the region ${{\mathrm{Re}}}(s)\ge 1/2$, see Proposition \[prop: special case\] and Theorem \[thm4.2\]. ]{}
[In [@M14], Miyauchi developed a formula of ramified Whittaker functions for ${{\mathrm{GL}}}_n$ (in particular for ${{\mathrm{GL}}}_3$), associated with local new forms defined by the group $K_n'$. But Miyauchi’s formula does not meet our purpose. It seems that there is no apparent relationship between our formula and Miyauchi’s formula, although $K_n$ is conjugate to $K_n'$. ]{}
We will prove Theorem \[thm1.3\] in the next section. The proof is by brute force computation. It would be interesting to know whether a proof which is more conceptual, and perhaps more amenable to generalization, can be found. As a preparation of the proof of the above theorem, we record the following useful lemma
\[lem1.4\] In the integrals appearing in the following, we fix the measure $dx$ on $F$ such that ${{\mathrm{vol}}}({{\mathfrak{o}}})=1$, and thus ${{\mathrm{vol}}}({{\mathfrak{o}}}^\times)={{\mathrm{vol}}}({{\mathfrak{o}}})-{{\mathrm{vol}}}({{\mathfrak{p}}})=1-q^{-1}$.
1. Let $k$ be an integer, then $$\nonumber
\int_{{{\mathfrak{p}}}^{k}-{{\mathfrak{p}}}^{k+1}}\psi^{-1}(x)dx=\left\{\begin{array}{lll}q^{-k}(1-q^{-1}), &k\ge 0 \\
-1, & k=-1\\ 0, & k\le -2. \end{array} \right.$$
2. Let $i$ be an integer. We have $$\nonumber
\int_{{{\mathfrak{o}}}^\times}\mu_2(1+\varpi^i x)dx=\left\{\begin{array}{lll}1-q^{-1} & i\ge c\\ -q^{-1} & i=c-1, \\ 0, & i<c-1. \end{array}\right.$$
The first statement follows from $$\int_{{{\mathfrak{p}}}^k}\psi^{-1}(x)dx=\left\{\begin{array}{lll}q^{-k},& k\ge 0, \\ 0, & k<0.\end{array} \right.$$ The second statement can be proved in the same manner if $i>0$. We consider the case $i\le 0$. If $i<0$, we have $$1+\varpi^ix=\varpi^i(x+\varpi^{-i}).$$ The set ${\left\{{x+\varpi^{-i}:x\in {{\mathfrak{o}}}^\times}\right\}}$ is exactly ${{\mathfrak{o}}}^\times$. Thus $\int_{{{\mathfrak{o}}}^\times}\mu_2(1+\varpi^ix)du=0$. If $i=0$, we have $$\int_{{{\mathfrak{o}}}^\times}\mu_2(1+x)dx=\int_{{{\mathfrak{o}}}}\mu_2(1+x)dx-\int_{{{\mathfrak{p}}}}\mu_2(1+x)dx=\int_{{{\mathfrak{o}}}}\mu_2(x)dx-\int_{{{\mathfrak{p}}}}\mu_2(1+x)dx.$$ Here $\int_{{{\mathfrak{o}}}}\mu_2(x)dx$ is understood as $\sum_{k\ge 0}\int_{\varpi^k{{\mathfrak{o}}}^\times}\mu_2(x)dx$, which is convergent to 0 since $\mu_2$ is ramified. Thus $\int_{{{\mathfrak{o}}}^\times}\mu_2(1+x)dx=-\int_{{{\mathfrak{p}}}}\mu_2(1+x)dx$, which is $-q^{-1}$ if $c=1$ and $0$ if $c>1$. The assertion follows.
Proof of Theorem \[thm1.3\]
---------------------------
We are going to compute $$W_f({{\mathrm{diag}}}(a,1,b))=\int_{U_3}^{st} f(wu{{\mathrm{diag}}}(a,1,b))\psi^{-1}(u)du,$$ with $|a|=q^{-m}, |b|=q^n$. We first fix the measure $du$. For $u={{\mathfrak{u}}}(x,y,z),x,y,z\in F$, we take $du=dxdydz$, where $dx,dy,dz$ are additive measures on $F$ such that ${{\mathrm{Vol}}}({{\mathfrak{o}}})=1$.
We can write the above integral as $$W_f({{\mathrm{diag}}}(a,1,b))=\int_{F^3}^{st} f(w{{\mathfrak{u}}}(x,y,z){{\mathrm{diag}}}(a,1,b))\psi^{-1}(x+y)dxdydz.$$ Since $$w{{\mathfrak{u}}}(x,y,z){{\mathrm{diag}}}(a,1,b)={{\mathrm{diag}}}( b,1,a)w{{\mathfrak{u}}}(a^{-1} x, by, a^{-1}bz),$$ we get $$\nonumber
W_f({{\mathrm{diag}}}(a,1,b))=\mu_1(b)\mu_3(a)|b/a|\int_{F^3}^{st} f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz))\psi^{-1}(x+y)dxdydz. \nonumber$$ Let $U^k={\left\{{{{\mathfrak{u}}}(x,y,z): x,y\in {{\mathfrak{p}}}^{-k}, z\in {{\mathfrak{p}}}^{-2k}}\right\}}$. Consider the integral $$\label{eq1.2}
I^k=I^k(m,n)=\int_{U^k} f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz))\psi^{-1}(x+y)dxdydz.$$ Since ${{\mathrm{diag}}}({{\mathfrak{o}}}^\times,1,{{\mathfrak{o}}}^\times)\subset K_c$, the right hand side of the above integral only depends on $m,n,k$, which justifies the notation $I^k(m,n)$. We then have $$\label{eq1.3}W_f({{\mathrm{diag}}}(a,1,b))=q^{n+m}t_1^{-n}t_3^m\lim_{k{\rightarrow}\infty} I^k.$$ To compute $I^k$, we will frequently use the following identity $$\begin{aligned}
\label{eq1.4}
w{{\mathfrak{u}}}(x,y,z)=&{{\mathrm{diag}}}(-1/(z-xy),1-(xy)/z ,z) \\
&\cdot {{\mathfrak{u}}}\left(\frac{x(z-xy)}{z}, \frac{zy}{z-xy},-z+xy\right) \hat {{\mathfrak{u}}}\left(-\frac{y}{z-xy}, \frac{x}{z},\frac{1}{z}\right), \nonumber\end{aligned}$$ if $z\ne 0, z-xy\ne 0$. Take $x=y=0,z=\varpi^{-c}$ in Eq.(\[eq1.4\]), we get $$w{{\mathfrak{u}}}(0,0,\varpi^{-c})={{\mathrm{diag}}}(-\varpi^c,1,\varpi^{-c}){{\mathfrak{u}}}(0,0,-\varpi^{-c})\hat {{\mathfrak{u}}}(0,0,\varpi^c).$$ Since ${{\mathfrak{u}}}(0,0,\varpi^{-c}),\hat {{\mathfrak{u}}}(0,0,\varpi^c)\in K_c$, we can get $$\label{eq1.5}f(w)=(t_1t_3^{-1}q^{-2})^c$$ from the above equation.
We now start to compute the integral $I^k$. In the following computation, $k$ is sufficiently large. We have $$\begin{aligned}
I^k&=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}} \int_{{{\mathfrak{p}}}^{-k}} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b(z-xy) {{\mathfrak{u}}}(0,yb,0))) \psi^{-1}(x+y)dy dxdz \label{eq: I^k}\\
&=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}} \int_{{{\mathfrak{p}}}^n} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b(z-xy) {{\mathfrak{u}}}(0,yb,0))) \psi^{-1}(x+y)dy dxdz \nonumber \\
&+\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}}\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^n}f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b(z-xy) {{\mathfrak{u}}}(0,yb,0))) \psi^{-1}(x+y)dy dxdz \nonumber \\
:&=I_1^k+I_2^k. \nonumber\end{aligned}$$ Since $yb\in {{\mathfrak{o}}},$ we have $ {{\mathfrak{u}}}(0,yb,0)\in K_c$, for $y\in{{\mathfrak{p}}}^{n}$. By changing variable on $z$, we have $$\begin{aligned}
I_1^k&=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}} \int_{{{\mathfrak{p}}}^n} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b(z-xy) {{\mathfrak{u}}}(0,by,0))) \psi^{-1}(x+y)dy dxdz\\
&=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}} \int_{{{\mathfrak{p}}}^n} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz ){{\mathfrak{u}}}(0,by,0)) \psi^{-1}(x+y)dy dxdz\\
&=q^{-n}\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b z))\psi^{-1}(x)dxdz\\
&=q^{-n}\int_{{{\mathfrak{p}}}^{-2k}} \left(\int_{{{\mathfrak{p}}}^m} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b z))dx\right)dz\\
&+q^{-n}\int_{{{\mathfrak{p}}}^{-2k}} \left(\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^m} f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b z))\psi^{-1}(x)dx\right)dz\\
:&=I_3^k+I_4^k.\end{aligned}$$ Since for $x\in {{\mathfrak{p}}}^m, a^{-1}x\in {{\mathfrak{o}}}$, we get $$\begin{aligned}
I_3^k&=q^{-m-n}\int_{{{\mathfrak{p}}}^{-2k}}f(w{{\mathfrak{u}}}(0,0,a^{-1}bz))dz\\
&=q^{-m-n}\left(\int_{{{\mathfrak{p}}}^{m+n-c}}f(w{{\mathfrak{u}}}(0,0,a^{-1}bz))dz \right.\\
&\left.+ \sum_{i= 1}^{2k+m+n-c}\int_{{{\mathfrak{p}}}^{m+n-c-i}-{{\mathfrak{p}}}^{m+n-c-i+1}}f(w{{\mathfrak{u}}}(0,0,a^{-1}bz))dz \right).\end{aligned}$$ Since $f$ is $K_c$-invariant, by Eq.(\[eq1.5\]), we have $$\int_{{{\mathfrak{p}}}^{m+n-c}}f(wu(0,0,a^{-1}bz))dz=q^{-m-n}(t_1t_3^{-1}q^{-1})^c.$$ Note that for $z\in {{\mathfrak{p}}}^{m+n-c-i}-{{\mathfrak{p}}}^{m+n-c-i+1}$, we have $ab^{-1}z^{-1}\in \varpi^{c+i}{{\mathfrak{o}}}^\times$. By Eq.(\[eq1.4\]), we have $$w{{\mathfrak{u}}}(0,0,a^{-1}bz)={{\mathrm{diag}}}(-ab^{-1}z^{-1},1,a^{-1}bz) {{\mathfrak{u}}}(0,0,-a^{-1}bz)\hat {{\mathfrak{u}}}(0,0,ab^{-1}z^{-1}).$$ For $i\ge 1$, we have $ \hat {{\mathfrak{u}}}(0,0,ab^{-1}z^{-1})\in K_c$ and $$f(w{{\mathfrak{u}}}(0,0,a^{-1}bz))=\mu_1(ab^{-1}z^{-1})\mu_3(a^{-1}bz^{-1})|ab^{-1}z^{-1}|^2=(t_1t_3^{-1}q^{-2})^{c+i}.$$ Thus $$\label{eq1.6}I_3^k=q^{-2(m+n)}\left((t_1t_3^{-1}q^{-1})^c+(1-q^{-1})\sum_{i=1}^{2k+m+n-c}(t_1t_3^{-1}q^{-1})^{c+i}\right).$$
We next consider $I_4^k.$ We have $$\begin{aligned}
I_4^k&=q^{-n}\int_{{{\mathfrak{p}}}^{m+n-c}}\left(\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{m}}f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz))\psi^{-1}(x)dx\right)dz\\
&+q^{-n}\int_{{{\mathfrak{p}}}^{-2k}-{{\mathfrak{p}}}^{m+n-c}}\left(\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{m}}f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz))\psi^{-1}(x)dx\right)dz\\
:&=I_5^k+I_6^k.\end{aligned}$$ For $z\in {{\mathfrak{p}}}^{m+n-c}$, we have $a^{-1}bz\in {{\mathfrak{p}}}^{-c}$ and thus ${{\mathfrak{u}}}(0,0,a^{-1}bz)\in K_c$. We get $$\begin{aligned}
I_5^k=q^{-m-2n+c} \int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{m}}f(w{{\mathfrak{u}}}(a^{-1}x,0,0))\psi^{-1}(x)dx.\end{aligned}$$ We have the identity $$w{{\mathfrak{u}}}(a^{-1}x,0,0)={{\mathrm{diag}}}(1,-ax^{-1},a^{-1}x){{\mathfrak{u}}}(0,-a^{-1}x,1)w'(ax^{-1}),$$ see Lemma \[lem1.1\](2) for the definition of $w'(r)$. By Lemma \[lem1.1\](2), we get $f(w{{\mathfrak{u}}}(a^{-1}x,0,0))=0$ and thus $I_5^k=0$.
We next consider $I_6^k$. We have $$\begin{aligned}
I_6^k&=q^{-n}\sum_{i=1}^{2k+m+n-c}\int_{{{\mathfrak{p}}}^{m+n-c-i}-{{\mathfrak{p}}}^{m+n-c-i+1}}\\
&\cdot \left(\sum_{j=1}^{k+m}\int_{{{\mathfrak{p}}}^{m-j}-{{\mathfrak{p}}}^{m-j+1}}f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz))\psi^{-1}(x)dx\right)dz\end{aligned}$$ By Eq.(\[eq1.4\]) again, we have $$w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz)={{\mathrm{diag}}}(-ab^{-1}z^{-1},1,a^{-1}bz){{\mathfrak{u}}}(a^{-1}x,0,-a^{-1}bz)\hat {{\mathfrak{u}}}(0,b^{-1}xz^{-1},ab^{-1}z^{-1} ).$$ For $z\in {{\mathfrak{p}}}^{m+n-c-i}-{{\mathfrak{p}}}^{m+n-c-i+1},x\in {{\mathfrak{p}}}^{m-j}-{{\mathfrak{p}}}^{m-j+1}$, we have $ab^{-1}z^{-1}\in \varpi^{c+i}{{\mathfrak{o}}}^\times$ and $b^{-1}xz^{-1}\in \varpi^{c+i-j}$. Since $i\ge 1$, we get $\hat {{\mathfrak{u}}}(0,0,ab^{-1}z^{-1})\in K_c$. Thus $$f( w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz))=(t_1t_3^{-1}q^{-2})^{c+i} f(\hat {{\mathfrak{u}}}(0,b^{-1}xz^{-1},0)).$$ By Lemma \[lem1.1\] (1), if $i<j$, we get $f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz) )=0.$ If $i\ge j\ge 1$, we have $$f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}bz) )=(t_1t_3^{-1}q^{-2})^{c+i}.$$ Thus, we get $$\begin{aligned}
I_6^k&=q^{-n}\sum_{i=1}^{2k+m+n-c}\int_{{{\mathfrak{p}}}^{m+n-c-i}-{{\mathfrak{p}}}^{m+n-c-i+1}}(t_1t_3^{-1}q^{-2})^{c+i}\\
&\cdot \left(\sum_{j=1}^{\min{\left\{{i,k+m}\right\}}}\int_{{{\mathfrak{p}}}^{m-j}-{{\mathfrak{p}}}^{m-j+1}}\psi^{-1}(x)dx\right)dz.\end{aligned}$$ By Lemma \[lem1.4\](1), we have $\int_{{{\mathfrak{p}}}^{m-j}-{{\mathfrak{p}}}^{m-j+1}}\psi^{-1}(x)=0 $ if $j>m+1$. Then in the above expression, $j$ satisfies the condition $1\le j\le \min {\left\{{i,m+1}\right\}}$. The double sum in the expression of $I_6^k$ then can be divided into 3 parts: $$\sum_{i=1}^m \sum_{j=1}^i, \quad \sum_{i=m+1}^{2k+m+n-c} \sum_{j=1}^m, \quad \sum_{i=m+1}^{2k+m+n-c}\sum_{j=m+1}^{m+1}.$$ Thus we get $$\begin{aligned}
I_4^k=I_6^k&=q^{-2m-2n}(1-q^{-1})^2\sum_{i=1}^m(t_1t_3^{-1}q^{-1})^{c+i}\sum_{j=1}^i q^j \label{eq1.7}\\
&+q^{-2m-2n}(1-q^{-1})^2\sum_{i=m+1}^{2k+m+n-c}(t_1t_3^{-1}q^{-1})^{c+i}\sum_{j=1}^m q^j \nonumber\\
&-q^{-m-2n}(1-q^{-1})\sum_{i=m+1}^{2k+m+n-c}(t_1t_3^{-1}q^{-1})^{c+i} \nonumber\\
&=q^{-2m-2n}(1-q^{-1})\sum_{i=1}^m (t_1t_3^{-1}q^{-1})^{c+i}(q^i-1) \nonumber\\
&-q^{-2m-2n}(1-q^{-1}) \sum_{i=m+1}^{2k+m+n-c}(t_1t_3^{-1}q^{-1})^{c+i}. \nonumber\end{aligned}$$ From the above calculation, Eq.(\[eq1.6\]) and Eq.(\[eq1.7\]), we get $$\begin{aligned}
\label{eq1.8}
I_1^k&=I_3^k+I_4^k \\
&=q^{-2m-2n}(1-q^{-1})\sum_{i=1}^m (t_1t_3^{-1}q^{-1})^{c+i}(q^i-1) \nonumber \\
&+q^{-2m-2n}(t_1t_3^{-1}q^{-1})^c \nonumber\\
&+q^{-2m-2n}(1-q^{-1}) \sum_{i=1}^{m}(t_1t_3^{-1}q^{-1})^{c+i} \nonumber\\
&=q^{-2m-2n}(t_1t_3q^{-1})^c+q^{-2m-2n}(1-q^{-1})\sum_{i=1}^m(t_1t_3^{-1}q^{-1})^{c+i}q^i, \nonumber\end{aligned}$$ which is independent of $k$ as long as $k$ is sufficiently large. To emphasize the dependence on $m,n$, we will write the above expression as $I_1(m,n)$.
We next consider $$I_2^k=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}}\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^n}f(w{{\mathfrak{u}}}(a^{-1}x,0,a^{-1}b(z-xy) ){{\mathfrak{u}}}(0,by,0)) \psi^{-1}(x+y)dy dxdz.$$ We have $$\begin{aligned}
I_2^k&=\int_{{{\mathfrak{p}}}^{-2k}\times {{\mathfrak{p}}}^{-k}}\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^n}f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz) ) \psi^{-1}(x+y)dy dxdz\\
&=\int_{{{\mathfrak{p}}}^{-2k}\times ({{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{n})}\int_{{{\mathfrak{p}}}^{-k}}f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz) ) \psi^{-1}(x+y)dx dydz\\
&=\int_{{{\mathfrak{p}}}^{-2k}\times ({{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{n})}\int_{{{\mathfrak{p}}}^m}f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz) ) \psi^{-1}(x+y)dx dydz\\
&+\int_{{{\mathfrak{p}}}^{-2k}\times ({{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{n})}\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^m}f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz) ) \psi^{-1}(x+y)dx dydz\\
:&=I_7^k+I_8^k.\end{aligned}$$ The calculation of $I_7^k$ is the same as that of $I_4^k$, Eq.(\[eq1.7\]). We only write the result here $$\begin{aligned}
I_{7}^k&=q^{-2m-2n}(1-q^{-1})\sum_{i=1}^n(t_1t_3^{-1}q^{-1})^{c+i}(q^i-1) \label{eq1.9}\\
&-q^{-2m-2n}(1-q^{-1})\sum_{i=n+1}^{2k+m+n-c} (t_1t_3^{-1}q^{-1})^{c+i}. \nonumber\end{aligned}$$
We next compute $$I_8^{k}=\int_{{{\mathfrak{p}}}^{-2k}\times ({{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^{n})}\int_{{{\mathfrak{p}}}^{-k}-{{\mathfrak{p}}}^m}f(w{{\mathfrak{u}}}(a^{-1}x,by,a^{-1}bz) ) \psi^{-1}(x+y)dx dydz.$$ Using Eq.(\[eq1.4\]) we get $$\begin{aligned}
I_8^k&=\sum_{l=-\infty}^{2k+m+n-c}\int_{{{\mathfrak{p}}}^{m+n-c-l}-{{\mathfrak{p}}}^{m+n-c-l+1}} \sum_{j=1}^{k+n}\int_{{{\mathfrak{p}}}^{n-j}-{{\mathfrak{p}}}^{n-j+1}} \sum_{i=1}^{k+m}\int_{{{\mathfrak{p}}}^{m-i}-{{\mathfrak{p}}}^{m-i+1}}\\
& \cdot \psi^{-1}(x+y)\cdot \mu_1(ab^{-1}(z-xy)^{-1})\mu_2(1-(xy)/z)\mu_3(a^{-1}bz)|ab^{-1}(z-xy)^{-1}|\cdot |a^{-1}bz|^{-1}\\
&\cdot f\left(\hat {{\mathfrak{u}}}\left(\frac{-ay}{z-xy},\frac{x}{bz},\frac{a}{bz} \right) \right)dxdydz.\end{aligned}$$ For $z\in {{\mathfrak{p}}}^{m+n-c-l}-{{\mathfrak{p}}}^{m+n-c-l+1}, y\in {{\mathfrak{p}}}^{n-j}-{{\mathfrak{p}}}^{n-j+1}, x\in {{\mathfrak{p}}}^{m-i}-{{\mathfrak{p}}}^{m-i+1}$, we have $$ab^{-1}z^{-1}\in \varpi^{c+l}{{\mathfrak{o}}}^\times, b^{-1}xz^{-1}\in \varpi^{c+l-i}, (xy)/z\in \varpi^{c+l-i-j}{{\mathfrak{o}}}^\times.$$ For fixed $x,y$, to make sure the integral of $\mu_2(1-(xy)/z)$ with respect to $z$ is non-zero, we need $l\ge i+j-1$, see Lemma \[lem1.4\] (2) (one can check that the other quantities in $I_8^k$ involving $z$ only depends on the absolute value of $z$). If $l\ge i+j-1$, we have $ay/(z-xy)\in {{\mathfrak{p}}}^c, ab^{-1}z^{-1}\in {{\mathfrak{p}}}^c, b^{-1}xz^{-1}\in {{\mathfrak{p}}}^c$. Moreover, we have $$ab^{-1}(z-xy)^{-1}\in \varpi^{c+l}{{\mathfrak{o}}}^\times.$$ Thus we get $$\begin{aligned}
\label{eq1.10}
I_8^k
&= \sum_{j=1}^{k+n}\int_{{{\mathfrak{p}}}^{n-j}-{{\mathfrak{p}}}^{n-j+1}}\psi^{-1}(y)dy \sum_{i=1}^{k+m}\int_{{{\mathfrak{p}}}^{m-i}-{{\mathfrak{p}}}^{m-i+1}}\psi^{-1}(x)dx\\
&\cdot \left( \sum_{l=i+j}^{2k+m+n-c}q^{-(m+n)} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}q^{-m-n}(t_1t_3^{-1}q^{-1})^{c+i+j-1} \right). \nonumber\end{aligned}$$ Similar to the calculation of $I_6^k$, the right side of the double sum on $i,j$ in Eq.(\[eq1.10\]) can be divided into the following 4 parts $$\sum_{i=1}^m\sum_{j=1}^n, \quad \sum_{i=1}^m \sum_{j=n+1}, \quad \sum_{i=m+1}\sum_{j=1}^n, \quad \sum_{i=m+1}\sum_{j=n+1}.$$ The term corresponding to $\sum_{i=1}^{m}\sum_{j=1}^n$ is $$\begin{aligned}
\label{eq1.11}
&q^{-2m-2n}(1-q^{-1})^2\sum_{i=1}^{m}\sum_{j=1}^n q^{i+j}\left( \sum_{l=i+j}^{2k+m+n-c} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+j-1} \right).\end{aligned}$$ The second term is $$\begin{aligned}
\label{eq1.12}
-q^{-2m-2n}(1-q^{-1})\sum_{i=1}^mq^{i+n}\left( \sum_{l=i+n+1}^{2k+m+n-c} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+n} \right).\end{aligned}$$ The third term is $$\begin{aligned}
\label{eq1.13}
-q^{-2m-2n}(1-q^{-1})\sum_{j=1}^nq^{i+m}\left( \sum_{l=i+m+1}^{2k+m+n-c} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+m} \right).\end{aligned}$$ The last term is $$\begin{aligned}
\label{eq1.14}
q^{-m-n}\left( \sum_{l=m+n+2}^{2k+m+n-c} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+m+n+1} \right).\end{aligned}$$
We consider the integral $I_2^k=I_7^k+I_8^k$. From Eq.(\[eq1.9\]-\[eq1.14\]), the coefficient of the term involes $k$ is $$\begin{aligned}
&q^{-2m-2n}(1-q^{-1})\\
&\cdot \left((1-q^{-1})^2 \sum_i^m\sum_j^n q^{i+j}-(1-q^{-1})\sum_{i=1}^mq^{i+n}-(1-q^{-1})\sum_{j=1}^nq^{j+m}+q^{m+n}-1\right),\end{aligned}$$ which is zero. Thus $I_2^k$ is in fact independent on $k$ as long as $k$ is sufficiently large. We then get $$\begin{aligned}
I_2^k&=I_7^k+I_8^k \label{eq: I2}\\
&=q^{-2m-2n}(1-q^{-1})\sum_{i=1}^n(t_1t_3^{-1}q^{-1})^{c+i}(q^i-1) \nonumber\\
&+q^{-2m-2n}(1-q^{-1}) \sum_{i=1}^{n}(t_1t_3^{-1}q^{-1})^{c+i} \nonumber \\
&-q^{-2m-2n}(1-q^{-1})^2\sum_{i=1}^{m}\sum_{j=1}^n q^{i+j}\left( \sum_{l=1}^{i+j-1} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} +q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+j-1} \right) \nonumber\\
&-q^{-2m-2n}(1-q^{-1})\sum_{i=1}^mq^{i+n}\left(- \sum_{l=1}^{i+n} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+n} \right) \nonumber \\
&-q^{-2m-2n}(1-q^{-1})\sum_{i=1}^nq^{i+m}\left(- \sum_{l=1}^{i+m} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+i+m} \right) \nonumber \\
&+q^{-m-n}\left( -\sum_{l=1}^{m+n+1} (1-q^{-1})(t_1t_3^{-1}q^{-1})^{c+l} -q^{-1}(t_1t_3^{-1}q^{-1})^{c+m+n+1} \right). \nonumber\end{aligned}$$ To emphasize the dependence on $m,n$, we denote the above expression by $I_2(m,n)$. Let $I(m,n)=\lim_{k{\rightarrow}\infty}I^k(m,n)$. By Eq.(\[eq: I\^k\]), we have $$I(m,n)=I_1(m,n)+I_2(m,n),$$ where the right hand side was given by Eq.(\[eq1.8\]) and Eq.(\[eq: I2\]).
A simple calculation shows that $$W_f(1)=I(0,0)=I_1(0,0)+I_2(0,0)=(t_1t_3^{-1}q^{-1})^c(1-t_1t_3^{-1}q^{-1}).$$ Since $\mu_1\mu_3^{-1}\ne |~|^{\pm}$ (irreduciblity condition), we have $t_1t_3^{-1}q^{-1}\ne 1$. Thus $W_f(1)\ne 0$.
By a careful symbolic calculation, one can get that $$I(m,n)=q^{-2m-2n}\frac{1-(qX)^{m+n+1}}{1-qX}W_f(1),$$ where $X=t_1t_3^{-1}q^{-1}$. Now Theorem \[thm1.3\] follows from Eq.(\[eq1.3\]).
Calculation of Ginzburg’s local zeta integral in a special case {#computation of the local zeta integral in a special case}
===============================================================
In this section, as in Section \[notion on F\], we still let $F$ be non-archimedean local field, ${{\mathfrak{o}}}$ be the ring of integers of $F$, $\varpi$ be a uniformizer of $F$, and $q$ be the number of the residue field of $F$. We also fix a nonzero square $\rho,$ and identify the group $H_\rho$ with ${{\mathrm{SL}}}_3.$ As in section \[notation I(mu)\], we now consider the induced representation $\pi={{\mathrm{Ind}}}_{B_3}^{{{\mathrm{GL}}}_3(F)}(\mu_1\otimes\mu_2\otimes \mu_3)$ of ${{\mathrm{GL}}}_3(F)$ with, $\mu_1,\mu_3$ unramified, and $\mu_2$ ramified with conductor $c$. We further require that $\mu_1\mu_3=1$. Let $t_1=\mu_1(\varpi)$. Let $W_c\in {{\mathcal {W}}}(\pi,\psi)$ be the Whittaker function defined by . We know that $W_c$ is right invariant under the group $K_c\subset {{\mathrm{GL}}}_3(F)$. By Lemma \[lem: root killing\] and Theorem \[thm1.3\] we have $W_c(1)\ne 0$ and $$\label{eq3.1}
W_c({{\mathrm{diag}}}(\varpi^m,1,\varpi^{-m}))=\left\{\begin{array}{lll}\frac{q^{-2m}}{t_1-t_1^{-1}}(t_1^{2m+1}-t_1^{-(2m+1)})W_c(1),& m\ge 0,\\ 0, & m<0. \end{array}\right.$$
Let $\chi$ be an unramified quasi-character of $F^\times$, $s\in {{\mathbb {C}}}$. Recall that, by the exact sequence (\[eq2.2\]), the induced representation $I(s,\chi)$ has a subspace ${{\mathrm{ind}}}_{N_{2,\rho}\cdot {{\mathrm{GL}}}_1}^{H_\rho}(\chi'_s)$.
For a subset $A\subset F, $ we denote ${{\mathrm{ch}}}_A$ the characteristic function of $A$. Consider the following function $f_s^c$ on $H_\rho(F)\cong {{\mathrm{SL}}}_3(F)$: $$f_s^c(g)=\left\{\begin{array}{lll}0,& \textrm{ if } g\notin B_3K_c\cap H_\rho (F),\\ \chi'_s(nh(a,1)){{\mathrm{ch}}}_{1+{{\mathfrak{p}}}^c}(b){{\mathrm{ch}}}_{{\mathfrak{o}}}(x), & \textrm{ if } g=nh(a,1)n(x)h(1,b) k, n\in N_{2,\rho}, k\in K_c\cap H_\rho(F), \end{array}\right.$$ where $$n(x)=\begin{pmatrix}1& x& 0\\ &1&0\\ &&1 \end{pmatrix}\in H_\rho(F).$$ Note that for $n\in N_{2,\rho}$, if $nh(a,1)n(x)h(1,b)\in B_3\cap K_c\cap {{\mathrm{SL}}}_3(F)$, we can get $a\in {{\mathfrak{o}}}^\times, b\in 1+{{\mathfrak{p}}}^c, x\in {{\mathfrak{o}}}$, and hence $\chi'_s(nh(a,1)){{\mathrm{ch}}}_{1+{{\mathfrak{p}}}^c}(b){{\mathrm{ch}}}_{{\mathfrak{o}}}(x)=1$. Thus $f_s^c$ is well-defined. Note that $f_s^c$ is right $K_c$-invariant. By definition, $f_s^c\in {{\mathrm{ind}}}_{N_{2,\rho}\cdot {{\mathrm{GL}}}_1}^{H_\rho(F)}(\chi'_s)$ and hence defines an element $\tilde f_s^c\in I(s,\chi)$ by the exact sequence (\[eq2.2\]).
\[prop: special case\] We have $$Z(W_c,\tilde f_s^c)=D_c\frac{1+\chi(\varpi)q^{-(3s-1)}}{(1-t_1^2\chi(\varpi)q^{-(3s-1)})(1-t_1^{-2}\chi(\varpi)q^{-(3s-1)})},$$ where $D_c={{\mathrm{vol}}}(K_c){{\mathrm{vol}}}(1+{{\mathfrak{p}}}^c)W_c(1)$.
We have $$Z(W_c,\tilde f_s^c)=\int_{N_{2,\rho}\setminus H_\rho(F)}W_c(g)f_s^c(g)dg.$$ Since ${\mathrm{supp}}(f_s^c)\subset B_3K_c$ and both $W_c$ and $f_s^c$ are left $N_{2,\rho}$-invariant and right $K_c$-invariant, we get $$\begin{aligned}
Z(W_c, \tilde f_s^c)&={{\mathrm{vol}}}(K_c)\int_{F^\times} \int_{1+{{\mathfrak{p}}}^c}\int_{{{\mathfrak{o}}}}W_c(h(a,1)n(x)h(1,b))\chi'_s(h(a,1))|a|^{-3}dx d^*b d^*a\\
&={{\mathrm{vol}}}(K_c){{\mathrm{vol}}}(1+{{\mathfrak{p}}}^c)\int_{F^\times}W_c(h(a,1))\chi(a)|a|^{3s-3}d^*a.\end{aligned}$$ By , we get $$\begin{aligned}
Z(W_c,\tilde f_s^c)&=D_c\sum_{m\ge 0} \frac{q^{-2m}}{t_1-t_1^{-1}}(t_1^{2m+1}-t_1^{-(2m+1)})(\chi(\varpi) q^{-(3s-3)})^m\\
&=D_c\sum_{m\ge 0}\frac{1}{t_1-t_1^{-1}}(t_1^{2m+1}-t_1^{-(2m+1)})(\chi(\varpi)q^{-(3s-1)})^m\\
&=D_c\frac{1+\chi(\varpi)q^{-(3s-1)}}{(1-t_1^2\chi(\varpi)q^{-(3s-1)})(1-t_1^{-2}\chi(\varpi)q^{-(3s-1)})}.\end{aligned}$$ This concludes the proof.
Holomorphy of adjoint Lg-function for Lg {#sec: split main theorem}
========================================
In this section, let $F$ be a global field and ${{\mathbb {A}}}$ be the ring of adeles of $F$. Let $\pi=\otimes \pi_v$ be an irreducible cuspidal automorphic representation of ${{\mathrm{GL}}}_3({{\mathbb {A}}})$. Let $\chi=\otimes\chi_v$ be a unitary Hecke character of $F^\times\setminus {{\mathbb {A}}}^\times$. Then one can consider the twisted adjoint $L$-function $$L(s,\pi,{{\mathrm{Ad}}}\times \chi)=\frac{L(s,(\chi\otimes \pi)\times \tilde \pi)}{L(s,\chi)}.$$
For a fixed $\pi$ and $\chi$, let $S=S(\pi,\chi)$ be the finite set of places consisting of all archimedean places and all finite places $v$ such that either $\pi_v$ or $\chi_v$ is ramified. The partial twisted adjoint $L$-function $L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)$ is defined by $$L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)=\prod_{v\notin S}L(s,\pi_v, {{\mathrm{Ad}}}\times \chi_v)=\prod_{v\notin S}\frac{L(s,(\chi_v\otimes \pi_v)\times \tilde \pi_v)}{L(s,\chi_v)}.$$
After the pioneering work of Ginzburg [@G], and Ginzburg-Jiang [@GJ], the following result was obtained:
\[thm: hundley\] The partial twisted adjoint $L$-function $L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)$ has no poles in the half plane ${{\mathrm{Re}}}(s)\ge \frac{1}{2}$, except possibly for a simple pole at ${{\mathrm{Re}}}(s)=1$ when $\chi$ is non-trivial and $\pi \cong \pi \otimes \chi$. (This forces $\chi$ to be cubic.) Every other pole of the complete $L$-function $L(s,\pi,{{\mathrm{Ad}}}\times \chi)$ in ${{\mathrm{Re}}}(s)\ge \frac{1}{2}$ is a zero of the Hecke $L$-function $L(s,\chi)$ and a pole of $\prod_{v\in S}L_v(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$.
The analogous result for quasisplit unitary groups was also obtained. We want to extend the above result to $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi)$, where $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi)=\prod_{v\notin S_\infty}L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ is the finite part of the $L$-function, where $S_\infty$ is the set of infinite places of $F$. In order to do this, we must treat all of the finite places $v\in S-S_{\infty}$, i.e., either $\pi_v$ or $\chi_v$ is ramified. The cases where $\pi_v$ is ramified can be split up according to whether $\pi_v$ is tempered or non-tempered. In the non-tempered case, we are able to exploit the classification of unitary representations, to say that a representation which is ramified, unitary, and non-tempered is of a fairly specific form. In the tempered case, the arguments which we use for ${{\mathrm{GL}}}_3$ work equally well in the unitary group case, and we therefore record them in this generality.
\[lem:tempered\] Let $v$ be a place of $F,$ $\rho$ an element of $F^\times,$ and $\pi_v$ an irreducible generic tempered representation of $H_\rho(F_v).$ Then $L(s, \pi_v, {{\mathrm{Ad}}}\times \chi_v)$ has no poles in ${{\mathrm{Re}}}(s) > 0.$
If $H_\rho$ is isomorphic to ${{\mathrm{SL}}}_3$ over $F_v,$ then $L(s, \pi_v, {{\mathrm{Ad}}}\times \chi) =L(s, \pi_v \times ({\widetilde}\pi_v \otimes \chi_v))/L(s, \chi_v).$ Since $L(s, \chi_v)^{-1}$ has no poles at all, it suffices to prove that $L(s, \pi_v \times {\widetilde}\pi_v \times \chi_v)$ has no poles in ${{\mathrm{Re}}}(s) > 0.$ This may be deduced from [@JPSS83 Proposition 8.4, page 451], since $({\widetilde}\pi_v \otimes \chi_v)$ is again tempered.
If $H_\rho$ is not split over $F_v$, then it determines a quadratic extension field $F_v(\tau)$ of $F_v.$ Let $\chi_{F_v(\tau)/F_v}$ denote the corresponding quadratic character. Then $$L(s, \pi_v, {{\mathrm{Ad}}}\times \chi) =L(s, {{\operatorname}{sbc}}(\pi_v), {{\operatorname}{Asai}}\times (\chi_v\chi_{F_v(\tau)/F_v}))/L(s, \chi_v),$$ where ${{\operatorname}{sbc}}$ denotes the stable base change lift of Kim and Krishnamurthy [@KimKrishOdd]. The twisted Asai $L$-function $L(s, {{\operatorname}{sbc}}(\pi_v), {{\operatorname}{Asai}}\times \chi_v\chi_{F_v(\tau)/F_v})$ may also be realized Asai $L$-function of the twist: $L(s, {{\operatorname}{sbc}}(\pi_v)\otimes {\widetilde}\chi, {{\operatorname}{Asai}})$ where ${\widetilde}\chi$ is any character of $F_v(\tau)^\times$ whose restriction to $F_v^\times$ is $\chi_v \chi_{F_v(\tau)/F_v}.$ Thus it suffices to show that the stable base change lift of a tempered representation is again tempered, and that the Asai $L$-function of a tempered representation has no pole in ${{\mathrm{Re}}}(s) > 0. $
The fact that the local stable base change lift of a generic tempered representation is again tempered is proved for quasisplit ${{\mathrm {U}}}_{2n}$ in Proposition 8.6 of [@KimKrishEven]. The argument adapts to ${{\mathrm {U}}}_{2n+1}$ in a straightforward manner, using the results of [@KimKrishOdd]. Holomorphy of the Asai $L$-function for tempered representations in ${{\mathrm{Re}}}(s)> 0$ then follows from Proposition 7.2 of [@Shahidi-Plancharel].
\[prop: calculation of zeta\] Fix a finite place $v\in S-S_{\infty}$ of $F.$ Let $\pi_v$ be a nontempered irreducible unitary representation of ${{\mathrm{GL}}}_3(F_v)$ and $\psi_v$ be an additive character of $F_v$. Define $\psi_{\rho, v}:U_\rho(F_v) \to {\mathbb{C}}^\times$ by and assume that $\pi_v$ is $\psi_{\rho,v}$-generic. Let $I(s,\chi_v)$ be the induced representation of $G_2(F_v)$ defined in $\S\ref{defn of I(s,chi)}$. Then there exists a Whittaker function $W_v\in {{\mathcal {W}}}(\pi_v,\psi_{\rho,v})$ and a standard section $f_{s,v}\in I(s,\chi_v)$ such that $$L(3s,\chi_v)L(6s-2,\chi_v^2)L(9s-3,\chi_v^3)\frac{Z(W_v,f_{s,v})}{L(3s-1,\pi_v,{{\mathrm{Ad}}}\times \chi_v)}$$ has no zeros on the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2}$.
By proposition \[prop:dependence on psi\], it suffices to treat the case when $\psi$ is also unramified. By the classification of unitary representation of ${{\mathrm{GL}}}_3$, [@JPSS79 $\S$6], the representation has the form $\pi_v=\eta_v\otimes \sigma_v$, where $\eta_v$ is a unitary character of $F_v^\times$ and $\sigma_v$ is of the form $${{\mathrm{Ind}}}_{B_{3}(F_v)}^{{{\mathrm{GL}}}_3(F_v)}(|~|^{\alpha}\otimes \mu_2\otimes |~|^{-\alpha}),$$ where $\mu_2$ is a unitary character of $F_v^\times$ and $\alpha$ is a real number with $0<\alpha<\frac{1}{2}$.
Note that $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)=L(s,\sigma_v,{{\mathrm{Ad}}}\times \chi_v)$ and for $W_v\in {{\mathcal {W}}}(\pi_v,\psi_v)$, $W_v|_{{{\mathrm{SL}}}_3(F_v)}$ is a Whittaker function for $\sigma_v$. Since Ginzburg’s integral $Z(W_v,f_{s,v})$ only depends on $W_v|_{{{\mathrm{SL}}}_3(F_v)}$, we can assume that $\eta_v=1$, i.e., $\pi_v=\sigma_v$.
Note that the assumption $v\in S-S_\infty$ implies that either $\mu_2$ is ramified or $\chi_v$ is ramified.
We first consider the case when $\mu_2$ is unramified. If $\chi_v$ is ramified, from the equation $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)=L(s,(\chi_v\pi_v)\otimes {\widetilde}\pi_v)/L(s,\chi_v)$ and [@JPSS83 Theorem 3.1], one can check that $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)=1$. Thus the Claim follows from the fact that one can find $W_v,f_{s,v}$ such that $Z(W_v,f_{s,v})$ is a nonzero constant, see Lemma \[lemnonvanishing\].
We next consider the case when $\mu_2$ is ramified and $\chi_v$ is also ramified. In this case by [@CPS Theorem 4.1], one can check that $$L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)=L(s,\chi_v\mu_2^{-1}|~|^\alpha)L(s,\chi_v\mu_2^{-1}|~|^{-\alpha})L(s,\chi_v\mu_2|~|^\alpha)L(s,\chi_v\mu_2|~|^{-\alpha}).$$ Note that if $\chi_v\mu_2$ and $\chi_v\mu_2^{-1}$ are ramified, then $L(s,\pi_v,{{\mathrm{Ad}}}\times\chi_v)=1$ and thus it has no pole. If either $\chi_v\mu_2$ or $\chi_v\mu_2^{-1}$ is unramified, then $L(s,\pi_v,{{\mathrm{Ad}}}\times\chi_v)$ has no pole on the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2},$ because $\chi_v$ and $\mu_2$ are unitary and $\alpha<1/2.$ Hence $L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ has no pole in the region ${{\mathrm{Re}}}(s)\ge \frac{1}{2}.$ The assertion then also follows from the nonvanishing of the local Ginzburg’s integrals, see Lemma \[lemnonvanishing\].
Finally, we consider the case when $\mu_2$ is ramified and $\chi_v$ is unramified. In this case, by [@CPS Theorem 4.1] again, we have $$L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v)=L(s,\chi_v)^2L(s,\chi_v|~|^{2\alpha})L(s,\chi_v|~|^{-2\alpha}).$$ By Proposition \[prop: special case\], we can take $W_v\in {{\mathcal {W}}}(\pi_v,\psi_v)$ and a standard section $f_{s,v}\in I(s,\chi_v)$ such that $$\begin{aligned}
Z(W_v,f_{s,v})&=\frac{1+\chi_v(\varpi)q^{-(3s-1)}}{(1-\chi_v(\varpi)|\varpi|^{2\alpha}q^{-(3s-1)})(1-\chi_v(\varpi)|\varpi|^{-2\alpha}q^{-(3s-1)})}\\
&=\frac{1-\chi_v(\varpi)^2q^{-(6s-2)}}{(1-\chi_v(\varpi)q^{-(3s-1)})(1-\chi_v(\varpi)|\varpi|^{2\alpha}q^{-(3s-1)})(1-\chi_v(\varpi)|\varpi|^{-2\alpha}q^{-(3s-1)})}\\
&=\frac{L(3s-1,\chi_v)L(3s-1,\chi_v|~|^{2\alpha})L(3s-1,\chi_v|~|^{-2\alpha})}{L(6s-2,\chi_v^2)}.\end{aligned}$$ Thus $$\begin{aligned}
\frac{Z(W_v,f_{s,v})}{L(3s-1,\pi_v,{{\mathrm{Ad}}}\times \chi_v)}&=\frac{1}{L(3s-1,\chi_v)L(6s-2,\chi_v^2)},\end{aligned}$$ and $$\begin{aligned}
&L(3s,\chi_v)L(6s-2,\chi_v^2)L(9s-3,\chi_v^3)\frac{Z(W_v,f_{s,v})}{L(3s-1,\pi_v,{{\mathrm{Ad}}}\times \chi_v)}\\
=&L(3s,\chi_v)\frac{L(9s-3,\chi_v^3)}{L(3s-1,\chi_v)}\\
=&\frac{1}{(1-\chi_v(\varpi)q^{-3s})(1+\chi_v(\varpi)q^{-(3s-1)}+\chi_v(\varpi)^2q^{-(6s-2)})}.\end{aligned}$$ The last expression clearly has no zeros. The assertion follows.
\[thm4.2\] Let $\chi$ be a unitary Hecke character of $F^\times\setminus {{\mathbb {A}}}^\times$ and $\pi$ be an irreducible cuspidal automorphic representation of ${{\mathrm{GL}}}_3({{\mathbb {A}}})$, then the finite part of the adjoint $L$-function $L_f(s,\pi, {{\mathrm{Ad}}}\times \chi)$ is holomorphic on the region ${{\mathrm{Re}}}(s)\ge 1/2$, except for a simple pole at ${{\mathrm{Re}}}(s)=1$ when $\chi$ is non-trivial and $\pi \cong \pi \otimes \chi$ (which forces $\chi$ to be cubic).
Note that we can write $\pi=|\det|^{\zeta}\pi_0$ with $\zeta\in {{\mathbb {C}}}$ and $\pi_0$ unitary. From the relation $L_f(s,\pi,{{\mathrm{Ad}}}\times\chi)=L_f(s,(\chi\otimes\chi)\times \tilde \pi)/L_f(s,\chi)$, we can get $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi)=L(s,\pi_0,{{\mathrm{Ad}}}\times\chi)$. Replacing $\pi$ by $\pi_0$ if necessary, we can assume that $\pi$ is unitary.
We first show that $L_f(s,\pi, {{\mathrm{Ad}}}\times \chi)$ is holomorphic for ${{\mathrm{Re}}}(s)\ge 1/2$.
Let $S=S(\pi,\chi)$. Then we can write $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi)=\prod_{v\in S-S_\infty}L(s,\pi_v,{{\mathrm{Ad}}}\times \chi_v) \cdot L^S(s,\pi,{{\mathrm{Ad}}}\times \chi)$. By Theorem \[thm: hundley\], it suffices to consider $L(s,\pi_v, {{\mathrm{Ad}}}\times \chi_v)$ for every place $v\in S-S_\infty$. Note that each $\pi_v$ is unitary since $\pi$ is cuspidal. Note that if $\pi_v$ is tempered, $L(s,\pi_v, {{\mathrm{Ad}}}\times \chi_v)$ has no pole on ${{\mathrm{Re}}}(s)>0$ by Lemma \[lem:tempered\].
We next assume that $\pi_v$ is non-tempered. Given a global section $f_s\in I(s,\chi)$, one can consider the normalized Eisenstein series $$E(g,f_s)=L(3s,\chi)L(6s-2,\chi^2)L(9s-3,\chi^3)\sum_{\lambda\in P(F)\setminus G_2(F)}f_s(\lambda g).$$ Given a cusp form $\varphi$ in the space of $\varphi$, which is assumed to correspond to a pure tensor, and a pure tensor $f_s=\otimes f_{s,v}\in I(s,\chi)$, Ginzburg [@G] defined the global integral $$Z(\varphi, f_s)=\int_{{{\mathrm{SL}}}_3(F)\setminus {{\mathrm{SL}}}_3({{\mathbb {A}}})}\varphi(g)E(g,f_s)dg,$$ and showed that it is Eulerian: $$Z(\varphi,f_s)=\prod_{v} Z^*(W_v,f_{s,v}),$$ where $Z^*(W_v,f_{s,v})=L(3s,\chi_v)L(6s-2,\chi_v^2)L(9s-3,\chi_v^3)Z(W_v,f_{s,v})$, $W_v$ is the $v$-th component of the Whittaker function of $\varphi.$ In [@G], it is also showed that if $v$ is a place such that $\pi_v, \chi_v$ and $\psi_v$ are all unramified, then $Z^*(W_v,f_{s,v})=L(3s-1,\pi_v,{{\mathrm{Ad}}}\times \chi_v)$ for spherical $W_v,f_{s,v}$. (Both [@G] and [@H12] only treat the case when $\chi$ is trivial, but the extension to nontrivial $\chi$ is direct.) The same identity holds up to an exponential factor at places when $\pi_v$ and $\chi_v$ are unramified but $\psi_v$ is ramified. Thus we get $$\begin{aligned}
Z(\varphi,f_s)&=
\chi_s(h(1,a))\prod_{v\in S_\infty}Z^*(W_v,f_{s,v})\cdot \prod_{v\in S-S_\infty}Z^*(W_v,f_{s,v})\cdot L^S(3s-1,\pi,{{\mathrm{Ad}}}\times \chi) \label{eq4.1}\\
&=\chi_s(h(1,a))\prod_{v\in S_\infty}Z^*(W_v,f_{s,v})\cdot \prod_{v\in S-S_\infty}\frac{Z^*(W_v,f_{s,v})}{L(3s-1,\pi_v,{{\mathrm{Ad}}}\times \chi_v)}\cdot L_f(3s-1,\pi,{{\mathrm{Ad}}}\times \chi), \nonumber\end{aligned}$$ for a certain idèle $a$ determined by the finite places $v$ such that $\psi_v$ is ramified and $\pi_v$ and $\chi_v$ are not. After the work of [@GJ], it is shown in [@H18] that for a flat section $f_s\in I(s,\chi)$, $Z(\varphi,f_s)$ has no pole on the region ${{\mathrm{Re}}}(s)\ge 1/2$ except for a possible simple pole at ${{\mathrm{Re}}}(s)=2/3$ which can occur only when $\chi$ is cubic. By Eq.(\[eq4.1\]), the non-vanishing results of the local zeta integrals $Z^*(W_v,f_{s,v})$ [@H18 Theorem 5.1] and Proposition \[prop: calculation of zeta\], we obtained that $L_f(3s-1,\pi,{{\mathrm{Ad}}}\times \chi)$ has no pole on the region ${{\mathrm{Re}}}(s)\ge 1/2$ except for a possible simple pole at $s=2/3$ in the case when $\chi$ is cubic.
Thus $L_f(s,\pi, {{\mathrm{Ad}}}\times \chi)$ has no pole on the region ${{\mathrm{Re}}}(s)\ge 1/2$ except for a possible simple pole at $s=1$ in the case $\chi$ is cubic. By Proposition 3.6 of [@JS-EulerProductsII], combined with [@JPSS83 Proposition 8.4, page 451], $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi )$ has a pole (which is simple when it exists) at $s=1$ if and only if $\chi\pi \cong \pi.$ By Tate’s thesis or the original work of Hecke, $L_f(s, \chi)$ has a pole (which is simple when it exists) at $s=1$ if and only if $\chi$ is trivial (in which case $L_f(s, \chi)$ is just the Dedekind zeta function of $F$).
It follows that $L_f(s,\pi,{{\mathrm{Ad}}}\times \chi )$ has a pole at $s=1$ if and only if $\pi \cong \chi\pi$ (which implies $\chi^3=1$), and $\chi$ is nontrivial.
Further discussion of poles in the nonsplit case {#sec: quasisplit}
================================================
For the remainder of the paper we devote our attention to a detailed study of the poles of $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi)$ when $\rho$ is a non-square. Here $S$ is a finite set of places, containing all Archimedean places, such that $\pi_v$ and $\chi_v$ are unramified for $v \notin S.$ By [@H18], theorem 6.4, $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi)$ has no poles in ${{\mathrm{Re}}}(s) \ge \frac 12,$ and may have a pole at $s=1$ only if $\chi$ is nontrivial cubic or $\chi$ is quadratic and $\pi$ is distinguished with respect to a group $H'_\rho,$ which we may think of as ${{\mathrm{SL}}}_2,$ or more suggestively as ${{\mathrm{SU}}}_{1,1}$ embedded into ${{\mathrm{SU}}}_{2,1}.$ See [@H18] for details. (In view of Lemma \[lem:tempered\], this information about the poles of $L^S(s, \pi, {{\mathrm{Ad}}}\times \chi)$ remains true even if we remove places $v$ such that $\pi_v$ is tempered from $S.$) Let $H^\flat_\rho$ denote the quasisplit unitary group ${{\mathrm {U}}}_{1,1}$, embedded into $H_\rho$ so that $H_\rho'$ is the derived group. Determinant maps ${{\mathrm {U}}}_{1,1}$ to ${{\mathrm {U}}}_1$ and we can choose a splitting to write $H_\rho^\flat$ as the internal semidirect product of $H_\rho'$ and a subgroup isomorphic to ${{\mathrm {U}}}_1.$ Factoring the Haar measure on $H_\rho^\flat$ accordingly shows that any representation distinguished with respect to $H_\rho'$ must also support the period $$\varphi \mapsto
{{\int\limits_}{{H_\rho^\flat(F){\backslash}H_\rho^\flat({\mathbb{A}})}}}\varphi(h)\eta^{-1}(\det h)\ dh$$ for some character $\eta$ of ${{\mathrm {U}}}_1({\mathbb{A}})$ (in which case we say that $\pi$ is $(H_\rho^\flat,\eta)$-distinguished). It is proved in [@GeRoSo2] that this forces the $L$-packet $\{\!\{ \pi \}\! \}$ of $\pi$ to be the image, under the endoscopic transfer(s) constructed in [@RogawskiBook], of some $L$-packet $\rho = \rho_2\times \rho_1$ of ${{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1,$ (where $\rho_2$ is an $L$-packet of representations of ${{\mathrm {U}}}_{1,1}({\mathbb{A}})$ and $\rho_1$ is a character of ${{\mathrm {U}}}_1({\mathbb{A}}) \approx {\mathbb{A}}_E^1$). (The $L$-homomorphism to which this transfer is attached depends on some choices; varying the choice varies the packet $\rho_2$ such that $\rho_2\times \rho_1$ lifts to $\{\!\{\pi\}\!\},$ without changing the overall image of transfer.) It is then proved in [@GeRoSo1] that if $\pi$ is $(H_\rho^\flat,\eta)$-distinguished, then at least one of the $L$-packets $\rho=\rho_2\times \rho_1$ which transfers to $\{\!\{\pi\}\!\}$ satisfies $\rho_1 = \eta.$
Thus, any pole of $L^S(s, \pi, {{\mathrm{Ad}}}'\times \chi)$ at $s=1,$ when $\chi$ is either trivial or quadratic, indicates that $\pi$ is endoscopic. In order to complete our treatment of the nonsplit case, we would like to address the case when $\chi$ is cubic, and to study poles in the case when $\pi$ is assumed to be endoscopic. Further investigation requires that we study Rogawski’s transfer(s) in more detail.
Weil forms of Lg-groups
-----------------------
### A technical point regarding Lg-groups {#ss: tech pt}
For purposes of discussing Rogawski’s transfer(s) the finite Galois form of the $L$-group will not suffice; we must consider the Weil form. We briefly explain the reason. We may realize the finite Galois form of ${}^L{{\mathrm {U}}}_{2,1}$ as in [@H12] and [@H18], as ${{\mathrm{GL}}}_3({\mathbb{C}}) \rtimes {{\mathrm{Gal}}}(E/F),$ with the nontrivial element of ${{\mathrm{Gal}}}(E/F)$ acting by $g\mapsto {}_tg^{-1}.$ Here $_tg=J{}^tgJ,$ as in [@H12; @H18]. This automorphism of ${{\mathrm{GL}}}_3({\mathbb{C}})$ preserves the subgroup subgroup$$\label{GL2 x GL1 --> GL3}
\left \{
{\begin{pmatrix}}a& &b \\ &t& \\ c&&d {\end{pmatrix}}\right \} \cong {{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}}) \subset {{\mathrm{GL}}}_3({\mathbb{C}})$$ of ${{\mathrm{GL}}}_3({\mathbb{C}}),$ and hence we obtain a subgroup of $^L{{\mathrm {U}}}_{2,1}$ which is the semidirect product of ${{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ and ${{\mathrm{Gal}}}(E/F).$
The finite Galois form of the $L$-group of ${{\mathrm {U}}}_{1,1}\times {{\mathrm {U}}}_1$ is also a semidirect product of ${{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ and ${{\mathrm{Gal}}}(E/F),$ but it may not be identified with this subgroup of $^L{{\mathrm {U}}}_{2,1}.$ Indeed, the nontrivial element of ${{\mathrm{Gal}}}(E/F)$ acts on $^L({{\mathrm {U}}}_{1,1}\times {{\mathrm {U}}}_1)$ by an automorphism of order two. As described in [@BorelCorvallis], this automorphism must satisfy certain conditions. The identity component of $^L({{\mathrm {U}}}_{1,1}\times {{\mathrm {U}}}_1)$ inherits, from its definition via based root data, a choice of split maximal torus and Borel subgroup. We may fix the isomorphism $^L({{\mathrm {U}}}_{1,1}\times {{\mathrm {U}}}_1)^0\to {{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ so that they consist of the diagonal and upper triangular elements, respectively. The automorphism must map these to themselves in a manner determined by duality and the action of ${{\mathrm{Gal}}}({\overline}F/F)$ on the maximal torus of ${{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1.$ The last condition, given in [@BorelCorvallis], 1.2, states that for a general reductive group there must be a set $\{ x_\alpha\}$ of representatives for the root subgroups attached to the simple roots, whose elements are permuted amongst themselves. In our case, where there is only one simple root, this last condition means that the action on the standard maximal unipotent subgroup of ${{\mathrm{GL}}}_2({\mathbb{C}})$ must be trivial. Since $g \mapsto {}_tg^{-1}$ does not act trivially on the maximal unipotent subgroup of it follows that the semidirect product of ${{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ and ${{\mathrm{Gal}}}(E/F),$ which sits naturally inside of $^L{{\mathrm {U}}}_{2,1},$ may not be identified with $^L({{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1).$
### Definition of the Weil forms of the $L$-group
The Weil form of the $L$-group of ${{\mathrm {U}}}_{2,1}$ is the semidirect product of ${{\mathrm{GL}}}_3({\mathbb{C}})$ and the Weil group, $W_F$ of $F,$ (see [@Tate-NumberTheoreticBackground]) with the action implicit in the semidirect product being defined using the canonical mapping $W_F/W_E \to {{\mathrm{Gal}}}(E/F).$ Thus, elements of $W_E$ commute with ${{\mathrm{GL}}}_3({\mathbb{C}})$ while elements of $W_F\smallsetminus W_E$ act by $g \mapsto {}_tg^{-1}.$ The Weil form of the $L$-group of ${{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1$ is defined similarly with ${{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ replacing ${{\mathrm{GL}}}_3({\mathbb{C}}).$ We may take the involution of ${{\mathrm{GL}}}_2({\mathbb{C}}) \times {{\mathrm{GL}}}_1({\mathbb{C}})$ to be $$(g, a) \mapsto \left( {\begin{pmatrix}}& 1 \\ -1& {\end{pmatrix}}{}^tg^{-1} {\begin{pmatrix}}& -1 \\ 1& {\end{pmatrix}}, a^{-1}\right)
= (\det g^{-1} \cdot g, a^{-1} ).$$
### Weil forms and the Satake parametrization
We recall the parametrization of unramified representations in [@BorelCorvallis] which is suited to dealing with Weil forms of $L$-groups. It is based on $L$-homomorphisms from the Weil group $W_F \to {}^L\!G.$ We do not need the full definition of an $L$-homomorphism, only the notion of an [*unramified*]{} $L$-homomorphism of $W_F$ for $F$ local nonarchimedean.
In this case $W_F$ is a dense subgroup of ${{\mathrm{Gal}}}({\overline}F/F)$ and comes equipped with a homomorphism ${{\mathrm{ord}}}_{W_F}: W_F \to {\mathbb{Z}}.$ An unramified $L$-homomorphism $W_F \to G^\vee({\mathbb{C}}) \rtimes W_F$ is a homomorphism which sends $w \in W_F$ to $(t^{{{\mathrm{ord}}}_{W_F}(w)}, w)$ for some semisimple element $t$ of $G^\vee({\mathbb{C}}).$ Note that $t$ must be ${{\mathrm{Gal}}}({\overline}F/ F)$-fixed in order for this to be a homomorphism, and, since we work up to conjugacy, we may assume $t$ is in $T^\vee({\mathbb{C}}).$ Thus conjugacy classes of unramified $L$-homomorphisms are in bijection with Galois-fixed elements of $T^\vee({\mathbb{C}}),$ which correspond to unramified characters as usual.
### Weil forms and $L$-functions
Let $F$ be nonarchimedean and local. We briefly recall the definition of $L(s, \varphi)$ for $\varphi: W_F \to {{\mathrm{GL}}}(V)$ a finite dimensional representation of $W_F.$
The kernel of ${{\mathrm{ord}}}_{W_F}$ is a normal subgroup of $W_F$ called the inertia group. We denote it $I_F.$ As $I_F$ is normal, its fixed subspace $V^{I_F}$ is $W_F$-invariant. We define $$L(s, \varphi) : = \det(I - q^{-s} \varphi(w)\big|_{V^{I_F}} )^{-1},
\quad \text{ for } w \in W_F \text{ with }{{\mathrm{ord}}}_{W_F}(w) = 1.$$ (The expression on the right-hand-side is independent of the choice of $w$.)
This then permits us to define $L(s, \pi, r)$ for $\pi$ an unramified representation and $r$ a finite dimensional representation of $G^\vee({\mathbb{C}}) \rtimes W_F.$ Indeed $\pi$ is attached to an unramified $L$-homomorphism $\varphi_t(w) = (t^{{{\mathrm{ord}}}_{W_F}(w)}, w),$ with $t \in T^\vee({\mathbb{C}})^{W_F},$ and $L(s, \pi, r)$ is defined as $$L(s, \pi, r)=L(s, r \circ \varphi_t)= \det(I - q^{-s} r(t)r(w)\big|_{V^{I_F}} )^{-1},\quad \text{ for } w \in W_F \text{ with }{{\mathrm{ord}}}_{W_F}(w) = 1.$$
Adjoint representations
-----------------------
For $H={{\mathrm {U}}}_{2,1}$ (resp. ${{\mathrm {U}}}_{1,1}$) the action of ${}^L\!H$ on itself by conjugation determines an action on ${\mathfrak}{sl}_3({\mathbb{C}})$ (resp. ${\mathfrak}{sl}_2({\mathbb{C}})$) which we denote ${{\mathrm{Ad}}}.$ Since each $w \in W_F {\smallsetminus}W_E$ acts on ${{\mathrm{GL}}}_3({\mathbb{C}})$ by $g \mapsto {}_tg^{-1},$ it will act on ${\mathfrak}{sl}_3({\mathbb{C}})$ by $X \mapsto -{}_tX.$ Note that this conflicts with the notation of [@H12].
Let ${{\mathrm{Ad}}}'$ denote the representation where ${{\mathrm{GL}}}_3({\mathbb{C}})$ acts by conjugation and each $w \in W_F {\smallsetminus}W_E$ acts by $X \mapsto {}_tX.$ This is the representation denoted ${{\mathrm{Ad}}}$ in [@H12]. It can also be described as the twist of ${{\mathrm{Ad}}}$ by the quadratic character $\chi_{E/F}$ attached to $E/F.$
Base Change and Automorphic Induction
-------------------------------------
Base change for the quadratic extension $E/F$ is the functorial lifting attached to the $L$-homomorphism $${{\operatorname}{bc}}:{}^L({{\mathrm{GL}}}_n) = {{\mathrm{GL}}}_n({\mathbb{C}}) \times W_F \to {}^L({{\mathrm{Res}}}_{E/F} {{\mathrm{GL}}}_n) = ({{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}}) ) \rtimes W_F$$ which sends $w \in W_F$ to itself, and $g \in {{\mathrm{GL}}}_n({\mathbb{C}})$ to $(g,g) \in ({{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}}) ).$ Automorphic induction for the quadratic extension $E/F$ is the functorial lifting attached to the $L$-homomorphism $${{\operatorname}{AI}_{E/F}}: {}^L({{\mathrm{Res}}}_{E/F} {{\mathrm{GL}}}_n) = ({{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}}) ) \rtimes W_F
\to ^L({{\mathrm{GL}}}_{2n})= {{\mathrm{GL}}}_{2n}({\mathbb{C}}) \times W_F$$ $${{\operatorname}{AI}_{E/F}}(g_1, g_2) = {\begin{pmatrix}}g_1&\\ & g_2 {\end{pmatrix}}, \qquad {{\operatorname}{AI}_{E/F}}(w) = \begin{cases}{\begin{pmatrix}}&I_n \\ I_n& {\end{pmatrix}}, & w \in W_F {\smallsetminus}W_E\\ I_{2n} & w \in W_E.
\end{cases}$$ Both of these cases of functoriality are proved (in greater generality) in [@ArthurClozel].
Stable Base Change and its image
--------------------------------
The stable base change lifting of Kim and Krishnamurthy has already been mentioned a couple of times. It lifts globally generic automorphic representations of the quasisplit group ${{\mathrm {U}}}_n$ (Elsewhere in the paper, we denoted ${{\mathrm {U}}}_3$ by ${{\mathrm {U}}}_{2,1}$ and ${{\mathrm {U}}}_2$ by ${{\mathrm {U}}}_{1,1}$ to emphasize the quasisplit nature. In the case of general $n$ this notation seems cumbersome.) attached to a quadratic extension $E/F$ to automorphic representations of ${{\mathrm{Res}}}_{E/F} {{\mathrm{GL}}}_n.$ The $L$-group of ${{\mathrm {U}}}_n$ is ${{\mathrm{GL}}}_n({\mathbb{C}}) \rtimes W_F,$ and $W_F{\smallsetminus}W_E$ acts on ${{\mathrm{GL}}}_n({\mathbb{C}})$ by a nontrivial involution which we denote $g \mapsto g^*.$ There is some freedom to the choice of involution by it must preserve the torus and the borel and permute a collection of elements $\{ x_\alpha\}$ as in [@BorelCorvallis §1.2]. We can take $g^* = {}_tg^{-1}$ when $n$ is odd but not when $n$ is even (cf. §\[ss: tech pt\]). In the even case we can take $g^* = d_0{}_tg^{-1} d_0,$ where $d_0$ is a diagonal matrix with alternating $1$’s and $-1$’s on the diagonal.
The $L$-group of ${{\mathrm{Res}}}_{E/F}{{\mathrm{GL}}}_n$ is $({{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}}) ) \rtimes W_F,$ and $W_F{\smallsetminus}W_E$ acts by permuting the factors. Stable base change is the functorial lifting which corresponds to the $L$-homomorphism $${{\operatorname}{sbc}}(g,w)= (g, {}_tg^{-1}, w).$$ It is closely related to the lifting from ${{\mathrm {U}}}_n$ to ${{\mathrm{Res}}}_{E/F}{{\mathrm {U}}}_n$ which is considered in [@RogawskiBook] (where it is called “base change”).
### Asai representations
The representation of ${{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}})$ on ${{\operatorname}{Mat}}_{n\times n}({\mathbb{C}})$ by $(g_1, g_2).X = g_1X{}_tg_2$ is irreducible. Thus it has two distinct extensions to a representation of $({{\mathrm{GL}}}_n({\mathbb{C}}) \times {{\mathrm{GL}}}_n({\mathbb{C}}))\rtimes {{\mathrm{Gal}}}(E/F),$ the finite Galois form of the $L$-group of ${{\mathrm{Res}}}_{E/F}{{\mathrm{GL}}}_n.$ We denote them ${{\operatorname}{Asai}}^\pm,$ such that ${{\operatorname}{Asai}}^{\pm}(w).X = \pm {}_tX$ for $w \in W_F \smallsetminus W_E.$
It is perhaps more conventional to define the Asai representation using the usual transpose as opposed to the lower transpose. The two representations are isomorphic, with an isomorphism given by $X \mapsto XJ.$ In the usual Asai representation, the sign is $+.$
### Asai of stable base change
It is readily verified that $${{\operatorname}{Asai}}^+ \circ {{\operatorname}{sbc}}=\begin{cases} {{\mathrm{Ad}}}' \oplus \mathbf{1}, n \text{ odd,}\\
{{\mathrm{Ad}}}\oplus \chi_{E/F}, n \text{ even},
\end{cases}
\qquad
{{\operatorname}{Asai}}^- \circ {{\operatorname}{sbc}}= \begin{cases}{{\mathrm{Ad}}}\oplus \chi_{E/F},& n \text{ odd,}\\
{{\mathrm{Ad}}}'\oplus \mathbf{1}, & n \text{ even}
\end{cases}$$ where $\mathbf{1}$ is the one dimensional trivial representation, and ${{\mathrm{Ad}}}$ and ${{\mathrm{Ad}}}'$ are defined as in [@H18]. Thus if we let $(-)^k=+$ for $k$ even and $-$ for $k$ odd, then $$\label{eq relating Ad,Ad' Asai for U21}
L^S( s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}^{(-)^n}) =L^S(s, \chi_{E/F}) L^S(s, \pi, {{\mathrm{Ad}}}) \qquad
L^S( s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}^{(-)^{n+1}}) = \zeta^S(s) L^S(s, \pi, {{\mathrm{Ad}}}').$$
### Image of stable base change {#ss:im sbc}
If $\pi$ is a globally generic automorphic representations of $U_n,$ we denote the stable base change lift by ${{\operatorname}{sbc}}(\pi).$ The image of this lifting has been characterized in [@GRSBook], Theorem 11.2.
\[thm: image of sbc\] An automorphic representation of ${{\mathrm{GL}}}_n({\mathbb{A}}_E)$ is in the image of ${{\operatorname}{sbc}}$ if and only if it is an isobaric sum of distinct cuspidal representations $\tau_1 \boxplus \dots \boxplus \tau_r$ of ${{\mathrm{GL}}}_{n_1}({\mathbb{A}}_E),
\dots {{\mathrm{GL}}}_{n_2}({\mathbb{A}}_E)$ such that $L^S(s, \tau_i, {{\operatorname}{Asai}}^{(-)^n+1})$ has a pole at $s=1$ for all $i$ (this pole is necessarily simple).
[As explained in [@GRSBook], $L^S(s, \pi, {{\operatorname}{Asai}}^+) L^S(s, \pi, {{\operatorname}{Asai}}^-)$ is the partial Rankin-Selberg convolution $L$-function $L^S(s, \pi \times \pi\circ {\operatorname{Fr}})$ of $\pi$ and the representation obtained by composing $\pi$ with the nontrivial element ${\operatorname{Fr}}$ of ${{\mathrm{Gal}}}(E/F).$ It follows that at most one of $L^S(s, \pi, {{\operatorname}{Asai}}^+),$ and $L^S(s, \pi, {{\operatorname}{Asai}}^-)$ may have a pole at $s=1.$ Moreover, since both are nonvanishing on the line ${{\mathrm{Re}}}(s)=1$ by theorem 5.1 of [@Sha81], it follows that if either has a pole then $\pi\circ {\operatorname{Fr}}= {\widetilde}\pi.$ This implies that the central character of $\pi$ is trivial on ${\mathbb{A}}^\times \subset {\mathbb{A}}_E^\times.$ Finally, if ${\widetilde}\pi \cong \pi \circ {\operatorname{Fr}},$ then either ${{\operatorname}{Asai}}^+$ or ${{\operatorname}{Asai}}^-$ has a pole.]{}
Rogawski’s liftings
-------------------
### Two families of $L$-homomorphisms
We describe two families of $L$-homomorphisms introduced in [@RogawskiBook].
Fix $\mu: E^\times {\backslash}{\mathbb{A}}_E^\times \to {\mathbb{C}}^\times$ satisfying $\mu|_{{\mathbb{A}}_F^\times} = \chi_{E/F},$ and fix $w_0 \in W_F \smallsetminus W_E.$ There is a unique $L$-homomorphism $\xi^{(2,1)}_{w_0, \mu}: {}^L({{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1)\to {}^L{{\mathrm {U}}}_{2,1}$ such that $$\xi^{(2,1)}_{w_0, \mu}\left( {\begin{pmatrix}}a&b\\ c&d {\end{pmatrix}}, t \right)
= {\begin{pmatrix}}a&&b\\&t&\\c&&d {\end{pmatrix}},\qquad \xi_{w_0, \mu}(w_0) =
{\begin{pmatrix}}1&&\\&1&\\&&-1 {\end{pmatrix}}w_0,$$ and $$\xi^{(2,1)}_{w_0, \mu}(w) =
{\begin{pmatrix}}\mu(w)&&\\&1&\\&&\mu(w) {\end{pmatrix}}w, \qquad w \in W_E.$$ There is a unique $L$-homomorphism $\xi_{w_0, \mu}^{(1,1)}:{}^L({{\mathrm {U}}}_1\times {{\mathrm {U}}}_1 ) \to {}^L{{\mathrm {U}}}_{1,1}$ such that $$\xi^{(1,1)}_{w_0, \mu}\left( a,b \right)
= {\begin{pmatrix}}a&\\&b {\end{pmatrix}},\qquad \xi_{w_0, \mu}(w_0) =
{\begin{pmatrix}}&-1\\1& {\end{pmatrix}}w_0,$$ and $$\xi^{(1,1)}_{w_0, \mu}(w) =
{\begin{pmatrix}}\mu^{-1}(w)&\\&\mu^{-1}(w) {\end{pmatrix}}w, \qquad w \in W_E.$$
### Rogawski’s Liftings
Each of the $L$-homomorphisms $\xi^{(2,1)}_{w_0, \mu}$ determines a conjectural functorial transfer map, taking automorphic $L$-packets on ${{\mathrm {U}}}_{1,1}({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}})$ to automorphic $L$-packets on $U_{2,1}({\mathbb{A}}).$ Likewise, each of the $L$-homomorphisms $\xi^{(1,1)}_{w_0, \mu}$ determines a conjectural functorial transfer map, taking automorphic $L$-packets on ${{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}})$ to automorphic $L$-packets on ${{\mathrm {U}}}_{1,1}({\mathbb{A}}).$ The existence of these transfer maps is proved in [@RogawskiBook]. Varying the choice of $w_0$ does not change the transfer mapping. Varying the choice of $\mu$ permutes the elements of the image.
An automorphic representation of ${{\mathrm {U}}}_{1,1}({\mathbb{A}})$ or of ${{\mathrm {U}}}_{2,1}({\mathbb{A}})$ is said to be endoscopic if it is in the image of the Rogawski liftings.
Applying the lifting attached to $\xi^{(2,1)}_{w_0, \mu}$ to a packet obtained from $\xi^{(1,1)}_{w_0, \mu}$ gives a packet on ${{\mathrm {U}}}_{2,1}({\mathbb{A}}).$ This construction is functorial. We describe the associated $L$-homomorphism.
Define $$\xi^{(1,1,1)}(a,b,c) = {\begin{pmatrix}}a&&\\&b&\\&&c {\end{pmatrix}},
\qquad
\xi^{(1,1,1)}(w) = m(w) w, \qquad m(w) :=\begin{cases}
I, w \in W_E, \\J, w \notin W_E.
\end{cases}$$ (Here $I$ is the identity matrix and $J$ is the matrix with ones on the diagonal from top right to lower left and zeros elsewhere.)
Take $\eta_1, \eta_2, \eta_3$ three automorphic characters of ${{\mathrm {U}}}_1({\mathbb{A}}).$ Let $\pi_1$ be the representation of ${{\mathrm {U}}}_{1,1}({\mathbb{A}})$ obtained from $\eta_1 \otimes \eta_3$ using the Rogawski lifting attached to $\xi^{(1,1)}_{w_0, \mu}$ and let $\pi$ be the representation of ${{\mathrm {U}}}_{2,1}({\mathbb{A}})$ obtained from $\pi_1\otimes \eta_2$ using the Rogawski lifting attached to $\xi^{(2,1)}_{w_0, \mu}.$ Then $\pi$ is the weak functorial lift of $\eta_1\otimes \eta_2\otimes \eta_3$ relative to the $L$-homomorphism $\xi^{(1,1,1)}.$
[Note that $\xi_{w_0,\mu}^{(1,1)}$ and $\xi_{w_0,\mu}^{(2,1)}$ depend on the choice of $w_0$ and $\mu,$ but $\xi^{(1,1,1)}$ does not.]{}
Let $v$ be a finite unramified place. Each of our representations is determined by an unramified $L$-homomorphism from $W_{F_v}$ into the relevant $L$-group. Fix $w$ an element of $W_{F_v}$ of norm $1.$ Then each unramified $L$-homomorphism from $W_{F_v}$ is determined by its image on $w.$ We regard $W_{F_v}$ as a subgroup of $W_F$ by some choice of embedding as in [@Tate-NumberTheoreticBackground]. Let us refer to this image as the Satake parameter of the representation. At a split place, the Satake parameter of $\eta_i$ is $(t_i,w).$ From considering the isomorphism $W_F/W_E \cong {{\mathrm{Gal}}}(E/F)$ we see that $w\in W_E$ if and only if $E$ splits over $v.$ When this is not the case, $t_i$ must be $1$ for all $i.$
First assume $v$ is split. Then the Satake parameter of $\pi_1$ is ${\left(\begin{smallmatrix}}t_1\mu^{-1}(w)&\\ & t_3 \mu^{-1}(w) {\end{smallmatrix}\right)}w.$ Hence, the Satake parameter of $\pi_1 \otimes \eta_2$ is $({\left(\begin{smallmatrix}}t_1\mu^{-1}(w)&\\ & t_3 \mu^{-1}(w) {\end{smallmatrix}\right)}, \ t_2)w,$ and that of $\pi$ is $${\begin{pmatrix}}t_1\mu^{-1}(w)&&\\&t_2&\\ && t_3 \mu^{-1}(w) {\end{pmatrix}}{\begin{pmatrix}}\mu(w) &&\\&1& \\ && \mu(w) {\end{pmatrix}}w
= {\begin{pmatrix}}t_1&&\\&t_2&\\ && t_3 {\end{pmatrix}}w.$$
At an inert place, the Satake parameter of $\pi_1$ is ${\left(\begin{smallmatrix}}&-1\\ 1& {\end{smallmatrix}\right)}w,$ that of $\pi_1\times \eta_2$ is $({\left(\begin{smallmatrix}}&-1\\ 1& {\end{smallmatrix}\right)}, 1)w,$ and that of $\pi$ is $${\begin{pmatrix}}&&-1\\ &1&\\1&& {\end{pmatrix}}{\begin{pmatrix}}1&&\\&1&\\&&-1 {\end{pmatrix}}w
= Jw.$$
For $\pi_1$ a cuspidal representation of ${{\mathrm {U}}}_{1,1}({\mathbb{A}})$ and $\eta$ a character of ${{\mathrm {U}}}_1({\mathbb{A}})$ we denote the corresponding Rogawski lift by ${{\operatorname}{Rog}}^{(2,1)}_{w_0,\mu}(\pi_1 \otimes \eta).$ The lift attached to three characters $\eta_1, \eta_2, \eta_3$ is denoted ${{\operatorname}{Rog}}^{(1,1,1)}(\eta_1\otimes \eta_2\otimes \eta_3).$
### Rogawski liftings and stable base change
It’s clear from the formulae for $\xi^{(2,1)}_{w_0, \mu}$ and $\xi^{(1,1,1)}$ that $${{\operatorname}{sbc}}({{\operatorname}{Rog}}^{(2,1)}(\sigma\otimes \eta)
=({{\operatorname}{sbc}}(\sigma)\otimes \mu) \boxplus {{\operatorname}{sbc}}(\eta),
\quad
{{\operatorname}{sbc}}({{\operatorname}{Rog}}^{(1,1,1)}(\eta_1\otimes \eta_2 \otimes \eta_3))={{\operatorname}{sbc}}(\eta_1)\boxplus {{\operatorname}{sbc}}(\eta_2)\boxplus {{\operatorname}{sbc}}(\eta_3).$$
### Rogawski liftings and descent
For globally generic representations, the combination of the Kim-Krishnamurthy lifting and the method of descent gives an alternate construction of Rogawski’s endoscopic liftings, and a generalization. A representation is endoscopic if and only if its stable base change is non-cuspidal. In this situation, descent may be applied to each summand of the stable base change, and the original representation is the endoscopic lift of the collection of representations thus obtained. See [@GRSBook], section 11.3.
Recall that the isobaric summands of the stable base change of a cusp form are all distinct. From this we may deduce that if ${{\operatorname}{sbc}}(\eta_1), {{\operatorname}{sbc}}(\eta_2)$ and ${{\operatorname}{sbc}}(\eta_3)$ are not distinct, then ${{\operatorname}{Rog}}^{(1,1,1)}(\eta_1 \otimes \eta_2\otimes \eta_3)$ is not cuspidal.
These endoscopic liftings have also been studied by the trace formula method in [@Mok].
### Adjoint $L$-functions of Rogawski lifts
We regard ${{\mathrm{Ad}}}$ as an action of ${{\mathrm{GL}}}_3({\mathbb{C}}) \rtimes W_F$ on $\mathfrak{sl}_3{\mathbb{C}}$ by composing with the canonical projection to the finite Galois form. Thus $w \in W_F$ acts by $X \mapsto {}-{}_t\,\!X$ if $w \notin W_E$ and trivially if $w \in W_E.$ In this section we consider ${{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu}: {}^L({{\mathrm {U}}}_{1,1} \times {{\mathrm {U}}}_1) \to {{\mathrm{GL}}}( \mathfrak{sl}_3({\mathbb{C}})).$ Let $$X\left(
{\begin{pmatrix}}x_1 & x_2 \\ x_3 & - x_1 {\end{pmatrix}}, z, {\begin{pmatrix}}u_1\\ u_2 {\end{pmatrix}}, {\begin{pmatrix}}v_1&v_2{\end{pmatrix}}\right)
= {\begin{pmatrix}}x_1-z & u_1 & x_2 \\ v_1 & 2z & v_2\\ x_3 & u_2 & -x_1-z {\end{pmatrix}}.$$ Then for $x \in \mathfrak{sl}_2({\mathbb{C}}), z \in {\mathbb{C}}, u \in {{\operatorname}{Mat}}_{2\times 1}({\mathbb{C}}), v \in {{\operatorname}{Mat}}_{1\times 2}({\mathbb{C}}),$ we have $$\begin{aligned}
{{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu}(g,t). X(x, z,u,v)
&= X( gxg^{-1} , z, gut^{-1}, tvg^{-1}), & (g \in {{\mathrm{GL}}}_2({\mathbb{C}}), \ t \in {\mathbb{C}}^\times)
\\
{{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu}(w).X(x,z,u,v)
&= X(x,z, \mu(w) u , \mu(w)^{-1} v),& (w \in W_E)\\
{{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu}(w_0).X(x,z,u,v)
&= X\left(x, -z, {\begin{pmatrix}}-v_2\\ v_1{\end{pmatrix}}, {\begin{pmatrix}}-u_2& u_1 {\end{pmatrix}}\right).
\end{aligned}$$ Note that $${\begin{pmatrix}}-v_2\\ v_1{\end{pmatrix}}= {\begin{pmatrix}}-1 &\\ &1 {\end{pmatrix}}{}_tv, \qquad \qquad
{\begin{pmatrix}}-u_2& u_1 {\end{pmatrix}}= {}_tu {\begin{pmatrix}}-1 &\\ &1 {\end{pmatrix}}$$
Take $\pi_1$ an irreducible automorphic representation of $U_{1,1}({\mathbb{A}})$ and $\eta$ an irreducible automorphic representation (necessarily a character) of ${{\mathrm {U}}}_1({\mathbb{A}}).$ Let $\pi = \pi_1 \otimes \eta.$ Let $S$ be a finite set of places of $F,$ including all archimedean places and all places where either $\pi_1$ or $\eta$ is ramified. Then $$\label{Ad of rog21}
\begin{aligned}
L^S( s, \pi, {{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu})
&= L^S(s, \pi_1, {{\mathrm{Ad}}}) L^S(s, \chi_{E/F})
L^S(s, \mu \otimes {{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1\otimes \eta^{-1}))\\
L^S( s, \pi, {{\mathrm{Ad}}}' \circ \xi^{(2,1)}_{w_0, \mu})
&= L^S(s, \pi_1, {{\mathrm{Ad}}}') \zeta^S(s)
L^S(s,\mu \otimes {{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1\otimes \eta^{-1}) )
\end{aligned}$$
The proofs of the two statements are parallel. We treat only the first. From the computations above we see that ${{\mathrm{Ad}}}\circ \xi^{(2,1)}_{w_0, \mu}$ is the direct sum of three irreducible components, corresponding to the variable $x,$ the variable $z,$ and the pair $(u,v).$ These three components give rise to the three factors above: we match local $L$-factors at both split and inert unramified finite places. We discuss only the third component in detail, as the first two are easier. Denote this representation $r_{\mu, w_0}.$ If $v$ is split then $\pi_1$ gives a diagonal matrix ${\begin{pmatrix}}t_1&\\ & t_2 {\end{pmatrix}}$ and $\eta$ gives a nonzero scalar $c.$ We take $w \in W_{F_v}$ with ${{\mathrm{ord}}}_{W_{F_v}}(w) = 1.$ Since $G$ is split at $v,$ ${{\mathrm{Gal}}}( {\overline}F_v/F_v)$ acts trivially on $T^\vee({\mathbb{C}}).$ So, the image of $w$ in $W_F$ lies in $W_E.$ With respect to a suitable basis, the matrix of the operator $$r_{\mu, w_0}\left( \left({\begin{pmatrix}}t_1&\\ & t_2 {\end{pmatrix}}, \ c\right), \ w\right)$$ is the matrix ${{\mathrm{diag}}}( \mu(w) \frac{t_1} c, \mu(w) \frac{t_2} c, \mu(w)^{-1} \frac{c}{t_1} , \mu(w)^{-1} \frac c{t_2} ).$ Tracing through the definitions, this is exactly the matrix attachd to $\mu \otimes {{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1 \otimes \eta^{-1}).$
If $v$ is inert then $\pi_1$ gives a diagonal matrix which is stable under the action of the Galois group, i.e., of the form ${\begin{pmatrix}}t&\\ &t^{-1} {\end{pmatrix}},$ while $\eta_v$ (an unramified character of a compact group) must be trivial. Write $\varphi_{t}$ for the corresponding unramified $L$-homomorphism $W_F \to G^\vee({\mathbb{C}}) \rtimes W_F.$ For $w \in W_F$ with ${{\mathrm{ord}}}_{W_F}(w) = 1,$ we have $$r_{w_0, \mu}\circ \varphi_{t}(w).( u, v) = \left(\mu(ww_0^{-1}) {\begin{pmatrix}}-t &\\ &t^{-1} {\end{pmatrix}}{}_tv, \ \ \
\mu^{-1}(ww_0^{-1})\ {}_tu {\begin{pmatrix}}-t^{-1} &\\ &t {\end{pmatrix}}\right)$$ Thus, the matrix of the operator $r_{w_0, \mu}\circ \varphi_{t}(w)$ relative to a suitable choice of basis is $${\begin{pmatrix}}0&0&0&-\mu(ww_0^{-1})t\\ 0&0& \mu(ww_0^{-1})t^{-1}\\
0&-\mu^{-1}(ww_0^{-1})t^{-1} &0&0\\
\mu^{-1}(ww_0^{-1})t&0&0&0
{\end{pmatrix}},$$ and the relevant $L$-factor is $$(1+t^2 q^{-2s})(1+t^{-2} q^{-2s})$$
The local $L$-factor for $L(s, {{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1\otimes \eta^{-1}))$ is $(1-t^2q^{-2s})(1-t^{-2} q^{-2s}).$ Twisting by $\mu$ flips the signs, as required, because $\mu$ is an extension of $\chi_{E/F}.$ Thus, $\mu$ maps the uniformizer $\varpi_v$ of $F_v$ to $1$ if $v$ splits and $-1$ if $v$ is inert. But for inert $v$ there is a unique completion $E_w$ of $E$ over $F_v$ and it has the same uniformizer.
The corresponding formulae for the lift from ${{\mathrm {U}}}_1\times {{\mathrm {U}}}_1\times {{\mathrm {U}}}_1$ are similar and proved in the same way.
Take $\eta_1, \eta_2,$ and $\eta_3$ be irreducible automorphic representations of ${{\mathrm {U}}}_1({\mathbb{A}}),$ and let $\eta$ denote the representation $\eta_1 \otimes \eta_2 \otimes \eta_3$ of ${{\mathrm {U}}}_1({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}}).$ Let $S$ be a finite set of places of $F,$ including all archimedean places and all places where any of $\eta_1, \eta_2,$ and $\eta_3$ is ramified. Let $T$ denote the set of places of $E$ lying above $S$ and let ${{\operatorname}{sbc}}$ denote the stable base change lifting from $U_1({\mathbb{A}}_F)$ to ${\mathbb{A}}_E^\times.$ Then $$\label{Ad of eta111 lift}
\begin{aligned}
L^S( s, \eta, {{\mathrm{Ad}}}\circ \xi^{(1,1,1)})
&= L^S(s, \chi_{E/F})^2
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_1}{\eta_2}))
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_1}{\eta_3}))
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_2}{\eta_3}))
\\
L^S( s, \eta, {{\mathrm{Ad}}}' \circ \xi^{(1,1,1)})
&= \zeta^S(s)^2
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_1}{\eta_2}))
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_1}{\eta_3}))
L^T(s, {{\operatorname}{sbc}}(\frac{\eta_2}{\eta_3}))
\end{aligned}$$
Conclusions
-----------
From formulas , and it’s easy to see that $L^S(s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}^+)$ has a pole at $s=1$ of order equal to the number of isobaric summands in ${{\operatorname}{sbc}}(\pi),$ while $L^S(s, \pi, {{\mathrm{Ad}}}')$ has a pole of one lower order. Also, $L^S(s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}^-)$ and $L^S(s, \pi, {{\mathrm{Ad}}})$ are holomorphic and nonvanishing at $s=1.$ Indeed, it suffices to see that the other $L$-functions appearing in , are holomorphic and nonvanishing on the line ${{\mathrm{Re}}}(s)=1.$ For the $L$-functions $L^T(s, {{\operatorname}{sbc}}(\eta_i/\eta_j))$ this follows from the fact that ${{\operatorname}{sbc}}(\eta_1), {{\operatorname}{sbc}}(\eta_2),$ and ${{\operatorname}{sbc}}(\eta_3)$ are distinct. For $L^S(s, \mu \otimes {{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1\otimes \eta^{-1}))$ it follows from the fact that ${{\operatorname}{AI}_{E/F}}{{\operatorname}{sbc}}(\pi_1\otimes \eta^{-1})$ is either an irreducible cuspidal automorphic representation of ${{\mathrm{GL}}}_4({\mathbb{A}}),$ or an isobaric sum of two irreducible cuspidal automorphic representations of ${{\mathrm{GL}}}_2({\mathbb{A}}),$ [@ArthurClozel].
\[thm: main for nonsplit case\] Take $\pi$ a globally generic irreducible cuspidal automorphic representation of ${{\mathrm {U}}}_{2,1}({\mathbb{A}}).$ Then $L^S(s, \pi, {{\mathrm{Ad}}})$ is holomorphic and nonvanishing at $s=1,$ while $L^S(s, \pi, {{\mathrm{Ad}}}')$ can have at most a double pole. More precisely, we have three sets of equivalent conditions.
1. \[nonzplit main: 3\] The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}')$ is holomorphic and nonvanishing at $s=1$
2. ${{\operatorname}{sbc}}(\pi)$ is cuspidal
3. $\pi$ is not endoscopic
4. $\pi$ is not $H'_\rho$-distinguished.
2. \[nonsplit main:(2,1)\] The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}')$ has a simple pole at $s=1$
2. ${{\operatorname}{sbc}}(\pi)$ is the isobaric sum of a character and a cuspidal representation of ${{\mathrm{GL}}}_2({\mathbb{A}}_E)$
3. $\pi$ is an endoscopic lift from ${{\mathrm {U}}}_{1,1}({\mathbb{A}}) \times {{\mathrm {U}}}_1({\mathbb{A}}),$ but not from ${{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}).$
3. \[nonsplit main:(1,1,1)\] The following are equivalent:
1. $L^S(s, \pi, {{\mathrm{Ad}}}')$ has a double pole at $s=1$
2. ${{\operatorname}{sbc}}(\pi)$ is the isobaric sum of three characters of ${\mathbb{A}}_E^\times,$
3. $\pi$ is an endoscopic lift from ${{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}) \times {{\mathrm {U}}}_{1}({\mathbb{A}}).$
Other poles
-----------
The previous theorem gives fairly complete results for $L^S(s, \pi, {{\mathrm{Ad}}}' \times \chi)$ when $\chi$ is trivial or $\chi_{E/F}.$ Combining it with our earlier result, we have a gap: we do not know whether $L^S(s, \pi, {{\mathrm{Ad}}}' \times \chi)$ can have a pole at $s=1$ when $\chi$ is cubic, or when $\chi$ is a quadratic character other than $\chi_{E/F}.$ It turns out that the best way to proceed is by cases, based on the number of isobaric summands in the stable base change lift ${{\operatorname}{sbc}}(\pi).$
\[thm:other poles; type3\] Let $\pi$ be a globally generic, irreducible cuspidal automorphic representation of ${{\mathrm {U}}}_{2,1}({\mathbb{A}})$ such that ${{\operatorname}{sbc}}(\pi)$ is cuspidal, and $\chi$ a character of ${\mathbb{A}}^\times.$ Then the following are equivalent
1. \[ad’ pole\]$L^S(s, \pi, {{\mathrm{Ad}}}'\times \chi) $ has a pole at $s=1$
2. \[asai+ pole\]$L^S(s, {{\operatorname}{sbc}}(\pi) \otimes {\widetilde}\chi, {{\operatorname}{Asai}}^+)$ has a pole at $s=1$ for some/any character $\widetilde \chi$ of ${\mathbb{A}}_E^\times$ whose restriction to ${\mathbb{A}}^\times$ is $\chi,$
3. \[twist in im sbc\]$ {{\operatorname}{sbc}}(\pi) \otimes {\widetilde}\chi$ is itself in the image of stable base change for some/any character $\widetilde \chi$ of ${\mathbb{A}}_E^\times$ whose restriction to ${\mathbb{A}}^\times$ is $\chi.$
The meaning of is the same if “some” is replaced by “any” because $$\label{Asai+chi/chi = Ad'chi}
\frac{ L^S(s, {{\operatorname}{sbc}}(\pi) \otimes {\widetilde}\chi, {{\operatorname}{Asai}}^+)}{L^S(s, \chi)}= \frac{L^S(s, {{\operatorname}{sbc}}(\pi), {{\operatorname}{Asai}}^+ \times \chi)}{L^S(s, \chi)}
=
L^S(s, \pi, {{\mathrm{Ad}}}'\times \chi) ,$$ for any ${\widetilde}\chi: {\mathbb{A}}^\times_E \to {\mathbb{C}}^\times$ with ${\widetilde}\chi\big|_{{\mathbb{A}}^\times} = \chi.$ The meaning of is the same if “some” is replaced by “any” because a character of ${\mathbb{A}}_E^\times$ which is trivial on ${\mathbb{A}}^\times$ is in the image of stable base change from ${{\mathrm {U}}}_1({\mathbb{A}}),$ and twisting by such a character preserves the image of stable base change from ${{\mathrm {U}}}_{2,1}({\mathbb{A}}).$
Equation also makes the first equivalence of and clear. The equivalence of and follows from theorem \[thm: image of sbc\].
[If $ {{\operatorname}{sbc}}(\pi) \otimes {\widetilde}\chi$ is in the image of stable base, then $ {{\operatorname}{sbc}}(\pi)$ and $ {{\operatorname}{sbc}}(\pi) \otimes {\widetilde}\chi$ both have central characters which are trivial on ${\mathbb{A}}^\times {\hookrightarrow}{\mathbb{A}}_E^\times.$ This implies that $\chi^3 =1.$ Thus, this case is very similar to the split case. ]{}
\[thm:other poles: type 2+1\] Let $\pi$ be a globally generic, irreducible cuspidal automorphic representation of ${{\mathrm {U}}}_{2,1}({\mathbb{A}})$ such that ${{\operatorname}{sbc}}(\pi)=\pi_1 \boxplus \chi_1$, for some irreducible cuspidal automorphic representation $\pi_1$ of ${{\mathrm{Res}}}_{E/F}{{\mathrm{GL}}}_2({\mathbb{A}})$ and character $\chi_1$ of ${\mathbb{A}}_E^\times$, and let $\chi$ be a nontrivial character of ${\mathbb{A}}^\times.$ So $\pi = {{\operatorname}{Rog}}^{(2,1)}_{w_0, \mu}(\sigma \otimes \eta_1)$, $\chi_1 = {{\operatorname}{sbc}}(\eta_1)$ and $\pi_1 = \mu \otimes {{\operatorname}{sbc}}( \sigma),$ for some irreducible cuspidal representation $\sigma$ of ${{\mathrm {U}}}_{1,1}({\mathbb{A}}),$ and character $\eta_1$ of ${\mathbb{A}}_E^\times.$ Then the following are equivalent
1. $L^S(s, \pi, {{\mathrm{Ad}}}' \times \chi)$ has a pole at $s=1$
2. $L^S(s, \sigma, {{\mathrm{Ad}}}' \times \chi)$ has a pole at $s=1$
3. $L^S(s, \pi_1, {{\operatorname}{Asai}}^- \times \chi)$ (which equals $L^S(s, \pi_1\otimes {\widetilde}\chi, {{\operatorname}{Asai}}^-)$ for any ${\widetilde}\chi: {\mathbb{A}}^\times_E \to {\mathbb{C}}^\times$ with ${\widetilde}\chi\big|_{{\mathbb{A}}^\times} = \chi$) has a pole at $s=1$
4. ${{\operatorname}{sbc}}(\sigma) \otimes{\widetilde}\chi$ is itself in the image of ${{\operatorname}{sbc}}$ for any ${\widetilde}\chi: {\mathbb{A}}^\times_E \to {\mathbb{C}}^\times$ with ${\widetilde}\chi\big|_{{\mathbb{A}}^\times} = \chi.$
If we twist all the $L$-functions in by $\chi$ then all are holomorphic and nonvanishing except possibly for $L^S(s, \sigma, {{\mathrm{Ad}}}' \times \chi).$ This proves the first equivalence. The second equivalence follows from twisting . The third follows from theorem \[thm: image of sbc\].
[As before, if ${{\operatorname}{sbc}}(\sigma) \otimes{\widetilde}\chi$ is itself in the image of ${{\operatorname}{sbc}},$ then $\chi^2=1.$ ]{}
To analyze the case when ${{\operatorname}{sbc}}(\pi)$ is the isobaric sum of three characters, i.e., that $\pi={{\operatorname}{Rog}}^{(1,1,1)}(\eta_1\otimes \eta_2\otimes \eta_3),$ we need to note that $L^T (s, {{\operatorname}{sbc}}(\eta_i/\eta_j)=L^S(s, {\operatorname}{AI}_{E/F}{{\operatorname}{sbc}}(\frac{\eta_i}{\eta_j})\times \chi).$ $$\label{twisted Ad' for 111}
L^S(s, \pi, {{\mathrm{Ad}}}' \times \chi) = L^S(s, \chi)^2
\prod_{1\le i < j \le 3}
L^S(s, {\operatorname}{AI}_{E/F}{{\operatorname}{sbc}}(\frac{\eta_i}{\eta_j})\times \chi).$$
By [@ArthurClozel], ${\operatorname}{AI}_{E/F}{{\operatorname}{sbc}}(\frac{\eta_i}{\eta_j})$ is a cuspidal representation of ${{\mathrm{GL}}}_2({\mathbb{A}})$ unless ${{\operatorname}{sbc}}(\frac{\eta_i}{\eta_j})$ is also in the image of the lifting ${{\operatorname}{bc}}.$
If a character ${\widetilde}\chi$ of ${\mathbb{A}}_E^\times$ is in the image of stable base change from ${{\mathrm {U}}}_1^{E/F}({\mathbb{A}}),$ it is in the image of base change from ${\mathbb{A}}^\times,$ if and only if it is quadratic.
Write ${\operatorname{Fr}}$ for the nontrivial element of ${{\mathrm{Gal}}}(E/F).$ We view it as an automorphism of ${\mathbb{A}}_E^\times.$ A character is in the image of base change if and only if it satisfies ${\widetilde}\chi \circ {\operatorname{Fr}}={\widetilde}\chi.$ (Cf. [@ArthurClozel Theorem 4.2, 5.1].) It is in the image of stable base change if and only if its restriction to ${\mathbb{A}}^\times$ is trivial. (This is a special case of the characterization in section \[ss:im sbc\]: note that for ${{\mathrm{Res}}}_{E/F}{{\mathrm{GL}}}_1,$ we have $L^S(s, \chi, {{\operatorname}{Asai}}) = L^S(s, \chi|_{{\mathbb{A}}^\times}).$) Now, given that ${\widetilde}\chi$ is trivial of ${\mathbb{A}}^\times$ we get ${\widetilde}\chi(a {\operatorname{Fr}}(a))=1$ for all $a \in {\mathbb{A}}_E^\times.$ Hence ${\widetilde}\chi \circ {\operatorname{Fr}}= {\widetilde}\chi^{-1}.$ Hence ${\widetilde}\chi \circ {\operatorname{Fr}}= {\widetilde}\chi \iff {\widetilde}\chi = {\widetilde}\chi^{-1}.$
\[prop: other poles, type 1+1+1\] The expression can have at most a simple pole. It has a simple if ${{\operatorname}{sbc}}\chi_i/\chi_j = {{\operatorname}{bc}}\chi$ for some (necessarily unique) $1 \le i < j \le 3.$
This follows from the previous lemma. We must show that the equality ${{\operatorname}{sbc}}\chi_i/\chi_j = {{\operatorname}{bc}}\chi$ can not hold for more than one pair $(i,j).$ Since we know that the components ${{\operatorname}{sbc}}(\chi_i) , 1\le i \le j$ are distinct, it follows that ${{\operatorname}{sbc}}\chi_1/\chi_3 \ne {{\operatorname}{sbc}}\chi_1/\chi_2, {{\operatorname}{sbc}}\chi_2/\chi_3.$
Moreover, if ${{\operatorname}{sbc}}\chi_1/\chi_2 = {{\operatorname}{sbc}}\chi_2/\chi_3 = \xi,$ then $\xi$ must not be quadratic, for if it were, then ${{\operatorname}{sbc}}\chi_1$ would equal ${{\operatorname}{sbc}}\chi_3.$ But then, since $\xi$ is in the image of ${{\operatorname}{sbc}}$ and not quadratic, it is not in the image of ${{\operatorname}{bc}}.$
[XXXX]{}
J. Arthur and L. Clozel, [*Simple algebras, base change, and the advanced theory of the trace formula*]{}, Annals of Math. Studies, no.120, Princeton University Press, (1989).
J. Bernstein, A.V. Zelevinski, [*Representations of the group ${{\mathrm{GL}}}(n,F)$, where $F$ is non-archimedian local field,*]{} Russian Math Surveys [**31:3**]{}, (1976), 1-68.
J. Bernstein, A.V. Zelevinski, [*Induced representations of reductive ${{\mathfrak{p}}}$-adic groups. I*]{}, Annales Scientifiques de l.É.N.S, **10** (1977), 441-472.
A. Borel, [*Automorphic L-functions. Automorphic forms, representations and L-functions*]{} (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 2, pp. 27–61, Proc. Sympos. Pure Math., XXXIII, Amer. Math. Soc., Providence, R.I., 1979.
D. Bump, D. Ginzburg, [*The adjoint L-function of $GL(4).$*]{} J. Reine Angew. Math. **505** (1998), 119–172.
J. Buttcane, F. Zhou, [*Plancherel distribution of Satake parameters of Maass cusp forms on $GL_3$*]{}, International Mathematics Research Notices, published online, https://doi.org/10.1093/imrn/rny061
W. Casselman, J. Shalika, [*The unramified principal series of $p$-adic groups, II, the Whittaker functions*]{}, Compositio Math. **41** (1980), 207-231.
B. Chang, [*The Conjugate Classes of Chevalley Groups of Type ($G_2$),*]{} J. Algebra, **9**, (1968), 190-211.
J. Cogdell. I.I. Piatetski-Shapiro, [*Derivatives and $L$-functions for ${{\mathrm{GL}}}_n$*]{}, in “Representation Theory, Number Theory and Invariant Theory", pp115-173, Progress in Mathematics, Vol 323, 2017.
Y. Flicker, [*The adjoint representation $L$-function for $GL(n)$*]{}, Pacific J. Math. **154** (1992), no. 2, 231–244.
S. Gelbart and H. Jacquet, [*A relation between automorphic representations of $GL(2)$ and $GL(3)$.*]{} Ann. Sci. École Norm. Sup. (4) **11** (1978), no. 4, 471–542.
S. S. Gelbart and A. W. Knapp, [*L-indistinguishability and R groups for the special linear group,*]{} Adv. in Math. **43** (1982), 101-121.
S. Gelbart, J. Rogawski, D. Soudry, [*On periods of cusp forms and algebraic cycles for ${{\mathrm {U}}}(3)$*]{}, Israel J. Math. **83** (1993), no. 1-2, 213–252.
S. Gelbart, J. Rogawski, D. Soudry, [*Endoscopy, theta-liftings, and period integrals for the unitary group in three variables.* ]{} Ann. of Math. (2) **145** (1997), no. 3, 419–476.
D. Ginzburg, [*A Rankin-Selberg integral for the adjoint representation of ${{\mathrm{GL}}}_3$*]{}, Invent. math. **105** (1991), 571-588.
D. Ginzburg, J. Hundley, [*Multivariable Rankin-Selberg integrals for orthogonal groups*]{} Int. Math. Res. Not. 2004, no. **58**, 3097-3119.
D. Ginzburg, J. Hundley, [*On Spin L-functions for $GSO_{10}$,* ]{} J. Reine Angew. Math. **603** (2007), 183-213.
D. Ginzburg, J. Hundley, [*The adjoint L-function for ${{\mathrm{GL}}}_5.$*]{} (English summary) Electron. Res. Announc. Math. Sci. **15** (2008), 24–32.
D. Ginzburg, D. Jiang, [*Siegel-Weil identity for $G_2$ and poles of $L$-functions*]{}, Journal of Number Theory **82** no.2(2000), 256-287.
D. Ginzburg, S. Rallis, and D. Soudry, [*Periods, poles of L-functions and symplectic-orthogonal theta lifts*]{}, J. Reine Angew. Math. **487** (1997), 85–114.
D. Ginzburg, S. Rallis, and D. Soudry, [*The descent map from automorphic representations of ${{\mathrm{GL}}}(n)$ to classical groups*]{}, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2011.
R. Godement, H. Jacquet, [*Zeta functions of simple algebras*]{}, Lecture Notes in Mathematics, **260**, Springer-Verlag, 1972.
J. Hundley, [*The adjoint L-function of ${{\mathrm {S}}}{{\mathrm {U}}}_{2,1}$*]{}, Multiple Dirichlet series, L-functions and automorphic forms, 193-204, Progr. Math., 300, Birkhauser/Springer, New York, 2012.
J. Hundley, [*Holomorphy of adjoint $L$-functions for quasi-split $A_2$*]{}, Research in Number Theory **4**:44 (2018),
J. Hundley and E. Sayag, [*Descent construction for GSpin groups*]{}, Mem. Amer. Math. Soc. 243 (2016), no. 1148,
H. Jacquet, [*Generic representations*]{}, in [*Non-commutative harmonic analysis (Actes Colloq., Marseille-Luminy, 1976)*]{}, pp. 91–101. Lecture Notes in Math., Vol. 587, Springer, Berlin, 1977.
H. Jacquet, [*A correction to Conducteur des représentations du groupe linéaire*]{}, Pacific Journal of Math, **260** (2012), 515-525.
H. Jacquet and J.A. Shalika, [*On Euler products and the classification of automorphic representations. I.*]{} Amer. J. Math. **103** (1981), no. 3, 499–558.
H. Jacquet and J.A. Shalika, Exterior square $L$ functions, Automorphic Forms, Shimura Varieties and $L$-Functions, Vol.2 Academic Press (1990), 143-226.
H. Jacquet, I. Piatetski-Shapiro, and J. Shalika, [*Automorphic forms on ${{\mathrm{GL}}}(3)$, I*]{}, Annals of Math, second series, **109**, No. 1 (1979), 169-212.
H. Jacquet, I. Piatetski-Shapiro, and J. Shalika, [*Conducteur des représentations du groupe linéaire*]{}, Math. Ann., **256**(2) (1981), 199-214.
H. Jacquet, I. Piatetski-Shapiro, and J. Shalika, [*Rankin-Selberg convolutions*]{}, American Journal of Mathematics **105** No. 2 (1983), 367-464.
H. Jacquet and D. Zagier [*Eisenstein series and the Selberg trace formula. II.*]{} Trans. Amer. Math. Soc. **300** (1987), no. 1, 1–48.
D. Jiang and S. Rallis, [*Fourier coefficients of Eisenstein series of the exceptional group of type $G_2$*]{}, Pacific J. Math. **181** (1997), 281-314.
H. Kim and M. Krishnamurthy, [*Base change lift for odd unitary groups*]{}. Functional analysis VIII, 116–125, Various Publ. Ser. (Aarhus), 47, Aarhus Univ., Aarhus, 2004.
H. Kim and M. Krishnamurthy, [*Stable base change lift from unitary groups to ${{\mathrm{GL}}}_n$*]{}. IMRP Int. Math. Res. Pap. 2005, no. 1, 1–52 N. Matringe, [*Essential Whittaker functions for ${{\mathrm{GL}}}(n)$*]{}, Doc. Math. **18** (2013), 1191-1214.
M. Miyauchi, [*On local new forms for unramified ${{\mathrm {U}}}(2,1)$*]{}, Manuscripta Mathematica, **141**, Issue 1 (2013), 149-169.
M. Miyauchi, [*Whittaker functions associated to newforms for GL(n) over p-adic fields,*]{} Journal of the Mathematical Society of Japan, **66** (2014), 17-24. C.P. Mok, [*Endoscopic classification of representations of quasi-split unitary groups.* ]{} Mem. Amer. Math. Soc. 235 (2015), no. 1108,
R. Ree, [*A Family of Simple Groups Associated with the Simple Lie Algebra of Type ($G_2$)*]{}, American Journal of Mathematics, Vol. **83**, No. 3 (Jul., 1961), pp. 432-462
J. Rogawski, [*Automorphic representations of unitary groups in three variables*]{}, Annals of Mathematics Studies, 123. Princeton University Press, Princeton, NJ, 1990.
A. Selberg, [*Old and new conjectures and results about a class of Dirichlet series*]{}, Proceedings of the Amalfi Conference on Analytic Number Theory (Maiori, 1989), 367–385, Univ. Salerno, Salerno, 1992.
F. Shahidi, [*Functional equations satisfied by certain L-functions*]{}, Compositio Math. **37** 2 (1978), 171-207.
G. Shimura, [*On the holomorphy of certain Dirichlet series*]{}, Proc. London Math. Soc. (3) **31** (1975), no. 1, 79–98.
T. Shintani, [*On an explicit formula for class-1 “Whittaker functions" on ${{\mathrm{GL}}}_n$ over ${{\mathfrak{P}}}$-adic fields*]{}. Proc. Japan Acad., **52**(1976) 180–182.
F. Shahidi, [*On Certain $L$-functions*]{}, Am. J. Math **103**, (1981), no. 2, pp. 297-355.
F. Shahidi, [*A proof of Langlands’ conjecture on Plancherel measures; complementary series for $p$-adic groups*]{}, Ann. of Math. (2) **132** (1990), no. 2, 273-330.
D. Soudry, [*On the Archimedean theory of Rankin-Selberg convolutions for $SO_{2l+1} \times GL_n$*]{}, Annales scientifiques de l’ É.N.S. 4$^e$ série, tome 28, n$^o$ 2 (1995), p. 161-224.
T.A. Springer, [*Linear Algebraic Groups, 2nd Ed.,*]{} Birkhäuser, 1998.
J. Tate, [*Number theoretic background*]{}, in Automorphic forms, representations and $L$-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 2, pp. 3–26, Proc. Sympos. Pure Math., XXXIII, Amer. Math. Soc., Providence, R.I., 1979. MR0546607
F. Tian, [*On the local theory of certain global zeta integrals and related problems*]{}, Thesis, University Minnesota, 2018.
F. Tian, [*On the Archimedean Local Gamma Factors for Adjoint Representation of ${{\mathrm{GL}}}3$, Part I*]{}, arXiv:1811.02752.
F. Tian, [*On the Archimedean Local Gamma Factors for Adjoint Representation of ${{\mathrm{GL}}}3$, Part II*]{}, preprint.
N. R. Wallach, [*Real Reductive Groups II*]{}, Volume in *Pure and Applied Mathematics*, Academic Press. Inc., (1992).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate different ways of generating approximate solutions to the pairwise Markov random field (MRF) selection problem. We focus mainly on the inverse Ising problem, but discuss also the somewhat related inverse Gaussian problem because both types of MRF are suitable for inference tasks with the belief propagation algorithm (BP) under certain conditions. Our approach consists in to take a Bethe mean-field solution obtained with a maximum spanning tree (MST) of pairwise mutual information, referred to as the *Bethe reference point*, for further perturbation procedures. We consider three different ways following this idea: in the first one, we select and calibrate iteratively the optimal links to be added starting from the Bethe reference point; the second one is based on the observation that the natural gradient can be computed analytically at the Bethe point; in the third one, assuming no local field and using low temperature expansion we develop a dual loop joint model based on a well chosen fundamental cycle basis. We indeed identify a subclass of planar models, which we refer to as *Bethe-dual graph models*, having possibly many loops, but characterized by a singly connected dual factor graph, for which the partition function and the linear response can be computed exactly in respectively O(N) and $O(N^2)$ operations, thanks to a dual weight propagation (DWP) message passing procedure that we set up. When restricted to this subclass of models, the inverse Ising problem being convex, becomes tractable at any temperature. Experimental tests on various datasets with refined $L_0$ or $L_1$ regularization procedures indicate that these approaches may be competitive and useful alternatives to existing ones.'
author:
- Cyril Furtlehner
- Yufei Han
- 'Jean-Marc Lasgouttes'
- Victorin Martin
bibliography:
- 'invising.bib'
title: Pairwise MRF Calibration by Perturbation of the Bethe Reference Point
---
Introduction
============
The problem at stake is a model selection problem, in the MRF families, where $N$ variables are observed pair by pair. The optimal solution is the MRF with maximal entropy obeying moment constraints. For binary variables, this happens then to be the Ising model with highest log-likelihood. It is a difficult problem, where both the graph structure and the values of the fields and couplings have to be found. In addition, we wish to ensure that the model is compatible with the fast inference algorithm “belief propagation” (BP) to be useful at large scale for real-time inference tasks. This leads us to look at least for a good trade-off between likelihood and sparsity.
Concerning the Inverse Ising Problem (IIP), the existing approaches fall mainly in the following categories:
- Purely computational efficient approaches rely on various optimization schemes of the log likelihood [@LeGaKo] or on pseudo-likelihood [@HoTi] along with sparsity constraints to select the only relevant features.
- Common analytical approaches are based on the Plefka expansion [@Plefka] of the Gibbs free energy by making the assumption that the coupling constants $J_{ij}$ are small. The picture is then of a weakly correlated uni-modal probability measure. For example, the recent approach proposed in [@CoMo] is based on this assumption.
- Another possibility is to assume that relevant coupling $J_{ij}$ have locally a tree-like structure. The Bethe approximation [@YeFrWe] can then be used with possibly loop corrections. Again this corresponds to having a weakly correlated uni-modal probability measure and these kind of approaches are referred to as pseudo-moment matching methods in the literature for the reason explained in the previous section. For example the approaches proposed in [@KaRo; @WeTe; @MoMe; @YaTa] are based on this assumption.
- In the case where a multi-modal distribution is expected, then a model with many attraction basins is to be found and Hopfield-like models [@Hopfield; @CoMoSe] are likely to be more relevant. To be mentioned also is a recent mean-field methods [@NgBe2] which allows one to find in some simple cases the Ising couplings of a low temperature model, i.e. displaying multiple probabilistic modes.
On the side of inverse Gaussian problem, not surprisingly similar methods have been developed by explicitly performing $L_0$ and $L_1$ matrix norm penalizations on the inverse covariance matrices, so as to determine sparse non-zero couplings in estimated inverse covariance matrices for large-scale statistical inference applications [@FrHaTi; @HSDR] where direct inversion is not amenable. In our context the goal is a bit different. In general cases, the underlying inverse covariance matrix is not necessarily sparse. What we aim to find is a good sparse approximation to the exact inverse covariance matrix. Furthermore, sparsity constraint is not enough for constructing graph structure to be used in conjunction with BP. Known sufficient conditions celled walk-summability [@MaJoWi] (WS) are likely to be imposed instead of (or in addition to) the sparsity constraint. To the best of our knowledge not much work is taking this point into consideration at the noticeable exception of [@AnTaHuWi] by restricting the class of learned graph structures. To complete this overview, let us mention also that some authors proposed information based structure learning methods [@NeBaSaSh] quite in line with some approaches to be discussed in the present paper.
In some preceding work dealing with a road traffic inference application, with large scale and real time specifications [@FuLaFo; @FuLaAu; @FuHaLa], we have noticed that these methods could not be used blindly without drastic adjustment, in particular to be compatible with belief propagation. This led us to develop some heuristic models related to the Bethe approximation. The present work is an attempt to give a theoretical basis and firmer ground to these heuristics and to develop new ones.
The paper is organized as follows: in Section \[sec:prelim\] we review some standard statistical physics approaches to the IIP, mainly based on perturbation expansions, and derive some new and useful expressions of susceptibility coefficients in Section \[sec:cumulant\]. In Section \[sec:iip\] we explain what we mean by the *Bethe reference point* and develop an iterative proportional scaling (IPS) based method to incrementally, link by link, refine this approximate solution, both for the inverse Ising and inverse GMRF problems. Section \[sec:perturbation\] explores a second way to refine the Bethe reference point, based on the Plefka’s expansion and on results of Section \[sec:cumulant\]. A third method, based on strong coupling expansion and leading to a dual weight propagation algorithm (DWP) is presented in Section \[sec:dwp\]. Merits of these methods differs, which makes them complementary to each other. The first one is particularly useful when the underlying graph structure is not given; the second one, by giving explicitly the natural gradient direction, can be used to reduce the number of parameters to tune; finally the third one can be fast and exact for given very sparse structure, assuming in addition no local fields. For sake of comparison, we explain in Section \[sec:penal\] how to adapt standard $L_0$ and $L_1$ norm-penalized sparsity inducing optimization framework for finding relevant approximate solutions to the inverse GMRF problem in our context. Comparison is then made in Section \[sec:experiments\] with our own IPS based approach, in addition to other numerical experiments illustrating performances of the other methods.
Preliminaries {#sec:prelim}
=============
Inverse Ising problem
---------------------
In this section we consider binary variables ($x_i\in\{0,1\}$), which at our convenience may be also written as spin variables $s_i = 2x_i-1\in\{-1,1\}$. We assume that from a set of historical observations, the empirical mean $\hat m_i$ (resp. covariance $\hat \chi_{ij}$) is given for each variable $s_i$ (resp. each pair of variable $(s_i,s_j)$). In this case, from Jayne’s maximum entropy principle [@Jaynes], imposing these moments to the joint distribution leads to a model pertaining to the exponential family, the so-called Ising model: $$\label{def:ising}
{\mathcal{P}}({\bf s}) = \frac{1}{Z[{{\bf J}},{{\bf h}}]}\exp\bigl(\sum_ih_is_i+\sum_{i,j}J_{ij}s_is_j\bigr)$$ where the local fields ${{\bf h}}=\{h_i\}$ and the coupling constants ${{\bf J}}=\{J_{ij}\}$ are the Lagrange multipliers associated respectively to mean and covariance constraints when maximizing the entropy of ${\mathcal{P}}$. They are obtained as minimizers of the dual optimization problem, namely $$\label{eq:iipb}
({{{\bf h}}}^\star,{{{\bf J}}}^\star) = \operatorname*{argmin}_{({{{\bf h}}},{{{\bf J}}})} {\mathcal{L}}[{{\bf h}},{{\bf J}}],$$ with $$\label{def:LL}
{\mathcal{L}}[{{\bf h}},{{\bf J}}] = \log Z[{{\bf h}},{{\bf J}}]-\sum_ih_i{\hat m}_i-\sum_{ij}J_{ij}\hat m_{ij}$$ the log likelihood. This leads to invert the linear response equations: $$\begin{aligned}
\frac{\partial \log Z}{\partial h_i}[{{\bf h}},{{\bf J}}] &= {\hat m}_i\label{eq:cei}\\[0.2cm]
\frac{\partial \log Z}{\partial J_{ij}}[{{\bf h}},{{\bf J}}] &= \hat m_{ij},\label{eq:ccij}\end{aligned}$$ $\hat m_{ij} = {\hat m}_i{\hat m}_j+{{\hat \chi}}_{ij}$ being the empirical expectation of $s_is_j$. As noted e.g. in [@CoMo], the solution is minimizing the cross entropy, a Kullback-Leibler distance between the empirical distribution ${{\hat P}}$ based on historical data and the Ising model: $$\label{eq:dkl}
D_{KL} [\hat{\cal P}\Vert {\cal P}] = \log Z[{{\bf h}},{{\bf J}}]-\sum_ih_i{\hat m}_i-\sum_{i<j}J_{ij}\hat m_{ij} - S(\hat{\cal P}).$$ The set of equations (\[eq:cei\],\[eq:ccij\]) cannot be solved exactly in general because the computational cost of $Z$ is exponential. Approximations resorting to various mean field methods can be used to evaluate $Z[{{\bf h}},{{\bf J}}]$.
#### Plefka’s expansion {#sec:plefka}
To simplify the problem, it is customary to make use of the Gibbs free energy, i.e. the Legendre transform of the free energy, to impose the individual expectations ${{\bf m}}= \{{\hat m}_i\}$ for each variable: $$G[{{\bf m}},{{\bf J}}] = {{\bf h}}^{T}({{\bf m}}){{\bf m}}+ F[{{\bf h}}({{\bf m}}),{{\bf J}}]$$ (with $F[{{\bf h}},{{\bf J}}] {\stackrel{\mbox{\upshape\tiny def}}{=}}-\log Z[{{\bf h}},{{\bf J}}]$, ${{\bf h}}^{T}{{\bf m}}$ is the ordinary scalar product) where ${{\bf h}}({{\bf m}})$ depends implicitly on ${{\bf m}}$ through the set of constraints $$\frac{\partial F}{\partial h_i} = -m_i.\label{eq:constraints}$$ Note that by duality we have $$\label{eq:dual1}
\frac{\partial G}{\partial m_i} = h_i({{\bf m}}),$$ and $$\label{eq:dual2}
\Bigl[\frac{\partial^2 G}{\partial m_i\partial m_j}\Bigr] = \left[\frac{d{{\bf h}}}{d{{{\bf m}}}}\right]_{ij}
=\left[\frac{d{{\bf m}}}{d{{{\bf h}}}}\right]^{-1}_{ij}
= -\Bigl[\frac{\partial^2 F}{\partial h_i\partial h_j}\Bigr]^{-1} = \bigl[\chi^{-1}\bigr]_{ij}.$$ i.e. the inverse susceptibility matrix. Finding a set of $J_{ij}$ satisfying this last relation along with (\[eq:dual1\]) yields a solution to the inverse Ising problem since the $m$’s and $\chi$’s are given. Still a way to connect the couplings directly with the covariance matrix is given by the relation $$\label{eq:eq2}
\frac{\partial G}{\partial J_{ij}} = - m_{ij}.$$ The Plefka expansion is used to expand the Gibbs free energy in power of the coupling $J_{ij}$ assumed to be small. Multiplying all coupling $J_{ij}$ by $\alpha$ yields the following cluster expansion: $$\begin{aligned}
G[{{\bf m}},\alpha{{\bf J}}] &= {{\bf h}}^{T}({{\bf m}},\alpha){{\bf m}}+ F[{{\bf h}}({{\bf m}},\alpha),\alpha{{\bf J}}]\label{eq:plefka1}\\[0.2cm]
&= G_0[{{\bf m}}] + \sum_{n=0}^\infty\frac{\alpha^n}{n!}G_n[{{\bf m}},{{\bf J}}]\label{eq:plefka2}\end{aligned}$$ where each term $G_n$ corresponds to cluster contributions of size $n$ in the number of links $J_{ij}$ involved, and ${{\bf h}}({{\bf m}},\alpha)$ depends implicitly on $\alpha$ in order to always fulfill (\[eq:constraints\]). This precisely is the Plefka expansion, and each term of the expansion (\[eq:plefka2\]) can be obtained by successive derivation of (\[eq:plefka1\]). We have $$G_0[{{\bf m}}] = \sum_i\frac{1+m_i}{2}\log\frac{1+m_i}{2}+\frac{1-m_i}{2}\log\frac{1-m_i}{2}.$$ Letting $$H_J {\stackrel{\mbox{\upshape\tiny def}}{=}}\sum_{i<j} J_{ij}s_is_j,$$ using (\[eq:constraints\]), the two first derivatives of (\[eq:plefka1\]) w.r.t $\alpha$ read $$\begin{aligned}
\label{eq:dG1}
\frac{dG[{{\bf m}},\alpha{{\bf J}}]}{d\alpha} &= -{\mathbb E}_\alpha\bigl(H_J\bigr), \\
\label{eq:dG2}
\frac{d^2G[{{\bf m}},\alpha{{\bf J}}]}{d\alpha^2} &= -\operatorname{Var}_\alpha\bigl(H_J\bigr)
-\sum_i\frac{dh_i({{\bf m}},\alpha)}{d\alpha}\operatorname{Cov}_\alpha\bigl(H_J,s_i\bigr),\end{aligned}$$ where subscript $\alpha$ indicates that expectations, variance and covariance are taken at given $\alpha$. To get successive derivatives of ${{\bf h}}({{\bf m}},\alpha)$ one can use (\[eq:dual1\]). Another possibility is to express the fact that ${{\bf m}}$ is fixed, $$\begin{aligned}
\frac{dm_i}{d\alpha} = 0 & = -\frac{d}{d\alpha}\frac{\partial F[{{\bf h}}(\alpha),\alpha{{\bf J}}]}{\partial h_i}\nonumber\\[0.2cm]
& = \sum_{j}h'_j(\alpha)\operatorname{Cov}_\alpha(s_i,s_j)+\operatorname{Cov}_\alpha(H_J,s_i), \end{aligned}$$ giving $$\label{eq:dha}
h'_i(\alpha) = - \sum_j[\chi_\alpha^{-1}]_{ij}\operatorname{Cov}_\alpha(H_J,s_j),$$ where $\chi_\alpha$ is the susceptibility delivered by the model when $\alpha\ne 0$. To get the first two terms in the Plefka expansion, we need to compute these quantities at $\alpha=0$: $$\begin{aligned}
\operatorname{Var}\bigl(H_J\bigr) &= \sum_{i<k,j} J_{ij}J_{jk}m_im_k(1-m_j^2)+\sum_{i<j}J_{ij}^2(1-m_i^2m_j^2),\\[0.2cm]
\operatorname{Cov}\bigl(H_J,s_i\bigr) &= \sum_{j}J_{ij}m_j(1-m_i^2),\\[0.2cm]
h'_i(0) &= - \sum_{j} J_{ij}m_j,\end{aligned}$$ (by convention $J_{ii}=0$ in these sums). The first and second orders then finally read: $$G_1[{{\bf m}},{{\bf J}}] = -\sum_{i<j}J_{ij}m_im_j,\qquad\qquad G_2[{{\bf m}},{{\bf J}}] = -\sum_{i<j}J_{ij}^2(1-m_i^2)(1-m_j^2),$$ and correspond respectively to the mean field and to the TAP approximation. Higher order terms have been computed in [@GeYe].
At this point we are in position to find an approximate solution to the inverse Ising problem, either by inverting equation (\[eq:dual2\]) or (\[eq:eq2\]). To get a solution at a given order $n$ in the coupling, solving (\[eq:eq2\]) requires $G$ at order $n+1$, while it is needed at order $n$ in (\[eq:dual2\]).
Taking the expression of $G$ up to second order gives $$\frac{\partial G}{\partial J_{ij}} = -m_im_j-J_{ij}(1-m_i^2)(1-m_j^2),$$ and (\[eq:eq2\]) leads directly to the basic mean-field solution: $$\label{eq:JMF}
J_{ij}^{MF} = \frac{{{\hat \chi}}_{ij}}{(1-{\hat m}_i^2)(1-{\hat m}_j^2)}.$$ At this level of approximation for $G$, using (\[eq:dual1\]) we also have $$h_i = \frac{1}{2}\log\frac{1+m_i}{1-m_i}-\sum_jJ_{ij}m_j+\sum_j J_{ij}^2m_i(1-m_j^2)$$ which corresponds precisely to the TAP equations. Using now (\[eq:dual2\]) gives $$\frac{\partial h_i}{\partial m_j}= [\chi^{-1}]_{ij} =
\delta_{ij}\bigl(\frac{1}{1-m_i^2}+\sum_kJ_{ik}^2(1-m_k^2)\bigr) - J_{ij}-2J_{ij}^2m_im_j.$$ Ignoring the diagonal terms, the TAP solution is conveniently expressed in terms of the inverse empirical susceptibility, $$\label{eq:JTAP}
J_{ij}^{TAP} = -\frac{2[\hat \chi^{-1}]_{ij}}{1+\sqrt{1-8{\hat m}_i{\hat m}_j[\hat \chi^{-1}]_{ij}}},$$ where the branch corresponding to a vanishing coupling in the limit of small correlation i.e. small $\hat \chi_{ij}$ and $[\hat \chi^{-1}]_{ij}$ for $i\ne j$, has been chosen.
#### Bethe approximate solution
When the graph formed by the pairs $(i,j)$ for which the correlations $\hat \chi_{ij}$ are given by some observations is a tree, the following form of the joint probability corresponding to the Bethe approximation: $$\label{eq:ansatz}
{\cal P}({{\mathbf{x}}}) = \prod_{i<j}\frac{{\hat p}_{ij}(x_i,x_j)}{{\hat p}(x_i){\hat p}(x_j)}\prod_i{\hat p}_i(x_i),$$ yields actually an exact solution to the inverse problem (\[eq:iipb\]), where the ${\hat p}$ are the single and pair variables empirical marginal given by the observations. Using the following identity $$\begin{aligned}
\log\frac{\hat p_{ij}(s_i,s_j)}{\hat p_i(s_i)\hat p_j(s_j)} &=
\frac{(1+s_i)(1+s_j)}{2}\log\frac{\hat p_{ij}^{11}}{\hat p_i^1\hat p_j^1}
+\frac{(1+s_i)(1-s_j)}{2}\log\frac{\hat p_{ij}^{10}}{\hat p_i^1\hat p_j^0}\\[0.2cm]
&+\frac{(1-s_i)(1+s_j)}{2}\log\frac{\hat p_{ij}^{01}}{\hat p_i^0\hat p_j^1}
+\frac{(1-s_i)(1-s_j)}{2}\log\frac{\hat p_{ij}^{00}}{\hat p_i^0\hat p_j^0}\end{aligned}$$ where the following parametrization of the $\hat p$’s $$\begin{aligned}
{\hat p}_i^{x} &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\hat p}\bigl(\frac{1+s_i}{2} = x\bigr) = \frac{1}{2}(1+\hat m_i(2x-1)),\label{eq:pi}\\[0.2cm]
{\hat p}_{ij}^{xy} &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\hat p}\bigl(\frac{1+s_i}{2}=x,\frac{1+s_j}{2}=y\bigr)\nonumber\\[0.2cm]
&= \frac{1}{4}(1+\hat m_i(2x-1)+\hat m_j(2y-1)+\hat m_{ij}(2x-1)(2y-1)\label{eq:pij}\end{aligned}$$ relating the empirical frequency statistics to the empirical “magnetizations” $m\equiv \hat m$, can be used. Summing up the different terms gives us the mapping onto an Ising model (\[def:ising\]) with $$\begin{aligned}
h_i &= \frac{1-d_i}{2}\log\frac{{\hat p}_i^1}{{\hat p}_i^0}
+\frac{1}{4}\sum_{j\in\partial i}\log\Bigl(\frac{{\hat p}_{ij}^{11}\
{\hat p}_{ij}^{10}}{{\hat p}_{ij}^{01}\ {\hat p}_{ij}^{00}}\Bigr),\label{eq:hi}\\[0.2cm]
J_{ij} &= \frac{1}{4}\log\Bigl(\frac{{\hat p}_{ij}^{11}\ {\hat p}_{ij}^{00}}{{\hat p}_{ij}^{01}\ {\hat p}_{ij}^{10}}\Bigr),
\qquad\forall\ (i,j)\in{\mathcal{E}},\label{eq:Jij}\end{aligned}$$ where $d_i$ is the number of neighbors of $i$, using the notation $j\in\partial i$ for “$j$ neighbor of $i$”. The partition function is then explicitly given by $$\label{eq:zbethe}
Z_{Bethe}[{\hat p}] = \exp\biggl[-\frac{1}{4}\sum_{(i,j)\in{\mathcal{E}}}\log\bigl({\hat p}_{ij}^{00}\ {\hat p}_{ij}^{01}\ {\hat p}_{ij}^{10}\ {\hat p}_{ij}^{11}\bigr)
-\sum_i\frac{1-d_i}{2}\log({\hat p}_i^0\ {\hat p}_i^1)\biggr]$$ The corresponding Gibbs free energy can then be written explicitly using (\[eq:hi\]–\[eq:zbethe\]). With fixed magnetizations $m_i$’s, and given a set of couplings $\{J_{ij}\}$, the parameters $m_{ij}$ are implicit function $$m_{ij} = m_{ij}(m_i,m_j,J_{ij}),$$ obtained by inverting the relations (\[eq:Jij\]). For the linear response, we get from (\[eq:hi\]) a result derived first in [@WeTe]: $$\begin{aligned}
\frac{\partial h_i}{\partial m_j} &= \Bigl[\frac{1-d_i}{1-m_i^2}\\[0.2cm]
&+ \frac{1}{16}\sum_{k\in \partial i}\Bigl(
\bigl(\frac{1}{{\hat p}_{ik}^{11}}+\frac{1}{{\hat p}_{ik}^{01}}\bigr)\bigl(1+\frac{\partial m_{ik}}{\partial m_i}\bigr)+
\bigl(\frac{1}{{\hat p}_{ik}^{00}}+
\frac{1}{{\hat p}_{ik}^{10}}\bigr)\bigl(1-\frac{\partial m_{ik}}{\partial m_i}\bigr)\Bigr)\Bigr]\delta_{ij}\\[0.2cm]
&+\frac{1}{16}\Bigl(
\bigl(\frac{1}{{\hat p}_{ij}^{11}}+\frac{1}{{\hat p}_{ij}^{10}}\bigr)\bigl(1+\frac{\partial m_{ij}}{\partial m_i}\bigr)+
\bigl(\frac{1}{{\hat p}_{ij}^{00}}+
\frac{1}{{\hat p}_{ij}^{01}}\bigr)\bigl(1-\frac{\partial m_{ij}}{\partial m_i}\bigr)\Bigr)\Bigr]\delta_{j\in\partial i}.\end{aligned}$$ Using (\[eq:Jij\]), we can also express $$\frac{\partial m_{ij}}{\partial m_i} =
-\frac{\frac{1}{{\hat p}_{ij}^{11}}+\frac{1}{{\hat p}_{ij}^{01}}-\frac{1}{{\hat p}_{ij}^{10}}-\frac{1}{{\hat p}_{ij}^{00}}}
{\frac{1}{{\hat p}_{ij}^{11}}+\frac{1}{{\hat p}_{ij}^{01}}+\frac{1}{{\hat p}_{ij}^{10}}+\frac{1}{{\hat p}_{ij}^{00}}},$$ so that with little assistance of Maple, we may finally reach the expression given in [@NgBe] $$\begin{aligned}
[\hat \chi^{-1}]_{ij} = \Bigl[\frac{1-d_i}{1-m_i^2} &+\sum_{k\in \partial i}\frac{1-m_k^2}{(1-m_i^2)(1-m_k^2)
-\hat \chi_{ik}^2}\Bigr]\delta_{ij}\nonumber\\[0.2cm]
&-\frac{\hat \chi_{ij}}{(1-m_i^2)(1-m_j^2)-\hat \chi_{ij}^2}\ \delta_{j\in \partial i},\label{eq:invchis}\end{aligned}$$ equivalent to the original one derived in [@WeTe], albeit written in a different form, more suitable to discuss the inverse Ising problem. This expression is quite paradoxical since the inverse of the $[\chi]_{ij}$ matrix, which coefficients appear on the right hand side of this equation, should coincide with the left hand side, given as input of the inverse Ising problem. The existence of an exact solution can therefore be checked directly as a self-consistency property of the input data $\hat\chi_{ij}$: for a given pair $(i,j)$ either:
- $[\hat\chi^{-1}]_{ij}\ne 0$, then this self-consistency relation (\[eq:invchis\]) has to hold and $J_{ij}$ is given by (\[eq:Jij\]) using $\hat m_{ij}= \hat m_i\hat m_j+\hat \chi_{ij}$.
- $[\hat\chi^{-1}]_{ij} = 0$ then $J_{ij}=0$ but $\hat\chi_{ij}$, which can be non-vanishing, is obtained by inverting $[\hat\chi^{-1}]$ defined by (\[eq:invchis\]).
Finally, complete consistency of the solution is checked on the diagonal elements in (\[eq:invchis\]). If full consistency is not verified, this equation can nevertheless be used to find approximate solutions. Remark that, if we restrict the set of equations (\[eq:invchis\]), e.g. by some thresholding procedure, in such a way that the corresponding graph is a spanning tree, then, by construction, $\chi_{ij} \equiv \hat\chi_{ij}$ will be solution on this restricted set of edges, simply because the BP equations are exact on a tree. The various methods proposed for example in [@MoMe; @YaTa] actually correspond to different heuristics for finding approximate solutions to this set of constraints. As noted in [@NgBe], a direct way to proceed is to eliminate $\chi_{ij}$ in the equations obtained from (\[eq:Jij\]) and (\[eq:invchis\]): $$\begin{aligned}
\chi_{ij}^2+2\chi_{ij}(m_im_j-\coth(2J_{ij}))+(1-m_i^2)(1-m_j^2) = 0\\[0.2cm]
\chi_{ij}^2-\frac{\chi_{ij}}{[\chi^{-1}]_{ij}}-(1-m_i^2)(1-m_j^2) = 0.\end{aligned}$$ This leads directly to $$\label{eq:JBETHE}
J_{ij}^{Bethe} = -\frac{1}{2}\operatorname{atanh}\Bigl(\frac{2[\hat\chi^{-1}]_{ij}}{\sqrt{1+4(1-{\hat m}_i^2)
(1-{\hat m}_j^2)[\hat\chi^{-1}]_{ij}^2}-2{\hat m}_i{\hat m}_j[\hat\chi^{-1}]_{ij}}\Bigr),$$ while the corresponding computed of $\chi_{ij}$, instead of the observed one $\hat \chi_{ij}$, has to be inserted in (\[eq:hi\]) to be fully consistent. Note that $J_{ij}^{Bethe}$ and $J_{ij}^{TAP}$ coincide at second order in $[\hat\chi^{-1}]_{ij}$.
#### Hopfield model
As mentioned in the introduction when the distribution to be modeled is multi-modal, the situation corresponds to finding an Ising model in the low temperature phase with many modes, referred to as Mattis states in the physics literature. Previous methods assume implicitly a high temperature where only one single mode, “the paramagnetic state” is selected. The Hopfield model, introduced originally to model auto-associative memories, is a special case of an Ising model, where the coupling matrix is of low rank $p\le N$ and corresponds to the sum of outer products of $p$ given spin vectors $\{\xi^k,k=1\ldots p\}$, each one representing a specific attractive pattern: $$J_{ij} = \frac{1}{p}\sum_{k=1}^N\xi_i^k\xi_j^k$$ In our inference context, these patterns are not given directly, the input of the model being the covariance matrix. In [@CoMoSe] these couplings are interpreted as the contribution stemming from the $p$ largest principle axes of the correlation matrix. This lead in particular the authors to propose an extension of the Hopfield model by introducing repulsive patterns to take as well into account the smallest principal axes. Assuming small patterns coefficients $|\xi^k| <1/\sqrt{N}$, they come up with the following couplings with highest likelihood: $$J_{ij}^{Hopfield} \equiv \frac{1}{\sqrt{(1-m_i^2)(1-m_j^2)}}
\sum_{k=1}^p\Bigl(\bigl(1-\frac{1}{\lambda_k}\bigr)v_i^kv_j^k - \bigl(\frac{1}{\lambda_{N-k}}-1\bigr)v_i^{N-k}v_j^{N-k}\Bigr)$$ at first order of the perturbation. At this order of approximation the local fields are given by $$h_i = \tanh(m_i)-\sum_j J_{ij}^{Hopfield}m_j.$$ In a previous study [@FuLaAu] we found a connection between the plain direct BP method with the Hopfield model, by considering a $1$-parameter deformation of the Bethe inference model (\[eq:ansatz\]) $$\label{eq:bpalpha}
{\cal P}({{\mathbf{x}}}) = \prod_{i<j}\Bigl(\frac{{\hat p}_{ij}(x_i,x_j)}{{\hat p}(x_i){\hat p}(x_j)}\Bigr)^\alpha\prod_i{\hat p}_i(x_i),$$ with $\alpha\in[0,1]$. We observed indeed that when the data corresponds to some multi-modal measure with well separated components, this measure coincides asymptotically with an Hopfield model made only of attractive pattern, representative of each component of the underlying measure. $\alpha$ represents basically the inverse temperature of the model and is easy to calibrate in practice.
More on the Bethe susceptibility {#sec:cumulant}
--------------------------------
The explicit relation (\[eq:invchis\]) between susceptibility and inverse susceptibility coefficients is not the only one that can be obtained. In fact, it is the specific property of a singly connected factor graph that two variables $x_i$ and $x_j$, conditionally to a variable $x_k$ are independent if $k$ is on the path between $i$ and $j$ along the tree: $$p(x_i,x_j,x_k) = p(x_i\vert x_k)p(x_j\vert x_k)p(x_k) = \frac{p(x_i,x_k)p(x_j,x_k)}{p(x_k)}$$ Using the parametrization (\[eq:pi\],\[eq:pij\]) with $m_{ij}=m_im_j+\chi_{ij}$ yields immediately $$\label{eq:chirec}
\chi_{ij} = \frac{\chi_{ik}\chi_{jk}}{1-m_k^2},\qquad\forall\ k\in(i,j)\ \text{along}\ {\mathcal{T}}.$$ By recurrence we get, as noticed in e.g. [@Mora], given the path $i_0=i,i_1,\ldots,i_{n+1}=j$ between $i$ and $j$ along the tree ${\mathcal{T}}$ $$\chi_{ij} = \frac{\prod_{a=0}^n\chi_{i_ai_{a+1}}}{\prod_{a=1}^n(1-m_{i_a}^2)},$$ reflecting the factorization of the joint measure. This expression actually coincides with (\[eq:invchis\]) only on a tree. On a loopy graph, this last expression should be possibly replaced by a sum over paths.
Higher order susceptibility coefficients are built as well in terms of elementary building blocks given by the pairwise susceptibility coefficients $\chi_{ij}$. The notations generalize into the following straightforward manner: $$\begin{aligned}
m_{ijk} &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb E}(s_is_js_k) {\stackrel{\mbox{\upshape\tiny def}}{=}}m_im_jm_k + m_i\chi_{jk}+m_j\chi_{ik}+m_k\chi_{ij}+\chi_{ijk} \\[0.2cm]
m_{ijkl} &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb E}(s_is_js_ks_l) {\stackrel{\mbox{\upshape\tiny def}}{=}}m_im_jm_km_l \\[0.2cm]
&+m_im_j\chi_{kl}+m_im_k\chi_{jl}+m_im_l\chi_{jk}+m_jm_k\chi_{il}+m_jm_l\chi_{ik}+m_km_l\chi_{ij}\\[0.2cm]
&+m_i\chi_{jkl}+m_j\chi_{ikl}+m_k\chi_{ijl}+m_l\chi_{ijk}+\chi_{ijkl}, \end{aligned}$$ where $\chi_{ijk}$ and $\chi_{ijkl}$ are respectively three and four point susceptibilities. The quantities being related to the corresponding marginals similarly to (\[eq:pi\],\[eq:pij\]): $$\begin{aligned}
p(s_i,s_j,s_k) &= \frac{1}{8}\bigl(1+m_is_i+m_js_j+m_ks_k\\[0.2cm]
&+m_{ij}s_is_j+m_{ik}s_is_k+m_{jk}s_js_k+m_{ijk}s_is_js_k\bigr)\\[0.2cm]
p(s_i,s_j,s_k,s_l) &= \frac{1}{16}\bigl(1+m_is_i+m_js_j+m_ks_k+m_ls_l\\[0.2cm]
&+m_{ij}s_is_j+m_{ik}s_is_k+m_{il}s_is_l+m_{jk}s_js_k+m_{jl}s_js_l+m_{kl}s_ks_l\\[0.2cm]
&+m_{ijk}s_is_js_k+m_{ijl}s_is_js_l+m_{ikl}s_is_ks_l+m_{jkl}s_js_ks_l+m_{ijkl}s_is_js_ks_l\bigr)\end{aligned}$$ Using the basic fact that, on the tree $$p(s_i,s_j,s_k) = \frac{p(s_i,s_j)p(s_j,s_k)}{p(s_i)}$$ when $j$ is on the path ${\wideparen{ik}}$ given by ${\mathcal{T}}$, and $$p(s_i,s_j,s_k) = \sum_{s_l}\frac{p(s_i,s_l)p(s_j,s_l)p(s_k,s_l)}{p(s_l)^2}$$ when path $\wideparen{ij}$, $\wideparen{ik}$ and ${\wideparen{jk}}$ along ${\mathcal{T}}$ intersect on vertex $l$, we obtain $$\begin{aligned}
\chi_{ijk} = \begin{cases}
\DD -2\frac{m_l}{(1-m_l^2)^2}\chi_{il}\chi_{jl}\chi_{kl} & \text{with}\ \{l\} =
(i,j)\cap (i,k)\cap(j,k)
\ \text{along}\ {\mathcal{T}},\\[0.2cm]
\DD -2m_j\chi_{ik}& \text{if}\ j\in(i,k)\ \text{along}\ {\mathcal{T}}.
\end{cases}\end{aligned}$$ For the fourth order, more cases have to be distinguished. When $i$, $j$, $k$ and $l$ are aligned as on Figure \[fig:vertex\].c, in this order on the path ${\wideparen{il}}$ along ${\mathcal{T}}$ we have $$p(s_i,s_j,s_k,s_l) = \frac{p(s_i,s_j)p(s_j,s_k,s_l)}{p(s_j)^2}$$ which leads to $$\chi_{ijkl} = 4m_km_j\chi_{il} - \chi_{ik}\chi_{jl}-\chi_{il}\chi_{jk}.$$ When path ${\wideparen{ij}}$, ${\wideparen{ik}}$ and ${\wideparen{jk}}$ along ${\mathcal{T}}$ intersect on vertex $l$ as on Figure \[fig:vertex\].d, we obtain instead [^1] $$\chi_{ijkl} = 2\frac{3m_l^2-1}{1-m_l^2}\chi_{ij}\chi_{kl}.$$ For the situation corresponding to Figure \[fig:vertex\].e, we have $$p_(s_i,s_j,s_k,s_l) = \sum_{s_q}\frac{p(s_i,s_j,s_q)p(s_q,s_k,s_l)}{p(s_q)^2}$$ which leads to $$\chi_{ijkl} = 2\frac{3m_q^2-1}{1-m_q^2}\chi_{ij}\chi_{kl}.$$ Finally, for the situation corresponding to Figure \[fig:vertex\].f, we have $$p_(s_i,s_j,s_k,s_l) = \sum_{s_q}\frac{p(s_i,s_j)p(s_j,s_k,s_l)}{p(s_j)^2}$$ leading to $$\chi_{ijkl} = -\chi_{ik}\chi_{jl}-\chi_{jk}\chi_{il}+4\frac{m_km_q}{1-m_q^2}\chi_{ij}\chi_{lq}.$$
Sparse inverse estimation of covariance matrix
----------------------------------------------
Let us leave the Ising modeling issue aside for a while and introduce another related graph selection problem, named sparse inverse covariance estimation, which is defined on real continuous random variables. This method aims at constructing a sparse factor graph structure by identifying conditionally independent pairs of nodes in the graph, given empirical covariances of random variables. Assuming that all nodes in the graph follow a joint multi-variate Gaussian distribution with mean $\mu$ and covariance matrix $C$, the existing correlation between the nodes $i$ and $j$ given all the other nodes in the graph are indicated by the non-zero $ij$th entry of the precision matrix $C^{-1}$, while zero entries correspond to independent pairs of variables. Therefore, under the joint normal distribution assumption, selection of factor graph structures amounts to finding the sparse precision matrix that best describes the underlying data distribution, given the fixed empirical covariance matrix. When the derived inverse estimation is sparse, it becomes easier to compute marginal distribution of each random variable and conduct statistical inference. To achieve that goal, optimizations methods have been developed based on $L_0$ or $L_1$ norm penalty for the estimation of $C^{-1}$, to enhance its sparse structure constraint on the estimated inverse of covariance matrix and discover underlying conditionally independent parts.
Let $\hat C \in \mathbb{R}^{n \times n}$ be the empirical covariance matrix of $n$ random variables (represented as the nodes in the graph model). The sparsity penalized maximum likelihood estimation $A$ of the precision matrix $C^{-1}$ can be derived by solving the following positive definite cone program: $$\label{eq:logdetopt}
A = \operatorname*{argmin}_{ X \succ 0} -{\cal L}(X) + \lambda P(X)$$ where $$\label{def:LLcov}
{\cal L}(A) {\stackrel{\mbox{\upshape\tiny def}}{=}}\log\det(A)-\operatorname{Tr}(A\hat C),$$ is the log likelihood of the distribution defined by $A$, $\log\det$ being the logarithm of the determinant, and $P(A)$ is a sparsity inducing regularization term [@FrHaTi]. $\lambda$ is the regularization coefficient balancing the data-fitting oriented likelihood and sparsity penalty. Since the precision matrix of joint normal distribution should be positive definite, any feasible solution to this optimization problem is thus required to locate within a positive definite cone. The penalty term $P(A)$ is typically constructed using sparsity inducing matrix norm, also known as sparse learning in the domain of statistical learning. There are two typical configurations of $P(A)$:
- The $L_0$ norm $\|A\|_{L_0}$ of the matrix $X$, which counts the number of non-zero elements in the matrix. It is also known as the cardinality or the non-zero support of the matrix. Given its definition, it is easy to find that $L_0$ norm based penalty is a strong and intuitive appeal for sparsity structure of the estimated precision matrix. However, it is computationally infeasible to solve exact $L_0$-norm minimization directly, due to the fact that exact $L_0$ norm penalty is discontinuous and non-differentiable. In practice, one either uses a continuous approximation to the form of the $L_0$-penalty, or solve it using a greedy method. Due to the non-convexity of the exact $L_0$ norm penalty, only a local optimum of the feasible solution can be guaranteed. Nevertheless, $L_0$ norm penalty usually leads to much sparser structure in the estimation, while local optimum is good enough for most practical cases.
- The $L_1$ matrix norm $\|A\|_{L_1} = \sum_{i ,j}^n |A_{ij}|$. $L_1$ norm penalty was firstly introduced into the standard least square estimation problem by Tibshirani [@TibLASSO], under the name "Lasso regression”. Minimizing the $L_1$ norm based penalty encourages sparse non-zero entries in the estimated precision matrix $A$, which achieves a selection of informative variables for regression and reduces complexity of regression model efficiently. Further extension of the basic $L_1$-norm penalty function allows one assigning different non-negative weight values $\lambda_{ij}$ to different entries $A_{ij}$, as $\sum_{i ,j}^n\lambda_{ij}|A_{ij}|$. This weighted combination can constrain the sparse penalties only on the off-diagonal entries, so as to avoid unnecessary sparsity on the diagonal elements. Furthermore, this extension allows us to introduce prior knowledge about the conditional independence structure of the graph into the joint combination problem.
![Demonstration of $L_1$ norm, exact $L_0$ norm and Seamless $L_0$ norm based penalty function.[]{data-label="fig:L0L1penfuncf"}](L1L0SELO.pdf){width="80.00000%"}
For further understanding of the relation between the exact $L_0$ and $L_1$ norm penalty, we illustrate them with respect to one scalar variable in Figure \[fig:L0L1penfuncf\]. As we can see, within $[-1,1]$, $L_1$-norm penalty plays as a convex envelope of the exact $L_0$-norm penalty. Due to the convexity property of $L_1$ norm penalty, the global optimum of the convex programming problem can be achieved with even linear computational complexity [@TibLASSO; @FrHaTi]. However, although $L_1$ norm based penalty leads to computationally sound solutions to the original issue, it also introduces modeling bias into the penalized maximum likelihood estimation. As illustrated in the figure, when the underlying true values of the matrix entries are sufficiently large, the corresponding $L_1$ norm based regularization performs linearly increased penalty to those entries, which thus results in a severe bias w.r.t. the maximum likelihood estimation [@FanASA]. In contrast, $L_0$ norm based penalty avoids such issue by constraining the penalties of non-zero elements to be constant. It has been proved in [@TibLASSO] that the $L_1$ norm penalty discovers the underlined sparse structure when some suitable assumptions are satisfied. However, in general cases, the quality of the solutions is not clear.
Iterative proportional scaling from the Bethe reference point {#sec:iip}
=============================================================
Bethe reference point and optimal $1$-link correction {#sec:onelink}
-----------------------------------------------------
As observed in the previous section, when using the Bethe approximation to find an approximate solution to the IIP, the consistency check should then be that either the factor graph be sparse, nearly a tree, either the coupling are small. There are then two distinct ways of using the Bethe approximation:
- the direct way, where the form of the joint distribution (\[eq:ansatz\]) is assumed with a complete graph. There is then by construction a belief propagation fixed point for which the beliefs satisfy all the constraints. This solution to be meaningful requires small correlations, so that the belief propagation fixed point be stable and unique, allowing the corresponding log likelihood to be well approximated. Otherwise, this solution is not satisfactory, but a pruning procedure, which amounts to select a sub-graph based on mutual information, can be used. The first step is to find the maximum spanning tree (MST) with mutual information taken as edges weights. How to add new links to this baseline solution in a consistent way will be the subject of the next section.
- the indirect way consists in first inverting the potentially non-sparse correlation matrix. If the underlying interaction matrix is actually a tree, this will be visible in the inverse correlation matrix, indicated directly by the non-zero entries. Corresponding couplings are then determined through (\[eq:invchis\]). This procedure seems to work better than the previous one also when no sparsity but weak coupling is assumed. It corresponds in fact to the equations solved iteratively by the susceptibility propagation algorithm [@MoMe].
Let us first justify the intuitive assertion concerning the optimal model with tree like factor graphs, valid for any type of MRF. Suppose that we are given a set of single and pairwise empirical marginals $\hat p_i$ and $\hat p_{ij}$ for a set of $N$ real variables $\{x_i,i=1\ldots N\}$. If we start from an empty graph with no link, the joint probability distribution is simply the product form $${\mathcal{P}}^{(0)}(x_i) = \prod_{i=1}^N \hat p_i(x_i).$$ Adding a link $(ij)$ to the empty graph is optimally done by multiplying ${\mathcal{P}}^{(0)}$ by $\hat p_{ij}/\hat p_i\hat p_j$. The gain in log likelihood is then simply the mutual information between $x_i$ and $x_j$. Thus, as long as no loop get closed by the procedure, the best candidate link corresponds to the pair of variables with maximum mutual information and the measure reads after $n$ steps $${\mathcal{P}}^{(n)}({{\mathbf{x}}}) = \prod_{(ij)\in{\mathcal{G}}^{(n)}} \frac{p_{ij}(x_i,x_j)}{p_i(x_i)p_j(x_j)} \prod_{i=1}^n
p_i(x_i).$$ This suggests that a good initialization point for the algorithm is the maximum spanning tree with edges weights given by the relevant mutual information. This corresponds to the classical results of [@ChoLee] concerning inference using dependence trees. It is optimal in the class of singly connected graphical models. In the following, we will refer in the text to this specific approximate solution as the *Bethe reference point*.
Starting from this point, we want now to add one factor $\psi_{ij}$ to produce the distribution $$\label{eq:1link}
{\mathcal{P}}^{(n+1)}({{\mathbf{x}}}) = {\mathcal{P}}^{(n)}({{\mathbf{x}}})\times \frac{\psi_{ij}(x_i,x_j)}{Z_\psi}$$ with $$Z_\psi = \int dx_idx_j p_{ij}^{(n)}(x_i,x_j)\psi_{ij}(x_i,x_j).$$ The log likelihood corresponding to this new distribution reads $${\mathcal{L}}' = {\mathcal{L}}+ \int d{{\mathbf{x}}}\hat{\mathcal{P}}({{\mathbf{x}}})\log\psi_{ij}(x_i,x_j)-\log Z_\psi.$$ Since the the functional derivative w.r.t. $\psi$ is $$\frac{\partial{\mathcal{L}}'}{\partial \psi_{ij}(x_i,x_j)} = \frac{\hat p(x_i,x_j)}{\psi_{
ij}(x_i,x_j)} -
\frac{p_{ij}^{(n)}(x_i,x_j)}{Z_\psi},$$ $\forall (x_i,x_j)\in \Omega^2$, the maximum is attained for $$\label{eq:1linkfactor}
\psi_{ij}(x_i,x_j) = \frac{\hat p_{ij}(x_i,x_j)}{p_{ij}^{(n)}(x_i,x_j)}\text{
with } Z_\psi=1,$$ where $p^{(n)}(x_i,x_j)$ is the reference marginal distribution obtained from ${\mathcal{P}}^{(n)}$. The correction to the log-likelihood can then be rewritten as $$\label{eq:onelink}
\Delta {\mathcal{L}}= D_{KL}(\hat p_{ij}\Vert p_{ij}^{(n)}).$$ Sorting all the links w.r.t. this quantity yields the (exact) optimal $1$-link correction to be made. The interpretation is therefore immediate: the best candidate is the one for which the current model yields the joint marginal $p_{ij}^{(n)}$ that is most distant target $\hat p_{ij}$. Note that the update mechanism is indifferent to whether the link has to be added or simply modified.
Iterative proportional scaling for GMRF model selection
-------------------------------------------------------
In the statistics literature this procedure is referred to as the *iterative proportional scaling* (IPS) procedure, originally proposed for contingency table estimations and extended further to MRF maximum likelihood estimation [@DaRa; @DelPieLa]. Assuming the structure of the graph is known, it appears not to be very efficient [@Malouf] when compared to other gradient based methods. The problem of the method is that it requires the knowledge of all pairwise marginals $\{p_{ij},(ij)\notin {\mathcal{G}}^{(n)}\}$ at each iteration step $n$. In the Ising case this is done by a sparse matrix inversion through equation (\[eq:invchis\]), which potentially renders the method a bit expensive and rapidly inaccurate after many links have been added. For Gaussian MRF, the situation is different, because in that case the correction to the log likelihood can be evaluated directly by another means. Making full use of the incremental characteristics of IPS we propose to construct the sparse GMRF graph structure link by link based on IPS in an efficient way.
Indeed, the correction factor (\[eq:1linkfactor\]) reads in that case $$\psi_{ij}(x_i,x_j) = \exp\Bigl(-\frac{1}{2}(x_i,x_j)\bigl(\hat
C_{\{ij\}}^{-1}-C_{\{ij\}}^{-1}\bigr)(x_i,x_j)^T\Bigr),$$ where $C_{\{ij\}}$ (resp. $\hat C_{\{ij\}}$) represents the restricted $2\times 2$ covariance matrix corresponding to the pair $(x_i,x_j)$ of the current model specified by precision matrix $A = C^{-1}$ (resp. the reference model). Let $[C_{\{ij\}}]$ denote the $N\times N$ matrix formed by completing $C_{\{ij\}}$ with zeros. The new model obtained after adding or changing link $(ij)$ reads $$\label{eq:perturbation}
A' = A + [\hat C_{\{ij\}}^{-1}]-[C_{\{ij\}}^{-1}] {\stackrel{\mbox{\upshape\tiny def}}{=}}A+[V].$$ with a log likelihood variation given by: $$\Delta {\mathcal{L}}= \frac{C_{ii}\hat C_{jj}+C_{jj}\hat C_{ii}-2C_{ij}\hat
C_{ij}}{\det(C_{\{ij\}})}- 2 -\log\frac{\det(\hat
C_{\{ij\}})}{\det(C_{\{ij\}})}.$$ Let us notice the following useful formula: $$\begin{aligned}
\label{eq:inverse}
(A+[V])^{-1} &=
A^{-1}-A^{-1}[V](1+A^{-1}[V])^{-1}A^{-1}
\nonumber\\[0.2cm]
&= A^{-1}-A^{-1}[V](1+[C_{\{ij\}}][V])^{-1}A^{-1},\end{aligned}$$ valid for a $2\times 2$ perturbation matrix $V = V_{\{ij\}}$. Using this formula, the new covariance matrix reads $$\label{eq:covupdate}
C' = A'^{-1} = A^{-1}- A^{-1}[C_{\{ij\}}^{-1}]\bigl(1-[\hat
C_{\{ij\}}][C_{\{ij\}}^{-1}]\bigr)A^{-1}.$$ The number of operations needed to maintain the covariance matrix after each addition is therefore ${\cal O}(N^2)$. This technical point is determinant to render our approach useful in practice.
Let us now examine under which condition adding or modifying links in this way lets the covariance matrix remain positive semi-definite. By adding a $2\times 2$ matrix, we expect a quadratic correction to the determinant: $$\begin{aligned}
\det(A') &= \det(A)\det(1+A^{-1}V)\\[0.2cm]
&= \det A \times \frac{\det(C_{\{ij\}})}{\det(\hat C_{\{ij\}})},\end{aligned}$$ which is obtained directly because $A^{-1}V$ has non zero entries only on column $i$ and $j$. Multiplying $V$ by some parameter $\alpha\ge 0$, define $$P(\alpha) {\stackrel{\mbox{\upshape\tiny def}}{=}}\det(1+\alpha A^{-1}V)
= \alpha^2\det\bigl(C_{\{ij\}}\hat
C_{\{ij\}}^{-1}-\frac{\alpha-1}{\alpha}\bigr).$$ $P(\alpha)$ is proportional to the characteristic polynomial of the matrix $C_{\{ij\}}\hat C_{\{ij\}}^{-1}$ of argument $(\alpha-1)/\alpha$. So $P$ will have no root in $[0,1]$, and $A'$ be definite positive, iff $C_{\{ij\}}\hat C_{\{ij\}}^{-1}$ is definite positive. Since the product of eigenvalues, $\det(C_{\{ij\}}\hat C^{-1}_{\{ij\}})$, is positive, one has to check their sum, given by $\operatorname{Tr}(C_{\{ij\}}\hat C_{\{ij\}}^{-1})$: $$\label{eq:tracepos}
C_{ii}\hat C_{jj}+C_{jj}\hat C_{ii}-2C_{ij}\hat C_{ij} > 0.$$ $C_{\{ij\}}$ and $\hat C_{\{ij\}}$ are individually positive definite: $$C_{ii}C_{jj}-{C_{ij}}^2 > 0 \qquad\text{and}\qquad \hat C_{ii}\hat C_{jj}-\hat
C_{ij}^2 > 0,$$ from which we deduce that $$\Bigl(\frac{C_{ii}\hat C_{jj}+C_{jj}\hat C_{ii}}{2}\Bigr)^2 > C_{ii}\hat C_{jj}
C_{jj}\hat C_{ii} > {C_{ij}}^2 {\hat C_{ij}}^2,$$ giving finally that (\[eq:tracepos\]) is always fulfilled when both $C_{\{ij\}}$ and $\hat C_{\{ij\}}$ are non-degenerate.
#### Removing one link
To use this in an algorithm, it is also desirable to be able to remove links, so that, with help of a penalty coefficient per link, the model can be optimized with a desired connectivity level.
For the Gaussian model, if $A$ is the coupling matrix, removing the link $(i,j)$ amounts to chose a factor $\psi_{ij}$ in (\[eq:1link\]) of the form: $$\psi_{ij}(x_i,x_j) = \exp\bigl(A_{ij} x_i x_j\bigr)$$ ($x_i$ and $x_j$ are assumed centered as in the preceding section). Again, let $V$ denote the perturbation in the precision matrix such that $A'
=A+V$ is the new one. The corresponding change in the log likelihood reads $$\Delta {\mathcal{L}}= \log\det(1+A^{-1}V)-\operatorname{Tr}(V\hat C).$$ Rearranging this expression leads to $$\Delta {\mathcal{L}}= \log\bigl(1-2A_{ij}C_{ij}-A_{ij}^2\det(C_{\{ij\}})\bigr) +
2A_{ij}\hat
C_{ij},$$ and using again (\[eq:inverse\]), we get for the new covariance matrix $$\label{eq:covupdate2}
C' = C - \frac{A_{ij}}{1-2A_{ij}C_{ij}-A_{ij}^2\det(C_{\{ij\}})}\ C [B_{\{ij\}}]
C,$$ with $$B_{\{ij\}} {\stackrel{\mbox{\upshape\tiny def}}{=}}\left[
\begin{matrix}
A_{ij}C_{jj} & 1-A_{ij}C_{ij}\\[0.2cm]
1-A_{ij}C_{ij} & A_{ij}C_{ii}
\end{matrix}\right].$$ To check for the positive-definiteness property of $A'$, let us observe first that $$\det(A') = \det(A)\times P(A_{ij}),$$ with $$P(x) = \bigl(1-x(C_{ij}-\sqrt{C_{ii}C_{jj}})\bigr)\bigl(1-x(C_{ij}+\sqrt{C_{ii}C_{jj}})\bigr).$$ When $x$ varies from $0$ to $A_{ij}$, $P(x)$ should remains strictly positive to insure that $A'$ is definite positive. This results in the following condition: $$\frac{1}{ C_{ij} - \sqrt{C_{ii}C_{jj}}} < A_{ij} < \frac{1}{\sqrt{C_{ii}C_{jj}}
+ C_{ij}}.$$
Imposing walk-summability
-------------------------
Having a sparse GMRF gives no guarantee about its compatibility with GaBP. In order to be able to use the GaBP algorithm for inference, more strict constraints on graph connectivity need to be imposed. The most precise condition known for convergence and validity of GaBP is called walk-summability (WS) and is extensively described by [@MaJoWi]. The two necessary and sufficient conditions for WS that we consider here are:
(i) The matrix $\operatorname{Diag}(A) - |R(A)|$ is definite positive, where $R(A) {\stackrel{\mbox{\upshape\tiny def}}{=}}A
- \operatorname{Diag}( A)$ contains the off-diagonal terms of $A$.
(ii) $\rho(|R'(A)|) <1$, with $R'(A)_{ij}{\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{R(A)_{ij}}{\sqrt{A_{ii}A_{jj}}}$,
with $\rho(A)$, called spectral radius of $A$, denotes the maximal modulus of the eigenvalues of the matrix $A$.
#### Adding or Modifying one link
If we wish to increase the likelihood by adding or modifying a link $(i,j)$ to the graph, the model is perturbed according to . Assuming that $A$ is WS, we want to express conditions under which $A'$ is also WS.
Using the definition (i) of walk summability we can derive a sufficient condition for WS by imposing that $W(A') {\stackrel{\mbox{\upshape\tiny def}}{=}}\operatorname{Diag}(A') - |R(A')|$ remains definite positive. For $\alpha \in [0,1]$ we have $$W(A+\alpha V) = W(A) + [\phi(\alpha V,A)],$$ with $$\phi(V,A) {\stackrel{\mbox{\upshape\tiny def}}{=}}\left[\begin{array}{cc}
V_{ii} & |A_{ij}|-|V_{ij}+A_{ij}|\\
|A_{ij}|-|V_{ji}+A_{ji}| & V_{jj}\\
\end{array}\right]$$ We express the determinant of $W(A+\alpha V)$ as $$\begin{aligned}
\det (W(A+\alpha V)) &= \det \left( W\right)\det \left(1+W^{-1}[\phi(\alpha V,
A)]
\right)\\
&= \det \left( W\right) \Theta(\alpha).\end{aligned}$$ shortening $W(A)$ in $W$. $\Theta(\alpha)$ is a degree $2$ polynomial by parts: $$\begin{aligned}
\Theta(\alpha) {\stackrel{\mbox{\upshape\tiny def}}{=}}&\det (W^{-1}_{\{ij\}})\left(
\alpha^2V_{ii}V_{jj}-B(\alpha)^2\right)\\
&+ \alpha\left(W^{-1}_{ii}V_{ii}+W^{-1}_{jj}V_{jj}\right) + 2B(\alpha)W_{ij},\end{aligned}$$ with $B(\alpha) {\stackrel{\mbox{\upshape\tiny def}}{=}}|A_{ij}|-|\alpha V_{ji}+A_{ji}|$. Our sufficient condition for WS of $A'$ is simply $$\label{eq:WSSC}
\Theta(\alpha)>0,\forall \alpha \in [0,1],$$ which boils down in the worst case to compute the roots of $2$ polynomials of degree $2$. Note that checking this sufficient condition imposes to keep track of the matrix $W^{-1}$ which requires $\mathcal{O}(N^2)$ operations at each step, using again .
#### Removing one link
Removing one link of the graph will change the matrix $A$ in $A'$ such as $|R'(A')|
\leq |R'(A)|$ where $\leq$ denotes element-wise comparison. Then, using elementary results on positive matrices, $\rho(|R'(A')|) \leq \rho(|R'(A)|)$ and thus removing one link of a WS model yields a new WS model.
As we shall see, WS might be a too strong constraint to impose since it is easy to find non WS models that are compatible with GaBP. The framework we present here allows us to impose weaker spectral constraint, e.g. we could only ensure that $\operatorname{Diag}(A) - R(A)$ remains definite positive. This is equivalent to replace (i) by $\rho(R'(A))<1$.
IPS based GMRF model selection algorithms {#sec:greedyalgo}
-----------------------------------------
We are now equipped to define algorithms based on additions/modifications/deletions of links.
#### Algorithm 1: Incremental graph construction by link addition
- INPUT: the MST graph, and corresponding covariance matrix $C$.
- : select the link with highest $\Delta{\mathcal{L}}$, compatible with if we wish to impose WS. Update $C$ according to (\[eq:covupdate\]).
- : repeat S2 until convergence (i) or until a target connectivity is reached (ii).
- : if (ii) repeat S2 until convergence by restricting the link selection in the set of existing ones.
Each single addition/modification involves $N^2$ operations due to the covariance matrix update (\[eq:covupdate\]). Without S4, the complexity is therefore ${\cal O}(KN^3)$ if $K$ is the final mean connectivity, i.e. ${\cal O}(N^3)$ in the sparse domain and ${\cal O}(N^4)$ in the dense one. In practice, if many links are added, S4 has to be run for several intermediate values of $K$ with a few updates for each existing link, so its contribution to the complexity is also of the same order ${\cal O}(KN^3)$.
#### Algorithm 2: IPS with backtracking by link addition/deletion
- INPUT: the MST graph, and corresponding covariance matrix $C$, a link penalty coefficient $\nu$.
- : select the modification with highest $\Delta{\mathcal{L}}-s\nu$, with $s=+1$ for an addition, $s=-1$ for a deletion and $s=0$ for a modification, compatible with the WS preserving condition of $A'$. Update $C$ according to (\[eq:covupdate\]) and (\[eq:covupdate2\]) respectively for an addition/modification and a deletion.
- : repeat S2 until convergence.
- : repeat S2 until convergence by restricting the link selection in the set of existing ones.
In absence of penalty ($\nu=0$) the algorithm will simply generate a model for any value of the mean connectivity, hence delivering an almost continuous Pareto set of solutions, with all possible trade-offs between sparsity and likelihood, as long as walk summability is satisfied. With a fixed penalty, the algorithm convergences instead towards a solution with the connectivity depending implicitly on $\nu$; it corresponds roughly to the point $K^\star$ where the slope is $$\frac{\Delta{\mathcal{L}}}{N\Delta K}(K^\star) = \nu.$$
If we want to use the backtracking mechanism allowed by the penalty term without converging to a specific connectivity, we may also let $\nu$ be adapted dynamically. A simple way is to adapt $\nu$ with the rate of information gain by letting $$\nu = \eta\Delta{\mathcal{L}}_{\text{add}},\quad\text{with}\quad\eta\in[0,1[,$$ where $\Delta{\mathcal{L}}_{\text{add}}$ corresponds to the gain of the last link addition. With such a setting, $\nu$ is always kept just below the information gain per link, allowing thus the algorithm to carry on toward higher connectivity. This heuristic of course assumes a concave Pareto front.
Perturbation theory near the Bethe point {#sec:perturbation}
========================================
Linear response of the Bethe reference point
--------------------------------------------
The approximate Boltzmann machines described in the introduction are obtained either by perturbation around the trivial point corresponding to a model of independent variables, the first order yielding the Mean-field solution and the second order the TAP one, either by using the linear response delivered in the Bethe approximation. We propose to combine in a way the two procedures, by computing the perturbation around the Bethe model associated to the MST with weights given by mutual information. We denote by ${\cal T}\subset{\cal E}$, the subset of links corresponding to the MST, considered as given as well as the susceptibility matrix $[\chi_{Bethe}]$ given explicitly by its inverse through (\[eq:invchis\]), in term of the empirically observed ones $\hat\chi$. Following the same lines as the one given in Section \[sec:prelim\], we consider again the Gibbs free energy to impose the individual expectations ${{\bf m}}= \{{\hat m}_i\}$ given for each variable. Let ${{\bf J}^{Bethe}}= \{K_{ij},(i,j)\in{\cal T}\}$ the set of Bethe-Ising couplings, i.e. the set of coupling attached to the MST s.t. corresponding susceptibilities are fulfilled and ${{\bf J}}= \{J_{ij},(i,j)\in{\mathcal{E}}\}$ a set of Ising coupling corrections. The Gibbs free energy reads now $$G[{{\bf m}},{{\bf J}}] = {{\bf h}}^{T}({{\bf m}}){{\bf m}}+ F\bigl[{{\bf h}}({{\bf m}}),{{\bf J}^{Bethe}}+{{\bf J}}\bigr]$$ where ${{\bf h}}({{\bf m}})$ depends implicitly on ${{\bf m}}$ through the same set of constraints (\[eq:constraints\]) as before. The only difference resides in the choice of the reference point. We start from the Bethe solution given by the set of coupling ${{\bf J}^{Bethe}}$ instead of starting with an independent model.
The Plefka expansion is used again to expand the Gibbs free energy in power of the coupling $J_{ij}$ assumed to be small. Following the same lines as in Section \[sec:plefka\], but with $G_0$ now replaced by $$G_{Bethe}[{{\bf m}}] = {{\bf h}}^{T}({{\bf m}}){{\bf m}}- \log Z_{Bethe}\bigl[{{\bf h}}({{\bf m}}),{{\bf J}}^{Bethe}\bigr],$$ and $h_i$, $J^{Bethe}$ and $Z_{Bethe}$ given respectively by (\[eq:hi\],\[eq:Jij\],\[eq:zbethe\]) where ${\mathcal{E}}$ is now replaced by ${\mathcal{T}}$, letting again $$H^1 {\stackrel{\mbox{\upshape\tiny def}}{=}}\sum_{i<j} J_{ij}s_is_j,$$ and following the same steps (\[eq:dG1\],\[eq:dG2\],\[eq:dha\]) leads to the following modification of the local fields $$h_i = h_i^{Bethe} - \sum_j[\chi_{Bethe}^{-1}]_{ij}\operatorname{Cov}_{Bethe}(H^1,s_i)\qquad\forall i\in{\mathcal{V}}$$ to get the following Gibbs free energy at second order in $\alpha$ (after replacing $H^1$ by $\alpha H^1$): $$\begin{aligned}
G[{{\bf m}},\alpha J] &= G_{Bethe}({{\bf m}})-\alpha{\mathbb E}_{Bethe}(H^1)\\[0.2cm]
-\frac{\alpha^2}{2}\Bigl(&\operatorname{Var}_{Bethe}(H^1)-
\sum_{ij}[\chi_{Bethe}^{-1}]_{ij}\operatorname{Cov}_{Bethe}(H^1,s_i)\operatorname{Cov}_{Bethe}(H^1,s_j)\Bigr)+o(\alpha^2).\end{aligned}$$ This is the general expression for the linear response near the Bethe reference point that we now use. $$\begin{aligned}
\label{eq:gblr}
G_{BLR}[{{\bf J}}] &{\stackrel{\mbox{\upshape\tiny def}}{=}}-{\mathbb E}_{Bethe}(H^1)\\[0.2cm]
&-\frac{1}{2}\Bigl(\operatorname{Var}_{Bethe}(H^1)-
\sum_{i,j}[\chi_{Bethe}^{-1}]_{ij}\operatorname{Cov}_{Bethe}(H^1,s_i)\operatorname{Cov}_{Bethe}(H^1,s_j)\Bigr).\end{aligned}$$ represents the Gibbs free energy at this order of approximation. It it is given explicitly through $$\begin{aligned}
{\mathbb{E}_{Bethe}}(H^1) & = \sum_{i<j} J_{ij}m_{ij}\\[0.2cm]
\operatorname{Var}_{Bethe}(H^1) & = \sum_{i<j,k<l} J_{ij}J_{kl}\bigl(m_{ijkl}- m_{ij}m_{kl}\bigl) \\[0.2cm]
\operatorname{Cov}_{Bethe}(H^1,s_k) & = \sum_{i<j} J_{ij}(m_{ijk} - m_{ij}m_k)\end{aligned}$$ where $$\begin{aligned}
m_i &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb{E}_{Bethe}}(s_i),\qquad
m_{ij} {\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb{E}_{Bethe}}(s_is_j)\\[0.2cm]
m_{ijk} &{\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb{E}_{Bethe}}(s_is_js_k),\qquad
m_{ijkl} {\stackrel{\mbox{\upshape\tiny def}}{=}}{\mathbb{E}_{Bethe}}(s_is_js_ks_l)\end{aligned}$$ are the moments delivered by the Bethe approximation. With the material given in Section \[sec:cumulant\] these are given in closed form in terms of the Bethe susceptibility coefficients $\chi_{Bethe}$. Concerning the log-likelihood, it is given now by: $$\label{eq:llo2}
{\mathcal{L}}[{{\bf J}}] = -G_{Bethe}({{\bf m}})-G_{BLR}[{{\bf J}}]-\sum_{ij}(J_{ij}^{Bethe}+J_{ij})\hat m_{ij}+o(J^2).$$ $G_{BLR}$ is at most quadratic in the $J$’s and contains the local projected Hessian of the log likelihood onto the magnetization constraints (\[eq:constraints\]) with respect to this set of parameters. This is nothing else than the Fisher information matrix associated to these parameter $J$ which is known to be positive-semidefinite, which means that the log-likelihood associated to this parameter space is convex. Therefore it makes sense to use the quadratic approximation (\[eq:llo2\]) to find the optimal point.
Line search along the natural gradient in a reduced space
---------------------------------------------------------
Finding the corresponding couplings still amounts to solve a linear problem of size $N^2$ in the number of variables which will hardly scale up for large system sizes. We have to resort to some simplifications which amounts to reduce the size of the problem, i.e. the number of independent couplings. To reduce the problem size we can take a reduced number of link into consideration, i.e. the one associated with a large mutual information or to partition them in a way which remains to decide, into a small number $q$ of group ${\mathcal{G}}_\nu, \nu=1,\ldots q$. Then, to each group $\nu$ is associated a parameter $\alpha_\nu$ with a global perturbation of the form $$H^1 = \sum_{\nu=1}^q \alpha_\nu H_\nu$$ where each $H_\nu$ involves the link only present in ${\mathcal{G}}_\nu$. $$H_\nu {\stackrel{\mbox{\upshape\tiny def}}{=}}\sum_{(i,j)\in{\mathcal{G}}_\nu} J_{ij} s_is_j,$$ and the $J_{ij}$ are fixed in some way to be discussed soon. The corresponding constraints, which ultimately insures a max log-likelihood in this reduced parameter space are then $$\frac{\partial G_{BLR}}{\partial \alpha_\nu} = -\hat{\mathbb E}(H_\nu).$$ This leads to the solution: $$\alpha_\mu = \sum_{\nu=1}^q\ {\mathcal{I}}^{-1}_{\mu\nu}\bigl(\hat{\mathbb E}(H_\nu) - {\mathbb{E}_{Bethe}}(H_\nu)\bigr)$$ where the Fisher information matrix ${\mathcal{I}}$ has been introduced and which reads in the present case $$\label{eq:Fisher}
{\mathcal{I}}_{\mu\nu} = \bigl[\operatorname{Cov}_{Bethe}(H_\mu,H_\nu)-\sum_{i\ne j\atop(i,j)\in{\mathcal{T}}}
[\chi_{Bethe}^{-1}]_{ij}\operatorname{Cov}_{Bethe}(H_\mu,s_i)\operatorname{Cov}_{Bethe}(H_\nu,s_j)\bigl]$$ The interpretation of this solution is to look in the direction of the natural gradient [@Amari; @ArAuHaOl] of the log likelihood. The exact computation of the entries of the Fisher matrix involves up to 4th order moments and can be computed using results of Section \[sec:cumulant\]. At this point, the way of choosing the groups of edges and the perturbation couplings $J_{ij}$ of the corresponding links, leads to various possible algorithms. For example, to connect this approach to the one proposed in Section \[sec:onelink\], the first group of links can be given by the MST, with parameter $\alpha_0$ and their actual couplings $J_{ij}=J_{ij}^{Bethe}$ at the Bethe approximation; making a short list of the $q-1$ best links candidates to be added to the graph, according to the information criteria \[eq:onelink\], defines the other groups as singletons. It is then reasonable to attach them the value $$J_{ij} = \frac{1}{4}\log\frac{\hat p_{ij}^{11}}{p_{ij}^{11}}\frac{\hat p_{ij}^{00}}{p_{ij}^{00}}\frac{p_{ij}^{01}}{\hat p_{ij}^{01}}
\frac{p_{ij}^{10}}{\hat p_{ij}^{10}},$$ of the coupling according to (\[eq:1linkfactor\]), while the modification of the local fields as a consequence of (\[eq:1linkfactor\]) can be dropped since the Gibbs free energy take it already into account implicitly, in order to maintain single variable magnetization $m_i=\hat m_i$ correctly imposed.
Reference point at low temperature
----------------------------------
Up to now we have considered the case where the reference model is supposed to be a tree and is represented by a single BP fixed point. From the point of view of the Ising model this corresponds to perturb a high temperature model in the paramagnetic phase. In practice the data encountered in applications are more likely to be generated by a multi-modal distribution and a low temperature model with many fixed points should be more relevant. In such a case we assume that most of the correlations are already captured by the definition of single beliefs fixed points and the residual correlations is contained in the co-beliefs of each fixed point. For a multi-modal distribution with $q$ modes with weight ${w_k,k=1\ldots q}$ and a pair of variables $(s_i,s_j)$ we indeed have $$\begin{aligned}
\chi_{ij} &= \sum_{k=1}^q w_k \operatorname{Cov}(s_i,s_j\vert k) +
\sum_{k=1}^q w_k ({\mathbb E}(s_i\vert k)-{\mathbb E}(s_i))({\mathbb E}(s_j\vert k)-{\mathbb E}(s_j))\\[0.2cm]
&{\stackrel{\mbox{\upshape\tiny def}}{=}}\chi_{ij}^{intra}+\chi_{ij}^{inter},\end{aligned}$$ where the first term is the average intra cluster susceptibility while the second is the inter cluster susceptibility. All the preceding approach can then be followed by replacing the single Bethe susceptibility and higher order moments in equations (\[eq:gblr\],\[eq:Fisher\]) in the proper way by their multiple BP fixed point counterparts. For the susceptibility coefficients, the inter cluster susceptibility coefficients $\chi^{inter}$ are given directly from the single variable belief fixed points. The intra cluster susceptibilities $\chi^{k}$ are treated the same way as the former Bethe susceptibility. This means that the co-beliefs of fixed points $k\in\{1,\ldots q\}$ are entered in formula (\[eq:invchis\]) which by inversion yields the $\chi^{k}$’s, these in turn leading to $\chi^{intra}$ by superposition. Higher order moments are obtain by simple superposition. Improved models could be then searched along the direction indicated by this natural gradient.
Weights propagation on Bethe-dual graphs {#sec:dwp}
========================================
Low temperature expansion
--------------------------
When the local fields are zero, which in many cases may be obtained with a proper definition of spin variables in a given inference problem, a traditional way to deal with the low temperature regime is given by the high coupling expansion. This is obtained by rewriting $$\label{eq:highJ}
e^{J_{ij}s_is_j} = \cosh(J_{ij})(1+\tanh(J_{ij})s_is_j).$$ Using this simple identity the partition function rewrites $$Z({{\bf J}}) = Z_0\times \sum_{\{\tau_{ij}\in\{0,1\}\}}\prod_{ij}\bigl(\bar \tau_{ij}+
\tau_{ij}\tanh(J_{ij})\bigr)\prod_i{{\leavevmode\hbox{\rm \small1\kern-0.35em\normalsize1}}_{\{\sum_{j\in \partial i}\tau_{ij}=\ 0\mod 2\}}},$$ with $$Z_0 = \prod_{(ij)}\cosh(J_{ij}).$$ The summation over bond variables $\tau_{ij}\in\{0,1\}$, corresponds to choosing one of the 2 terms in the factor (\[eq:highJ\]). The summation over spin variables then selects bonds configurations having an even number of bonds $\tau_{ij}=1$ attached to each vertex $i$. From this condition it results that the paths formed by these bonds must be closed. The contribution of a given path is simply the product of all bond factor $\tanh(J_{ij})$ along the path. As such the partition function is expressed as $$Z({{\bf J}}) = Z_0\times Z_{loops}$$ with $$Z_{loops} {\stackrel{\mbox{\upshape\tiny def}}{=}}\sum_{\ell}Q_\ell,$$ where the last sum runs over all possible closed loops ${\mathcal{G}}_\ell$, i.e. subgraphs for which each vertex has an even degree, including the empty graph and $$Q_\ell {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{(ij)\in{\mathcal{E}}_\ell}\tanh(J_{ij}),$$ where ${\mathcal{E}}_\ell$ denotes the set of edges involved in loop ${\mathcal{G}}_\ell$. This is a special case of the loop expansion around a belief propagation fixed point proposed by Chertkov and Chernyak in [@chertkov1]. In their case $Z_0$ is replaced by $Z_{Bethe}$ and the loop corrections runs over all generalized loops, i.e. all subgraphs containing no vertex with degree $1$.
If the graph has no loop, the partition function reduces to $Z_0$. If there are loops and $k({\mathcal{G}})$ connected components in the graph, we may define a set $\{{\mathcal{G}}_c,c=1,\ldots, C({\mathcal{G}})\}$ of independent cycles, with $C({\mathcal{G}})=\vert{\mathcal{E}}\vert-\vert{\mathcal{V}}\vert + k({\mathcal{G}})$ the so-called cyclomatic number [@Berge] of graph ${\mathcal{G}}$. Spanning the set $\{0,1\}^{C({\mathcal{G}})}$ yields all possible loops with the convention that edges are counted modulo 2 for a given cycle superposition (see Figure \[fig:loops\]). The partition function can therefore be written as a sum over dual binary variables $\tau_c\in\{0,1\}$ attached to each cycle $c\in\{1,\ldots \vert C\vert\}$: $$\label{eq:dualmeasure}
Z_{loops} = \sum_{\tau}Q_{{\mathcal{G}^\star}}(\tau),$$ where $Q_{{\mathcal{G}^\star}}(\tau)$ represents the weight for any loop configuration specified by $\{\tau_c\}$ on the dual (factor-)graph ${\mathcal{G}^\star}$ formed by the cycles. For instance, when the primal graph ${\mathcal{G}}$ is a $2$-d lattice, the dual one is also $2$-d and the Kramers-Wannier duality allows one to express the partition function at the dual coupling $J^\star{\stackrel{\mbox{\upshape\tiny def}}{=}}\log(\tanh(J))$ of the associated Ising model on this graph, with spin variable $s_c=2\tau_c-1$ attached to each plaquette representing an independent cycle $c$.
For general primal graph ${\mathcal{G}}$, different cases as illustrated on Figure \[fig:graphs\] can then be considered by increasing levels of complexity, depending on the properties of ${\mathcal{G}^\star}$. The vertices of this graph, corresponding to the elements of the cycles basis are connected via edges or more generally factors which correspond to common edges in ${\mathcal{G}}$ shared by these cycles.
If there exists a basis of disjoint cycles sharing no link in common, the partition function then factorizes as $$Z_{loops} = Z_1 {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c=1}^{C({\mathcal{G}})}(1+Q_c),$$ with $$Q_c {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{(ij)\in \ell_c}\tanh(J_{ij}),$$ the weight attached to each cycle $c$.
If one cannot find such a cycle basis, but still assuming there exists a basis such that each link belongs to at most 2 cycles and each cycle has a link in common with at most one other cycle, the partition function then reads $$\begin{aligned}
Z_{loops} &= \sum_{\tau}\prod_{c=1}^{C({\mathcal{G}})}(\bar\tau_c+\tau_cQ_c)
\prod_{c,c'}\bigl(\bar\tau_c\bar\tau_{c'}+\tau_c\bar\tau_{c'}+\bar\tau_c\tau_{c'}+
\tau_c\tau_{c'}Q_{cc'}\bigr),\label{eq:Zdual}\\[0.2cm]
&= Z_1\prod_{cc'}\Bigl(1+\frac{Q_c Q_{c'}(Q_{cc'}-1)}{(1+Q_c)(1+Q_{c'})}\Bigr){\stackrel{\mbox{\upshape\tiny def}}{=}}Z_1Z_2,\label{eq:Z2}\end{aligned}$$ where $$Q_{cc'}{\stackrel{\mbox{\upshape\tiny def}}{=}}\Bigl(\prod_{(ij)\in{\mathcal{G}}_c\cap{\mathcal{G}}_{c'}}\tanh(J_{ij})\Bigr)^{-2}.$$ When the dual graph, i.e. the graph of loops has higher interactions levels, these expressions constitute the first and second orders of approximation of a systematic cluster expansion taking into account cycle clusters of any size. The more general case where some links are common to more than 2 cycles at a time, leads to models with higher interaction order than pairwise factors. Since the interaction between cycles variables involves $\tanh(J_{ij})$ factors, we expect this dual cluster approximation to work better when the primal couplings get stronger.
To get the linear response, i.e. the susceptibility matrix, we derive the log partition function with respect to the couplings $J_{ij}$. The $0$’th order simply reads: $$\frac{\partial \log(Z_0)}{\partial J_{ij}} = \tanh(J_{ij}).$$ The first order reads: $$\frac{\partial \log(Z_1)}{\partial J_{ij}} = \frac{1-\tanh^2(J_{ij})}{\tanh(J_{ij})}
\sum_{c,(ij)\in\ell_c}\frac{Q_c}{1+Q_c}.$$ At second order different terms arise depending on whether $(ij)$ is part of one or two cycles at a time. $$\begin{aligned}
\frac{\partial \log(Z_2)}{\partial J_{ij}} &= \frac{1-\tanh^2(J_{ij})}{\tanh(J_{ij})}\Bigl(
\sum_{c,c',{\mathcal{G}}_c\cap{\mathcal{G}}_{c'}\ne\emptyset,\atop (ij)\notin{\mathcal{G}}_c\cap{\mathcal{G}}_{c'}}
\frac{Q_cQ_{c'}(Q_{cc'}-1)}{(1+Q_c)(1+Q_c+Q_{c'}+Q_cQ_{c'}Q_{cc'})}\\[0.2cm]
&-\sum_{c,c'\atop (ij)\in\ell_c\cap\ell_c'}
\frac{Q_cQ_{c'}}{1+Q_c+Q_{c'}+Q_cQ_{c'}Q_{cc'}}\bigl(
\frac{1+Q_cQ_{cc'}}{1+Q_c}+\frac{1+Q_{c'}Q_{cc'}}{1+Q_{c'}}
\bigr)\Bigr).\end{aligned}$$ This last expression involves terms of different approximation order at low temperature, since $Q_{cc'}-1\approx o(1)$ in that case. To be consistent let us keep only the second one. Summing up to order one all the preceding contributions gives finally a set of constraints to be solved of the form $$\tanh(J_{ij})+\frac{1-\tanh^2(J_{ij})}{\tanh(J_{ij})}R_{ij} = \hat \chi_{ij},$$ with $$R_{ij} =
\begin{cases}
\DD \frac{Q_c}{1+Q_c}\hspace{3.5cm}\ if\ \exists !\ c,\ s.t. (ij)\in{\mathcal{G}}_c\\[0.2cm]
\DD \frac{Q_c+Q_{c'}}{1+Q_c+Q_{c'}+Q_cQ_{c'}Q_{cc'}}\qquad (ij)\in{\mathcal{G}}_c\cap{\mathcal{G}}_{c'}.
\end{cases}$$ This quantity $R_{ij}$ is build solely on the loops containing the link $(ij)$ restricted to the subset of basic cycles and pair-combinations of basic cycles.
Pairwise cycle weights propagation
----------------------------------
This simple cluster expansion might break down rapidly when independent cycles start to accumulate to form large sized connected components in the dual graph. Nevertheless, if this graph remains singly connected, we can set up a message passing procedure to compute the exact weights. Let us first restrict the discussion to the case where there exists a cycle basis such that the dual cycle graph ${\mathcal{G}^\star}$ is pairwise. From Mac Lane’s planarity criterion this is actually equivalent to having ${\mathcal{G}}$ planar.
Since the sign of $Q_{{\mathcal{G}^\star}}(\tau)$ is not guaranteed to be positive, there is possibly no probability interpretation for these weights. Nevertheless, we can proceed analogously to ordinary belief propagation. First define the single and pair cycle’s weights: $$\begin{aligned}
q_c &{\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{1}{Z_{loops}}\sum_{\tau}\tau_c Q_{{\mathcal{G}^\star}}(\tau)\\[0.2cm]
q_{cc'} &{\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{1}{Z_{loops}}\sum_{\tau}\tau_c\tau_{c'} Q_{{\mathcal{G}^\star}}(\tau).\end{aligned}$$ From (\[eq:Zdual\]) we have $$\begin{aligned}
\chi_{ij} &= \frac{\partial\log(Z_0)}{\partial J_{ij}} +
\sum_c \frac{\partial\log(Z_{loops})}{\partial Q_c}\frac{\partial Q_c}{\partial J_{ij}}
+\sum_{cc'} \frac{\partial\log(Z_{loops})}{\partial Q_{cc'}}\frac{\partial Q_{cc'}}{\partial J_{ij}}\nonumber\\[0.2cm]
&=\tanh(J_{ij})+\sum_c \frac{q_c}{Q_c}\frac{\partial Q_c}{\partial J_{ij}}
+\sum_{cc'} \frac{q_{cc'}}{Q_{cc'}}\frac{\partial Q_{cc'}}{\partial J_{ij}}\nonumber\\[0.2cm]
&= \tanh(J_{ij})+\frac{1-\tanh^2(J_{ij})}{\tanh(J_{ij})}
\Bigl(\sum_{c,\atop (ij)\in{\mathcal{G}}c}q_c - 2\sum_{cc',\atop (ij)\in {\mathcal{G}}_c\cap{\mathcal{G}}_{c'}}q_{cc'}\Bigr).\label{eq:xijpwdual}\end{aligned}$$ Then, to compute the cycles weights, the corresponding message passing procedure involves messages of the form $$m_{c'\to c}(\tau_c) = (1-m_{c'\to c})\bar\tau_c+m_{c'\to c}\tau_c,$$ which update rules are given by $$m_{c\to c'} = \frac{1+r_{c\to c'}Q_cQ_{cc'}}{2+r_{c\to c'}Q_c(1+Q_{cc'})},$$ where $$r_{c\to c'} = \prod_{c''\in\partial c\backslash c'}\frac{m_{c''\to c}}{1-m_{c''\to c}},$$ $\partial c$ representing the neighborhood of $c$ in ${\mathcal{G}^\star}$. Finally, letting $$\nu_{c\to c'}{\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{m_{c\to c'}}{1-m_{c\to c'}},$$ leads to the following cycle weights propagation update rules: $$\begin{aligned}
\nu_{c\to c'} &\longleftarrow \frac{1+r_{c\to c'}Q_cQ_{cc'}}{1+r_{c'\to c}Q_c},\\[0.2cm]
r_{c\to c'} &\longleftarrow \prod_{c''\in\partial c\backslash c'}\nu_{c''\to c}.\end{aligned}$$ From these messages, we obtain the following expressions for the cycles weights: $$\begin{aligned}
q_c &= \frac{Q_cr_c}{1+Q_cr_c}\label{eq:qc}\\[0.2cm]
q_{cc'} &= \frac{Q_cQ_{c'}Q_{cc'}r_{c'\to c}r_{c\to c'}}
{1+Q_cr_{c\to c'}+Q_{c'}r_{c'\to c}+Q_cQ_{c'}Q_{cc'}r_{c'\to c}r_{c\to c'}},\label{eq:qcc}\end{aligned}$$ with $$r_c {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c'\in\partial c}\nu_{c'\to c}.$$ Another useful expression resulting from belief propagation equation is the partition function in terms of the single and pairwise beliefs normalizations. Introducing also $$s_c{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c'\in\partial c}(1+\nu_{c'\to c})\qquad\text{and}\qquad
s_{c'\to c}{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c''\in\partial c'\backslash c}(1+\nu_{c''\to c'}),$$ we have $$Z_{loops} = \prod_cZ_c\prod_{(cc')\in\cal T}\frac{Z_{cc'}}{Z_cZ_{c'}}.$$ with $$\begin{aligned}
Z_c &= \frac{1+Q_cr_c}{s_c}\label{eq:localzc}\\[0.2cm]
Z_{cc'} &= \frac{1+Q_cr_{c\to c'}+Q_{c'}r_{c'\to c}+Q_cQ_{c'}Q_{cc'}r_{c'\to c}r_{c\to c'}}
{s_{c\to c'}s_{c'\to c}}.\label{eq:localzcc}\end{aligned}$$
General cycle weights propagation
---------------------------------
Note that any two cycles having vertices but no edges in common do not interact, so it is maybe not that obvious that a graph having a singly connected dual graph must be planar. However since the complete graph $K_5$ shown on Figure \[fig:graphs\].c and also the bipartite graph $K_{3,3}$ have non-planar dual graphs, from Kuratowski characterisation of planar graphs, this is indeed likely to be the case. For planar graphs, exact methods have been proposed in the literature based on Pfaffian’s decompositions of the partition function [@GlJa; @chertkov4] with a computational cost of $O(N^3)$. At least, for the subclass of factor graph that we consider, the computational cost becomes linear in the number of cycles which scales anyway like $O(N)$ for planar graphs. As far as exact determination of $Z$ is concerned, the pairwise cycle weights propagation described in the preceding section should be suitable in all relevant cases. However, in practice, finding the proper cycle basis might not be always an easy task. Also, by analogy with loopy belief propagation, we don’t want to limit ourselves to exact cases, and propagating weights on loopy dual graphs could lead possibly to interesting approximate results even for non-planar graphs.
So let us consider the case where some edges are shared by more than two cycles. The dual factor graph is constructed by associating one factor to each such edge in addition to the ones already shared by exactly two cycles. The dual loop partition function then reads $$Z_{loops} = \sum_{\tau}\prod_{c=1}^{C({\mathcal{G}})}(\bar\tau_c+\tau_cQ_c)
\times\prod_{e\in{\mathcal{E}}}
\left[\sum_{k=0}^{d^\star(e)}
\delta\bigl(k - \sum_{c,{\mathcal{G}}_c\ni e}\tau_c\bigr)\tanh(J_e)^{-2\lfloor k/2\rfloor}\right],\label{eq:Zloops}$$ where $\lfloor x\rfloor$ denotes the entire part of $x$, $e$ indexes any edge in the original graph ${\mathcal{G}}$ with $J_e$ the corresponding coupling, while $d^\star(e)$ is the degree of the factor associated to $e$ in the dual graph ${\mathcal{G}^\star}$, that is the number of cycles containing $e$. In this expression the factor $\tanh(J_e)^{-2\lfloor k/2\rfloor}$ is there to compensate for overcounting the edge factor $\tanh(J_e)$ when $k$ cycles containing this edge are taken into account. Note that if two or more edges are shared exactly by the same set of cycles, they should be gathered into a single factor, with $\tanh(J_e)$ simply replaced by the product of hyperbolic tangents corresponding to these edges in the above formula.
From the expression (\[eq:Zloops\]), the susceptibility coefficient associated to any edge $e\in{\mathcal{E}}$, reads $$\chi_e = \tanh(J_e) + \frac{1-\tanh^2(J_e)}{\tanh(J_e)}\Bigl(
\sum_{c\in e}q_c - \sum_{k=2}^{d^\star(e)}2\lfloor \frac{k}{2}\rfloor q_{e,k}\Bigr),\label{eq:linearesponse1}$$ where in this case joint weights of interest are given by $$q_{e,k} {\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{1}{Z}\sum_{\tau} Q_{{\mathcal{G}^\star}}(\tau)\delta\bigl(k - \sum_{c,{\mathcal{G}}_c\ni e}\tau_c\bigr).$$ This generalizes the expression (\[eq:xijpwdual\]) obtained for the pairwise dual graph to arbitrary dual factor graph.
The corresponding message passing algorithm allowing one to compute the weights $q_c$ is now expressed in terms of messages $\nu_{e\to c}$ and $r_{c\to e}$ (defined as before from the original messages $m_{e\to c}$) respectively from factors to nodes and from nodes to factors. The update rules read $$\begin{aligned}
\nu_{e\to c} &\longleftarrow \frac{1+\sum_{k=1}^{d^\star(e)-1}
\sum_{c_1,\ldots c_{k}\backslash c\atop e\in {\mathcal{G}}_{c_i}}\prod_{i=1}^k Q_{c_i}r_{c_i\to e}\ \tanh(J_e)^{-2\lfloor (k+1)/2\rfloor}}
{1+\sum_{k=1}^{d^\star(e)-1}
\sum_{c_1,\ldots c_{k}\backslash c\atop e\in {\mathcal{G}}_{c_i}}
\prod_{i=1}^k Q_{c_i}r_{c_i\to e}\ \tanh(J_e)^{-2\lfloor k/2\rfloor}},\label{eq:gcwpup}\\[0.2cm]
r_{c\to e} &\longleftarrow \prod_{e'\ni c\backslash e}\nu_{e\to c}.\nonumber\end{aligned}$$ In terms of the messages, after adapting the definition of $r_c$, the expression (\[eq:qc\]) of $q_c$ remains valid while (\[eq:qcc\]) generalizes to the new weights: $$q_{e,k} = \frac{\sum_{c_1,\ldots c_{k}\atop e\in {\mathcal{G}}_{c_i}}\prod_{i=1}^k Q_{c_i}r_{c_i\to e}\ \tanh(J_e)^{-2\lfloor (k+1)/2\rfloor}}
{1+\sum_{k=1}^{d^\star(e)}
\sum_{c_1,\ldots c_{k}\atop e\in {\mathcal{G}}_{c_i}}
\prod_{i=1}^k Q_{c_i}r_{c_i\to e}\ \tanh(J_e)^{-2\lfloor k/2\rfloor}}$$ which coincides with (\[eq:qcc\]) for $k=2$. Concerning local partition functions, (\[eq:localzc\]) remains unchanged after adapting the definition of $r_c$ and $s_c$, while (\[eq:localzcc\]) generalizes to $$Z_e = \frac{1+ \sum_{k=1}^{d^\star(e)}
\sum_{c_1,\ldots c_{k}\atop e\in {\mathcal{G}}_{c_i}}\prod_{i=1}^k Q_{c_i}r_{c_i\to e}\ \tanh(J_e)^{-2\lfloor (k+1)/2\rfloor}}
{\prod_{i=1}^k s_{c_i\to e}}$$ so that we finally have $$Z_{loops} = \prod_e\frac{Z_e}{\prod_{c\in e}Z_c}\prod_c Z_c,$$ when the dual graph is singly connected.
Extended pairwise dual-graph and dual weight propagation
--------------------------------------------------------
When the degree of factors get larger, the combinatorial burden to evaluate $q_{e,k}$ and associated messages for $k\gg 1$, becomes intractable. Coming back to (\[eq:Zloops\]) let us remark first the following simplification in the way to write each factor: $$\begin{aligned}
\sum_{k=0}^{d^\star(e)}
\delta\bigl(k - \sum_{c,e\in{\mathcal{G}}_c}\tau_c\bigr)t_e^{-2\lfloor k/2\rfloor} &= \frac{1}{2}\Bigl[
\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c+\tau_ct_e^{-1})+\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c-\tau_ct_e^{-1})\\[0.2cm]
&+ t_e\Bigl(\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c+\tau_ct_e^{-1})-\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c-\tau_ct_e^{-1})\Bigr)\Bigr],\\[0.2cm]
= \frac{1}{2}\Bigl[\prod_{c,e\in{\mathcal{G}}_c}&(\bar\tau_c+\tau_ct_e^{-1})(1+t_e)+
\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c-\tau_ct_e^{-1})(1-t_e)\Bigr],\end{aligned}$$ after separating the odd and even part in $k$, with the notation $t_e{\stackrel{\mbox{\upshape\tiny def}}{=}}\tanh(J_e)$. This suggests the introduction of an additional binary variable $\sigma_e\in\{-1,1\}$ associated to each factor $e$, such that the loop partition function now reads $$Z_{loops} = \sum_{\tau,\sigma} Q_{{\mathcal{G}}^\star}(\tau,\sigma),\label{eq:Zdwp}$$ with $$Q_{{\mathcal{G}}^\star}(\tau,\sigma) {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c=1}^{C({\mathcal{G}})}(\bar\tau_c+\tau_cQ_c)
\times\prod_{e\in{\mathcal{E}}}\frac{1+\sigma_e t_e}{2}
\prod_{c,e\in{\mathcal{G}}_c}(\bar\tau_c+\tau_c\frac{\sigma_e}{t_e}),\label{def:Qtausigma}$$ i.e. expressing it as a sum over cycles and edges binary variables, of a joint weight measure corresponding to an extended pairwise factor graph, containing cycle-edges interactions. Concerning the susceptibility this new formulation leads to a simpler expression. Indeed, for any edge $(ij)=e\in{\mathcal{G}}$, after deriving $Z({{\bf J}})$ with respect to $J_e$ and arranging the terms we finally obtain $$\chi_{e} = 2q_e -1,\label{eq:linearesponse2}$$ with $$q_e {\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{1}{Z_{loops}} \sum_{\tau,\sigma}\frac{1+\sigma_e}{2} Q_{{\mathcal{G}}^\star}(\tau,\sigma).$$ In the case where $e$ is part of a larger factor $f$ containing other edges shared exactly by the same set of cycles, this formula should be slightly modified. The susceptibility $\chi_e$ is then expressed in terms of the weight $q_f$ associated to this factor and reads: $$\chi_e = \frac{t_e^2-t_f^2}{t_e(1-t_f^2)}+\frac{t_f}{t_e}\frac{1-t_e^2}{1-t_f^2}\bigl(2q_f-1),$$ where $t_f$ is now the product of $\tanh(J_{e'})$ of all edges $e'$ represented by this factor, including $e$. Note that there is no approximation at this point. $2q_e-1$ represents the dual "magnetization” associated to variables $\sigma_e$. ${\mathcal{G}^\star}$ has the same structure as before, except that factors have been replaced by vertices.
Assuming ${\mathcal{G}^\star}$ singly connected we may again settle a message passing procedure in order to compute these weights. We have now to distinguish between messages $m_{c\to e}(\sigma_e)$ sent by vertex’s cycles to vertex’s edges and $m_{e\to c}(\tau_c)$ sent by edges to cycles vertices. Letting $$\nu_{e\to c} {\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{m_{e\to c}(\tau_c=1)}{m_{e\to c}(\tau_c=0)}\qquad\text{and}\qquad
\nu_{c\to e} {\stackrel{\mbox{\upshape\tiny def}}{=}}\frac{m_{c\to e}(\sigma_e=1)}{m_{c\to e}(\sigma_c = -1)},$$ we come up with the following update rules: $$\begin{cases}
\DD \nu_{e\to c} \longleftarrow \frac{r_{e\to c}e^{2J_e}-1}{r_{e\to c}e^{2J_e}+1}\ \coth(J_e),\\[0.4cm]
\DD \nu_{c\to e} \longleftarrow \frac{1+r_{c\to e} Q_c\coth(J_e)}{1-r_{c\to e} Q_c\coth(J_e)},\\[0.4cm]
\DD r_{e\to c} \longleftarrow \prod_{c'\in\partial e\backslash c}\nu_{c'\to e}.\\[0.4cm]
\DD r_{c\to e} \longleftarrow \prod_{e'\in\partial c\backslash e}\nu_{e'\to c}.
\end{cases}\label{eq:dwpupdate}$$ After convergence we get for the edge weights: $$q_e = \frac{e^{2J_e}\ r_e}{1+e^{2J_e}\ r_e},\qquad\text{with}\qquad r_e {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c\in\partial e}\nu_{c\to e}.$$ Finally to compute the dual partition function we need as before to compute local ones which read: $$\begin{cases}
\DD Z_c = \frac{1+Q_cr_c}{s_c}\\[0.4cm]
\DD Z_e = \frac{e^{-J_e}+e^{J_e}r_e}{2s_e\cosh(J_e)}\\[0.4cm]
\DD Z_{ec} =
\frac{e^{-J_e}+e^{J_e}r_{e\to c}-Q_ce^{-J_e} \coth(J_e)r_{c\to e}+Q_ce^{J_e}\coth(J_e)r_{e\to c}r_{c\to e}}{2s_{e\to c}s_{c\to e}\cosh(J_e)},
\end{cases}$$ with now $$\begin{aligned}
&s_c{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{e\in\partial c}(1+\nu_{e\to c}),\qquad
s_e{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c\in\partial e}(1+\nu_{c\to e}),\qquad
r_{c\to e}{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{e'\in\partial c\backslash e}\nu_{e'\to c},\\[0.2cm]
&s_{c\to e}{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{e'\in\partial c\backslash e}(1+\nu_{e'\to c}),\qquad\text{and\qquad}
s_{e\to c}{\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{c'\in\partial e\backslash c}(1+\nu_{c'\to e}),\end{aligned}$$ to obtain $$Z_{loops} = \prod_{(e,c)\in{\mathcal{E}}^\star}\frac{Z_{ec}}{Z_eZ_c}\prod_{c\in{\mathcal{G}}^\star}Z_c\prod_{e\in{\mathcal{G}}^\star}Z_e.$$ In practice when the connectivity $d^\star(e)$ of an edge variable $e$ is not greater than $2$, the corresponding variables $\sigma_e$ maybe summed out beforehand, so that finally the joint weight measure associated to $Z_{loops}$ involves cycle-cycle and cycle-edges interactions given in (\[eq:Zdual\]) and (\[def:Qtausigma\]).
Linear response theory
----------------------
Using this extended pairwise dual model and the corresponding fixed point of DWP, we can derive the linear response theory for any pair of nodes $i\in{\mathcal{G}}$ and $j\in{\mathcal{G}}$, when the dual graph ${\mathcal{G}^\star}\equiv {\mathcal{T}}$ forms a tree.
First, for any edge $e\in{\mathcal{E}}$, either of expression (\[eq:xijpwdual\]), (\[eq:linearesponse1\]) and (\[eq:linearesponse2\]) can be used in principle to determine $\chi_e$, except that the first one requires a good choice of the cycle basis, while the others two need to care less about it, as long as the dual graph remains singly connected.
If $e\notin{\mathcal{E}}$, then let us give the index $0$ to the new independent cycle which is formed by adding $e$ to the initial graph ${\mathcal{G}}$ with some arbitrary coupling $J_e$ (see Figure \[fig:lineresp\]). The corresponding susceptibility then reads $$\chi_e = \frac{d\log Z_{loops}(J_e)}{dJ_e}\Big\vert_{J_e=0},$$ where, from (\[eq:Zdwp\],\[def:Qtausigma\]), $$Z_{loops}(J_e) = \sum_{\tau,\sigma,\tau_0} Q_{{\mathcal{G}}^\star}(\tau,\sigma)\bigl(\bar\tau_0+\tau_0Q_0(J_e)\bigr)
\prod_{e'\in{\mathcal{G}}_0\backslash e}\bigl(\bar\tau_0+\tau_0\frac{\sigma_{e'}}{t_{e'}}\bigr),$$ with cycle factor $$Q_0(J_e) {\stackrel{\mbox{\upshape\tiny def}}{=}}\prod_{(ij)\in{\mathcal{G}}_0} \tanh(J_{ij}).$$ Let $$Q_{0\backslash e} {\stackrel{\mbox{\upshape\tiny def}}{=}}Q_0(J_e)/\tanh(J_e),$$ the cycle weight, where the edge $e$ is not taken into account, which by definition is independent of $J_e$. We have now $$\begin{aligned}
\chi_e &= \frac{Q_{0\backslash e}}{Z_{loops(0)}}\sum_{\tau,\sigma}Q_{{\mathcal{G}}^\star}(\tau,\sigma)
\prod_{e'\in{\mathcal{G}}_0\backslash e}\frac{\sigma_{e'}}{t_{e'}}\\[0.2cm]
&{\stackrel{\mbox{\upshape\tiny def}}{=}}<\prod_{e'\in{\mathcal{G}}_0\backslash e}\sigma_{e'}>_{Q_{{\mathcal{G}}^\star}},\end{aligned}$$ where this last quantity represents “pseudo moment” of $\sigma_{e'}$ variables under the weight measure $Q_{{\mathcal{G}}^\star}$ given in (\[def:Qtausigma\])).
Various cases may be considered, depending on how the additional cycle node is connected to ${\mathcal{G}^\star}$. In absence of connection, all variables $\sigma_{e'}$ in the above expression are independent and free, i.e. $2q_{e'}-1 = \tanh(J_{e'})$ so we get $$\chi_e = Q_{0\backslash e}.$$ Next, if the extra cycle variable is connected to a single factor $f$, the susceptibility reads $$\chi_e = \frac{(2q_f-1)}{t_f} Q_{0\backslash e}.$$ When more than one factor $f_1,\ldots,f_k$ are involved, as in the example of Figure \[fig:lineresp\], corresponding variables $\sigma_{f_i}$ are not independent and $$\chi_e \approx Q_{0\backslash e}\prod_{i=1}^k \frac{2q_{f_i}-1}{t_{f_i}},$$ can lead to a poor approximation if the dependencies between these variables are not small. Nevertheless, it is still possible from the analysis of Section \[sec:cumulant\] to take into account these “correlations”. The explicit formulas given in this latter section make only use of the factorization (\[eq:ansatz\]) of the joint measure in addition to the tree structure property of the factor graph. They are limited to $4$-points susceptibility coefficients, but could be in principle generalized to any cluster size. They are still formally valid in the present context, where the knowledge of the DWP fixed point gives directly exact dual pairwise weight’s “susceptibilities” $\chi_{cc'}^{\star}$ or $\chi_{ce}^{\star}$, for any pair $(c,c')\in{\mathcal{E}}^\star$ or $(c,e)\in{\mathcal{E}}^\star$ forming an edge in ${\mathcal{G}}^\star$. Therefore, in principle the complete determination of the linear response can be done with DWP, with a computational cost not exceeding $O(N^2)$. The limitation here is simply the size of the largest cycles, which then possibly requires to have explicit expressions of dual susceptibilities of corresponding cluster size, which in effect is arbitrarily limited to $4$ in the present work.
$L_0$ norm penalized sparse inverse estimation algorithm {#sec:penal}
========================================================
We propose here to use the Doubly Augmented Lagrange (DAL) method [@IusemDAL; @EcksteinMOR; @DongJSC] to solve the penalized log-determinant programming in (\[eq:logdetopt\]). For a general problem defined as follows: $$\label{eq:dalbasicpro}
\min_x F(x) = f(x) + g(x)$$ where $f(x)$ and $g(x)$ are both convex. DAL splits the combination of $f(x)$ and $g(x)$ by introducing a new auxiliary variable $y$. Thus, the original convex programming problem can be formulated as : $$\label{eq:dalbasicsplit}
\begin{aligned}
\min_{x,y} F(x) = f(x) + g(y) \\
\text{s.t. } \; \; {x - y = 0}
\end{aligned}$$ Then it advocates an augmented Lagrangian method to the extended cost function in (\[eq:dalbasicsplit\]). Given penalty parameters $\mu$ and $\gamma$, it minimizes the augmented Lagrangian function $$\label{eq:dalbasicsfunc}
L(x,y,\nu,\tilde{x},\tilde{y})
= f(x) + g(y) + \langle \nu, x-y \rangle
+ \frac{\mu}{2} \|x-y\|_{2}^2
+ \frac{\gamma}{2} \|x-\tilde{x}\|_{2}^2
+ \frac{\gamma}{2} \|y-\tilde{y}\|_{2}^2$$ where $\tilde{x}$ and $\tilde{y}$ are the prior guesses of $x$ and $y$ that can obtained either from a proper initialization or the estimated result in the last round of iteration in an iterative update procedure. Since optimizing jointly with respect to $x$ and $y$ is usually difficult, DAL optimizes $x$ and $y$ alternatively. That gives the following iterative alternative update algorithm with some simple manipulations: $$\label{eq:dalaltfunc}
\begin{aligned}
x^{k+1} &= \min_x f(x) + \frac{\mu}{2}\|{x - y^k + {\tilde{\nu}}^k}\|_{2}^2
+ \frac{\gamma}{2}\|x-x^k\|_{2}^2 \\
y^{k+1} &= \min_y g(y) + \frac{\mu}{2}\|{x^{k+1} - y + {\tilde{\nu}}^k}\|_{2}^2 + \frac{\gamma}{2}\|y-y^k\|_{2}^2 \\
{\tilde{\nu}}^{k+1} &= {\tilde{\nu}}^k + x^{k+1} - y^{k+1}
\end{aligned}$$
where $\tilde{\nu} = {\frac{1}{\mu}}{\nu}$. As denoted in [@DongJSC] and [@IusemDAL], DAL improves basic augmented Lagrangian optimization by performing additional smooth regularization on estimations of $x$ and $y$ in successive iteration steps. As a result, it guarantees not only the convergence of the scaled dual variable $\tilde{\nu}$, but also that of the proximal variables $x^k$ and $y^k$, which could be divergent in basic augmented Lagrangian method.
We return now to the penalized log-determinant programming in sparse inverse estimation problem, as seen in (\[eq:logdetopt\]). The challenge of optimizing the cost function is twofold. Firstly, the exact $L_0$-norm penalty is non-differentiable, making it difficult to find an analytic form of gradient for optimization. Furthermore, due to the log-determinant term in the cost function, it implicitly requires that any feasible solution to the sparse approximation $A$ of the precision matrix should be strictly positive definite. The gradient of the log-determinant term is given by $\hat C - A^{-1}$, which is not continuous in the positive definite domain and makes it impossible to obtain any second-order derivative information to speed up the gradient descent procedure. We hereafter use $S_{++}$ as the symmetric positive definite symmetric matrices that form the feasible solution set for this problem. By applying DAL to the cost function (\[eq:logdetopt\]), we can derive the following formulation: $$\label{eq:dalopt}
\begin{aligned}
\hat{J}(A,Z,\tilde{A}, \tilde{Z},\nu)
=& -\log\det(A) + \operatorname{Tr}(\hat C A) + \lambda P(Z) + \langle \nu , A-Z \rangle\\
& + \frac{\mu}{2} \|A-Z\|_{2}^2
+ \frac{\gamma}{2} \|A-\tilde{A}\|_{2}^2
+ \frac{\gamma}{2} \|Z-\tilde{Z}\|_{2}^2 \\
\text{s.t.} & \;\; {A, Z} \in S_{++}
\end{aligned}$$ where $Z$ is the auxiliary variable that has the same dimension as the sparse inverse estimation $A$. $\tilde{A}$ and $\tilde{Z}$ are the estimated values of $A$ and $Z$ derived in the last iteration step. The penalty parameter $\gamma$ controls the regularity of $A$ and $Z$. By optimizing $A$ and $Z$ alternatively, the DAL procedure can be easily formulated as an iterative process as follows, for some $\delta > 0$: $$\label{eq:daliter}
\begin{aligned}
A^{k+1} &=\operatorname*{argmin}_A -\log\det(A) + \operatorname{Tr}(\hat C A) + \lambda P({Z}^{k}) + \langle {\nu}^{k} , A-{Z^k} \rangle \\
&+ \frac{\mu}{2} \|A-{Z^k}\|_{2}^2 + \frac{\gamma}{2}\|A-{A^k}\|_{2}^2 \\
Z^{k+1} &=\operatorname*{argmin}_Z \lambda P(Z) + \langle \nu^{k} , A^{k+1}-Z\rangle
+ \frac{\mu}{2} \|A^{k+1}-Z\|_{2}^2 \\
&+ \frac{\gamma}{2}\|Z-{Z^k}\|_{2}^2 \\
& {\nu}^{k+1} = {\nu}^{k} + \delta( A^{k+1} - Z^{k+1}) \\
\text{s.t.} & \; \; A^{k+1}, Z^{k+1} \in S_{++}
\end{aligned}$$
By introducing the auxiliary variable $Z$, the original penalized maximum likelihood problem is decomposed into two parts. The first one is composed mainly by the convex log-determinant programming term. Non-convex penalty is absorbed into the left part. Separating the likelihood function and the penalty leads to the simpler sub-problems of solving log-determinant programming using eigenvalue decomposition and $L_0$ norm penalized sparse learning alternatively. Each sub-problem contains only one single variable, making it applicable to call gradient descent operation to search local optimum. Taking $\tilde{\nu} = {\frac{1}{\mu}}{\nu}$, we can derive the following scaled version of DAL for the penalized log-determinant programming: $$\label{eq:dalscale}
\begin{aligned}
A^{k+1} &=\operatorname*{argmin}_A -\log\det(A) + \operatorname{Tr}({\hat C}A)
+ \frac{\mu}{2} \bigl\|A-Z^{k} + \tilde{\nu}^k\bigr\|_{2}^2
+ \frac{\gamma}{2} \bigl\|A-A^{k}\bigr\|_{2}^2 \\
Z^{k+1} &=\operatorname*{argmin}_Z \frac{\mu}{2} \bigl\|A^{k+1}-Z + \tilde{{\nu}}^{k}\bigr\|_{2}^2
+ \frac{\gamma}{2} \bigl\|Z-Z^{k}\bigr\|_{2}^2 + \lambda P(Z) \\
\tilde{{\nu}}^{k+1} &= \tilde{{\nu}}^{k} + A^{k+1} - Z^{k+1} \\
\text{s.t.} & \; \; {A^{k+1}, Z^{k+1}} \in S_{++}
\end{aligned}$$
To attack the challenge caused by non-differentiability of the exact $L_0$ norm penalty, we make use of a differentiable approximation to $L_0$-norm penalty in the cost function $\hat{J}$, named as "seamless $L_0$ penalty” (SELO) in [@DickerSELO]. The basic definition of this penalty term is given as: $$\label{eq:selopenalty}
P_{\text{SELO}}(Z) = \sum_{i,j} {\frac{1}{\log(2)}} \log( 1 + {\frac{|Z_{i,j}|}{|Z_{i,j}|+{\tau}}} )$$ where $Z_{i,j}$ denotes individual entry in the matrix $Z$ and ${\tau}
> 0$ is a tuning parameter. As seen in Figure \[fig:L0L1penfuncf\], as $\tau$ gets smaller, $P(Z_{i,j})$ approximates better the $L_0$ norm $I(Z_{i,j} \not= 0)$. SELO penalty is differentiable, thus we can calculate the gradient of $P(Z)$ explicitly with respect to each $Z_{i,j}$ and make use of first-order optimality condition to search local optimum solution. Due to its continuous property, it is more stable than the exact $L_0$ norm penalty in optimization. As proved in [@DickerSELO], the SELO penalty has the oracle property with proper setting of $\tau$. That’s to say, the SELO penalty is asymptotically normal with the same asymptotic variance as the unbiased OLS estimator in terms of Least Square Estimation problem. Furthermore, if we perform local first-order Taylor expansion to the SELO penalty term, we can find the intrinsic relation between the SELO penalty between robust M-estimator, which explains the stability of SELO against noise. We describe this part in the followings.
The first two steps in (\[eq:dalscale\]) are performed with the positive definite constrains imposed on $A$ and $Z$. The minimizing with respect to $A$ is accomplished easily by performing Singular Vector Decomposition (SVD). By calculating the gradient of $\hat{J}$ with respect to A in (\[eq:dalscale\]), based on the first-order optimality, we derive: $$\label{eq:svdsol1}
{\hat C} - A^{-1} + {\mu} (A - Z^{k} + \tilde{{\nu}}^{k}) + {\gamma}(A - A^{k}) = 0$$
Based on generalized eigenvalue decomposition, it is easy to verify that ${A^{k+1}} = V\operatorname{Diag}(\beta){V^{T}}$, where $V$ and $\{d_i\}$ are the eigenvectors and eigenvalues of ${\mu} (Z^k - \tilde{\nu}^k) - {\hat C} + {\gamma}{A^k}$. ${\beta}_{i}$ is defined as: $$\label{eq:svdsol2}
{\beta}_{i} = \frac{d_i + \sqrt[]{{d_i}^2 + 4(\tilde{{\nu}}+{\gamma})}}{2(\tilde{\nu} + {\gamma})}$$
Imposing $Z\in S_{++}$ directly in minimizing the cost function with respect to $Y$ make the optimization difficult to solve. Thus, instead, we can derive a feasible solution to $Z$ by a continuous search on $\mu$. Based on spectral decomposition, it is clear that $X^{k+1}$ is guaranteed to be positive definite, while it is not necessarily sparse. In contrast, $Z$ is regularized to be sparse while not guaranteed to be positive definite. $\mu$ is the regularization parameter controlling the margin between the estimated $X^{k+1}$ and the sparse $Z^{k+1}$. Increasingly larger $\mu$ during iterations makes the sequences $\{X^{k}\}$ and $\{Z^{k}\}$ converge to the same point gradually by reducing margin between them. Thus, with enough iteration steps, the derived $Z^k$ follows the positive definite constraint and sparsity constraint at the same time. We choose here to increase $\mu$ geometrically with a positive factor ${\eta} > 1$ after every $N_{\mu}$ iterations until its value achieves a predefined upper bound $\mu_{\max}$. With this idea, the iterative DAL solution to the $L_0$ norm penalty is given as: $$\label{eq:dalselo}
\begin{aligned}
A^{k+1} &=\operatorname*{argmin}_A -\log\det(A) + \operatorname{Tr}({\hat C}A)
+ \frac{\mu}{2} \bigl\|A-Z^{k} + \tilde{{\nu}}^{k}\bigr\|_{2}^2
+ \frac{\gamma}{2} \bigl\|A-A^{k}\bigr\|_{2}^2, \\
Z^{k+1} &=\operatorname*{argmin}_Z \frac{\mu}{2} \bigl\|A^{k+1}-Z + \tilde{{\nu}}^{k}\bigr\|_{2}^2
+ \frac{\gamma}{2} \bigl\|Z-Z^{k}\bigr\|_{2}^2 + \lambda P(Z)\\
\tilde{{\nu}}^{k+1} &= \tilde{{\nu}}^{k} + A^{k+1} - Z^{k+1} \\
\mu^{k+1} &= \min\bigl(\mu\,\eta^{\lfloor k/N_\mu\rfloor},\mu_{\max}\bigr).
\end{aligned}$$
By alternatively optimizing w.r.t $A$ and the auxiliary matrix $Z$, we reduce margin between $A$ and $Z$ gradually, which finally leads to estimation of sparse graph structure and maximizing likelihood of the sparse graph simultaneously. As we can find in the second step of (\[eq:dalselo\]), the guess of sparse graph structure is obtained by solving penalized least square regression w.r.t $Z$. Besides its analytical convenience, this compact functional form also provides a clear view about intrinsic link between the SELO penalized sparse inverse and robust M-estimator, which indicates superior stability of the approximated L0 norm penalty over lasso penalization. Due to concavity of the SELO penalty , we can make use of a Taylor expansion of the SELO penalty as its tight upper bound within a local neighborhood. Therefore, minimizing the linear upper bound given good initialization point can generate the same optimization path as minimizing the penalty term directly. Based on this equivalency relation, we replace $P(Z)$ in (\[eq:dalselo\]) with its first order Taylor expansion $g$, which reads:
$$\label{eq:robustselo}
\begin{aligned}
Z^{k+1} &=\operatorname*{argmin}_Z \frac{\mu}{2} \bigl\|Z - A^{k+1} + \tilde{{\nu}}^{k}\bigr\|_{2}^2
+ \frac{\gamma}{2} \bigl\|Z-Z^{k}\bigr\|_{2}^2 + \lambda \sum_{i,j} {g(Z_{i,j})}\\
\end{aligned}$$
where $g(Z_{i,j})$ is defined as $$g(Z_{i,j}) = P(Z_{i,j}^{\text{old}}) + \frac{\tau sgn(Z_{i,j}^{\text{old}})}{\log(2)
(\tau + \|Z_{i,j}^{\text{old}}\|) (\tau + 2\|Z_{i,j}^{\text{old}}\|)} (Z_{i,j} - Z_{i,j}^{\text{old}})$$
$sgn(Z_{i,j})$ is the sign of the scalar value $Z_{i,j}$ and $Z_{i,j}^{\text{old}}$ is a given initializing guess of $Z_{i,j}$, which is a constant term. Minimizing the first-order Taylor expansion w.r.t $Z$ is equal to solve a weighted linear regression, with the weight of each $Z_{i,j}$ $w_{i,j} = \frac{\tau sgn(Z_{i,j}^{\text{old}})}{\log(2) (\tau + \|Z_{i,j}^{\text{old}}\|) (\tau + 2\|Z_{i,j}^{\text{old}}\|)}$. The entry-wise weight value declines to zero quickly when the magnitude of the corresponding $Z_{i,j}$ increases. With such weight configuration, only $Z_{i,j}$ approaching to zero has strong influence to the gradient direction, which pursues a self-adaptive pruning of entries with small magnitudes. $Z_{i,j}$ with large magnitudes are free from unnecessary sparsity penalization. This weight setting is intrinsically consistent with the basic idea of robust M-estimator. According to [@ChoQUIC], robust M-estimator is designed to avoid penalizing outliers inducing large bias to the model. Through under-weighting outlier training data, the derived M-estimator gains stability against noisy data. In our case, $Z_{i,j}$ with underlying large non-zero magnitude encodes critical correlation between random variables. Forcing these terms to be sparse thus results in artifacts in the graph structure estimation. In this sense, the SELO penalty benefits from the consistent design to the robust statistics and obtains stable estimation of sparse correlation structure in the graph. In fact, the weight setting in the first order Taylor expansion is a close approximate to the robust M-estimator with its $\psi$ function defined by Cauchy distribution. It verifies the intrinsic robustness of SELO penalty in a further step. In contrast, the first order expansion of the $L1$ norm penalty $\|Z\|$ gives constant weight to all entries in the linear term. Therefore, compared with the SELO penalty, $L1$ norm penalty is likely to result in bias in graph structure estimation.
In the second step of (\[eq:dalselo\]), we calculate the gradient of the cost function with respect to $Z$ and achieve the local minimum by performing the first-order optimum condition on it. Therefore, the updated value of each entry of $Z$ is given by a root of a cubic equation, as defined below: $$\label{eq:dalzupdate}
\begin{aligned}
\text{if} & \; {Z_{i,j} > 0}, \; Z_{i,j}\text{ is the positive root of} \\
& 2{Z_{i,j}}^{3} + (3{\tau} - 2{{\theta}_{i,j}}){Z_{i,j}}^{2} +
({\tau}^2 - 3{\tau}{\theta_{i,j}}){Z_{i,j}} - {{\tau}^{2}}{\theta_{i,j}} + \frac{{\lambda}{\tau}}{{\mu} + {\gamma}} = 0\\
\text{if} & \; {Z_{i,j} < 0}, \; Z_{i,j}\text{ is the negative root of} \\
& 2{Z_{i,j}}^{3} - (3{\tau} + 2{{\theta}_{i,j}}){Z_{i,j}}^{2} +
({\tau}^2 + 3{\tau}{\theta_{i,j}}){Z_{i,j}} - {{\tau}^{2}}{\theta_{i,j}} - \frac{{\lambda}{\tau}}{{\mu} + {\gamma}} = 0\\
&\text{else} \; {Z_{i,j} = 0}
\end{aligned}$$ where $Z_{i,j}$ is one single entry of $Z$ and $$\theta_{i,j} =\frac{\gamma Z_{i,j}^k + \mu(A_{i,j}^{k+1} + {\tilde{\nu}}^k)}
{\mu +\gamma}.$$ Solving the cubic equations can be done rapidly using Cardano’s formula within a time cost $O(n^2)$. Besides, the spectral decomposition procedure has the general time cost $O(n^3)$. Given the total number of iterations $K$, theoretical computation complexity of DAL is $O(K{n}^3)$. For our experiments, we initialize $\mu$ to $0.06$, the multiplier factor $\eta$ to $1.3$ and the regularization penalty parameter $\gamma$ to $10^{-4}$. To approximate the $L_0$ norm penalty, $\tau$ is set to be $5\cdot 10^{-4}$. In our experiment, to derive the Pareto curve of the optimization result, we traverse different values of $\lambda$. Most learning procedures converge with no more than $K=500$ iteration steps.
To validate performance of sparse inverse estimation based on the $L_0$ norm penalty, we involve an alternative sparse inverse matrix learning method using $L_1$ norm penalization for comparison. Taking $P(A)$ in (\[eq:logdetopt\]) to be the $L_1$ matrix norm of $A$, we strengthen conditional dependence structure between random variables by jointly minimizing the negative log likelihood function and the $L_1$ norm penalty of the inverse matrix. Since $L_1$ norm penalty is strictly convex, we can use a quadratic approximation to the cost function to search for the global optimum, which avoids singular vector decomposition with complexity of $O(p^3)$ and improves the computational efficiency of this solution to $O(p)$, where $p$ is the number of random variables in the GMRF model. This quadratic approximation based sparse inverse matrix learning is given in [@ChoQUIC], named as QUIC. We perform it directly on the empirical covariance matrix with different settings of the regularization coefficient $\lambda$. According to works in compressed sensing, the equality between $L_1$ norm penalty and $L_0$ norm penalty holds if and only if the design matrix satisfies restricted isometry property. However, restricted isometry property is sometimes too strong in practical case. Furthermore, to our best knowledge, there is no similar necessary condition guaranteeing equivalence between $L_1$ and $L_0$ norm penalty in sparse inverse estimation problem. Therefore, in our case, $L_1$ norm penalized log-determinant programming is highly likely to be biased from the underlying sparse correlation structure in the graph, which leads to much denser inverse matrices.
Experiments {#sec:experiments}
===========
In this section, various solutions based on the different methods exposed before are compared.
#### Inverse Ising problem
Let us start with the inverse Ising problem. The first set of experiments illustrates how the linear-response approach exposed in Section \[sec:prelim\] works when the underlying model to be found is itself an Ising model.
![\[fig:invising\] Comparison between various approximate solutions to the inverse Ising problem. RMSE errors as a function of the temperature are plotted in (a) and (b) for the couplings $J_{ij}$, in (c) and (d) for the susceptibility matrix $\chi_{ij}$ obtained from the corresponding BP fixed point. Local fields $h_i$ are zero in (a) and (c) and finite but zero in average in (b) and (d).](Invers_Ising_Jij_N=10_J0=0_hi=0 "fig:"){width="0.49\columnwidth"} ![\[fig:invising\] Comparison between various approximate solutions to the inverse Ising problem. RMSE errors as a function of the temperature are plotted in (a) and (b) for the couplings $J_{ij}$, in (c) and (d) for the susceptibility matrix $\chi_{ij}$ obtained from the corresponding BP fixed point. Local fields $h_i$ are zero in (a) and (c) and finite but zero in average in (b) and (d).](Invers_Ising_Jij_N=10_J0=0_hi=05 "fig:"){width="0.49\columnwidth"}\
(a)(b)\
![\[fig:invising\] Comparison between various approximate solutions to the inverse Ising problem. RMSE errors as a function of the temperature are plotted in (a) and (b) for the couplings $J_{ij}$, in (c) and (d) for the susceptibility matrix $\chi_{ij}$ obtained from the corresponding BP fixed point. Local fields $h_i$ are zero in (a) and (c) and finite but zero in average in (b) and (d).](Invers_Ising_Xij_N=10_J0=0_hi=0 "fig:"){width="0.49\columnwidth"} ![\[fig:invising\] Comparison between various approximate solutions to the inverse Ising problem. RMSE errors as a function of the temperature are plotted in (a) and (b) for the couplings $J_{ij}$, in (c) and (d) for the susceptibility matrix $\chi_{ij}$ obtained from the corresponding BP fixed point. Local fields $h_i$ are zero in (a) and (c) and finite but zero in average in (b) and (d).](Invers_Ising_Xij_N=10_J0=0_hi=05 "fig:"){width="0.49\columnwidth"}\
(c)(d)\
The quality of the solution can then be assessed directly by comparing the couplings $J_{ij}$ found with the actual ones. Figure \[fig:invising\] are obtained by generating at random $10^3$ Ising models of small size $N = 10$ either with no local fields ($h_i=0,\forall i=1\ldots N$) or with random centered ones $h_i=U[0,1]-1/2$ and with couplings $J_{ij} = \frac{J}{\sqrt{N/3}}(2*U[0,1]-1)$, centered with variance $J^2/N$, $J$ being the common rescaling factor corresponding to the inverse temperature. A glassy transition is expected at $J=1$. The couplings are then determined using (\[eq:JMF\]), (\[eq:JTAP\]), (\[eq:Jij\]) and (\[eq:JBETHE\]) respectively for the mean-field, TAP, BP and Bethe (equivalent to susceptibility propagation) solutions. Figure \[fig:invising\].a shows that the Bethe approximation yields the most precise results in absence of local fields while it is equivalent to TAP when a local field is present as shown on Figure \[fig:invising\].b. Since we want to use these methods in conjunction with BP we have also compared the BP-susceptibilities they deliver. To do that, we simply run BP to get a set of belief and co-beliefs in conjunction with equation (\[eq:invchis\]) which after inversion yields a susceptibility matrix to be compared with the exact ones. The comparison shown on Figure \[fig:invising\].c indicates that Bethe and TAP yield the best results in absence of local field, but are less robust when compared to the more naive BP method when local fields are present as seen on Figure \[fig:invising\].d. This is due to the fact that BP delivers exact beliefs when model (\[eq:Jij\],\[eq:hi\]) is used, which is not necessarily the case for other methods when the local fields are non-vanishing. It is actually not a problem of accuracy but of BP compatibility which is raised by this plot.
#### Sparse inverse models
Let us now test the IPS based approach proposed in Section \[sec:onelink\] to build a model link by link for comparison with more conventional optimization schema based on $L_0$ and $L_1$ penalizations.
We show the results of tests only for the Gaussian case where the link modification can be treated exactly. In the Ising case the method can be used only marginally to propose little correction on the maximum spanning tree or any other sparse model. In general we cannot expect the method to be able to compete with the inverse susceptibility propagation schema i.e. what we call here the Bethe inverse model (\[eq:JBETHE\]). The reason is that the LL gain given by one link is more costly to assess than in the Gaussian case and it is also only approximate. So the stability of the schema is more difficult to control when many links have to be added because the condition of the validity of the Bethe approximation are not controlled without paying an additional computational price. For the Gaussian case instead the situation is much more favorable because the gain can be computed exactly with low computational cost even when the graph is dense. The test we show on Figure \[fig:invGauss\] are done on simulated data produced by the traffic simulator METROPOLIS [@dPMa], our original motivation for this work being related to traffic inference [@FuLaFo; @FuHaLa]. The first set corresponds to a small traffic network called Sioux-Fall consisting of 72 links, from which we extract the $N=60$ most varying ones (the other one being mostly idle). The second set (IAU) is obtained for a large scale albeit simplified network of the Paris agglomeration of size 13626 links, out of which we extracted a selection of the $N=1000$ most varying ones. Each sample data is a $N$-dimensional vector of observed travel times $\{\hat t_i,i=1\ldots N\}$, giving a snapshot of the network at a given time in the day. The total number of samples is $S=3600$ for Sioux-Falls and $S=7152$ for IAU, obtained by generating many days of various traffic scenarios. Then for each link the travel time distribution is far from being Gaussian, having heavy tails in particular. So to deal with normal variables (when taken individually) we make the following standard transformation: $$\label{def:Gaencode}
y_i = F_{Gauss}^{-1}\hat F_i(t_i),\qquad\forall i=1\ldots N$$ which map the travel time $t_i$ to a genuine Gaussian variable $y_i$, where $\hat F_i$ and $F_{Gauss}$ are respectively the empirical cdf of $t_i$ and of a centered normal variable. The input of the different algorithms under study is then the covariance matrix $\operatorname{Cov}(y_i,y_j)$. This mapping will actually be important in the next section when using the devised MRF for inference tasks.
Figure \[fig:invGauss\] displays the comparison between various methods. Performances of the greedy method are comparable to the $L_0$ penalized optimization.
To generate one solution both methods are comparable also in term of computational cost, but the greedy is faster in very sparse regime, and since it is incremental, it generate a full Pareto subset for the cost of one solution. On this figure we see also that the $L_1$ method is simply not adapted to this problem. From the figures, we can see that the estimated inverse matrix derived based on $L_1$ norm penalty needs distinctively more non-zero entries to achieve similar log-likelihood level as the $L_0$ penalty, indicating its failure of discovering the underlying sparse structure, the thresholding of small non-zero entries being harmful w.r.t. positive definiteness. The reason might be that is adapted to situations where a genuine sparse structure exists, which is not the case in the present data.
As expected WS is a very strict constraint and stops to early with very sparse low likelihood models. Relaxing this constraints into $\rho(R') < 1$ yields better GaBP compatibility, but provides no guarantee about it. The justification about the constraints $\rho(R') < 1$ is based on the convergence of the series $\sum_k R^k$ [@MaJoWi].
#### Dual weight propagation for Inverse Ising:
we have performed some numerical checks of the DWP equations presented in Section \[sec:dwp\]. These experiments are done on a sparse bipartite random Ising model to test the accuracy of the partition function estimation on loopy-dual graphs. To be able to compare with the true value of the partition function we have considered a sparse bipartite graph, with a reduced number ($10$) of variables on the top layer so that complete enumeration of these variables state can be done. The links are chosen randomly with the constraint that the graph be connected and that the degree of bottom layer’s variables do not vary by more than one unit. Couplings are independent centered random variables with absolute mean $J$.
For each experimental point, the cycle basis is chosen with a simple greedy heuristic, starting first with the fundamental cycles of a random spanning tree; these cycles are then randomly mixed two by two by linear combinations, such as to reduce the mean size of the basic cycles; the procedure stops when a local minimum is obtained. Note indeed that the dual cyclomatic number is given by $$C({\mathcal{G}}^\star) = \sum_{c=1}^{C({\mathcal{G}})}d^\star(c)-C({\mathcal{G}})-\vert{\mathcal{E}}\vert+k({\mathcal{G}}^\star),$$ with $d^\star(c)$ the degree of a cycle node in the dual graph, i.e. its size in terms of edges, assuming that no two edges are gathered to form a single factor. So reducing the mean size of basic cycles contributes to reduce the cyclomatic number $C({\mathcal{G}}^\star)$ of the dual graph.
Figure \[fig:bipartite\] show results concerning the partition function estimation. Different levels of approximation in loop contributions given in Section \[sec:dwp\] are compared to the one obtained with DWP, when varying either the primal cyclomatic number $C({\mathcal{G}})$ or the mean coupling $J$. The convergence and the results delivered by DWP are very sensitive to the quality of the cycle basis which is used, and this is reflected on Figure \[fig:bipartite\], where the fluctuations that we see are mainly due the difference between cycles basis. Nevertheless, we can see on these plots that when it converges, DWP delivers rather satisfactory values, very often close to the true one, and never beyond a factor of $5$, which represent a few percent of missing information for a system of size $100$. As expected results becomes very accurate in the low temperature regime (high value of $J$) as seen on Figure \[fig:bipartite\].b. even though convergence of DWP seems more delicate in this regime.
![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Z_10_90_J2.pdf "fig:"){width="49.00000%"} ![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Z_10_90_C30.pdf "fig:"){width="49.00000%"}\
(a)(b) ![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Xij_10_90_J2.pdf "fig:"){width="49.00000%"} ![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Xij_10_90_C30.pdf "fig:"){width="49.00000%"}\
(c)(d) ![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Jij_10_90_J2.pdf "fig:"){width="49.00000%"} ![\[fig:bipartite\]Partition function (a) and (b), error on the susceptibility coefficients (c) and (d) and on the couplings (e) and (f) for of a sparse bipartite random Ising model with $10$ variables on the top layer and $90$ on the bottom one. On (a), (c) and (e) the number of fundamental cycles in the primal graph is varied with fixed mean absolute coupling $J=2$. On (b), (d) and (f) the coupling is varied with fixed number of primary cycles $C({\mathcal{G}})=30$. Dual cyclomatic number $C({\mathcal{G}}^\star)$ is also plotted. The experiment is repeated $3$ times for each value. Points are plotted only when DWP converges. Fluctuations correspond to different choice of the primal cycle basis. On (a) and (b) $Z_r$ is the reference exact partition function. The value $Z_{dwp}$ obtained by dual weight propagation (DWP) is compared to $Z_0$, $Z_1$ and $Z_2$ corresponding respectively to no loop, independent cycles and pair of cycles approximations. On (b), (c), (d) and (e) errors on the values obtained by dual weight propagation (DWP) are compared to bare one (BP) and to the error made by the Bethe estimation.](bipartite_Jij_10_90_C30.pdf "fig:"){width="49.00000%"}\
(e)(f)
On the same figure, the use of DWP for the IIP, is illustrated on the same bipartite sparse model. Figure \[fig:bipartite\].c and d shows that the precision on susceptibility coefficients corresponding to edges of ${\mathcal{G}}$ when the couplings $J_{ij}$ are given. The comparison is made with the bare BP approximation ($\chi_{ij} = \tanh(J_{ij})$) and the Bethe one (equivalent to susceptibility propagation) explained at the beginning of this section and based on (\[eq:invchis\]). As seen, the precision can be increased by several orders of magnitude both a low and large temperature with relatively loopy graphs. We can expect this to be reflected in the quality of the solutions of the IIP problem delivered by DWP. As seen on Figure \[fig:bipartite\].e and f, performing a gradient descent on the log likelihood, which in principle is strictly valid in absence of dual loop, yields comparatively rather deceptive improvements at very low temperature (high $J$), when compared again with direct BP and Bethe based methods, except when $C({\mathcal{G}}^\star)$ is small enough. Beyond possible numerical problems, this is probably possibly due to an insufficient convergence rate of DWP during the gradient descent, which has then to be interrupt to far away from the good solution, when $C({\mathcal{G}}^\star)$ is to high. This problem should be addressed in future work.
![\[fig:multimodes\]Comparison of decimation curves in the case of a multi-modal distribution with five cluster for $N=500$ variables (a). Projection of the dataset in the $3$-d dominant PCA space along with the corresponding BP fixed points projections obtained with the Ising model.](synth_500_C5_a01_b01 "fig:"){width="0.49\columnwidth"} ![\[fig:multimodes\]Comparison of decimation curves in the case of a multi-modal distribution with five cluster for $N=500$ variables (a). Projection of the dataset in the $3$-d dominant PCA space along with the corresponding BP fixed points projections obtained with the Ising model.](bp-proj0.png "fig:"){width="0.49\columnwidth"} (a)(b)
#### Inverse models for inference:
we turn now to experiments related to the original motivation of this work, which is to use calibrated model for some inference tasks. The experiments goes as follows: we have an historical data set consisting of a certain number of samples, each one being a $N$-dimensional variable vector, say travel times, which serves to build these models [^2]. Given a sample test data we want to infer the $(1-\rho)N$ hidden variables when a certain fraction $\rho$ of the variables are revealed. In practice we proceed gradually on each test sample by revealing one by one the variables in a random order and plot as a function of $\rho$ the $L_1$ error made by the inference model on the hidden variables. Both for Ising and Gaussian MRF, the inference is not performed in the original variable space, but in an associated one obtained through a mapping (a traffic index) using the empirical cumulative distribution of each variable. For the Gaussian model the inference is performed in the index space defined previously by (\[def:Gaencode\]). For the Ising models we have studied a variety of possible mapping [@MaLaFu] in order to associate a binary variable to a real one such that a belief associated to a binary state can be converted back into a travel time prediction. Without entering into the details (see [@MaLaFu] for details), to define in practice this binary state $\sigma_i$, either we make use of the median value $x_i^{1/2} = F_i^{-1}(1/2)$ in the distribution of $x_i$ for all $i=1\ldots N$: $$\sigma_i = {{\leavevmode\hbox{\rm \small1\kern-0.35em\normalsize1}}_{\{x_i>x_i^{1/2}\}}}\qquad (i).$$ Either we perform a soft mapping using the cdf: $$P(\sigma_i=1) = \hat F_i(x_i)\qquad (ii),$$ the last one having the advantage of being functionally invertible if $\hat F_i^{-1}$ is defined, while the former one being inverted using Bayes rule. The data we are considering are “low temperature” data in the sense that correlations are too strong for an Ising model with one single fixed point. This is reflected in the fact that none of the basic methods given in the Section \[sec:prelim\] is working. To overcome this we use a simple heuristic which consists in to add a parameter $\alpha\in[0,1]$ in the BP model like e.g. in (\[eq:bpalpha\]) or to multiply the $J_{ij}$ by $\alpha$ for the MF, TAP and Bethe models, the local field being consequently modified owing to their dependency on the $J_{ij}$. Concerning the factor-graph we have considered various graph selection procedures. All are based on the mutual information given empirically between variables. A global/local threshold can be used to construct the graph, the parameter being the mean/local connectivity $K$; the MST can be used conveniently as a backbone and additional links are obtained through the thresholding selection procedures. These two parameter $\alpha$ and $K$ are calibrated such as to optimize the performance for each type of model so that fair comparisons can be made afterward.
![\[fig:metrodecim\]Comparison of decimation curves between various MRF for Sioux-Falls data (a) and IAU data (b)](SF-60-MRF-i3 "fig:"){width="0.49\columnwidth"} ![\[fig:metrodecim\]Comparison of decimation curves between various MRF for Sioux-Falls data (a) and IAU data (b)](IAURIF-1000-MST-e0 "fig:"){width="0.49\columnwidth"} (a)(b)
One important difference between the Ising model and the Gaussian one is that multiple fixed points may show up in the Ising case while only a single one, stable or not stable, is present in the Gaussian case. This can be an advantage in favor of the Ising model when the data have well separated clusters. Figure \[fig:multimodes\] illustrates this point. The data are sampled from a distribution containing $5$ modes, each one being a product form over $N$ random bimodal distributions attached to each link. On Figure \[fig:multimodes\].a which displays the error as a function of the fraction of revealed variables we see that the Ising model obtained with (\[eq:bpalpha\]), encoded with the median value (i), gives better prediction than the exact Gaussian model or the approximated GaBP compatible one. Indeed Figure \[fig:multimodes\].b shows a projection of the data in the most relevant $3$-d PCA space along with the projected position of BP fixed points (given by their sets of beliefs) delivered by the Ising model. As we see, the model is able to attach one BP fixed point to each component of the distribution. Ideally we would like a perfect calibration of the Ising model in order that these fixed points be located at the center of each cluster. The method proposed in Section \[sec:perturbation\] could help to do this, but has not been implemented yet. On Figure \[fig:multimodes\].a we see also that the <span style="font-variant:small-caps;">knn</span> predictor performs optimally in this case, since the error curve coincides exactly with the one given by the hidden generative model of the data (ground truth). Figure \[fig:metrodecim\] shows how the different models compare on the data generated by the traffic simulator. On the Sioux-Falls network, the Gaussian model gives the best results, and a sparse version obtained with the greedy algorithm of section \[sec:greedyalgo\] reach the same level of performance and outperforms <span style="font-variant:small-caps;">knn</span>. The best Ising model is obtained with the (\[eq:bpalpha\]) with type (ii) encoding. For IAU the full Gaussian model is also competitive w.r.t <span style="font-variant:small-caps;">knn</span>, but the best sparse GaBP model is not quite able to follow. In fact the correlations are quite high in this data, which explain why the best Ising model shows very poor performance. The best Ising model in that case corresponds to the plain BP model with type (ii) encoding and MST graph.
Conclusion
==========
This paper is based on the observation that in many cases, the Bethe approximation can be a good starting point for building inverse models from data observations. We have developed here three different ways of perturbing such a mean-field solution. One of them (see Section \[sec:onelink\]), based on IPS, is valid both for binary and Gaussian variables, and leads to an efficient algorithm in the Gaussian case to generated sparse approximation models compatible with BP. The additional requirement that the model be compatible with BP for large scale applications discards dense models and simplifies in a way the search space on model selection. This method should be also used in complementary to the two others (natural gradient and DWP) for the Ising case, because the susceptibility evaluation step which was considered too expensive and not precise enough with the Bethe method, can be done efficiently in $O(N^2)$ thanks to DWP in the very sparse regime. Hence, constructing the graph with IIP, while keeping it DWP-compatible could lead to interesting approximate solutions. This point will certainly be considered for future work. Concerning DWP, various possible improvements will be addressed in future studies, concerning respectively the cycle basis choice, the DWP update scheduling itself and the gradient descent, to render this method efficient for a broader class of problems as the simple one considered here. Its extension to problems will local fields should also be addressed experimentally.
#### Acknowledgments
This work was supported by the French National Research Agency (ANR) grant N° ANR-08-SYSC-017.
[^1]: This apparently non-symmetric expression can be symmetrized with help of (\[eq:chirec\]).
[^2]: In fact the pairwise MRF models exploit only pairwise observations but for sake of comparison with a <span style="font-variant:small-caps;">knn</span> predictor we generate complete historical sample data.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present LOFAR high-band data over the frequency range $115$–$189$ MHz for the X-ray binary SS433, obtained in an observing campaign from 2013 February – 2014 May. Our results include a deep, wide-field map, allowing a detailed view of the surrounding supernova remnant W50 at low radio frequencies, as well as a light curve for SS433 determined from shorter monitoring runs. The complex morphology of W50 is in excellent agreement with previously published higher-frequency maps; we find additional evidence for a spectral turnover in the eastern wing, potentially due to foreground free-free absorption. Furthermore, SS433 is tentatively variable at $150$ MHz, with both a debiased modulation index of $11$ per cent and a $\chi^2$ probability of a flat light curve of $8.2 \times 10^{-3}$. By comparing the LOFAR flux densities with contemporaneous observations carried out at $4800$ MHz with the RATAN-600 telescope, we suggest that an observed $\sim$$0.5$–$1$ Jy rise in the $150$-MHz flux density may correspond to sustained flaring activity over a period of approximately six months at $4800$ MHz. However, the increase is too large to be explained with a standard synchrotron bubble model. We also detect a wealth of structure along the nearby Galactic plane, including the most complete detection to date of the radio shell of the candidate supernova remnant G38.7$-$1.4. This further demonstrates the potential of supernova remnant studies with the current generation of low-frequency radio telescopes.'
title: 'LOFAR 150-MHz observations of SS433 and W50'
---
stars: individual: SS433 – ISM: individual objects: W50 – ISM: jets and outflows – ISM: supernova remnants – radio continuum: ISM – radio continuum: stars
\[firstpage\]
Introduction
============
The W50 nebula [@westerhout58] is a large, non-thermal, Galactic radio source, at the centre of which is the X-ray binary system SS433 [@stephenson77], a known source of relativistic jets [@fabian79; @margon79a; @margon79b; @liebert79; @milgrom79; @abell79]. These jets, with mean velocity $0.26$$c$, precess about an axis with a period of $162.4$ d, where the half-opening angle of the precession cone is $21^{\circ}$ and the precession axis is inclined at $78^{\circ}$ to the line of sight (@eikenberry01; also see @margon84). This creates a corkscrew-shaped trace on the sky as the jet components move ballistically outwards [@hjellming81a; @hjellming81b].
At radio frequencies, W50 takes the form of a circular shell of diameter $58$ arcmin with lateral extensions (the ‘wings’ or ‘ears’) of $47$ arcmin and $25$ arcmin to the east and west, respectively [e.g. $327.5$- and $1465$-MHz maps in @dubner98 henceforth D98]. Originally classified as a supernova remnant (SNR) by @holden69, it was suggested with the discovery of the jets in SS433 that, given the remarkable alignment between the jet axis and the major axis of W50, they are responsible for the elongated ($\sim$ $2\degr \times 1\degr$) shape of the nebula [@begelman80]. Hydrodynamical simulations by @goodall11a suggested that W50 is about $20\,000$ years old, and that several distinct episodes of jet activity, with different precession characteristics to the current jets, can explain the morphology of W50; this is consistent with the kinematical investigation by @goodall11b. The asymmetry in the sizes of the ears is usually ascribed to the known density gradient associated with the Galactic plane; SS 433 is located at Galactic coordinates $(l,b) = (39\fdg7, -2\fdg2)$, and the major axis of W50 is oriented at $20$$^{\circ}$ from the normal to the plane (e.g. @downes81a; @goodall11a).
In observations at frequencies $\gtrsim 1$ GHz, flaring activity from SS433 has been found to occur on time-scales $\lesssim 100$–$200$ days [e.g. @seaquist82; @fiedler87; @vermeulen93; @trushkin03; @pal06; @trushkin12; @trushkin14; @trushkin16]. During flaring, flux densities of $\sim$$0.5$–$4$ Jy are measured, on average a factor of a few above the quiescent baseline value. Flaring has also been detected at $408$ and $843$ MHz [e.g. @bonsignori86; @vermeulen93]; at $408$ MHz, the source is typically about $1$ Jy brighter than during quiescence [$3.4$ Jy versus $2.3$ Jy; @bonsignori86].
At low radio frequencies ($< 300$ MHz), however, SS433, and indeed W50 as well, have been relatively poorly characterized. A brief summary of previous observations is as follows. Firstly, using $80$- and $160$-MHz data from the Culgoora Circular Array (CCA), as well as near-simultaneous higher-frequency measurements, @seaquist80 suggested that the quiescent radio spectrum of SS433 turns over at approximately $300$ MHz. From two epochs, spread three weeks apart, the average $160$-MHz flux density was $1.9$ Jy, and the average $5\sigma$ upper limit at $80$ MHz was $1.6$ Jy. Secondly, @pandey07 monitored SS433 at $235$ MHz with the Giant Metrewave Radio Telescope (GMRT) over the period 2002 July – 2005 January (observational cadence range $2$–$270$ days). The average $235$-MHz flux density was $2.71$ Jy and the modulation index $21$ per cent. Using simultaneous $610$-MHz GMRT observations, the two-point spectral index[^1] ${\alpha}^{610}_{235}$ varied between $-1.05$ and $-0.69$, where the uncertainty per individual measurement is much less than this range. This is possibly due to optical depth variations at different stages of flaring [discussion in @pandey07]. Lastly, @millerjones07 detected both SS433 and W50 with the Very Large Array (VLA) at $74$ MHz; the root-mean-square (rms) noise level in their map was $192$ mJy beam$^{-1}$, with an angular resolution of about $100$ arcsec. The flux density of SS433 was $1.9 \pm 0.2$ Jy (J. Miller-Jones, priv. comm.). Using these data, as well as previously published observations, including $83$- and $102$-MHz measurements from the Pushchino Radio Astronomy Observatory [@kovalenko94], @millerjones07 demonstrated that the integrated radio spectrum of W50 can be modelled with a single power law between $74$ and $4750$ MHz, where $\alpha=-0.51 \pm 0.02$ [also see D98 and @gao11].
With the advent of a new generation of novel, low-frequency radio telescopes, including the Low-Frequency Array [LOFAR; @vanhaarlem13], the Murchison Widefield Array [MWA; @tingay13] and the Long Wavelength Array [LWA; @ellingson13], there is now an excellent opportunity to embark on a much more thorough exploration of the low-frequency synchrotron radio emission from jet-producing sources such as SS433. Topics of particular interest include constraining models for intrinsic variability over wide frequency ranges (including frequency-dependent time lags), investigating absorption processes that result in a spectral turnover, as well as the presence of a low-frequency cutoff in the synchrotron spectrum. In this paper, we present the results of a LOFAR high-band observing campaign of SS433 and W50, carried out over the frequency range $115$–$189$ MHz during the period 2013 February – 2014 May. In Section \[observations\], we describe the observations, and how the data were calibrated and imaged. A deep continuum map is presented in Section \[deep map\], while in Section \[spectral index discussion\] we combine our data with previous observations at lower and higher frequencies to investigate the radio spectra of the components of W50. Our low-frequency SS433 monitoring observations are analysed in Section \[SS433 variability\], including a comparison with contemporaneous data at $4800$ MHz from the RATAN-600 telescope. Then, in Section \[SNR candidate\], we discuss the detection of a nearby candidate SNR, G38.7$-$1.4, in our deep image. Finally, we conclude in Section \[conclusions\].
LOFAR observations and data reduction {#observations}
=====================================
A detailed description of LOFAR can be found in @vanhaarlem13. In this section, we provide a summary of our observations, carried out with the high-band antennas (HBA; full operating frequency range $110$–$240$ MHz). An observing log is given in Table \[table:observations\].
[llccrcr]{} & & & & $t_{\rm int}$ & &\
& & & & & &\
Deep & 2013 February 18 & $56341.38$ & $2.08 \pm 0.21$ & $176$ & $24$, $13$, $0$ & L$95041$–L$95072$\
M1 & 2013 March 20 & $56371.28$ & $2.09 \pm 0.21$ & $22$ & $24$, $13$, $0$ & L$106054$–L$106057$\
M2 & 2013 April 25 & $56407.22$ & $1.65 \pm 0.17$ & $22$ & $16$, $13$, $8$ & L$128512$–L$128515$\
M3 & 2013 May 23 & $56435.12$ & $1.69 \pm 0.17$ & $22$ & $22$, $13$, $0$ & L$136070$–L$136073$\
M4 & 2013 July 25/26 & $56499.00$ & $1.49 \pm 0.15$ & $22$ & $19$, $13$, $0$ & L$166060$–L$166063$\
M5 & 2013 August 22 & $56526.87$ & $1.60 \pm 0.16$ & $22$ & $24$, $14$, $0$ & L$169277$–L$169280$\
M6 & 2013 September 20 & $56555.86$ & $1.74 \pm 0.18$ & $22$ & $24$, $14$, $0$ & L$175340$–L$175343$\
M7 & 2013 September 28 & $56563.83$ & $1.64 \pm 0.17$ & $22$ & $24$, $14$, $0$ & L$178910$–L$178913$\
M8 & 2013 November 16 & $56612.63$ & $2.14 \pm 0.21$ & $19$ & $24$, $14$, $0$ & L$188650$–L$188651$\
M9 & 2013 December 11 & $56637.55$ & $2.13 \pm 0.21$ & $26$ & $24$, $14$, $0$ & L$193989$–L$193992$\
M10 & 2014 February 11 & $56699.39$ & $2.62 \pm 0.26$ & $26$ & $24$, $11$, $0$ & L$204224$–L$204227$\
M11 & 2014 March 12 & $56728.30$ & $2.06 \pm 0.21$ & $26$ & $24$, $14$, $0$ & L$207137$–L$207140$\
M12 & 2014 April 24 & $56771.25$ & $1.89 \pm 0.19$ & $26$ & $24$, $14$, $0$ & L$215295$–L$215298$\
M13 & 2014 May 10 & $56787.10$ & $1.89 \pm 0.19$ & $26$ & $24$, $14$, $0$ & L$228277$–L$228280$\
\
\
\
\
\
\
\
\
\[table:observations\]
Deep continuum observation and monitoring runs
----------------------------------------------
Our first HBA observation of SS433 and W50 was a deep $4$-h run carried out with the ‘HBA Dual’ mode, centred on transit (elevation $42\degr$). The frequency range was $115$–$163$ MHz, covered by a total of $244$ sub-bands, each with a bandwidth of $195.3$ kHz. We used the Dutch stations only: $24$ stations in the core, each with two closely-spaced sub-stations, and $13$ remote stations. The projected baselines range from about $29$ m to $79$ km. The primary beam full width at half maximum (FWHM) at $150$ MHz is $3\fdg80$ for the core sub-stations and $2\fdg85$ for the remote stations.
The observation was carried out in blocks of $15$ min: $2$ min on the primary flux density calibrator, 3C380, followed by $11$ min on the target. Although LOFAR is a software telescope, rapid electronic beam switching was not possible when our study took place: the remaining $2 \times 1$ min in the block was needed to switch between the calibrator and target (and vice versa). By the end of the run, $16$ snapshots of SS433 and W50 had been obtained, corresponding to a total integration time of $176$ min.
The primary goal of the shorter runs was to monitor the flux density of SS433. The setup was very similar, except that the frequency coverage was extended to $115$–$189$ MHz, spread over $380$ sub-bands. In general, $2 \times 11$-min or $2 \times 13$-min snapshots, near transit, were obtained per run, along with $2 \times 2$-min scans of 3C380. However, the data from 2013 November 16 were obtained with a slightly different setup: one $19$-min snapshot, approximately centred on transit, preceded by a $1.5$-min observation of 3C380.
Pre-processing, calibration and imaging {#calibration and imaging}
---------------------------------------
For all observations, data were recorded with a time-step of $2$ s and $64$ channels per sub-band. Pre-processing was carried out using standard methods [e.g. @vanhaarlem13]. Firstly, radio-frequency interference (RFI) was removed using [aoflagger]{} [@offringa10; @offringa12], and the two edge channels at both ends of each sub-band were completely flagged, reducing the bandwidth per sub-band to an effective value of $183.1$ kHz. Secondly, because the very bright ‘A-team’ source Cygnus A is only $17\fdg$8 from 3C380, the ‘demixing’ algorithm [@vandertol07] was used to subtract the response of the former from the visibilities of the latter. Lastly, for practical reasons concerning data volume, and the computing time required for calibration and imaging, the data were averaged in time and frequency: to $10$ s per time-step and either $4$ frequency channels per sub-band (deep run) or $1$ frequency channel per sub-band (monitoring runs). For the baseline ranges considered in this paper, the effects of bandwidth and time-average smearing [e.g. @bridle99] in the averaged data are well within the calibration uncertainty (Section \[in band flux scale\]), even at the edges of the field.
Calibration and imaging were also carried out using standard practices. Each 3C380 sub-band was calibrated using a model of the source defined in @scaife12; the absolute flux density scale is that of @roger73. The gain amplitudes were transferred to the target sub-bands in the same observing block. We then performed phase-only calibration on the target field using data from the global sky model developed by @scheers11, within a radius of $10$ deg from the position of SS433. The basis for our model of the field was the $74$-MHz VLA Low-Frequency Sky Survey [VLSS(r); @cohen07; @lane14], with spectral index information being obtained by cross-correlating the relevant VLSS catalogue entries with the $1.4$-GHz NRAO VLA Sky Survey [NVSS; @condon98]. To increase the signal-to-noise and hence the accuracy of the phase-only calibration step, solutions were computed for at most $20$ sub-bands at a time ($\Delta\nu \approx 4$ MHz). Furthermore, we restricted the projected baseline range over which the phase solutions were determined to $250$–$5000$ $\lambda$ (i.e. $500$ m – $10$ km at $150$ MHz), so as to avoid the shortest and longest baselines where our sky model will generally be an inadequate representation of the true sky brightness distribution.
Prior to imaging each run, the sub-bands were combined into six, approximately evenly split bands, and the data were visually inspected; any further errant data points were manually flagged. A primary-beam-corrected image was then generated for each band with [awimager]{} [@tasse13]; we used a robust weighting parameter [@briggs95] of $0$. At the time of analysis, the [awimager]{} did not fully account for variations of flux density with frequency in the deconvolution process [e.g. @sault94]; by making multiple images across the full bandwidth, we kept the fractional bandwidth of the data being imaged to $<10$ per cent in all cases, which helps to minimize the possible reduction in image quality due to effects associated with this issue.
For the deep run, a maximum projected baseline of $4000$ $\lambda$ was used when imaging; for the monitoring runs, the baseline range was $100$–$3000$ $\lambda$. In the latter case, the lower cutoff reduced the contribution from the diffuse emission from W50, while both higher cutoffs gave the most reliable images for the given $uv$ coverage and relatively simple calibration procedure. A multi-scale [clean]{} option had not yet been implemented in the [awimager]{}; it will be possible in the future to more accurately image the significant amount of extended emission in this field.
To make a final map for each run, the six separate images were first convolved to a common angular resolution using the task [convol]{} in [miriad]{} [@sault95]. We then stacked and averaged these resulting maps together with the [miriad]{} tasks [imcomb]{} and [avmaths]{}. Equal weights were used; the effective frequency is about $140$ MHz for the deep run and $150$ MHz for the monitoring runs.
Lastly, after a first pass at imaging, we included an additional round of phase-only self-calibration with a model generated from the LOFAR image. The source finder [pybdsf]{} [@mohan15] was used to generate the model; it was restricted to sources sufficiently compact such that they could be well described with either a single Gaussian or a moderate number ($\leq 6$) of Gaussians. While we subsequently used these self-calibration solutions (determined over the same projected baseline range as before), we found that they did not make a significant difference to the image properties.
In-band flux density accuracy {#in band flux scale}
-----------------------------
At the time of analysis, an issue existed with the standard LOFAR HBA beam model such that the in-band spectral index could be artificially steep. The magnitude of the effect is dependent on the elevation of the target in a given observation, and the distance between the primary calibrator and the target. Unfortunately, this currently precludes an investigation of how the radio properties of SS433 and W50 potentially vary within the LOFAR high band. However, to ensure that the average HBA flux density scale is as accurate as possible, approximate corrections can be made based on a preliminary analysis of full simulations of the electromagnetic properties of the dipoles in the array, including the effects of mutual coupling and grating lobe attenuation (G. Heald, priv. comm.).
For the deep observation, such corrections were applied to each of the six images before they were convolved to a common resolution and averaged together; we divided by factors ranging from $1.6$–$0.8$ between $115$–$163$ MHz. To first order, this process removes an artificially steep in-band spectral index of approximately $-2.5$ in this case. Unfortunately, the simulations are currently incomplete above $163$ MHz, meaning that we are not able to fully assess the effect on the monitoring runs. However, based on an extrapolation of the available data between $115$–$163$ MHz, we estimate that the average correction across a bandwidth of $115$–$189$ MHz should, to first order, fall within the general calibration uncertainty: for all runs, based on the flux stability of sources in the target field, along with 3C380 itself, we estimate that the internal flux density calibration uncertainty is $\sim$$10$ per cent. Hence, we have made no flux density scale adjustments to the images from the monitoring runs.
Absolute flux density calibration {#absolute flux calibration}
---------------------------------
We also cross-checked the band-averaged LOFAR flux densities of the SNRs 3C396 and 3C397 (also known as G039.2$-$00.3 and G041.1$-$00.3, respectively; e.g. @green14), as well as 4C$+$06.66, with flux densities measured at $160$ MHz with the CCA [@slee95]. Using the information presented in @slee95 and @scaife12, both our LOFAR data, and the data from @slee95, are tied to absolute flux density scales that agree to within $\sim$$3$ per cent. We measured the LOFAR flux densities from our deep image, which has the more complete $uv$ and baseline coverage, and, additionally, we interpolated the $160$-MHz CCA flux densities to $140$ MHz using the $80$-MHz CCA flux densities and two-point spectral indices $\alpha^{160}_{80}$ that were also reported in @slee95.
We find that our LOFAR flux densities for 3C396 and 4C$+$06.66 agree with the interpolated $140$-MHz flux densities from @slee95 to within approximately $\pm 2$ per cent, giving us confidence in the absolute flux density calibration. However, the LOFAR measurement for 3C397 is about $25$ per cent lower. One possibility to explain at least some of this discrepancy is that the @slee95 measurement at $160$ MHz has recorded more diffuse emission from the source, and/or there is significant confusion in this case from the Galactic plane, owing to the lower angular resolution, which is roughly twice as coarse as in our study.
To investigate further whether this finding for 3C397 is a potential outlier, we conducted a similar comparison for 3C398 (G043.3$-$00.2), which we also detect, albeit well into a sidelobe of the primary beam; this source is $4\fdg1$ to the north of SS433. Despite the fact that a flux density measured so far from the phase centre may inherently be unreliable, interestingly the agreement with the adjusted measurement from @slee95 is well within the $\pm 2$ per cent margin described above for 3C396 and 4C$+$06.66.
Given the above results for 3C396 and 4C$+$06.66 (and 3C398), we conclude that our data have been correctly tied to the @roger73 absolute flux density scale, and that the main calibration uncertainty is an internal one, as outlined above in Section \[in band flux scale\]. Moreover, we will show throughout this paper that there is a good agreement between our LOFAR flux densities and previously published flux density measurements. Of our key conclusions (Section \[conclusions\]), further evidence for spectral curvature in the eastern wing of W50 (Section \[spectral index discussion\]), as well as the corresponding fit parameters, may be affected if the calibration uncertainty is larger than we have calculated.
Generally speaking, it is important to note that absolute flux density calibration is well known to be challenging at low frequencies with aperture arrays such as LOFAR, particularly near and along the Galactic plane. Due to the excellent sensitivity to diffuse extended emission, future planned releases of Galactic plane catalogues from the LOFAR Multifrequency Snapshot Sky Survey [MSSS; @heald15], and the GaLactic and Extragalactic All-sky Murchison Widefield Array survey [GLEAM; @hurleywalker17], will be highly valuable comparison datasets.
A LOFAR map of SS433 and W50 {#deep map}
============================
{width="18.5cm"}
{width="17cm"}
Our deep continuum map is presented in Figure \[fig:deepmap\], with a zoomed-in view of SS433 and W50 shown in Figure \[fig:deepmap2\]. This is the most detailed image of W50 made thus far at frequencies below $300$ MHz. The angular resolution is $78$ arcsec $\times$ $55$ arcsec (beam position angle BPA $= 20^{\circ}$; throughout this paper we use the convention that the BPA is measured north through east), comparable to that of the $327.5$- and $1465$-MHz VLA images from D98; in particular, the resolution of the latter map from D98 is $56$ arcsec $\times$ $54$ arcsec (BPA $-61\fdg4$). Indeed, at a similar angular resolution, there is a clear consistency between the LOFAR and D98 maps: we also observe the three main components of the nebula (i.e. the central shell, as well as the eastern and western wings) and the complex structure contained within. A similar correspondence can be seen when comparing the LOFAR data with the recent study by @farnes17, who conducted a full-Stokes broadband analysis of SS433 and W50, centred at about $2.3$ GHz, with the Australia Telescope Compact Array (ATCA); the angular resolution of their map is $2.5$ arcmin $\times$ $1.9$ arcmin (BPA $75\fdg9$). Within the wide field of view, we have also detected a wealth of structure near and along the Galactic plane; we return to this topic in Section \[SNR candidate\].
In the vicinity of W50, the rms noise level is $\sim$$4$–$6$ mJy beam$^{-1}$. Source and sidelobe confusion so near the Galactic plane, together with the presence of undersampled diffuse emission, are responsible for the noise level being much higher than the theoretical value of $\sim$$0.08$ mJy beam$^{-1}$. Moreover, there are clear variations off-source (henceforth referred to as the ‘background’), including ‘negative bowls’, due to the undersampling of extended emission.
Flux densities {#W50 fluxes}
--------------
A Gaussian fit (using [imfit]{} in [miriad]{}) to SS433 yields a peak flux density of $2.01 \pm 0.20$ Jy beam$^{-1}$ and an integrated flux density of $2.25 \pm 0.23$ Jy. Note that the $10$ per cent calibration uncertainty is the dominant error term. There is very marginal evidence of source extension (at about the $1.6\sigma$ level along both the major and minor axes of the fitted Gaussian), but the deconvolved position angle, approximately $30^{\circ}$, is not consistent with the jet axis and precession cone. Intriguingly, however, this position angle is consistent with the general direction of the low-surface-brightness ‘ruff’ emission observed at GHz frequencies [e.g. @blundell01], albeit on much smaller angular scales (tens of mas) than in our map.
Assuming that W50 has a central shell with radius $29$ arcmin (D98), we calculated the integrated flux densities of the various components of the nebula by summing the pixels in the respective regions (indicated in Figure \[fig:deepmap2\]) and making corrections for the non-zero background levels. Each background level correction value was determined by measuring the mean flux density in a series of boxes around the component and then taking the average of this set of means; moreover, the standard deviation of the set was used as the uncertainty for the correction. The mean offsets that were subtracted (per pixel) range from $-15.5$ to $6$ mJy beam$^{-1}$. Twisted-plane fits were also investigated, but the results were deemed to be less reliable. Henceforth in this section, as well as Sections \[W50 morphology\] and \[nearby features\], all quoted integrated flux densities and surface-brightness values are background corrected, unless stated otherwise.
The integrated flux densities for the eastern wing, central shell and western wing are $48 \pm 13$, $121 \pm 25$ and $38 \pm 8$ Jy, respectively. The flux density errors were calculated by combining the $10$ per cent calibration uncertainty and the uncertainty in the background level in quadrature; there is also a much smaller statistical uncertainty from the noise level. The $2.25$-Jy contribution from SS433 was subtracted from the central shell, but we do not make any corrections for possibly unassociated compact sources in the three regions; their effects are essentially negligible in comparison to the other flux density uncertainties.
Summing the integrated flux densities for the three components stated in the previous paragraph, and taking into account the relative contributions of systematic and statistical uncertainties when propagating the errors, our total integrated flux density for W50 at $140$ MHz is $207 \pm 33$ Jy. This value is consistent with the previously calculated integrated radio spectrum (see discussion in Section \[introduction\]), giving us further confidence in our calibration and background subtraction procedures. However, depending on the relative contributions of the negative bowls and diffuse Galactic emission in Figures \[fig:deepmap\] and \[fig:deepmap2\], one possibility is that the mean background corrections should be more negative in value (i.e. the flux densities of the three components of W50 should be higher), and that the respective uncertainties are overestimated. As will become apparent in Section \[integrated flux densities spectra\], this would move the flux densities of the eastern wing and central shell upwards towards the extrapolated single power-law radio spectra determined by D98 at higher frequencies. In the former case, this would reduce the evidence for spectral curvature at low frequencies. However, it seems very unlikely that an incorrect background correction alone could explain the deviation from a single power-law radio spectrum for the eastern wing, as there is insufficient evidence in Figures \[fig:deepmap\] and \[fig:deepmap2\] for the significant negative bowl (mean $\approx -46$ mJy beam$^{-1}$) that would be needed in the vicinity of this component. Furthermore, if the background correction for the western wing were to be a larger negative value, then the corrected $140$-MHz flux density would quickly become overestimated.
Various second-order effects may result from the use of the @clark80 [clean]{} algorithm in the [awimager]{} on a target such as W50, including residual deconvolution sidelobes and artefacts, as well as possible [clean]{} bias [e.g. @condon98]. We plan to re-image our data with multi-scale [clean]{} and maximum entropy algorithms once they are implemented in the [awimager]{}; an alternative would be to use sparse image reconstruction methods [e.g. @garsden15; @dabbech15; @girard15; @pratley18]. Other techniques, such as the one outlined in @braun85, could also be considered as part of future work.
Low-frequency morphology of W50 {#W50 morphology}
-------------------------------
![The spur, or ‘chimney’, at $140$ MHz. To outline its shape, we have overlaid a radio contour at the $30$ mJy beam$^{-1}$ level (not corrected for the non-zero background).[]{data-label="fig:chimney"}](fig3.pdf){width="8.5cm"}
The arcs and filaments observed at higher frequencies are also prominent in our map. These include the thin filament at a position angle of $80^{\circ}$ (labelled in Figure \[fig:deepmap2\]), discussed in D98 as being positionally associated with the precession cone of SS433, with a $140$-MHz surface brightness ranging from about $50$–$90$ mJy beam$^{-1}$. In addition, we see the set of north-south filaments which are suggestive of large-scale traces of the precessing jets; the brighter of these are also labelled in Figure \[fig:deepmap2\], with hints of similar fainter features in between. The peak flux densities of the eastern and western wings are approximately $150$ and $170$ mJy beam$^{-1}$, respectively; they occur in the brightest north-south filament and near the edge-brightened western boundary, respectively.
We detect the ‘chimney’ (Figure \[fig:chimney\]), previously observed at $1465$ MHz by D98 and also seen by @farnes17. This is a spur-like feature to the north of the bright filament in the eastern wing, near the position RA $19^{\rm h}$$16^{\rm m}$$18^{\rm s}$, Dec. $+04^{\circ}$$59\arcmin\,30\arcsec$ (J2000). D98 suggested that it may be analogous in morphology to a spur, approximately an order of magnitude smaller, at the northern edge of the Crab Nebula [e.g. @fesen86; @rudie08; @black15]. The approximate dimensions at $140$ MHz are very similar to those at $1465$ MHz: $\sim$$6$ arcmin in length and $\sim$$7$ arcmin in width at the base. This corresponds to a projected linear size of about $10$ pc $\times$ $11$ pc.[^2] The average surface brightness at $140$ MHz is $\sim$$15$–$20$ mJy beam$^{-1}$. Using the $1465$-MHz image from D98, convolved to the resolution of our data (see Section \[spectral index discussion map\] for further details, where we construct a two-point spectral index $\alpha^{1465}_{140}$ map of W50), we estimate a two-point spectral index $\alpha^{1465}_{140} \sim -0.5$, with an error of about $30$–$40$ per cent due to the faintness of the emission and the uncertainties in estimating the background level at both frequencies.
@farnes17 identified a potential second chimney, $20$ arcmin to the west of the chimney described above. However, it is not visible in Figure \[fig:chimney\]. Our non-detection would be consistent with the suggestion of @farnes17 that this feature is likely to be a low-level artefact from the deconvolution process.
Related to the previous paragraph, there are, however, some generally low-level artefacts in our LOFAR map, likely due to residual calibration and/or deconvolution errors. For example, $3.7$ arcmin to the north-north-west of SS433, at a position angle of $-20^{\circ}$, there is a spurious feature consisting of two components, with a peak flux density of approximately $80$ mJy beam$^{-1}$. A similar structure can be seen below the central shell of W 50 near the compact radio source NVSS J191252+043104.
The H[ii]{} region S74 {#nearby features}
----------------------
S74 [@crampton78], a foreground H[ii]{} region [distance $\sim$$2$–$3$ kpc; e.g. @forbes89; @brand93; @lockman07; @alves10] to the north-west of W50, is also detected at $140$ MHz. The peak flux density is $\sim$ $50 \pm 20$ mJy beam$^{-1}$, with the quoted $1\sigma$ error dominated by the uncertainty in the background level. In the convolved $1465$-MHz image from D98, as mentioned above in Section \[W50 morphology\], we estimate a peak flux density of $75 \pm 10$ mJy beam$^{-1}$; thus, a representative two-point spectral index $\alpha^{1465}_{140}$ at the position of peak intensity is $0.2 \pm 0.2$. Although H[ii]{} regions can become optically thick at low frequencies [e.g. @kassim89; @copetti91; @subrahmanyan95; @omar02], our estimated spectral index for S74 is consistent with the canonical spectral index of $-0.1$ in the optically-thin regime: the difference is only $1.5\sigma$. The radio spectrum may in fact be less inverted than our estimate depending on the relative contribution to the brightness temperature at $140$ MHz from the Galactic emission behind S74 (e.g. formulae in @nord06; also see sky maps in e.g. @haslam82, @guzman11, @remazeilles15 and @zheng17).
In a $30.9$-MHz Galactic plane survey, conducted by @kassim88 with the Clark Lake Teepee-Tee Telescope, S74 appears as a negative absorption hole against the Galactic background (we refer the reader to @nord06 and @su17a [@su17b] for other similar cases). There is evidence for such behaviour at $74$ MHz as well (J. Miller-Jones, priv. comm.). Although further measurements at low frequencies are needed to establish exactly where the spectral turnover occurs, we can make some predictions based on an estimated turnover frequency. Firstly, the linear emission measure $EM$ can be calculated approximately using the expression $$\label{equation emission measure}
\tau_{\nu} \approx 3.28 \times 10^{-7} \left(\frac{T_{\rm e}}{10^4 \, \rm{K}}\right)^{-1.35} \left(\frac{\nu}{\rm{GHz}}\right)^{-2.1} \left(\frac{EM}{\rm{pc\,cm}^{-6}}\right),$$ where $\tau_{\nu}$ is the optical depth at frequency $\nu$, and $T_{\rm e}$ is the electron temperature of the ionized gas in the H[ii]{} region [e.g. @mezger67]. Taking a turnover frequency of $\sim$$100$ MHz, where the optical depth is unity, and assuming $T_{\rm e} \sim 6000-7000$ K for S74 (e.g. see derived relationships between electron temperature and Galactocentric distance for H[ii]{} regions in @shaver83, @paladini04, and @alves12), then $EM \sim 1.2$–$1.5 \times 10^4$ pc cm$^{-6}$.
Secondly, due to the diffuse Galactic emission in the vicinity of the source, it is somewhat difficult to assess the full extent of S74 in our LOFAR map, but a rough estimate is $\sim$ $7$ arcmin $\times$ $13$ arcmin ($\sim$ $5.1$ pc $\times$ $9.5$ pc at a distance of $2.5$ kpc; also see e.g. @altenhoff70, @geldzahler80, and @kuchar97 for other size estimates). Then, using the simple approximation that $EM \approx n_e^2 L$, where $n_{e}$ is the electron number density in S74, and $L$ is the path length through the source along the line of sight, $n_{\rm e}$ $\sim$ $40$–$45$ cm$^{-3}$ assuming that the path length is roughly the average of the projected dimensions on the sky (i.e. $L\sim7.3$ pc). Referring to the sub-classes in e.g. @kurtz05, our very crude analysis suggests that S74 is most consistent with ‘classical’ H[ii]{} regions.
Radio spectra {#spectral index discussion}
=============
--------------- -- -- -- -- --------------------- ------------------------ ------------------------- ----------------------- --
Component $\alpha^{140}_{74}$ $\alpha^{327.5}_{140}$ $\alpha^{1465}_{327.5}$ $\alpha^{1465}_{140}$
Eastern wing
Central shell
Western wing
--------------- -- -- -- -- --------------------- ------------------------ ------------------------- ----------------------- --
\[table:fluxes\_components\]
{width="18cm"}
As mentioned in Section \[introduction\], @millerjones07 found that the integrated radio spectrum of W50 is well described by a single power law between $74$ and $4750$ MHz ($\alpha=-0.51 \pm 0.02$). Using their data at $74$ MHz, as well as the previously published data from D98 at $327.5$ and $1465$ MHz, they also demonstrated that the only component of W50 that shows evidence for significant spectral turnover at low frequencies is the eastern wing.
Our LOFAR $140$-MHz map not only bridges the frequency gap in the @millerjones07 analysis, but has sufficient dynamic range and angular resolution to permit us to study changes in the spectral index across the W50 nebula, either using the integrated flux densities for the specific components, or pixel by pixel. We now discuss our findings using a combination of both approaches.
Integrated flux densities {#integrated flux densities spectra}
-------------------------
Table \[table:fluxes\_components\] lists the integrated flux densities for the eastern wing, central shell and western wing at $74$, $140$, $327.5$ and $1465$ MHz. While our LOFAR data are on the @roger73 flux density scale, the $74$-MHz observations used in the @millerjones07 study were calibrated with the @baars77 scale. Hence, we took the $74$-MHz flux densities for the three components (J. Miller-Jones, priv. comm.), scaling each measurement up by $10$ per cent and propagating the corresponding uncertainty term of $6$ per cent with the previous uncertainty; these values were found to be appropriate in an analogous rescaling of the $74$-MHz VLSSr [@lane14].
In Table \[table:fluxes\_components\], we also give two-point spectral indices calculated from the integrated flux densities, although we note that the $\alpha^{1465}_{327.5}$ values are from D98 and were calculated with a linear regression method using the pixel values in specified regions of the source (further details in Section \[spectral index discussion map\]). In addition, broadband radio spectra are presented in Figure \[fig:spectra\].
As can be seen in Figure \[fig:spectra\], our LOFAR flux densities are consistent with the previously published measurements, in particular strengthening the evidence for a spectral turnover in the eastern wing. We have fitted each spectrum with either a free-free absorption model (eastern wing), or a linear fit in $\log(S_{\nu})$–$\log(\nu)$ space (central shell and western wing). The best-fitting parameters are reported in Table \[table:fluxes\_components2\], and the fits are overlaid in the panels in Figure \[fig:spectra\]. The reduced $\chi^2$ value for the western wing ($0.035$) suggests that our uncertainties are possibly overestimated in this case.
In the case of the eastern wing, our model is of the form $$\label{eqn free free}
S_{\nu} = S_{140,{\rm u}}\left(\frac{\nu}{140\,\rm{MHz}}\right)^{\alpha}\exp\left[-(\tau_{140})\left(\frac{\nu}{140\,\rm{MHz}}\right)^{-2.1}\right],$$ where $\tau_{140}$ is as before, $S_{140,{\rm u}}$ is the unabsorbed flux density at $140$ MHz for optically-thin emission, and $\alpha$ is the optically-thin spectral index [e.g. @weiler86]. We assume that $\alpha=-0.82$, using the $\alpha^{1465}_{327.5}$ value from D98. The turnover frequency (at about $120$ MHz, where the fitted integrated flux density is approximately $80$ Jy) is near the bottom of the LOFAR high band, although more low-frequency measurements are needed to confirm this. Further discussion can be found in Section \[alpha free-free\].
W50 spectral index map {#spectral index discussion map}
----------------------
[lccc]{}\
\
\
& $S_{140,{\rm u}}$ (Jy) & $\tau_{140}$ & Reduced $\chi^2$\
Eastern wing & $99 \pm 8$ & $0.27 \pm 0.03$ & $2.57$\
\
\
\
\
\
& $a$ & $b$ $(\equiv \alpha)$ & Reduced $\chi^2$\
Central shell & $3.46 \pm 0.09$ & $-0.58 \pm 0.04$ & $1.23$\
Western wing & $2.62 \pm 0.10$ & $-0.48 \pm 0.04$ & $0.035$\
\[table:fluxes\_components2\]
{width="17cm"}
We generated a two-point spectral index map for W50 using the $140$-MHz image shown in Figures \[fig:deepmap\] and \[fig:deepmap2\], as well as the $1465$-MHz image from D98; see Figure \[fig:alpha\_map\]. Firstly, to account for the fact that the angular resolutions are slightly different, we convolved the $1465$-MHz image (resolution $56$ arcsec $\times$ $54$ arcsec, BPA $-61.4^{\circ}$; figures 1a and 1b in D98), to the resolution of our LOFAR map ($78$ arcsec $\times$ $55$ arcsec, BPA $20^{\circ}$). Secondly, for consistency with the $1465$-MHz map, we regridded the LOFAR image from $10$ arcsec pixel$^{-1}$ to $15$ arcsec pixel$^{-1}$. Lastly, we calculated the spectral index values: $\log(S_{140}/S_{1465})/\log(140/1465)$ per pixel, with average background corrections of $-4.5$ and $3$ mJy beam$^{-1}$ at $140$ and $1465$ MHz, respectively, applied first.
Generally, there is a very good agreement between the astrometry of the two maps, although there are some examples of non-systematic offsets of order $1$–$2$ pixels for moderately bright, compact sources in the field of view. Investigating this further, it appears as though our astrometry is generally more consistent with, for example, the $1.4$-GHz NVSS. Given the limitations in our analysis, which we describe below, we have not attempted to make differential alignment corrections to account for these relatively small offsets.
One difficulty with the method described above for calculating the spectral index map is that varying, non-zero background levels will cause biases; in particular, this is an issue for our LOFAR image. Hence, the results are best interpreted in a qualitative sense. As in the $\alpha^{1465}_{327.5}$ spectral index map (figure 4 in D98), we too observe regions of flatter spectral indices towards the south-west of the source, either side of the steeper, bright filament. However, in general, the spectral index gradients are not as pronounced in our map; moreover, while the apparent steepening from west to east across the source is consistent with the D98 spectral index map, this is, at least in part, likely to result from biases related to the background levels (see below). The brightest filamentary structure in the eastern wing may slightly flatten immediately to the east and west of the central region, although higher-resolution maps are needed to confirm this. The very flat spectral indices along some of the boundary regions are artificial: the negative residual backgrounds at $140$ MHz result in underestimated $140$-MHz flux densities.
There is a hole-like feature approximately centred at RA $19^{\rm h}$$10^{\rm m}$$58\fs1$, Dec. $+04^{\circ}$$39\arcmin$$57\farcs8$ (J2000), in the south-western region of the central shell below the bright filament, that has an inverted spectrum (maximum value $\sim$0.2). In the LOFAR band, it is $\sim$$7$–$8$ times fainter than the diffuse emission surrounding it; the feature is labelled in Figure \[fig:deepmap2\]. In D98, figures 1b and 2 suggest possibly similar behaviour at $327.5$ and $1465$ MHz, but not as pronounced in the latter case (also see figure 1 in @farnes17). This may be a significantly absorbed region, although a corresponding foreground H[ii]{} region at or near these coordinates could not be found in the SIMBAD [@wenger00] and VizieR [@ochsenbein00] online databases.[^3] Moreover, @farnes17 presented a continuum-corrected H$\alpha$ mosaic of the W50 region, using data from the Isaac Newton Telescope Photometric H$\alpha$ Survey of the Northern Galactic Plane [IPHAS; @drew05]; however, an excess of emission at the position of the feature is not observed (figure 6 in @farnes17).
As mentioned above, generating a spectral index map using the standard formula, pixel by pixel, can potentially give highly misleading quantitative results. Therefore, following @anderson93, as well as D98 (also see e.g. @leahy91 and @leahy98), we investigated a linear regression method. Various box sizes ($5$–$35$ pixels, i.e. $75$ arcsec – $8.75$ arcmin, per side) were slid over both maps in increments of one pixel in both directions. We then used the average slope from the regressions of $S_{1465}$ versus $S_{140}$ and $S_{140}$ versus $S_{1465}$ (taking the inverse of the calculated slope in the latter case) to determine the spectral index for each area, which was assigned to the central pixel. Regions off-source were masked out. Rather than considering the ratio of the flux densities at $140$ and $1465$ MHz, this approach, also often referred to as the $T$–$T$ plot method (where $T$ is the brightness temperature), removes any bias which is constant over the box size, but averages over the spectra of small-scale variations within the box (e.g. structure within filaments). Larger box sizes will have a range of faint and bright emission to provide more accurate constraints when determining each slope.
After using the technique described in the previous paragraph, the apparent $\alpha^{1465}_{140}$ spectral index gradient visible in Figure \[fig:alpha\_map\] is significantly reduced. Each of the three components of W50 has about the same mean spectral index, $\sim$$-0.4$, a value consistent with the two-point spectral indices $\alpha^{1465}_{140}$ derived from the integrated flux densities in Table \[table:fluxes\_components\]. This mean is not affected significantly by the choice of the dimensions of the box, apart from the typical $1\sigma$ error, which decreases from $\sim$$0.3$ to $\sim$$0.1$ with increasing box size. The effects of correlated noise, which can, for example, be reduced by only considering pixels at regularly spaced intervals [e.g. @green90], are second order here. Unlike between $327.5$ and $1465$ MHz, an approximately constant average spectral index $\alpha^{1465}_{140}$ across W50 simply follows from the fact that there is spectral turnover in the eastern wing at low frequencies (Table \[table:fluxes\_components\] and Figure \[fig:spectra\]), and so $\alpha^{1465}_{140}$ is significantly flatter than $\alpha^{1465}_{327.5}$ for the eastern wing.
Spectral index asymmetry between the wings {#spectral index asymmetry}
------------------------------------------
D98 suggested that the density gradient due to the nearby Galactic plane can explain the difference in the radio spectra of the two wings between $327.5$ and $1465$ MHz: the western wing has a flatter spectrum because the jet is expanding into a denser region, leading to stronger shocks and higher levels of compression. However, our results in Section \[spectral index discussion map\] are less consistent with this interpretation, albeit with the possible complication of absorption effects at low frequency. On the other hand, in Figure \[fig:alpha\_map\], the steepening of the spectral index from west to east across W50 is a relatively smooth. Some of this is highly likely to be due to the corresponding gradient in the background level at $140$ MHz, but we cannot determine with certainty whether there is a residual gradient after accounting for the varying background levels: the uncertainties in the spectral index maps generated with the linear regression method are too large.
We now explore an alternative explanation for any spectral index asymmetry between the wings. Expanding into the higher-density environment closer to the Galactic plane, the western wing has been less affected by adiabatic expansion, and so its magnetic field will consequently be stronger than that in the more expanded eastern wing of the nebula. Electrons emitting synchrotron radiation at a specific frequency $\nu$ in an ambient magnetic field of strength $B$ have a Lorentz factor $\gamma$ given by $$\label{lorentz factor}
\gamma = \left( \frac{2 \pi m_{\rm e} \nu}{e B} \right)^{1/2},$$ where $e$ is the elementary charge and $m_{\rm e}$ is the electron rest mass. Therefore, in the western wing, where the magnetic field is stronger, we would probe lower Lorentz factors at a fixed radio frequency. Moreover, for a curved distribution of the underlying electron Lorentz factors, we should also measure a flatter spectral index [e.g. @scheuer68].
If this hypothesis is correct, then we may expect the equipartition magnetic field $B_{\rm eq}$ to be stronger in the western wing. However, X-ray observations of W50 suggest similar ranges of $B_{\rm eq}$ in the eastern and western wings: @safi97 calculated values of $7$–$60$ $\upmu$G (eastern wing) and $7$–$49$ $\upmu$G (western wing; also see @moldowan05). The equipartition magnetic fields can also be estimated from radio data, although the calculation can involve significant uncertainties; revisions to the classical approach [e.g. @longair94] have been suggested [e.g. @beck05; @arbutina12; @arbutina13; @duffy12]. In particular, @arbutina12 [@arbutina13] developed modified equipartition calculations for the particular case of SNRs, although it is possible that further modifications would be needed for W50, given both its distinct morphology and the relatively flat spectral index of the western wing (their model is valid for $-1 \leq \alpha \leq -0.5)$.
While it is beyond the scope of this paper to carry out a detailed comparison between the different methods, we now briefly consider the formalism developed by @duffy12. Despite the fact that the focus of their study is the lobes of radio galaxies and quasars, the framework described is based upon fundamental synchrotron physics. The key quantities in equation (26) of @duffy12 are as follows: (i) a spectral curvature term (that is intrinsic curvature and not from absorption effects); (ii) the frequency of the peak of the radio spectrum; and (iii) the emissivity at this frequency. All terms vary slowly, however, due to power-law dependencies of $1/7$ or $2/7$. We assume that the majority of the radio emission in the eastern wing can be modelled as a cylinder, height $47$ arcmin and diameter $22$ arcmin; similarly, we model the western wing as a cone with height $25$ arcmin and base diameter $27$ arcmin. Some straightforward calculations can be made to show that, very tentatively, and in a qualitative sense, $B_{\rm eq}$ may be at least slightly stronger in the western wing if the associated spectral curvature is milder than for the eastern wing. However, the comparison will be affected by spectral curvature that is caused by foreground free-free absorption (Section \[alpha free-free\]).
In addition, despite the limitations mentioned above, given that the radio spectrum of the western wing could have a spectral index slightly steeper than $-0.5$ within the $1\sigma$ uncertainties (Table \[table:fluxes\_components2\]), then the online calculator[^4] for the @arbutina12 [@arbutina13] model also possibly suggests a slightly stronger value of $B_{\rm eq}$ for the western wing, if the non-spherical geometries of both wings are compensated for, to first order. Potentially significant differences in the depths and filling factors of the two wings are additional sources of uncertainty, both for this model and the @duffy12 formalism. More complete radio spectra are also needed.
Low-frequency spectral curvature {#alpha free-free}
--------------------------------
Below 327.5 MHz, it is the eastern wing and not the western wing that has a flatter spectrum; this is contrary to our suggestion in Section \[spectral index asymmetry\]. One possibility is that the underlying electron energy spectra in the two wings are different, with perhaps shock acceleration in the denser ambient environment of the western wing accounting for the lack of spectral curvature, at least over the frequency range considered in this paper. Alternatively, the radio spectrum of the eastern wing may correspond to enhanced low-frequency free-free absorption along the line of sight to this component of the nebula. This hypothesis is supported by the fact that @farnes17 found a global anti-correlation for W50 between the linearly polarized flux density at $2.3$ GHz and the strength of the diffuse H$\alpha$ emission, suggesting that this is due to a foreground Faraday screen of warm, ionized plasma (also see the polarization measurements in @downes81b, and @downes86). Inspecting the H$\alpha$ mosaic from this study (their figure 6), bright emission is coincident with the position of the eastern wing.
In Section \[integrated flux densities spectra\], we fitted the radio spectrum of the eastern wing with a free-free absorption model. Taking the fitted value $\tau_{140} = 0.27 \pm 0.03$ (Table \[table:fluxes\_components2\]), and using Equation \[equation emission measure\] with the assumption that $T_{\rm e} \sim 10^4$K, then the corresponding linear emission measure is $EM \sim 1.3 \times 10^4$ pc cm$^{-6}$. However, as mentioned before, more low-frequency data points are needed (e.g. within the LOFAR high band), particularly in light of the relatively poor reduced $\chi^2$ value of $2.57$.
The potential effects of a synchrotron ageing break also need to be considered. Radiative losses from the emitting electrons result in a break in the synchrotron spectrum, and, for continuous particle injection, the spectrum is steepened from $\nu^{\alpha}$ to $\nu^{\alpha-0.5}$. The break frequency $\nu_{\rm B}$ is given by $$\nu_{\rm B} = \frac{1610^2\,{\rm GHz}}{(B/\upmu{\rm G})^3(t_{\rm s}/{\rm Myr})^2},$$ where $t_{\rm s}$ is the ‘synchrotron age’ of the distribution, i.e. the time since it was a power law out to infinite frequency [e.g. @carilli91]. If the age of W50 is approximately $20\,000$ yr [@goodall11a], then a break frequency of e.g. $\sim$$100$–$300$ MHz would require a magnetic field of strength $\sim$$2800$–$4000$ $\upmu$G. These values are $1.7$–$2.8$ dex larger than the equipartition magnetic fields that were discussed above in Section \[spectral index asymmetry\]. Therefore, assuming that the lobes do not deviate significantly from equipartition, the effects of synchrotron ageing can be neglected.
In the $30.9$-MHz Galactic plane survey by @kassim88, two thirds of a sample of $32$ SNRs exhibit low-frequency spectral turnovers [@kassim89a; @kassim89b]. W50 was detected, although it is not part of the study presented in @kassim89a [@kassim89b]; the eastern half of the source falls outside of the survey area, due to the Galactic latitude cutoff imposed. The brightest part of the western wing and the northern part of the central shell are clearly detected, with corresponding flux densities of $35.2$ and $23.7$ Jy, respectively (flux density uncertainty approximately $20$ per cent; angular resolution $13.0$ arcmin $\times$ $11.1$ arcmin). However, the more southerly regions of the nebula are at best marginally detected. Given that the $30.9$-MHz flux densities are on the @baars77 scale, we adjusted the $74$-MHz flux densities in Table \[table:fluxes\_components\] back to this scale (dividing by a factor of $1.1$; see Section \[integrated flux densities spectra\]), as well as making an additional correction for the $74$-MHz flux density of the central shell: about $25$ per cent of the flux density orginates from the same spatial region as at $30.9$ MHz (J. Miller-Jones, priv. comm.). We therefore obtain approximate spectral indices $\alpha^{74}_{30.9}$ for the western wing and central shell of $\sim$$0.4$ and $\sim$$1.0$, respectively. These spectral indices, coupled with the distinct change in morphology, provide further evidence for a low-frequency spectral turnover in W50, possibly due to free-free absorption along the line of sight. LOFAR low-band measurements, which cover the frequency range $30$–$80$ MHz, would allow a far more detailed investigation to be made.
Finally, we briefly consider whether a spectral turnover at low frequencies could also be related to a low-frequency cutoff in the electron energy spectrum. In this case we would expect a $S_{\nu} \propto \nu^{0.3}$ dependence [e.g. @scheuer68]. Using Equation \[lorentz factor\] and assuming the equipartition magnetic field strengths calculated by @safi97, $\gamma \sim 420$–$1240$ at $30$ MHz. It is not entirely clear whether a low-frequency cutoff would therefore be discernible with LOFAR, particularly given the upper end of this range [e.g. discussion in @harris05]. Further measurements between $30$ and $80$ MHz would also potentially allow contributions from free-free absorption, a low-frequency cutoff and intrinsic spectral curvature to be disentangled.
Low-frequency variability of SS433 {#SS433 variability}
==================================
150-MHz light curve {#SS433 variability_low}
-------------------
{width="12cm"}
A light curve of SS433 from all $14$ observations (Table \[table:observations\]) is shown in Figure \[fig:lightcurve\]. For the deep run, we have not used the SS433 flux density reported in Section \[deep map\]. Instead, we made another averaged map from the same baseline range as that used when imaging the monitoring runs, and over a similar, more restricted $uv$ coverage. We then divided by a correction factor of $1.2$ to shift the effective average frequency from $140$ to $150$ MHz, bringing the flux density scale into line with the monitoring observations (see Section \[in band flux scale\] for further details).
For the monitoring runs and the regenerated map from the first run described above, the angular resolution is about $140$ arcsec $\times$ $100$ arcsec (median BPA $\approx$ $60^{\circ}$). Each Gaussian fit to the emission from SS433 was point-like. Note that the new measured flux density from the deep run (as given in Table \[table:observations\]) agrees with the value given in Section \[deep map\] within the $1\sigma$ uncertainties.
In Figure \[fig:lightcurve\], SS433 shows tentative evidence for variability in the LOFAR band. To quantify its magnitude and significance, we used two statistics. Firstly, the modulation index $m$ is given by $$m = \frac{\sigma_{S}}{\overline{S}},$$ where $\sigma_{S}$ is the standard deviation of the combined set of flux densities from Table \[table:observations\], and $\overline{S}$ is the mean flux density. We find that $m$ is $16$ per cent, which is comparable to the modulation index at $235$ MHz [$21$ per cent; @pandey07]. If we take into account the errors on each data point and debias the modulation index (as, for example, is described in @bell14), then $m$ is reduced to $11$ per cent.
Secondly, we calculated the $\chi^2$ probability that the flux density remained constant over the period covered by our observations [e.g. @gaensler00; @bell14]. The quantity $$\label{eqn_light_curve_constant_check}
x^2 = \sum\limits_{i=1}^n \frac{(S_{i} - \tilde{S})^2}{\sigma^2_i},$$ where $S_{i}$ is the $i$th flux density measurement, $\tilde{S}$ is the weighted mean flux density, and $\sigma_i$ is the uncertainty for the $i$th flux density measurement, follows a $\chi^2$ distribution with $n-1$ degrees of freedom, assuming that the uncertainties are drawn from a normal distribution. The weighted mean is calculated using the formula $$\label{eqn_light_curve_constant_check2}
\tilde{S} = \sum\limits_{i=1}^n \left(\frac{S_{i}}{\sigma^2_i}\right) \bigg/ \sum\limits_{i=1}^n \left(\frac{1}{\sigma^2_i}\right).$$ The assumed 10 per cent calibration uncertainty is the dominant error term. Background corrections off-source, including from W50, are second-order effects because of the $uv$ range chosen, and hence we do not consider the small contributions from the associated uncertainties.
Using Equations \[eqn\_light\_curve\_constant\_check\] and \[eqn\_light\_curve\_constant\_check2\] for the $n=14$ observations in Table \[table:observations\], we find that the probability $P$ of $x^2$ exceeding our calculated value by chance (i.e. $P(x^2 > 28.3)$ for $13$ degrees of freedom) is $8.2 \times 10^{-3}$. This probability value may be too large to be regarded as a significant indicator of variability [e.g. the criteria discussed in @gaensler00], but we suggest that it is sufficiently small that it could be interpreted as being at least marginally significant. Furthermore, in Figure \[fig:lightcurve\], there are hints of a general consistency in the calibration for several of the more closely spaced observations, suggesting that the calculated $\chi^2$ probability does not simply result from a light curve with large scatter, but rather genuine variability. Reducing the calibration uncertainty further would be needed to better quantitatively assess the magnitude of the intrinsic variability of the source.
Comparison with 4800-MHz data {#SS433 variability_high}
-----------------------------
In Figure \[fig:lightcurve\], we have also plotted $4800$-MHz flux densities from contemporaneous monitoring observations (angular resolution $5$ arcmin $\times$ $50$ arcsec) carried out with the RATAN-600 telescope.[^5] The same broad trends are observed at both $150$ and $4800$ MHz: an initial decrease in flux density to a relatively constant baseline level (approximately $1.6$ and $0.44$ Jy at $150$ and $4800$ MHz, respectively), followed by increased activity at later epochs, before a second decline in flux density at the end of the sampling period.
### Individual flares {#individual flares}
At $4800$ MHz, there were bright flares on MJD $56442.99$ and $56724.22$, which peaked at $1.40$ and $1.28$ Jy, respectively. However, neither of them have obvious counterparts at $150$ MHz. There was also activity before and shortly after our LOFAR observing campaign began, with a maximum $4800$-MHz flux density of $1.40$ Jy on MJD $56264.48$.
We now discuss whether we would have expected to see the outbursts in the LOFAR band. Firstly, if the plasmons associated with a flare are well described by, for example, the @vanderlaan66 synchrotron bubble model (also see e.g. @hjellming88, and @ball93), then the rise time of an outburst is longer at lower frequencies (owing to opacity effects), and the peak flux density is often lower (relative to the quiescent baseline level). The ratio of the peak flux densities at $150$ and $4800$ MHz is given by $$\label{equation vdL}
\frac{S_{{\rm max},150}}{S_{{\rm max},4800}} = \left(\frac{4800}{150}\right)^{-(7p+3)/(4p+6)},$$ where $p = 1-2\alpha$ is the negative of the power-law index of the electron energy ($E$) distribution (i.e. $N(E)\mathrm{d}E \propto E^{-p}\mathrm{d}E$). We use a fiducial value of $1$ Jy for the $4800$-MHz flares (after subtracting the quiescent component), and, for simplicity, assume that the spectral index is in the range $-1$ and $0$, which is broadly consistent with long-term monitoring data between $960$ MHz and $21.7$ GHz from RATAN-600 [@trushkin03]. Then, for the resulting values of $p$ ($1$–$3$), we would expect the peak flux density at $150$ MHz to be about $1$–$3$ per cent of the $4800$-MHz value, i.e. approximately $10$–$30$ mJy. This is much less than the typical calibration uncertainty (Table \[table:observations\]). Conversely, a hypothetical flare of, say, $500$ mJy at $150$ MHz, marginally detectable given our current calibration uncertainty, would be $\sim$tens of Jy at $4800$ MHz. Such a bright outburst has not been observed before from SS433.
An alternative approach is as follows. Firstly, @trushkin03 empirically derived an average power-law relationship between the maximum flux density $S_{\rm max}$ of a flare (after the quiescent component has been subtracted) and the observing frequency $\nu$, such that $$\label{equation flare flux}
S_{\rm max} = 1.3 \nu^{-0.46},$$ where $S_{\rm max}$ is in units of Jy, and $\nu$ is in GHz. Equation \[equation flare flux\] is inconsistent with the predictions of the @vanderlaan66 model for the values of $p$ that we considered above; this equation only agrees with Equation \[equation vdL\] for a significantly inverted spectral index $\alpha \approx 0.83$. However, a negative power-law dependence is, for example, expected in the model developed by @marti92, at frequencies where the outburst is initially optically thin; this model assumes particle injection into twin jets that are adiabatically expanding in the lateral direction, exponentially at first.
From Equation \[equation flare flux\], a hypothetical $1$-Jy flare at $4800$ MHz is brighter than what we would expect on average at this frequency ($0.63$ Jy). Nonetheless, assuming that the average power-law index still holds, a $1$-Jy flare at $4800$ MHz would be about $2$ Jy at $960$ MHz, the lowest RATAN-600 frequency used in @trushkin03. Moreover, in the absence of a spectral turnover, an extrapolation to lower frequencies would suggest a flux density of around $5$ Jy in the LOFAR high band. Despite our limited sampling, there is no evidence of such a large flux density increase at $150$ MHz in Figure \[fig:lightcurve\].
However, if we assume that a flare from SS433 will initially be optically thick below $\sim$$1$ GHz, then we can use equations (29) and (30) in @marti92 to estimate the flux density ratio of the flare between $150$ and $960$ MHz. These equations are $$\label{eqn_marti_ff}
S_{\rm max} \propto \nu^{(2.78p+3.46)/9}$$ in the case of free-free absorption, and $$\label{eqn_marti_ssa}
S_{\rm max} \propto \nu^{(13p+2)/(7p+8)}$$ for synchrotron self-absorption. For the value of $p$ corresponding to Equation \[equation flare flux\], i.e. $p=1.92$, we would therefore expect the $\sim$$2$-Jy flare at $960$ MHz to have a peak flux density of either $\sim$$340$ or $\sim$$200$ mJy at $150$ MHz. Repeating the entire alternative approach for $p = 1$–$3$, including modifying the spectral index from Equation \[equation flare flux\] accordingly, the expected peak flux density at $150$ MHz varies between $\sim$$160$–$440$ mJy for a $1$-Jy flare at $4800$ MHz. The flare, although now brighter than was predicted using Equation \[equation vdL\], would still not be detectable with the current calibration uncertainty.
Also note that as a rough consistency check, if one were to start with the average maximum flux density at $960$ MHz from Equation \[equation flare flux\], i.e. $1.3$ Jy, and do the same analysis as described in the previous paragraph, then the expected peak flux density at $408$ MHz would be either $\sim$$560$ or $\sim$$440$ mJy. On average, flares at $408$ MHz have an excess flux density of $1110 \pm 410$ mJy above the quiescent baseline level [@bonsignori86], which overlaps with our calculated values within at most $1.6\sigma$ (also see @vermeulen93). If the initial transition between optically-thin and optically-thick flares occurs below $\sim$$1$ GHz, then the flares at $408$ MHz would generally be brighter than predicted above using Equations \[eqn\_marti\_ff\] and \[eqn\_marti\_ssa\], and the chances of a detection at $150$ MHz would also increase.
The fact that the $150$- and $4800$-MHz observations were independent observing programmes also needs to be considered. @trushkin03 empirically derived a power-law relationship between the time at which the maximum flux density occurs during a flare, $t_{S_{\rm max}}$, and the observing frequency, such that $$\label{equation flare delay}
t_{S_{\rm max}} = 5.1 \nu^{-0.43},$$ where $t_{S_{\rm max}}$ is in units of days, and $\nu$ is in GHz. Equation \[equation flare delay\] suggests a delay of about $9$ days between $4800$ and $150$ MHz, assuming that it is still valid at low frequencies. The first LOFAR observation after the $4800$-MHz flare on MJD $56442.99$ occurred $56$ days afterwards, while there was a LOFAR observation $4.1$ days after the flare on MJD $56724.22$. While in theory we may have been able to detect the flare rise with LOFAR in the latter case, the extended $4800$-MHz activity before this particular flare may complicate the analysis at low frequencies (Section \[extended flaring\]). Closely coordinating observations across a wide range of radio frequencies will be the most robust approach for studying future flaring activity from SS433.
### Extended flaring activity {#extended flaring}
The onset of $4800$-MHz activity at MJD $\sim$56560 is accompanied by a sustained $\sim$$0.5$–$1$ Jy rise in flux in the LOFAR band. There is also a hint of a similar flux density increase when comparing the $150$- and $4800$-MHz data taken at times up to and including the epoch corresponding to the second LOFAR data point. It is possible that the numerous separate high-frequency bursts have become blended together at $150$ MHz, owing to the longer associated rise time-scales at low frequencies. This is also very tentatively suggested by the apparent difference between the $150$-MHz flux densities at the start and end of the $4800$-MHz flaring activity beginning at MJD $\sim$56560: approximately $1.63$ and $1.89$ Jy, respectively.
The median peak brightness of these flares at $4800$ MHz is about $360$ mJy, after correcting for the quiescent levels on either side of the period of activity. A few tens of flares, spaced $\sim$5 days apart, would not replicate the observed rise in the LOFAR flux density if the @vanderlaan66 model is assumed: a plateau can be modelled, but it is over an order of magnitude too faint and decays too rapidly. Although beyond the scope of this paper, if individual flares are instead much brighter at $150$ MHz, as, for example, was suggested above in Section \[individual flares\] when considering the @marti92 model, then a sustained rise of order several hundred mJy or more is a possibility. More complex scenarios previously outlined in the literature [e.g. @vermeulen93; @blundell11; @jeffrey16] may also be relevant.
### Possible effects of refractive interstellar scintillation {#scintillation}
Another potential explanation for the relatively slow variability that we observe at $150$ MHz is refractive interstellar scintillation [RISS; e.g. @rickett86]. @pandey07 investigated whether RISS could explain the variability of SS433 in their GMRT data at $235$ and $610$ MHz, concluding that the source is most likely affected by this phenomenon. Taking their calculated time-scales for RISS and making a standard assumption that this time-scale is proportional to $\lambda^{2.2}$ (where Kolmogorov turbulence is assumed), then the associated time-scale at $150$ MHz is $\sim$$2.3$ yr. This value is roughly double the length of our monitoring campaign, and so it seems unlikely that we are seeing a corresponding effect in our data. A much longer time baseline at $150$ MHz would be beneficial.
Quiescent flux densities
------------------------
Considering the data points between MJD $56407$–$56560$, before the onset of the extended flaring activity, the median $4800$-MHz flux density, $0.44$ Jy, is close to the long-term average quiescent value of $0.43$ Jy (@trushkin03; also note that the median flux density at the end of the flaring activity, $0.40$ Jy, is very similar). In addition, the average flux density of the five LOFAR observations over this period is $1.63 \pm 0.07$ Jy, which is consistent with both $160$ MHz quiescent flux density measurements from @seaquist80: $1.6 \pm 0.2$ Jy (1979 July 1) and $2.2 \pm 0.3$ Jy (1979 July 21 and 22). These data were obtained at a similar angular resolution to the LOFAR observations ($2.3$ arcmin $\times$ $1.9$ arcmin; BPA $0^{\circ}$) and also have an absolute flux density scale that is consistent with our data [@roger73; @seaquist80; @scaife12].
@seaquist80 suggested that the quiescent radio spectrum of SS433 turns over near $300$ MHz, although there was a gap in their data between $160$–$408$ MHz. In Table \[table:quiescent\_fluxes\], we list the average flux densities from this study, along with more recent measurements, including our LOFAR data. We have estimated the quiescent flux densities at $235$ and $610$ MHz from the monitoring data presented by @pandey07; given the possibility of flaring activity and/or the effects of interstellar scintillation during this observing campaign (discussion in @pandey07), our estimate in each case is the range between the minimum flux density and the median flux density.
Figure \[figure:quiescent\_fluxes\] shows the quiescent radio spectrum at low and mid frequencies. We have also plotted an extrapolated empirical power-law fit that was derived by @trushkin03 using higher-frequency RATAN-600 data, where the spectral index $\alpha = -0.60 \pm 0.14$. The spectrum still appears to be consistent with a turnover at around $300$ MHz, as suggested by @seaquist80, and our LOFAR data further suggest that the spectrum has truly turned over at $150$–$160$ MHz. However, there may be discrepancies resulting from the different angular resolutions of the various datasets, which would affect the relative level of contamination from W50. Again, in-band LOFAR flux densities could be used to obtain a more complete picture of the low-frequency properties of the quiescent radio spectrum of SS433. It may also be particularly useful to obtain new HBA measurements at the upper end of the operating frequency range ($210$–$240$ MHz) and GMRT data at $325$ MHz.
[ccc]{} & &\
&\
$80$ & $<1.6^{a}$ & (1)\
$150$ & $1.63 \pm 0.07$ & (2)\
$160$ & $1.9 \pm 0.2$ & (1)\
$235$ & $1.73$–$2.74$ & (3)\
$408$ & $1.1$–$2.6$ & (4,1)\
$408$ & $2.30 \pm 0.32$ & (5)\
$610$ & $0.73$–$1.23$ & (3)\
$843$ & $1.13 \pm 0.09$ & (6)\
\
\
\
\
\
\[table:quiescent\_fluxes\]
![The low- and mid-frequency radio spectrum for SS433 during quiescence, using the flux densities from Table \[table:quiescent\_fluxes\]. The dashed line is the RATAN-600 quiescent flux density empirical power-law fit from @trushkin03, i.e. $S_{\nu} = 1.1\nu^{-0.6}$ (where $S_{\nu}$ and $\nu$ are in units of Jy and GHz, respectively), extrapolated to lower frequencies. Error bars indicate $\pm 1\sigma$ uncertainties, the solid lines between asterisks indicate flux density ranges at $235$, $408$ and $610$ MHz, and the triangle represents the average $5\sigma$ upper limit at $80$ MHz.[]{data-label="figure:quiescent_fluxes"}](fig7.pdf){width="8.55cm"}
Variability of nearby sources {#variability nearby sources}
-----------------------------
We searched for transients, as well as other variable sources in the field, by running all of our LOFAR maps through the [transients pipeline]{} [[trap]{}; @swinbank15]. No transients were found over the course of our monitoring campaign. Moreover, SS433 has the most significant variability statistics of all the detected sources. Given the discussion in Section \[SS433 variability\_low\], our data are therefore not sufficiently precise to allow us to draw conclusions concerning the potential low-frequency variability of other sources in the field.
D98 detected a $\sim$$340$-mJy point source at $327.5$ MHz, close to the bright filaments in the eastern wing of W50 (coordinates RA $19^{\rm h}$$15^{\rm m}$$49.8^{\rm s}$, Dec. $+04^{\circ}$$51\arcmin$$45\arcsec$; J2000). We have labelled the position of this source with a white cross in Figure \[fig:deepmap2\]. Given several non-detections or marginal detections at higher frequencies, D98 suggested that it may be variable. We do not detect this source in any of our LOFAR datasets (e.g. $3\sigma$ upper limit $\sim$$25$ mJy beam$^{-1}$ in the deep observation, after taking into account the background level from W50). Furthermore, the source appears not to be present in the @farnes17 ATCA map. There are also non-detections in four $327.5$-MHz VLA maps over the period 2001 January–October, as well as the 1996 August 19 run used in the D98 study; the $3\sigma$ upper limit from the map generated using all five observations is $\sim$$17$ mJy beam$^{-1}$ (J. Miller-Jones, priv. comm.). This result suggests that the source may have only been present in the second $327.5$-MHz dataset that D98 obtained on 1997 August 18. Another possibility may be that the source is an artefact generated during the imaging process described in D98, which involved source subtraction and re-addition steps; see section 2.2 in D98.
Candidate supernova remnant G38.7$-$1.4 {#SNR candidate}
=======================================
{width="9.0cm"} {width="7.8cm"}
In Figure \[fig:deepmap\], west-south-west of SS433 and W50, there is a diffuse radio source that is coincident with the candidate SNR G38.7$-$1.4 (also known as SNR G38.7$-$1.3). This candidate was first suggested by @schaudel02, who searched the *ROSAT* all-sky survey [@voges99] for extended X-ray sources that could be unidentified SNRs. They found a $12$ arcmin $\times$ $8$ arcmin source that is coincident with an incomplete radio shell observed in the Effelsberg Galactic Plane Survey at $11$ cm [@reich84; @reich90], favouring a SNR interpretation. An independent identification was also made at optical wavelengths in the IPHAS H$\alpha$ survey by @sabin13, who further supported their claims through comparisons with the same radio shell, as seen in other archival radio datasets.
Additional evidence for a SNR identification was found by @huang14. Using *XMM-Newton* and *Chandra X-ray Observatory* observations, these authors suggested that the X-ray emission is from shock-heated plasma. In addition, their analysis of archival VLA radio observations at $1.4$ GHz revealed a coincident broken radio shell, consistent with that reported by @schaudel02. @huang14 estimated that G38.7$-$1.4 is at a distance of $\sim$$4$ kpc, with an age of $\sim$$14\,000$ yr, and also suggested that the source belongs to the SNR mixed-morphology category [@rho98].
Although previously detected as a broken radio shell, our deep observation of SS433 and W50 has sufficient surface-brightness sensitivity and angular resolution to reveal a more complete radio shell morphology (left-hand panel of Figure \[fig:snr\]). In particular, the candidate supernova remnant is significantly larger than had been indicated by the X-ray observations, with a diameter of approximately $40$ arcmin $\times$ $30$ arcmin, centred at Galactic coordinates $(l,b) = (38\fdg7, -1\fdg1)$. However, accurate flux density measurements are hampered by the fact that the source sits in a significant negative bowl. After making a first-order background correction and assuming that the compact sources coincident with the shell are unrelated, the average $140$-MHz surface brightness of the diffuse emission is $\sim$$30$ mJy beam$^{-1}$.
Traditionally, the confirmation of a SNR is achieved by demonstrating that the radio emission is non-thermal and/or linearly polarized. A polarimetric analysis of our LOFAR dataset was not possible at the time of writing. Also, while the south-eastern part of the shell (labelled ‘SE’ in the left-hand panel of Figure \[fig:snr\]) falls within the ATCA $2.3$-GHz mosaic from the @farnes17 study, no associated linearly polarized emission is detected, which is suggested to be due to foreground depolarization. Therefore, the potential non-thermal nature of G38.7$-$1.4 can instead be tested by using the selection criteria outlined in @brogan06: other than having a shell-like or partial-shell-like morphology, the candidate must have a negative spectral index (where $S_{\nu} \propto \nu^{\alpha}$, as before) and must show a lack of associated mid-infrared $8$-$\upmu$m emission.
The VLA Galactic Plane Survey [VGPS; @stil06], conducted at $1420$ MHz with $60$-arcsec angular resolution, covers the region containing G38.7$-$1.4. While only a hint of the south-eastern part of the shell is detected in the VGPS map, the north-western part of the shell (labelled ‘NW’ in the left-hand panel of Figure \[fig:snr\]) is slightly brighter. Therefore, in order to calculate a rough spectral index for G38.7$-$1.4, we first measured the flux density from the brightest knot of diffuse radio emission that we assume to be associated with the north-western part of the shell, indicated in the left-hand panel of Figure \[fig:snr\]. Our estimate is $\alpha^{1420}_{140} \sim -0.8 \pm 0.1$, where the $1\sigma$ uncertainty is dominated by the uncertainty in the background subtraction in the LOFAR map. The fainter diffuse emission in the vicinity of this feature has a spectral index ranging from about $-1.0$ to $-0.5$, although the $1\sigma$ spectral index uncertainty can be as large as about $0.3$. Moreover, using a $3 \sigma$ upper limit at $1420$ MHz, an approximate spectral index constraint for the south-eastern part of the shell is $\alpha^{1420}_{140} \lesssim -0.6$. Therefore, given these various values, there is no significant evidence to suggest a discrepancy in relation to the general SNR spectral index distribution (e.g. $\overline{\alpha} \approx -0.48$ and $\sigma_{\alpha} \approx 0.15$ in the @green14 SNR catalogue), and, in addition, our measurements are indicative of non-thermal radiation.
Referring to the third selection criterion from @brogan06, at the position of G38.7$-$1.4, there is no evidence for associated mid-infrared emission at $8$ $\upmu$m, using Galactic plane survey data collected by the *Midcourse Space Experiment* [*MSX*; @price01]. The MSX map is shown in the right-hand panel of Figure \[fig:snr\]. This result further suggests a non-thermal nature for the shell, strengthening the potential SNR identification.
We also investigated the possibility of nearby pulsar associations by searching the Australia Telescope National Facility Pulsar Catalogue [@manchester05].[^6] The two closest known pulsars are PSR J1904$+$0451 (angular distance $\approx$ 21 arcmin from the centre of the nebula), with a $1400$-MHz flux density of $0.117$ mJy, and PSR J1906$+$0414 (angular distance $\approx$ $26$ arcmin), with a $1400$-MHz flux density of $0.23$ mJy. Their positions are indicated in the left-hand panel of Figure \[fig:snr\]. At $140$ MHz, the $3\sigma$ flux density upper limits are $\sim$$15$–$20$ mJy beam$^{-1}$ (background corrected). PSR J1904$+$0451 is an isolated millisecond pulsar with a spin period of $6.09$ ms; the dispersion measure distance and characteristic age are $4.7$ kpc (using the @cordes02 model) and $1.7 \times 10^{10}$ yr, respectively [@scholz15a; @scholz15b]. PSR J1906$+$0414 is also not part of a binary system and has a spin period of $1.04$ s; furthermore, the dispersion measure distance and characteristic age are $7.5$ kpc and $1.4 \times 10^{6}$ yr, respectively [@lorimer06]. While there is certainly a very large discrepancy between the characteristic ages of the two pulsars and the estimated age of the SNR candidate, the potentially significant uncertainties associated with the various age and distance estimates do not allow any strong conclusions concerning a pulsar-SNR association to be made [e.g. following the guidelines in @kaspi96]. Moreover, assuming an age of roughly $14\,000$ yr, an association in either case would imply that the pulsar has a projected velocity in the plane of the sky of up to several thousand km s$^{-1}$, well into the tail of the velocity distribution for isolated pulsars [e.g. @arzoumanian02].
@brogan06 demonstrated through $330$-MHz VLA observations of the Galactic plane that deep, mid-frequency surveys can significantly increase the number of known SNRs, helping to solve the long-standing problem of ‘missing’ SNRs in the Galaxy [e.g. @helfand89; @brogan04]. Our detection of G38.7$-$1.4 with LOFAR is an encouraging result in this sense: although this source was previously known, we have shown that low-frequency observations with competitive low-surface-brightness sensitivity can also aid in the detection of diffuse emission. This is particularly so given that the surface brightness values will generally increase with decreasing frequency for the SNR spectral index distribution stated above. Galactic plane surveys conducted with the current generation of low-frequency radio telescopes therefore have the potential to reveal a hitherto missing population.
Conclusions
===========
In this paper, we have presented the results from a LOFAR high-band observing campaign of the X-ray binary SS433 and its associated supernova remnant W50. We have drawn the following conclusions from our study:
1. Our $140$-MHz wide-field image of SS433 and W50 is the deepest low-frequency map of this system to date. The morphology of W50, as well as the flux densities of the various components, are consistent with previous investigations. Foreground free-free absorption may explain the curvature of the radio spectrum of the eastern wing, although more low-frequency data points are needed to confirm this result.
2. We have tentatively detected variability from SS433, with both a debiased modulation index of $11$ per cent and a $\chi^2$ probability of a flat light curve of $8.2 \times 10^{-3}$. The $\sim$$0.5$–$1$ Jy rise in the $150$-MHz flux density may correspond to extended flaring activity over a period of about six months, observed at $4800$ MHz in contemporaneous RATAN-600 observations. The flux density increase is too large to explain with a standard synchrotron bubble model, although a model along the lines of the one presented in @marti92, involving laterally expanding twin jets, may potentially offer an alternative explanation. While LOFAR could be a key trigger for other facilities if there is a prolonged period of activity, low-frequency detections of individual flares will require more accurate calibration, as well as higher-cadence sampling.
3. Within the large field of view, a significant amount of extended structure has been detected near and along the Galactic plane. We have described the most complete detection thus far of the radio shell of SNR candidate G38.7$-$1.4.
Our findings demonstrate the potential of LOFAR, as well as other low-frequency Square Kilometre Array precursors and pathfinders, for studying not only X-ray binaries, but supernova remnants as well.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous referee for their insightful, detailed comments, which were valuable in refining our analysis and presentation. We also acknowledge the considerable efforts of the ASTRON Radio Observatory, particularly Carmen Toribio, in setting up the observations and pre-processing the data. In addition, we warmly thank George Heald and Christian Kaiser for useful discussions, as well as Gloria Dubner for kindly providing us with a $1465$-MHz map of SS433 and W50.
JWB, RPF, AJS, GEA, TDS and MP acknowledge support from European Research Council Advanced Grant 267697 ‘4 Pi Sky: Extreme Astrophysics with Revolutionary Radio Telescopes’. AR, JDS, AJvdH, DC, GJM and RAMJW acknowledge support from European Research Council Advanced Grant 247295 ‘AARTFAAC’. JCAM-J is the recipient of an Australian Research Council Future Fellowship (FT140101082). SAT is grateful to the Russian Base Research Foundation for financial support (grant number 12-02-0812a) and acknowledges support through the Russian Government Program of Competitive Growth of Kazan Federal University. RPB acknowledges support from the European Research Council under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 715051; ‘Spiders’). SC acknowledges financial support from the UnivEarthS Labex program of Sorbonne Paris Cité (ANR-10-LABX-0023 and ANR-11-IDEX-0005-02). JWTH acknowledges funding from an NWO Vidi fellowship and European Research Council Starting Grant “DRAGNET” (337062). JvL acknowledges funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 617199. The financial assistance of the South African SKA Project (SKA SA) towards this research is hereby acknowledged. Opinions expressed and conclusions arrived at are those of the authors and are not necessarily to be attributed to the SKA SA.
This paper is based (in part) on data obtained with the International LOFAR Telescope (ILT). LOFAR [@vanhaarlem13] is the Low Frequency Array designed and constructed by ASTRON. It has facilities in several countries, that are owned by various parties (each with their own funding sources), and that are collectively operated by the ILT foundation under a joint scientific policy.
This research has made use of the SIMBAD database and the VizieR catalogue access tool, operated at CDS, Strasbourg, France. This research has also made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
This research made use of data products from the *Midcourse Space Experiment*. Processing of the data was funded by the Ballistic Missile Defense Organization with additional support from NASA Office of Space Science. This research has also 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.
Abell G. O., Margon B., 1979, Nature, 279, 701 Altenhoff W. J., Downes D., Goad L., Maxwell A., Rinehart R., 1970, A&AS, 1, 319 Alves M. I. R., Davies R. D., Dickinson C., Davis R. J., Auld R. R., Calabretta M., Staveley-Smith L., 2010, MNRAS, 405, 1654 Alves M. I. R., Davies R. D., Dickinson C., Calabretta M., Davis R., Staveley-Smith L., 2012, MNRAS, 422, 2429 Anderson M. C., Rudnick L., 1993, ApJ, 408, 514 Arbutina B., Uro[š]{}evi[ć]{} D., Andjeli[ć]{} M. M., Pavlovi[ć]{} M. Z., Vukoti[ć]{} B., 2012, ApJ, 746, 79 Arbutina B., Uro[š]{}evi[ć]{} D., Vu[č]{}eti[ć]{} M. M., Pavlovi[ć]{} M. Z., Vukoti[ć]{} B., 2013, ApJ, 777, 31 Arzoumanian Z., Chernoff D. F., Cordes J. M., 2002, ApJ, 568, 289 Baars J. W. M., Genzel R., Pauliny-Toth I. I. K., Witzel A., 1977, A&A, 61, 99 Ball L., Vlassis M., 1993, PASA, 10, 342 Beck R., Krause M., 2005, Astron. Nachr., 326, 414 Begelman M. C., Hatchett S. P., McKee C. F., Sarazin C. L., Arons J., 1980, ApJ, 238, 722 Bell M. E. et al., 2014, MNRAS, 438, 352 Black C. S., Fesen R. A., 2015, MNRAS, 447, 2540 Blundell K. M., Bowler M. G., 2004, ApJ, 616, L159 Blundell K. M., Mioduszewski A. J., Muxlow T. W. B., Podsiadlowski P., Rupen M. P., 2001, ApJ, 562, L79 Blundell K. M., Schmidtobreick L., Trushkin S., 2011, MNRAS, 417, 2401 Bonsignori-Facondi S. R., Padrielli L., Montebugnoli S., Barbieri R., 1986, A&A, 166, 157 Brand J., Blitz L., 1993, A&A, 275, 67 Braun R., Walterbos R. A. M., 1985, A&A, 143, 307 Bridle A. H., Schwab F. R., 1999, in Taylor G. B., Carilli C. L., Perley R. A., eds, ASP Conf. Ser. Vol. 180, Synthesis Imaging in Radio Astronomy II, Astron. Soc. Pac., San Francisco, p. 371 Briggs D. S., 1995, PhD thesis, New Mexico Institute of Mining and Technology Brogan C. L., Devine K. E., Lazio T. J., Kassim N. E., Tam C. R., Brisken W. F., Dyer K. K., Roberts M. S. E., 2004, AJ, 127, 355 Brogan C. L., Gelfand J. D., Gaensler B. M., Kassim N. E., Lazio T. J. W., 2006, ApJ, 639, L25 Carilli C. L., Perley R. A., Dreher J. W., Leahy J. P., 1991, ApJ, 383, 554 Clark B. G., 1980, A&A, 89, 377 Cohen A. S., Lane W. M., Cotton W. D., Kassim N. E., Lazio T. J. W., Perley R. A., Condon J. J., Erickson W. C., 2007, AJ, 134, 1245 Condon J. J., Cotton W. D., Greisen E. W., Yin Q. F., Perley R. A., Taylor G. B., Broderick J. J., 1998, AJ, 115, 1693 Copetti M. V. F., Schmidt A. A., 1991, MNRAS, 250, 127 Cordes J. M., Lazio T. J. W., 2002, preprint (arXiv:astro-ph/0207156) Crampton D., Georgelin Y. M., Georgelin Y. P., 1978, A&A, 66, 1 Dabbech A., Ferrari C., Mary D., Slezak E., Smirnov O., Kenyon J., 2015, A&A, 576, A7 Downes A. J. B., Salter C. J., Pauls T., 1981a, A&A, 97, 296 Downes A. J. B., Salter C. J., Pauls T., 1981b, A&A, 103, 277 Downes A. J. B., Pauls T., Salter C. J., 1986, MNRAS, 218, 393 Drew J. E. et al., 2005, MNRAS, 362, 753 Dubner G. M., Holdaway M., Goss W. M., Mirabel I. F., 1998, AJ, 116, 1842 (D98) Duffy P., Blundell K. M., 2012, MNRAS, 421, 108 Eikenberry S. S., Cameron P. B., Fierce B. W., Kull D. M., Dror D. H., Houck J. R., Margon B., 2001, ApJ, 561, 1027 Ellingson S. W. et al., 2013, IEEE Trans. Ant. Prop., 61, 2540 Fabian A. C., Rees M. J., 1979, MNRAS, 187, 13P Farnes J. S., Gaensler B. M., Purcell C., Sun X. H., Haverkorn M., Lenc E., O’Sullivan S. P., Akahori T., 2017, MNRAS, 467, 4777 Fesen R. A., Gull T. R., 1986, ApJ, 306, 259 Fiedler R. L. et al., 1987, AJ, 94, 1244 Forbes D., 1989, A&AS, 77, 439 Gaensler B. M., Hunstead R. W., 2000, PASA, 17, 72 Gao X. Y., Han J. L., Reich W., Reich P., Sun X. H., Xiao L., 2011, A&A, 529, A159 Garsden H. et al., 2015, A&A, 575, A90 Geldzahler B. J., Pauls T., Salter C. J., 1980, A&A, 84, 237 Girard J. N., Garsden H., Starck J. L., Corbel S., Woiselle A., Tasse C., McKean J. P., Bobin J., 2015, JINST, 10, C08013 Goodall P. T., Blundell K. M., Bell Burnell S. J., 2011b, MNRAS, 414, 2828 Goodall P. T., Alouani-Bibi F., Blundell K. M., 2011a, MNRAS, 414, 2838 Green D. A., 1990, AJ, 100, 1927 Green D. A., 2014, Bull. Astron. Soc. India, 42, 47 Guzm[á]{}n A. E., May J., Alvarez H., Maeda K., 2011, A&A, 525, A138 Harris D. E., 2005, in Kassim N., Perez M., Junor M., Henning P., eds, ASP Conf. Ser. Vol. 345, From Clark Lake to the Long Wavelength Array: Bill Erickson’s Radio Science, Astron. Soc. Pac., San Francisco, p. 254 Haslam C. G. T., Salter C. J., Stoffel H., Wilson W. E., 1982, A&AS, 47, 1 Heald G. H. et al., 2015, A&A, 582, A123 Helfand D. J., Velusamy T., Becker R. H., Lockman F. J., 1989, ApJ, 341, 151 Hjellming R. M., Johnston K. J., 1981a, Nature, 290, 100 Hjellming R. M., Johnston K. J., 1981b, ApJ, 246, L141 Hjellming R. M., Johnston K. J., 1988, ApJ, 328, 600 Holden D. J., Caswell J. L., 1969, MNRAS, 143, 407 Huang R. H. H., Wu J. H. K., Hui C. Y., Seo K. A., Trepl L., Kong A. K. H., 2014, ApJ, 785, 118 Hurley-Walker N. et al., 2017, MNRAS, 464, 1146 Jeffrey R. M., Blundell K. M., Trushkin S. A., Mioduszewski A. J., 2016, MNRAS, 461, 312 Kaspi V. M., 1996, in Johnston S., Walker M. A., Bailes M., eds, ASP Conf. Ser. Vol. 105, Pulsars: problems and progress, Astron. Soc. Pac., San Francisco, p. 375 Kassim N. E., 1988, ApJS, 68, 715 Kassim N. E., 1989a, ApJ, 347, 915 Kassim N. E., 1989b, ApJS, 71, 799 Kassim N. E., Weiler K. W., Erickson W. C., Wilson T. L., 1989, ApJ, 338, 152 Kovalenko A. V., Pynzar’ A. V., Udal’Tsov V. A., 1994, AZh, 71, 110 Kuchar T. A., Clark F. O., 1997, ApJ, 488, 224 Kurtz S., 2005, in Cesaroni R., Felli M., Churchwell E., Walmsley M., eds, Proc. IAU Symp. 227, Massive star birth: A crossroads of Astrophysics, Cambridge: Cambridge University Press, p. 111 Lane W. M., Cotton W. D., van Velzen S., Clarke T. E., Kassim N. E., Helmboldt J. F., Lazio T. J. W., Cohen A. S., 2014, MNRAS, 440, 327 Leahy D. A., Roger R. S., 1991, AJ, 101, 1033 Leahy D. A., Xizhen Z., Xinji W., Jiale L., 1998, A&A, 339, 601 Liebert J., Angel J. R. P., Hege E. K., Martin P. G., Blair W. P., 1979, Nature, 279, 384 Lockman F. J., Blundell K. M., Goss W. M., 2007, MNRAS, 381, 881 Longair M. S., 1994, High energy astrophysics. Vol.2: Stars, the galaxy and the interstellar medium. Cambridge: Cambridge University Press, 2nd ed. Lorimer D. R. et al., 2006, MNRAS, 372, 777 Manchester R. N., Hobbs G. B., Teoh A., Hobbs M., 2005, AJ, 129, 1993 Margon B., 1984, ARA&A, 22, 507 Margon B., Ford H. C., Katz J. I., Kwitter K. B., Ulrich R. K., Stone R. P. S., Klemola A., 1979a, ApJ, 230, L41 Margon B., Grandi S. A., Stone R. P. S., Ford H. C., 1979b, ApJ, 233, L63 Martí J., Paredes J. M., Estalella R., 1992, A&A, 258, 309 Mezger P. G., Henderson A. P., 1967, ApJ, 147, 471 Milgrom M., 1979, A&A, 76, L3 Miller-Jones J. C. A., Kapinska A. D., Blundell K. M., Stappers B., Braun R., 2007, Proc. Bursts, Pulses and Flickering: Wide-Field Monitoring of the Dynamic Radio Sky, PoS(Dynamic2007)011 Mohan N., Rafferty D., 2015, Astrophysics Source Code Library, record ascl:1502.007 Moldowan A., Safi-Harb S., Fuchs Y., Dubner G., 2005, ASR, 35, 1062 Nord M. E., Henning P. A., Rand R. J., Lazio T. J. W., Kassim N. E., 2006, AJ, 132, 242 Ochsenbein F., Bauer P., Marcout J., 2000, A&AS, 143, 23 Offringa A. R., de Bruyn A. G., Biehl M., Zaroubi S., Bernardi G., Pandey V. N., 2010, MNRAS, 405, 155 Offringa A. R., van de Gronde J. J., Roerdink J. B. T. M., 2012, A&A, 539, A95 Omar A., Chengalur J. N., Anish Roshi D., 2002, A&A, 395, 227 Pal S., Chakrabarti S. K., Kraus A., Mandal S., 2006, Bull. Astron. Soc. India, 34, 1 Paladini R., Davies R. D., De Zotti G., 2004, MNRAS, 347, 237 Pandey M., Rao A. P., Ishwara-Chandra C. H., Durouchoux P., Manchanda R. K. 2007, A&A, 463, 567 Pratley L., McEwen J. D., d’Avezac M., Carrillo R. E., Onose A., Wiaux Y., 2018, MNRAS, 473, 1038 Price S. D., Egan M. P., Carey S. J., Mizuno D. R., Kuchar T. A., 2001, AJ, 121, 2819 Reich W., Fuerst E., Haslam C. G. T., Steffen P., Reif K., 1984, A&AS, 58, 197 Reich W., Fuerst E., Reich P., Reif K., 1990, A&AS, 85, 633 Remazeilles M., Dickinson C., Banday A. J., Bigot-Sazy M.-A., Ghosh T., 2015, MNRAS, 451, 4311 Rho J., Petre R., 1998, ApJ, 503, L167 Rickett B. J., 1986, ApJ, 307, 564 Roger R. S., Bridle A. H., Costain C. H., 1973, AJ, 78, 1030 Rudie G. C., Fesen R. A., Yamada T., 2008, MNRAS, 384, 1200 Sabin L. et al., 2013, MNRAS, 431, 279 Safi-Harb S., [Ö]{}gelman H., 1997, ApJ, 483, 868 Sault R. J., Wieringa M. H., 1994, A&AS, 108, 585 Sault R. J., Teuben P. J., Wright M. C. H., 1995, in Shaw R. A., Payne H. E., Hayes J. J. E., eds, ASP Conf. Ser. Vol. 77, Astronomical Data Analysis Software and Systems IV, Astron. Soc. Pac., San Francisco, p. 433 Scaife A. M. M., Heald G. H., 2012, MNRAS, 423, L30 Schaudel D., Becker W., Voges W., Aschenbach B., Reich W., Weisskopf M., 2002, in Slane P. O., Gaensler B. M., eds, ASP Conf. Ser. Vol. 271, Neutron Stars in Supernova Remnants, Astron. Soc. Pac., San Francisco, p. 391 Scheers L. H. A., 2011, PhD thesis, Astronomical Institute Anton Pannekoek, Univ. Amsterdam Scheuer P. A. G., Williams P. J. S., 1968, ARA&A, 6, 321 Scholz P. et al., 2015a, ApJ, 800, 123 Scholz P. et al., 2015b, ApJ, 805, 85 Seaquist E. R., Gilmore W., Nelson G. J., Payten W. J., Slee O. B., 1980, ApJ, 241, L77 Seaquist E. R., Gilmore W. S., Johnston K. J., Grindlay J. E., 1982, ApJ, 260, 220 Shaver P. A., McGee R. X., Newton L. M., Danks A. C., Pottasch S. R., 1983, MNRAS, 204, 53 Slee O. B., 1995, Aust. J. Phys., 48, 143 Spencer R. E., 1979, Nature, 282, 483 Stephenson C. B., Sanduleak N., 1977, ApJS, 33, 459 Stil J. M. et al., 2006, AJ, 132, 1158 Su H. et al., 2017a, MNRAS, 465, 3163 Su H. et al., 2017b, MNRAS, 472, 828 Subrahmanyan R., Goss W. M., 1995, MNRAS, 275, 755 Swinbank J. D. et al., 2015, A&C, 11, 25 Tasse C., van der Tol S., van Zwieten J., van Diepen G., Bhatnagar S., 2013, A&A, 553, A105 Tingay S. J. et al., 2013, PASA, 30, 7 Trushkin S. A., Bursov N. N., Nizhelskij N. A., 2003, Bull. Spec. Astrophys. Obs., 56, 57 Trushkin S. A., Nizhelskij N. A., Tsybulev P. G., 2012, ATel, 4484 Trushkin S. A., Nizhelskij N. A., Tsybulev P. G., 2014, ATel, 6492 Trushkin S. A., Nizhelskij N. A., Tsybulev P. G., Zhekanis G. V., 2016, ATel, 9481 van der Laan H., 1966, Nature, 211, 1131 van der Tol S., Jeffs B. D., van der Veen A.-J., 2007, IEEE Trans. Sig. Process., 55, 4497 van Haarlem M. P. et al., 2013, A&A, 556, A2 Vermeulen R. C., McAdam W. B., Trushkin S. A., Facondi S. R., Fiedler R. L., Hjellming R. M., Johnston K. J., Corbin J., 1993, A&A, 270, 189 Voges W. et al., 1999, A&A, 349, 389 Watson M. G., Willingale R., Grindlay J. E., Seward F. D., 1983, ApJ, 273, 688 Weiler K. W., Sramek R. A., Panagia N., van der Hulst J. M., Salvati M., 1986, ApJ, 301, 790 Wenger M. et al., 2000, A&AS, 143, 9 Westerhout G., 1958, Bull. Astron. Inst. Netherlands, 14, 215 Zheng H. et al., 2017, MNRAS, 465, 2901
\[lastpage\]
[^1]: In this paper, we use the convention $S_{\nu} \propto \nu^{\alpha}$, where $S_{\nu}$ is the flux density at frequency $\nu$, and $\alpha$ is the spectral index.
[^2]: In this paper, we assume that the distance to SS433 and W50 is $5.5$ kpc [@blundell04; @lockman07]. At this distance, an angular size of $1$ arcmin corresponds to a projected linear size of $1.6$ pc.
[^3]: http://simbad.u-strasbg.fr/simbad/ and http://vizier.u-strasbg.fr/viz-bin/VizieR
[^4]: http://poincare.matf.bg.ac.rs/[.17ex]{}arbo/eqp/
[^5]: http://www.sao.ru/cats/cgi-bin/ss433.cgi
[^6]: http://www.atnf.csiro.au/people/pulsar/psrcat/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Pion dissociation in a medium of hot quark matter is studied. The decay width of pion is found to be large but finite at temperatures much higher than the so called critical temperature of chiral or deconfinement transition. Consequently, pions should coexist with quarks and gluons at such high temperatures. The result is in agreement with the lattice calculations. The implication of the above result in the study of Quark-Gluon plasma is discussed. PACS No. : 24.85.+p, 25.70.-z,12.38.Mh'
---
plus 10pt
Department of Physics, Bose Institute\
93/1, A.P.C.Road\
Calcutta 700 009, India
A strong prediction of Quantum Chromodynamics (QCD), [*the underlying theory of strong interaction*]{}, is that at very high temperature and/or density, the bulk properties of strongly interacting matter would be governed by the quarks and gluons, rather than the usual hadrons. Such a phase is called quark gluon plasma (QGP) [@a] in the literature and the search for such a novel phase of matter constitutes a major area of current research in the field of high energy physics.
The properties and dynamics of QGP are obviously governed by QCD. This conceptually straight forward task is, however, quite formidable in practice, particularly because of the failure of perturbative QCD already in the temperature range in the vicinity of $\Lambda_{QCD}$ ($\sim$ few hundred MeV) [@a1]. Analytical non-perturbative methods are not yet sufficiently developed to be of much use in this context and as such, the lattice formulation of QCD has developed into the primary vehicle for the study of QGP [@c]. In addition to the intensive computation, both in terms of CPU time and numerical complexities, one can only address static properties in the lattice. As a result, the space - time evolution of the system formed in the ultrarelativistic heavy ion collisions remains unapproachable in the framework of the lattice; thus the alternate, classical picture of hydrodynamic evolution, which accounts for the overall energy - momentum conservation in a collective manner and not much else, has been used quite extensively to study the evolution of the QGP [@d]. QCD inputs enter into such a picture through the equation of state of the QGP, preferably evaluated on the lattice ( but more often, through a phenomenological bag model [@b]).
An inescapable feature of the collision process is that the quarks and gluons must, at some epoch, turn into hadrons which would ultimately be detected, never the individual quarks and gluons. The actual process of hadronisation, however, continues to elude us. It has been widely postulated that there could be an actual phase transition (the order of which is an open issue), separating the QGP phase from the hadronic phase [@f]. The recent results, showing the lack of thermodynamic equilibrium [@g] in the quark-gluon phase in ultrareletivistic heavy ion collisions, indicate that such an ideal situation is unlikely. It should also be noted at this juncture that although the persistence of non-perturbative effects till very high temperatures was suggested in the literature quite early on [@h], it is only recently that the lattice results have confirmed that non-perturbative hadron like excitations could survive at temperatures far above the chiral phase transition temperature [@i]. It is thus imperative to understand the behaviour of such hadronic resonances, their formation, stability and so on, in a quark gluon medium at high temperature. In this work we confine our attention to the case of pions alone, which is naturally a prototype of all hadrons. Being the lightest hadron, pions account for the bulk of the multiplicity.
Formation of pions, a bound state of light relativistic quarks, is an extremely difficult problem to handle in QCD. This is where all the troublesome features of non-perturbative QCD would make their presence felt. We therefore employ the usual practice of looking at the pion as a Goldstone boson arising from the spontaneous breaking of the chiral symmetry. The coupling of the pion to the quarks can then be obtained in a straightforward manner, by starting with the free lagrangian, imposing a chiral transformation and demanding invariance under such a transformation. Explicitly, the chiral angle is associated with the pion field and the quark field is rotated by the chiral [@k] transformation, $$q^{'}= exp \left[i {{\vec\pi.{\vec \tau} \gamma^{5}}\over {2 f_{\pi}}}\right ] q
\label{eq:1}$$ Expanding the exponential to first power in ${1 \over {f_\pi}}$, we obtain the pion-quark coupling. The interaction term is given by [@k], $${\cal L}_{int} = {{m_q} \over {f_\pi}}{\bar q} \gamma_5 {\vec \tau} .
{\vec \pi} q
\label{eq:intt}$$ where $m_q$ is the quark mass, $f_{\pi}$ is the usual pion decay constant (=93 MeV), $q$ is quark field and $\tau$ is the usual Pauli matrix. 0.3in
[cl]{}
0.2in
[ Figure 1 : Temperature dependence of pion mass from lattice data.]{}
0.3in
Even with such an interaction, the formation of pions from quarks and gluons would require an involved analysis through the Bethe-Salpeter equation. Such a study is very much on our agenda but we do not address this issue here. In the present work, we concern ourselves with the decay of pionic excitations, the properties of which we assume to be given by the lattice calculations. It should be reiterated that at temperatures above the critical temperature, these pionic excitations are more like resonances with large effective masses [@i; @j]. The variation of the pion mass with temperature, as calculated in the lattice [@j], is shown in figure 1. In the following, we study the decay width of such pionic excitations in the hot quark medium as a function of temperature, starting with the interaction given above in equation (\[eq:intt\]).
The quark mass $m_{q}$ appearing in eq. (\[eq:intt\]) is a very important ingredient in our calculation. In the absence of any medium and/or dynamic effect, $m_{q}$ should assume the value of the current quark mass. On the other hand, we know that due to the spontaneous breakdown of the chiral symmetry, quarks attain the value of the constituent quark mass [@k]. However, we are investigating the behaviour of pions in quark medium at very high temperatures ($>>$ chiral symmetry restoration temperature) where quarks pick up a large thermal mass [@l] due to medium effects. So, in our calculation, we have taken, for the sake of completeness, three different values for $m_{q}$, namely the current quark mass ($\sim 10$MeV), the constituent quark mass ($\sim 300$MeV) as well as the thermal mass (see below).
The decay width of a pion in its rest frame is given by the usual expression, $$\Gamma = \int {{d^3p_1} \over {2p_1^0(2\pi)^3}}
{{d^3p_2} \over {2p_2^0(2\pi)^3}} {{(2\pi)^4 \delta^4(Q - p_1 - p_2)}
\over {2Q_0}} |M|^2 (1-f(p_{1})) (1-f(p_{2}))$$ where $M$ is the matrix element, $Q$ is the momentum of the pion and $p_1$ and $p_2$ are the momenta of $q$ and ${\bar q}$. $f(p_1)$ and $f(p_2)$ are the usual Fermi-Dirac distribution functions accounting for the Pauli blocking of final state quarks. The matrix element is given by $$M={{m_q} \over {f_\pi}}\bar {q} \gamma_{5} q \tau$$ From (3) and (4), the final expression for the decay width of the pion to a quark anti-quark pair, in the rest frame of the pion, is given by, $$\Gamma = {1 \over {4\pi {m_{\pi}}^2}} \left[{m_{q}\over f_{\pi}}\right]^{2}
({m_{\pi}}^{2} - 4 {m_{q}}^{2})^{3/2}
(1 - f(E_{1})) (1 - f(E_{2}))$$ where $E_{1}$ and $E_{2}$ are the quark energies. The $\delta$-function in equation (3) yields $E_1 = E_2 = m_{\pi}/2$.
The thermal quark mass is defined as [@l] $$m_{th} = \sqrt{m_{curr}^2 + {{g_s^2T^2} \over 9}}$$ where [@l] $$\begin{aligned}
g_s &=& \sqrt{4 \pi \alpha_s} \nonumber\\
\alpha_s &=& {{6 \pi} \over 29} ln \left({{3T} \over \Lambda}\right)\end{aligned}$$ In the above equation, $\Lambda$ is the QCD parameter and $m_{curr}$ the current quark mass. We have considered three values of $\Lambda$, $0.3$ , $0.2$ and $0.1$ GeV. The variation of the pion decay width with temperature for different quark masses and $\Lambda$ are shown in figure 2. The decay width for the current quark mass is not included in the figure, as for this case, the decay width is very small; in fact it is practically zero in the scale of the present figure.
Figure 2 shows that the decay width is very high at high temperature ($\sim 0.3$ GeV) and decreases with decreasing temperature, going to zero at around $T = 0.16$ GeV. It is worth noticing that at around the same temperature, the effective pion mass attains the value of the free pion mass (figure 1). The dependence on $\Lambda$ is very clear. The decay width increases with increase in $\Lambda$. However, the temperature at which the decay width goes to zero does not depend sensitively on $\Lambda$. The decay width is found to be maximum for the constituent quark mass. The fact that for the current quark mass the decay width is very small, whereas for the constituent quark mass it is fairly high, leads one to infer that the decay width decreases as the quark mass decreases. This explains the dependence on $\Lambda$ too, as larger $\Lambda$ corresponds to larger $m_q$ at any given temperature. In figure 2 one should also note the ’shoulder’ like structure in decay width around $T = 200 MeV$. This ’shoulder’ is a reflection of the temperature variation of pion mass (figure 1), which also has a ’shoulder’ around that temperature. 0.3in
[cl]{}
0.2in
Figure 2 : Temperature dependence of pion decay width; (a) for the
constituent quark mass, (b), (c) and (d) are for the thermal
quark masses for $\Lambda = 0.3,$ $0.2$ and $0.1$ GeV respectively.
0.3in
The increases of the pion decay width with increasing $m_q$ is somewhat against the common notion that as the difference in the masses of pion and quark increases the decay width should increase. For the interaction used here, the decay width is proportional to $m_q^2$ (equation 2) and as a result the decay width increases with the increase in quark mass. The decay width is maximum when the quark mass is taken to be its constituent mass [*[i.e.]{}*]{} Its value at $T = 290$ MeV is $0.8$ GeV; as a result, the pions formed at that temperature will decay immediately to $q {\bar q}$ pairs.
Our results will have a strong bearing on the study of hadronisation. As already mentioned, the lack of thermodynamic equilibrium in a QGP system implies that one may not get a clear-cut phase transition from QGP to hadrons. Thus, to understand the process of hadronisation, one should really start from a very high temperature ($>>$ expected $T_c$) and then let the system evolve dynamically towards lower temperatures. Here what one would find, as indicated from our present calculation, is that initially a very small number of pions would be present in the system along with quarks and gluons. Then, even if additional pions are formed through $q {\bar q}$ fusion and/or bound state formation, the total number of pions should not increase very fast, as most of them must decay immediately due to the large decay width at such high temperatures. Only in the vicinity of $T \sim 170 MeV$, where the decay width is small, the number of pions would start increasing significantly and gradually become dominant compared to the number of quarks at some lower temperature. However, the exact value of the temperature, at which the decay width goes to zero, will depend on the value of the quark mass considered.
To summarise, we have calculated, for the first time, the decay of pions, which is a prototype of all hadrons, in a hot quark medium. The most interesting and noteworthy feature is that, even without any consideration of the detailed evolution and dynamics of the system, the pionic modes are found to dominate around a temperature of $160$ MeV. Though the question whether this is a signature of a phase transition cannot be addressed within the framework of the present work, the fact that most of the pions decay into quarks, owing to a large decay width at temperatures higher than $T_c$, is a remarkable finding.
It will be interesting to compare the pion decay width obtained here with the decay widths of other mesons. Qualitatively, the same conclusion should hold but it remains to be seen if all hadronic modes start becoming important at about the same temperature. Work in this direction is in progress.
The work of AB and SKG have been supported, in part, by the Department of Atomic Energy (Government of India) and Council of Scientific and Industrial Research (Government of India), respectively. 0.4in
[99]{} E.V.Shuryak, [*Phys. Rep.*]{} [**61**]{}, 71 (1980). See, for example, [*Quark-Gluon Plasma*]{}, Ed. R.C.Hwa, World Scientific, Singapore (1990) F.Karsch, [*Z. Phys.*]{} [**C38**]{}, 147 (1988). J.Alam, S.Raha and B.Sinha, [*Phys. Rep.*]{} [**273**]{}, 243 (1996). P.Hasenfratz and J.Kuti, [*Phys. Rep.*]{} [**40**]{}, 73 (1978). It should be noted that, in the context of QGP or hot hadronic matter, one comes across two kinds of phase transitions, namely, the chiral phase transition and the deconfinement phase transition. The current consensus is that although these two phase transitions are certainly distinct from each other, the critical temperatures may not be much different. We have taken them to be same for the present purpose. E.V.Shuryak, [*Phys. Rev. Lett.*]{} [**68**]{}, 3270 (1992); S.Chakraborty, S.Raha and B.Sinha, [*Mod. Phys. Lett.*]{} [**A7**]{}, 927 (1992); J.Alam, S.Raha and B.Sinha, [*Phys. Rev. Lett.*]{} [**73**]{}, 1895 (1994); T.S.Biro, B.M[ü]{}ller, M.H.Thoma and X.N.Wang, [*Nucl. Phys.*]{} [**A566**]{}, 543c (1994). M.Pl[ü]{}mer, S.Raha and R.M.Weiner, [*Phys. Lett.*]{} [**B139**]{}, 198 (1984); [*Nucl. Phys.*]{} [**A418**]{}, 549c (1984). K.Born [*et al.*]{}, [*Phys. Rev. Lett*]{} [**67**]{}, 302 (1991). T.Frederico and G.A.Miller, [*Phys. Rev.*]{} [**D45**]{}, 4207 (1992). A.Gocksh, [*Phys.Rev.Lett.*]{} [**67**]{}, 1701 (1991). T.Altherr and D.Seibert, [*Phys. Rev.*]{} [**C49**]{}, 1684 (1994).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the bending of jets in binary stellar systems. A compact companion accretes mass from the slow wind of the mass-losing primary star, forms an accretion disk, and blows two opposite jets. These fast jets are bent by the slow wind. Disregarding the orbital motion, we find the dependence of the bending angle on the properties of the slow wind and the jets. Bending of jets is observed in planetary nebulae which are thought to be the descendants of interacting binary stars. For example, in some of these planetary nebulae the two bubbles (lobes) which are inflated by the two opposite jets, are displaced to the same side of the symmetry axis of the nebula. Similar displacements are observed in bubble pairs in the center of some clusters and groups of galaxies. We compare the bending of jets in binary stellar systems with that in clusters of galaxies.'
author:
- Noam Soker and Gili Bisker
title: 'BUBBLES IN PLANETARY NEBULAE AND CLUSTERS OF GALAXIES: JET BENDING'
---
INTRODUCTION
============
The nebular gas in planetary nebulae (PNs) originates in the envelope of asymptotic giant branch (AGB) stars that are the descendants of intermediate mass stars (initial masses $\sim 1-8 M_\odot$). Such stars rotate very slowly, and their mass loss is expected to be spherical. Indeed, AGB stellar winds usually consist of a more or less spherically symmetric outflow at rates of $\sim 10^{-7}-10^{-5} M_\odot \yr ^{-1}$. Most PNs, though, possess a global axisymmetrical structure rather than a spherical structure in their inner region, indicating a non-spherical shaping process. Among the several PN shaping models (Balick & Frank 2002), one of the most successful is the jet-shaping model. If the jets are not well collimated they are termed collimated fast wind (CFW). The presence of jets in PNs was deduced from observations more than 20 years ago (e.g., Feibelman 1985). Gieseking et al. (1985) found collimated outflow in the PN NGC 2392, and noted the similarity of these jets with that of young stellar objects, and speculated that such outflows exist in many similar PNs. On the theoretical side, Morris (1987) suggested that two jets blown by an accreting companion (the secondary star) can form bipolar nebulae. This model is strongly supported by the similarity of bipolar PNs to many bipolar symbiotic nebulae which are known to be shaped by jets (e.g., Schwarz et al. 1989; Corradi & Schwarz 1995). Soker (1990) proposed that the two fast low-ionization emission blobs (FLIERs or [*ansae*]{}) along the symmetry axis of many elliptical PNs are formed by jets blown during the last phase of the AGB or the post-AGB phase of the PN progenitor. The high quality HST images led Sahai & Trauger (1998) to suggest that in many PNs the non-spherical structures are formed solely by jets.
Projecting from similar astronomical objects, the formation of massive jets, to distinguish from magnetized low density pulsar jets, require the presence of accretion disks. The only source of angular momentum sufficient to form accretion disks in evolved stars is the orbital angular momentum of a stellar (or in some cases substellar) companion. The disk can be formed around the progenitor during the late post-AGB phase, when it is already small (Soker & Livio 1994), or, more likely, around a stellar companion accreting mass, forming an accretion disk, and blowing two jets.
The past seven years have seen further consolidation of the bipolar jet-shaping model in binary systems, addressed both in observations (e.g., Parthasarathy et al. 2000; Sahai & Nyman 2000; Miranda et al. 2001a,b; Corradi et al. 2001; Guerrero et al. 2001; Vinkovic et al. 2004; Huggins et al. 2004; Pena et al. 2004; Balick & Hajian 2004; Arrieta et al. 2005; Oppenheimer et al. 2005; Sahai et al. 2005) and in theory (e.g., Soker 2002, 2005; Lee & Sahai 2003, 2004; Livio & Soker 2001; Garcia-Arredondo& Frank 2004; Velazquez et al. 2004; Riera et al. 2005). Many of the PNs in the observations listed above posses point symmetric morphology, i.e., several symmetry axes rotate with respect to each other through a common origin, indicating precessing jets. The most likely explanation for precession is an accretion disk in the presence of a companion. Soker & Rappaport (2000) further discussed the jet shaping process and have shown that the statistical distribution of bipolar PNs can be accounted for in the binary model. Further support for the formation of jets in binary systems comes from X-ray observations hinting at jets in a PN (Kastner et al. 2003) similar to X-ray jets in symbiotic systems (Kellogg et al. 2001; Galloway & Sokoloski 2004). Garcia-Arredondo & Frank (2004) were the first to conduct 3D numerical simulations of the interaction of jets launched by a secondary star with the slow primary wind. Their high quality results strengthen the general stellar-binary jets model, and in particular the conjecture (Soker & Rappaport 2000) that a narrow waist can be formed by jets. It should be stressed that not all PNs are shaped by jets, but bubble pairs are formed by jets. X-ray images of active galactic nuclei in clusters of galaxies indeed show that double-jets, observed in the radio band, can form a bubble pair with a narrow waist between them, similar to narrow waists in PNs with no need for enhanced equatorial mass loss rate, although enhanced equatorial mass loss rate might occur in many PNs.
The subject of the similarity between some morphological structures in clusters of galaxies, as revealed via X-ray observations, and in PNs, as revealed in the visible band, was studied in a series of four papers. ( Soker 2003b). In that paper ( see also Soker 2003a, and section 5 in Soker 2004c) the similarity in morphological structures was discussed [^1]. This similarity is not trivial. Two opposite jets are observed in many young stellar objects (YSOs), however, bubbles pairs similar to those in PNs and in clusters of galaxies are not usually observed around YSOs. (Soker 2004a). It was found that to inflate fat, more or less spherical, bubbles the opening angle of the jets should be large; the half opening angle measured from the symmetry axis of each jet should typically be $\alpha \gtrsim 40 ^\circ$, or the jets might precess. (Soker 2004b). Paper 3 studies the stability of off-center low-density fat bubbles in clusters of galaxies and in PNs to the Rayleigh-Taylor instability. (Pizzolato & Soker 2005). Pizzolato & Soker examined the point symmetric structure of the bubble pair in the cluster MS 0735.6+7421 (McNamara et al. 2005) and compared it to the point symmetric structure of PNs. Point symmetric PNe are thought to be shaped by stellar binary interactions; namely, the presence of a companion to the PN’s progenitor star is required. Pizzolato & Soker (2005) suggested that similar point-symmetric structures in the X-ray deficient cavities of galaxy clusters might be associated with the presence of massive binary black holes.
In this paper, the fifth in the series, we examine the bending of the two jets, and the subsequent bending of the two bubbles inflated by the jets, to the same side of their original symmetry axis (jets’ axis). Such displacement relative to the symmetry axis of bubbles touching the center is seen, for example, in the Perseus cluster of galaxies (Fabian et al. 2000; Dunn et al. 2006), and in the PN NGC 3587 (Guerrero et al. 2003). Dunn et al. (2006) discuss the departure of the two bubbles from their alignment along a cluster center and explain this departure by the two opposite bubbles detaching from the precessing jets at different times. We consider this displacement to result from the ram pressure of the intra cluster medium (ICM). Displacement of bubbles at a distance from the center are seen in the PN NGC 6886 (Terzian & Hajian 2000), and the group of galaxies HCG 62 (Vrtilek et al. 2002). We focus on PNs and related binary stellar objects, e.g., the massive binary stellar system $\eta$ Carinae ($\S 2.1$). The departure of PNs and related binary systems from axisymmetry has been previously studied (Soker & Hadar 2002 and references therein). Our goal here is to derive a simple expression for the bending of jets in binary stellar systems ($\S 2.2$). This expression is not a substitute for future numerical simulations. The results for typical binary systems ($\S 2.2$) can account for some morphological structures in PNs and related systems. Readers interested in only using the relations and the results, can skip $\S 2.1$ and go directly to $\S 2.2$. In $\S 3$ we compare the situation with jet bending in cooling flow clusters, and $\S 4$ is a summary.
BENDING IN A BINARY STELLAR SYSTEM
==================================
Assumptions and Equations
-------------------------
When a compact secondary star accretes from the AGB (or post-AGB) stellar wind only part of the AGB wind is accreted, and the rest expands outward and forms the medium that the jets expand into. In addition, when the jet is still close to the binary system, the AGB wind hits the jet on its side, causing the jet to deflect (Soker & Rappaport 2000). Like precession, this can have large effects on the descendant PN morphology. However, while precession leads to point-symmetric nebula, the deflection of the two oppositely ejected jets is to the same side, leading the two opposite lobes to be bent to the same side; this is the [*bent*]{} departure from axisymmetry according to the classification of Soker & Hadar (2002). The bending interaction can clear the way to radiation, possibly ionizing radiation, from the central binary system to more strongly illuminate the same side in both lobes (bubbles). Due to the orbital motion, this structure forms a revolving light source. Livio & Soker (2001) suggested such a revolving ionizing source model to explain the positional shift of the bright knots in the inner nebular lobes of the M2-9 nebula (Doyle et al. 2000). Soker & Rappaport (2000) derived a simple expression for the bending angle of a narrow jet. In this section we relax some of the assumptions made by Soker & Rappaport and derive a more accurate expression for the bending angle, while still keeping the expression simple. The goal is to derive a simple approximate relation that will give the jet’s bending angle upon specifying the jet’s parameters and slow wind parameters.
The bending interaction is drawn schematically by Soker & Rappaport (2000) and Livio & Soker (2001), and it is shown in Figure \[draw1\]; 3D images of numerical simulations are presented by Garcia-Arredondo & Frank (2004). The slow wind has a spherical mass loss rate of $\dot M_s$ and a relative speed to the primary star of $v_s$. A small fraction of this wind is accreted by the secondary star, forms an accretion disk that blows two jets, with a mass loss rate of $\dot M_j$ into the two jets together, and with a speed of $v_j$ perpendicular to the equatorial plane relative to the secondary star. Although the jets can have a large opening angle and, in many cases, are likely to have a large opening angle, in the present study we assume a narrow jet with a half opening angle $\alpha \ll 1$, and also assume that the jet is bent as one entity (sound crossing time across the jet is very short). The density per unit length along the jet axis is $$m_j = \frac {\dot M_j}{2 v_j}
\label{mj}$$ (recall that $\dot M_j$ is the mass loss rate into the two jets together).
-2.2 cm -3.2 cm
We move to a frame of reference attached to the secondary star in its orbital motion, with a velocity relative to the primary of ${\bf {v}}_{\rm orb}=v_r \hat r + v_\theta \hat \theta$, where $v_\theta \simeq r \dot \theta$, $\theta$ is the relative angle of the two stars in the equatorial plane, and $r$ is the projected distance from the primary to the jet on the equatorial plane; $v_r<0$ when the two stars approach each other. We consider a narrow jet’s segment at a height $z$ above (or below) the equatorial plane. The slow wind segment that hit this segment left the primary at an angle $\beta$ to the equatorial plane (see fig. \[draw1\]) $$\sin \beta = \frac{z} {(z^2+r^2)^{1/2}}.
\label{stheta}$$ The slow wind that hits the jet at a high $z$ above the equatorial plane has a relative velocity to the jet of $$v_{rel}= [v_\theta^2+(v_s \cos \beta-v_r)^2+(v_s \sin \beta)^2]^{1/2}.
\label{vrel}$$ We consider a fast jet $v_j \gg v_s$ that initially expands perpendicularly to the orbital plane, but is then bent by the ram pressure of the slow wind and acquires a velocity parallel to the equatorial plane $v_p$. The ram pressure exerted by the slow wind on the jet in a direction parallel to the equatorial plane is $$P_{ram}=
\rho \left\{ [v_\theta^2+(v_s \cos \beta-v_r)^2]^{1/2}-v_p \right\}^2,
\label{pram1}$$ where the density of the slow wind $$\rho = \frac {\dot M_s}{4 \pi v_s (r^2+z^2)}.
\label{rho}$$
The equation for accelerating the jet in a direction parallel to the equatorial plane (perpendicular to the initial direction of the jet), under the assumption of a fast jet, $v_p \ll v_j$, reads $$\frac{dv_p}{dt}= \frac{P_{ram} 2 z \tan \alpha}{m_j}
\label{dvpdt}$$ Under the assumption of a fast jet, $z=v_j t$ and $dt=dz/v_j$. We also scale velocities by the slow wind speed $$u_r \equiv \frac{v_r}{v_s}; \quad u_\theta \equiv \frac{v_\theta}{v_s}; \quad
u_p \equiv \frac{v_p}{v_s}; \quad u_j \equiv \frac{v_j}{v_s}.
\label{defvs}$$ The equation of motion reads $$\frac{du_p}{dz}= A \left\{
\left[ u_\theta^2+\left( \frac{r}{\sqrt{r^2+z^2}}-u_r \right)^2 \right]^{1/2}
-u_p \right\}^2 \frac{z}{r^2+z^2},
\label{dvpdz}$$ where $$A= \frac{\tan \alpha}{\pi} \frac{\dot M_s}{\dot M_j}
\label{adef}$$
The meaning of the different terms in equation (\[dvpdz\]) are as follows. (1) The factor $A$ is proportional to the ratio of colliding masses. Bending efficiency increases with $A$. (2) The terms $u_\theta$ and $u_r$ result from the orbital motion of the secondary star, which blows the jets, relative to the slow wind. (3) The term $r/(r^2+z^2)^{1/2}$ results from the ram pressure of the slow wind on the jet. The slow wind moves at a velocity $v_s$; but since velocity was scaled by $v_s$, a factor of unity multiplies this term. (4) The numerator in the last term is due to the increase in the jet cross section, and it increases the bending efficiency as the jet expands. (5) The denominator in the last term is the decrease in the slow wind density, and it makes bending less efficient as distance from the primary star grows.
Results for Impulsive Jets
--------------------------
We consider a case in which the jets are blown by a secondary star that is less massive than the primary star. companion. The slow wind is blown by the primary star, residing close to the center of mass of the binary system. The formulation derived above is applicable to continuously blown jets, or jets blown impulsively. However, for the bubbles in PNs, or similar object, to be significantly displaced by the mechanism discussed in $\S 2.1$, the jet should be blown during a short time compare to the orbital period. (Significant displacement from axisymmetry for continuously blown jets can be acquired if the binary system has a large eccentricity; see references in Soker & Hadar 2002.). In many PNs, the jets’ ejection (PN jets refer to the jets blown by the PN progenitor) can take place over a short time period (e.g., Meaburn 2006), which we take to be shorter than the orbital period. For example, the orbital period can be 5-50 years (orbital separation of $\sim 3-20 \AU$), and the ejection event a few years, as in symbiotic-nova outbursts on an accreting WD companion. The mass accretion rate from the primary stellar wind, $\dot M_2$, by a companion of mass $M_2$ at an orbital separation $r_0$ is $$\frac{\dot M_2}{\dot M_s} \simeq
0.05
\left( \frac {M_2}{0.6 M_\odot} \right)^{2}
\left( \frac {v_{rel}}{15 \km \s^{-1}} \right)^{-4}
\left( \frac {r_0}{10 \AU} \right)^{-2}.$$ If in impulsive jets’ ejection $\dot M_j \sim 0.2 \dot M_2$, then for the above mass accretion rate $A \simeq 5.6 \alpha/10^\circ$. In short eruption events, like disk instability or nova-like outbursts on an accreting WD, it might be that $\dot M_j > 0.2 \dot M_2$, and $r_0$ span a range of $\sim 1-30 \AU$. Therefore, we consider $A$ to be be in the range $A \sim 0.1-100$.
The jet is bent, according to equation (\[dvpdz\]), and $u_p$, the velocity component parallel to the equatorial plane and perpendicular to initial velocity of the jet reaches an asymptotic velocity of $$u_{pa} = \left[ u_\theta^2+\left( \frac{r}{\sqrt{r^2+z^2}}-u_r \right)^2 \right]^{1/2}.
\label{upa}$$ where $u_{pa}$ is in unit of the slow wind speed $v_s$. The asymptotic (final) velocity $u_p$ due to the orbital tangential velocity $u_\theta$ does not depend on the factor $A$ or the jet speed $v_j$ (or $u_j=v_j/v_s$). This is approximately true for the radial orbital component $u_r$ as well, meaning that the initial jet velocity component along the secondary stellar orbital motion is quite efficiently reduced to zero. The departure from axisymmetry due to the orbital motion of the star blowing the jet will be small. Therefore, in imposing a noticeable large-scale departure from axisymmetry, where the two jets are bent to the same side, the bending due to the slow wind outflow from the primary star must be considered. This bending is less efficient because the slow wind velocity is not perpendicular to the jet velocity after the jet leaves the equatorial plane, as seen by the decreasing of the term $r/(r^2+z^2)^{1/2}$. Ignoring the orbital motion, equation (\[dvpdz\]) reads $$\frac{du_p}{dz}= A
\left[ \frac{r}{(r^2+z^2)^{1/2}} -u_p \right]^2 \frac{z}{r^2+z^2},
\label{dup}$$ This equation is supplemented by another equation for the jet propagation along the direction perpendicular to the equatorial plane. For a fast jet, $v_j \gg v_s$, this reads, $$\frac{dr}{dz}= \frac{u_p}{u_j}.
\label{dr}$$
Figure \[upf1\] presents the numerical solutions of the last two coupled equations for initial jet’s speed $u_j\equiv v_j/v_s=6$ and for three values of $A$ as function of the distance from the equatorial plane $z$ in units of the orbital separation $r_0$. The velocity $u_p$ is plotted in the upper panel, in the middle panel the projection of the jet distance on the equatorial plane $r$ is drawn (in units of $r_0$), while the lower panel presents the acceleration $du_p/dz$. In Figure \[upaf\], we show the asymptotic velocity $u_{pa}$ as a function of $A$ for $u_j=6$ (the thick line).
-3.2 cm -1.30 cm -3.2 cm
Changing the initial jet’s speed $u_j$ does not change the solution for $u_p$, while the quantity $r-r_0$ is proportional to $u_j^{-1}$, because the bending angle is given by $\tan \phi =u_p/u_j$, so that for faster jets the bending angle decreases; the dependence is $\phi \propto v_j^{-1}$. This can be understood as follows. As the jet speed $v_j$ increases, the time of accelerating the jet by the slow wind’s ram pressure along a distance $dz$ decreases as $v_j^{-1}$; however, the density in the jet decreases as $v_j^{-1}$ as well. Hence the total change in $v_p$ (or $u_p$) along a distance $dz$ does not depend on $v_j$ under our assumptions, in particular the assumption $v_j$ (or $u_j$) is constant and is not influenced by the interaction with the slow wind.
-4.2 cm -3.2 cm
As seen from Figure \[upf1\], for a fast jet (and for typical values used here) $u_p$ almost reaches its terminal speed while $r \simeq r_0$, where $r_0$ is the initial orbital separation. Note that most of the bending occurs for $r$ not much larger than $r_0$. For $r=r_0$ equation (\[dup\]) reads then $$\frac{du_p}{dz}= A
\left[ \frac{r_0}{(r_0^2+z^2)^{1/2}} -u_p \right]^2 \frac{z}{r_0^2+z^2},
\label{dup0}$$ The solution near the origin, when $u_p \ll 1$ is $$u_{p0} (r \sim r_0) \simeq
\frac{A}{2} \frac{z^2}{r_0^2+z^2}.
\label{up0}$$ The asymptotic velocity is reached when the numerical value inside the square brackets in equation (\[dup0\]) is very small, or $$u_{pac} (r \sim r_0) \simeq \frac{r_0}{(r_0^2+z^2)^{1/2}}.
\label{upaa}$$ The change of behavior between the solution near the jets’ origin and the asymptotic solution takes place when $u_{p0} \sim u_{pa}$, which by equations (\[up0\]) and (\[upaa\]) is $$u_{pac} \sim \left( 1+ \frac{1}{A^2} \right)^{1/2} -\frac{1}{A}.
\label{upas}$$ This very crude expression for the asymptotic transverse velocity is drawn by a thin line on Figure \[upaf\].
As the jet leaves the launching accretion disk, it is very dense and no bending occurs, namely, $du_p/dz$ is very small. At large distances from the jet’s origin, the angle $\delta$ is small and bending is no longer efficient. The bending is most efficient at some intermediate value of $z$, after the density of the jets decreases as they expand, but before the angle $\delta$ decreases much. Practically, this intermediate value of $z$ is quite close to the jet’s origin, $z \la r_0$, as is seen in the lower panel of Figure \[upf1\].
If the jet pair in a binary system is known to be blown in a time period much shorter than the orbital period (an impulsive jet pair), and a bending is observed, the bending angle can be used with the thick line in Figure \[upaf\] to find the constant $A$ given in equation (\[adef\]). Thus the relation between the three quantities: the primary stellar mass loss rate $\dot M_s$, the secondary stellar mass loss rate to the two jet $\dot M_j$, and the half opening angle of the jets $\alpha$, can be found.
BENDING IN CLUSTERS {#clben}
===================
Many jet pairs blown by radio galaxies are observed to be bent as a result of the relative motion of the galaxy and the ICM (e.g., Bliton et al. 1998). Radio jets which are strongly bent are called narrow-angle tailed (NAT) radio galaxies, while those with slightly bent jets are called wide-angle tailed (WAT) radio galaxies. Many of the WAT radio galaxies are dominant galaxies in clusters, like cD galaxies (Owen & Rudnick 1976; Burns et al. 1979). A bulk motion of the IC, e.g., as a result of cluster merger, can efficiently bend radio jets (Bliton et al. 1998). A bulk ICM motion relative to the central cD galaxy can exist as a result of merging with a sub-cluster (group of galaxies), as found in several cases (Dupke & Bregman 2005, 2006; Fujita et al. 2006).
The bending process of jets by the ICM was extensively studied (e.g. Balsara & Norman 1992); the calculations are not repeated here. Basically, the bending of jets in clusters is characterized by the curvature radius $R_c$ of the bent jet. Because the ambient density changes slowly with distance from the cluster center, unlike the case in PNs, a constant ambient density is assumed in the region where most of the jet’s bending occurs. An approximate expression for the radius of curvature is (Sarazin et al. 1995) $$R_{\rm {curv}} \sim 2
\left( \frac {L_{2j}}{3 \times 10^{43} \erg \s^{-1}} \right)
\left( \frac {R_j}{1 \kpc} \right)^{-1}
\left( \frac {v_j}{0.1 c} \right)^{-1}
\left( \frac {v_a}{300 \km \s^{-1}} \right)^{-2}
\left( \frac {n_e}{0.1 \cm^{-3}} \right)^{-1} \kpc
\label{rcur}$$ where $n_e$ is the ambient electron density, $R_j$ is the jet’s radius where most bending takes place, the total mechanical (kinetic) power of the two jets, $L_{2j}$, was scaled according to Birzan et al. (2004), and the relative ambient to jet speed $v_a$ according to Malumuth (1992), with a $90^\circ$ angle between the initial jet velocity and the ambient flow. To achieve a noticeable bending in cD clusters we require $R_{\rm curv} \la 10 \kpc$. This implies that even if the relative velocity of the cD galaxy to the ambient medium component perpendicular to the jet axis is $\sim 100 \km \s^{-1}$ we get the required bending, as observed in some jets, or bubbles, blown by cD galaxies. This required bending will occur also for a narrower jet or a faster jet with $v_j$ close to the speed of light $c$,
Our primary interest is to compare the bending process in binary stars to that of jets blown by the dominant cD galaxies at the centers of cooling flow clusters (or galaxies). Common to bending of jets in binary progenitors of PNs (referred to as bending in PNs) and bending of jets at the centers of clusters of galaxies is that the jets are bent by the ram pressure due to the relative motion of the medium the jets expand to and interact with, both in PNs and clusters. This is unlike point-symmetric structures which result from precession; in both PNs and clusters precession is due to the accretion disk that launches the jets and not due to the ambient medium. There are some differences between bending in clusters and bending in binary systems, such as PNs’ progenitors.
1. [*Relative velocity.*]{} The bending considered in PNs is due to the outflow velocity of the slow wind blown by the primary star. This implies that the angle between the jets velocity and the ambient medium velocity, $\delta$ in Figure \[draw1\], decreases very fast. It decreases even if bending does not occur. In clusters the velocity is due to the motion of the cD galaxy relative to the ICM. The angle decreases only because of the bending, and it decreases slowly.
2. [*Densities.*]{} The jet’s density decreases as the jet moves outward, both in clusters and in stellar binary systems. However, in PNs the density of the bending slow wind (see Fig. \[draw1\]) decreases as well, $\rho_s \propto (r^2+z^2)$ (denominator of last term in eq. \[dvpdz\]), while near the center of clusters the ICM density profile is much shallower and decreases slowly with increasing distance.
3. [*Asymmetry in clusters.*]{} The bending process in PNs is the same for the two opposite jets, but this is not necessarily the case in clusters of galaxies. If the jets are not blown perpendicular to the relative velocity between the ICM and the galaxy, then the jet expanding to the same direction the galaxy moves to will feel a larger ram pressure opposing its expansion velocity, and it will be slowed down more efficiently. More important, as this jet is bent, the angle of the relative velocity between the ICM and the galaxy to the jet’s axis will increase to $90^\circ$ before decreasing. In the opposite jet this angle decreases continuously. Therefore, the jet expanding against the ICM motion will be bent more than the other jet. Asymmetry between the two jets in clusters can be also caused by the presence of asymmetric strong magnetic fields in the ICM (Soker 1997), and/or density inhomogeneities such as clouds (Sarazin et al. 1995).
4. [*Late stages of the bending process.* ]{} After reaching their asymptotic bending angle $\phi$, jets in binary stellar systems will not bend any more. If a jet inflates a bubble, it will move outward radially along the streaming slow wind material. In clusters the situation is different because of the flow structure mentioned in points (1) and (2) above and because the low density bubbles buoyant outward. The result is that although the radio jets of cD galaxies are not bent much, after they become subsonic the bending is very efficient (Eilek et al. 1984; Odea & Owen 1986), and the asymmetry between the two sides can substantially increase (Burns et al. 1986).
5. [*The effective bending location.*]{} From the differences in points 1,2, and 4 it turns out that in stellar binary systems most of the bending occurs close to the jets’ origin (lower panel of fig. \[upf1\]). In clusters the bending becomes more efficient as the jet expands and slows down. In particular, if the jet inflate bubbles, they move slowly, have very low density, and large cross section. Thus, in clusters the departure from axisymmetry will be most noticeable in bubbles.
Despite the differences listed above, there are some striking morphological similarities of bubbles displaced from the symmetry axis in PNs and clusters; two cases are mentioned in $\S 1$ (see appendix of the astro-ph version of the paper).
SUMMARY
=======
In recent years the jet shaping model for many PNs and similar objects, like the massive binary star $\eta$ Carinae, acquired considerable acceptance. It should be stressed that not all PNs were shaped by jets, and not all morphological structure in PNs were formed by jets. Bubble pairs, though, are most likely inflated by double jets, and the jets are probably blown by a stellar secondary star. The secondary star accretes mass from the primary’s slow wind, forms an accretion disk and blow two jets, either continuously or impulsively, that is, during a time shorter than the orbital period. In some PNs, the line joining the centers of the two bubbles in a pair does not pass through the center of the nebula, meaning that the bubbles are displaced such that the nebular structure departs from axisymmetry. The explanation is that the two jets that inflate the bubbles were bent to the same side by the ram pressure of the slow wind (Fig. \[draw1\]).
We therefore set the goal of deriving a simple and approximate relation between the bending angle of the jets and the properties of secondary stellar jets and the primary slow wind. For fast jets, $v_j \gg v_s$, the important factor is the quantity $A$ defined in equation (\[adef\]). The relation between the jet’s asymptotic transverse speed $v_{pa}$ (see fig. \[draw1\]) and $A$ is presented by the thick line in Figure \[upaf\], and a very crude approximation is given in equation (\[upas\]) ($v_{pa}$ is in units of the slow wind speed $v_s$). If the jets are impulsive, then the bending will be easier to observe; otherwise it is averaged over different directions as the binary system rotate. If $A$ is not too small, and the orientation of the nebula is such that the bending is not along the line of sight, then observations may reveal the two jets or the bubbles (lobes) inflated by the jets to be displaced to the same side of the symmetry axis. Examples of such PNs are listed and classified by Soker & Hadar (2002).
In some clusters, X-ray-deficient bubble (cavity) pairs that were inflated by jets blown by the central cD galaxy, show displacement from axisymmetry similar to visible-deficient bubble (lobe) pairs observed in PNs (see appendix of the astro-ph version of this paper). We therefore set a second goal of comparing the bending process of jets in these two groups of objects. Two factors of the bending process are common to these two classes of objects: 1) the bending results from the ram pressure perpendicular to the jet axis, and 2) the ram pressure is exerted by the same external medium the jets expand to and interact with.
However, there are some significant differences listed in §\[clben\]. (1) Because the bending in binary systems results from the slow wind blown by the primary star, the ambient density decreases faster with distance than the ambient density of the ICM in the centers of clusters. (2) Also, the angle $\delta$ (see Fig. \[draw1\]) between the jet velocity and ambient slow wind velocity in binary systems decreases with distance along the jet axis, even when bending does not occur. In clusters the relative velocity is due to the bulk ICM motion and changes only because of bending. (3) In binary systems the two opposite jets are likely to be blown perpendicular to the orbital plane, thus they will be bent in the same way. In clusters, the jets’ axis need not be perpendicular relative to the bulk motion of the ICM relative to the central black hole that blows the jets, so the jet facing the ICM flow will be bent more efficiently. In binary systems such an asymmetry between the two jets can occur if the jets (more specifically the accretion disk that launches the jets) precess. (4) In clusters, after the bubbles (cavities; lobes) are inflated, they buoy outward. They are more susceptible than the jets to the ram pressure, and departure from axisymmetry may substantially increase. This process does not exist in binary systems because the circumbinary ambient matter is not in hydrostatic equilibrium, but rather the ambient matter expands at a high Mach number. (5) In binary stars most of the bending occurs when the jets are at a distance $z \la r_0$, where $r_0$ is the orbital separation. In clusters the bending becomes more efficient at larger and larger distances.
We hope that the study presented in this paper will motivate researchers to pay more attention to the departure from axisymmetry of bubble (cavity; lobe) pairs in both clusters of galaxies and PNs.
This research was supported in part by the Asher Fund for Space Research at the Technion.
Arrieta, A., Torres-Peimbert, S., & Georgiev, L. 2005, ApJ, 623, 252
Balick, B., & Frank, A. 2002, ARA&A, 40, 439
Balick, B., & Hajian, A. R. 2004, AJ, 127, 2269
Balsara, D., & Norman, M. L. 1992, ApJ, 393, 631
Birzan, L., Rafferty, D. A., McNamara, B. R., Wise, M. W., & Nulsen, P. E. J. 2004, ApJ, 607, 800
Bliton, M.; Rizza, E.; Burns, J. O.; Owen, F. N.; Ledlow, M. J. 1998, MNRAS, 301, 609B
Burns, J. O., Gregory, S. A., Odea, C. P., & Balonek, T. J. 1986, ApJ, 307, 73
Burns, J. O., Owen, F. N., & Rudnick, L. 1979, AJ, 84, 1683
Corradi, R. L. M., Livio, M., Balick, B., Munari, U., & Schwarz, H. E. 2001, ApJ, 553, 211
Corradi, R. L. M., & Schwarz, H. E. 1995, A&A, 293, 871
Doyle, S., Balick, B., Corradi, R. L. M., & Schwarz, H. E. 2000, AJ, 119, 1339
Dunn, R. J. H., Fabian, A. C., & Sanders, J. S. 2006, MNRAS, in press astro-ph/0512022)
Dupke, R. A., & Bregman, J. N. 2005, ApJS, 161, 224
Dupke, R. A., & Bregman, J. N. 2006, ApJ, in press (astro-ph/0512045)
Eilek, J. A., Burns, J. O., Odea, C. P., Owen, F. N. 1984, ApJ, 278, 37
Fabian, A. C., et al. 2000, MNRAS, 318, L65 Feibelman, W. A., 1985, AJ, 90, 2550
Fujita, Y., Sarazin., C. L., & Sivakoff, G. R. 2006, PASJ, in press (astro-ph/0512308)
Galloway, D. K., & Sokoloski, J. L. 2004, ApJ, 613, L61
Garcia-Arredondo, F., & Frank, A. 2004, ApJ, 600, 992
Gieseking, F., Becker, I., & Solf, J. 1985, ApJ, 295, L17
Guerrero, M. A., Chu, Y.-H., Manchado, A., Kwitter, K. B. 2003, AJ, 125, 3213
Guerrero, M. A., Miranda, L. F., Chu, Y.-H., Rodriguez, M., & Williams, R. M. 2001, ApJ, 563, 883
Huggins, P. J., Muthu, C., Bachiller, R., Forveille, T., & Cox, P. 2004, A&A, 414, 581
Kastner, J. H., Balick, B., Blackman, E. G., Frank, A., Soker, N., Vrtilek, S. D., & Li, J. 2003, ApJ, 591, L37
Kellogg, E., Pedelty, J. A., & Lyon, R. G. 2001, ApJ, 563, L151
Lee, C.-F., & Sahai, R. 2003, ApJ, 586, 319
Lee, C.-F., & Sahai, R. 2004 ,ApJ, 606, 483
Livio, M., & Soker, N. 2001, ApJ, 552, 685
Malumuth, E. M. 1992, ApJ, 386, 420
McNamara, B. R., Nulsen, P. E. J., Wise, M. W., Rafferty, D. A., Carilli, C., Sarazin, C. L. & Blanton, E. L. 2005, Nature, 433 ,45
Meaburn, J. 2006, in ‘Diffuse Matter from Star Forming Regions to Active Galaxies’, Eds. T. W. Harquist, J. M. Pittard and S. A. E. G. Falle (astro-ph/0512099)
Miranda, L. F., Guerrero, M. A., & Torrelles, J. M. 2001a, MNRAS, 322, 195
Miranda, L. F., Torrelles, J. M., Guerrero, M. A., Vazquez, R., & Gomez, Y. 2001b, MNRAS, 321, 487
Morris, M. 1987, PASP, 99, 1115
Odea, C. P., & Owen, F. N. 1986, ApJ, 301, 841
Oppenheimer, B. D., Bieging, J. H., Schmidt, G. D., Gordon, K. D., Misselt, K. A., & Smith, P. S. 2005, ApJ, 624, 957
Owen, F. N., & Rudnick, L. 1976, ApJ, 205, L1
Parthasarathy, M., Garcia-Lario, P., Pottasch, S. R., de Martino, D., Surendiranath, R. 2000, A&A, 355, 720
Pena, M., Hamann, W.-R., Ruiz, M. T., Peimbert, A., & Peimbert, M. 2004, A&A, 419, 583
Pizzolato, F. & Soker, N. 2005, AdSpR, 36, 762 (astro-ph/0501658) (Paper-4)
Riera, A., Raga, A. C., & Alcolea, J. 2005, RMxAA, 41, 147
Sahai, R., Le Mignant, D., Sanchez Contreras, C., Campbell, R. D., & Chaffee, F. H. 2005, ApJ, 622, L53
Sahai, R., & Nyman, L.-A. 2000, ApJ, 538, L145
Sahai, R., & Trauger, J. T. 1998, AJ, 116, 1357
Sarazin, C. L., Burns, J. O., Roettiger, K., & McNamara, B. R. 1995, ApJ, 447, 559
Schwarz, H. E., Aspin, C., & Lutz, J. H. 1989, ApJ, 344, L29
Soker, N. 1990, AJ, 99, 1869
Soker, N. 1997, ApJ, 488, 572
Soker, N. 2002, ApJ, 568, 726 Soker, N. 2003a, Nature, 426, 236 Soker, N. 2003b, PASP, 115, 1296 (Paper-1) Soker, N. 2004a, A&A, 414, 943 (Paper-2) Soker, N. 2004b, NewA, 9, 285 (Paper-3) Soker, N. 2004c, in Asymmetrical Planetary Nebulae III: Winds, Structure and the Thunderbird, eds. M. Meixner, J. H. Kastner, B. Balick, & N. Soker, ASP Conf. Series, 313, (ASP, San Francisco), p. 562 (see the extended version on astro-ph/0309228)
Soker, N. 2005, AJ, 129, 947 Soker, N., & Hadar, R. 2002, MNRAS, 331, 731
Soker, N., & Livio M. 1994, ApJ, 421, 219
Soker, N., & Rappaport, S. 2000, ApJ, 538, 241
Terzian, Y., & Hajian, A. R. 2000, “Asymmetrical Planetary Nebulae II: From Origins to Microstructures,” eds. J.H. Kastner, N. Soker, & S. Rappaport, ASP Conf. Ser. Vol. 199, p. 33 (see PNs images at: http://ad.usno.navy.mil/pne/gallery.html)
Velazquez, P. F., Riera, A., & Raga, A. C. 2004, A&A, 419, 991
Vinkovic, D., Blöcker, T., Hofmann, K.-H., Elitzur, M., & Weigelt, G. 2004, MNRAS, 352, 852
Vrtilek, J. M., Grego, L., David, L. P., Ponman, T. J., Forman, W., Jones, C., & Harris, D. E. 2002, APS, APRB, 17107
**APPENDIX**
In this [*appendix*]{} we summarize the morphological similarities between X-ray deficient bubbles (cavities) in clusters or groups of galaxies and visible-light deficient bubbles (lobes) in PNs. The Table below is based on that given in Paper-1, with some additional comparisons. In Paper 1, Soker also pointed out similar values of some non-dimensional quantities between clusters and PNs. These similarities led Soker to postulate a similar formation mechanism, thereby strengthening models for PN shaping by jets, although not all PNs are shaped by jets; the jets in PN progenitors are likely blown by binary companions.
The images of the objects listed in Table 1 of Paper-1 are summarized in a [*PowerPoint*]{} file Soker presented at the Asymmetrical Planetary Nebulae III meeting (2003) at http://www.astro.washington.edu/balick/APN/APN\_talks\_posters.html go the ‘ppt’ file in the “Discussion” of Session 13. Other images were added here. All image sources are listed, and references not in the list of the paper are listed after the table.
[lll]{} Butterfly shape; faint & Abell 478 \[1\]$^4$ & Roberts 22 \[2\]\
along symmetry axis$^1$ & (Sun et al. 2003) & (Sahai et al. 1999)\
Pairs of fat spherical & Perseus \[3\] & NGC 3587 \[4\]\
bubbles near center & (Fabian et al. 2000) & (Guerrero et al. 2003)\
Closed bubbles connected & Abell 2052 \[5\] & VV 171 \[6\]\
at the equatorial plane & (Blanton et al. 2001) & (Sahai 2001)\
Open bubbles & M 84 & He 2-104\
connected at the & (Finoguenov & Jones & (Sahai & Trauger,\
equatorial plane & 2001) \[7\] & 1998) \[8\]\
Pair of bubbles detached & HCG 62 \[9\] & Hu 2-1 \[10\]\
from a bright center & (Vrtilek et al. 2002) & (Miranda et al. 2001b)\
Point symmetry; & MS 0735.6+7421 \[11\] & Hb 5 \[12\]$^3$\
suggesting precession$^{2}$ & (McNamara et al. 2005) & (Terzian & Hajian 2000)\
Bending to & HCG 62 \[9\] & NGC 6886 \[13\]\
one side & (Vrtilek et al. 2002) & (Terzian & Hajian 2000)\
point-symmetric & Hydra A \[14\] & NGC 6537 \[15\]\
elongated lobes & (McNamara et al. 2000) & (Balick 2000)\
Pairs of bright bullets & Cygnus A \[16\] & NGC 7009 \[17\]\
along the symmetry axis & (Smith et al. 2002) & (Balick et al. 1998)\
Ripples & Perseus \[18\] & M 57 (NGC 6720) \[19\]\
& (Fabian et al. 2003;06) & (Hora et al. 2005)\
Comments to the table in the [*Appendix*]{}: Similar images of bubbles in clusters of galaxies and planetary nebulae (PNs). In clusters these are X-ray images (e.g., with X-ray deficient bubbles), while in PNs they are visible-light images (e.g., with visible-light deficient bubbles). In the first eight pairs of images the similarity is of high degree. In the last two pairs of images the similarity between the cluster and the PN is of lesser degree. See Pizzolato & Soker (2005) for more detail. The low-resolution image of the same object is from the catalogue of Schwarz et al. (1992) Free access to images are at these sites: http://arxiv.org/PS\_cache/astro-ph/pdf/0210/0210054.pdf http://ad.usno.navy.mil/pne/images/rob22.jpg http://arxiv.org/PS\_cache/astro-ph/pdf/0007/0007456.pdf http://arxiv.org/PS\_cache/astro-ph/pdf/0303/0303056.pdf http://arxiv.org/PS\_cache/astro-ph/pdf/0107/0107221.pdf http://ad.usno.navy.mil/pne/images/vv171.jpg http://arxiv.org/PS\_cache/astro-ph/pdf/0010/0010450.pdf http://ad.usno.navy.mil/pne/images/he2\_104.jpg http://chandra.harvard.edu/photo/cycle1/hcg62/index.html http://arxiv.org/PS\_cache/astro-ph/pdf/0009/0009396.pdf see also: \[10b\] http://ad.usno.navy.mil/pne/images/hu21\_ha.gif http://arxiv.org/PS\_cache/astro-ph/pdf/0411/0411553.pdf http://ad.usno.navy.mil/pne/images/hb5.jpg http://ad.usno.navy.mil/pne/images/ngc6886.jpg http://arxiv.org/PS\_cache/astro-ph/pdf/0001/0001402.pdf http://ad.usno.navy.mil/pne/images/ngc6537.jpg http://www.journals.uchicago.edu/ApJ/journal/issues/ApJ/ v565n1/54312/54312.web.pdf http://ad.usno.navy.mil/pne/images/ngc7009.jpg see also (Goncalves et al. 2004, fig. 1): http://arxiv.org/PS\_cache/astro-ph/pdf/0307/0307265.pdf http://chandra.harvard.edu/photo/2003/perseus/perseus\_ripple\_illustration\_nolabel.jpg http://www.spitzer.caltech.edu/Media/releases/ssc2005-07/ssc2005-07a.shtml
Balick, B. 2000, “Asymmetrical Planetary Nebulae II: From Origins to Microstructures,” eds. J.H. Kastner, N. Soker, & S. Rappaport, ASP Conf. Ser. Vol. 199, p. 41
Balick, B., Alexander, J., Hajian, A. R., Terzian, Y., Perinotto, M., & Patriarchi, P. 1998, AJ, 116, 360
Blanton, E. L., Sarazin, C. L., McNamara, B. R., & Wise, M. W. 2001, ApJ, 558, L15
Fabian, A. C., Sanders, J. S., Allen, S. W., Crawford, C. S., Iwasawa, K., Johnstone, R. M., Schmidt, R. W., & Taylor, G. B. 2003, MNRAS, 344, L43
Fabian, A. C., Sanders, J. S., Taylor, G. B., Allen, S. W., Crawford, C. S., Johnstone, R. M., & Iwasawa, K., 2006, MNRAS, in press (astro-ph/0510476)
Finoguenov, A., & Jones, C. 2001, ApJ, 547, L107
Goncalves, D. R., Mampaso, A., Corradi, R. L. M., Perinotto, M., Riera, A., & Lopez-Martin, L. 2004, MNRAS, 355, 37
Hora, J. L., Latter, W. B., Marengo, M., Fazio, G. G., Allen, L. E., & Pipher, J. L. 2005, AAS, 206, 3901
McNamara, B. R., et al. 2000, ApJ, 534, L135 Sahai, R. 2001, in Galactic Structure, Stars and the Interstellar Medium, Eds. Charles E. Woodward, Michael D. Bicay, and J. Michael Shull, ASP Conf. Ser. Vol. 231 (San Francisco: ASP), 528
Sahai, R., Zijlstra, A., Bujarrabal, V., Te Lintel Hekkert, P. 1999, AJ, 117, 1408
Schwarz, H. E., Corradi, R. L. M., & Melnick, J. 1992, A&A Suppl. Ser., 96, 23.
Smith, D. A.; Wilson, A. S., Arnaud, K. A., Terashima, Y., & Young, A. J. 2002, ApJ, 565 195
Sun, M., Jones, C., Murray, S. S., Allen, S. W., Fabian, A. C., & Edge, A. C. 2003, ApJ, 587, 619
-1.2 cm
[^1]: The similar morphologies are compared in the appendix of the astro-ph version of the present paper.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we consider the transmission of confidential messages over slow fading wireless channels in the presence of an eavesdropper. We propose a transmission scheme that employs a single reconfigurable antenna at each of the legitimate partners, whereas the eavesdropper uses a single conventional antenna. A reconfigurable antenna can switch its propagation characteristics over time and thus it perceives different fading channels. It is shown that without channel side information (CSI) at the legitimate partners, the main channel can be transformed into an ergodic regime offering a *secrecy capacity* gain for strict outage constraints. If the legitimate partners have partial or full channel side information (CSI), a sort of selection diversity can be applied boosting the maximum secret communication rate. In this case, fading acts as a friend not a foe.'
author:
- 'Ahmed M. Alaa [^1] [^2]'
title: 'Defeating the Eavesdropper: On the Achievable Secrecy Capacity using Reconfigurable Antennas'
---
[A. M. Alaa : Secret Selection Diversity using Confidential Reconfigurable Antenna State Switching Patterns]{}
Channel state information (CSI), outage probability, outage secrecy capacity, reconfigurable antennas, secrecy capacity.
Introduction
============
theoretic security was quantified by Shannon’s notion of *perfect secrecy*. Perfect information-theoretic secrecy requires the signal received by the eavesdropper not to provide any additional information about the transmitted message [@IEEEhowto:kopka1]. The conventional secret communications scheme includes two legitimate parties, commonly known as Alice and Bob, communicating over a wireless slow fading channel. A malicious third party, known as Eve, eavesdrops on the wireless medium and tries to decode the transmitted signal. In a block fading channel, the channel gain is constant over a codeword, thus the channel is characterized by an outage behavior. The achievable secrecy rate was obtained in terms of the outage probability in [@IEEEhowto:kopka2], where it is shown that for a fading channel, poor secret rates are achieved for strict outage constraints.
Recently, improving the outage secrecy capacity by using multiple antennas has been studied [@IEEEhowto:kopka11] [@IEEEhowto:kopka12] [@IEEEhowto:kopka13]. However, the usage of multiple antennas is inhibited by the space limitations in many wireless transceivers. In addition to that, multiple antennas require multiple RF chains which increases the cost and complexity of the wireless transceiver. In this work, we propose a novel secret communications scheme that employs *reconfigurable antennas*; a class of antennas capable of changing one of its characteristics (polarization, operating frequency and radiation pattern) over time [@IEEEhowto:kopka3] [@IEEEhowto:kopka5] using a single RF chain. Each configuration is known as a *radiation state* and corresponds to an independent channel realization. Previous research work utilized reconfigurable antennas in authentication and secret key generation [@IEEEhowto:kopka15] [@IEEEhowto:kopka17]. However, the achievable capacity bounds for reconfigurable antenna schemes were never obtained before. We propose two modes of legitimate communication via reconfigurable antennas: *state switching* and *state selection*. State switching is applied by the CSI is not available at the transmitter/receiver and relies on switching the antenna *radiation state* over time manipulating the wireless channel and creating artificial channel fluctuations. On the other hand, state selection is applied by selecting the “best" radiation state per codeword for a block fading channel based on the CSI at the transmitter/receiver. It is shown that when strict outage constraints are imposed on the system, state switching can offer an ergodic capacity that exceeds the achievable outage capacity. Moreover, state selection based on partial or full CSI can offer a secrecy capacity that exceeds that of the AWGN channel, thus fading acts as a friend not a foe. State selection resembles opportunistic transmission in a fast fading channel but with power allocation in the *state domain* rather than the time domain, thus supporting both *delay constrained* and *delay tolerant* applications.
As shown in figure 1, we modify the conventional secrecy communications scheme by employing a reconfigurable antenna at both of the legitimate parties. A message $W^k$ is mapped to a codeword $X^n$. The codeword is then transmitted from Alice to Bob via a rayleigh fading channel $\gamma^n_{M} = \{\gamma^n_{M}(1), \ldots, \gamma^n_{M}(n)\}$, and Additive White Gaussian Noise (AWGN) = $\{n^n_{M}(1), \ldots, n^n_{M}(n)\}$, where $n^n_{M}(i)$ $\sim$ $\mathcal{CN}(0, 1)$. The estimated message by the decoder is obtained by demapping the received signal $Y^n_{M}$ to $\hat{W}^k$. Eve, an eavesdropper, receives the signal via a similar channel $\gamma^n_{W}$, and noise $n^n_{W}$. While Eve uses a conventional single antenna, both Bob and Alice use reconfigurable antennas with $Q_{R}$ and $Q_{T}$ propagation modes respectively. The realizations $\gamma^n_{M}(i)$ and $\gamma^n_{W}(i)$ are the legitimate and eavesdropper channel realizations for the $i^{th}$ symbol within a codeword of length $n$. For a slow fading channel, both are constant over a codeword. However, a reconfigurable antenna is capable of switching the channel state once per symbol, thus there are $Q_{R}Q_{T}$ possible realizations for the main channel $\big(\gamma^n_{M}(i) \in \{\gamma_{M}(1),\ldots,\gamma_{M}(Q_{R}Q_{T})\}\big)$ and $Q_{T}$ possible realizations for the eavesdropper channel. Note that all channels are assumed to be Rayleigh fading channels, thus the *probability density functions* (pdfs) of the channels are $f_{\gamma}(\gamma_{M}) = \frac{1}{\overline{\gamma}_{M}} e^{-\frac{\gamma_{M}}{\overline{\gamma}_{M}}}$ and $f_{\gamma}(\gamma_{W}) = \frac{1}{\overline{\gamma}_{W}} e^{-\frac{\gamma_{W}}{\overline{\gamma}_{W}}}$, where $\overline{\gamma}_{W}$ and $\overline{\gamma}_{M}$ are the average SNR values for the main and eavesdropper channels. We define the error probability $P_{e}^n$ as the average probability of erroneous decoding of the received message, and the equivocation rate $R_{e}$ as the entropy rate of the transmitted message conditioned on the channel outputs at the eavesdropper, i.e., $R_{e} \triangleq \frac{1}{n} H(W^k|Y^n_{W})$.
Our goal is to maximize both the transmission rate between Alice and Bob in addition to Eve’s uncertainty about the message (equivocation rate). A relaxed secrecy condition is letting $\frac{1}{n} H(W^k)-R_{e} \leq \epsilon$ and $P_{e}^n \leq \epsilon$, where $\epsilon \to 0$ as $n \to \infty$. The secrecy capacity is defined as the maximum achievable secrecy rate for all sequences of $(2^{nR_{s}},n)$ codes $$\label{eqn_1}
C_{s} \triangleq \sup\limits_{P_{e}^n \leq \epsilon} R_{s}.$$ The secrecy capacity for AWGN channels (or fixed $\gamma_{M}$ and $\gamma_{W}$ realizations) is given by $$\label{eqn_2}
C_{s} = \big\{\log(1+\gamma_{M})-\log(1+\gamma_{W})\big\}^{+},$$ where $\{x\}^{+} = \max\{x, 0\}$. Thus, the AWGN secrecy capacity is given by the difference between legitimate and eavesdropper channel capacities. By defining the outage probability as $P_{out} = P(C_{s} \leq R_{s})$, the $\epsilon$-outage secrecy capacity is the value of $R_{s}$ that satisfies $P_{out} = \epsilon$.
Secrecy Capacity of Transmission Schemes under Study
====================================================
Conventional scheme
-------------------
In this scheme, all parties are using single antennas and the channel is a block fading channel. Assuming that the legitimate and eavesdropper channel realizations are $\gamma_{M}$ and $\gamma_{W}$ respectively, the secrecy capacity for one realization of both channels is given by (\[eqn\_2\]). For a quasi-static fading channel, we characterize the performance via outage secrecy capacity and outage probability. The outage probability is $P_{out}(R_{s}) = P(C_{s}<R_{s})$, and can be written as $$P_{out}(R_{s}) = P(C_{s}<R_{s}|\gamma_{M} > \gamma_{W}) P(\gamma_{M} > \gamma_{W})$$ $$+ P(C_{s}<R_{s}|\gamma_{M} \leq \gamma_{W}) P(\gamma_{M} \leq \gamma_{W}).$$ From [@IEEEhowto:kopka2] we know that $P(\gamma_{M} > \gamma_{W}) = \frac{\overline{\gamma}_{M}}{\overline{\gamma}_{M}+\overline{\gamma}_{W}}$, $P(C_{s}<R_{s}|\gamma_{M} > \gamma_{W}) = 1-\frac{\overline{\gamma}_{M} + \overline{\gamma}_{W}}{\overline{\gamma}_{M} + 2^{R_{s}}\overline{\gamma}_{W}} e^{-\frac{2^{R_{s}}-1}{\overline{\gamma}_{M}}}$, and $P(C_{s}<R_{s}|\gamma_{M} \leq \gamma_{W}) = 1$. Thus, the $\epsilon$-outage secrecy capacity is the value of $R_{s}$ that sets the outage probability to $\epsilon$. The $\epsilon$-outage secrecy capacity $R_{s}$ is obtained by solving the transcendental equation $(1-\epsilon) \big(1+\frac{\overline{\gamma}_{W}}{\overline{\gamma}_{M}} 2^{R_{s}}\big) = e^{-\frac{2^{R_{s}}-1}{\overline{\gamma}_{M}}},$ which can be put in a closed-form in terms of the *Lambert W function*. Thus, the $\epsilon$-outage secrecy capacity is given by $$\label{eqn_3}
R_{s} = \overline{\gamma}_{M} \times \bigg\{\mathcal{W}_{o}\bigg(\frac{e^{\frac{1}{\overline{\gamma}_{M}}+\frac{1}{\overline{\gamma}_{W}}}}{\overline{\gamma}_{W}(1-\epsilon)}\bigg) - \frac{1}{\overline{\gamma}_{W}} \bigg\}^{+},$$ where $\mathcal{W}_{o}(x)$ is the single valued Lambert W function.
State switching scheme (Reconfigurable antennas without CSI)
------------------------------------------------------------
In this scheme, the legitimate transmitter and receiver utilize reconfigurable antennas with $Q_{T}$ and $Q_{R}$ radiation states respectively. We assume large codeword lengths and that both $Q_{T}$ and $Q_{R}$ are comprable to $n$. The legitimate channel has $Q_{T}Q_{R}$ possible independent realizations per codeword, each realization correspond to a certain transmitter-receiver antenna state selection. We switch the antenna states such that the channel realization changes every symbol within a codeword. A codeword of length $n$ artificially experiences $n$ coherence intervals as long as $n < Q_{T}Q_{R}$. Thus, the legitimate channel capacity $C_{M}$ for specific $Q_{T}Q_{R}$ legitimate channel realizations with $Q_{T}Q_{R} > n$ is given by $C_{M} = \frac{1}{n} \sum_{i=1}^{n} \log(1+\gamma_{M}^{n}(i))$, for $n \to \infty$ and invoking the *law of large numbers*, we have $C_{M} = E\{\log(1+\gamma_{M})\}$ where $\gamma_{M}$ is an exponential random variable with an average of $\overline{\gamma}_{M}$. By averaging the legitimate channel capacity over the exponential pdf, the legitimate channel can be defined by an *ergodic capacity* as $C_{M} = e^{\frac{1}{\overline{\gamma}_{M}}} \operatorname{Ei}\bigg(\frac{1}{\overline{\gamma}_{M}}\bigg)$ [@IEEEhowto:kopka1], where $\operatorname{Ei}(x) = - \int_{-x}^{\infty}\frac{e^{-t}}{t} dt$ is the exponential integral function. Similarly, the eavesdropper channel has $Q_{T}$ possible channel realizations, and the eavesdropper channel capacity can be written as $C_{W} = e^{\frac{1}{\overline{\gamma}_{W}}} \operatorname{Ei}\bigg(\frac{1}{\overline{\gamma}_{W}}\bigg)$. Therefore, recalling (\[eqn\_2\]), the ergodic secrecy capacity is given by $$\label{eqn_5}
C_{s} = \Bigg\{e^{\frac{1}{\overline{\gamma}_{M}}} \operatorname{Ei}\bigg(\frac{1}{\overline{\gamma}_{M}}\bigg) - e^{\frac{1}{\overline{\gamma}_{W}}} \operatorname{Ei}\bigg(\frac{1}{\overline{\gamma}_{W}}\bigg)\Bigg\}^{+}.$$
![Outage capacity of the conventional scheme versus the ergodic capacity of the state switching scheme.[]{data-label="fig_sim"}](Figsec1.eps){width="5"}
Note that the ergodic definition for the secrecy capacity in (\[eqn\_5\]) describes two artificial fast fading legitimate and eavesdropper channels. Although the wireless channel is quasi-static, reconfigurable antennas can be used to induce channel fluctuations over time by switching the radiation states to emulate fast fading. We are interested in studying whether the ergodic capacity in (\[eqn\_5\]) can be larger than the $\epsilon$-outage secrecy capacity. Figure 2 depicts the outage secrecy capacity (solid lines) plotted versus $\epsilon$ for $\overline{\gamma}_{M}$ = 10 dB together with the ergodic capacity (dashed lines) for different eavesdropper channel average SNR values. Note that the ergodic capacity exceeds the outage capacity for tight outage constraints (i.e. small values of $\epsilon$). For instance, when $\overline{\gamma}_{W}$ = -10 dB, the state switching scheme outperforms the conventional scheme as long as $\epsilon < 0.25$. Moreover, for $\overline{\gamma}_{W}$ = 5 dB, the state switching scheme is better for $\epsilon < 0.4$. Besides, the outage capacity does not exist for $\epsilon < 0.25$, thus the state switching scheme is definitely beneficial for strict outage constraints. On the other hand, when $\overline{\gamma}_{M} \leq \overline{\gamma}_{W}$, the ergodic capacity does not exist and only outage capacity with relaxed outage probability constraints is realizable.
State selection with partial CSI
--------------------------------
In this scheme, the legitimate transmitter has the legitimate channel CSI but is not provided with the eavesdropper channel CSI. Thus, the transmitter and the receiver can agree on the adopted radiation states at both parties once per codeword. The secrecy capacity is given by $$C_{s} = \sup\limits_{1 \leq j \leq Q_{T}Q_{R}} \Bigg\{\log(1+\gamma_{M, j}) - \log(1+\gamma_{W})\Bigg\}^{+},$$ where $\gamma_{M, j}$, with $j \in \{1, 2, \ldots, Q_{R}Q_{T}\}$, is one of the $Q_{R}Q_{T}$ independent identical channel realizations obtained by different combinations of the transmission and reception radiation states [@IEEEhowto:kopka5], and $\gamma_{W}$ is a Rayleigh random variable and represents the corresponding eavesdropper channel obtained from a certain selection of the radiation states. Intuitively, the secrecy capacity for a certain set of $Q_{R}Q_{T}$ channel realizations is given by $$\label{eqn_7}
C_{s} = \Bigg\{\log(1+\gamma_{M, max}) - \log(1+\gamma_{W})\Bigg\}^{+},$$ where $\gamma_{M, max} = \max\{\gamma_{M, 1}, \gamma_{M, 2}, \ldots, \gamma_{M, Q_{R}Q_{T}}\}$. The pdf of $\gamma_{M, max}$ is [@IEEEhowto:kopka7] $$f_{\gamma}(\gamma_{M, max}) = Q_{T}Q_{R} \sum_{i=0}^{Q_{T}Q_{R}} \binom{Q_{T}Q_{R}-1}{i} \frac{(-1)^i}{\overline{\gamma}_{M}} e^{-\frac{\gamma_{M, max}}{\overline{\gamma}_{M}/(i+1)}},$$ thus, it can be easily shown that $P(\gamma_{M, max} > \gamma_{W}) = Q_{T}Q_{R} \sum_{i=0}^{Q_{T}Q_{R}} \binom{Q_{T}Q_{R}-1}{i} \frac{(-1)^i}{i+1} \frac{1}{1+\frac{(i+1) \overline{\gamma}_{W}}{\overline{\gamma}_{M}}}$, whereas $P(C_{s}<R_{s} | \gamma_{M, max}>\gamma_{W})$ is given in (\[eqn\_dbl\_x1\]). As demonstrated before, $P_{out}(R_{s}) = P(C_{s}<R_{s}|\gamma_{M, max} > \gamma_{W}) P(\gamma_{M, max} > \gamma_{W})+ P(C_{s}<R_{s}|\gamma_{M, max} \leq \gamma_{W}) P(\gamma_{M, max} \leq \gamma_{W})$ and the $\epsilon$-outage secrecy capacity is obtained by solving the transcendental equation in (\[eqn\_dbl\_x2\]) for $R_{s}$.
$$\label{eqn_dbl_x1}
P(C_{s}<R_{s} | \gamma_{M, max}>\gamma_{W}) = Q_{T}Q_{R} \sum_{i=0}^{Q_{T}Q_{R}} \binom{Q_{T}Q_{R}-1}{i} \frac{(-1)^i}{(i+1)} \Bigg( \frac{1}{1+\frac{(i+1)\overline{\gamma}_{W}}{\overline{\gamma}_{M}}} - \frac{e^{\frac{-(2^{R_{s}}-1)(i+1)}{\overline{\gamma}_{M}}}}{1+\frac{(i+1) 2^{R_{s}}\overline{\gamma}_{W}}{\overline{\gamma}_{M}}}\Bigg).$$
$$\frac{1-\epsilon}{\Bigg(Q_{T}Q_{R} \sum_{i=0}^{Q_{T}Q_{R}} \binom{Q_{T}Q_{R}-1}{i} \frac{(-1)^i}{i+1} \frac{1}{1+\frac{(i+1) \overline{\gamma}_{W}}{\overline{\gamma}_{M}}}\Bigg)}=$$ $$\label{eqn_dbl_x2}
\Bigg(1- Q_{T}Q_{R} \sum_{i=0}^{Q_{T}Q_{R}} \binom{Q_{T}Q_{R}-1}{i} \frac{(-1)^i}{(i+1)} \Bigg( \frac{1}{1+\frac{(i+1)\overline{\gamma}_{W}}{\overline{\gamma}_{M}}} - \frac{e^{\frac{-(2^{R_{s}}-1)(i+1)}{\overline{\gamma}_{M}}}}{1+\frac{(i+1) 2^{R_{s}}\overline{\gamma}_{W}}{\overline{\gamma}_{M}}}\Bigg)\Bigg).$$
State selection with full CSI
-----------------------------
Assume that both the legitimate and the eavesdropper channel CSI are available at the legitimate parties. In this case, state selection will be applied such that the legitimate channel is maximized while the eavesdropper channel is minimized. Because the legitimate channel depends on the selection of one of $Q_{T}$ transmitter radiation states, and one of $Q_{R}$ receiver radiation states, we have a total of $Q_{T}Q_{R}$ possible independent channel realizations. On the other hand, the eavesdropper channel depends only on the transmitter radiation state and thus has one of $Q_{T}$ possible channel realizations. Let the legitimate channel be denoted by $\gamma_{M}^{i,j}$ where $i \in \{1,\ldots,Q_{T}\}$ and denotes the selected transmitter radiation state, whereas $j \in \{1,\ldots,Q_{R}\}$ and denotes the selected receiver state. Similarly, the eavesdropper channel is $\gamma_{W}^{i}$ where $i \in \{1,\ldots,Q_{T}\}$, thus we note that the selection of a transmitter radiation state dictates an eavesdropper channel and a set of possible $Q_{R}$ legitimate channels, from where a single realization is picked based on the receiver state. The achievable secrecy capacity for a certain set of legitimate and eavesdropper channel realizations corresponds to the supremum of all selections for transmitter and receiver radiation states $$\label{eqn_8}
C_{s} = \sup\limits_{1 \leq i \leq Q_{T}, 1 \leq j \leq Q_{R}} \Bigg\{\log(1+\gamma_{M}^{i,j}) - \log(1+\gamma_{W}^{i})\Bigg\}^{+}.$$ Equation (\[eqn\_8\]) suggests that we do not only improve the legitimate channel, but also use the CSI to select the radiation state that undermines the eavesdropper channel. Numerical results for the $\epsilon$-outage secrecy capacity $R_{s}$ are obtained in section IV.
Numerical results
=================
In figure 3, we investigate the achievable $\epsilon$-outage secrecy capacity for different schemes. By setting $\epsilon$ = 0.1, we plot the secrecy capacity versus $\overline{\gamma}_{M}$ for low and high values of $\overline{\gamma}_{W}$ (-10 and 20 dB respectively). Note that adopting single antennas (conventional scheme) causes an SNR loss of around 10 dB for $\overline{\gamma}_{W}$ = 20 and -10 dB compared to the AWGN secrecy capacity. Reconfigurable antennas provide considerable SNR gains when the CSI is available. Without CSI, the state switching scheme provides poor ergodic capacity for $\overline{\gamma}_{W}$ = 20 dB. For $\overline{\gamma}_{W}$ = -10 dB, the ergodic capacity of the state switching scheme outperforms the outage capacity of the conventional single antenna system for $\overline{\gamma}_{M} <$ 25 dB. In this case, an SNR gain of about 5 dB is achieved. Thus, the state switching scheme offers an SNR gain only for low values of $\overline{\gamma}_{M}$. On the other hand, the state selection with partial CSI scheme offers a significant gain for all legitimate and eavesdropper SNR ranges. The number of radiation states involved in calculations are $Q_{T}$ = $Q_{R}$ =5. For $\overline{\gamma}_{W}$ = -10 dB, partial CSI offer 5 dB SNR gain compared to the AWGN capacity and 15 dB compared to the single antenna system in Rayleigh fading. For high eavesdropper average SNR ($\overline{\gamma}_{W}$ = 20 dB), an SNR gain of 2 dB compared to the AWGN capacity and 12 dB compared to the conventional scheme. It is worth mentioning that the achievable gain is higher for lower values of $\overline{\gamma}_{W}$ as this scheme is not provided with the eavesdropper channel CSI. Moreover, the state selection scheme with full CSI provide superior secrecy capacity compared to all other schemes. This gain is most notable for large $\overline{\gamma}_{W}$, i.e. for $\overline{\gamma}_{W}$ = 20 dB, where and SNR gain of 20 dB compared to the AWGN capacity and 30 dB compared to the single antenna fading channel capacity. The reason for such impressive performance boost is that knowledge of the eavesdropper CSI and the selection of the “worst" eavesdropper channel is most effective when the eavesdropper channel enjoys high SNR. The gain achieved for $\overline{\gamma}_{W}$ = -10 dB is about 8 dB compared to the AWGN channel. The gain decreases for low values of $\overline{\gamma}_{W}$ because undermining the eavesdropper channel becomes of less effectivness.
Figure 4 demonstrates the secrecy capacity normalized to the AWGN secrecy capacity for $\epsilon$ = 0.1 and $\overline{\gamma}_{W}$ = -10, 0 and 10 dB. Focusing on the state selection scheme with full CSI, we note that the capacity gain is maximum at low SNR. This is similar to the effect of optimal water filling power allocation over a fast fading channel, where the highest capacity gain is obtained at low SNR. In our case power allocation is applied across the radiation state domain rather than the time domain. Besides, instead of water filling, we allocate all power to the best radiation state. Thus, we are able to achieve considerable capacity gains regardless of the time latency of the applications, i.e. delay tolerant and delay sensitive applications are both supported by the reconfigurable antenna scheme.
![Comparison between all secret communications schemes.[]{data-label="fig_sim"}](Figsec5.eps){width="5in"}
![Secrecy capacity gain for different schemes.[]{data-label="fig_sim"}](Figsec2.eps){width="5in"}
[1]{} P. K. Gopala, L. Lai and H. El Gamal, “On the Secrecy Capacity of Fading Channels," *IEEE Trans. Inf. Theory*, vol. 54, no. 10, pp. 4687-4697, Oct. 2008. J. Barros and M. R. D. Rodrigues,“Secrecy capacity of wireless channels," *in Proc. IEEE Int. Symp. Information Theory*, Seattle, WA, pp. 356?60, Jul. 2006. A. Khandani, G. Bagherikaram, and A. Motahary, “The Secrecy Capacity Region of the Gaussian MIMO Broadcast Channel," *IEEE Trans. Inf. Theory*, vol. 59, no. 5, pp. 2673-2682, Oct. 2013. A. Khisti and G. W. Wornell, “Secure Transmission With Multiple Antennas?art II: The MIMOME Wiretap Channel," *IEEE Trans. Inf. Theory*, vol. 56 , no. 11, pp. 5515-5532, Oct. 2010. C. Y. Wu, P. C. Lan, P. C. Yeh, C. H. Lee and C. M. Cheng, “Practical Physical Layer Security Schemes for MIMO-OFDM Systems Using Precoding Matrix Indices," *IEEE J. Selected Areas Commun*, vol. 31, no. 9, pp. 1687-1700, Sept. 2013. Yaxing Cai and Zhengwei Du,“A Novel Pattern Reconfigurable Antenna Array for Diversity Systems," *IEEE Antennas and Wireless Propagation Letters,*, vol. 8, pp. 1227-1230, 2009. P. A. Martin, P. J. Smith, and R. Murch, “Improving Space-Time Code Performance in Slow Fading Channels using Reconfigurable Antennas," *IEEE Communications Letters*, vol. 16, pp. 494-497, Apr. 2012. P. Mookiah and K. R. Dandekar, “A Reconfigurable Antenna-Based Solution for Stationary Device Authentication in Wireless Networks," *International Journal of Antennas and Propagation*, vol. 2012, Article ID 545783, 11 pages, 2012. R. Mehmood, and J. W. Wallace, “Wireless security enhancement using parasitic reconfigurable aperture antennas," *Proceedings of the 5th European Conference on Antennas and Propagation (EUCAP)*, Rome, pp. 2761-2765, April 2011. Ning Kong, “Performance Comparison among Conventional Selection Combining, Optimum Selection Combining and Maximal Ratio Combining," *Proceedings of IEEE International Conference on Communications (ICC’09), Dresden, Germany*, pp. 1-6, June 2009.
[^1]: The authors are with XX
[^2]: Manuscript received XXXX XX, 2013; revised XXXX XX, 201X.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we show that two constructions form stacks: Firstly, as one varies the $\infty$-topos, ${\mathcal}{X}$, Lurie’s homotopy theory of higher categories internal to ${\mathcal}{X}$ varies in such a way as to form a stack over the $\infty$-category of all $\infty$-topoi.
Secondly, we show that Haugseng’s construction of the higher category of iterated spans in a given $\infty$-topos (equipped with local systems) can be used to define various stacks over that $\infty$-topos.
As a prerequisite to these results, we discuss properties which limits of $\infty$-categories inherit from the $\infty$-categories comprising the diagram. For example, Riehl and Verity have shown that possessing (co)limits of a given shape is hereditary. Extending their result somewhat, we show that possessing Kan extensions of a given type is heriditary, and more generally that the adjointability of a functor is heriditary.
author:
- 'David Li-Bland'
bibliography:
- 'basicbib.bib'
title: 'The stack of higher internal categories and stacks of iterated spans.'
---
Introduction
============
Building upon the ideas of Rezk [@Rezk:2001ey] and Barwick [@Barwick:2005wl; @Barwick:2011wl], Lurie constructed a model for the homotopy theory of higher categories internal to an $\infty$-topos ${\mathcal}{X}$ [@Lurie:2009uz]. More precisely, he constructs an $\infty$-category ${CSS}_k({\mathcal}{X})$ of complete $k$-fold Segal objects in an arbitrary $\infty$-topos ${\mathcal}{X}$. Our first result (cf. Theorem \[thm:CSSaSheaf\]) in this paper is to show that the construction $$\label{eq:CSSfunctor}{\mathcal}{X}\mapsto {CSS}_k({\mathcal}{X})$$ satisfies a certain descent condition: suppose that ${\mathcal}{X}_i\to {\mathcal}{X}$ is an étale cover of ${\mathcal}{X}$ indexed by a small simplicial set $i\in I$, i.e. $${\mathcal}{X}\cong {{{\operatorname}{co}\lim}}_{i\in I}{\mathcal}{X}_i,$$ then $${CSS}_k({\mathcal}{X})\cong \lim_{i\in I}{CSS}_k({\mathcal}{X}_i).$$ In other words, defines a stack (cf. [@Lurie:2009un Notation 6.3.5.19]). Given an $\infty$-category ${\mathcal}{C}$, Barwick [@Barwick:2013th] showed how to construct an $(\infty,1)$-category ${\operatorname}{Span}({\mathcal}{C})$ which has the same space of objects as ${\mathcal}{C}$, but whose morphisms between two objects $c_0,c_1\in {\mathcal}{C}$ is the space of diagrams in ${\mathcal}{C}$ of the form $$\begin{tikzpicture}
\node (m-1-1) at (0,1) {$x$};
\node (m-1-2) at (1,0) {$c_0$};
\node (m-2-1) at (-1,0) {$c_1$};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\end{tikzpicture}$$ That is, spans $c_0\nrightarrow c_1$ in ${\mathcal}{C}$. Composition of two such morphisms is given by taking the fibred product. Given an $\infty$-topos ${\mathcal}{X}$ and a complete $k$-fold Segal object $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$, Haugseng extends this construction in [@Haugseng:2014vw] to produce an $(\infty,k)$-category ${\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})\in{CSS}_k(\hat{\mathscr{S}})$ of iterated $k$-fold spans with local systems valued in $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$ (here $\hat{\mathscr{S}}$ is the $\infty$-category of (not necessary small) spaces).
Our second main result is to show that the assignment of ${\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})$ to $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$ depends continuously on ${\mathcal}{X}$ and $X_{\bullet,\dots,\bullet}$ (i.e. it preserves small limits). This result is somewhat more subtle than it first appears: for example, the functor $${CSS}_k({\mathcal}{X})\xrightarrow{X_{\bullet,\dots,\bullet}\mapsto {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$ is not continuous - it preserves neither products nor the terminal object. To correctly understand the continuity of ${\operatorname}{Span}_k$, we need to work in a larger context: we assemble all the $\infty$-categories ${CSS}_k({\mathcal}{X})$ into one large $\infty$-category $\int{CSS}_k$, whose (roughly speaking)[^1]
- objects are pairs $({\mathcal}{X},X_{\bullet,\dots,\bullet})$, where ${\mathcal}{X}$ is an arbitrary $\infty$-topos and $X_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{X})$ is a higher category internal to ${\mathcal}{X}$, and
- morphisms $({\mathcal}{X},X_{\bullet,\dots,\bullet})\to ({\mathcal}{Y},Y_{\bullet,\dots,\bullet})$ consist of a geometric morphism of $\infty$-topoi, $(f^*\dashv f_*):{\mathcal}{X}\leftrightarrows{\mathcal}{Y}$, together with a morphism $X_{\bullet,\dots,\bullet}\to (f_*)_!Y_{\bullet,\dots,\bullet}$ in ${CSS}_k({\mathcal}{X})$.
In Theorem \[thm:SpanCont\], we prove that the functor $$\int{CSS}_k\xrightarrow{({\mathcal}{X},X_{\bullet,\dots,\bullet})\mapsto {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$ is continuous (preserves small limits).
Consequently, suppose that for every $U\in {\mathcal}{X}$ we assign (in a natural way) a complete Segal object $\sigma(U)\in {CSS}_k({\mathcal}{X}_{/U})$, in a manner which depends locally on $U\in{\mathcal}{X}$: that is, for any colimit diagram $U_i\to U$ in ${\mathcal}{X}$ indexed by a small simplicial set $i\in I$, $$\sigma(U)=\lim_{i\in I}\sigma(U_i)$$ (where the latter limit is taken in $\int{CSS}_k$), then
\[eq:IntroSpanStack\] $${\mathcal}{X}^{op}\xrightarrow{U\mapsto {\operatorname}{Span}_k({\mathcal}{X}_{/U},\sigma(U))}{CSS}_k(\hat{\mathscr{S}})$$ defines an $(\infty,k)$-stack over ${\mathcal}{X}$ (cf. Theorem \[thm:StackofSpans\]). As a first example, ${\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})$ itself forms a stack $${\mathcal}{X}^{op}\xrightarrow{U\mapsto {\operatorname}{Span}_k({\mathcal}{X}_{/U},U\times X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$
over ${\mathcal}{X}$. As a second example, taking ${\mathcal}{X}={\operatorname}{dSt}_{{\mathbb}{K}}$ to be derived stacks over a field ${\mathbb}{K}$ of characteristic zero, and $X_{\bullet,\dots,\bullet}$ to be trivial, the fact that forms a stack implies that the *derived* composition of spans depends continuously (algebraically, in fact) on the spans involved.
Our motivation for these results comes from mathematical physics: the success of the *Lagrangian Creed*:
*“everything is a Lagrangian correspondence”[^2],*
places Lagrangian correspondences between symplectic manifolds at the centre of classical mechanics. Lagrangian correspondences have two major flaws however: **firstly** they fail to compose in general, i.e. given two Lagrangian correspondences $$U\overset{L}{\nleftarrow} V\text{ and }V\overset{L'}{\nleftarrow} W,$$ their set theoretic composite $$\label{eq:LagComp}U\overset{L\circ L'}{\nleftarrow} W,$$ often fails to be smooth, and - **secondly** - when the composite exists as a Lagrangian correspondence, it may not depend continuously on $L$ and $L'$.
The first of these issues was essentially resolved by Pantev, Toën, Vaquié, and Vessozi [@Pantev:2011uz], and Calaque [@Calaque:2013un] using derived geometry. Building upon this, Haugseng [@Haugseng:2014vw] then gave an embedding of Weinstein’s symplectic ‘category’ [@Weinstein82] whose morphisms are the Lagrangian correspondences, as a subcategory of ${\operatorname}{Span}_1({\operatorname}{dSt}_{{\mathbb}{K}}, {\mathcal}{A}^2_{{\operatorname}{cl}})$, spans of derived stacks with local systems valued in closed 2-forms. The fact that is a stack is a first step towards a deeper understanding of what it means to restore the continuity of composition using derived geometry.
Moreover, the second issue - the failure of composition to be continuous - is closely related to the failure to quantize classical mechanics functorially: After quantizing pairs where the composite fails to depend continuously on $L$ and $L'$, one is typically trying to multiply Dirac $\delta$-functions in the corresponding quantization. In work in progress with Gwilliam, Haugseng, Johnson-Freyd, Scheimbauer, and Weinstein, we show that *at least to first order* (i.e. after linearizing),
- the *derived* composition of Lagrangian correspondences depends continuously on $L$ and $L'$ (cf. [@LiBland:2014tf]), and
- there is a functorial quantization.
In order to show that and define stacks, we first need to examine limits of $\infty$-categories. Suppose that ${\mathcal}{C}_k,\quad k\in K$ is some diagram of $\infty$-categories indexed by a simplicial set $K$. In [@Riehl:2014ut], Riehl and Verity show that if each ${\mathcal}{C}_k$ has all (co)limits of shape $I$ (where $I$ is some small simplicial set), and for each arrow $k\to k'$ in $K$, the corresponding functor ${\mathcal}{C}_k\to {\mathcal}{C}_{k'}$ preserves all (co)limits of shape $I$, then the limit $\infty$-category $\lim_{k\in K}{\mathcal}{C}_k$ also has all (co)limits of shape $I$. After providing an alternate proof of this result (cf. Theorem \[thm:LimInLim\]), we extend their result to show that if each ${\mathcal}{C}_k$ possesses all Kan extensions along a functor ${\mathcal}{I}\to{\mathcal}{I}'$, and each functor ${\mathcal}{C}_k\to {\mathcal}{C}_{k'}$ preserves those Kan extensions, then the limit $\infty$-category $\lim_{k\in K}{\mathcal}{C}_k$ also has all Kan extensions along ${\mathcal}{I}\to{\mathcal}{I}'$ (cf. Corollary \[cor:KanClosed\]). More generally, suppose that $$F_k:{\mathcal}{C}_k \rightleftarrows{\mathcal}{D}_k:G_k, \quad k\in K$$ is diagram of adjunctions coherently indexed by a small simplical set $K$, then we prove there is an adjunction $$\lim_{k\in K}F_k:\lim_{k\in K}{\mathcal}{C}_k \rightleftarrows\lim_{k\in K}{\mathcal}{D}_k:\lim_{k\in K}G_k$$ between the corresponding limit $\infty$-categories (cf. Theorem \[thm:limadj\]).
Acknowledgements
----------------
We would like to thank Owen Gwilliam, Theo Johnson-Freyd, Claudia Scheimbauer, and Alan Weinstein for many important discussions surrounding the content of this paper. We would also like to thank Thomas Nikolaus for a very helpful introduction to descent theory, and Omar Antol[í]{}n Camarena for a number of helpful conversations about higher category theory. The author was supported by an NSF Postdoctoral Fellowship DMS-1204779.
Notation
--------
We generally use the notation and terminology developed by Lurie (cf. [@Lurie:2009un]). In particular, by an *$\infty$-category*, we mean a quasicategory, i.e. a simplicial set satisfying certain horn filling conditions. In addition, we use the following notation, some of which differs from Lurie’s:
- ${\mathbb{\Delta}}$ denotes the simplicial indexing category whose objects are non empty fnite totally ordered sets $[n] := \{0, 1, . . . , n\}$ and morphisms are order-preserving functions between them. $\Delta^n:{\mathbb{\Delta}}^{op}\to {\operatorname}{Sets}$ is the simplicial set represented by $[n]$.
- We denote generic $\infty$-categories by upper-case caligraphic letters, ${\mathcal}{A},{\mathcal}{B},{\mathcal}{C},{\mathcal}{D},\dots$. We typically denote elements $c\in {\mathcal}{C}$ of a generic $\infty$-category by lowercase versions of the same letter.
- We let ${\operatorname}{Fun}({\mathcal}{C},{\mathcal}{D})$ denote the $\infty$-category of functors between $\infty$-categories, and ${\operatorname}{Map}_{{\mathcal}{C}}(c,c')$ denote the mapping space between two objects $c,c'\in {\mathcal}{C}$.
- If ${\mathcal}{C}$ is an $\infty$-category, we write $\iota{\mathcal}{C}$ for the *interior* or *classifying space of objects* of ${\mathcal}{C}$, i.e. the maximal Kan complex contained in ${\mathcal}{C}$.
- If $f : {\mathcal}{C} \to {\mathcal}{D}$ is left adjoint to a functor $g : {\mathcal}{D} \to{\mathcal}{C}$, we will refer to the adjunction as $f \dashv g$.
- ${{\mathscr{C}\mathrm{at}}}_\infty$ denotes the $\infty$-category of small $\infty$-categories, and the $\infty$-category of spaces, ${\mathscr{S}}\subset {{\mathscr{C}\mathrm{at}}}_\infty$, is the full subcategory spanned by the Kan complexes.
- If ${\mathcal}{C}$ is an $\infty$-category, we let $${\!\text{\usefont{U}{min}{m}{n}\symbol{'210}}\!}:{\mathcal}{C}\to {\mathscr{P}}({\mathcal}{C}):={\operatorname}{Fun}({\mathcal}{C},{\mathscr{S}})$$ denote the Yoneda embedding[^3].
- Suppose that $p_0:X_0\to K$ and $p_1:X_1\to K$ are two morphisms of simplicial sets and that $p_1$ is a (co)Cartesian fibration. We let ${\operatorname}{Fun}_{K}(X_0,X_1)$ denote the simplicial subset of all simplicial maps between $X_0$ and $X_1$ spanned by those maps which intertwine $p_0$ and $p_1$. Note that ${\operatorname}{Fun}_{K}(X_0,X_1)$ is automatically an $\infty$-category (cf. [@Lurie:2009un Remark 3.1.3.1]).
When $p_0$ and $p_1$ are both (co)Cartesian fibrations, then we let $${\operatorname}{Fun}^{{\operatorname}{(co)Cart}}_K(X_0,X_1)\subseteq {\operatorname}{Fun}_{K}(X_0,X_1)$$ denote the subcategory spanned by those maps which preserve the (co)Cartesian edges.[^4]
Properties Inherited by Limit $\infty$-Categories
=================================================
Let ${{\mathscr{C}\mathrm{at}}}_\infty$ denote the $\infty$-category of small $\infty$-categories, let $K$ be a small simplicial set, and consider a diagram $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$. We will be interested in the limit $\infty$-category, $\lim p'\in {{\mathscr{C}\mathrm{at}}}_\infty$.
To compute such limits, consider the functor
$$\label{eq:DiagFunct}\Delta:{{\mathscr{C}\mathrm{at}}}_\infty\to {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty),$$
which sends an $\infty$-category ${\mathcal}{C}\in {{\mathscr{C}\mathrm{at}}}_\infty$ to the constant diagram: $$\Delta_{{\mathcal}{C}}:k\to {\mathcal}{C},\quad\text{ for any }k\in K.$$ The right adjoint to is the functor which sends a diagram $p'\in {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty)$ to the corresponding limit $\infty$-category $\lim p'\in {{\mathscr{C}\mathrm{at}}}_\infty$.
Now, fix a second diagram $p_0'\in {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty)$ and consider the functor $$\label{eq:DiagTensorFunctor}\Delta\times p_0':{{\mathscr{C}\mathrm{at}}}_\infty \to {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty),$$
which sends any $\infty$-category ${\mathcal}{C}\in {{\mathscr{C}\mathrm{at}}}_\infty$ to the functor $$\Delta_{\mathcal}{C}\times p_0':k\to {\mathcal}{C}\times p_0'(k),\quad\text{ for any }k\in K.$$ The right adjoint to $\Delta\times p_0'$ sends any diagram $p'\in {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty)$ to the $\infty$-category ${\operatorname}{Nat}_{K}(p_0',p')\in {{\mathscr{C}\mathrm{at}}}_\infty$ of natural transformations between $p_0'$ and $p'$. Since in the special case that $p_0'=\Delta_\ast$ is the constant diagram at the terminal $\infty$-category[^5] both functors and coincide, in particular, for any diagram $p'\in {\operatorname}{Fun}(K,{{\mathscr{C}\mathrm{at}}}_\infty)$, we have an equivalence $$\lim p'\cong {\operatorname}{Nat}_{K}(\Delta_\ast,p')$$ between the limit of $p'$ and the $\infty$-category of natural transformations from the trivial diagram to $p'$.
In practice, often the best description of diagrams in ${{\mathscr{C}\mathrm{at}}}_\infty$ is in terms of (co)Cartesian fibrations, as developed by Lurie [@Lurie:2009un § 2.4]. Briefly, given an inner fibration between $\infty$-categories ${\mathcal}{C}\to {\mathcal}{D}$ a edge $f:c\to c'$ in ${\mathcal}{C}$ is called *coCartesian* if $${\mathcal}{C}_{f/}\to {\mathcal}{C}_{c/}\times_{{\mathcal}{D}_{p(c)/}}{\mathcal}{D}_{p(f)/}$$ is an equivalence.[^6] The fibration ${\mathcal}{C}\to {\mathcal}{D}$ is called *coCartesian* if there is a coCartesian edge over any edge in ${\mathcal}{D}$ starting at any vertex of ${\mathcal}{C}$. The fibration ${\mathcal}{C}\to{\mathcal}{D}$ is called Cartesian if ${\mathcal}{C}^{op}\to {\mathcal}{D}^{op}$ is coCartesian. Functors $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$ correspond to coCartesian fibrations $p:X\to K$ via the straightening/unstraightening construction [@Lurie:2009un]; for every $k\in K$, the fibre $X_k:=p^{-1}(k)$ is equivalent to $p'(k)$, and for every edge $k\to k'$ in $K$, the corresponding functor $X_k\to X_{k'}$ is equivalent to one sending any $x\in X_k$ to the target of a coCartesian edge over $k\to k'$ starting at $x$.
Given a second coCartesian fibration $p_0:X_0\to K$ corresponding to a diagram $p_0':K\to {{\mathscr{C}\mathrm{at}}}_\infty$, Gepner, Haugseng, and Nikolaus identify the $\infty$-category of natural transformations $${\operatorname}{Nat}_{K}(p_0',p')\cong {\operatorname}{Fun}_K^{{\operatorname}{coCart}}(X_0,X)$$ with the $\infty$-category of coCartesian maps $X_0\to X$, i.e. those maps $X_0\to X$ over $K$ which preserve the coCartesian edges (cf. [@Gepner:2015ww Proposition 6.9]). In particular, Lurie shows that an elegant model of $\lim p'$ is the $\infty$-category $$\lim p'\cong {\operatorname}{Nat}_{K}(\Delta_\ast,p')\cong {\operatorname}{Fun}_K^{{\operatorname}{coCart}}(K,X)$$ of coCartesian sections of $p$ [@Lurie:2009un Corollary 3.3.3.2] (note that every edge in the trivial fibration $K\to K$ is coCartesian).
Let ${{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{Str}}$ denote the ordinary category whose objects are (small) $\infty$-categories, and suppose that $C$ is a (small) category. Given a strict functor $p':C\to {{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{Str}}$, the corresponding coCartesian fibration can be computed via the relative nerve construction (cf. [@Lurie:2009un § 3.2.5]). Forming the resulting $\infty$-category of coCartesian sections, one sees that a model for $\lim p'$ is the simplicial set whose $k$ simplices consist of the following data:
- for every functor $(x,y):[n]\to C\times [k]$, a choice of $n$-simplex, $\tau_{(x,y)}:\Delta^n\to p'\big(x(n)\big)$, such that
1. for every $f:[m]\to [n]$, the following diagram commutes: $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{\Delta^m & &p'\big( x(f(m))\big)\\ \Delta^n&&p'\big( x(n)\big)\\};
\draw[->] (m-1-1) -- node {$f$} (m-2-1);
\draw[->] (m-1-1) -- node {$\tau_{f^*(x,y)}$} (m-1-3);
\draw[->] (m-2-1) -- node {$\tau_{(x,y)}$} (m-2-3);
\draw[->] (m-1-3) -- node {$p'\big[x\big(f(m)\to n\big)\big]$} (m-2-3);
\end{tikzpicture}$$ and
2. whenever $y(i)=y(j)$, then $\tau_{(x,y)}(\Delta^{\{i,j\}})\in p'\big(x(n)\big)$ is an equivalence.
(co)Limits in Limit $\infty$-categories
---------------------------------------
Let $I$ be a second (small) simplicial set, and suppose that
1. for each vertex $k\in K$, the $\infty$-category $p'(k)$ admits (co)limits for all diagrams indexed by $I$.
2. for each edge $(k\to k')\in K$, the functor $p'(k)\to p(k')$ preserves (co)limits for all diagrams indexed by $I$.
then Riehl and Verity [@Riehl:2014ut] have shown that the limit $\infty$-category $\lim p' \in {{\mathscr{C}\mathrm{at}}}_\infty$ admits (co)limits for all diagrams indexed by $I$, and that those (co)limits are preserved by the functors in the limit cone. We now provide an alternate proof of this result, based on Lurie’s (co)Cartesian fibrations.
To disambiguate our presentation, we will prove our results only for *colimits* (rather than *limits*) in the limit $\infty$-category $\lim p'$; the duality between colimits in $\lim p'$ and limits in $(\lim p')^{op}$ imply that the corresponding results hold equally for limits as well. We begin with the special case where $I$ is the empty set, in which case we have the following variant of [@Lurie:2009un Proposition 2.4.4.9]:
\[lem:LimOfInit\] Given a functor $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$, if
1. for each $k\in K$, the $\infty$-category $p'(k)$ admits an initial object $t\in X_k$, and
2. for each edge $k\to k'$ in $K$ the functor $p'(k)\to p'(k')$ preserves initial objects,
then
A
: the limit $\infty$-category $\lim p'$ admits an initial object $t_\infty$, and
B
: an object $t\in \lim p'$ is initial if and only if for each $k\in K$, the object $\pi_k(t)\in X_k$ is initial where $\pi_k:\lim p'\to p'(k)$ is the functor appearing in the limit cone.
We will find it easier to model our functor $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$ in terms of a *Cartesian fibration* $p:X\to K^{op}$ (rather than a *coCartesian fibration*).
Let $p:X\to K^{op}$ be a Cartesian fibration of simplicial sets classified by the functor $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$. By assumption
1. for each $k\in K$, the $\infty$-category $X_k\cong p'(k)$ admits an initial object $t\in X_k$, and
2. for each $p$-Cartesian edge $f:t'\to t$ over $p(f):k'\to k$ the object $t'\in X_{k'}$ is initial whenever $t\in X_k$ is,
Let $X'\subseteq X$ be the simplical subset spanned by those vertices $t\in X$ which are initial objects of $X_{p(t)}\cong p'\circ p(t)$. Then (as we shall show), every edge $f:t'\to t$ in $X'$ is $p$-Cartesian (when seen as an edge of $X$). To see this suppose that $f:t'\to t$ is such an edge. Let $f':t''\to t$ be a $p$-Cartesian edge in $X$ over $p(f)$; then (cf. [@Lurie:2009un Remark 2.4.1.4.]) there exists a 2-simplex $\sigma:\Delta^2\to X$ such that $$\sigma(\Delta^{\{1,2\}})=f',\quad \sigma(\Delta^{\{0,2\}})=f,\quad\text{and}\quad p\big(\sigma(\Delta^{\{0,1\}})\big)=p(s_0(t')),$$ where $s_0:K_0\to K_1$ is the degeneracy map.
By assumption $t'',t'\in X_{p(s)}$ are both initial, and hence $\sigma(\Delta^{\{0,1\}})\in X_{p(t')}$ is an equivalence. In particular $\sigma(\Delta^{\{0,1\}})\in X$ is a $p$-Cartesian morphism. It follows from [@Lurie:2009un Proposition 2.4.1.7.] that $f=\sigma(\Delta^{\{0,2\}})$ is $p$-Cartesian.
Now, by [@Lurie:2009un Proposition 2.4.4.9.], there exists a section $t_\infty:K^{op}\to X'$; and by the previous discussion, $t_\infty$ is a Cartesian section.
Now $\lim p'\cong {\operatorname}{Fun}_{K^{op}}^{{\operatorname}{Cart}}({K^{op}},X)$ is the $\infty$-category of Cartesian sections of $p$ (cf. [@Lurie:2009un Corollary 3.3.3.2]). Thus, we can identify $t_\infty$ with an element of $\lim p'$. We now claim that $t_\infty\in \lim p'$ is an initial object: Notice that $\lim p'$ is the full subcategory of ${\operatorname}{Fun}_{K^{op}}(K^{op},X)$ spanned by the Cartesian sections. Suppose we have a diagram $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{\partial \Delta^n & {\operatorname}{Fun}_{K^{op}}^{{\operatorname}{Cart}}({K^{op}},X) & {\operatorname}{Fun}_{K^{op}}({K^{op}},X)\\ \Delta^n & &\\};
\draw[->] (m-1-1) -- node {$f$} (m-1-2);
\draw[right hook->] (m-1-2) -- (m-1-3);
\draw[right hook->] (m-1-1) -- (m-2-1);
\draw[->,dashed] (m-2-1) -- node[swap] {$\tilde f$} (m-1-3);
\draw[->,dashed] (m-2-1) -- node {$\tilde f'$} (m-1-2);
\end{tikzpicture}$$ such that $f\rvert_{\{0\}}=t_\infty$. Then by [@Lurie:2009un Proposition 2.4.4.9.], the arrow $\tilde f$ exists (making the diagram commute), but since ${\operatorname}{Fun}_{K^{op}}^{{\operatorname}{Cart}}({K^{op}},X)$ is the full subcategory of ${\operatorname}{Fun}_{K^{op}}({K^{op}},X)$ spanned by the Cartesian sections, and all the vertices of $\tilde f$ lie in the image of $f$, and hence in ${\operatorname}{Fun}_{K^{op}}^{{\operatorname}{Cart}}({K^{op}},X)$, it follows that $\tilde f$ factors through a map $\tilde f':\Delta^n\to {\operatorname}{Fun}_{K^{op}}^{{\operatorname}{Cart}}({K^{op}},X)$.
Thus $t_\infty\in \lim p'$ is an initial object.
By construction, for every $k\in K$, $\pi_k(t_\infty)\in p'(k)$ is the image of the initial object $t_\infty(k)\in X_k$ under the equivalence of $\infty$-categories $X_k\xrightarrow{\cong} p'(k)$. Thus (2) follows from the uniqueness of initial objects.
Now, we can interpret arbitrary colimits in terms of initial objects using the concept of an undercategory, as follows: Suppose that ${\mathcal}{C}\in {{\mathscr{C}\mathrm{at}}}_\infty$ is an $\infty$-category and $q:I\to {\mathcal}{C}$ is a diagram of shape $I$ (where $I$ is a small simplicial set). Then a colimit diagram for $q$ is equivalent to an initial object of the undercategory, ${\mathcal}{C}^{q/}$.
Suppose that $K$ is a (small) simplicial set and $\bar p':K^{\lhd}\to {{\mathscr{C}\mathrm{at}}}_\infty$ is a diagram with cone point ${\mathcal}{C}\in {{\mathscr{C}\mathrm{at}}}_\infty$. For any vertex $k\in K^\lhd$, let $\pi_k:{\mathcal}{C}\to p'(k)$ denote the corresponding functor in the cone. Given any diagram $q:I\to {\mathcal}{C}$, since the formation of undercategories is natural, there exists a diagram $(\bar p')^{q/}:K^\lhd\to {{\mathscr{C}\mathrm{at}}}_\infty$ indexing the undercategories: $$(\bar p')^{q/}(k)\cong \big(\bar p'(k)\big)^{\pi_k\circ q/},$$ for all $k\in K^{\lhd}$, along with a natural transformation $(\bar p')^{q/}\to \bar p'$ which restricts at every $k\in K^\lhd$ to the canonical functors $\big(\bar p'(k)\big)^{\pi_k\circ q/}\to \bar p'(k)$.
\[lem:LimComUnder\] If $\bar p':K^{\lhd}\to {{\mathscr{C}\mathrm{at}}}_\infty$ is a limit diagram, then so is $(\bar p')^{q/}:K^\lhd\to {{\mathscr{C}\mathrm{at}}}_\infty$.
Let $\bar p:\bar X \to K^\lhd$ be a coCartesian fibration classified by $\bar p'$; and let $\ast\in (K^{op})^\lhd$ denote the cone point. Since $\{\ast\}^\sharp\subseteq \big(K^\lhd\big)^\sharp$ is marked anodyne, we have a natural equivalence of the $\infty$-category of coCartesian sections: $${\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)\xrightarrow{\cong}{\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(\{\ast\},\bar X)\cong \bar X\times_{K^\lhd}\{\ast\}\cong {\mathcal}{C}.$$ Therefore, we may lift $q:I\to {\mathcal}{C}$ to a diagram $$\tilde q:I\to {\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)$$ in a homotopically unique way. Choose a factorization $I \to I' \to {\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)$ of $\tilde q$, where $I \to I'$ is inner anodyne (and therefore a categorical equivalence) and $I' \to {\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)$ is an inner fibration (so that $I'$ is an $\infty$-category). The map $I \to I'$ is a categorical equivalence, and therefore cofinal. We are free to replace $I$ by $I'$, and may thereby assume that $I$ is an $\infty$-category.
Given two morphisms of simplicial sets $Y\to K^\lhd$, and $Z\to K^\lhd$, recall that $Z\diamond_{K^\lhd} Y$ denotes the relative (alternate) join of the simplicial sets $Z$ and $Y$, $$Z\diamond_{K^\lhd} Y:= Z\coprod_{Z\times_{K^\lhd} Y\times \{0\}}(Z\times_{K^\lhd} Y\times \Delta^1)\coprod_{Z\times_{K^\lhd} Y\times \{1\}}Y$$ (cf. [@Lurie:2009un § 4.2.2] and [@Joyal:2008wr]).
Let $q_{K^\lhd}:I\times K^\lhd\to \bar X$ denote the composite $$I\times K^\lhd\xrightarrow{q\times {\operatorname}{Id}_{K^\lhd}}{\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)\times K^\lhd\xrightarrow{{\operatorname}{ev}} \bar X.$$ As in [@Lurie:2009un § 4.2.2] we define $\bar X^{q_{K^\lhd}/}\to K^\lhd$ to be the simplicial set satisfying the universal property that for any morphism of simplicial sets $Y\to K^\lhd$, commutative diagrams of the form $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{I\times K^\lhd &\bar X\\(I\times K^\lhd)\diamond_{K^\lhd} Y& K^\lhd\\};
\draw[->] (m-1-1) -- node {$q_{K^\lhd}$} (m-1-2);
\draw[right hook->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- node {$\bar p$} (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\draw[->] (m-2-1) -- (m-1-2);
\end{tikzpicture},$$ correspond to diagrams of the form $$\begin{tikzpicture}\node (y) at (-2,1) {$Y$};
\node (q) at (2,1) {$\bar X^{q_{K^\lhd}/}$};
\node (css) at (0,0) {$K^\lhd$};
\draw[->] (y) -- (q);
\draw[->] (y) -- (css);
\draw[->] (q) -- (css);
\end{tikzpicture}$$
Note that $p\circ q_{K^\lhd}:I\times K^\lhd\to K^\lhd$ is just the projection, so $p\circ q_{K^\lhd}$ is a Cartesian fibration; and by [@Lurie:2009un Proposition 4.2.2.4.] $\bar X^{q_{K^\lhd}/}\to K^\lhd$ is a coCartesian fibration classified by $(\bar p')^{q/}:({K})^\lhd \to {{\mathscr{C}\mathrm{at}}}_\infty$. In particular, the fibre of $\bar X^{q_{K^\lhd}/}$ over any $k\in K^\lhd$ may be identified with the undercategory $\big(\bar p'(k)\big)^{\pi_k\circ q/}$ (cf. [@Lurie:2009un § 4.2.2]). Let $X=\bar X\times_{K^\lhd} K$, and $q_{K}=q_{K^\lhd}\rvert_{I\times K}:I\times K\to X$. Then $\bar X^{q_{K^\lhd}/}\times_{K^\lhd} K\to K$ is canonically isomorphic to $X^{q_K/}\to K$. Consequently, by [@Lurie:2009un Proposition 3.3.3.1], it suffices to show that whenever
$$\label{eq:thetaequiv}\theta:{\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X)\to {\operatorname}{Fun}_{K}^{{\operatorname}{coCart}}(K,X)$$
is an equivalence of $\infty$-categories, so is $$\label{eq:thetaqequiv}{\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X^{q_{K^\lhd/}})\to {\operatorname}{Fun}_{K}^{{\operatorname}{coCart}}(K^\sharp,X^{q_K/}).$$
Using the identification $(I\times {K^\lhd})\diamond_{K^\lhd} (\Delta^n\times {K^\lhd})\cong (I\diamond \Delta^n)\times {K^\lhd}$, one sees that the $n$ simplices of ${\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X^{q_{K^\lhd/}})$ are lifting diagrams of the form $$\protect\tikz{
\node (m-1-1) at (0,2) {$I\times {K^\lhd}$};
\node (m-1-2) at (3,2) {$\bar X$};
\node (m-2-1) at (0,0) {$(I\diamond \Delta^n)\times {K^\lhd}$};
\node (m-2-2) at (3,0) {${K^\lhd}$};
\draw[->] (m-1-1) -- node {$q_{K^\lhd}$} (m-1-2);
\draw[right hook->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- node {$p$} (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\draw[dashed,->] (m-2-1) -- node {$\sigma$} (m-1-2);
}$$ such that for each vertex $v$ of $I\diamond \Delta^n$, the restriction $\sigma\rvert_{\{v\}\times {K^\lhd}}:{K^\lhd}\to \bar X$ is coCartesian. Thus, ${\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}(K^\lhd,\bar X^{q_{K^\lhd/}})\cong \big({\operatorname}{Fun}_{K^\lhd}^{{\operatorname}{coCart}}({K^\lhd},\bar X)\big)^{\tilde q/}$.
Similarly, ${\operatorname}{Fun}_K^{{\operatorname}{coCart}}K,X^{q_{K/}})\cong \big({\operatorname}{Fun}_K^{{\operatorname}{coCart}}({K}, X)\big)^{\theta\circ \tilde q/}$. It follows that is an equivalence whenever is.
Combining Proposition \[lem:LimOfInit\] and Lemma \[lem:LimComUnder\] yields the general case:
\[thm:LimInLim\]Let $I$ and $K$ be small simplicial sets, and suppose $p':K\to {{\mathscr{C}\mathrm{at}}}_\infty$ is a functor. Let $\lim p'\in {{\mathscr{C}\mathrm{at}}}_\infty$ denote the limit $\infty$-category and for each $k\in K$, let $\pi_k:\lim p'\to p'(k)$ denote the corresponding functor in the limit cone. Suppose that $q:I\to \lim p'$ is a diagram indexed by $I$, and that
1. for each vertex $k\in K$, the composite diagram $\pi_k\circ q:I\to p'(k)$ has a (co)limit diagram, and
2. for each edge $f:k\to k'$ of $K$, the functor $p'(f):p'(k)\to p'(k')$ takes (co)limit diagrams extending $\pi_k\circ q$ to (co)limit diagrams extending $\pi_{k'}\circ q$.
Then:
A
: there exists a map $\bar q:I^\rhd\to \lim p'$ which extends $q$ and such that each composite $\pi_{k'}\circ \bar q:I^\rhd\to p'(k)$ is a (co)limit (co)cone, and
B
: an arbitrary extension $\bar q:I^\rhd\to \lim p'$ of $q$ is a (co)limit diagram extending $q$ if and only if each composite $\pi_{k'}\circ \bar q:I^\rhd\to p'(k)$ is a (co)limit diagram extending $\pi_{k'}\circ q$.
In particular, if
1. for each vertex $k\in K$ the $\infty$-category $p'(k)$ admits (co)limits for all diagrams indexed by $I$, and
2. for each edge $f:k\to k'$ of $K$, the functor $p'(f):p'(k)\to p'(k')$ preserves (co)limits for all diagrams indexed by $I$.
then the limit $\infty$-category $\lim p'$ admits all (co)limits of shape $I$, and the functors $\pi_k:\lim p'\to p'(k)$ fitting into the limit cone preserve all (co)limits of shape $I$.
Let $\bar p':K^\lhd\to {{\mathscr{C}\mathrm{at}}}_\infty$ be a limit cone extending $p'$ which maps the cone point $\infty\in K^\lhd$ to $\lim p'$ and the cone edge $\infty\to k$ to $\pi_k$ for each $k\in K^\lhd$. Let $(\bar p')^{q/}:K^\lhd\to {{\mathscr{C}\mathrm{at}}}_\infty$ denote corresponding diagram of undercategories, as in Lemma \[lem:LimComUnder\]. Then $(\bar p')^{q/}:K^\lhd\to {{\mathscr{C}\mathrm{at}}}_\infty$ is a limit cone which (by assumptions (1) and (2)) satisfies the assumptions for Proposition \[lem:LimOfInit\]. Now for any $k\in K^\lhd$, the $\infty$-category of diagrams $I^\rhd\to \bar p'(k)$ extending $\pi_k\circ q$ is equivalent to the undercategory $\big(\bar p'(k)\big)^{\pi_k\circ q}$, and this equivalence identifies colimit diagrams with initial objects of the undercategory. Moreover, for any edge $k\to k'$ in $K^\lhd$, the functor $\big(\bar p'(k)\big)^{\pi_k\circ q/}\to \big(\bar p'(k')\big)^{\pi_{k'}\circ q/}$ preserves initial objects if and only if the functor $\bar p'(k)\to \bar p'(k')$ takes colimits diagrams extending $\pi_k\circ q$ to colimit diagrams extending $\pi_{k'}\circ q$.
Therefore statements **A** and **B** follow from Proposition \[lem:LimOfInit\].
Adjunctions and Kan Extensions
------------------------------
### Limits of adjunctions.
Our first application of Theorem \[thm:LimInLim\] is to prove the following result:
\[thm:limadj\] Suppose $$F_k:{\mathcal}{C}_k \rightleftarrows{\mathcal}{D}_k:G_k, \quad k\in K$$ is diagram of adjunctions coherently indexed by a small simplical set $K$, i.e. given by a diagram $(f\dashv g):K\to {\operatorname}{Adj}$ into the $\infty$-category of adjunctions. Let ${\mathcal}{C}=\lim_{k\in K}{\mathcal}{C}_k$ and ${\mathcal}{D}=\lim_{k\in K}{\mathcal}{D}_k,$ and suppose that
- there is a functor $F:{\mathcal}{C}\to{\mathcal}{D}$ which fits into the cone edge of a diagram $\bar f:K^\lhd\times \Delta^1\to {{\mathscr{C}\mathrm{at}}}_\infty$ extending $f:K\times \Delta^1\to {{\mathscr{C}\mathrm{at}}}_\infty$, and
- there is a functor $G:{\mathcal}{D}\to {\mathcal}{C}$ which fits into the cone edge of a diagram $\bar g:K^\lhd\times \Delta^1\to {{\mathscr{C}\mathrm{at}}}_\infty$, extending $g:K\times \Delta^1\to {{\mathscr{C}\mathrm{at}}}_\infty$,
such that
1. the restrictions $\bar f\rvert_{K^\lhd\times\{0\}}$ and $\bar g\rvert_{K^\lhd\times\{1\}}$ are limit cones for ${\mathcal}{C}$, and
2. the restrictions $\bar f\rvert_{K^\lhd\times\{1\}}$ and $\bar g\rvert_{K^\lhd\times\{0\}}$ are limit cones for ${\mathcal}{D}$.
Then $F$ and $G$ form a pair of adjoint functors $$F:{\mathcal}{C} \rightleftarrows{\mathcal}{D}:G$$
We defer the proof until later: we will first need to give a precise definition of the $\infty$-category, ${\operatorname}{Adj}$, of adjunctions. To do this, we will use the framework for adjunctions of $\infty$-categories in terms of *pairing of $\infty$-categories*, as developed in [@Lurie:ZsNVU6Ru]. For now we give an immediate corollary:
\[cor:KanClosed\] Suppose that $\delta:I\to I'$ is a morphism of simplicial sets. Let ${{\mathscr{C}\mathrm{at}}}_\infty^\delta\subset{{\mathscr{C}\mathrm{at}}}_\infty$ be the subcategory consisting of
- those $\infty$-categories ${\mathcal}{C}$ which admit left (right) Kan extensions along $\delta$ for any functor $f:I\to {\mathcal}{C}$, and
- those functors ${\mathcal}{C}\to{\mathcal}{C}'$ which preserve left (right) Kan extensions along $\delta$.
Then ${{\mathscr{C}\mathrm{at}}}_\infty^\delta\subset{{\mathscr{C}\mathrm{at}}}_\infty$ is closed under (small) limits.
Let $\phi:{{\mathscr{C}\mathrm{at}}}_\infty\to{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$ be the functor which sends an $\infty$-category ${\mathcal}{C}$ to the pullback-functor $$\label{eq:rKanExt}{\operatorname}{Fun}(I,{\mathcal}{C})\leftarrow {\operatorname}{Fun}(I',{\mathcal}{C}):\delta^*.$$ Evaluating $\phi$ at either endpoint of $\Delta^1$ $$\begin{aligned}
{\operatorname}{ev}_{\{0\}}\circ \phi:{{\mathscr{C}\mathrm{at}}}_\infty\xrightarrow{{\mathcal}{C}\mapsto {\operatorname}{Fun}(I',{\mathcal}{C})}{{\mathscr{C}\mathrm{at}}}_\infty\\
{\operatorname}{ev}_{\{1\}}\circ \phi:{{\mathscr{C}\mathrm{at}}}_\infty\xrightarrow{{\mathcal}{C}\mapsto {\operatorname}{Fun}(I,{\mathcal}{C})}{{\mathscr{C}\mathrm{at}}}_\infty\end{aligned}$$ results in continuous functors (they are right adjoints). Therefore [@Lurie:2009un Corollary 5.1.2.3] implies that $\phi$ is continuous.
Now ${\mathcal}{C}$ admits left Kan extensions along $\delta$ if and only if is a right adjoint. In particular, we may identify ${{\mathscr{C}\mathrm{at}}}_\infty^\delta$ with the pullback $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{{\mathscr{C}\mathrm{at}}}_\infty^\delta& {{\mathscr{C}\mathrm{at}}}_\infty\\{\operatorname}{Adj}&{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)\\};
\draw[->] (m-1-2) -- node {$\phi$} (m-2-2);
\draw[->] (m-2-1) -- node {$R$} (m-2-2);
\draw[dashed,->] (m-1-1) -- (m-1-2);
\draw[dashed,->] (m-1-1) -- (m-2-1);
\end{tikzpicture}$$ Where $R:{\operatorname}{Adj}\to{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$ is the functor which sends an adjunction $(F\dashv G)$ to its right adjoint $G$. By Theorem \[thm:limadj\], the functor $R$ is continuous. Therefore, by Theorem \[thm:LimInLim\], ${{\mathscr{C}\mathrm{at}}}_\infty^\delta$ admits all small limits and the functor ${{\mathscr{C}\mathrm{at}}}_\infty^\delta\to {{\mathscr{C}\mathrm{at}}}_\infty$ is continuous.
### Pairings of $\infty$-categories
We recall the theory of *pairings of $\infty$-categories*; essentially all this material is taken from [@Lurie:ZsNVU6Ru], though we provide proofs for certain details that will be important to us when discussing adjunctions. Recall that the $\infty$-category of pairings $${\operatorname}{CPair}\subseteq {\operatorname}{Fun}(\Lambda_0^2,{{\mathscr{C}\mathrm{at}}}_\infty)$$ is the full subcategory consisting of diagrams $$\begin{tikzpicture}
\node (m-1-1) at (0,1) {${\mathcal}{M}$};
\node (m-1-2) at (1,0) {${\mathcal}{D}^{op}$};
\node (m-2-1) at (-1,0) {${\mathcal}{C}$};
\draw[->] (m-1-1) -- node {$\lambda_D$} (m-1-2);
\draw[->] (m-1-1) -- node [swap] {$\lambda_C$} (m-2-1);
\end{tikzpicture}$$ such that $\lambda:{\mathcal}{M}\xrightarrow{\lambda_C\times\lambda_D} {\mathcal}{C}\times{\mathcal}{D}^{op}$ is equivalent to a right fibration.
Given such a right fibration, $\lambda$ is classified by a functor (cf. [@Lurie:2009un § 2.2.1]) $${\mathcal}{D}\times{\mathcal}{C}^{op}\to {\mathscr{S}}$$ to the $\infty$-category, ${\mathscr{S}}$, of spaces; or equivalently a functor $$\label{eq:lD'}\lambda':{\mathcal}{D}\to {\operatorname}{Fun}({\mathcal}{C}^{op},{\mathscr{S}})=:{\mathscr{P}}({\mathcal}{C})$$ to the $\infty$-category of presheaves over ${\mathcal}{C}$. Here $\lambda'$ takes each vertex $d\in {\mathcal}{D}$ to the right fibration $$\label{eq:Pred}{\mathcal}{M}\times_{{\mathcal}{D}^{op}}\{d\}\to {\mathcal}{C}.$$
As in [@Lurie:ZsNVU6Ru], we call an object of $m\in{\mathcal}{M}$ *right universal* if it is a terminal object of ${\mathcal}{M}\times_{{\mathcal}{D}^{op}}\{\lambda_D(m)\}$ and we call a right fibration $$\label{eq:RFibrPair}\lambda:{\mathcal}{M}\to {\mathcal}{C}\times{\mathcal}{D}^{op}$$ a *right representable pairing*, if for each $d\in{\mathcal}{D}^{op}$, there exists a right universal object in the fibre ${\mathcal}{M}\times_{{\mathcal}{D}^{op}}\{d\}$ over $d$. In this case, for each $d\in{\mathcal}{D}$, the right fibration is representable (cf. [@Lurie:2009un Proposition 4.4.4.5]). The Yoneda embedding[^7] ${\!\text{\usefont{U}{min}{m}{n}\symbol{'210}}\!}:{\mathcal}{C}\to {\mathscr{P}}({\mathcal}{C})$ identifies ${\mathcal}{C}$ with the full subcategory of ${\mathscr{P}}({\mathcal}{C})$ spanned by the representable presheaves (cf [@Lurie:2009un Proposition 5.1.3.1]); whence it follows that $\lambda'$ factors through ${\mathcal}{C}$, $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{D}&&{\mathcal}{C}\\&{\mathscr{P}}({\mathcal}{C})&\\};
\draw[->] (m-1-1) -- node[swap] {$\lambda'$} (m-2-2);
\draw[right hook->] (m-1-3) -- node {${\!\text{\usefont{U}{min}{m}{n}\symbol{'210}}\!}$} (m-2-2);
\draw[dashed,->] (m-1-1) -- node {$\lambda^R$} (m-1-3);
\end{tikzpicture}$$ The Yoneda lemma implies that we have a weak equivalence of spaces $${\operatorname}{Hom}_{\mathcal}{C}\big(c,\lambda^R(d)\big)\cong \{c\}\times_{{\mathcal}{C}}{\mathcal}{M}\times_{{\mathcal}{D}^{op}}\{d\},$$ which depend naturally on $(c,d)\in {\mathcal}{C}^{op}\times{\mathcal}{D}$.
Similarly, an object of $m\in{\mathcal}{M}$ is called *left universal* if it is a terminal object of ${\mathcal}{M}\times_{{\mathcal}{C}}\{\lambda_C(m)\}$, and the right fibration is called a *left representable pairing*, if for each $c\in{\mathcal}{C}$, there exists a left universal object in the fibre ${\mathcal}{M}\times_{\mathcal}{C}\{c\}$ over $c$. As before, this determines a functor $\lambda^L:{\mathcal}{C}\to {\mathcal}{D}$; and the yoneda Lemma implies that we have weak equivalences of spaces $${\operatorname}{Hom}_{\mathcal}{C}\big(\lambda^L(c),d\big)\cong \{c\}\times_{{\mathcal}{C}}{\mathcal}{M}\times_{{\mathcal}{D}^{op}}\{d\}\cong{\operatorname}{Hom}_{\mathcal}{C}\big(c,\lambda^R(d)\big)$$ depending naturally on $(c,d)\in {\mathcal}{C}^{op}\times{\mathcal}{D}$. Indeed, $\lambda^R$ is a right adjoint to $\lambda^L$ (cf. [@Lurie:ZsNVU6Ru] or [@Lurie:2009un § 5.2.6] for more details).
Suppose that ${\mathcal}{M}\to {\mathcal}{C}\times{\mathcal}{D}^{op}$ and ${\mathcal}{M}'\to {\mathcal}{C}'\times{\mathcal}{D}'^{op}$ are two right representable right fibrations of $\infty$-categories, then a morphism of diagrams $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{M}&{\mathcal}{M}'\\{\mathcal}{C}\times{\mathcal}{D}^{op}&{\mathcal}{C}'\times{\mathcal}{D}'^{op}\\};
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[dashed,->] (m-1-1) -- node {$\gamma$} (m-1-2);
\draw[dashed,->] (m-2-1) -- node {$\alpha\times\beta$} (m-2-2);
\end{tikzpicture}$$ is called *right representable* if it takes right universal objects to right universal objects.
The $\infty$-category of right-representable pairings ${\operatorname}{CPair}^R\subseteq{\operatorname}{CPair}$ is defined to be the subcategory whose objects are equivalent to right representable pairings, and whose morphisms are equivalent to right representable morphisms. The $\infty$-category of left-representable pairings ${\operatorname}{CPair}^L\subseteq{\operatorname}{CPair}$ is defined analogously.
\[lem:CPairClosed\] Let ${\operatorname}{CPair}^R\subseteq{\operatorname}{CPair}\subseteq {\operatorname}{Fun}(\Lambda_0^2,{{\mathscr{C}\mathrm{at}}}_\infty)$ be the $\infty$-categories defined in [@Lurie:ZsNVU6Ru]. Then both subcategories are closed under small limits.
Since ${\operatorname}{CPair}\subseteq {\operatorname}{Fun}(\Lambda_0^2,{{\mathscr{C}\mathrm{at}}}_\infty)$ is a reflective localization (cf. [@Lurie:ZsNVU6Ru Remark 4.2.9]), it is closed under small limits; so we need only show that ${\operatorname}{CPair}^R\subseteq{\operatorname}{CPair}$ is also closed under small limits.
Let $p:K^{op}\to {\operatorname}{CPair}^R$ be a diagram (for which we wish to compute the limit). The composite functor $K^{op}\to {\operatorname}{CPair}^R\to {\operatorname}{CPair}$ is classified by a diagram of simplicial sets $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{M}_p & {\mathcal}{D}_p^{op}\\ {\mathcal}{C}_p & K^{op}\\};
\draw[->] (m-1-1) -- node {$\lambda_C$} (m-1-2);
\draw[->] (m-1-1) -- node [swap] {$\lambda_D$} (m-2-1);
\draw[->] (m-1-2) -- node {$\tilde p_D$} (m-2-2);
\draw[->] (m-2-1) -- node [swap]{$\tilde p_C$} (m-2-2);
\end{tikzpicture}$$ where $\lambda_p=\lambda_C\times_K\lambda_D:{\mathcal}{M}_p\to {\mathcal}{C}_p\times_K{\mathcal}{D}_p$ is a right fibration and $\tilde p_C$ and $\tilde p_D$ are Cartesian fibrations. The limit of $p$ is a right fibration (cf. [@Lurie:ZsNVU6Ru Remark 4.2.9.]) $$\label{eq:LimPair}{\mathcal}{M}:=\lim p_M\to \lim p_C\times \lim p_D=:{\mathcal}{C}\times{\mathcal}{D},$$ where $p_M:K\to {{\mathscr{C}\mathrm{at}}}_\infty$ is the functor classified by $(\tilde p_C\times_K\tilde p_D)\circ\lambda_p$, and $p_C$ and $p_D$ are classified by $\tilde p_C$ and $\tilde p_D$ respectively. We need to show that is right representable and that the canonical morphisms to are right representable.
Now a vertex $d\in \lim p_D={\mathcal}{D}$ can be identified with a Cartesian section $\tilde d:K\to {\mathcal}{D}_p$ of $\tilde p_D$ (cf. [@Lurie:2009un Corollary 3.3.3.2.]). Let ${\mathcal}{M}_{p,d}\xrightarrow{\lambda_d} K$ be the Cartesian fibration which fits into the pullback square $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{M}_{p,d}& {\mathcal}{M}_p\\K& {\mathcal}{D}_p\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- node {$\tilde d$} (m-2-2);
\end{tikzpicture}$$ and let $q:K^{op}\to {{\mathscr{C}\mathrm{at}}}_\infty$ be the corresponding functor. Then $\lim q\cong {\mathcal}{M}_d:={\mathcal}{M}\times_{{\mathcal}{D}}\{d\}$ (since taking pullbacks commutes with taking limits). To show that is right representable, we need to show that ${\mathcal}{M}_d$ has a final object. However, since $p$ takes values in the $\infty$-category ${\operatorname}{CPair}^R$ of right representable pairings, for each $k\in K$, the pullback $${\mathcal}{M}_{p,d,k}:={\mathcal}{M}_{p,d}\times_{K}\{k\}$$ has a final object, and for each morphisms $(k\to k')$ in $K^{op}$, the corresponding functor ${\mathcal}{M}_{p,d,k}\to {\mathcal}{M}_{p,d,k'}$ takes final objects to final objects. Thus, by [@Riehl:2014ut Theorem. 3.16.] (or Proposition \[thm:LimInLim\]), the limit ${\mathcal}{M}_d\cong \lim q$ has a final object, and moreover the canonical morphisms ${\mathcal}{M}_d\to {\mathcal}{M}_{p,d,k}$ preserve final objects.
It follows that $\lim p\in {\operatorname}{CPair}$ is in fact an element of ${\operatorname}{CPair}^R$ and that the limit cone is a diagram in ${\operatorname}{CPair}^R$; i.e. ${\operatorname}{CPair}^R\subseteq {\operatorname}{CPair}$ is closed under small limits.
\[prop:FunInfCat\] There are equivalences of $\infty$-categories $$\begin{aligned}
{\operatorname}{CPair}^L&\cong {\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty),\\
{\operatorname}{CPair}^R&\cong {\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty),\end{aligned}$$ which associate a left representable pairing $\lambda:{\mathcal}{M}\to {\mathcal}{C}\times{\mathcal}{D}^{op}$ to the functor $\lambda^L:{\mathcal}{C}\to {\mathcal}{D}$, and a right representable pairing $\lambda:{\mathcal}{M}\to {\mathcal}{C}\times{\mathcal}{D}^{op}$ to the functor $\lambda^R:{\mathcal}{D}\to {\mathcal}{C}$.
As in [@Lurie:ZsNVU6Ru], we say that a right fibration is a perfect pairing if it is both left and right representable, and an object $m\in {\mathcal}{M}$ is left universal if and only if it is right universal. Let ${\operatorname}{CPair}^{\textrm{perf}}\subset{\operatorname}{CPair}^L$ be the full subcategory spanned by the perfect pairings. Let $\phi:{\operatorname}{CPair}^{\textrm{perf}}\to{{\mathscr{C}\mathrm{at}}}_\infty$ denote the forgetful functor which sends a perfect pairing to ${\mathcal}{C}$; and let $\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}$ denote the $\infty$-category fitting into the pullback square $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}&{\operatorname}{CPair}^{\textrm{perf}}\\{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)&{{\mathscr{C}\mathrm{at}}}_\infty\\};
\draw[dashed,->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- node {$\phi$} (m-2-2);
\draw[dashed,->] (m-1-1) -- (m-1-2);
\draw[->] (m-2-1) -- node {${\operatorname}{ev}_1$} (m-2-2);
\end{tikzpicture}$$ Note: since the bottom arrow is a Cartesian fibration (cf. [@Lurie:2009un Corollary 2.4.7.11]), this homotopy pullback can be computed as a pullback of simplicial sets (cf. [@Lurie:2009un Corollary 3.3.1.4]). Since $\phi$ is an equivalence of $\infty$-categories (cf. [@Lurie:ZsNVU6Ru Remark 4.2.12]), the left arrow defines an equivalence between $\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}$ and ${\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$.
The inclusion ${\operatorname}{CPair}^{\textrm{perf}}\subset {\operatorname}{Fun}(\Lambda_0^2,{{\mathscr{C}\mathrm{at}}}_\infty)$ allows us to identify $\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}$ with diagrams of the form $$\label{eq:tilfun}\begin{tikzpicture}
\node (c) at (-1,1) {${\mathcal}{C}$};
\node (d) at (0,0) {${\mathcal}{D}$};
\node (p) at (1,1) {${\mathcal}{P}$};
\node (d1) at (2,0) {$\tilde{{\mathcal}{D}}^{op}$};
\draw[->] (c)-- node {$f$} (d);
\draw[->] (p)-- (d);
\draw[->] (p)-- (d1);
\end{tikzpicture}$$ where ${\mathcal}{P}\to {\mathcal}{D}\times\tilde{{\mathcal}{D}}^{op}$ is a perfect pairing. Taking the limit of such a diagram yields $$\begin{tikzpicture}
\node (m) at (1,2) {${\mathcal}{M}$};
\node (c) at (-1,1) {${\mathcal}{C}$};
\node (d) at (0,0) {${\mathcal}{D}$};
\node (p) at (1,1) {${\mathcal}{P}$};
\node (d1) at (2,0) {$\tilde{{\mathcal}{D}}^{op}$};
\draw[->] (c)-- (d);
\draw[->] (p)-- (d);
\draw[->] (p)-- (d1);
\draw[dashed,->] (m) -- (c);
\draw[dashed,->] (m) -- (d);
\draw[dashed,->] (m) -- (p);
\draw[dashed,->] (m) -- (d1);
\end{tikzpicture}$$ where $$\label{eq:lam'}\lambda:{\mathcal}{M}\to {\mathcal}{C}\times\tilde{{\mathcal}{D}}^{op}$$ is a left-representable pairing. Thus we get a functor $$A:\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}\to {\operatorname}{CPair}^L,$$ sending an object of the form to the left representable pairing . Notice that, by construction, the functor $f$ appearing in is equivalent to $\lambda^L:{\mathcal}{C}\to \tilde{{\mathcal}{D}}$.
It remains to show that $A$ is an equivalence of categories. The essential surjectivity of $A$ is explained in[@Lurie:ZsNVU6Ru Remark 4.2.13]. We argue that $A$ is fully faithful: Suppose that $$\begin{tikzpicture}
\node at (-2,.5) {$\tilde{f}=$};
\node (m) at (1,2) {${\mathcal}{M}$};
\node (c) at (-1,1) {${\mathcal}{C}$};
\node (d) at (0,0) {${\mathcal}{D}$};
\node (p) at (1,1) {${\mathcal}{P}$};
\node (d1) at (2,0) {$\tilde{{\mathcal}{D}}^{op}$};
\draw[->] (c)-- node {$f$} (d);
\draw[->] (p)-- (d);
\draw[->] (p)-- (d1);
\draw[dashed,->] (m) -- (c);
\draw[dashed,->] (m) -- (d);
\draw[dashed,->] (m) -- (p);
\draw[dashed,->] (m) -- (d1);
\node at (3,.5) {and};
\node at (4,.5) {$\tilde{f}'=$};
\node (m') at (7,2) {${\mathcal}{M}'$};
\node (c') at (5,1) {${\mathcal}{C}'$};
\node (d') at (6,0) {${\mathcal}{D}'$};
\node (p') at (7,1) {${\mathcal}{P}'$};
\node (d1') at (8,0) {$\tilde{{\mathcal}{D}}'^{op}$};
\draw[->] (c')-- node {$f'$} (d');
\draw[->] (p')-- (d');
\draw[->] (p')-- (d1');
\draw[dashed,->] (m') -- (c');
\draw[dashed,->] (m') -- (d');
\draw[dashed,->] (m') -- (p');
\draw[dashed,->] (m') -- (d1');
\end{tikzpicture}$$ are a pair of objects in $\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}$, with $$A(\tilde{f})=\big(\lambda:{\mathcal}{M}\to {\mathcal}{C}\times\tilde{{\mathcal}{D}}^{op}\big),\text{ and }A(\tilde{f}')=\big(\lambda':{\mathcal}{M}'\to {\mathcal}{C}'\times\tilde{{\mathcal}{D}}'^{op}\big).$$ We need to show that the natural map between the mapping spaces $$\label{eq:AFullFaith}A:{\operatorname}{Map}_{\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}}(\tilde f,\tilde f')\to {\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{M}')$$ is a homotopy equivalence.[^8]
On the one hand, $${\operatorname}{Map}_{\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}}(\tilde f,\tilde f')
\cong {\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{C}')\times^h_{{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{D}')}{\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{P},{\mathcal}{P}')$$ But, since ${\mathcal}{P}'\to {\mathcal}{D}'\times\tilde{{\mathcal}{D}}'^{op}$ is a perfect pairing, [@Lurie:ZsNVU6Ru Proposition 4.2.10] shows we have homotopy equivalences of mapping spaces $${\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{P},{\mathcal}{P}')\xrightarrow{\cong}{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}(\tilde{{\mathcal}{D}},\tilde{{\mathcal}{D}}')\xleftarrow{\cong}{\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{P}').$$ Consequently,
\[eq:HomEquivPerfPair\]$${\operatorname}{Map}_{\widetilde{{\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)}}(\tilde f,\tilde f')
\cong {\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{C}')\times^h_{{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{D}')}{\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{P}')$$
On the other hand, since ${\mathcal}{M}'$ is a pullback of ${\mathcal}{P}'$, we have a homotopy equivalence of mapping spaces $${\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{M}')\cong {\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{C}')\times^h_{{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{D}')}{\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{P}').$$
It follows from that is a homotopy equivalence. In particular, $A$ is fully faithful.
### The $\infty$-category of Adjunctions, and the proof of Theorem \[thm:limadj\]
We are now in a position to define the $\infty$-category of adjunctions and to prove Theorem \[thm:limadj\].
The $\infty$-category of adjunctions, ${\operatorname}{Adj}$ is defined as the pullback of $\infty$-categories $$\label{eq:AdjPull}\begin{tikzpicture}
\node (adj) at (-7,0) {${\operatorname}{Adj}$};
\node (fun1L) at (-4,-1) {${\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$};
\node (fun1R) at (-4,1) {${\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$};
\node (cpL) at (-1,-1) {${\operatorname}{CPair}^L$};
\node (cpR) at (-1,1) {${\operatorname}{CPair}^R$};
\node (cp) at (1,0) {${\operatorname}{CPair}$};
\draw[->] (fun1L) -- node {$\cong$} (cpL);
\draw[->] (fun1R) -- node {$\cong$} (cpR);
\draw[->] (cpL) -- (cp);
\draw[->] (cpR) -- (cp);
\draw[dashed,->] (adj) -- (fun1L);
\draw[dashed,->] (adj) -- (fun1R);
\end{tikzpicture}$$
By Lemma \[lem:CPairClosed\] and Proposition \[thm:LimInLim\], each of the categories in the diagram are complete, and each of the functors in the diagram preserve small limits. Thus, any diagram $(f\dashv g):K\to {\operatorname}{Adj}$ admits a limit. Moreover, the limiting left adjoint $F:{\mathcal}{C}\to {\mathcal}{D}$ is a limiting functor for the diagram $f:K\to {\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$. In particular, $F$ can be characterized as in the statement of the theorem (cf. [@Lurie:2009un Corollary 5.1.2.3]).
Similarly, the limiting right adjoint $G:{\mathcal}{D}\to{\mathcal}{C}$ is a limiting functor for the diagram $g:K\to {\operatorname}{Fun}(\Delta^1,{{\mathscr{C}\mathrm{at}}}_\infty)$; so $G$ can be characterized as in the statement of the theorem.
Complete $k$-fold Segal objects.
================================
Let ${\mathbb{\Delta}}$ denote the simplex category, and for any simplicial set $K$, let ${\mathbb{\Delta}}_{/K}\to {\mathbb{\Delta}}$ denote the corresponding category of simplices[^9] of $K$. The spine of the standard $n$-simplex is the subsimplicial set $${\operatorname}{Sp}(n)=\overset{n}{\overbrace{\Delta^{\{0,1\}}\coprod_{\Delta^{\{1\}}}\cdots \coprod_{\Delta^{\{n-1\}}}\Delta^{\{n-1,n\}}}}\subseteq \Delta^n.$$ generated by the 1-simplices $\Delta^{\{i,i+1\}}\subseteq\Delta^n$. The inclusion ${\operatorname}{Sp}(n)\subseteq \Delta^n$ is a categorical equivalence,[^10] and a simplical object $X_\bullet:{\mathbb{\Delta}}\to {\mathcal}{X}$ in an $\infty$-category ${\mathcal}{X}$ is called a *category object* if it satisfies the so-called *Segal conditions* (cf. [@Rezk:2001ey]): i.e. for each $n\geq 0$, the natural map $$\label{eq:Segal}X_n\to \lim_{{\mathbb{\Delta}}^{op}_{/{\operatorname}{Sp}(n)}} X_\bullet \cong \overset{n}{\overbrace{X_1\times_{X_0}\cdots \times_{X_0}X_1}}$$ is an equivalence.
Given a category object $X_\bullet$ in ${\mathcal}{X}$, one should think of $X_0\in {\mathcal}{X}$ as describing the objects of an $(\infty,1)$-category internal to ${\mathcal}{X}$, $X_1\in {\mathcal}{X}$ as describing the morphisms of an $(\infty,1)$-category internal to ${\mathcal}{X}$, $X_i\in{\mathcal}{X}$ as describing the object classifying composable $i$-tuples of morphisms, and the various structural maps between the $X_i$’s as describing the homotopy-associative composition and units.
Now suppose that ${\mathcal}{X}$ is an $\infty$-topos. We let ${\operatorname}{Cat}({\mathcal}{X})\subseteq {\operatorname}{Fun}({\mathbb{\Delta}}^{op},{\mathcal}{X})$ denote the full subcategory spanned by the category objects. Unfortunately, ${\operatorname}{Cat}({\mathcal}{X})$ doesn’t describe the correct homotopy theory of $(\infty,1)$-categories internal to ${\mathcal}{X}$; one must localize with respect to an appropriate class of “fully faithful and essentially surjective functors”. In order to describe this phenomena in more detail, we recall that a category object $X_\bullet\in {\operatorname}{Cat}({\mathcal}{X})$ is called a groupoid object if all it’s morphisms are invertible, i.e. $$X_2\to \lim_{{\mathbb{\Delta}}^{op}_{/\Lambda^2_0}} X_\bullet$$ is an equivalence, where $$\Lambda^2_0=\Delta^{0,1}\coprod_{\Delta^{0}}\Delta^{0,2}\subset \Delta^2.$$ We let ${\operatorname}{Gpd}({\mathcal}{X})\subseteq{\operatorname}{Cat}({\mathcal}{X})$ denote the full subcategory spanned by the groupoid objects. The *underlying groupoid functor* ${\operatorname}{Gp}:{\operatorname}{Cat}({\mathcal}{X})\to {\operatorname}{Gpd}({\mathcal}{X})$ is any right adjoint to the inclusion. For a category object $X_\bullet$, one should think of ${\operatorname}{Gp} X_\bullet$ as describing the “maximal groupoid contained in $X_\bullet$”, which classifies the “objects” of the internal $(\infty,1)$-category $X_\bullet$.
The *fully faithful and essentially surjective morphisms* (cf. [@Lurie:2009uz Definition 1.2.12]), are those morphisms of category objects $X_\bullet\to Y_\bullet$ in ${\mathcal}{X}$ which are
fully faithful
: the diagram $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{X_1&Y_1\\
X_0\times X_0&Y_0\times Y_0\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\end{tikzpicture}$$ is a pullback square, and
essentially surjective
: the map $${\lvert{\operatorname}{Gp}X_\bullet\rvert}\to{\lvert{\operatorname}{Gp}Y_\bullet\rvert}$$ between the classifying spaces of objects is an equivalence, where ${\lvert-\rvert}$ denotes the geometric realization: $${\lvertZ_\bullet\rvert}={{{\operatorname}{co}\lim}}_{{\mathbb{\Delta}}^{op}}Z_\bullet$$ for any $Z_\bullet:{\mathbb{\Delta}}^{op}\to {\mathcal}{X}$.
Localizing along the fully faithful and essentially surjective morphisms of category objects, one obtains ${CSS}({\mathcal}{X})\subseteq {\operatorname}{Cat}({\mathcal}{X})$, the correct homotopy theory of $(\infty,1)$-category objects in ${\mathcal}{X}$. Following Rezk [@Rezk:2001ey] Lurie proves [@Lurie:2009uz Theorem 1.2.13] that ${CSS}({\mathcal}{X})\subseteq {\operatorname}{Cat}({\mathcal}{X})$ is equivalent to the full subcategory spanned by the *complete Segal objects*: those category objects $X_\bullet\in {\operatorname}{Cat}({\mathcal}{X})$ such that ${\operatorname}{Gp} X_\bullet$ is essentially constant (i.e. ${\operatorname}{Gp} X_\bullet:{\mathbb{\Delta}}^{op}\to {\mathcal}{X}$ is equivalent to a constant functor).
To describe $(\infty,k)$-category objects in ${\mathcal}{X}$, will be interested in the following full subcategories of multisimplicial objects $$\label{eq:CSShier}{CSS}_k({\mathcal}{X})\subseteq {\operatorname}{Seg}_k({\mathcal}{X})\subseteq {\operatorname}{Cat}^k({\mathcal}{X})\subseteq {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big).$$
Here ${\operatorname}{Cat}^k({\mathcal}{X})$ is spanned by the $k$-uple category objects, i.e those multisimplicial objects $X_{\bullet,\dots,\bullet}$ such that for any $1\leq i\leq k$ and any $n_1,\dots,\hat n_i,\dots, n_k\geq 0$, the simplicial object
\[eq:kfoldSeg0\]$$\label{eq:kupleCat}X_{n_1,\dots, n_{i-1},\bullet,n_{i+1},\dots,n_k}:{\mathbb{\Delta}}^{op}\to{\mathcal}{X}$$ is a category object. As before $X_{0,\dots,0}$ encodes the objects of the $k$-uple category internal to ${\mathcal}{X}$, but now each of $X_{1,0,\dots,0}$, $X_{0,1,0,\dots,0}$, $\dots,X_{0,\dots,0,1}$ encodes a different type of 1-morphism; while each of $X_{i_1,\dots,i_k}$ (with $0\leq i_1,\dots,i_k\leq 1$) represents a different type of $(i_1+\dots+i_k)$-morphism. As before, ${\operatorname}{Cat}^k({\mathcal}{X})$ does not model the correct homotopy theory of $k$-uple categories internal to ${\mathcal}{X}$; one must localize with respect to an appropriate class of “fully faithful and essentially surjective functors”.
Next, ${\operatorname}{Seg}_k({\mathcal}{X})$ is spanned by the $k$-fold Segal objects (cf. [@Barwick:2005wl]), i.e. those $k$-uple category objects $X_{\bullet,\dots,\bullet}$ such that for every $1\leq i\leq k$, and any $n_1,\dots,n_{i-1}\geq 0$ the multisimplicial object $$\label{eq:kfoldSeg} X_{n_1,\dots,n_{i-1},0,\bullet,\dots,\bullet}$$
is equivalent to a constant functor. The idea behind this condition is that while a $k$-uple category object has ${k\choose i}$ different types of $i$-morphisms, there is only one non-trivial type of $i$-morphism in a $k$-fold Segal object. More specifically, $X_{(\bullet,\dots,\bullet)}:({\mathbb{\Delta}}^{op})^n\to {\mathcal}{X}$ encodes the data of an $(\infty,k)$-category as follows:
- $X_{(0,\dots,0)}$ encodes the objects,
- $X_{(1,0,\dots,0)}$ encodes the $1$-morphisms,
- $X_{(1,1,0,\dots,0)}$ encodes the $2$-morphisms,
- $\dots$
- and $X_{(1,\dots,1)}$ encodes the $k$-morphisms.
the remaining objects $X_{(n_1,\dots,n_k)}$ encode composable configurations of morphisms, while the homotopy coherent associative composition and unit are encoded in the various structural maps between the spaces $X_{(n_1,\dots,n_k)}$.
Note that ${\operatorname}{Seg}_k({\mathcal}{X})$ does not model the correct homotopy theory of $k$-fold categories internal to ${\mathcal}{X}$. However, when ${\mathcal}{X}$ is an $\infty$-topos (e.g. ${\mathcal}{X}={\mathscr{S}}$), we may localize ${\operatorname}{Seg}_k({\mathcal}{X})$ with respect to an appropriate class of “fully faithful and essentially surjective functors”, to obtain ${CSS}_k({\mathcal}{X})$, which is spanned by those $k$-fold Segal objects which satisfy a certain *completeness condition*; we refer the reader to [@Barwick:2005wl; @Barwick:2011wl; @Lurie:2009uz; @Haugseng:2014vw] for more details.
\[lem:ContOfCatObj\] Suppose ${\mathcal}{D}$ is an $\infty$-category, ${\mathcal}{X}$ is a presentable $\infty$-category, and $\tilde{{\mathcal}{X}}$ is any reflective localization of ${\operatorname}{Cat}^k({\mathcal}{X})$ (the two main examples being $\tilde{{\mathcal}{X}}={\operatorname}{Seg}_k({\mathcal}{X})$, or when ${\mathcal}{X}$ is an $\infty$-topos, $\tilde{{\mathcal}{X}} = {CSS}_k({\mathcal}{X})$). Then a functor $$\label{eq:CSSFunct}{\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(\bullet,\dots,\bullet)}}\tilde{{\mathcal}{X}}$$ is continuous if and only if each of the composite functors $$\label{eq:CSSMorphFunct}{\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(i_1,\dots,i_k)}} {\mathcal}{X},\quad 0\leq i_1,\dots, i_k\leq 1$$ obtained by evaluating at $(i_1,\dots,i_k)\in {\mathbb{\Delta}}^k$ for $0\leq i_1,\dots, i_k\leq 1$, are continuous.
Without loss of generality, we may take $\tilde{{\mathcal}{X}}={\operatorname}{Cat}^k({\mathcal}{X})$. Recall that ${\operatorname}{Cat}^k({\mathcal}{X})$ is a reflective localization of ${\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)$ (cf. [@Lurie:2009uz]), so is continuous if and only if the composite functor $${\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(\bullet,\dots,\bullet)}}{\operatorname}{Cat}^k({\mathcal}{X})\hookrightarrow {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)$$ is continuous. Since limits in functor $\infty$-categories are detected pointwise (cf. [@Lurie:2009un Corollary 5.1.2.3]), it follows that is continuous if and only if the composite functors $$\label{eq:EvalAtNs}{\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(n_1,\dots,n_k)}} {\mathcal}{X}$$ obtained by evaluating at any $(n_1,\dots,n_k)\in {\mathbb{\Delta}}^k$ are continuous. This proves the only if part of the statement.
Now we prove the if part of the statement. Let $i:{\mathbb}{Morph}^k\hookrightarrow {\mathbb{\Delta}}^k$ denote the inclusion of the full subcategory spanned by $(i_1,\dots,i_k)\in {\mathbb{\Delta}}^k$, where $0\leq i_1,\dots,i_k\leq 1$. Then by assumption, and hence the restricted functor $${\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(\bullet,\dots,\bullet)}}{\operatorname}{Cat}^k({\mathcal}{X})\hookrightarrow {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)\xrightarrow{i^*} {\operatorname}{Fun}\big(({\mathbb}{Morph}^k)^{op},{\mathcal}{X}\big)$$ is continuous.
A map of simplices $(\phi:[n]\to [m])\in{\mathbb{\Delta}}$ is said to be *inert* if it is the inclusion of a full sub-interval, i.e. $\phi(i+1)=\phi(i)+1$ for every $i\in [n]$ (cf. [@Haugseng:2014vw; @Barwick:2013vc]). We let $j:{\mathbb{\Delta}}_{{\operatorname}{int}}\hookrightarrow{\mathbb{\Delta}}$ denote the inclusion of the wide subcategory containing only the inert maps. For any $(n_1,\dots,n_k)\in {\mathbb{\Delta}}^k$, let $${\mathbb}{Spine}(n_1,\dots,n_k)={\mathbb}{Morph}^k\times_{{\mathbb{\Delta}}_{{\operatorname}{int}}^k}\big({\mathbb{\Delta}}_{{\operatorname}{int}}^k\big)_{/(n_1,\dots,n_k)}.$$ Then the Segal conditions imply that for any $(n_1,\dots,n_k)\in {\mathbb{\Delta}}^k$, and $d\in {\mathcal}{D}$, the object $F(d)_{(n_1,\dots,n_k)}\in {\mathcal}{X}$ is a limit for the composite functor $${\mathbb}{Spine}(n_1,\dots,n_k)^{op}\to ({\mathbb{\Delta}}^k)^{op}\xrightarrow{F(d)}{\mathcal}{X}$$ (cf. [@Haugseng:2014vw Lemma 2.27]). Equivalently, the restriction $F(d)\rvert_{\big({\mathbb{\Delta}}_{{\operatorname}{int}}^k)^{op}}$ is a right Kan extension along $\delta:{\mathbb}{Spine}(n_1,\dots,n_k)^{op}\to ({\mathbb{\Delta}}_{{\operatorname}{int}}^k)^{op}$.
Let $${\operatorname}{Fun}\big(({\mathbb{\Delta}}_{{\operatorname}{int}}^k)^{op},{\mathcal}{X}\big)\xrightarrow{\delta_*} {\operatorname}{Fun}\big(({\mathbb}{Morph}^k)^{op},{\mathcal}{X}\big)$$ denote the right adjoint (the global right Kan extension) to the pullback $\delta^*$, then the composite $${\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(\bullet,\dots,\bullet)}}{\operatorname}{Cat}^k({\mathcal}{X})\hookrightarrow {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)\xrightarrow{i^*} {\operatorname}{Fun}\big(({\mathbb}{Morph}^k)^{op},{\mathcal}{X}\big)\xrightarrow{\delta_*}{\operatorname}{Fun}\big(({\mathbb{\Delta}}_{{\operatorname}{int}}^k)^{op},{\mathcal}{X}\big)$$ is continuous. However, by assumption, this functor is equivalent to the restricted functor $${\mathcal}{D}\xrightarrow{d\mapsto F(d)_{(\bullet,\dots,\bullet)}}{\operatorname}{Cat}^k({\mathcal}{X})\hookrightarrow {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)\xrightarrow{j^*} {\operatorname}{Fun}\big(({\mathbb{\Delta}}_{{\operatorname}{int}}^k)^{op},{\mathcal}{X}\big).$$ It follows that each is continuous, whence is continuous.
The Sheaf of Complete $k$-Fold Segal Objects
--------------------------------------------
Given two morphisms of simplicial sets $X\to S$ and $Y\to S$, we let $Y^X\to S$ denote the simplicial set satisfying the universal property that for any morphism of simplicial sets $K\to S$, commutative diagrams of the form $$\begin{tikzpicture}\node (y) at (-2,1) {$K\times_S X$};
\node (q) at (2,1) {$Y$};
\node (css) at (0,0) {$S$};
\draw[->] (y) -- (q);
\draw[->] (y) -- (css);
\draw[->] (q) -- (css);
\end{tikzpicture}$$ correspond to diagrams of the form $$\begin{tikzpicture}\node (y) at (-2,1) {$K$};
\node (q) at (2,1) {$Y^X$};
\node (css) at (0,0) {$S$};
\draw[->] (y) -- (q);
\draw[->] (y) -- (css);
\draw[->] (q) -- (css);
\end{tikzpicture}$$ In particular, when $Y\to S$ is a coCartesian fibration, and $X\to S$ is a Cartesian fibration, then $Y^X\to S$ is a coCartesian fibration satisfying $${\operatorname}{Fun}_{S}(K,Y^X)\cong{\operatorname}{Fun}_S(K\times_S X,Y),$$ (see [@Lurie:2009un Corollary 3.2.2.13] for more details).
Let $\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ denote the $\infty$-category of (not necessarily small) $\infty$-categories, and $\imath:{{\mathscr{LT}\mathrm{op}}}\hookrightarrow\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ denote the subcategory consisting of $\infty$-topoi and geometric morphisms (functors which preserve small colimits and finite limits). Notice that $\imath$ factors through the subcategory of presentable $\infty$-categories and left adjoints. Let $\imath^*{\mathcal}{Z}\to {{\mathscr{LT}\mathrm{op}}}$ denote the (cannonical) presentable fibration classified by $\imath$ (cf. [@Lurie:2009un Proposition 5.5.3.3]).[^11] We define a presentable fibration $$k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}):=\imath^*{\mathcal}{Z}^{({{\mathscr{LT}\mathrm{op}}}\times ({\mathbb{\Delta}}^k)^{op})}\xrightarrow{p} {{\mathscr{LT}\mathrm{op}}},$$ whose fibre over any $\infty$-topos ${\mathcal}{X}$ is equivalent to ${\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)$ and which associates to any geometric morphism of $\infty$-topoi $f^*:{\mathcal}{X}\leftrightarrows{\mathcal}{Y}:f_*$ the adjunction given by composition with $f^*$ (resp. $f_*$) $$(f^*)_!:{\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}\big)\leftrightarrows {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{Y}\big):(f_*)_!.$$
Suppose that $X_{\bullet,\dots,\bullet}\in k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}})$ is a vertex lying over ${\mathcal}{X}=p(X_{\bullet,\dots,\bullet})$. We say that $X_{\bullet,\dots,\bullet}$ is a *complete Segal object* if it lies in the essential image of ${CSS}_k({\mathcal}{X})\hookrightarrow p^{-1}({\mathcal}{X})$. We define ${\int{CSS}_k}$ to be the full subcategory of $k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}})$ spanned by the complete Segal objects.
\[lem:CSSPres\] ${\int{CSS}_k}\to {{\mathscr{LT}\mathrm{op}}}$ is a presentable fibration.
We begin by showing that ${\int{CSS}_k}$ is a Cartesian fibration. It suffices to show that for any complete Segal object $Y_{\bullet,\dots,\bullet}\in k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}})$ and any $p$-Cartesian morphism $\tilde f:X_{\bullet,\dots,\bullet}\to Y_{\bullet,\dots,\bullet}$, the vertex $X_{\bullet,\dots,\bullet}\in k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}})$ is also a complete Segal object. Let $f^*:{\mathcal}{X}\to{\mathcal}{Y}$ denote the image of $\tilde f$ under $p$, and let ${\mathcal}{X}\leftarrow {\mathcal}{Y}:f_*$ denote a right adjoint to $f^*$. Then $X_{\bullet,\dots,\bullet}\cong (f_*)_!(Y_{\bullet,\dots,\bullet})$ by the construction of $p$, and the latter is a complete Segal object by [@Haugseng:2014vw Proposition 2.20].
Indeed, this shows that ${\int{CSS}_k}\to{{\mathscr{LT}\mathrm{op}}}$ is a Cartesian fibration classifed by a functor $$\chi:{{\mathscr{LT}\mathrm{op}}}^{op}\to \widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$$ such that
- for every $\infty$-topos ${\mathcal}{X}\in{{\mathscr{LT}\mathrm{op}}}$, the image $\chi({\mathcal}{X})$ is equivalent to the presentable $\infty$-category ${CSS}_k({\mathcal}{X})$, and
- for every geometric morphism $f^*:{\mathcal}{X}\leftrightarrows{\mathcal}{Y}:f_*$, the functor $${CSS}_k({\mathcal}{X})\leftarrow {CSS}_k({\mathcal}{Y}):\chi(f^*)$$ is equivalent to $${CSS}_k({\mathcal}{X})\leftarrow {CSS}_k({\mathcal}{Y}):(f_*)_!,$$ which has a left adjoint (cf. [@Haugseng:2014vw Proposition 2.20]).
It follows that ${\int{CSS}_k}\to {{\mathscr{LT}\mathrm{op}}}$ is a presentable fibration (cf. [@Lurie:2009un Proposition 5.5.3.3]).
By construction, the objects $X_{\bullet,\dots,\bullet}$ in ${\int{CSS}_k}$ over an $\infty$-topos ${\mathcal}{X}$ can be identified with complete Segal objects in ${\mathcal}{X}$, and morphisms $X_{\bullet,\dots,\bullet}\to Y_{\bullet,\dots,\bullet}$ in ${\int{CSS}_k}$ over a geometric morphism $f^*:{\mathcal}{X}\leftrightarrows{\mathcal}{Y}:f_*$ of $\infty$-topoi can be identified with either
- morphisms $X_{\bullet,\dots,\bullet}\to (f_*)_!(Y_{\bullet,\dots,\bullet})$ in ${CSS}_k({\mathcal}{X})$, or
- morphisms $L_{k,{\mathcal}{Y}}(f^*)_!(X_{\bullet,\dots,\bullet})\to Y_{\bullet,\dots,\bullet}$ in ${CSS}_k({\mathcal}{Y})$,
where $L_{k,{\mathcal}{Y}}:{\operatorname}{Seg}_k({\mathcal}{Y})\to {CSS}_k({\mathcal}{Y})$ is the localization functor which sends a $k$-fold Segal object in ${\mathcal}{Y}$ to its completion. The equivalence between morphisms of types (a) and (b) is given by the adjunction $$L_{k,{\mathcal}{Y}}(f^*)_!:{CSS}_k({\mathcal}{X})\leftrightarrows{CSS}_k({\mathcal}{Y}):(f_*)_!$$ of [@Haugseng:2014vw Proposition 2.20].
Recall that a geometric morphism $f^*:{\mathcal}{X}\rightleftarrows {\mathcal}{Y}:f_*$ is said to be étale if it admits a factorization $$f^*:{\mathcal}{X}\underset{f'_*}{\overset{f'^*}{\rightleftarrows}}{\mathcal}{X}_{/U}\cong {\mathcal}{Y}:f_*$$ for some object $U\in {\mathcal}{X}$. We let ${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\subset{{\mathscr{LT}\mathrm{op}}}$ denote the subcategory spanned by the étale geometric morphisms, and we define $\int^{\acute{e}t}{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ to be the presentable fibration fitting into the pullback square: $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{\int^{\acute{e}t}{CSS}_k&\int{CSS}_k\\
{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}&{{\mathscr{LT}\mathrm{op}}}\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\end{tikzpicture}$$
Since $\int{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}$ and $\int^{\acute{e}t}{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ are both presentable fibrations[@Lurie:2009un Corollary 4.3.1.11] implies that they both admit all small relative limits and colimits. Since ${{\mathscr{LT}\mathrm{op}}}$ admits all small limits and colimits (cf. [@Lurie:2009un § 6.3]), it follows that $\int{CSS}_k$ admits all small limits and colimits, and that the functor $\int{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}$ preserves those limits and colimits (cf. [@Gepner:2015ww Lemma 9.8]).
Finally, [@Lurie:2009un Theorem 6.3.5.13] implies that ${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\subset{{\mathscr{LT}\mathrm{op}}}$ is closed under small limits, which implies that $\int^{\acute{e}t}{CSS}_k\subset\int{CSS}_k$ is also closed under small limits.
As explained in [@Lurie:2009un Remark 6.3.5.10] for any $\infty$-topos ${\mathcal}{X}$, the Cartesian fibration $${\operatorname}{Fun}(\Delta^1,{\mathcal}{X})\to {\operatorname}{Fun}(\{1\},{\mathcal}{X})\cong{\mathcal}{X}$$ is classified by a functor
\[eq:FXtopostoEtale\]$$\label{eq:EtaleOverX}{\mathcal}{X}^{op}\xrightarrow{\begin{array}{rcl}U&\mapsto& {\mathcal}{X}_{/U}\\(f:U\to V)&\mapsto &(f^*:{\mathcal}{X}_{/V}\leftrightarrows{\mathcal}{X}_{/U}:f_*)\end{array}} {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t},$$ which factors as $$\label{eq:EtaleOverXEquiv}{\mathcal}{X}^{op}\xrightarrow{\cong} ({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})_{{\mathcal}{X}/}\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t},$$
where the first functor is an equivalence of categories.
Given a functor $F:{{\mathscr{LT}\mathrm{op}}}\to {\mathcal}{C}$, let $F_{{\mathcal}{X}}:{\mathcal}{X}^{op}\to {\mathcal}{C}$ denote the composite $${\mathcal}{X}^{op}\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\subseteq {{\mathscr{LT}\mathrm{op}}}\xrightarrow{F}{\mathcal}{C}.$$ We say that $F$ is a *sheaf* if for every $\infty$-topos ${\mathcal}{X}$, the composite functor $F_{{\mathcal}{X}}$ preserves small limits.
\[thm:CSSaSheaf\] The functor
$$\label{eq:CSSetFun}{CSS}_k:{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\xrightarrow{{\mathcal}{X}\mapsto {CSS}_k({\mathcal}{X})}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$$
classifying $\int^{\acute{e}t}{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ preserves small limits. In particular, $$\label{eq:CSSFun}{CSS}_k:{{\mathscr{LT}\mathrm{op}}}\xrightarrow{{\mathcal}{X}\mapsto {CSS}_k({\mathcal}{X})}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$$
is a sheaf.
By definition, for any étale geometric morphism $f^*:{\mathcal}{X}\rightleftarrows {\mathcal}{X}_{/U}:f_*$, the canonical projection $f_!:{\mathcal}{X}_{/U}\to{\mathcal}{X}$ forms part of an adjoint triple, $$(f_!\dashv f^*\dashv f_*):{\mathcal}{X}_{/U}\underset{\xrightarrow{f_*}}{\overset{\xrightarrow{f_!}}{\xleftarrow{f^*}}}{\mathcal}{X}.$$ Moreover, the forgetful functor from the over category $f_!:{\mathcal}{X}_{/U}\to {\mathcal}{X}$ preserves pullbacks, so that $f_!:{\mathcal}{X}_{/U}\leftrightarrows{\mathcal}{X}:f^*$ is a *pseudo-geometric morphism* (cf. [@Haugseng:2014vw]). In particular, $(f^*)_!:{\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X})\to {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{\mathcal}{X}_{/U})$ preserves complete $k$-fold Segal objects (cf. [@Haugseng:2014vw Proposition 2.20]); and consequently the inclusion $$\begin{tikzpicture}
\node (m-1-1) at (-1.5,1.5) {$\int^{\acute{e}t}{CSS}_k$};
\node (m-1-2) at (1.5,1.5) {$k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})$};
\node (m-2-2) at (0,0) {${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-2);
\draw[->] (m-1-2) -- (m-2-2);
\end{tikzpicture}$$ preserves both Cartesian and coCartesian edges. Therefore,
- ${CSS}_k:{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\to \widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ is a (fully faithful) subfunctor of the composite functor $$k\textrm{-Simpl}':{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\hookrightarrow\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}\xrightarrow{{\mathcal}{C}\mapsto {\operatorname}{Fun}(({\mathbb{\Delta}}^k)^{op},{\mathcal}{C})}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty},$$ classifying $k\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$, and
- this latter functor is continuous (cf. [@Lurie:2009un Proposition 6.3.2.3, Theorem 6.3.5.13]).
We will leverage these facts to show that ${CSS}_k\rvert_{{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}}$ is continuous. For simplicity of exposition, we restrict to the case that $k=1$.
Suppose that $q:I\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ is a diagram. Then we may identify the limit of $1\textrm{-Simpl}'\circ q$ with the $\infty$-category $$\lim \big(1\textrm{-Simpl}'\circ q\big)\subset {\operatorname}{Fun}_I\bigg(I, q^*\big(1\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\big)\bigg)$$ of coCartesian sections of the pulled-back presentable fibration $q^*\big(1\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\big)\to I$. Similarly, we may identify the limit of ${CSS}_1\circ q$ with the $\infty$-category $$\lim \big({CSS}_1\circ q\big)\subset {\operatorname}{Fun}_I\bigg(I, q^*\big(\int^{\acute{e}t}{CSS}_1\big)\bigg)$$ of coCartesian sections of the pulled-back presentable fibration $q^*\big(\int^{\acute{e}t}{CSS}_1)\big)\to I$.
Now let ${\mathcal}{X}\cong\lim q\in {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ be the limit of $q$. Then ${CSS}_1({\mathcal}{X})$ is the accessible localization of ${\operatorname}{Fun}\big(({\mathbb{\Delta}})^{op},{\mathcal}{X})\cong\lim \big(1\textrm{-Simpl}'\circ q\big)$ spanned by those objects which satisfy
1. the Segal conditions which specify the category objects,[^12] and
2. the completeness conditions; namely (in the case that $k=1$) that ${\operatorname}{Gp} X_\bullet:{\mathbb{\Delta}}^{op}\to {\mathcal}{X}$ is equivalent to the constant functor.
So we have full and faithful inclusions of both ${CSS}_1({\mathcal}{X})$ and $\lim({CSS}_1\circ q)$ into the $\infty$-category, ${\operatorname}{Fun}\big(({\mathbb{\Delta}})^{op},{\mathcal}{X})$, of co-Cartesian sections of $q^*\big(1\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\big)\to I$. Using the universal property for the limit yields a diagram of full and faithful inclusions: $${CSS}_1({\mathcal}{X})\hookrightarrow\lim({CSS}_1\circ q)\hookrightarrow{\operatorname}{Fun}\big(({\mathbb{\Delta}})^{op},{\mathcal}{X}).$$ Thus, it suffices to show that any coCartesian section of $q^*\big(\int^{\acute{e}t}{CSS}_1)\big)\to I$ lies in the essential image of the leftmost functor - i.e. satisfies conditions (1) and (2). As a first step, notice that $q^*\big(\int^{\acute{e}t}{CSS}_1)\big)\to I$ satisfies conditions (1) and (2) fibrewise.
For every $i\in I$, let $\pi_i^*:{\mathcal}{X}\leftrightarrows q(i):{\pi_i}_*$ denote the étale geometric morphism fitting into the limit cone. Recall that left adjoints of étale geometric morphisms $f^*:{\mathcal}{Y}\to {\mathcal}{Z}$ are continuous. Now, since the conditions for a simplicial object $X_{\bullet}\in {\operatorname}{Fun}\big(({\mathbb{\Delta}})^{op},{\mathcal}{X})$ to be a category object are given in terms of limits, Theorem \[thm:LimInLim\] implies that $X_{\bullet}$ is a category object if and only of each of the simplicial objects $(\pi_i^*)_!\big(X_{\bullet}\big)$ are category objects. Next, [@Haugseng:2014vw Proposition 2.20] implies that left adjoints of étale geometric morphisms $f^*:{\mathcal}{Y}\to {\mathcal}{Z}$ commute with the *underlying groupoid* functors, i.e. $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\operatorname}{Seg}({\mathcal}{Y})&{\operatorname}{Seg}({\mathcal}{Z})\\
{\operatorname}{Gpd}({\mathcal}{Y})&{\operatorname}{Gpd}({\mathcal}{Z})\\};
\draw[->] (m-1-1) -- node {$(f^*)_!$} (m-1-2);
\draw[->] (m-1-1) -- node[swap] {${\operatorname}{Gp}$} (m-2-1);
\draw[->] (m-1-2) -- node {${\operatorname}{Gp}$} (m-2-2);
\draw[->] (m-2-1) -- node[swap] {$(f^*)_!$} (m-2-2);
\end{tikzpicture}$$ commutes. Suppose now that $X_{\bullet}\in{\operatorname}{Seg}({\mathcal}{X})$ is a category object, which we may identify with a coCartesian section $X_\bullet':I\to q^*\big(1\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\big)$. Applying Theorem \[thm:limadj\], we see that the underlying groupoid ${\operatorname}{Gp}X_{\bullet}$ can be identified with the coCartesian section ${\operatorname}{Gp}\circ X_\bullet':I\to q^*\big(1\textrm{-Simpl}({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})\big)$ obtained by applying the underlying groupoid functor fibrewise.[^13] Consequently, for every $i\in I$ we have $$(\pi_i^*)_!{\operatorname}{Gp}X_{\bullet}\cong {\operatorname}{Gp}\big((\pi_i^*)_!X_{\bullet}\big).$$ Therefore $${\operatorname}{Gp}X_{\bullet}:{\mathbb{\Delta}}^{op}\to {\mathcal}{X}$$ is essentially constant if and only if each $${\operatorname}{Gp}\big((\pi_i^*)_!X_{\bullet}\big):{\mathbb{\Delta}}^{op}\to q(i)$$ is essentially constant.
Thus, we have shown ${CSS}_1({\mathcal}{X})\hookrightarrow\lim \big({CSS}_1\circ q\big)$ is an equivalence, which proves that preserves small limits.
Now for any $\infty$-topos ${\mathcal}{X}$, the functor ${\mathcal}{X}^{op}\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ given by factors as an equivalence followed by the forgetful functor from an undercategory ; hence it preserves small limits (cf. [@Lurie:2009un Proposition 1.2.13.8]). It follows that the composite $${\mathcal}{X}^{op}\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\xrightarrow{{CSS}_k}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$$ also preserves small limits, so is a sheaf.
$\infty$-categories of spans.
=============================
Let ${\mathcal}{C}$ be an $\infty$-category with pullbacks. In [@Barwick:2013th], Barwick introduces the $\infty$-category ${\operatorname}{Span}({\mathcal}{C})$, which has the same space of objects as ${\mathcal}{C}$, but whose morphisms between two objects $c_0,c_1\in {\mathcal}{C}$ is the space of diagrams in ${\mathcal}{C}$ of the form $$\begin{tikzpicture}
\node (m-1-1) at (0,1) {$x$};
\node (m-1-2) at (1,0) {$c_0$};
\node (m-2-1) at (-1,0) {$c_1$};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\end{tikzpicture}$$ That is, spans $c_0\nrightarrow c_1$ in ${\mathcal}{C}$. Composition of two such morphisms is given by taking the fibred product. Haugseng [@Haugseng:2014vw] extends this construction, introducing an $(\infty,k)$-category ${\operatorname}{Span}_k({\mathcal}{C})$ of iterated spans in ${\mathcal}{C}$, whose 2-morphisms are spans between spans, and so forth. In this section, we show that the functor ${\mathcal}{C}\to {\operatorname}{Span}_k({\mathcal}{C})$ depends continuously on ${\mathcal}{C}$.
Continuity of the formation of $\infty$-categories of iterated spans.
---------------------------------------------------------------------
We now briefly recall Haugseng’s construction. Let ${\mathbb{\Sigma}}^n$ denote the partially ordered set whose objects are pairs of numbers $(i,j)$ such that $0\leq i\leq j\leq n$, and $(i,j)\leq (i',j')$ if $i\leq i'$ and $j'\leq j$. We may picture the poset ${\mathbb{\Sigma}}^n$ (using Barwick’s notation $\bar p=n-p$) as follows:
For any map of totally ordered sets $\phi:[n]\to [m]$, the map $(i,j)\to \big(\phi(i),\phi(j)\big)$ induces a monotone map ${\mathbb{\Sigma}}^n\to{\mathbb{\Sigma}}^m$; and thus we have a functor ${\mathbb{\Sigma}}^\bullet:{\mathbb{\Delta}}\to{{\mathscr{C}\mathrm{at}}}_\infty$. Similarly, taking $k$-fold product, ${\mathbb{\Sigma}}^{n_1,\dots,n_k}:={\mathbb{\Sigma}}^{n_1}\times\cdots\times{\mathbb{\Sigma}}^{n_k}$ defines a functor $$\label{eq:BSigFunct}{\mathbb{\Sigma}}^{\bullet,\dots,\bullet}:{\mathbb{\Delta}}^k\to{{\mathscr{C}\mathrm{at}}}_\infty.$$
Suppose that ${\mathcal}{C}$ is an $\infty$-category with finite limits. We will be interested in functors $f:{\mathbb{\Sigma}}^{n_1,\dots,n_k}\to {\mathcal}{C}$. We let ${\mathbb{\Lambda}}^k\subseteq {\mathbb{\Sigma}}^k$ denote the full subcategory
spanned by those pairs $(i,j)$ with $j-i\leq1$. Similarly, we define ${\mathbb{\Lambda}}^{n_1,\dots,n_k}:={\mathbb{\Lambda}}^{n_1}\times\cdots\times{\mathbb{\Lambda}}^{n_k}$, and let $\iota_{n_1,\dots,n_k}:{\mathbb{\Lambda}}^{n_1,\dots,n_k}\to {\mathbb{\Sigma}}^{n_1,\dots,n_k}$ denote the inclusion.
\[def:CartFunct\] We say that a functor $f:{\mathbb{\Sigma}}^{n_1,\dots,n_k}\to {\mathcal}{C}$ is *Cartesian* if it is a right Kan extension of $f\circ \iota_{n_1,\dots,n_k}$, and we let ${\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)\subseteq {\operatorname}{Fun}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)$ denote the full subcategory spanned by the Cartesian functors. We let $${\operatorname}{Map}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big):=\iota{\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)\subset {\operatorname}{Fun}^{Cart}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)$$ denote the classifying space of Cartesian functors.[^14]
For example, when $k=1$, a Cartesian functor $f:{\mathbb{\Sigma}}^{n_1,\dots,n_k}\to {\mathcal}{C}$ is a diagram of the form
where each square is a pullback in ${\mathcal}{C}$. Such a diagram is to be understood as a composable sequence of spans $$c_{00}\overset{c_{01}}{\nrightarrow}c_{11}\overset{c_{12}}{\nrightarrow}c_{22}\nrightarrow\cdots\nrightarrow c_{nn}$$ where for $i<j<k$ each $c_{ik}$ is the composite (fibre product) of $c_{ii}\overset{c_{ij}}{\nrightarrow}c_{jj}\overset{c_{jk}}{\nrightarrow}c_{kk}$. Recall that ${{\mathscr{C}\mathrm{at}}}_\infty$ is a Cartesian closed $\infty$-category, in particular, there is an internal mapping object bi-functor (cf. [@Lurie:0uBkkKfz Remark 4.2.1.31]), $$\label{eq:2Yon}{{\mathscr{C}\mathrm{at}}}_\infty^{op}\times{{\mathscr{C}\mathrm{at}}}_\infty\xrightarrow{({\mathcal}{D},{\mathcal}{C})\mapsto {\operatorname}{Fun}({\mathcal}{D},{\mathcal}{C})}{{\mathscr{C}\mathrm{at}}}_\infty,$$ which is separately continuous in either variable. Composing with ${\mathbb{\Sigma}}^{\bullet,\dots,\bullet}:{\mathbb{\Delta}}^k\to{{\mathscr{C}\mathrm{at}}}_\infty$ yields a functor $({\mathbb{\Delta}}^k)^{op}\times{{\mathscr{C}\mathrm{at}}}_\infty\to{{\mathscr{C}\mathrm{at}}}_\infty,$ or equivalently, a functor
$$\label{eq:barSPAN+}
\begin{split}
\overline{{\operatorname}{SPAN}}_k^+:{{\mathscr{C}\mathrm{at}}}_\infty&\to{\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{{\mathscr{C}\mathrm{at}}}_\infty\big),\\
{\mathcal}{C}&\to \bigg[(n_1,\dots,n_k)\to {\operatorname}{Fun}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)\bigg]
\end{split}$$
which is continuous (by [@Lurie:2009un Corollary 5.1.2.3] and the continuity of in the second variable).
Let ${{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}\subset {{\mathscr{C}\mathrm{at}}}_\infty$ consist of those $\infty$-categories with finite limits and functors preserving finite limits. Suppose that ${\mathcal}{C}\in{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}$ has finite limits, and $f:{\mathbb{\Sigma}}^{n_1\dots,n_k}\to {\mathcal}{C}$ is Cartesian (in the sense of Definition \[def:CartFunct\]); then for any finite limit preserving functor $F:{\mathcal}{C}\to{\mathcal}{D}$, the composite $F\circ f:{\mathbb{\Sigma}}^{n_1\dots,n_k}\to {\mathcal}{D}$ is also Cartesian. Therefore, following [@Haugseng:2014vw], we may define $$\label{eq:SPAN+}
\begin{split}
{\operatorname}{SPAN}_k^+:{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}&\to{\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{{\mathscr{C}\mathrm{at}}}_\infty\big),\\
{\mathcal}{C}&\to \bigg[(n_1,\dots,n_k)\to {\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)\bigg]
\end{split}$$
to be the subfunctor of $\overline{{\operatorname}{SPAN}}_k^+\rvert_{{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}}$ which assigns to each ${\mathcal}{C}\in{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}$ and each $(n_1,\dots,n_k)\in{\mathbb{\Delta}}^k$ the full subcategory spanned by the Cartesian functors ${\mathbb{\Sigma}}^{n_1\dots,n_k}\to {\mathcal}{C}$. As explained in [@Haugseng:2014vw] the functor takes values in $k$-uple category objects (see also [@Barwick:2013th]).
\[lem:SPAN+Cont\] The functor $${\operatorname}{SPAN}_k^+:{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}\to{\operatorname}{Cat}^k({{\mathscr{C}\mathrm{at}}}_\infty)$$ is continuous.
Following Lemma \[lem:ContOfCatObj\], we need only show that the composite
$$\label{eq:UnderlyMorph}{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}\xrightarrow{{\operatorname}{SPAN}_k^+}{\operatorname}{Cat}^k({{\mathscr{C}\mathrm{at}}}_\infty) \xrightarrow{i^*} {\operatorname}{Fun}\big(({\mathbb}{Morph}^k)^{op},{{\mathscr{C}\mathrm{at}}}_\infty\big)$$
is continuous, where $i:{\mathbb}{Morph}^k\to {\mathbb{\Delta}}^k$ is as in Lemma \[lem:ContOfCatObj\]. But is equivalent to the composite $$\label{eq:UnderlyMorph2}{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}\hookrightarrow{{\mathscr{C}\mathrm{at}}}_\infty\xrightarrow{\overline{{\operatorname}{SPAN}}_k^+}{\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},{{\mathscr{C}\mathrm{at}}}_\infty\big) \xrightarrow{i^*} {\operatorname}{Fun}\big(({\mathbb}{Morph}^k)^{op},{{\mathscr{C}\mathrm{at}}}_\infty\big).$$
The first arrow in is continuous by [@Riehl:2014ut] (or Theorem \[thm:LimInLim\]), the second arrow is continuous since ${{\mathscr{C}\mathrm{at}}}_\infty$ is Cartesian closed (as explained above), and the final arrow is continuous by [@Lurie:2009un Corollary 5.1.2.3].
Next, let $\iota: {{\mathscr{C}\mathrm{at}}}_\infty\to{\mathscr{S}}$ denote the right adjoint to the inclusion, which sends an $\infty$-category ${\mathcal}{C}$ to its classifying space of objects, the largest Kan complex contained in ${\mathcal}{C}$. Then as in [@Haugseng:2014vw], we define $$\label{eq:SPAN}
\begin{split}
{\operatorname}{SPAN}_k:{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}&\to{\operatorname}{Cat}^k({\mathscr{S}}),\\
{\mathcal}{C}&\to \bigg[(n_1,\dots,n_k)\to {\operatorname}{Map}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)\bigg]
\end{split}$$ to be the composite $\iota\circ{\operatorname}{SPAN}_k^+$.
Finally let $U_{{\operatorname}{Seg}}:{\operatorname}{Cat}^k({\mathscr{S}})\to{\operatorname}{Seg}_k({\mathscr{S}})$ denote a right adjoint to ${\operatorname}{Seg}_k({\mathscr{S}})\hookrightarrow{\operatorname}{Cat}^k({\mathscr{S}})$. Then ${\operatorname}{Span}_k:=U_{{\operatorname}{Seg}}\circ {\operatorname}{SPAN}_k$ takes values in complete Segal spaces [@Haugseng:2014vw Corollary 3.18].
The functor $${\operatorname}{Span}_k:{{\mathscr{C}\mathrm{at}}}_\infty^{{\operatorname}{lex}}\to{CSS}_k({\mathscr{S}})$$ is continuous.
${\operatorname}{Span}_k$ is the composite $U_{{\operatorname}{Seg}}\circ \iota\circ{\operatorname}{SPAN}_k^+$, the first two functors are continuous (since they are right adjoints), and the last functor is continuous by Lemma \[lem:SPAN+Cont\].
Let ${\mathcal}{K}$ be the subcategory inclusions $i_{n_1,\dots,n_k}:{\mathbb{\Lambda}}^{n_1,\dots,n_k}\to {\mathbb{\Sigma}}^{n_1,\dots,n_k}$ used in the definition of a Cartesian functor (cf. Definition \[def:CartFunct\]). Let ${{\mathscr{C}\mathrm{at}}}_\infty^{\mathcal}{K}\subset {{\mathscr{C}\mathrm{at}}}_\infty$ denote the subcategory consisting of $\infty$-categories which admit all right Kan extensions along any $i_{n_1,\dots,n_k}:{\mathbb{\Lambda}}^{n_1,\dots,n_k}\to {\mathbb{\Sigma}}^{n_1,\dots,n_k}$ and of functors which preserve those right Kan extensions. Then ${{\mathscr{C}\mathrm{at}}}_\infty^{\mathcal}{K}$ is the maximal subcategory of ${{\mathscr{C}\mathrm{at}}}_\infty$ on which the functor ${\operatorname}{SPAN}_k^+$ may be defined. As a consequence of Corollary \[cor:KanClosed\], each of the functors $$\begin{aligned}
{\operatorname}{SPAN}_k^+:{{\mathscr{C}\mathrm{at}}}_\infty^{{\mathcal}{K}}&\to{\operatorname}{Cat}^k({{\mathscr{C}\mathrm{at}}}_\infty)\\
{\operatorname}{SPAN}_k:{{\mathscr{C}\mathrm{at}}}_\infty^{{\mathcal}{K}}&\to{\operatorname}{Cat}^k({\mathscr{S}})\\
{\operatorname}{Span}_k:{{\mathscr{C}\mathrm{at}}}_\infty^{{\mathcal}{K}}&\to{CSS}_k({\mathscr{S}})
\end{aligned}$$ are continuous.
The sheaf of iterated spans with local systems.
-----------------------------------------------
Suppose that ${\mathcal}{X}$ is an $\infty$-topos, and $X_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{X})$ is a complete $k$-fold Segal object in ${\mathcal}{X}$. In [@Haugseng:2014vw], Haugseng gave an elegant construction of the $(\infty,k)$-category $${\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})$$ of iterated $k$-fold spans in ${\mathcal}{X}$ with local systems valued in $X_{\bullet,\dots,\bullet}$,
- whose objects are objects in ${\mathcal}{X}$ equipped with a map to the objects of the local system, $X_{\bullet,\dots,\bullet}$,
- whose morphisms are spans in ${\mathcal}{X}$ equipped with compatible maps to the space of morphisms of the local system, $X_{\bullet,\dots,\bullet}$,
- $\dots$,
- and whose $i$-morphisms are $i$-fold spans in ${\mathcal}{X}$ equipped with compatible maps to the space of $i$-morphisms of the local system, $X_{\bullet,\dots,\bullet}$.
In this section, we show that for any continuous functor $\sigma:{\mathcal}{X}^{op}\to \int{CSS}_k$ over $${\mathcal}{X}^{op}\xrightarrow{U\mapsto {\mathcal}{X}_{/U}}{{\mathscr{LT}\mathrm{op}}},$$ the functor $${\mathcal}{X}^{op}\xrightarrow{U\mapsto {\operatorname}{Span}_k({\mathcal}{X}_{/U},\sigma(U))}{CSS}_k(\hat{\mathscr{S}})$$ forms an $(\infty,k)$-stack over ${\mathcal}{X}$.
We begin by describing the functor $$\int{CSS}_k\xrightarrow{({\mathcal}{X}, X_{\bullet,\dots,\bullet})\mapsto {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$ in more detail.
As in [@Haugseng:2014vw], we let $\hat{\mathbb{\Sigma}}\xrightarrow{q} {\mathbb{\Delta}}^{op}$ denote the Grothendieck fibration classified by the functor ${\mathbb{\Sigma}}^\bullet:{\mathbb{\Delta}}\xrightarrow{[n]\mapsto {\mathbb{\Sigma}}^n} {{\mathscr{C}\mathrm{at}}}$, whose objects are pairs $\big([n],(i,j)\big)$ with $0\leq i\leq j\leq n$, and whose morphisms $$\label{eq:hatBSig}\big([n],(i,j)\big)\xrightarrow{\begin{array}{rcl}[n]&\leftarrow &[m]:\phi\\(i,j)&\rightarrow &\big(\phi(i'),\phi(j')\big)\end{array}}\big([m],(i',j')\big)$$ are pairs of morphisms $\phi:[m]\to[n]$ in ${\mathbb{\Delta}}$ and $(i,j)\to \big(\phi(i'),\phi(j')\big)$ in ${\mathbb{\Sigma}}^n$.
Let $\imath^*{\mathcal}{Z}\to {{\mathscr{LT}\mathrm{op}}}$ denote the (cannonical) presentable fibration classified by the inclusion $\imath:{{\mathscr{LT}\mathrm{op}}}\hookrightarrow\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$. The functor $\overline{{\operatorname}{SPAN}}^+_k:{{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}\to \widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ classifies the coCartesian fibration (cf. [@Lurie:2009un Corollary 3.2.2.13]) $$(\imath^*{\mathcal}{Z}\times({\mathbb{\Delta}}^k)^{op})^{{{\mathscr{LT}\mathrm{op}}}\times \hat{\mathbb{\Sigma}}^k}\to {{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op},$$ whose fibre over any $\big({\mathcal}{X};(n_1,\dots,n_k)\big)\in {{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}$ is equivalent to $${\operatorname}{Fun}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{X}\big).$$ Similarly, ${\operatorname}{SPAN}^+_k:{{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}\to \widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ classifies the coCartesian fibration defined as the full subcategory $${\int{\operatorname}{SPAN}^+_k}\subset (\imath^*{\mathcal}{Z}\times({\mathbb{\Delta}}^k)^{op})^{{{\mathscr{LT}\mathrm{op}}}\times \hat{\mathbb{\Sigma}}^k}$$ spanned (over $\big({\mathcal}{X};(n_1,\dots,n_k)\big)\in {{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}$) by the Cartesian functors ${\mathbb{\Sigma}}^{n_1,\dots,n_k}\to {\mathcal}{X}$ (cf. Definition \[def:CartFunct\]).
There is a second functor $$\Pi:\hat{\mathbb{\Sigma}}\xrightarrow{\big([n],(i,j)\big)\mapsto [j-i]} {\mathbb{\Delta}}^{op}$$ which sends the map to $$[j-i]\xrightarrow{k\mapsto\phi(k+i')-i}[j'-i'].$$
The corresponding morphism of Cartesian fibrations $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathbb{\Delta}}^{op}\times{\mathbb{\Delta}}^{op}&&\hat{\mathbb{\Sigma}}\\&{\mathbb{\Delta}}^{op}&\\};
\draw[<-] (m-1-1) -- node[swap]{ $\Pi\times q$} (m-1-3);
\draw[->] (m-1-1) -- (m-2-2);
\draw[->] (m-1-3) -- (m-2-2);
\end{tikzpicture}$$ induces a morphism of coCartesian fibrations $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{(\imath^*{\mathcal}{Z}\times({\mathbb{\Delta}}^k)^{op})^{{{\mathscr{LT}\mathrm{op}}}\times ({\mathbb{\Delta}}^k)^{op}\times ({\mathbb{\Delta}}^k)^{op}}&&(\imath^*{\mathcal}{Z}\times({\mathbb{\Delta}}^k)^{op})^{{{\mathscr{LT}\mathrm{op}}}\times \hat{\mathbb{\Sigma}}^k}\\
&{{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}&\\};
\draw[->] (m-1-1) -- (m-1-3);
\draw[->] (m-1-1) -- (m-2-2);
\draw[->] (m-1-3) -- (m-2-2);
\end{tikzpicture}$$
By [@Haugseng:2014vw Lemma 4.3] this restricts to a morphism: $$\label{eq:sectS}\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{\bigg({\int{CSS}_k}\bigg)\times({\mathbb{\Delta}}^k)^{op}&&{\int{\operatorname}{SPAN}^+_k}\\
&{{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}&\\};
\draw[->] (m-1-1) -- node {$s_0$} (m-1-3);
\draw[->] (m-1-1) -- (m-2-2);
\draw[->] (m-1-3) -- (m-2-2);
\end{tikzpicture}$$ which sends any $\big({\mathcal}{X};(n_1,\dots,n_k)\big)\in {{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}$ to $$X_{\bullet,\dots,\bullet}\circ \Pi_{n_1,\dots,n_k}\in {\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{X}),$$ where $\Pi_{n_1,\dots,n_k}:=\Pi\rvert_{{\mathbb{\Sigma}}^{n_1,\dots,n_k}}$.
In turn, defines a section of the left hand arrow in the pullback square $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{Q} & {\int{\operatorname}{SPAN}^+_k}\\
\bigg({\int{CSS}_k}\bigg)\times({\mathbb{\Delta}}^k)^{op} & {{\mathscr{LT}\mathrm{op}}}\times({\mathbb{\Delta}}^k)^{op}\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\draw[dotted,->] (m-2-1) to[bend left] node{$s$}(m-1-1);
\end{tikzpicture}$$ For brevity, we denote ${\mathcal}{D}=\big({\int{CSS}_k}\big)\times({\mathbb{\Delta}}^k)^{op}$, and we define ${\mathcal}{Q}^{/s}\to {\mathcal}{D}\cong\big({\int{CSS}_k}\big)\times({\mathbb{\Delta}}^k)^{op}$ to be the simplicial set satisfying the universal property that for any morphism of simplicial sets $Y\to {\mathcal}{D}$, commutative diagrams of the form $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{D} &{\mathcal}{Q}\\
Y\diamond_{{\mathcal}{D}}{\mathcal}{D}& {\mathcal}{D}\\};
\draw[->] (m-1-1) -- node {$s$} (m-1-2);
\draw[right hook->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\draw[->] (m-2-1) -- (m-1-2);
\end{tikzpicture}$$ (where $Y\diamond_{{\mathcal}{D}}{\mathcal}{D}=Y\times\Delta^1\coprod_{Y\times\{1\}}{\mathcal}{D}$) correspond to diagrams of the form $$\begin{tikzpicture}\node (y) at (-2,1) {$Y$};
\node (q) at (2,1) {${\mathcal}{Q}^{/s}$};
\node (css) at (0,0) {${\mathcal}{D}$};
\draw[->] (y) -- (q);
\draw[->] (y) -- (css);
\draw[->] (q) -- (css);
\end{tikzpicture}$$
By [@Lurie:2009un Proposition 4.2.2.4.] ${\mathcal}{Q}^{/s}\to \big({\int{CSS}_k}\big)\times({\mathbb{\Delta}}^k)^{op}$ is a coCartesian fibration whose fibre over $\big({\mathcal}{X},X_{\bullet,\dots,\bullet};(n_1,\dots,n_k)\big)\in {\int{CSS}_k}\times({\mathbb{\Delta}}^k)^{op}$ is equivalent to the overcategory $${\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{X})_{/X_{\bullet,\dots,\bullet}\circ \Pi_{n_1,\dots,n_k}}$$
Let $$({\operatorname}{SPAN}^+_k)^{/s}:{\int{CSS}_k}\to {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}\big)$$ be a functor classifying ${\mathcal}{Q}^{/s}$. By, [@Haugseng:2014vw Proposition 4.5], $({\operatorname}{SPAN}^+_k)^{/s}$ takes values in $k$-uple category objects in $\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$. We define $${\operatorname}{Span}_k:=U_{{\operatorname}{Seg}}\circ \iota\circ ({\operatorname}{SPAN}^+_k)^{/s}:{\int{CSS}_k}\to {\operatorname}{Seg}_k(\widehat{{\mathscr{S}}}).$$ It follows from [@Haugseng:2014vw Proposition 4.8] that ${\operatorname}{Span}_k$ takes values in complete $k$-fold Segal spaces, i.e. we have a functor $$\label{eq:locSpan}{\operatorname}{Span}_k:{\int{CSS}_k}\to {CSS}_k(\widehat{{\mathscr{S}}})$$ which sends a complete $k$-fold Segal object $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$ to the complete $k$-fold Segal space $${\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})$$ of iterated spans in ${\mathcal}{X}$ with local systems valued in $X_{\bullet,\dots,\bullet}$ (cf. [@Haugseng:2014vw § 4]).
### Continuity of $({\mathcal}{X},X_{\bullet,\dots,\bullet})\to {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})$.
\[lem:SpanProduct\] The functor preserves small products.
Since $\int^{\acute{e}t}{CSS}_k\subseteq\int{CSS}_k$ is a continuous inclusion of a wide subcategory (i.e. it contains all the objects), it suffices to show that the restriction of to $\int^{\acute{e}t}{CSS}_k$ preserves small products.
Now suppose $\{X^j_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{X}^j)\}_{j\in J}$ is a set of complete Segal objects indexed by a small set $J$. Since $p:\int^{\acute{e}t}{CSS}_k\to {{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ is a presentable fibration and ${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ has small products, we may compute the product $$\prod_{j\in J}({\mathcal}{X}^j,X^j_{\bullet,\dots,\bullet})\in \int{CSS}_k$$ by first computing the product $\prod_{j\in J}{\mathcal}{X}^j$ in ${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$, and then computing the $p$-relative product of $\{X^j_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{X}^j)\}_{j\in J}$ over $\prod_{j\in J}{\mathcal}{X}^j$.
Note that [@Lurie:2009un Proposition 6.3.2.3 and Theorem 6.3.5.13] imply that $\prod_{j\in J}{\mathcal}{X}^j$ is just the product of the $\infty$-categories ${\mathcal}{X}^j$ (i.e. we can take this product in $\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ rather than ${{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$). Next, Theorem \[thm:CSSaSheaf\] implies that the fibre of $\int^{\acute{e}t}{CSS}_k$ over $\prod_{j\in J} {\mathcal}{X}^j$ is just $${CSS}_k(\prod_{j\in J} {\mathcal}{X}^j)\cong \prod_{j\in J} {CSS}_k({\mathcal}{X}^j),$$ where the right hand product is taken in $\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$. Consequently, the $p$-relative product of $\{X^j_{\bullet,\dots,\bullet}\}_{j\in J}$ over $\prod_{j\in J} {\mathcal}{X}^j$ is $$\prod_{j\in J}X^j_{\bullet,\dots,\bullet}\in \prod_{j\in J} {CSS}_k({\mathcal}{X}^j).$$
Next we argue that the restriction of $({\operatorname}{SPAN}^+_k)^{/s}$ to $\int^{\acute{e}t}{CSS}_k$ preserves small products. In view of Lemma \[lem:ContOfCatObj\], we need only show that for any $0\leq i_1,\dots, i_k\leq 1$, the functor $$\label{eq:overCatSpan1Prod}\int^{\acute{e}t}{CSS}_k\xrightarrow{({\mathcal}{X},X_{\bullet,\dots,\bullet})\mapsto {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X})_{/X_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}.$$ preserves small products.
Notice that the continuous functor $$\prod_{j\in J} {CSS}_k({\mathcal}{X}^j)\cong {CSS}_k(\prod_{j\in J} {\mathcal}{X}^j)\hookrightarrow {\operatorname}{Fun}\big(({\mathbb{\Delta}}^k)^{op},\prod_{j\in J} {\mathcal}{X}^j)\xrightarrow{\Pi_{i_1,\dots,i_k}^*} {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},\prod_{j\in J} {\mathcal}{X}^j)$$ takes $\prod_{j\in J}X^j_{\bullet,\dots,\bullet}$ to $$\big(\prod_{j\in J}X^j_{\bullet,\dots,\bullet}\big)\circ \Pi_{i_1,\dots,i_k}\cong \prod_{j\in J}\big(X^j_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}\big).$$ So $$\begin{gathered}
{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},\prod_{j\in J}{\mathcal}{X}^j)_{/\big((\prod_{j\in J}X^j_{\bullet,\dots,\bullet})\circ \Pi_{i_1,\dots,i_k}\big)}\\\cong \big(\prod_{j\in J}{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X}^j)\big)_{/\big(\prod_{j\in J}(X^j_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k})\big)}\\
\cong \prod_{j\in J}\bigg({\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X}^j)_{/X^j_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}\bigg),
\end{gathered}$$ which implies that preserves small products.
Finally, we have ${\operatorname}{Span}_k= U_{{\operatorname}{Seg}}\circ\iota\circ ({\operatorname}{SPAN}^+_k)^{/s}$, and since $U_{{\operatorname}{Seg}}$ and $\iota$ are both right adjoints, they are continuous, which implies the statement we wished to prove.
\[lem:SpanPullback\] The functor preserves pullbacks.
We begin by arguing that $({\operatorname}{SPAN}^+_k)^{/s}:\int{CSS}_k\to \widehat{{{\mathscr{C}\mathrm{at}}}_\infty}$ preserves pullbacks. In view of Lemma \[lem:ContOfCatObj\], we need only show that for any $0\leq i_1,\dots, i_k\leq 1$, the functor $$\label{eq:overCatSpan1Pullback}\int{CSS}_k\xrightarrow{({\mathcal}{X},X_{\bullet,\dots,\bullet})\mapsto {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X})_{/X_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}}\widehat{{{\mathscr{C}\mathrm{at}}}_\infty}.$$ preserves pullbacks.
Suppose we have a diagram
$$\label{eq:CSSpullback}({\mathcal}{X},X_{\bullet,\dots,\bullet})\rightarrow ({\mathcal}{Z},Z_{\bullet,\dots,\bullet})\leftarrow({\mathcal}{Y},Y_{\bullet,\dots,\bullet})$$
in $\int{CSS}_k$ (here $X_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{X})$, $Y_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{Y})$, and $Z_{\bullet,\dots,\bullet}\in {CSS}_k({\mathcal}{Z})$). To compute the pullback of we use [@Lurie:2009un Corollary 4.3.1.11]. That is, we first compute the pullback $$\label{eq:CSSpullback1}
\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{W} &{\mathcal}{Y}\\ {\mathcal}{X}& {\mathcal}{Z}\\};
\draw[->] (m-1-1) -- node {$b^*$} (m-1-2);
\draw[->] (m-1-1) -- node[swap] {$a^*$} (m-2-1);
\draw[->] (m-1-2) -- node {$g^*$} (m-2-2);
\draw[->] (m-2-1) -- node[swap] {$f^*$} (m-2-2);
\draw[->] (m-1-1) -- node {$c^*$} (m-2-2);
\end{tikzpicture}$$ in ${{\mathscr{LT}\mathrm{op}}}$ and then take the relative pullback $$\label{eq:CSSpullback2}
\begin{tikzpicture}
\node (m-1-1) at (-4,2) {$\overset{W}{\overbrace{(a_*)_!X\times_{(c_*)_!Z}(b_*)_!Y}}$};
\node (m-1-2) at (-1,2) {$Y$};
\node (m-2-1) at (-4,0) {$X$};
\node (m-2-2) at (-1,0) {$Z$};
\node (m-1-1') at (2,2) {${CSS}_k({\mathcal}{W})$};
\node (m-1-2') at (5,2) {${CSS}_k({\mathcal}{Y})$};
\node (m-2-1') at (2,0) {${CSS}_k({\mathcal}{X})$};
\node (m-2-2') at (5,0) {${CSS}_k({\mathcal}{Z})$};
\node (in) at (.5,1) {$\in $};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\draw[->] (m-1-1') -- (m-1-2');
\draw[->] (m-1-1') -- (m-2-1');
\draw[->] (m-1-2') -- (m-2-2');
\draw[->] (m-2-1') -- (m-2-2');
\end{tikzpicture}$$
in the fibre ${CSS}_k({\mathcal}{W})$ over ${\mathcal}{W}$ (here we have dropped the abstract multi-indices on $X_{\bullet,\dots,\bullet}$, $Y_{\bullet,\dots,\bullet}$, and $Z_{\bullet,\dots,\bullet}$).
For any $\infty$-category ${\mathcal}{C}$, let
$$\label{eq:preEmb}({{\mathscr{C}\mathrm{at}}}_\infty)_{/{\mathcal}{C}}\hookleftarrow{\mathscr{P}}({\mathcal}{C})$$
denote the functor which sends a presheaf over ${\mathcal}{C}$ to the corresponding right fibration over ${\mathcal}{C}$. Then [@Lurie:2009un Corollary 2.1.2.10] implies that is equivalent to a reflective left localization of $({{\mathscr{C}\mathrm{at}}}_\infty)_{/{\mathcal}{C}}$; in particular is continuous (see also [@Gepner:2015ww Theorem 4.5]). Since the Yoneda embedding is continuous, and the forgetful functor $({{\mathscr{C}\mathrm{at}}}_\infty)_{/{\mathcal}{C}}\to{{\mathscr{C}\mathrm{at}}}_\infty$ preserves pullbacks, it follows that the composite $$\label{eq:pullbackPreservingYon}
\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=1em,
column sep=1em, text height=1.5ex, text depth=0.25ex]}{{\mathcal}{C}&{\mathscr{P}}({\mathcal}{C})&({{\mathscr{C}\mathrm{at}}}_\infty)_{/{\mathcal}{C}}&{{\mathscr{C}\mathrm{at}}}_\infty\\
c&&&{\mathcal}{C}_{/c}\\};
\draw[->] (m-1-1) -- node {${\!\text{\usefont{U}{min}{m}{n}\symbol{'210}}\!}$} (m-1-2);
\draw[right hook->] (m-1-2) -- (m-1-3);
\draw[->] (m-1-3) -- (m-1-4);
\draw[->] (m-2-1) -- (m-2-4);
\end{tikzpicture}$$
preserves pullbacks.
Applying to $W_{\bullet,\dots,\bullet}$ (the top left corner of ) and using the continuity of yields a pullback diagram $$\label{eq:pullbackSpan+}\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/W_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}} &{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/(b_*)_!Y_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}\\ {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/(a_*)_!X_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}& {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/(c_*)_!Z_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- (m-2-2);
\end{tikzpicture}$$ Now, since $$(a^*)_!:{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})\leftrightarrows{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X}):(a_*)_!$$ is an adjunction, we have a pullback square $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/(a_*)_!X_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}} &{\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X})_{/X_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}\\ {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})& {\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{X})\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- node {$(a^*)_!$} (m-2-2);
\end{tikzpicture}$$ Similarly, the right hand terms of fit into analagous pullback squares. It follows that $${\operatorname}{Fun}({\mathbb{\Sigma}}^{i_1,\dots,i_k},{\mathcal}{W})_{/W_{\bullet,\dots,\bullet}\circ \Pi_{i_1,\dots,i_k}}$$ fits into a limit diagram $$\label{eq:pullbackSpan+1}\begin{tikzpicture}
\node (m-1-1-1) at (-2,4) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{W})_{/W\circ \Pi_I}$};
\node (m-1-1-2) at (-4,2) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{X})_{/X\circ \Pi_I}$};
\node (m-1-2-1) at (4,4) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{Y})_{/Y\circ \Pi_I}$};
\node (m-1-2-2) at (2,2) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{Z})_{/Z\circ \Pi_I}$};
\node (m-2-1-1) at (-2,0) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{W})$};
\node (m-2-1-2) at (-4,-2) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{X})$};
\node (m-2-2-1) at (4,0) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{Y})$};
\node (m-2-2-2) at (2,-2) {${\operatorname}{Fun}({\mathbb{\Sigma}}^{I},{\mathcal}{Z})$};
\draw[->] (m-1-1-1) -- (m-1-1-2);
\draw[->] (m-1-1-1) -- (m-1-2-1);
\draw[->] (m-1-1-1) -- (m-2-1-1);
\draw[->] (m-1-1-2) -- (m-1-2-2);
\draw[->] (m-1-1-2) -- (m-2-1-2);
\draw[->] (m-1-2-1) -- (m-1-2-2);
\draw[->] (m-1-2-1) -- (m-2-2-1);
\draw[->] (m-2-1-1) -- node[swap] {$a^*_!$} (m-2-1-2);
\draw[->] (m-2-1-1) -- node {$b^*_!$} (m-2-2-1);
\draw[->] (m-1-2-2) -- (m-2-2-2);
\draw[->] (m-2-1-2) -- node[swap] {$f^*_!$}(m-2-2-2);
\draw[->] (m-2-2-1) -- node {$g^*_!$} (m-2-2-2);
\end{tikzpicture}$$ where we have abbreviated the multi-index $i_1,\dots,i_k=I$ and dropped the abstract multi-indices on $W_{\bullet,\dots,\bullet}$, $X_{\bullet,\dots,\bullet}$, $Y_{\bullet,\dots,\bullet}$, and $Z_{\bullet,\dots,\bullet}$.
Since the bottom square in is already a pullback square, it follows that the top square is also a pullback square, which proves that preserves pullbacks.
Finally, we have ${\operatorname}{Span}_k= U_{{\operatorname}{Seg}}\circ\iota\circ ({\operatorname}{SPAN}^+_k)^{/s}$, and since $U_{{\operatorname}{Seg}}$ and $\iota$ are both right adjoints, they are continuous, which implies the statement we wished to prove.
\[thm:SpanCont\] The functor , $$\int{CSS}_k\xrightarrow{({\mathcal}{X}, X_{\bullet,\dots,\bullet})\mapsto {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}}),$$ preserves small limits.
The inclusion of a fibre ${CSS}_k({\mathcal}{X})\hookrightarrow\int{CSS}_k$ doesn’t preserve products or terminal objects (though it does preserve small limits with connected diagrams). So the functor $${CSS}_k({\mathcal}{X})\xrightarrow{X_{\bullet,\dots,\bullet}\mapsto {\operatorname}{Span}_k({\mathcal}{X},X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$ is *not* continuous: while it does preserve small limits with connected diagrams, it generally fails to preserve products or terminal objects.
According to [@Lurie:2009un Proposition 4.4.2.7], it suffices to prove this result for pullbacks and small products; thus the result follows from Lemmas \[lem:SpanProduct\] and \[lem:SpanPullback\].
\[thm:StackofSpans\] Suppose that ${\mathcal}{X}$ is an $\infty$-topos and $\sigma:{\mathcal}{X}^{op}\to \int{CSS}_k$ is a continuous functor fitting into the diagram $$\begin{tikzpicture}
\node (x) at (-2,1) {${\mathcal}{X}^{op}$};
\node (css) at (2,1) {$\int{CSS}_k$};
\node (ltop) at (0,-1) {${{\mathscr{LT}\mathrm{op}}}$};
\draw[->] (x) -- node {$\sigma$} (css);
\draw[->] (x) -- (ltop);
\draw[->] (css) -- (ltop);
\end{tikzpicture}$$ where the left diagonal arrow ${\mathcal}{X}^{op}\xrightarrow{U\mapsto {\mathcal}{X}_{/U}}{{\mathscr{LT}\mathrm{op}}}$ is . Then $$\label{eq:SpanLocStack}{\operatorname}{Span}_k\circ\sigma:{\mathcal}{X}^{op}\xrightarrow{U\mapsto {\operatorname}{Span}_k({\mathcal}{X}_{/U},\sigma(U))}{CSS}_k(\hat{\mathscr{S}})$$ forms an $(\infty,k)$-stack over ${\mathcal}{X}$.
In particular, given any complete $k$-fold Segal space $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$, iterated spans in ${\mathcal}{X}$ with local systems valued in $X_{\bullet,\dots,\bullet}$ form an $(\infty,k)$-stack $$\label{eq:ItSpanLocSheaf}{\mathcal}{X}^{op}\xrightarrow{U\mapsto {\operatorname}{Span}_k({\mathcal}{X}_{/U},U\times X_{\bullet,\dots,\bullet})}{CSS}_k(\hat{\mathscr{S}})$$ over ${\mathcal}{X}$.
The first statement is equivalent to the continuity of , which follows directly from Theorem \[thm:SpanCont\].
Let $F:{\mathcal}{X}^{op}\to{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}$ be defined by . Of course, we have ${\mathcal}{X}^{op}\cong ({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})_{{\mathcal}{X}/}$, so $F^*\big(\int^{\acute{e}t}{CSS}_k\big)$ is equivalent to the pullback $$\begin{tikzpicture}
{\matrix (m) [matrix of math nodes, row sep=3em,
column sep=3em, text height=1.5ex, text depth=0.25ex]}{F^*\big(\int^{\acute{e}t}{CSS}_k\big)&\int^{\acute{e}t}{CSS}_k\\
({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})_{{\mathcal}{X}/}&{{\mathscr{LT}\mathrm{op}}}_{\acute{e}t}\\};
\draw[->] (m-1-1) -- (m-1-2);
\draw[->] (m-1-1) -- (m-2-1);
\draw[->] (m-1-2) -- (m-2-2);
\draw[->] (m-2-1) -- node[swap] {$F$} (m-2-2);
\end{tikzpicture}$$ Since ${\mathcal}{X}\in ({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})_{{\mathcal}{X}/}$ is an initial object, [@Lurie:2009un Proposition 3.3.3.1] and Theorem \[thm:CSSaSheaf\] imply that the $\infty$-category of coCartesian sections of $F^*\big(\int^{\acute{e}t}{CSS}_k\big)\to ({{\mathscr{LT}\mathrm{op}}}_{\acute{e}t})_{{\mathcal}{X}/}\cong {\mathcal}{X}^{op}$ is equivalent to ${CSS}_k({\mathcal}{X})$. In particular, any complete $k$-fold complete Segal object $X_{\bullet,\dots,\bullet}\in{CSS}_k({\mathcal}{X})$ determines a coCartesian section $\sigma:{\mathcal}{X}^{op}\to F^*\big(\int^{\acute{e}t}{CSS}_k\big)$. By [@Lurie:2009un Lemma 6.3.3.5], $\sigma$ is continuous, which implies that is an $(\infty,k)$-stack.
[^1]: More precisely, $\int{CSS}_k$ is the lax colimit of the functor ${\mathcal}{X}\mapsto {CSS}_k({\mathcal}{X})$, (cf. [@Gepner:2015ww]); equivalently, $\int{CSS}_k$ is Lurie’s unstraightening of that functor. We provide a direct construction of $\int{CSS}_k$, however.
[^2]: Paraphrased from [@Weinstein:1979vn] “everything is a Lagrangian submanifold”.
[^3]: following Theodore Johnson-Freyd’s suggestion, we denote the Yoneda embedding by the first Hiragana character - pronounced ‘yo’ - of his name, .
[^4]: Note that ${\operatorname}{Fun}_{K}(X_0,X_1)$ is denoted by ${\operatorname}{Map}_{K}^\flat(X_0^\flat,X_1^\natural)$ in [@Lurie:2009un], while ${\operatorname}{Fun}^{{\operatorname}{(co)Cart}}_K(X_0,X_1)$ is denoted by ${\operatorname}{Map}_{K}^\flat(X^\natural_0,X^\natural_1)$. We choose this alternate notation (in line with [@Gepner:2015ww]) to emphasize that the resulting simplicial set is an $\infty$-category.
[^5]: The terminal $\infty$-category, $\ast$, has exactly one object, one 1-morphism (the identity) and one $n$-morphism for every $n$.
[^6]: Recall that for any diagram $q:I\to {\mathcal}{C}$, the *undercategory* ${\mathcal}{C}^{q/}$ satisfies the universal property that the space of maps $Y\to {\mathcal}{C}^{q/}$ classifies maps of the form $I\diamond Y\to {\mathcal}{C}$ (which restrict to $q$ along $I$), where $$I\diamond Y= I\coprod_{I\times Y\times\{0\}} (I\times Y\times \Delta^1)\coprod_{I\times Y\times \{1\}} Y.$$
[^7]: following Theodore Johnson-Freyd’s suggestion, we denote the Yoneda embedding by the first Hiragana character of his name, .
[^8]: For two left-representable pairings $\lambda:{\mathcal}{M}\to {\mathcal}{C}\times\tilde{{\mathcal}{D}}^{op}$ and $\lambda':{\mathcal}{M}'\to {\mathcal}{C}'\times\tilde{{\mathcal}{D}}'^{op}$, the mapping space ${\operatorname}{Map}_{{\operatorname}{CPair}^L}({\mathcal}{M},{\mathcal}{M}')$ is the subspace of $${\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{C},{\mathcal}{C}')\times^h_{{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{M},{\mathcal}{C}')}{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{M},{\mathcal}{M}')\times^h_{{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}({\mathcal}{M},\tilde{{\mathcal}{D}}')}{\operatorname}{Map}_{{{\mathscr{C}\mathrm{at}}}_\infty}(\tilde{{\mathcal}{D}},\tilde{{\mathcal}{D}}')$$ which preserves left universal objects. Notice that the homotopy pullbacks can be taken to be strict pullbacks when $\lambda$ and $\lambda'$ are right fibrations.
[^9]: The objects of ${\mathbb{\Delta}}_{/K}$ over $[n]\in {\mathbb{\Delta}}$ are simplicial maps $\Delta^n\to K$ from the standard $n$-simplex, and morphisms in ${\mathbb{\Delta}}_{/K}$ over a morphism $f:[m]\to [n]$ are commutative diagrams $$\protect\tikz{
\node (m) at (-1,1) {$\Delta^m$};
\node (n) at (1,1) {$\Delta^n$};
\node (k) at (0,0) {$K$};
\draw[->] (m) -- node {$f$} (n);
\draw[->] (m) -- (k);
\draw[->] (n) -- (k);
}$$ Equivalently, $\big({\mathbb{\Delta}}_{/K}\to {\mathbb{\Delta}}\big)=\big(\int^{{\mathbb{\Delta}}} K\to {\mathbb{\Delta}}\big)$ is the Grothendieck fibration (or category of elements) associated the functor $K:{\mathbb{\Delta}}^{op}\to {\operatorname}{Sets}$.
[^10]: In fact, the model structure on simplicial sets for $\infty$-category is the Cisinski model structure induced by the localizer which consists of the inclusions ${\operatorname}{Sp}(n)\subseteq \Delta^n$ (cf. [@Ara:2012uj]).
[^11]: Recall that a fibration is presentable if it is both a Cartesian and a coCartesian fibration each of whose fibres are presentable $\infty$-categories.
[^12]: when $k> 1$, one also has constancy conditions , which are likewise given as limits.
[^13]: See [@Lurie:0uBkkKfz Proposition 7.3.2.6] to confirm that this right adjoint can be applied fibrewise in a coherent manner.
[^14]: i.e. the largest Kan complex in ${\operatorname}{Fun}^{{\mathbb{\Sigma}}\textrm{-Cart}}\big({\mathbb{\Sigma}}^{n_1,\dots,n_k},{\mathcal}{C}\big)$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Atlantic Meridional Overturning Circulation (AMOC) transports substantial amounts of heat into the North Atlantic sector, and hence is of very high importance in regional climate projections. The AMOC has been observed to show multi-stability across a range of models of different complexity. The simplest models find a bifurcation associated with the AMOC “on” state losing stability that is a saddle node. Here we study a physically derived global oceanic model of Wood [*et al*]{} with five boxes, that is calibrated to runs of the FAMOUS coupled atmosphere-ocean general circulation model. We find the loss of stability of the “on” state is due to a subcritical Hopf for parameters from both pre-industrial and doubled CO${}_2$ atmospheres. This loss of stability via subcritical Hopf bifurcation has important consequences for the behaviour of the basin of attraction close to bifurcation. We consider various time-dependent profiles of freshwater forcing to the system, and find that rate-induced thresholds for tipping can appear, even for perturbations that do not cross the bifurcation. Understanding how such state transitions occur is important in determining allowable safe climate change mitigation pathways to avoid collapse of the AMOC.'
author:
- 'Hassan Alkhayuon$^{1}$, Peter Ashwin$^{1}$, Laura C Jackson$^{2}$, Courtney Quinn$^{1,3}$ and Richard A Wood$^{2}$'
title: 'Basin bifurcations, oscillatory instability and rate-induced thresholds for AMOC in a global oceanic box model'
---
$^{1}$Department of Mathematics, University of Exeter, Exeter EX4 4QF, U.K.\
$^{2}$Met Office Hadley Centre, Exeter, U.K.\
$^{3}$CSIRO Oceans and Atmosphere, Hobart, TAS, Australia
Introduction {#sec:intro}
============
Since Stommel’s realisation [@Stommel1961] that competing thermal and haline circulation effects may give rise to multiple stable states in the Atlantic Meridional Overturning Circulation (AMOC), several studies have found evidence of this across a wide range of scales in the model hierarchy: see for example [@Dijkstra2007; @Dijkstra2013; @hawkins2011; @Rahmstorf1996; @Rahmstorf1999; @rahmstorf2005thermohaline]. Dansgard-Oeschger (DO) events give strong palaeoclimatic evidence of repeated and significant changes to the transport properties of AMOC between an “on” state that transports significant amounts of heat to high latitudes and an “off” state that does not. In the palaeoclimate record, the time of residence in the states can be a matter of 1000s of years while the switching can be relatively rapid and may be associated with Heinrich events where there were large inputs of fresh water into the Atlantic from glaciers [@Bondetal:2013; @ClementPeterson2008]. It is possible that fresh water input at high latitudes (i.e. changes to precipitation and evaporation patterns or ice sheet melting) could give a perturbation that can move the AMOC over a tipping point from “on” to “off” state.
Tipping points are significant changes to the solutions of a system in response to a small perturbation, and there has been a concerted effort to understand their appearance and predictability [@Scheffer2009; @Lenton2011]. They are often thought of in terms of bifurcations for some slowly varying nonlinear system, though [@Ashwin2011] point out that, in addition to such bifurcation-induced tipping (B-tipping), noise fluctuations (N-tipping) and sufficiently rapid shifts of a parameter can give rise to similar phenomena, where the tipping is associated with critical rates [@Scheffer2008] (R-tipping) rather than specific bifurcations.
For the climate, a change in external forcing (e.g. atmospheric greenhouse gases or fresh water input from ice sheets) would be expected to change both the current equilibrium state (“on”) of the AMOC and its basin of attraction. In some cases the forcing could be strong enough that the “on” state is no longer a valid equilibrium. In that case the AMOC would be expected to collapse to the “off” state (B-tipping), although if the forcing were reversed quickly enough it could recover [@vdB2018] (with some “resilience time” [@JW17]). For weaker forcing, the “on” state may remain a viable equilibrium. If the forcing changes sufficiently slowly, the AMOC will keep up with the changing equilibrium and remains within its basin of attraction, whereas if the change in forcing is too fast, the current AMOC state may lie outside the basin of the new “on” equilibrium. In this latter case the AMOC may be expected to transition to an “off” equilibrium, even if the “on” state still exists and is stable. Such rate dependent tipping of the AMOC has already been observed in a an intermediate-complexity climate model [@stocker1997]: when atmospheric carbon dioxide was increased to a new, steady concentration value slowly, the AMOC remained on, but when it was increased more quickly to the same steady concentration, the AMOC collapsed.
For the AMOC, one hopes that monitoring fluctuations in its strength [@mccarthy2015measuring], or other related ocean variables [@RWSH17], will enable us to spot early warning signals of an approaching tipping point due to changing parameters in the system. These changing parameters are typically idealised as changes in freshwater input or “hosing” of regions of the North Atlantic. We note that some more complex models (e.g. [@Peltier:2014]) can simulate repeated switching between these AMOC states as slow and spontaneous relaxation oscillations without the need for any freshwater input.
For simple ocean box models such as Stommel’s [@Stommel1961], the ocean is divided into a small number of boxes and continuity of salt and heat fluxes is used to derive a number of dynamic equations. For these models it is possible to perform a fairly complete analysis of the dynamics and find steady bifurcation points where changes to model parameters give changes in the qualitative set of solutions, although even for these models there can be surprises, for example in the susceptibility to noise [@Lohmann1999]. These have been extended by some authors to include multiple-box models of the oceans [@LucariniStone2005; @Titzetal2002a] where the increased dimension of phase space means that bifurcations can involve dynamically more complex states. In particular, [@Titzetal2002b; @Titzetal2002a] find the destabilisation of the “on” state may involve Hopf and homoclinic bifurcations. At the other end of the modelling scale, the most realistic models use detailed global atmosphere-ocean general circulation models (AOGCMs) with realistic geometry, and earth system models of intermediate complexity (EMICs) which are similar to AOGCMs but use simplified physical schemes to speed up computation. A few studies have shown the existence of multiple AMOC equilibria and transitions between the states, usually in response to idealised “hosing” experiments in which additional fresh water input is applied to the North Atlantic. This is either slowly increased up to and beyond a critical value (e.g. [@hawkins2011; @jacksonfamous17; @rahmstorf2005thermohaline]), or applied as a large perturbation for a limited time (e.g. [@DeVries05; @JW17; @liu2012diagnostic; @manabe88]. A few AOGCM studies have begun to explore the transient AMOC dynamics as thresholds are crossed, including the resilience time when fresh water forcing is temporarily increased to a level where the AMOC “on” state is unsustainable [@jackson18ClDyn; @JW17; @sijp2012precise; @LucariniCalmanti2005; @LucariniCalmanti2007]. As mathematical models of AMOC across scales may have quite different structures and parametrizations, it can be a challenge to make direct quantitative comparisons between the models at the ends of the scale. Analysis of AOGCM solutions requires a substantial reduction of complexity, for example by considering time- and spatially-averaged flows. In an attempt to make direct quantitative comparison between the most complex and simple models, a simple but physically-based five box model for global ocean circulation that exhibits AMOC switching has been developed by Wood et al. [@RWSH17]. They show that the dynamics of the AMOC in a hosing experiment with the FAMOUS AOGCM [@Smith2012; @hawkins2011; @jacksonfamous17] are well described by the dynamics of the box model, when the boxes are calibrated to the water masses in the unperturbed AOGCM state. This allows direct quantitative comparisons between states and fluxes for the AOGCM runs and the much simpler five box model, so that the box model can be used to provide insights into the AOGCM’s behaviour.
In this study we investigate in detail the dynamics of this box model of [@RWSH17], using our analysis to gain insight into possible tipping behaviours of the AMOC and hence to interpret the behaviours seen in more complex AOGCM experiments. We use a timescale separation that further reduces the five box model of [@RWSH17] to a three box model that seems to include the qualitative and quantitative behaviour of interest for the five box model. We find notable departures from the dynamics of the two-box model [@Stommel1961], some of which have been found in other multi-box models [@LucariniStone2005; @Titzetal2002b; @Titzetal2002a]. The differences are:
- The first loss of stability of the on-state is via a subcritical Hopf bifurcation rather than a fold bifurcation. This distance between the subcritical Hopf and fold bifurcations increases on increasing CO$_2$. Although there is a fold nearby, it represents the meeting of two solutions that are both unstable: cf [@Titzetal2002b; @Titzetal2002a].
- There can be an oscillatory relaxation to the on-state found both in three box model and the AOGCM runs: cf [@LucariniStone2005].
- The basin of attraction of the “on” state goes from semi-infinite size to finite size at a homoclinic bifurcation.
These changes in the basin of attraction of the “on” state mean that the response of the system to changes in forcing can vary with time in a non-trivial manner. We find rate-dependent effects where thresholds involve interaction between the attractors of the system and the unstable states. In climate terms, these insights into the dynamics of the model system can potentially lead to observable indicators of when the AMOC is approaching, or has crossed, critical stability thresholds; for example trajectories of the subtropical and subpolar Atlantic salinity undergo a qualitative change in character when the current, strong AMOC becomes unsustainable. The possibility of rate-dependent tipping has important implications for determining safe emissions pathways to reach climate stabilisation targets.
The paper is organized as follows. In Section \[sec:concept\] we briefly review box models of the AMOC and their bifurcations. Section \[sec:fivebox\] introduces the five box model of [@RWSH17] and presents an empirical three box reduction to fit numerical observations that two of the boxes show very little in the way of dynamic variation. We perform a comparative bifurcation analysis of these models for parameters from [@RWSH17] that are fit to FAMOUS runs with pre-industrial ($1\times $CO$_2$) and double ($2\times $CO$_2$), and find good agreement in both cases. Section \[sec:tipping\] discusses some of the implications for tipping points caused by time dependent changes to the hosing that may lead to collapse of AMOC. We compare these time-dependent perturbations to similar AOGCM studies [@JW17; @RWSH17]. In particular, Section \[sec:Btip\] demonstrates that crossing the threshold need not inevitably lead to tipping: there may be temporary resilience as long as and crossing is reversed sufficiently rapidly, and we quantify how fast this reversal must be. Section \[sec:Rtip\] shows that on the other hand, too rapid a change in forcing can destabilise the system, even if the bifurcation is not crossed. In Section \[sec:gcm\] we relate the results of our box model analysis to relevant AOGCM experiments. Section \[sec:discussion\] is a further discussion of implications of this work on the stability of the AMOC for rate-dependent perturbed CO${}_2$ scenarios.
Box models of the AMOC {#sec:concept}
----------------------
The two-box model of Stommel [@Stommel1961] can be written in non-dimensional form as $$\begin{aligned}
\frac{dT}{dt}&=\eta_1-T(1+|T-S|) \label{eq:twobox},\\
\frac{dS}{dt}&=\eta_2-S(\eta_3+|T-S|)\nonumber,\end{aligned}$$ and it is an idealisation of the dynamical interaction between non-dimensional equator-to-pole temperature gradient $T$ and a corresponding salinity gradient $S$ [@Dijkstra2013]. There are three non-negative dimensionless parameters: $\eta_1$ and $\eta_2$ represent the relative strength of thermal and freshwater forcing and $\eta_3$ is the ratio of thermal to freshwater surface restoring times. Note that even though this system is composed of two coupled ODEs where each expression is quadratic, there remain fundamental unsolved problems about the dynamics of quadratic systems on the plane - in particular Hilbert’s 16th problem is unsolved - this conjectures that there can only be finitely many limit cycles for such a system [@Dumortier2000].
As recognised already by Stommel [@Stommel1961], the system (\[eq:twobox\]) allows two stable equilibrium states, and bifurcations (possibly non-smooth, due to the terms $|T-S|$ in (\[eq:twobox\]) that are continuous but not differentiable at $T=S$ [@Bernardo2008; @Kowalczyk2011]) that go between these states on varying a parameter (typically $\eta_2$). On varying $\eta_2$ in (\[eq:twobox\]) we can find a bifurcation diagram as shown in Figure \[fig:twoboxbifs\] which shows a region of hysteresis with two stable states. Note that the linear stability of any equilibrium $(T^*,S^*)$ for (\[eq:twobox\]) with $T^*\neq S^*$ and $\phi= \mathrm{sgn}(T^*-S^*)$ is determined by the Jacobian $$J=\begin{bmatrix}
-1+[-2T^*+S^*]\phi & T^*\phi\\
-S^*\phi & -\eta_3+[2S^*-T^*]\phi
\end{bmatrix}.$$ Note that $$\mathrm{Tr}(J)=-1-\eta_3+[-T^*+S^*]\phi= -1-\eta_3-3|T^*-S^*|.$$ The fact that this trace is negative for all positive $\eta_3$ implies that there can be no Hopf bifurcation of an equilibria for (\[eq:twobox\]): the only smooth local bifurcations that occur will have one dimensional centre manifold, in this case saddle-node. There are also non-smooth saddle-node bifurcations of equilibria with $T^*=S^*$ [@Kowalczyk2011]. Figure \[fig:twoboxbifs\] shows a typical bifurcation diagram on varying $\eta_2$ for this model: note that complex eigenvalues can appear at equilibria even though there is no Hopf bifurcation.
Modifications of this two-box model can be reduced to an Itô SDE for one variable $y(t)$ [@Dijkstra2013 Equations 10.19, 10.20] of the form $$dy=-V'(y,F)dt+\sigma dW_t,
\label{eq:twobox_reduced}$$ where $$V(y,F)=-Fy+y^2/2+\mu^2(y^4/4-2y^3/3+y^2/2)$$ is a potential with two wells corresponding to AMOC “on" and “off" states, $F$ represents an (possibly time-dependent) injection of a freshwater flux in the North Atlantic box corresponding to a “hosing" of this with fresh water, and $W_t$ is a noise process with amplitude $\sigma$ that represents a stochastic component to the freshwater flux [@cessi1994simple]. The model (\[eq:twobox\_reduced\]) for $\sigma=0$ effectively produces the same behaviour as (\[eq:twobox\]) and a bifurcation diagram similar to Figure \[fig:twoboxbifs\], namely a region of hysteresis between the “on” and “off” states, $X_{\mathrm{on}}$ and $X_{\mathrm{off}}$, terminated by saddle-node bifurcations with the unstable separating state $X_{\mathrm{saddle}}$. More complex models, such as the multiple box models studied in [@LucariniStone2005; @Titzetal2002b; @Titzetal2002a; @LucariniFaranda2012] show richer dynamics including Hopf bifurcations.
![ Bifurcation diagram showing $\Psi=T-S$ plotted against $\eta_2$ for $\eta_1=3.0$ and $\eta_3=0.3$ for the two-box model (\[eq:twobox\]): see for example [@Dijkstra2013 Figure 10.4]. Observe the upper and lower branches correspond to AMOC “on” and “off” states. Note the saddle node bifurcations at the end of each branch, the lower of which is a non-smooth saddle node. \[fig:twoboxbifs\] ](./bifdaig_2box.eps){width="60.00000%"}
Bifurcations of an AMOC box model {#sec:fivebox}
=================================
By examining the geometry and behaviour of ocean currents and salinity for the experiments of Hawkins et al [@hawkins2011] using the FAMOUS AOGCM [@Smith2012], Wood et al [@RWSH17] propose a realistic box model with five boxes that models the balance of salinity in the main ocean water masses. They consider North Atlantic ($N$) and Tropical Atlantic ($T$) boxes, as well as an Indo-Pacific ($IP$) box, and assume that the temperature of the North Atlantic box is dependent on the AMOC strength. These boxes are coupled together via boxes corresponding to the Southern Ocean ($S$) and the bottom waters ($B$). Figure \[fig:fiveboxflows\] schematically illustrates the salinity fluxes between boxes.
![ The flows between the five boxes representing (from top left) Southern Ocean, Tropical Atlantic, North Atlantic, Indo-Pacific and bottom waters, in the global ocean box model of [@RWSH17] (cf their Fig. 1). There are freshwater surface fluxes $F_X$ and exchange of salinity via the indicated arrows. The $K_X$ indicate wind-driven fluxes while the advective flux due to the AMOC, $q$, is driven by a balance of salinity and temperature gradients. There is an $S$/$B$ mixing parameter $\eta$. A proportion $\gamma$ of the flux $q$ follows a cold water path via the Southern Ocean, while the remainder follows the warm water path via the Indo-Pacific.[]{data-label="fig:fiveboxflows"}](./fig_fiveboxes.eps){width="60.00000%"}
The five box model is as follows. We assume that the AMOC flow $q$ is proportional to the density gradient between the $N$ and $S$ boxes (which in turn is determined by the temperatures and salinities of the boxes): $$q=\frac{\lambda [\alpha (T_S - T_0) + \beta (S_N - S_S)]}{1+\lambda\alpha\mu}.$$ For $q \geq 0$ conservation of salt for the five boxes gives:
$$\label{eq:AOMC_model_q>0}
\left.\begin{array}{rcl}
V_N \frac{dS_N}{dt} &=& q(S_T - S_N) + K_N (S_T - S_N) - F_N S_0, \\
V_T \frac{dS_T}{dt} &=& q [\gamma S_S + (1- \gamma) S_{IP} - S_T] + K_S (S_S - S_T)+K_N(S_N - S_T) - F_T S_0,\\
V_S \frac{dS_S}{dt} &=& \gamma q (S_B -S_S) + K_{IP} (S_{IP} - S_S) + K_S (S_T - S_S) +\eta (S_B - S_S) - F_S S_0, \\
V_{IP} \frac{dS_{IP}}{dt} &=& (1-\gamma)q(S_B - S_{IP}) + K_{IP} (S_S - S_{IP}) - F_{IP} S_0, \\
V_{B} \frac{d S_B}{dt} &=& q(S_N - S_B) + \eta (S_S - S_B),
\end{array}\right\}$$
while for $q < 0$ we have $$\label{eq:AOMC_model_q<0}
\left.\begin{array}{rcl}
V_N \frac{dS_N}{dt} &=& |q|(S_B - S_N) + K_N (S_T - S_N) - F_N S_0, \\
V_T \frac{dS_T}{dt} &=& |q| (S_N - S_T) + K_S (S_S - S_T)+K_N(S_N - S_T) - F_T S_0,\\
V_S \frac{dS_S}{dt} &=& \gamma |q| (S_T -S_S) + K_{IP} (S_{IP} - S_S) + K_S (S_T - S_S) +\eta (S_B - S_S) - F_S S_0, \\
V_{IP} \frac{dS_{IP}}{dt} &=& (1-\gamma)|q|(S_T - S_{IP}) + K_{IP} (S_S - S_{IP}) - F_{IP} S_0, \\
V_{B} \frac{d S_B}{dt} &=& \gamma |q| S_S + (1-\gamma) |q| S_{IP} - |q| S_B + \eta (S_S -S_B).
\end{array}\right\}$$
If we consider the total salt content $C$, $$C= V_N S_N+ V_T S_T+ V_S S_S+ V_{IP} S_{IP} + V_B S_B,
\label{eq:totalsalt}$$ then $$dC/dt=-(F_N+F_T+F_S+F_{IP})S_0.$$ So if the total surface fresh water fluxes satisfy $$F_N+F_T+F_S+F_{IP}=0,
\label{eq:fluxesbalance}$$ then $C$ is constant on trajectories of (\[eq:AOMC\_model\_q>0\],\[eq:AOMC\_model\_q<0\]). This means that we can eliminate one of the equations.
Note that if we assume $C$ is constant and solve (\[eq:totalsalt\]) to give $S_B$ in terms of the other variables this corresponds to the solutions of the original system as long as (\[eq:fluxesbalance\]) is satisfied. If we impose (\[eq:totalsalt\]) but (\[eq:fluxesbalance\]) is not satisfied, then we are effectively solving the system with an additional flux $F_B$ to the $V_B\frac{dS_B}{dt}$ equation, such that $F_B=-F_N-F_S-F_T-F_{IP}$.
In this paper we consider values of the model parameters that are calibrated to the $FAMOUS_B$ AOGCM model version [@Smith2012] with either pre-industrial $\mathrm{CO}_2$ concentrations (“$1 \times \mathrm{CO}_2$”, shown in Appendix \[sec:appdx1\], Table \[tab:oceans\] and Table \[tab:params\]), or with doubled $\mathrm{CO}_2$ concentrations (“$2 \times \mathrm{CO}_2$”, shown in Appendix \[sec:appdx1\], Table \[tab:oceans2CO2\] and Table \[tab:params2CO2\]). See [@RWSH17] for details.
A three box reduction of the five box model
-------------------------------------------
We consider an empirical three box reduction as follows. We note from [@RWSH17 Figure 2] that the variation of $S_S$ and $S_B$ are apparently small and slow compared to $S_N$ and $S_T$. Hence we fix $S_S$ and $S_B$ and consider only $S_N$, $S_T$ and $S_{IP}$ as dynamic variables. Conservation of salt (\[eq:totalsalt\]) means we can solve for one of the variables (we choose $S_{IP}$) to give $$\label{eq:AOMC_model_q>0_3box}
\left.\begin{array}{rcl}
V_N \frac{dS_N}{dt} &=& q(S_T - S_N) + K_N (S_T - S_N) - F_N S_0, \\
V_T \frac{dS_T}{dt} &=& q [\gamma S_S + (1- \gamma) S_{IP} - S_T] + K_S (S_S - S_T)+K_N(S_N - S_T) - F_T S_0,\\
\end{array}\right\}$$ for $q\geq 0$ and $$\label{eq:AOMC_model_q<0_3box}
\left.\begin{array}{rcl}
V_N \frac{dS_N}{dt} &=& |q|(S_B - S_N) + K_N (S_T - S_N) - F_N S_0, \\
V_T \frac{dS_T}{dt} &=& |q| (S_N - S_T) + K_S (S_S - S_T)+K_N(S_N - S_T) - F_T S_0,\\
\end{array}\right\}$$ for $q<0$, where we fix $S_S$, $S_B$ and use (\[eq:totalsalt\]) to determine the remaining variable $S_{IP}$. The baseline parameters (we refer to as $1\times \mathrm{CO}_2$ as they are tuned to FAMOUS runs with pre-industrial atmospheric $\mathrm{CO}_2$) are given in Appendix \[sec:appdx1\]. For the numerical studies presented in the following sections we will use scaled salinities to enable more accurate numerical computation: $\tilde{S}_i = 100(S_i - S_0)$ for $i\in\{N,T,S,IP,B\}$ (see Appendix \[sec:appdx2\] for the modified equations). To convert to the commonly used units of parts per thousand, $S_{ppt} = 35.0 + 10\tilde{S}$.
Bifurcations of the five box model
----------------------------------
We consider variation of the fresh water fluxes as in [@RWSH17]: we consider a “hosing” forcing $H$ which may depend on time and that affects all of the fresh water fluxes simultaneously. The pattern of hosing represents a surface fresh water input into the North Atlantic over 20-50$^{o}$N (into parts of both $N$ and $T$ boxes) and removal elsewhere in order to preserve fresh water content so that (\[eq:fluxesbalance\]) is satisfied, as in [@hawkins2011]. Table \[tab:hosingfluxes\] shows the surface fluxes corresponding to hosing $H$, for the $1\times \mathrm{CO}_2$ parameter set.
$$\begin{array}{l|rcl|r}
F_N & 0.384+ 0.1311 H &~~~& F_S & 1.078-0.2626 H\\
F_T & -0.723+0.6961 H &~~~& F_{IP} & -0.738-0.5646 H
\end{array}$$
We present xppaut [@xppautref] and COCO [@cocoref] continuation and bifurcation analysis of (\[eq:AOMC\_model\_q<0\],\[eq:AOMC\_model\_q>0\]) for the $1\times\mathrm{CO}_2$ scenario. Similarly to the two-box model, there is a hysteresis between two branches of stable equilibria $X_{\mathrm{on}}$ and $X_{\mathrm{off}}$, separated by saddles $X_{\mathrm{saddle}}$, on varying $H$ and imposing $C$ constant (ie. using (\[eq:totalsalt\]) to specify $S_{IP}$), as shown in Figure \[fig:fiveboxbifs1CO2\]. Figure \[fig:threebox\_sketchbifs\] schematically shows changes to the system on varying $H$. The behaviour is qualitatively the same for either $1\times\mathrm{CO}_2$ or $2\times\mathrm{CO}_2$ parameters. The hysteresis in this model is not as straightforward as appears in the two-box model but is more like that found in [@Titzetal2002a; @Titzetal2002b].
In particular, although the lower branch is destroyed at a saddle- node bifurcation, the upper branch loses stability at a subcritical Hopf bifurcation before it is destroyed. In addition there is a homoclinic bifurcation (see Figure \[fig:fiveboxbifs1CO2\]) that leads to a sudden collapse in the size of the basin of attraction before the Hopf bifurcation occurs.
Bifurcations of the three box model
-----------------------------------
The bifurcation structures for the five box model are qualitatively the same as those of the three box reduction (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) on varying $H$, and both are qualitatively the same for both $1\times \mathrm{CO}_2$ and $2\times \mathrm{CO}_2$ parameters. Indeed, Table \[tab:comparebifsH\] compares the locations of the four bifurcations for the two systems. With the assistance of computer algebra (Maple) it is possible to find all branches for $q\geq 0$ and $q<0$ respectively to arbitrary precision and the local (i.e. saddle-node and Hopf) bifurcation points for the three box model: these parameter values are included on Table \[tab:comparebifsH\]. The homoclinic bifurcations are approximated by following a large but fixed period orbit near the end of the branch that bifurcates from the Hopf bifurcation.
Comparing the bifurcations of the five and three box models on varying $H$, note there is close agreement qualitatively (and quantitatively for $1\times \mathrm{CO}_2$): see Table \[tab:comparebifsH\]. The bifurcation scenario is robust: for example on changing the parameter $\gamma$ we find the Hopf curve meets the upper saddle node bifurcations on reducing $\gamma$: see Figure \[fig:threebox\_2parambif\]. The two curves meet at a Bogdanov-Takens (BT) point at $(H,\gamma)=(0.2268,0.1559)$ that can be determined (with the aid of Maple) by simultaneous solution of an equilibrium of (\[eq:AOMC\_model\_q>0\_3box\]) with $q>0$ whose Jacobian has both determinant and trace zero. Note that the BT point represents a convergence not just of Hopf and Saddle node bifurcations but also we can infer [@Kuznetsov1998] that a homoclinic bifurcation continues to the BT point.
$$\begin{array}{r|rrr|rrr}
& & \mbox{$1\times \mathrm{CO}_2$} & & & \mbox{$2\times \mathrm{CO}_2$}& \\
& \mbox{Five box} & \mbox{Three box} & \mbox{Three box} & \mbox{Five box} &\mbox{Three box}&\mbox{Three box}\\
& \mbox{numerical} & \mbox{numerical} & \mbox{Maple} & \mbox{numerical} & \mbox{numerical}&\mbox{Maple}\\
\hline
H_{\mathrm{lsn}} & -0.07996 & -0.05446 & -0.05445 & -0.4056 & -0.3795& -0.3792 \\
H_{\mathrm{hom}} & 0.2165 & 0.2128 & \mbox{NA}& 0.4510 & 0.3555 & \text{NA} \\
H_{\mathrm{Hopf}} & 0.2191 & 0.2134 & 0.2133& 0.4789 & 0.3895& 0.3888 \\
H_{\mathrm{usn}} & 0.2214 & 0.2139 & 0.2138& 0.511 & 0.4236 & 0.4225 \\
\end{array}$$
Table \[tab:comparebifsH\] and Figure \[fig:threeboxbifs2CO2\] also show the bifurcations for the 3-box model with the $2\times \mathrm{CO}_2$ parameter set. We see that for $2\times \mathrm{CO}_2$ there is a larger separation (in $H$) between the Hopf and upper saddle node bifurcations. For this reason we use the $2\times \mathrm{CO}_2$ parameters from here on, for clarity of exposition. We expect qualitatively similar results for the $1\times \mathrm{CO}_2$ parameters.
Basin bifurcations
------------------
Figures \[fig:threebox\_sketchbifs\] and \[fig:threebox\_basins\_H\] shows changes that take place in the basin of attraction of the upper branch $X_{\mathrm{on}}$ on varying $H$ (3-box model, $2\times \mathrm{CO}_2$ parameters). We observe that (similarly to [@Titzetal2002a; @Titzetal2002b])
- The approach to the upper branch is oscillatory: for the entire stable part of the upper branch there is a complex pair of slowly attracting eigenvalues for the linearisation. These become real at a spiral to node transition between Hopf and Saddle-node points.
- The basin of the upper branch shrinks from unbounded to a small and finite size at a homoclinic bifurcation just before the Hopf bifurcation.
This has some counter-intuitive effects for perturbations, for example it predicts an oscillatory overshoot of $S_N$ at a perturbation that switches the system from the lower to upper branch. As the basin of the upper branch shrinks to a small size, small perturbations in any direction can take the system out of the basin of the upper branch.
\[0.45\] [![ The basins of attraction of the two stable equilibria for different values of $H$, system (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) ($2\times \mathrm{CO}_2$ parameters). The colour map represents the duration (in years) to get as close as $10^{-3}$ to the equilibria. The values of $H$ are (a) 0, (b) 0.32678, (c) 0.34696 and (d) 0.35503. The integration for basins of attraction use 4th order Runge-Kutta with stepsize $10$ years. [*Supplementary Movie [MOC\_basin\_2co2.mp4]{} shows an animation of this figure.*]{}[]{data-label="fig:threebox_basins_H"}](./basin_coloured1.png "fig:"){width="45.00000%"}]{} \[0.45\] [![ The basins of attraction of the two stable equilibria for different values of $H$, system (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) ($2\times \mathrm{CO}_2$ parameters). The colour map represents the duration (in years) to get as close as $10^{-3}$ to the equilibria. The values of $H$ are (a) 0, (b) 0.32678, (c) 0.34696 and (d) 0.35503. The integration for basins of attraction use 4th order Runge-Kutta with stepsize $10$ years. [*Supplementary Movie [MOC\_basin\_2co2.mp4]{} shows an animation of this figure.*]{}[]{data-label="fig:threebox_basins_H"}](./basin_coloured2.png "fig:"){width="45.00000%"}]{} \[0.45\] [![ The basins of attraction of the two stable equilibria for different values of $H$, system (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) ($2\times \mathrm{CO}_2$ parameters). The colour map represents the duration (in years) to get as close as $10^{-3}$ to the equilibria. The values of $H$ are (a) 0, (b) 0.32678, (c) 0.34696 and (d) 0.35503. The integration for basins of attraction use 4th order Runge-Kutta with stepsize $10$ years. [*Supplementary Movie [MOC\_basin\_2co2.mp4]{} shows an animation of this figure.*]{}[]{data-label="fig:threebox_basins_H"}](./basin_coloured3.png "fig:"){width="45.00000%"}]{} \[0.45\] [![ The basins of attraction of the two stable equilibria for different values of $H$, system (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) ($2\times \mathrm{CO}_2$ parameters). The colour map represents the duration (in years) to get as close as $10^{-3}$ to the equilibria. The values of $H$ are (a) 0, (b) 0.32678, (c) 0.34696 and (d) 0.35503. The integration for basins of attraction use 4th order Runge-Kutta with stepsize $10$ years. [*Supplementary Movie [MOC\_basin\_2co2.mp4]{} shows an animation of this figure.*]{}[]{data-label="fig:threebox_basins_H"}](./basin_coloured4.png "fig:"){width="45.00000%"}]{}
Tipping in response to time-dependent forcing {#sec:tipping}
=============================================
In this section we illustrate the potential tipping behaviour of the AMOC in response to time-dependent forcing. We use the 3-box model with the $2\times \mathrm{CO}_2$ parameter set and integrate using a 4th order Runge-Kutta scheme with stepsize between $1$ and $2$ years.
Time-dependent hosing
---------------------
We consider the influence of time-dependent perturbations of (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) where the freshwater fluxes $F_X$ are varied via $H(t)$ according to a given protocol of variation that allows exploration of various scenarios of moving around the bifurcation diagram. The time-dependent hosing can be considered as an idealised representation of different future climate change mitigation scenarios, or palaeo-climatic periods.
If $H(t)$ is varied slowly enough, standard arguments of adiabatic reduction [@Kuehn2015] can be used to predict the behaviour of the nonautonomous perturbed system. However, it is not always clear how slow this needs to be, and in many cases there are scientifically important questions that need exploration of a range of fast or slow perturbations; for example current anthropogenic forcing of the climate system is on a similar or faster timescale than the inherent timescales of AMOC adjustment.
The particular protocol we consider here is inspired by the question of how quickly changes to hosing need to be reversed to avoid tipping of the upper branch onto the lower branch [@Ritchie2019]. This represents an idealisation of a scenario in which climate forcing temporarily overshoots its final value.
To this end, we consider the effect of piecewise linear (PWL) perturbations $H(t)=H_{pwl}(t-t_0)$ (visualized in Figure \[fig:PWL\], ) on the system, where $$H_{pwl}(t) =\begin{cases} H_0 & t<0, \\
\alpha(t) & t\in[0,T_{\mathrm{rise}}], \\
H_{\mathrm{pert}} & t-T_{\mathrm{rise}}\in[0,T_{\mathrm{pert}}],\\
\beta(t) & t-T_{\mathrm{rise}}-T_{\mathrm{pert}}\in[0,T_{\mathrm{fall}}], \\
H_0 & t\geq T_{\mathrm{rise}}+T_{\mathrm{pert}}+T_{\mathrm{fall}},
\end{cases}
\label{eq:H_PWL}$$ where $\alpha(t)$ and $\beta(t)$ are linear functions such that $H$ is continuous. This depends on constants $T_{\mathrm{rise}}$, $T_{\mathrm{pert}}$, $T_{\mathrm{fall}}$, and levels $H_0$ and $H_{\mathrm{pert}}$. If we define rise and fall rates $$r_{\mathrm{rise}}=\frac{|H_0-H_{\mathrm{pert}}|}{T_{\mathrm{rise}}},~
r_{\mathrm{fall}}=\frac{|H_0-H_{\mathrm{pert}}|}{T_{\mathrm{fall}}},$$ then note that $\alpha(t)=r_{\mathrm{rise}}t$ and $\beta(t)=r_{\mathrm{fall}}(t-T_{\mathrm{rise}}-T_{\mathrm{pert}})$. Similar PWL forcing experiments to the AMOC are explored in [@LucariniCalmanti2007; @LucariniCalmanti2005] where $T_{\mathrm{fall}}$ is always very large. As an example of the sort of critical transition that can arise, Figure \[fig:timeseries\_HPWL\] shows the effect of two similar PWL perturbations that differ only in $T_{\mathrm{fall}}$. As can be seen in the phase plane (a) and time series (b) these lead to quite different outcomes at the end of the perturbation.
![Schematic diagram showing a piecewise linear time-dependent perturbation $H(t)=H_{\mathrm{pwl}}(t-t_0)$ where $t_0=50$, $T_{\mathrm{rise}}=10$, $T_{\mathrm{pert}}=140$, $T_{\mathrm{fall}}=30$, $H_0=0$ and $H_{\mathrm{pert}}=0.1$. Whether the system tips from one state to the other depends not only on $H_{\mathrm{pert}}$ but also the other parameters. In the special case $T_{\mathrm{rise}}=T_{\mathrm{fall}}=0$ we refer to the perturbation as a “press”.[]{data-label="fig:PWL"}](./fig_pwlpert_v3.png){width="70.00000%"}
\[0.45\] [![Trajectories of (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) for $2\times \mathrm{CO}_2$ parameters and piecewise linear time-dependent parameter $H(t)$ defined by , the parameter values are $T_{\mathrm{pert}} = 200$, $H_0=0$, $H_{\mathrm{pert}} = 0.5$, $T_{\mathrm{rise}} = 200, T_{\mathrm{fall}} = 100$ for the black trajectory, and $T_{\mathrm{fall}} = 200$, $T_{\mathrm{rise}} = 200$ for the dark red trajectory. (a) shows the phase portrait and (b) the time profile of $S_N$ (for time profile of $S_T$ see Figure \[fig:Btip\]). In all cases we start at an initial condition on $X_{\mathrm{on}}$.[]{data-label="fig:timeseries_HPWL"}](./PWL_phaseportrait.eps "fig:"){width="45.00000%"}]{} \[0.45\] [![Trajectories of (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) for $2\times \mathrm{CO}_2$ parameters and piecewise linear time-dependent parameter $H(t)$ defined by , the parameter values are $T_{\mathrm{pert}} = 200$, $H_0=0$, $H_{\mathrm{pert}} = 0.5$, $T_{\mathrm{rise}} = 200, T_{\mathrm{fall}} = 100$ for the black trajectory, and $T_{\mathrm{fall}} = 200$, $T_{\mathrm{rise}} = 200$ for the dark red trajectory. (a) shows the phase portrait and (b) the time profile of $S_N$ (for time profile of $S_T$ see Figure \[fig:Btip\]). In all cases we start at an initial condition on $X_{\mathrm{on}}$.[]{data-label="fig:timeseries_HPWL"}](./AMOC_Btip_SN.eps "fig:"){width="45.00000%"}]{}
Resilience time: B-tipping and avoided B-tipping for instantaneous forcing changes {#sec:Btipinst}
----------------------------------------------------------------------------------
A special case of the PWL perturbation (\[eq:H\_PWL\]) is when $T_{\mathrm{rise}}=T_{\mathrm{fall}}=0$: an instantaneous change from $H_0$ to the perturbed $H_{\mathrm{pert}}$ and then back. In all cases we start from the $X_{\mathrm{on}}$ state for $H=0$. This type of perturbation has been studied extensively in ecological settings [@ratajczak2017] where it is called a “press”. We explore the behaviour of this case when varying $H_{\mathrm{pert}}$ and $T_{\mathrm{pert}}$, fixing $H_0=0$: an example is depicted in Figure \[fig:pp\_Hpress\]. The black line shows the initial departure from the equilibrium state for $H=H_0$. All trajectories follow this path in the $H={H_{\mathrm{pert}}}$ phase plane for the time ${T_{\mathrm{pert}}}$ corresponding to the line colour. When $t>{T_{\mathrm{pert}}}$ trajectories then make an excursion back to one of the two stable equilibria $X_{\mathrm{on/off}}$ for $H=H_0$. If a trajectory returns to $X_{\mathrm{on}}$ we say it does not tip, while if it approaches $X_{\mathrm{off}}$ then we say it tips.
The critical situations that divide these “tipping” and “non-tipping” responses can be estimated as follows. Comparing Figures \[fig:pp\_Hpress\]a and \[fig:threebox\_basins\_H\] we note that the divergence corresponds to where the black trajectory intersects with the basin boundary for $H=0$ (also shown on Figure 10). This critical situation corresponds precisely to the situation where a trajectory starting at $X_{\mathrm{on}}$ is on the stable manifold of the saddle $X_{\mathrm{saddle}}$ when the perturbation turns off. This corresponds to there being a “sink-saddle connection” for the perturbation we consider. We illustrate this critical value (“resilience time”) for different values of $H_{\mathrm{pert}}$ (mostly $> H_{\mathrm{Hopf}}$) in the case $T_{\mathrm{rise}}=T_{\mathrm{fall}}=0$ and $H_0=0$ in Figure \[fig:Hpert\_Tcrit\]. Note that there is not an exact relationship between whether tipping occurs and the total amount of fresh water input $H_{\mathrm{pert}}\times T_{\mathrm{pert}}$. In some cases a short, intense pulse can cause tipping while a longer but weaker pulse containing the same total fresh water does not (compare red and black dashed curves in the Figure). Physically this is because for the long, weak pulse the circulation (including the gyre diffusion term $K_N(S_T - S_N)$) has longer to flush the extra fresh water away from the North Atlantic (N box) (see also [@JW17]).
![Trajectories of (\[eq:AOMC\_model\_q>0\_3box\],\[eq:AOMC\_model\_q<0\_3box\]) ($2\times \mathrm{CO}_2$ parameters) with $H$ defined by (\[eq:H\_PWL\]) for the “press” perturbation with ${H_{\mathrm{pert}}}=0.5$ and $T_{\mathrm{rise}}=T_{\mathrm{fall}}=0$. The initial hosing value is fixed at $H_0=0$, and $T_{\mathrm{pert}}$ (in years) is indicated by the line colour. An initial condition is taken close to $X_{\mathrm{on}}$ for $H=H_0$. Note that longer durations result in switching to $X_{\mathrm{off}}$, and there is a critical duration, $234$ years, that gives a connection to the saddle state $X_{\mathrm{saddle}}$ (not shown). The grey dashed line represents the stable manifold of the saddle equilibrium (off the top of the figure) for $H=0$. []{data-label="fig:pp_Hpress"}](./H_press_05_1.eps){width="80.00000%"}
We now consider some illustrative examples of the effect of more general perturbations of the form (\[eq:H\_PWL\]) starting at $H_0=0$ and the upper branch.
Resilience time: B-tipping and avoided B-tipping for gradual forcing changes {#sec:Btip}
----------------------------------------------------------------------------
For more general perturbations of the type $H_{\mathrm{pwl}}$ illustrated in Figure \[fig:PWL\] we highlight that not only the perturbation and duration spent there, but also the timescales of getting there and returning will influence whether tipping will occur. In particular, even if $H_{\mathrm{pert}}>H_{\mathrm{Hopf}}$ B-tipping may be avoided: there can be a temporary resilience of the system even after crossing a bifurcation point if the return is fast enough, as already seen in Section \[sec:Btipinst\].
In the case that $H_0<H_{\mathrm{Hopf}}<H_{\mathrm{pert}}$, the bifurcation diagram in Figure \[fig:threeboxbifs2CO2\] for $2\times \mathrm{CO}_2$ parameters shows that the lower branch is the unique attractor for $H=H_{\mathrm{pert}}$. This means that regardless of $T_{\mathrm{rise}}$ and $T_{\mathrm{fall}}$ we expect to see tipping onto the lower branch if $T_{\mathrm{pert}}$ is large enough. However, if some combination of $T_{\mathrm{rise}}$, $T_{\mathrm{fall}}$ and $T_{\mathrm{pert}}$ is sufficiently small we may avoid B-tipping. Figure \[fig:Btip\] illustrates these two scenarios, the black trajectory shows the avoiding tipping behaviour while the dark red trajectory shows B-tipping.
For fixed $T_{\mathrm{pert}}$ and $H_{\mathrm{pert}}=0.5$ we characterise in Figure \[fig:tip\_after\_Hopf\] the regions of $(T_{\mathrm{rise}}, T_{\mathrm{fall}})$ where we get tipping onto the lower branch as a result of the perturbation. For large enough $T_{\mathrm{pert}}$ we find only tipping. However for $T_{\mathrm{pert}}<235$ we find that a rapid enough combination of rise and fall can avoid tipping. Again, whether tipping occurs is not solely determined by the total amount of fresh water added, with shorter, more intense perturbations being more effective in inducing tipping.
\[0.4\] \[0.4\] \[0.4\] \[0.4\]
![Results of applying the piecewise linear perturbation $H_{\mathrm{pwl}}(t)$ starting from $X_{\mathrm{on}}$ with $H_{\mathrm{pert}}$ beyond the Hopf bifurcation, with $H_0=0$, $H_{\mathrm{pert}}=0.5$, $T_{\mathrm{pert}}=100$ (left), $150$ (middle), $200$ (right) and $2\times \mathrm{CO}_2$ parameters. Otherwise as in Figure \[fig:Hpert\_Tcrit\]: white indicates an eventual return to $X_{\mathrm{on}}$ (does not tip), while grey ends at the $X_{\mathrm{off}}$ state. The threshold is indicated in red. Observe that rapid enough rates (white regions of figures) gives temporary resilience, i.e. the system avoids tipping in spite of crossing the threshold. []{data-label="fig:tip_after_Hopf"}](./TrisVSTfall_Btipping_Hpert05_Tpert100.png "fig:"){width="32.00000%"} ![Results of applying the piecewise linear perturbation $H_{\mathrm{pwl}}(t)$ starting from $X_{\mathrm{on}}$ with $H_{\mathrm{pert}}$ beyond the Hopf bifurcation, with $H_0=0$, $H_{\mathrm{pert}}=0.5$, $T_{\mathrm{pert}}=100$ (left), $150$ (middle), $200$ (right) and $2\times \mathrm{CO}_2$ parameters. Otherwise as in Figure \[fig:Hpert\_Tcrit\]: white indicates an eventual return to $X_{\mathrm{on}}$ (does not tip), while grey ends at the $X_{\mathrm{off}}$ state. The threshold is indicated in red. Observe that rapid enough rates (white regions of figures) gives temporary resilience, i.e. the system avoids tipping in spite of crossing the threshold. []{data-label="fig:tip_after_Hopf"}](./TrisVSTfall_Btipping_Hpert05_Tpert150.png "fig:"){width="32.00000%"} ![Results of applying the piecewise linear perturbation $H_{\mathrm{pwl}}(t)$ starting from $X_{\mathrm{on}}$ with $H_{\mathrm{pert}}$ beyond the Hopf bifurcation, with $H_0=0$, $H_{\mathrm{pert}}=0.5$, $T_{\mathrm{pert}}=100$ (left), $150$ (middle), $200$ (right) and $2\times \mathrm{CO}_2$ parameters. Otherwise as in Figure \[fig:Hpert\_Tcrit\]: white indicates an eventual return to $X_{\mathrm{on}}$ (does not tip), while grey ends at the $X_{\mathrm{off}}$ state. The threshold is indicated in red. Observe that rapid enough rates (white regions of figures) gives temporary resilience, i.e. the system avoids tipping in spite of crossing the threshold. []{data-label="fig:tip_after_Hopf"}](./TrisVSTfall_Btipping_Hpert05_Tpert200.png "fig:"){width="32.00000%"}
R-tipping, avoided R-tipping and basin boundaries {#sec:Rtip}
-------------------------------------------------
The bifurcation diagram Figure \[fig:threeboxbifs2CO2\] might seem to suggest that a perturbation from $H_0$ to $H_{\mathrm{pert}}<H_{\mathrm{Hopf}}$ will not result in tipping to the lower branch for any choice of $T_{\mathrm{rise}},T_{\mathrm{fall}}$ and $T_{\mathrm{pert}}$. Indeed, Figure \[fig:Rtip\] shows the effect of a perturbation that increases and then decreases $H$ fast enough for there to be no tipping from the upper branch. On the other hand, for a less rapid decrease (the dark red trajectory in Figure \[fig:Rtip\]) we can find situations that switch to the lower branch.
\[0.4\] \[0.4\] \[0.4\] \[0.4\]
For fixed $T_{\mathrm{pert}}$ and $H_{\mathrm{pert}}=0.37$ we characterise in Figure \[fig:tip\_before\_hom\] the regions of $(T_{\mathrm{rise}}, T_{\mathrm{fall}})$ where we get tipping onto the lower branch as a result of the perturbation. For large enough $T_{\mathrm{pert}}$ we find tipping if $T_{\mathrm{rise}}$ is small enough, independent of $T_{\mathrm{fall}}$.
![ Eventual behaviour of the system as in Figure \[fig:Hpert\_Tcrit\] for $H_{\mathrm{pwl}}$ with $H_{\mathrm{pert}}$ before the homoclinic, $H_0=0$, $H_{\mathrm{pert}}=0.37$, $T_{\mathrm{pert}}=400$ (left), $600$ (middle), $800$ (right), and $2\times \mathrm{CO}_2$ parameters. As before, white: does not tip, grey: tips, red: threshold. Note that too rapid a rise of the perturbation can give rate-induced tipping, even when the perturbation is held for a long time and the return is very slow.[]{data-label="fig:tip_before_hom"}](./TrisVSTfall_Rtipping_Hpert037_Tpert400.png "fig:"){width="32.00000%"} ![ Eventual behaviour of the system as in Figure \[fig:Hpert\_Tcrit\] for $H_{\mathrm{pwl}}$ with $H_{\mathrm{pert}}$ before the homoclinic, $H_0=0$, $H_{\mathrm{pert}}=0.37$, $T_{\mathrm{pert}}=400$ (left), $600$ (middle), $800$ (right), and $2\times \mathrm{CO}_2$ parameters. As before, white: does not tip, grey: tips, red: threshold. Note that too rapid a rise of the perturbation can give rate-induced tipping, even when the perturbation is held for a long time and the return is very slow.[]{data-label="fig:tip_before_hom"}](./TrisVSTfall_Rtipping_Hpert037_Tpert600.png "fig:"){width="32.00000%"} ![ Eventual behaviour of the system as in Figure \[fig:Hpert\_Tcrit\] for $H_{\mathrm{pwl}}$ with $H_{\mathrm{pert}}$ before the homoclinic, $H_0=0$, $H_{\mathrm{pert}}=0.37$, $T_{\mathrm{pert}}=400$ (left), $600$ (middle), $800$ (right), and $2\times \mathrm{CO}_2$ parameters. As before, white: does not tip, grey: tips, red: threshold. Note that too rapid a rise of the perturbation can give rate-induced tipping, even when the perturbation is held for a long time and the return is very slow.[]{data-label="fig:tip_before_hom"}](./TrisVSTfall_Rtipping_Hpert037_Tpert800.png "fig:"){width="32.00000%"}
We can interpret this behaviour as follows: for $T_{\mathrm{pert}}$ large enough, tipping depends on whether we are in the basin of attraction of the lower branch at the end of the rise. Because the basin for $H=0.125$ is bounded, a fast enough rise can leave the system outside the upper branch basin.
Relationship to AMOC tipping in complex climate models {#sec:gcm}
======================================================
The analysis of Sections \[sec:fivebox\] and \[sec:tipping\] provides insight into the bifurcation structure and tipping behaviour of the five box model and its three box reduction. Given that the box model has been shown to capture the dynamics of AMOC hysteresis and tipping in the FAMOUS AOGCM [@RWSH17], it is natural to ask how relevant our insights are to the AMOC behaviour in complex climate models. For the present study we simply ask how the behaviour of the box model relates to existing GCM and EMIC studies. Future work will exploit the insights from the box model to target experimental design for further AOGCM studies, and hence make best use of limited computational resources.
A particular feature of the transient box model solutions can be seen in the ($S_N$,$S_T$) phase trajectories shown in idealised form in Figure \[fig:threebox\_sketchbifs\]: trajectories in the vicinity of the upper branch (strong AMOC) equilibrium oscillate in a clockwise sense in phase space (whether they are approaching or diverging from the upper branch), but trajectories that are eventually destined for the lower branch (weak/reversed AMOC) eventually show anticlockwise curvature (or a sharp left turn at the middle branch), and approach the lower branch in a non-oscillatory way. This suggests a simple and in principle observable diagnostic in a more complex model (or the real world): anticlockwise trajectories in phase space suggest that the system is on the attractor of the lower branch.
We use the above diagnostic to interpret the “hosing hysteresis" experiments described by [@hawkins2011; @RWSH17]. In these experiments hosing $H$ is slowly increased from $0$ to $1 \mathrm{Sv}$ over 2000 years, then decreased again. We know from [@RWSH17] that the five box model reproduces the dynamics of the FAMOUS AOGCM for these experiments, especially for the “ramp up" phase where $H$ is increasing. The usual assumption in such experiments is that the forcing change is slow enough that the system remains close to its equilibrium state throughout. However, [@hawkins2011] noted that the AMOC appeared to remain “on" for long after $H$ had passed the critical value where no “on" equilibrium existed, suggesting significant transient resilience. Looking at Figure \[fig:threebox\_basins\_H\] we see that for states just outside the basin of the “on” state, the time to reach the “off” state equilibrium is of order hundreds to thousands of years; since this timescale coincides with the timescale of the forcing increase it is not surprising that the AMOC persists long after the forcing threshold $H_{\mathrm{usn}}$ has been passed.
Figure \[fig:hysteresis\_trajectories\] shows the $(S_N,S_T)$ phase plane for the ramp-up phase of the hosing hysteresis experiment in the five box model (red). Points on the trajectory are joined by blue/green lines to the equilibrium point corresponding to their instantaneous hosing value (estimated by re-running the hysteresis experiment with a ten times slower rate of increasing $H$, black, [*[cf.]{}*]{} [@LucariniCalmanti2005]). We see that initially the transient solution remains close to its corresponding upper branch equilibrium. However when $H$ increases beyond $H_{\mathrm{usn}}$ (around 0.22 Sv for this parameter setting, Table \[tab:comparebifsH\]) and the upper branch ceases to exist, the transient solution remains in a strong AMOC state (large $S_N$, $S_T$), even though its equilibrium is now a weak AMOC state (low $S_N$, $S_T$). This change occurs at a point in the trajectory where $S_T$ starts to increase, while $S_N$ continues to decrease, an indication that a positive salinity advection feedback has started to dominate between the $T$ and $N$ boxes, reducing the amount of salty water carried northwards and so causing further MOC weakening([@LucariniCalmanti2005],[@RWSH17]). The trajectory for the corresponding FAMOUS experiment is qualitatively similar (orange curve): again $S_T$ starts to increase before the trajectory makes a sharp left turn towards the lower branch state. The change from clockwise to anticlockwise curvature indicates an approach to the lower branch state, but the critical value of hosing may have already been passed by the time the curvature changes; the earlier change from decreasing to increasing $S_T$ may be a better indicator of when $H_{\mathrm{usn}}$ has been passed, but this result may be scenario-dependent as it does not appear to apply for the ”press” scenario (see Figures \[fig:pp\_Hpress\], \[fig:GC2\_press\_traj\]).
We now consider the “press" forcing scenario studied for the box model in Section \[sec:Btipinst\]. Jackson and Wood [@JW17] have recently studied resilience times in a CMIP6-generation AOGCM, HadGEM3-GC2 under such a scenario, for a range of values of $H_{\mathrm{pert}}$ and $T_{\mathrm{pert}}$. Figure \[fig:res\_time\] shows the resilience times for the box model (with $1\times \mathrm{CO}_2$ parameters) and for HadGEM3-GC2. We do not expect quantitative comparison as HadGEM3-GC2 has different bifurcation points from FAMOUS (generally showing AMOC collapse at lower $H$ values). However, we do note that the forms of the two curves are the same, and the result that short, intense perturbations are more likely to cause tipping than long, weak perturbations is consistent across the models [@JW17]. Figure \[fig:GC2\_press\_traj\] shows the phase space trajectories for the various HadGEM3-GC2 experiments of [@JW17]. The AOGCM trajectories compare qualitatively with those for the box model in Figure \[fig:pp\_Hpress\]: cases for which the AMOC clearly recovers after the forcing is removed show the same clockwise trajectories, while the cases where there is not a clear recovery (shown in red) mostly appear to be close to the dividing trajectory between recovery and collapse ([*[cf.]{}*]{} dashed line in Figure \[fig:pp\_Hpress\]). In these cases the AOGCM AMOC appears to be in an intermediate state and it was not clear to [@JW17] whether that state would persist, revert to a strong AMOC or collapse (see Figure 1 of [@JW17]). Comparison to Figure \[fig:pp\_Hpress\] with a press scenario suggests that the AOGCM trajectories are not inconsistent with the box model but that the runs would need to be continued for longer to resolve this.
Overall the box model dynamics appears to provide insights into the behaviour and design of AOGCM hosing experiments. However as a caveat we note that recently Haskins et al. [@Haskins18] have analysed a “press" scenario in the FAMOUS AOGCM, in a case which shows an oscillatory AMOC recovery after the forcing is removed. While this oscillatory recovery looks qualitatively similar to our box model solutions, it appears to be driven by variations in the South Atlantic density, a processes that is not captured in the three box, and possibly not the five box model. Other processes may operate in AOGCMs that produce qualitatively similar looking dynamics to the box model solutions.
Finally we discuss the rate-dependent tipping observed in Section \[sec:Rtip\]. The concept of rate-dependent tipping in response to fresh water hosing was explored in a two-dimensional Boussinesq model in [@LucariniCalmanti2005; @LucariniCalmanti2007], where it was shown that fast perturbations can destabilize the typical advective (negative) feedback of the AMOC. Additionally, Stocker and Schmittner [@stocker1997] found rate-dependent AMOC tipping in an intermediate-complexity climate model (EMIC) in response to increasing greenhouse gases. Faster $\mathrm{CO}_2$ increase could result in tipping, while a more gradual increase toward the same steady state value did not. Although the $\mathrm{CO}_2$ forcing used by [@stocker1997] is not the same as the fresh water forcing we use, we note from Figure \[fig:tip\_before\_hom\] that similar behaviour can be seen in the box model, supporting the view that in order to predict the future stability of the AMOC we need to understand not just the sustainability of an equilibrium “on" state under climate forcing, but also the rate of forcing change (i.e. the emissions pathway, not just the final greenhouse gas levels).
Discussion {#sec:discussion}
==========
In this paper, we investigate the global ocean box model of [@RWSH17] and demonstrate, similarly to [@Titzetal2002a], that the switching from the “on" state is associated with subcritical Hopf rather than saddle-node bifurcation. We interpret time-dependent perturbations in terms of the bifurcation diagrams and show in particular that there can be a mixing of rate- and bifurcation-tipping.
While the forcing scenarios we investigate are highly idealised, they illustrate the possibility of rate-dependent tipping in the climate system, which is of particular interest as the international community considers options to limit dangerous climate change under the Paris climate accord. Our results reveal that tipping can occur in response to a rapid increase in forcing, even if slower increase to the same level of forcing does not result in tipping. Conversely if a forcing threshold is passed such that the AMOC is ultimately unsustainable, tipping may still be avoided if the forcing increase is reversed quickly enough. These results suggest a framework for the design of safe climate stabilisation pathways to avoid AMOC tipping. The characteristic direction and curvature of the phase trajectories in $(S_N,S_T)$ space is potentially observable in the real ocean and so could provide early warning that the AMOC is on the attractor of an “off” state.
Our work needs to be followed up using more realistic scenarios and models, but the theoretical understanding developed here can be used to guide experimentation with more comprehensive (but computationally expensive) models to address the questions “Under what circumstances might the AMOC switch to an off state?”, and “What would we need to observe to give the earliest possible warning of such a switch?”
We highlight that because the bifurcation to a finite basin occurs before loss of stability, this means that even perturbations that are not large enough to cause tipping may result in significant and long deviations, even if there is eventually a return. For example, in Figure \[fig:threebox\_basins\_H\] (c) if the state of the system is perturbed such that it gets close enough to the boundary of the basin of $X_\mathrm{on}$, it needs more than 2000 years to get back to equilibrium. For this reason it seems plausible that the present day AMOC may still retain significant memory of major perturbations that it underwent hundreds of years ago. For our investigations we only consider time-dependent perturbation of the “hosing" parameter $H$ where it is assumed that fluxes are balanced and so total ocean volume and salt content are preserved. Over longer timescales or for larger perturbations ([*[e.g.]{}*]{} glacial-interglacial transitions) this approximation will not necessarily be valid. In particular (\[eq:totalsalt\]) will no longer be constant and the associated dimension reduction not valid. Potentially this will give rise to new dynamical effects.
Finally we highlight the close relationship between R-tipping and nonautonomous stability: [@Alkhayuon2017; @Ashwin2017] show that R-tipping can be understood as a change in properties of a local pullback attractor and identify criteria for the appearance of R-tipping in simple systems. R-tipping can also appear in the presence of noise, for example [@Ritchie2016; @Ritchie2017] have studied the probability of the occurrence of R-tipping in the presence of noise, as well as some potential early warning indicators of such a phenomenon. Other recent work has examined in detail the crossing “points of no return” for systems with R-tipping [@OKeefe2019; @Ritchie2019], but in general the science of how reliable early warning signals may be for R-tipping is still to be worked out.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The simulation code used to generate figures for the box models, and the movies, are available via the code repository [https://github.com/peterashwin/AAJQW\_AMOC]{}.
HA was supported by the Higher Committee For Education Development in Iraq (HCED Iraq) via grant agreement No D13436. CQ was supported by the CRITICS Innovative training Network for funding via the European Union’s Horizon 2020 research and innovation programme under the Marie Sk[ł]{}odowska-Curie grant agreement No 643073. LJ and RW were supported by the Met Office Hadley Centre Climate Programme funded by BEIS and Defra. All authors thank the Past Earth Network (EPSRC grant number EP/M008363/1) and ReCoVER (EPSRC grant number EP/M008495/1) for supporting a workshop in 2017 where this work started.
Thanks also to Ulrike Feudel, Felix Mendelssohn, Paul Ritchie, Jan Sieber, Sebastian Wieczorek for inspiration and insightful comments.
[99]{}
Stommel H. 1961 Thermohaline Convection with Two Stable Regimes of Flow. [ *Tellus*]{} **13**, 224–230.
Dijkstra HA. 2007 Characterization of the multiple equilibria regime in a global ocean model. [*Tellus A*]{} **59**, 695–705.
Dijkstra HA. 2013 [*Nonlinear climate dynamics*]{}. Cambridge University Press, Cambridge.
Hawkins E, Smith RS, Allison LC, Gregory JM, Woollings TJ, Pohlmann H, De Cuevas B. 2011 Bistability of the Atlantic overturning circulation in a global climate model and links to ocean freshwater transport. [ *Geophysical Research Letters*]{} **38**.
Rahmstorf S. 1996 On the freshwater forcing and transport of the Atlantic thermohaline circulation. [*Climate Dynamics*]{} **12**, 799–811.
Rahmstorf S. 1999 Rapid transitions of the thermohaline ocean circulation. In [*Reconstructing Ocean History*]{} pp. 139–149. Springer.
Rahmstorf S, Crucifix M, Ganopolski A, Goosse H, Kamenkovich I, Knutti R, Lohmann G, Marsh R, Mysak LA, Wang Z, Weaver AJ. 2005 Thermohaline circulation hysteresis: A model intercomparison. [*Geophysical Research Letters*]{} **32**.
Bond GC, Showers W, Elliot M, Evans M, Lotti R, Hajdas I, Bonani G, Johnson S. 2013 pp. 35–58. In [*The North Atlantic’s 1-2 Kyr Climate Rhythm: Relation to Heinrich Events, Dansgaard/Oeschger Cycles and the Little Ice Age*]{}, pp. 35–58. American Geophysical Union (AGU).
Clement AC, Peterson LC. 2008 Mechanisms of abrupt climate change of the last glacial period. [*Reviews of Geophysics*]{} **46**.
Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, Dakos V, Held H, van Nes EH, Rietkerk M, Sugihara G. 2009 [Early-warning signals for critical transitions]{}. [*Nature*]{} **461**, 53–59.
Lenton TM. 2011 [Early warning of climate tipping points]{}. [*Nature Climate Change*]{} **1**, 201–209.
Ashwin P, Wieczorek S, Vitolo R, Cox P. 2012 [Tipping points in open systems: bifurcation, noise-induced and rate-dependent examples in the climate system]{}. [*Phil. Trans. R. Soc. A*]{} **370**, 1166–1184.
Scheffer M, [Van Nes]{} EH, Holmgren M, Hughes T. 2008 [Pulse-driven loss of top-down control: The critical-rate hypothesis]{}. [*Ecosystems*]{} **11**, 226–237.
Bolt, B and Nes, EH and Bathiany, S and Vollebregt, ME and Scheffer, M. 2018 Climate reddening increases the chance of critical transitions, [*Nature Climate Change*]{}, [**8**]{}:478.
Jackson L, Wood R. 2018 Hysteresis and Resilience of the AMOC in an Eddy-Permitting GCM. [*Geophysical Research Letters*]{} **45**, 8547–8556.
Stocker T, Schmittner A. 1997 Influence of $CO_2$ emission rates on the stability of the thermohaline circulation. [*Nature*]{} **388**, 862–865.
McCarthy GD, Smeed DA, Johns WE, Frajka-Williams E, Moat BI, Rayner D, Baringer MO, Meinen CS, Collins J, Bryden HL. 2015 [Measuring the Atlantic Meridional Overturning Circulation at 26$^o$N]{}. [*Progress in Oceanography*]{} **130**, 91–111.
Wood R, Rodriguez J, Smith R, Jackson L, Hawkins E. 2019 Observable, low-order dtnamical controls on thresholds of the Atlantic Meridional Overturning Circulation. [*Climate Dynamics (submitted)*]{}.
Richard PW, Guido V. 2014 Dansgaard-Oeschger oscillations predicted in a comprehensive model of glacial climate: A ’kicked’ salt oscillator in the Atlantic. [*Geophysical Research Letters*]{} **41**, 7306–7313.
Lohmann G, Schneider J. 1999 Dynamics and predictability of Stommel’s box model. A phase-space perspective with implications for decadal climate variability. [*Tellus A: Dynamic Meteorology and Oceanography*]{} **51**, 326–336.
Lucarini V, Stone PH. 2005 Thermohaline Circulation Stability: A Box Model Study. Part I: Uncoupled Model. [*Journal of Climate*]{} **18**, 501–513.
Lucarini V, Calmanti S, Artale V. 2005 Destabilization of the thermohaline circulation by transient changes in the hydrological cycle. [*Climate Dynamics*]{} **24**, 253-262.
Lucarini V, Calmanti S, Artale V. 2007 Experimental Mathematics: Dependence of the Stability Properties of a Two-Dimensional Model of the Atlantic Ocean Circulation on the Boundary Conditions [*Russ. J. Math. Phys*]{} **14**, 224–231.
Lucarini V, Faranda D, Willeit M. 2012 Bistable systems with stochastic noise: virtues and limits of effective one-dimensional Langevin equations [*Nonlin. Processes Geophys.*]{} **19**, 9–22.
Titz S, Kuhlbrodt T, Rahmstorf S, Feudel U. 2002a On freshwater-dependent bifurcations in box models of the interhemispheric thermohaline circulation. [*Tellus A*]{} **54**, 89–98.
Titz S, Kuhlbrodt T, Feudel U. 2002b Homoclinic bifurcation in an ocean circulation box model. [*International Journal of Bifurcation and Chaos*]{} **12**, 869–875.
Jackson L, Smith R, Wood R. 2017 Ocean and atmosphere feedbacks affecting AMOC hysteresis in a GCM. [*Climate Dynamics*]{} **49**, 173–191.
de Vries P, Weber SL. 2005 [The Atlantic freshwater budget as a diagnostic for the existence of a stable shut down of the Meridional Overturning Circulation]{}. [*Geophys. Res. Lett*]{} **32**.
Liu W, Liu Z. 2012 [A Diagnostic Indicator of the Stability of the Atlantic Meridional Overturning Circulation in CCSM3]{}. [*J. Climate*]{} **26**, 1926–1938.
Manabe S, Stouffer R. 1988 Two stable equilibria of a coupled ocean-atmosphere model. [*J. Climate*]{} **1**, 841–863.
Jackson L, Wood R. 2018 Timescales of AMOC decline in response to fresh water forcing. [*Climate Dyn.*]{} **51**, 1333–1350.
Sijp WP, England MH. 2012 Precise Calculations of the Existence of Multiple AMOC Equilibria in Coupled Climate Models. Part II: Transient Behavior. [ *Journal of Climate*]{} **25**, 299–306.
Smith RS. 2012 The FAMOUS climate model (versions XFXWB and XFHCC): description update to version XDBUA. [*Geoscientific Model Development*]{} **5**, 269–276.
Dumortier F, Herssens C, Perko L. 2000 Local Bifurcations and a Survey of Bounded Quadratic Systems. [*Journal of Differential Equations*]{} **165**, 430 – 467.
Bernardo M, Budd C, Champneys AR, Kowalczyk P. 2008 [*Piecewise-smooth dynamical systems: theory and applications*]{} vol. 163. Springer Science & Business Media.
Kowalczyk P, Glendinning P. 2011 Boundary-equilibrium bifurcations in piecewise-smooth slow-fast systems. [*Chaos: An Interdisciplinary Journal of Nonlinear Science*]{} **21**, 023126.
Cessi, P. 1994 A simple box model of stochastically forced thermohaline flow. [*Journal of physical oceanography*]{} **24**, 1911–1920.
Ermentrout B. 2002 [*Simulating, analyzing, and animating dynamical systems*]{} vol. 14[*Software, Environments, and Tools*]{}. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA. A guide to XPPAUT for researchers and students.
Dankowicz H, Schilder F. 2013 [*Recipes for continuation*]{} vol. 11[ *Computational Science & Engineering*]{}. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA.
Kuznetsov YA. 1998 [*Elements of Applied Bifurcation Theory (2nd Ed.)*]{}. Berlin, Heidelberg: Springer-Verlag.
Kuehn C. 2015 [*[Multiple Time Scale Dynamics]{}*]{}. Springer, New York.
Ritchie P, Karabacak Ö, Sieber J. 2019 [Inverse-square law between time and amplitude for crossing tipping thresholds]{}. [*Proceedings of the Royal Society A*]{} **475**, 20180504.
Ratajczak Z, D’odorico P, Collins SL, Bestelmeyer BT, Isbell FI, Nippert JB. 2017 The interactive effects of press/pulse intensity and duration on regime shifts at multiple scales. [*Ecological Monographs*]{} **87**, 198–218.
Haskins R, Oliver K, Jackson L, Drijfhout S, Wood R. 2018 Explaining asymmetry between weakening and recovery of the AMOC in a coupled climate model. [ *Climate Dyn.*]{}
Alkhayuon HM, Ashwin P. 2018 Rate-induced tipping from periodic attractors: Partial tipping and connecting orbits. [*Chaos*]{} **28**, 033608.
Ashwin P, Perryman C, Wieczorek S. 2017 [Parameter shifts for nonautonomous systems in low dimension: Bifurcation- and Rate-induced tipping]{}. [ *Nonlinearity*]{} **30**, 2185–2210.
Ritchie P, Sieber J. 2016 [Early-warning indicators for rate-induced tipping]{}. [*Chaos*]{} **26**.
Ritchie P, Sieber J. 2017 [Probability of noise-and rate-induced tipping]{}. [*Physical Review E*]{} **95**, 1–13.
O’Keefe PE, Wieczorek S. 2019 [Tipping Phenomena and Points of No Return in Ecosystems: Beyond Classical Bifurcations]{}. [*arXiv*]{}:1902.01796.
Model constants and parameters \[sec:appdx1\]
=============================================
The default constants and parameters for the five box model are listed in Tables \[tab:oceans\] and \[tab:params\].
$$\begin{array}{l|rrr}
& \mbox{Volume} & \mbox{Salinity}& \mbox{Flux} \\
\hline
\mbox{North Atlantic} & V_N= 0.3261 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_N = 0.034912 & F_N= 0.384 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Tropical Atlantic} & V_T= 0.7777 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_T = 0.035435 & F_T= -0.723 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Southern Ocean} & V_S= 0.8897 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_S = 0.034427 & F_S= 1.078 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Indo-Pacific} & V_{IP}= 2.2020 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_{IP} = 0.034668 & F_{IP}=-0.738 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Bottom Ocean} & V_{B}= 8.6490 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_{B} = 0.034538 &
\end{array}$$
$$\begin{array}{l|rlc||l|rl}
\mbox{Name} & \mbox{Default value} & \mbox{Units}&~~&\mbox{Name} & \mbox{Default value} & \mbox{Units}\\
\hline
\alpha & 0.12 &{\hspace{1mm}\mathrm{kg}}{\hspace{1mm}\mathrm{m}^{-3}}{\hspace{1mm}^{\circ}\mathrm{C}}^{-1}&&K_N & 5.456 &{\hspace{1mm}\mathrm{Sv}}\\
\beta & 790.0 &{\hspace{1mm}\mathrm{kg}}{\hspace{1mm}\mathrm{m}^{-3}}&&K_S & 5.447 &{\hspace{1mm}\mathrm{Sv}}\\
S_0 & 0.035& && K_{IP} & 96.817 &{\hspace{1mm}\mathrm{Sv}}\\
T_S & 4.773 &{\hspace{1mm}^{\circ}\mathrm{C}}&& \lambda & 2.79 {\mathrm{x}}10^{7} &\mathrm{m}^{6} {\hspace{1mm}\mathrm{kg}}^{-1} \mathrm{s}^{-1} \\
T_0 & 2.650 &{\hspace{1mm}^{\circ}\mathrm{C}}&&\gamma & 0.39&\\
\eta & 74.492 &{\hspace{1mm}\mathrm{Sv}}&& \mu & 5.5 & {\hspace{1mm}^{\circ}\mathrm{C}}^{-1} \mathrm{m}^{-3} \mathrm{s}{\mathrm{x}}10^{-8}
\end{array}$$
$$\begin{array}{l|rrr}
& \mbox{Volume} & \mbox{Salinity}& \mbox{Flux} \\
\hline
\mbox{North Atlantic} & V_N= 0.3683 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_N = 0.034912 & F_N= 0.486 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Tropical Atlantic} & V_T= 0.5418 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_T = 0.035435 & F_T= -0.997 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Southern Ocean} & V_S= 0.6097 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_S = 0.034427 & F_S= 1.265 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Indo-Pacific} & V_{IP}= 1.4860 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_{IP} = 0.034668 & F_{IP}=-0.754 {\hspace{1mm}\mathrm{Sv}}\\
\mbox{Bottom Ocean} & V_{B}= 9.9250 {\mathrm{x}}10^{17} {\hspace{1mm}\mathrm{m}^{3}}& S_{B} = 0.034538 &
\end{array}$$
$$\begin{array}{l|rlc||l|rl}
\mbox{Name} & \mbox{Default value} & \mbox{Units}&~~&\mbox{Name} & \mbox{Default value} & \mbox{Units}\\
\hline
\alpha & 0.12 &{\hspace{1mm}\mathrm{kg}}{\hspace{1mm}\mathrm{m}^{-3}}{\hspace{1mm}^{\circ}\mathrm{C}}^{-1}&&K_N & 1.762 &{\hspace{1mm}\mathrm{Sv}}\\
\beta & 790.0 &{\hspace{1mm}\mathrm{kg}}{\hspace{1mm}\mathrm{m}^{-3}}&&K_S & 1.872 &{\hspace{1mm}\mathrm{Sv}}\\
S_0 & 0.035& && K_{IP} & 99.977 &{\hspace{1mm}\mathrm{Sv}}\\
T_S & 7.919 &{\hspace{1mm}^{\circ}\mathrm{C}}&& \lambda & 1.62 {\mathrm{x}}10^{7} &\mathrm{m}^{6} {\hspace{1mm}\mathrm{kg}}^{-1} \mathrm{s}^{-1} \\
T_0 & 3.870 &{\hspace{1mm}^{\circ}\mathrm{C}}&&\gamma & 0.36&\\
\eta & 33.264 &{\hspace{1mm}\mathrm{Sv}}&& \mu & 22 & {\hspace{1mm}^{\circ}\mathrm{C}}^{-1} \mathrm{m}^{-3} \mathrm{s}{\mathrm{x}}10^{-8}
\end{array}$$
The scaled five box model \[sec:appdx2\]
========================================
In order to improve computational efficiency (particularly for the continuation software), we rescale both time and state variables of the models. The time unit $t$ of the original model is in seconds; we will consider a time unit of years $\tau = tY^{-1}$ where $Y=3.15{\mathrm{x}}10^{7}$. In addition we will consider the state variables as scaled perturbations from a background state, namely $$\label{eq:scaled_S}
\tilde{S}_i = 100(S_i - S_0), \qquad i\in[N,T,S,IP,B].$$ This leads to a modified equation for $q$, $$q=\lambda [\alpha (T_S - T_N) + \frac{\beta}{100} (\tilde{S}_N - \tilde{S}_S)],$$ and the following ODEs: For $q \geq 0$ we have $$\label{eq:AOMC_model_scaled_q>0}
\left.\begin{array}{rcl}
\frac{V_N}{Y} \frac{d\tilde{S}_N}{d\tau} &=& q(\tilde{S}_T - \tilde{S}_N) + K_N (\tilde{S}_T - \tilde{S}_N) - 100 F_N S_0, \\
\frac{V_T}{Y} \frac{d\tilde{S}_T}{d\tau} &=& q[\gamma \tilde{S}_S + (1- \gamma) \tilde{S}_{IP} - \tilde{S}_T] + K_S (\tilde{S}_S - \tilde{S}_T)+K_N(\tilde{S}_N - \tilde{S}_T) - 100 F_T S_0,\\
\frac{V_S}{Y} \frac{d\tilde{S}_S}{d\tau} &=& \gamma q (\tilde{S}_B -\tilde{S}_S) + K_{IP} (\tilde{S}_{IP} - \tilde{S}_S) + K_S (\tilde{S}_T - \tilde{S}_S) +\eta (\tilde{S}_B - \tilde{S}_S) - 100 F_S S_0, \\
\frac{V_{IP}}{Y} \frac{d\tilde{S}_{IP}}{d\tau} &=& (1-\gamma)q(\tilde{S}_B - \tilde{S}_{IP}) + K_{IP} (\tilde{S}_S - \tilde{S}_{IP}) - 100 F_{IP} S_0, \\
\frac{V_{B}}{Y} \frac{d\tilde{S}_B}{d\tau} &=& q (\tilde{S}_N - \tilde{S}_B) + \eta (\tilde{S}_S - \tilde{S}_B).
\end{array}\right\}$$ while for $q < 0$ we have $$\label{eq:AOMC_model_scaled_q<0}
\left.\begin{array}{rcl}
\frac{V_N}{Y} \frac{d\tilde{S}_N}{d\tau} &=& |q|(\tilde{S}_B - \tilde{S}_N) + K_N (\tilde{S}_T - \tilde{S}_N) - 100 F_N S_0, \\
\frac{V_T}{Y} \frac{d\tilde{S}_T}{d\tau} &=& |q| (\tilde{S}_N - \tilde{S}_T) + K_S (\tilde{S}_S - \tilde{S}_T)+K_N(\tilde{S}_N - \tilde{S}_T) - 100 F_T S_0,\\
\frac{V_S}{Y} \frac{d\tilde{S}_S}{d\tau} &=& \gamma |q| (\tilde{S}_T -\tilde{S}_S) + K_{IP} (\tilde{S}_{IP} - \tilde{S}_S) + K_S (\tilde{S}_T - \tilde{S}_S) +\eta (\tilde{S}_B - \tilde{S}_S) - 100 F_S S_0, \\
\frac{V_{IP}}{Y} \frac{d\tilde{S}_{IP}}{d\tau} &=& (1-\gamma)|q|(\tilde{S}_T - \tilde{S}_{IP}) + K_{IP} (\tilde{S}_S - \tilde{S}_{IP}) - 100 F_{IP} S_0, \\
\frac{V_{B}}{Y} \frac{d\tilde{S}_B}{d\tau} &=& \gamma |q| \tilde{S}_S + (1-\gamma) |q| \tilde{S}_{IP} - |q| \tilde{S}_B + \eta (\tilde{S}_S -\tilde{S}_B).
\end{array}\right\}$$ Note that the total salt content is still conserved with $$C= \frac{1}{100}(V_N \tilde{S}_N+ V_T \tilde{S}_T+ V_S \tilde{S}_S+ V_{IP} \tilde{S}_{IP} + V_B \tilde{S}_B)+S_0(V_N+V_T+V_S+V_{IP}+V_B).
\label{eq:totalsalt_scaled}$$ We have only shown the scaled version of the five box model but the three box model is scaled equally.
Supplementary Movies
====================
Supplementary Movie [MOC\_basin\_2co2.mp4]{} shows animations of Figure \[fig:threebox\_basins\_H\]. Supplementary Movies [MOC\_phas\_Btip.mp4]{} and [MOC\_phas\_Btrack.mp4]{} show animations corresponding to the red and black trajectories respectively for Figure \[fig:Btip\]. Supplementary Movies [MOC\_phas\_Rtip.mp4]{} and [MOC\_phas\_Rtrack.mp4]{} show animations corresponding to the red and black trajectories respectively for Figure \[fig:Rtip\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Precision measurements of parity non-conserving (PNC) interactions in atoms, molecules and ions can lead to the discovery of new physics beyond the standard model and understanding of weak-force induced interactions in the nucleus. In this paper, we propose and analyze a novel atomic parity violation measurement scheme for a forbidden transition where we combine a two-pathway coherent control mechanism with probe gain techniques. We detail a feasible experimental geometry for $6S_{1/2}\rightarrow 7S_{1/2}$ transitions in a cesium vapor cell, and consider the statistical noise of such a measurement under reasonable laboratory conditions. We estimate the signal-to-noise ratio to be approaching $\sim2.3/\sqrt{Hz}$. This scheme, with low expected systematic errors, would allow for precise measurements in cesium and other heavy metal systems.'
author:
- 'J. Choi$^{1,3}$, R. T. Sutherland[^1] $^{2}$, George Toh$^{1,3}$, A. Damitz$^{2}$ and D. S. Elliott$^{1,2,3}$'
bibliography:
- 'biblio.bib'
title: 'Gain measurement scheme for precise determination of atomic parity violation through two-pathway coherent control'
---
Introduction
============
Precision measurements of the weak interaction, first proposed in atoms by Bouchiat and Bouchiat [@BouchiatB74a; @BouchiatB74b], have been observed and are currently in progress on various parity non-conserving (PNC) transitions in numerous media including atoms, molecules, and ions. To date, the most accurate result is from the Boulder group’s 1997 experiment [@WoodBCMRTW97] in an atomic cesium beam measurement with a 0.35% uncertainty. This measurement, in concert with precise theoretical models of the cesium atom [@PorsevBD09; @PorsevBD10; @DzubaBFR12], allows for a precise determination of the weak charge $Q_w$. The theoretical efforts of Refs. [@PorsevBD09; @PorsevBD10; @DzubaBFR12] have yielded a sub-0.5% uncertainty calculation, and further development of an improved atomic structure model of cesium is underway [@Ginges2017; @Ginges2017b]. The Boulder group’s experiment also produced a measurement of the nuclear anapole moment, which results from the weak force within the nucleus [@Zel1958; @Flam1984]. Their nuclear-spin-dependent (NSD) measurement, however, is at odds with other measurements of the anapole moment, as discussed in Refs. [@Bouchiat1991; @FlambaumM97; @HaxtonW01; @Ginges2004]. No other significant determinations of nuclear anapole moments in atomic systems have been reported. In short, a new measurement of PNC transitions with a lower uncertainty is needed for probing of physics beyond the standard model [@MarcianoR90; @HaxtonW01; @Rosner02; @SafronovaBDKDC18] and resolving the discrepancy the Boulder group reported in their measurement of the nuclear anapole moment.
Several programs have recently reported exciting progress in high precision weak measurements. Antypas et al. [@AntypasFSTB18] reported 0.5% uncertainty measurements in the $6s^2~^1S_0\rightarrow5d6s^3D_1$ transition at 408 nm in four isotopes of ytterbium to show the dependence of the weak interaction on the neutron number. Their effort to measure the weak NSD interaction continues [@Antypas2017]. The TRIUMF collaboration [@Sheng2010; @AubinGBPSZCMFSOG11] have been developing techniques for trapping francium and have carried out preliminary spectroscopic measurements of this unstable alkali metal atom. Their goal is to probe the weak interaction in a chain of trapped francium isotopes. At Yale, the DeMille group has recently reported [@AltuntasACD18a; @AltuntasACD18b] progress in characterization and suppression of systematic effects in $^{138}$Ba$^{19}$F polar molecule measurements toward the weak NSD measurement in $^{137}$BaF.
Since the PNC transitions are so weak, their measurement must in each case be carried out using interference with a relatively stronger transition (e.g. magnetic dipole ($M1$), electric quadrupole ($E2$), or Stark-induced transitions). Optical rotation via PNC and $M1$ interference was carried out with a 1% uncertainty in thallium [@EdwardsPBN95; @VetterMMLF95] and in lead [@MeekhofVMLF95], and with a 2% uncertainty in Bismuth [@MacphersonZWSH91]. In atomic beam measurements, a Stark-PNC interference technique was used in cesium (e.g. [@WoodBCMRTW97]) and ytterbium (e.g. [@AntypasFSTB18; @Antypas2017]) with modulation of net transition rates detected through fluorescence detection. In addition, the group of M. Bouchiat [@Guena2003] has developed a pump-probe Stark-PNC interference technique for measurements in cesium where a high intensity pulse excites the forbidden transition and a moderate pulse probes the population asymmetry in the excited states via gain polarization rotation detection. This scheme has yielded 2.6% uncertainty measurements [@GuenaLB05a; @GuenaLB05b]. More recently, our group has developed a two-color coherent control scheme where an additional laser is added to “strongly” excite the weak transitions. This technique displayed shot-noise-limited detection in measurements of a weak $6S_{1/2}\rightarrow 8S_{1/2}$ Stark-induced transition [@GunawardenaE07a; @GunawardenaE07b], and was used to measure the magnetic dipole moment $M_1$ on the $6S_{1/2}\rightarrow 7S_{1/2}$ transition [@AntypasE13a; @AntypasE14] in cesium. We are also working on a two-color optical and rf interference experiment to directly probe the NSD interaction in the cesium ground hyperfine states [@Choi2016].
The novel technique that we outline in this paper is a pump-probe gain scheme where we excite the weak transition via two-color interfering interactions with cw lasers and directly monitor the excitation rate with a cw probe field through a stimulated emission process. It involves interference between a strong two-photon and weak one-photon (Stark-induced and PNC) transitions. The primary observable in this scheme is the modulation amplitude of the probe gain signal as a function of the relative phase difference between the two-photon and one-photon transitions. This gain differs in several ways from that observed previously [@Guena2003; @GuenaLB05a; @GuenaLB05b]. First, two-pathway coherent control techniques allow for direct modulation of the gain signal. Secondly, it is not based upon the asymmetry of the population of the excited state and, hence, the observable is not the rotation angle of the optical polarization of the probe beam. And finally, our scheme involves cw rather than pulsed lasers.
The paper is organized as follows; in Sec. \[sec:int\], we detail the two-color coherent control technique for a novel PNC-Stark interference measurement; in Sec. \[sec:gain\], we describe the pump-probe gain scheme in cesium with reasonable experimental parameters for $6S_{1/2}\rightarrow 7S_{1/2}$ transitions; in Sec. \[sec:analysis\] we analyze the systematic and statistical errors; and we summarize our findings in Sec. \[sec:sum\].
![(Color online) Abbreviated energy level diagram of cesium. The two-color excitation lasers excite the $|1 \rangle \rightarrow |2 \rangle$ ($6s \: ^2S_{1/2} \rightarrow 7s \: ^2S_{1/2}$) transition. One may interrogate the $7s$ excitation by measuring the gain in the probe laser beam, which is tuned to the $|2 \rangle \rightarrow |3 \rangle$ ($7s \: ^2S_{1/2} \rightarrow 6p \: ^2P_{3/2}$ or $7s \: ^2S_{1/2} \rightarrow 6p \: ^2P_{1/2}$) transition. The optical-phase-dependent population of the $7s \: ^2S_{1/2}$ state results in a modulation of the probe laser gain. []{data-label="fig:Cs_energy_levels_gain_exp"}](Cs_energy_levels_gain_exp.png "fig:"){width="7.5cm"}\
Interfering Interactions {#sec:int}
========================
We show a simplified energy level diagram of the cesium atom in Fig. \[fig:Cs\_energy\_levels\_gain\_exp\].
{width="11cm"}\
The mutually-coherent excitation laser field components, at wavelengths of $\lambda = 1079$ nm and 539.5 nm, drive the $6s \: ^2S_{1/2} \rightarrow 7s \: ^2S_{1/2}$ transition. We label these levels $|1 \rangle$ and $|2 \rangle$, respectively. The probe laser, which propagates parallel to the excitation beams, will experience gain when its frequency is resonant with the $|2 \rangle \rightarrow |3 \rangle$ transition due to the population in level $|2 \rangle$. State $|3 \rangle$ can be either the $6p \: ^2P_{3/2}$ level at a probe wavelength of $\lambda^{pr} = 1.47 \ \mu$m or the $6p \: ^2P_{1/2}$ level at $\lambda^{pr} = 1.36 \ \mu$m.
We have considered several different potential measurement geometries in order to evaluate their utility in this type of gain measurement. Several requirements must be satisfied. First, the two excitation beams (the 1079 nm beam and the 540 nm beam) must propagate co-linearly in order to maintain a constant phase difference between the various transition amplitudes for excitation of the $7s$ state throughout the interaction region. Second, $\Delta F$ and $\Delta m$ for each of these transitions must be the same so that the amplitudes interfere with one another. ($F$ and $m$ are the quantum numbers representing the total angular momentum and its projection onto the $z$-axis, respectively). After consideration of the selection rules for two-photon, Stark-induced, and PNC-induced transitions, with various states of laser polarization, we have determined that the static electric field $\mathbf{E}_0$ (that is, the Stark-mixing field) must be perpendicular to the propagation direction of the excitation lasers, and that the electric field polarization $\mbox{\boldmath$\varepsilon$}^{gr}$ of the green beam (at 540 nm) must be parallel to the static field $\mathbf{E}_0$. We assign this direction as the $z$-direction, and show the experimental geometry in Fig. \[fig:exp\_geom\_gain\]. For this geometry, the projection quantum number $m$ does not change for any of the excitations; that is, only $\Delta m = 0$ excitations are allowed. Similarly, only $\Delta m = 0$ transitions are allowed in this two-photon excitation using equal frequency photons, regardless of the polarization of the 1079 nm beam [@Cagnac1973].
The total transition amplitude for excitation of the $7s \: ^2S_{1/2}$ state is the sum of amplitudes for the individual distinct interactions. We show a representation of these amplitudes in Fig. \[fig:PhasorAmps\], including the two-photon amplitude $A_{2p}$ driven by the 1079 nm laser (represented by the long, red solid arrow); a Stark-induced amplitude $A_{St}$ driven by the 540 nm beam, (the intermediate length, green, dashed arrow); and a PNC amplitude $A_{PNC}$, also driven by the green laser (the short, blue, dotted arrow).
![(Color online) Representations of the transition amplitudes for excitation of two hyperfine components ($m<0$ on the left, $m>0$ on the right) of the cesium $7s \: ^2S_{1/2}$ state. The long red solid arrow is the two-photon amplitude $A_{2p}$ driven by the 1079 nm laser, while the green dashed arrow and blue dotted arrow show the Stark-induced amplitude $A_{St}$ and the PNC amplitude $A_{PNC}$, both driven by the 540 nm beam. $\Delta \phi = 2\phi^{IR} -\phi^{gr} $ is the phase difference between the IR and green beams. Note that the PNC amplitudes for the two diagrams are reversed, resulting in different net transition amplitudes of states $m<0$ and $m>0$. []{data-label="fig:PhasorAmps"}](PhasorAmps.png "fig:"){width="5cm"}\
For the geometry of the experiment described above, these various transition amplitudes can be written as $$\label{eq:A2pdef}
A_{2p} = \tilde{\alpha} \left( \varepsilon^{IR} \right)^2 e^{-2 i \phi^{IR}}$$ for two-photon excitation, $$\label{eq:AStdef}
A_{St} = \alpha E_0 \varepsilon^{gr} e^{- i \phi^{gr}}$$ for Stark-induced excitation, and $$\label{eq:APNCdef}
A_{PNC} = i \Im ( \mathcal{E}_{PNC}) C_{Fm}^{Fm} \varepsilon^{gr} e^{- i \phi^{gr}}$$ for the weak-force-induced amplitude, where $\mathcal{E}_{PNC}$ is the purely imaginary dipole moment induced by the weak force, and $\Im$ indicates the imaginary part. We use the notation of Gilbert and Wieman [@GilbertW86] for these transition amplitudes. $\varepsilon^{IR}$, $\varepsilon^{gr}$, and $E_0 $ represent the field amplitudes of the 1079 nm beam, the 539.5 nm beam, and the static electric field, respectively. We include the phases $\phi^{IR}$ and $\phi^{gr}$ of the time-varying fields, since these parameters are critical to the coherent sum of the amplitudes. The parameter $\alpha$ is the scalar Stark polarizability (see, for example, Ref. [@VasilyevSSB02]), and $\tilde{\alpha}$ is the two-photon moment. The Stark polarizability, calculated as $$\begin{aligned}
\alpha &=& \frac{e^2}{6} \: \sum_n \left[\rule{0in}{0.25in} \langle 7s || r || np_{1/2} \rangle \langle np_{1/2} || r || 6s \rangle \right. \nonumber \\
&& \left( \frac{1}{E_{7s} - E_{np_{1/2}}} + \frac{1}{E_{6s} - E_{np_{1/2}}} \right) \nonumber \\
&& \hspace{0.4in} - \rule{0in}{0.3in} \langle 7s || r || np_{3/2} \rangle \langle np_{3/2} || r || 6s \rangle \nonumber \\ && \left. \left( \frac{1}{E_{7s} - E_{np_{3/2}}} + \frac{1}{E_{6s} - E_{np_{3/2}}} \right) \right], \nonumber\end{aligned}$$ has played a central role in the determination of the weak charge $Q_w$ of the cesium atom. Its value (in atomic units) using the latest experimental [@YoungHSPTWL94; @RafacT98; @RafacTLB99; @AminiG03; @DereviankoP02; @BouloufaCD07; @ZhangMWWXJ13; @AntypasE13b; @Tanner14; @PattersonSEGBSK15] or theoretical [@VasilyevSSB02; @SafronovaSC16] values available for electric dipole matrix elements in cesium, is $\alpha = -264.4 \: (6) \: a_0^3$, where $a_0$ is the Bohr radius. (To convert to SI units, divide by $6.06511 \times 10^{40}$ J$^{-1}$(V/m)$^2$.) The two-photon moment, using the perturbation expansion for the two-photon interaction for a one-color laser beam tuned far from any one-photon interactions in Ref. [@Loudon] is $$\begin{aligned}
\label{eq:alphatilde}
\tilde{\alpha} &=& \frac{e^2}{6} \: \sum_n \left[ \frac{\langle 7s || r || np_{1/2} \rangle \langle np_{1/2} || r || 6s \rangle }{ E_{np_{1/2}} - \hbar \omega} \right. \nonumber \\
&& + \left. \frac{\langle 7s || r || np_{3/2} \rangle \langle np_{3/2} || r || 6s \rangle }{ E_{np_{3/2}} - \hbar \omega} \right] \nonumber\end{aligned}$$ The numerical value of the two-photon moment in our geometry is $\tilde{\alpha} = 1006 \: (2) \: a_0^3$. The coefficients $ C_{Fm}^{F'm'} $ come from the Clebsch-Gordon coefficients, and for the transitions of interest here are $$C_{Fm}^{F'm'} = \left\{ \begin{array}{cc}
+m/4 & \mbox{\hspace{0.5in} for $F=4$} \\ -m/4 & \mbox{\hspace{0.5in} for $F=3$}
\end{array} \right.$$ In addition to the two-photon, Stark, and PNC moments, the $6s \: ^2S_{1/2} \rightarrow 7s \: ^2S_{1/2}$ transition is also active through a magnetic dipole interaction and an electric quadrupole interaction. The former can be suppressed (with effort and care) using counter-propagating excitation beams, as discussed in general in Ref. [@BouchiatGHP82; @GilbertNWW85], and for two-pathway coherent control in particular in Ref. [@AntypasE14]. The latter is not active on a $\Delta F = 0$ transition. We note that it will also be necessary to inhibit the two-photon Doppler-free transition, as this signal cannot interfere with the Stark-induced or weak-force induced transition. The simplest means of doing this will be to use orthogonal polarizations for the two counter-propagating 1079 nm laser fields.
From these expressions, one can identify the key characteristics of the phasor representations of the transition amplitudes shown in Fig. \[fig:PhasorAmps\]. Under the conditions that we propose here, the two-photon amplitude $A_{2p}$ is much larger than $A_{St}$ or $A_{PNC}$. (The relative lengths of the Stark and PNC amplitudes are magnified in Fig. \[fig:PhasorAmps\] for visibility. They would be much smaller in practice.) The amplitude of $A_{St}$ is controllable, through variation of the static field strength $E_0$. As the phase difference $\Delta \phi = 2\phi^{IR} -\phi^{gr}$ between the green and infra-red beams is varied, the phase of $A_{St} + A_{PNC}$ relative to that of $A_{2p}$ varies, and the interference can be varied between constructive and destructive. That is, the net transition amplitude contains a large dc term (due to the two-photon amplitude alone), plus a small contribution that varies sinusoidally with phase $\Delta \phi$.
When a probe laser is tuned to the $7s \: ^2S_{1/2} \rightarrow 6p \: ^2P_{J}$ transition where $J$ = 3/2 or 1/2, this probe laser will stimulate a transition to the $6p \: ^2P_{J}$ state, and will be amplified as a result. The gain of this beam depends on the population of the $7s \: ^2S_{1/2}$ state, and varies with the magnetic component $m$, the amplitude of the electric field $E_0$, and the phase difference $\Delta \phi$. This gain is the basis for the measurement technique described here.
Estimate of probe laser gain {#sec:gain}
============================
In this section, we will evaluate the magnitude of the gain coefficient of the probe beam resulting from the population of the $7s \ ^2S_{1/2}$ state. When driven concurrently by the three interactions introduced above, the total excitation rate of a ground state atom to the $7s \: ^2S_{1/2}$ state is $$\mathcal{R} = \frac{2 \pi}{\hbar} | A_{2p} + A_{St} + A_{PNC} |^2 \tilde{\rho}_{7s}(E),$$ where $\tilde{\rho}_{7s}(E)$ is the density of states of the $7s \: ^2S_{1/2}$ state. On resonance, the density of states is $\tilde{\rho}_{7s}(0) = 2 / (\pi \hbar \Gamma)$, where $\Gamma$ is the decay rate of the $7s$ state, so the transition rate is $$\mathcal{R} = \frac{4}{\hbar^2 \Gamma} | A_{2p} + A_{St} + A_{PNC} |^2 .$$ The decay rate $\Gamma$ is $\tau_{7s}^{-1}$, where $\tau_{7s} = 48.28$ ns is the lifetime of the $7s$ state [@TohJGQSCWE18]. In steady state, the probability that an atom is in the excited state is $$\label{eq:rhoeqsqofsum}
\rho_{22} = \frac{\mathcal{R}}{\Gamma} = \frac{4 }{\hbar^2 \Gamma^2} | A_{2p} + A_{St} + A_{PNC} |^2 .$$
As shown in Meystre and Sargent [@MeystreS], the gain coefficient for the probe beam tuned to the frequency of the $|2\rangle \rightarrow |3\rangle$ transition $$\label{eq:gainceffmrho32}
\gamma = \frac{i \mu_{23} k}{\varepsilon_0} \frac{\rho_{32}}{\varepsilon^{pr}} \: n,$$ where $\mu_{23}$ is the electric dipole transition moment for the $|2 \rangle \rightarrow |3 \rangle$ transition, $k = 2\pi/\lambda^{pr}$ is the wavenumber of the probe beam, $\varepsilon_0 $ is the vacuum permittivity, $\rho_{32} $ is the off-diagonal matrix element for the probe transition, and $n$ is the number density of the cesium atoms. (The sign of Eq. (\[eq:gainceffmrho32\]) is opposite that given in Ref. [@MeystreS], since they present the *absorption* coefficient for the transition.) In steady-state, the coherence term of the density matrix for a three-level system is $$\label{eq:rho32a}
\rho_{32} = \frac{1} {\hbar} \frac{\mathcal{V}_{32} \mathcal{D}_{32} \left[ i \left( \rho_{33} - \rho_{22} \right) + \mathcal{D}_{31} \mathcal{V}_{12} \rho_{21} \right] }{1 + | \mathcal{V}_{12} / \hbar |^2 \mathcal{D}_{31} \mathcal{D}_{32}} ,$$ where $$\mathcal{D}_{ij} = \frac{1}{\gamma_{ij} + i \Delta_{ij}}$$ describes the variation of the atomic response with detuning from resonance $\Delta_{ij}$, $$\mathcal{V}_{ij} = \mu_{ij} \varepsilon$$ is the interaction energy with the laser field, and $\gamma_{ij}$ is the decay rate of the atomic coherence. When collisional effects are small (as in a low-density vapor cell or in an atomic beam), one can substitute $\gamma_{ij} \rightarrow \Gamma/2$. On resonance, therefore, where $\Delta_{ij}=0$, $\mathcal{D}_{32}$ is $\sim 2/\Gamma$. Presuming that ($i$) the probe laser intensity is below its saturation intensity $I_{sat}^{pr}$, and ($ii$) contributions to the gain from the second (coherence) term in the numerator of Eq. (\[eq:rho32a\]) are negligible, the off-diagonal element $\rho_{32}$ is $$\label{eq:rho32b}
\rho_{32} = \frac{ 2 i \mu_{32} \varepsilon^{pr}}{ \hbar \Gamma} \left( \tau_{6p} \ \Sigma \Gamma_{2 \rightarrow 3} -1 \right) \rho_{22}.$$ We have used the probability that an atom is in the $6p$ state following spontaneous decay from level $|2 \rangle$ is $\rho_{33} \sim \tau_{6p} (\Sigma \Gamma_{2 \rightarrow 3} ) \rho_{22}$, valid for low probe intensity. $\tau_{6p}$ is the lifetime of the cesium $6p$ state, and $\Sigma \Gamma_{2 \rightarrow 3}$ is the total spontaneous decay rate leading to population in the hyperfine component of level $|3 \rangle$ ($6p ^2P_{1/2}$ or $6p ^2P_{3/2}$) coupled by the probe beam to level $|2 \rangle$. The lifetime of the $6p \: ^2P_{3/2}$ state is $30.42 $ ns, and of the $6p \: ^2P_{1/2}$ state is $34.83 $ ns [@YoungHSPTWL94; @RafacT98; @RafacTLB99; @AminiG03; @DereviankoP02; @BouloufaCD07; @ZhangMWWXJ13; @PattersonSEGBSK15]. Inserting Eq. (\[eq:rho32b\]) for the off-diagonal matrix element into Eq. (\[eq:gainceffmrho32\]), the gain coefficient becomes $$\label{eq:gamma_rho22}
\gamma = \frac{2 n k}{\hbar \Gamma \varepsilon_0} | \mu_{32} |^2 \left( 1 - \tau_{6p} \ \Sigma \Gamma_{2 \rightarrow 3} \right) \rho_{22} .$$
![(Color online) Energy levels showing hyperfine states.[]{data-label="fig:Cs_6s7s6p_levels_gain"}](Cs_6s7s6p_levels_gain.png "fig:"){width="7.5cm"}\
In our multilevel system, with degenerate levels for the $6s$, $7s$, and $6p$ states, the PNC contributions to the gain coefficient from the different $m$-states tend to cancel one another. (Due to the factor $C_{Fm}^{F^{\prime}m^{\prime}}$ in the PNC amplitude, $A_{PNC}$ for the hyperfine component $m>0$ and $A_{PNC}$ for the hyperfine component $m<0$ are opposite in sign.) Therefore, we need to lift the degeneracy of the individual projection states by applying a dc magnetic field $\mathbf{B} = \mathbf{a}_z B_0$ to the atoms, and measure the gain on just one Zeeman component of the probe transition. The Zeeman shift of each level is given by $\Delta E = - m g_F \mu_B B_0$, where $g_F$ is the Landé g-factor and $\mu_B = 9.274 \times 10^{–-24}$ J/T is the Bohr magneton. We label the Landé g-factors for each of the relevant levels in Fig. \[fig:Cs\_6s7s6p\_levels\_gain\]. Note that the g-factors are the same for levels $|1\rangle$ and $|2\rangle$, so the transition frequency for the $\Delta F = 0$, $\Delta m = 0$ excitation of the $7s \ ^2S_{1/2}$ state is insensitive to the application of $\mathbf{B}$. For individual probe transitions $|2\rangle \rightarrow |3\rangle$, the Zeeman shifts of the upper and lower states differ, and the transition frequencies for the individual lines are separated from one another.
The Zeeman splitting $\Delta \nu_Z$ between adjacent lines of the probe spectrum is the least of $| g_{F^{\prime}} - g_{F^{\prime \prime}}| \mu_B B_0/h$, $2 |g_{F^{\prime \prime}}| \mu_B B_0/h $, and $| 3g_{F^{\prime}} - g_{F^{\prime \prime}}| \mu_B B_0/h$, where $h$ is Planck’s constant. We use primed notation for $7s ^2S_{1/2}$ quantities and double-primed notation for $6p ^2P_{J}$ quantities. This splitting must be larger than the linewidth of the transition, which for the probe laser is limited by the natural linewidth of the transition, $\sim$8 MHz. (The probe transition is Doppler free, even when the measurements are carried out in a heated vapor cell, since only atoms whose longitudinal velocity $v_y $ is zero are excited by the excitation fields when the frequency of the excitation field is tuned to line center.) Inserting the excited state population $\rho_{22}$ from Eq. (\[eq:rhoeqsqofsum\]), and assuming that the initial state population is uniformly distributed over each of the 16 hyperfine components of the ground state, the gain coefficient becomes $$\gamma = K_0 \chi | \mu_{32} |^2 | A_{2p} + A_{St} + A_{PNC} |^2 , \label{eq:gaincoefsqrs}$$ where the factor $K_0$ is $$K_0 =\frac{ n k }{2 (\hbar \Gamma)^3 \varepsilon_0} \label{eq:gaincoefkappa}$$ and $$\chi = 1 - \tau_{6p} \ \Sigma \Gamma_{2 \rightarrow 3} .$$
Since $A_{2p} \gg A_{St}$ and $A_{PNC}$, when we expand the square in Eq. (\[eq:gaincoefsqrs\]), we can drop the terms that contain $A_{St}^2$, $A_{St} A_{PNC}$, and $A_{PNC}^2$, and write $$\begin{aligned}
\gamma &=& K_0 \chi | \mu_{32} |^2
\left\{\rule{0in}{0.17in} | A_{2p}|^2 \right. \nonumber \\ && \left. + \left( A_{2p} \left[ \rule{0in}{0.13in} A_{St} + A_{PNC} \right]^{\ast} + c.c. \right) \rule{0in}{0.17in}\right\}. \nonumber\end{aligned}$$ The gain coefficient $\gamma$, which consists of three terms, depends on the polarization of the probe beam, the phase difference between the IR and green beams, and the direction and magnitude of the applied dc electric field $E_0$. We write the gain coefficient as $$\gamma = \gamma_{2p} + \gamma_{St} + \gamma_{PNC},$$ where $\gamma_{2p}$ is the gain coefficient due to the two-photon excitation of the 7s state alone, $$\gamma_{2p} = K_0 \chi | \mu_{32} |^2 | A_{2p}|^2 ,$$ $\gamma_{St}$ is the gain coefficient that arises from the interference between the two-photon excitation and the Stark-induced excitation $$\gamma_{St} = K_0 \chi | \mu_{32} |^2
\left\{\rule{0in}{0.17in} | A_{2p}| \ \alpha E_0 \varepsilon^{gr} e^{ i \Delta \phi } + c.c. \rule{0in}{0.17in} \right\},$$ and $\gamma_{PNC}$ is the gain coefficient that arises from the interference between the two-photon excitation and the weak-force-induced excitation $$\begin{aligned}
\gamma_{PNC} &=& K_0 \chi | \mu_{32} |^2 \left\{\rule{0in}{0.17in} | A_{2p}| \ i \Im ( \mathcal{E}_{PNC}) \right. \nonumber \\
& & \left. \hspace{0.7in} C_{Fm}^{F^{\prime}m^{\prime}} \varepsilon^{gr} e^{ i \Delta \phi } + c.c. \rule{0in}{0.17in} \right\}. \nonumber\end{aligned}$$
We list the transition moments, the population factor $\chi$, $ C_{Fm}^{F^{\prime}m^{\prime}}$, and the Zeeman peak separation $\Delta E_Z$ for selected lines in Table \[table:moments\].
[|c|c|c|c|c|c|c|]{} Probe &
------------------------------------------------------------------------
Probe
------------------------------------------------------------------------
&
------------------------------------------------------------------------
$F^{\prime}, m^{\prime} \rightarrow F^{\prime \prime}, m^{\prime \prime}$
------------------------------------------------------------------------
& $\displaystyle \left| \frac{\mu_{32}}{e \langle 7s || r || 6p_{J} \rangle} \right |^2$ &
------------------------------------------------------------------------
------------------------------------------------------------------------
$ \chi $
------------------------------------------------------------------------
&
------------------------------------------------------------------------
------------------------------------------------------------------------
$ C_{Fm}^{F^{\prime}m^{\prime}}$
------------------------------------------------------------------------
&
------------------------------------------------------------------------
$\Delta \nu_Z / B_0 \rule{0.10in}{0in} $\
transition & pol. & & & & & (MHz/G)\
& &
------------------------------------------------------------------------
$4, 4 \rightarrow 5, 5$ & 1/4 = 0.250 & $0.796$ & +1 & 0.210\
$7s \: ^2S_{1/2} \hspace{0.15in}$ & horiz. & $4, 3 \rightarrow 4, 4$ &
------------------------------------------------------------------------
7/240 = 0.029
------------------------------------------------------------------------
& $0.881$ & +3/4 & 0.023\
$\rightarrow 6p \: ^2P_{3/2}$ & & $3, 3 \rightarrow 2,2$ & 5/28 = 0.179 & $0.796$ & -3/4 & 0.117\
(1.47 $\mu$m) & vert. &
------------------------------------------------------------------------
$4, 4 \rightarrow 4, 4$ & 7/60 = 0.117 & $0.881$ & +1 & 0.023\
& & $3,3 \rightarrow 3,3$ & 9/64 = 0.141 & $0.847$ & -3/4 & 0.350\
$7s \: ^2S_{1/2}\hspace{0.15in} $ & horiz. & $4, 4 \rightarrow 3, 3$& 7/24 = 0.292 & $0.810$ &
------------------------------------------------------------------------
+1 & 0.467\
$\rightarrow 6p \: ^2P_{1/2}$ & & $3,3 \rightarrow 4,4$ & 7/24 = 0.292 & $0.852$ & -3/4 & 0.233\
(1.36 $\mu$m) & vert. &
------------------------------------------------------------------------
$4, 4 \rightarrow 4,4 $ & 1/6 = 0.167 & $0.894$ & +1 & 0.233\
As an example, we consider in detail the gain on the $7s \: ^2S_{1/2}, \: (4, \ 4) \rightarrow 6p \: ^2P_{3/2}, \: (5, \ 5)$ probe transition. (The numbers enclosed within the parentheses are $F$ and $m$ for the two states.) The PNC gain is largest on this line, and the Zeeman peak separation, while not maximal, is sufficient. The individual gain coefficients are $$\label{eq:gamma2p}
\gamma_{2p} = K_0 \chi \left( \frac{ e^2}{4} | \langle 7s || r || 6p_{3/2} \rangle |^2 \right) | A_{2p}|^2 ,$$ $$\begin{aligned}
\label{eq:gammaSt}
\gamma_{St} &=& K_0 \chi \left( \frac{ e^2}{4} | \langle 7s || r || 6p_{3/2} \rangle |^2 \right) \nonumber \\
&& \times | A_{2p}| \alpha E_0 \varepsilon^{gr} 2 \cos( \Delta \phi ) ,\end{aligned}$$ and $$\begin{aligned}
\gamma_{PNC} &=& K_0 \chi \left( \frac{ e^2}{4} | \langle 7s || r || 6p_{3/2} \rangle |^2 \right) \nonumber \\
& & \hspace{-0.1in} \times | A_{2p}| \Im ( \mathcal{E}_{PNC}) \varepsilon^{gr} (-2) \sin( \Delta \phi ) .\label{eq:gammaPNC}\end{aligned}$$ Each of these gain coefficients contains the factor $$K_0 e^2 |\langle 7s || r || np_{j} \rangle |^2 = \frac{ n k e^2 }{2 (\hbar \Gamma)^3 \varepsilon_0} |\langle 7s || r || np_{j} \rangle |^2 , \nonumber$$ so we will start by evaluating this. We use
- $n = 3.7 \times 10^{12}$ cm$^{-3}$. We derive this effective density using the equilibrium vapor density of cesium at a temperature of 180$^{\circ}$C ($8.3 \times 10^{14}$ cm$^{-3}$), reduced by the factor $\Delta \nu_n / \Delta \nu_D$ for the $6s \rightarrow 7s$ transition, where $\Delta \nu_n$ is the natural linewidth ($\Delta \nu_n$ = 3.3 MHz) for the transition and $\Delta \nu_D$ is its Doppler width ($\Delta \nu_D$ = 750 MHz).
- $k = 2\pi/\lambda^{pr} = 4.27 \times 10^4 $ cm$^{-1}$ for the 1.47 $\mu$m probe beam and $4.62 \times 10^4 $ cm$^{-1}$ for the 1.36 $\mu$m probe beam.
- $\Gamma = \tau_{7s}^{-1} = 2.1 \times 10^7$ s$^{-1}$ [@TohJGQSCWE18].
- $\langle 7s || r || 6p_{3/2} \rangle = 6.487 \: a_0$ and $\langle 7s || r || 6p_{1/2} \rangle = 4.245 \: a_0$, as determined from the lifetime of the $7s$ state $\tau_{7s} = 48.28$ ns [@TohJGQSCWE18] and presuming the ratio of transition moments is $ \langle 7s || r || 6p_{3/2} \rangle / \langle 7s || r || 6p_{1/2} \rangle = 1.528$ [@DzubaFKS89; @BlundellSJ92; @DzubaFS97; @SafronovaJD99].
The common factor $K_0 e^2 |\langle 7s || r || np_{j} \rangle |^2$ is $$\label{eq:multfacjeq3halves}
K_0 e^2 |\langle 7s || r || 6p_{3/2} \rangle |^2 = 2.6 \times 10^{59} \: {\rm J}^{-2}{\rm m}^{-1}$$ for the 1.47 $\mu$m line, and $$\label{eq:multfacjeq1half}
K_0 e^2 |\langle 7s || r || 6p_{1/2} \rangle |^2 = 1.2 \times 10^{59} \: {\rm J}^{-2}{\rm m}^{-1}$$ for the 1.36 $\mu$m line. These pre-factors show that the gain is typically larger on the $7s \: ^2S_{1/2} \rightarrow 6p \: ^2P_{3/2}$ probe transition than on the $7s \: ^2S_{1/2} \rightarrow 6p \: ^2P_{1/2}$ line.
Finally, we must determine the transition amplitudes $A_{2p}$, $A_{St}$, and $A_{PNC}$. Using $\tilde{\alpha} = 1006 \: a_0^3$, we determine the two-photon transition amplitude $A_{2p} = \tilde{\alpha} \left( \varepsilon^{IR} \right)^2 $ and the transition rate $\mathcal{R}_{2p}$. Assuming the power of the 1079 nm beam as 5 W, and a Gaussian beamshape with the radius of the beam as $w^{IR} \sim $ 1.0 mm, the field amplitude on the axis is $$\varepsilon^{IR} = 4.9 \times 10^4 \: {\rm V/m}.$$ Then the two photon amplitude, from Eq. (\[eq:A2pdef\]), is $$\label{eq:A2pval}
A_{2p}/\hbar = 3.8 \times 10^5 \: {\rm s}^{-1},$$ the two-photon excitation rate $\mathcal{R}_{2p}$ alone (that is, without the Stark or PNC contributions) is $$\mathcal{R}_{2p} = \frac{4}{\Gamma} \left(\frac{ A_{2p}}{\hbar} \right)^2 = 2.8 \times 10^4 \: {\rm s}^{-1},$$ and the net probability of finding an atom in the excited state is $$\mathcal{R}_{2p} \tau_{7s} \simeq 0.0013.$$ As required to avoid saturation effects, this probability is much less than unity.
We use the values for $K_0 e^2 |\langle 7s || r || np_{j} \rangle |^2 $ (Eq. (\[eq:multfacjeq3halves\])) and $A_{2p}$ (Eq. (\[eq:A2pval\])) to find the on-axis gain coefficient $\gamma_{2p}$. $$\label{eq:gamma2pval}
\gamma_{2p} = 81 \: {\rm m}^{-1} . \nonumber$$ We define the gain factor due to the two-photon excitation alone as $$G_{2p} \equiv \gamma_{2p} \ell_{\rm gain} = 8.1,$$ where we use a vapor cell length $\ell_{\rm gain}$ = 10 cm, assuming good beam overlap between the probe beam and the excitation beams over the full length.
We next evaluate the gain coefficient $\gamma_{PNC}$ which comes from the interference between the two-photon interaction and the PNC interaction. For this, we need the PNC moment $\Im ( \mathcal{E}_{PNC})$ and the field amplitude $\varepsilon^{gr}$ of the green beam. (The product of these two terms gives us the amplitude $|A_{PNC}|$ of the $6s \: ^2S_{1/2}, \: (4, \ 4) \rightarrow 7s \: ^2S_{1/2}, \: (5, \ 5)$ component.) The calculated value of $\mathcal{E}_{PNC}$ is about $i 0.9 \times 10^{-11} (-Q_w/N) \: ea_0$, where $(Q_w \sim - 76$ is the weak charge and $N = 78$ is the neutron number of cesium [@PorsevBD09; @PorsevBD10; @DzubaBFR12]. (This is converted to SI units using $e a_0 = 8.4735 \times 10^{-30}$ Cm.) Presuming a green laser power of 3 W and a beam radius of $w^{gr} = w^{IR}/\sqrt{2}$, the field amplitude of the green beam is $$\varepsilon^{gr} = 5.4 \times 10^4 \: {\rm V/m}.$$ The PNC amplitude for the $m=4$ level is then $$\label{eq:APNC}
|A_{PNC}/\hbar| = \Im ( \mathcal{E}_{PNC}) \varepsilon^{gr} /\hbar \approx 0.038 \: {\rm s}^{-1}.$$ Using Eqs. (\[eq:gammaPNC\], \[eq:multfacjeq3halves\], \[eq:A2pval\], and \[eq:APNC\]), the amplitude of the gain coefficient $\gamma_{PNC} $ is therefore $$|\gamma_{\rm PNC}| = 1.6 \times 10^{-5} \: {\rm m}^{-1} , \nonumber$$ and the gain factor is $$G_{\rm PNC} = |\gamma_{\rm PNC}| \ell_{\rm gain} = 1.6 \times 10^{-6}.$$
The gain $G_{\rm St}$ due to the Stark-induced amplitude is variable, since this gain depends on the applied dc electric field $E_0$. In a measurement, one would apply a field $E_0$ that produces a gain $G_{\rm St}$ comparable to the PNC gain. Therefore we rely on our evaluation of $G_{\rm PNC}$, and identify the range of values for $G_{\rm St}$ between zero and $\sim 4 \times G_{\rm PNC}$.
Measurement Scheme {#sec:analysis}
==================
In a measurement of the gain in this system, we need to be able to separate the three gain contributions $G_{\rm 2p}$, $G_{\rm St}$, and $G_{\rm PNC}$ from each other, and to use the Stark gain to calibrate the measurement of the PNC gain. In our proposed scheme, we examine the gain for a horizontally-polarized probe beam passing through the gain region, tuned to the selected Zeeman-shifted hyperfine component of the gain transition. In this geometry, one could detect the increase in the probe power, and exploit the following signatures of the three contributions to the gain to differentiate them. The two-photon gain $G_{2p}$ does not depend on $\Delta \phi$, the phase between the green and IR beams, while $G_{St}$ and $G_{PNC}$ do. And while $G_{St}$ and $G_{PNC}$ each modulate sinusoidally with $\Delta \phi$, these gains are $\pi/2$ out of phase with each other, $$\begin{aligned}
I_{\rm out}^{pr} &=& I_{\rm in}^{pr} \left( e^{G_{\rm 2p} + G_{\rm St}\cos(\Delta \phi) - G_{\rm PNC}\sin(\Delta \phi)} \right)^2 \nonumber \\ & \approx & I_{\rm in}^{pr} e^{2 G_{\rm 2p}} \left( \rule{0in}{0.15in} 1 + 2 G_{\rm St}\cos(\Delta \phi) \right. \nonumber \\
& & \hspace{0.6in} \left. - 2G_{\rm PNC}\sin(\Delta \phi) \rule{0in}{0.15in} \right).\nonumber\end{aligned}$$ Therefore, the amplitude of the sinusoidally-varying modulation of the net gain is the quadrature sum of these individual gain terms, $$G_{\rm mod} = \sqrt{G_{\rm St}^2 + G_{\rm PNC}^2}.$$ That is, the gain modulation amplitude grows hyperbolically with the static field strength $E_0$, as we show in Fig. \[fig:Mod\_Amp\_vs\_E\], and measurements of $G_{\rm mod}$ vs. $E_0$ can yield $\mathcal{E}_{PNC}/\alpha$. This is a feature that we have designed into our previous coherent control schemes as well.
![(Color online) A plot of the amplitude of the modulating gain $G_{\rm mod}$, normalized by $G_{\rm PNC}$, versus the static electric field strength $E_0$. The field amplitude is in units of $\mathcal{E}_{PNC}/\alpha$.[]{data-label="fig:Mod_Amp_vs_E"}](Mod_Amp_vs_E.png "fig:"){width="8.5cm"}\
A complete determination requires repeated measurements at each field strength $E_0$, and measurements at several different values of $E_0$; in all, perhaps 20–25 measurements.
We can estimate the signal-to-noise ratio (S/N) of such a measurement as follows. We assume that the intensity of the probe beam after amplification by the gain medium is one-tenth the saturation intensity $I_{\rm sat}$ of the probe transition, which we estimate as $$I_{\rm sat} \approx \frac{c \varepsilon_0\Gamma^2}{2 e^2 |\langle 7s | r | 6p_{j} \rangle |^2} \sim 0.2 \: {\rm mW/cm}^2$$ for the $7s \: ^2S_{1/2}, \: (4, \ 4) \rightarrow 6p \: ^2P_{3/2}, \: (5, \ 5)$ probe line, where $c$ is the speed of light in vacuum. (The probe beam intensity can influence the precision of the measurement, since saturation effects can suppress the gain. High precision measurements will require a probe intensity much less than $I_{\rm sat}$, but due to the gain within the vapor cell, saturation effects are not an issue at the entrance of the cell, but only towards the exit. The limiting input intensity is difficult to estimate in advance.) This beam will overlap and be parallel to the excitation beams; we choose a probe beam radius that matches that of the 540 nm beam, $w^{\rm pr} = w^{\rm gr}$, so the power of the probe beam after amplification in the vapor cell is $P_{out}^{pr} = I_{in}^{pr} e^{2G_{\rm 2p}} \pi (w^{pr})^2/2 = 0.16 \: \mu {\rm W}$. The photodiode signal $S$ that one would measure is $$S = S_{PD} P_{out}^{pr} \left( 2 \times G_{PNC} \right),$$ where $S_{PD}$ is the sensitivity of the photodiode used to measure the probe beam power (0.9 A/W for InGaAs photodiode from Hamamatsu, for example). This yields $$S = 0.5 \times 10^{-12} \: {\rm A}.$$
The noise of the measurement will be primarily shot noise arising from the large dc component of the probe beam power $P_{out}^{pr}$. This property is a characteristic of heterodyne detection, for which the signal consists of a large dc term (in this case due to two-photon absorption) alone, plus a modulation term due to the interference between the strong (two-photon amplitude) and the weak (Stark and/or PNC amplitudes) term. We estimate the level of the shot noise using $$i_N^2 = 2 e I \Delta \nu,$$ where $e = 1.6 \times 10^{-19} \: {\rm C}$ is the fundamental charge unit, $I = S_{PD} P_{out}^{pr} $ is the average photo-current generated by the photodetector, and $\Delta \nu$ is the bandwidth of the detector. The signal-to-noise ratio is $$S/N = \frac{ 2 S_{PD} P_{out}^{pr} G_{PNC} }{\sqrt{2 e (S_{PD} P_{out}^{pr} ) \Delta \nu}} = \sqrt{ \frac{ 2 S_{PD} P_{out}^{pr}}{ e \Delta \nu}} \: G_{\rm PNC}.$$ Using the parameters we have discussed above, the S/N ratio is $$\label{eq:SNratio}
S/N \sim 2.3\sqrt{t(s)},$$ where $t = 1/\Delta \nu$ is the integration time of a measurement. For example, to achieve a S/N ratio of 100 in a single measurement, an integration time of 21 minutes is required.
While direct comparison of the S/N ratio between measurements is difficult because of the many different parameters used, the potential of the proposed gain measurement technique is promising. The Bouchiat group’s pump-probe gain polarization rotation experiment [@Guena2003; @GuenaLB05a; @GuenaLB05b] yielded an S/N of $\sim0.9\sqrt{t(s)}$ with a reasonable pulse repetition rate and optical density. (We arrived at this expression using the equation in Section F of Ref. [@GuenaLB05a].) The Boulder group’s atomic beam measurement in 1997 [@WoodBCMRTW97; @Wood1999] yielded an S/N ratio of $\sim 0.6\sqrt{t(s)}$, and the ytterbium measurements by Antypas et al. [@AntypasFSTB18] reached a $S/N \sim 0.55 \sqrt{t(s)}$. Both of these measurements employ a power build-up cavity to enhance the field amplitude. We emphasize that we derived Eq. (\[eq:SNratio\]) for the present gain measurement for the case of no power build-up cavity. Further improvement in the S/N is possible with a dual-wavelength (1079 nm and 539.5 nm) power build-up cavity in our gain measurement scheme. A fine-temperature-controlled vapor cell with a minimal reflection loss at the windows [@Jahier2000] can be placed inside a high finesse cavity to increase the PNC gain by several times, which may enhance the S/N by the same factor. Such a cavity geometry may complicate the experimental setup, but we concluded that no additional systematic effects would be introduced due to the cavity.
It is interesting to speculate as to whether this probe gain detection technique can be competitively applied to an atomic beam experiment. We have evaluated the S/N ratio for our current experiment, replacing our current detection scheme based upon fluorescence detection on a cycling transition such as described in Refs. [@WoodBCMRTW97; @AntypasE13a] with a detection scheme based on probe gain, but keeping all other aspects of the experiment, such as beam density, the size and power of the excitation laser, the same, and find that the S/N ratio of a PNC measurement using the probe gain technique can be as much as twice that of the fluorescence detection method.
Atomic parity violation experiments are notorious for their sensitivity to systematic errors introduced by stray uncontrolled dc electric and magnetic fields and small alignment imperfections between $\mathbf{B}$, $\mathbf{E}_0$, and $\mbox{\boldmath$\varepsilon$}^{gr}$. We have previously reported on the expected systematics for this geometry of the excitation beams in Ref. [@AntypasE14]. In that work, we showed that many systematic effects encountered in previous measurements are reduced or eliminated in the two-pathway coherent control schemes through the use of ($i$) low amplitude static electric fields, and ($ii$) modulation of the frequency of the phase difference $\Delta \phi$, combined with phase-sensitive detection. As shown in Sec. IV of Ref. [@AntypasE14], the primary systematics that are expected in the geometry of the proposed measurement are due to stray static electric field components in the $x$-direction $\Delta E_x$ of the form $\alpha \Delta E_x (\varepsilon_x^{gr})^{\prime \prime}$, and $M_1 (\varepsilon_x^{gr})^{\prime \prime}$, where $(\varepsilon_x^{gr})^{\prime \prime}$ is the imaginary component of the 540 nm beam in the $x$-direction. (This component exists only when the optical polarization is slightly elliptical.) Reduction of the former requires weak stray fields $\Delta E_x$, while the latter is reduced by using counter-propagating beams of equal intensity. Both effects are reduced by using a highly-linearly-polarized 540 nm laser beam. We estimate that these unwanted contributions can reasonably be controlled to levels less than $10^{-4}$ of the PNC contribution.
Summary {#sec:sum}
=======
In this report, we have proposed and analyzed a new scheme in which one may detect the gain in a probe laser beam in two pathway coherent control to determine the amplitude of the parity non-conserving weak-force-induced electric dipole amplitude. While we have considered specifically the cesium $6s \ ^2S_{1/2} \rightarrow 7s \ ^2S_{1/2}$ transition in this work, with the probe beam tuned to the $7s \ ^2S_{1/2} \rightarrow 6p \ ^2P_{3/2}$ transition, the probe laser could in principle be tuned to any $7s \ ^2S_{1/2} \rightarrow np \ ^2P_{3/2}$, where $n > 6$, as well. Our choice of $n=6$ is guided by three factors: ($i$) For $n=6$, the probe beam is amplified, rather than attenuated as it would be for $n > 6$. The larger probe beam power after amplification aids in the detection sensitivity. ($ii$) The dipole moment for the transition to $n=6$ is larger, leading to relatively large gain. ($iii$) The wavelength for the probe transition to the $6p \ ^2P_{3/2}$ can be generated with a commercially-available diode laser.
We also suggest that the probe gain technique discussed in this paper could in principle be applied to PNC investigations in other heavy atomic systems. For instance, the TRIUMF collaboration has taken steps to measure the PNC amplitudes in rubidium in parallel with a measurement in a francium MOT [@Sheng2010]. Our technique can be easily adopted for the PNC measurement in a rubidium vapor cell, since it is the second heaviest stable alkali metal.
Acknowledgement
===============
This material is based upon work supported by the National Science Foundation under Grant Numbers PHY-1607603 and 1404419-PHY. We would like to acknowledge fruitful discussions with Marie-Anne Bouchiat, F. Robicheaux, Jun Ye, T. S. Zwier, J. S. M. Ginges, M. Safronova, and Jay Meikrantz at Precision Glassblowing.
[^1]: Current address: Lawrence Livermore National Lab, 7000 East Avenue, Livermore, CA, 94550
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $G$ be a simple algebraic group over an algebraically closed field with Lie algebra ${\mathfrak{g}}$. Then the orbits of nilpotent elements of ${\mathfrak{g}}$ under the adjoint action of $G$ have been classified. We describe a simple algorithm for finding a representative of a nilpotent orbit. We use this to compute lists of representatives of these orbits for the Lie algebras of exceptional type. Then we give two applications. The first one concerns settling a conjecture by Elashvili on the index of centralizers of nilpotent orbits, for the case where the Lie algebra is of exceptional type. The second deals with minimal dimensions of centralizers in centralizers.'
author:
- |
Willem A. de Graaf\
Dipartimento di Matematica\
Università di Trento\
Italy
title: Computing with nilpotent orbits in simple Lie algebras of exceptional type
---
Introduction {#sec:intro}
============
Let $G$ be a simple algebraic group over an algebraically closed field of characteristic $0$. Let ${\mathfrak{g}}$ denote its Lie algebra. Then $G$ acts on ${\mathfrak{g}}$ via the adjoint representation. It is a natural question what the $G$-orbits in ${\mathfrak{g}}$ are. Recall that an element $e\in {\mathfrak{g}}$ is said to be nilpotent if the map ${\mathrm{\mathop{ad}}}e : {\mathfrak{g}}\to {\mathfrak{g}}$ is nilpotent. Now the $G$-orbits of nilpotent elements in ${\mathfrak{g}}$ are called nilpotent orbits. These have drawn a lot of attention in the past decades. On some occasions it turns out that using conceptual arguments to prove their properties is a lot harder for the exceptional types than it is for the classical types. However, for the former an approach based on a case by case analysis is possible. It is the objective of this paper to describe how this can be carried out using computer calculations.
The nilpotent orbits in ${\mathfrak{g}}$ are classified in terms of so-called weighted Dynkin diagrams. The first problem that we consider is to find a nilpotent element in ${\mathfrak{g}}$ given the corresponding weighted Dynkin diagram. We describe a straightforward algorithm for this (Section \[sec:orb\]). Then the algorithm is used to compute lists of explicit representatives of the nilpotent orbits in the Lie algebras of exceptional type. They are listed in Appendix \[sec:tables\].
We use these lists to prove Elashvili’s conjecture for the exceptional types by computer calculations. This conjecture concerns the index of centralizers of nilpotent elements. The concept of index is defined as follows. Let $K$ be a finite-dimensional Lie algebra, and let $K^*$ denote the dual space. For $f\in K^*$ set $K^f = \{ x\in K \mid f([x,y])=0 \text{ for all } y\in K\}$. Then the index of $K$ is defined as the number $${\mathrm{\mathop{ind}}}(K) = \inf_{f\in K^*} \dim K^f.$$ For semisimple Lie algebras in characteristic zero it is known that the index is equal to the rank ([@dix], Proposition 1.11.12).
By $C_{\mathfrak{g}}(x)$ we denote the centralizer of of $x\in {\mathfrak{g}}$.
Let ${\mathfrak{g}}$ be a semisimple Lie algebra over an algebraically closed field of characteristic $0$. Let $x\in {\mathfrak{g}}$. Then ${\mathrm{\mathop{ind}}}( C_{\mathfrak{g}}(x) )$ is equal to the rank of ${\mathfrak{g}}$.
This conjecture has recently received renewed attention, cf. [@panpreya], [@panyushev], [@vinya]. Its proof immediately reduces to the case where ${\mathfrak{g}}$ is simple, and $x$ nilpotent (cf. [@panyushev], §3). Also an inequality of Vinberg states that ${\mathrm{\mathop{ind}}}( C_{\mathfrak{g}}(x) )$ is at least the rank of ${\mathfrak{g}}$ (see [@panyushev], 1.6, 1.7). The conjecture has been proved for ${\mathfrak{g}}$ of classical type in [@yakimova], see also the discussion in [@panpreya]. In Section \[sec:index\] we report on computer calculations that settle the conjecture for the exceptional types.
In [@sekiguchi] the question is considered whether for a given nilpotent $e\in {\mathfrak{g}}$ there exists $x\in C_{\mathfrak{g}}(e)$ such that the dimension of $C_{\mathfrak{g}}(e,x)$ equals the rank of ${\mathfrak{g}}$. There an example is given where such an $x$ does not exist, for the case where ${\mathfrak{g}}$ is of type $F_4$. In Section \[sec:dblcen\] we approach this question using our lists of representatives of nilpotent orbits. This way we are able to give a complete list of all $e$ for which such an $x$ does not exist, in all exceptional types. For the Lie algebra of type $E_8$ this solves an open problem from [@sekiguchi]. For type $G_2$ this corrects a statement in [@sekiguchi].
The paper ends with two appendices. The first contains the lists of representatives of nilpotent orbits. The second (Appendix \[sec:appB\]) has lists of positive roots as they appear in the computer algebra system [GAP]{}. They have been added to help reading the tables of Appendix \[sec:tables\].
All algorithms described in this paper have been implemented in the language of the computer algebra system [GAP]{}4. The implementations are available from
http://www.science.unitn.it/~degraaf/nilpotent-orbit.html
[**Acknowledgments:**]{} I thank Alexander Elashvili for suggesting all the topics of this paper to me, and for his enthusiastic advice while I was writing it. Also I would like to thank Karin Baur for several helpful email exchanges, and for her comments on earlier versions.
Preliminaries on nilpotent orbits {#sec:prelim}
=================================
In this section we give a short overview of the theory behind the classification of nilpotent orbits. For more detailed accounts we refer to [@cart], [@colmcgov].
Let $e\in{\mathfrak{g}}$ be a nilpotent element. Then by the Jacobson-Morozov theorem $e$ lies in a subalgebra of ${\mathfrak{g}}$ that is isomorphic to ${\mathfrak{\mathop{sl}}}_2$. In other words, there are elements $f,h\in {\mathfrak{g}}$ with $[e,f]=h$, $[h,f]=-2f$, $[h,e]=2e$. In this case we say that $(f,h,e)$ is an ${\mathfrak{\mathop{sl}}}_2$-triple.
Now let $(f,h,e)$ be an ${\mathfrak{\mathop{sl}}}_2$-triple. Then by the representation theory of ${\mathfrak{\mathop{sl}}}_2$ we get a direct sum decomposition ${\mathfrak{g}}= \oplus_{k\in {\mathbb{Z}}} {\mathfrak{g}}(k)$, where ${\mathfrak{g}}(k) = \{ x\in {\mathfrak{g}}\mid [h,x]=kx \}$. Fix a Cartan subalgebra $H$ of ${\mathfrak{g}}$ with $h\in H$. Let $\Phi$ be the corresponding root system of ${\mathfrak{g}}$. For $\alpha\in \Phi$ we let $x_\alpha$ be a corresponding root vector. For each $\alpha$ there is a $k\in {\mathbb{Z}}$ with $x_\alpha\in {\mathfrak{g}}(k)$. We write $\eta(\alpha)=k$. It can be shown that there exists a basis of simple roots $\Delta\subset \Phi$ such that $\eta(\alpha)\geq 0$ for all $\alpha\in\Delta$. Furthermore, for such a $\Delta$ we have $\eta(\alpha)\in \{0,1,2\}$ for all $\alpha\in\Delta$. Write $\Delta = \{\alpha_1,
\ldots,\alpha_l\}$. Then the Dynkin diagram of $\Phi$ has $l$ nodes, the $i$-th node corresponding to $\alpha_i$. Now to each node we add the label $\eta(\alpha_i)$; the result is called the weighted Dynkin diagram. It is denoted $\Delta(e)$, and it depends only on $e$, and not on the choice of ${\mathfrak{\mathop{sl}}}_2$-triple containing $e$.
Let $e,e'$ be two nilpotent elements in ${\mathfrak{g}}$. It can be shown that $e,e'$ lie in the same $G$-orbit if and only if $\Delta(e)=\Delta(e')$. So the weighted Dynkin diagram of $e$ uniquely identifies the nilpotent orbit $Ge$. The weighted Dynkin diagrams corresponding to nilpotent orbits have been classified. For the exceptional types there are explicit lists. For the classical types there is a classification in terms of partitions. In particular, the nilpotent orbits in ${\mathfrak{g}}$ have been classified.
Let $e\in {\mathfrak{g}}$ be a representative of a nilpotent orbit. We may assume that $e$ is a linear combination of root vectors, corresponding to positive roots. Let $\beta_1,\ldots,\beta_r$ be the positive roots involved in this linear combination. Let $x_{\beta_i}$ (respectively $y_{\beta_i}$) be the root vector corresponding to $\beta_i$ (respectively $-\beta_i$). Let ${\mathfrak{l}}\subset {\mathfrak{g}}$ be the subalgebra generated by $H$ along with the $x_{\beta_i}$ and $y_{\beta_i}$. Then ${\mathfrak{l}}$ is reductive, and $e\in {\mathfrak{l}}$. Let $(f,h,e)$ be an ${\mathfrak{\mathop{sl}}}_2$-triple containing $e$, contained in ${\mathfrak{l}}$. Then ${\mathfrak{l}}$ decomposes with respect to the action of ${\mathrm{\mathop{ad}}}h$ as ${\mathfrak{l}} = \oplus_{k\in {\mathbb{Z}}} {\mathfrak{l}}(k)$. Let ${\mathfrak{p}}=\oplus_{k\geq 0} {\mathfrak{l}}(k)$, which is a subalgebra of ${\mathfrak{l}}$. Now it can be shown that the nilpotent orbit containing $e$ is uniquely determined by the pair $({\mathfrak{l}},{\mathfrak{p}})$ (cf. [@colmcgov], Chapter 8). Corresponding to this the nilpotent orbit has the label $X_n(a_i)$, where $X_n$ is the type of the semisimple part of ${\mathfrak{l}}$, and $i$ is the number of simple roots in the semisimple part of ${\mathfrak{p}}$. If the latter algebra is solvable, then we omit the $a_i$. Furthermore, if the roots of ${\mathfrak{l}}$ are short (seen as roots of ${\mathfrak{g}}$), then a tilde is put over the $X_n$. On some occasions, two different orbits can have the same label. Then a $'$ is added to one of them, whereas the other gets $''$. We note that, although the pair $({\mathfrak{l}},{\mathfrak{p}})$ uniquely determines the nilpotent orbit, it is also true that the same nilpotent orbit can have more than one (non-isomorphic) such pair. So the same nilpotent orbit can have more than one label.
The nilpotent element $e$ from above also has a Dynkin diagram, which is simply the Dynkin diagram of the roots $\beta_i$. This diagram has $r$ nodes, and node $i$ is connected to node $j$ by $\langle \beta_i, \beta_j^\vee\rangle \langle \beta_j,\beta_i^\vee\rangle =0,1,2,3$ lines. Furthermore, if these scalar products are positive, then the lines are dotted. This only occurs when ${\mathfrak{p}}$ is not solvable.
Finding representatives of nilpotent orbits {#sec:orb}
===========================================
In this section we consider the problem of finding a nilpotent element in ${\mathfrak{g}}$ corresponding to a given weighted Dynkin diagram $D$. We write $D_i$ for the label at node $i$. Let $H$ be a fixed Cartan subalgebra of ${\mathfrak{g}}$.
Let $e\in {\mathfrak{g}}$ be a nilpotent element such that $\Delta(e)=D$. Then there is an ${\mathfrak{\mathop{sl}}}_2$-triple $(f,h,e)$, containing $e$. Since we can conjugate any Cartan subalgebra of ${\mathfrak{g}}$ to $H$ by an element of $G$, we may assume that $h\in H$. As in the previous section we write $\Delta=\{\alpha_1,\ldots,\alpha_l\}$ for a basis of simple roots. By choosing a Chevalley basis in ${\mathfrak{g}}$ we get basis elements $h_1,\ldots,h_l$ of $H$, and root vectors $x_{\alpha_i}$ with $[h_j,x_{\alpha_i}]=\langle \alpha_i,\alpha_j^\vee\rangle x_{\alpha_i}$.
Each $h\in H$ yields a decomposition ${\mathfrak{g}}= \oplus_{k\in {\mathbb{Z}}} {\mathfrak{g}}(k)$, and a weighted Dynkin diagram, as described in the previous section. This weighted Dynkin diagram is equal to $D$ if and only if $[h,x_{\alpha_i}]= D_i x_{\alpha_i}$ for $1\leq i\leq l$. But this happens if and only if $\sum_{j=1}^l \langle
\alpha_i,\alpha_j^\vee\rangle a_j =D_i$, where the $a_j$ are such that $h=\sum_j a_j h_j$. Let $C=(\langle \alpha_i,\alpha_j^\vee \rangle)_{1\leq i,j\leq l}$ be the Cartan matrix of $\Phi$. It follows that $h$ yields the weighted Dynkin diagram $D$ if and only if $C(a_1,\ldots,a_l)^t = (D_1,\ldots,D_l)$. Hence that there is a unique such $h$, and we can compute it by solving a system of linear equations. However, not every weighted Dynkin diagram corresponds to a nilpotent orbit. In other words, not every weighted Dynkin diagram yields a $h$ that lies in an ${\mathfrak{\mathop{sl}}}_2$-triple. The next two lemmas lead to a probabilistic algorithm to decide whether this is the case or not.
Let $h\in H$. Then $h$ belongs to an ${\mathfrak{\mathop{sl}}}_2$-triple if and only if there is an $x\in {\mathfrak{g}}(2)$ such that $h\in [x,{\mathfrak{g}}(-2)]$.
The condition is clearly necessary. If $h\in [x,{\mathfrak{g}}(-2)]$ then there is a $y\in {\mathfrak{g}}(-2)$ with $[x,y]=h$. Then $(y,h,x)$ is an ${\mathfrak{\mathop{sl}}}_2$-triple.
\[lem2\] Let $h\in H$ be contained in an ${\mathfrak{\mathop{sl}}}_2$-triple $(y,h,x)$. Let $E$ be the set of $x'\in {\mathfrak{g}}(2)$ such that $h\in [x',{\mathfrak{g}}(-2)]$. Then $E$ is Zariski dense in ${\mathfrak{g}}(2)$.
(cf. [@cart], Proposition 5.6.2). Let $G_h = \{g\in G\mid {\mathrm{\mathop{Ad}}}(g)(h)=h\}$ be the stabilizer of $h$ in $G$. Then $G_h$ is an algebraic subgroup of $G$. Now ${\mathrm{\mathop{Lie}}}(G_h) = \{u\in {\mathfrak{g}}\mid {\mathrm{\mathop{ad}}}(u)(h)=0\}$. This is the centralizer of $h$ in ${\mathfrak{g}}$. Hence ${\mathrm{\mathop{Lie}}}(G_h) = {\mathfrak{g}}(0)$. For $u\in {\mathfrak{g}}(2)$ and $g\in G_h$ we have $[h,{\mathrm{\mathop{Ad}}}(g)(u)] = {\mathrm{\mathop{Ad}}}(g)[{\mathrm{\mathop{Ad}}}(g^{-1})(h),u] = {\mathrm{\mathop{Ad}}}(g)[h,u]=2{\mathrm{\mathop{Ad}}}(g)(u)$. Hence ${\mathrm{\mathop{Ad}}}(g)$ stabilizes ${\mathfrak{g}}(2)$. Let $\varphi : G_h \to {\mathfrak{g}}(2)$ be the morphism defined by $\varphi(g)={\mathrm{\mathop{Ad}}}(g)(x)$. Then the image of $\varphi$ is the $G_h$-orbit of $x$ in ${\mathfrak{g}}(2)$. The differential of $\varphi$ is $d\varphi : {\mathfrak{g}}(0)\to {\mathfrak{g}}(2)$, $d\varphi(u) = [u,x]$. But this is surjective because $[{\mathfrak{g}}(0),x]={\mathfrak{g}}(2)$ (this follows from the representation theory of ${\mathfrak{\mathop{sl}}}2_2$). So $\varphi$ is a dominant morphism. Hence $\varphi(G_h)$ is a dense subset of ${\mathfrak{g}}(2)$. Furthermore $\varphi(G_h)\subset E$.
Based on this we have a probabilistic algorithm for finding a representative of a nilpotent orbit, given a weighted Dynkin diagram. First we determine the unique $h\in H$ corresponding to the diagram. Then we select a random $x\in {\mathfrak{g}}(2)$, in the following way. Let $x_1,\ldots,x_s$ be a basis of ${\mathfrak{g}}(2)$. Let $\Omega$ be a finite subset of ${\mathbb{Q}}$ and select $\mu_1,\ldots,\mu_s$ randomly, uniformly and independently from $\Omega$. Then set $x= \sum_i \mu_i x_i$. By the previous lemma the probability that $h\in [x,{\mathfrak{g}}(-2)]$ is high (and can be made arbitrarily close to $1$ by enlarging $\Omega$). If it happens to be the case that $h\not\in [x,{\mathfrak{g}}(-2)]$ then we select another $x$ and continue. This algorithm will terminate in very few steps.
The $x$ found by the algorithm above will have “ugly” coefficients with respect to a Chevalley basis. We can obtain an element with “nice” coefficients in the following way. We write $x$ with respect to a Chevalley basis of ${\mathfrak{g}}$. We fix every coefficient but the first. For the first coefficient we try the values $0,1,2,\ldots$. The lemma ensures that we will quickly find an $x'$ which is a representative of the same nilpotent orbit, with the first coefficient a nice integer. We continue this way until all coefficients are nice integers.
The above results also provide a probabilistic algorithm for testing whether a given weighted Dynkin diagram corresponds to a nilpotent orbit. We basically try the same algorithm a few times, and if it does not come up with an $x$ then the weighted Dynkin diagram does not correspond to a nilpotent orbit with high probability. In principle we can make this absolutely sure by using Gröbner bases. This works as follows. Let $x_1,\ldots,x_s$ and $y_1,\ldots,y_s$ be bases of respectively ${\mathfrak{g}}(2)$ and ${\mathfrak{g}}(-2)$. Let $a_1,\ldots,a_s,b_1,\ldots,b_s$ be indeterminates. Let $u_1,\ldots,u_r$ be a basis of ${\mathfrak{g}}(0)$, and write $[x_i,y_j] = \sum_k \gamma_{ij}^k u_k$, and $h= \sum_k \alpha_k u_k$. Then there is an $x\in {\mathfrak{g}}(2)$ with $h\in [x,{\mathfrak{g}}(-2)]$ if and only if the system of polynomial equations $$\sum_{i=1}^s \sum_{j=1}^s \gamma_{ij}^k a_ib_j -\alpha_k = 0 \text{ ~~ for } 1\leq k\leq r$$ has a solution. Now this system has a solution over ${\mathbb{C}}$ if and only if the reduced Gröbner basis of the ideal generated by the left hand sides of these equations is not $\{1\}$.
In [@popov] Popov has given an algorithm for determining the strata of the nullcone of a linear representation of a reductive algebraic group. This also yields an algorithm for classifying nilpotent orbits in reductive Lie algebras, and for finding representatives of them.
Calculating the index {#sec:index}
=====================
In this section we describe a simple algorithm that for a Lie algebra gives an upper bound for its index. If the Lie algebra is defined over a sufficiently large field (e.g., of characteristic $0$), then the probability that this upper bound is equal to the index can be made arbitrarily high. We use the same notation as in Section \[sec:intro\].
Let $K$ be a finite-dimensional Lie algebra with basis $\{x_1,\ldots,x_n\}$. Let $c_{ij}^k$ be the structure constants of $K$, i.e., $[x_i,x_j] = \sum_{k=1}^n
c_{ij}^k x_k$. Let $\{\psi_1,\ldots,\psi_n\}$ be the dual basis of $K^*$, i.e., $\psi_i(x_j)=
\delta_{ij}$. Let $f=\sum_i T_i \psi_i$ be an element of the dual space $K^*$. Let $x = \sum_i \alpha_i x_i\in K$. Then $x\in K^f$ if and only if $f([x,x_j])=0$ for $1\leq j\leq n$. Now this is equivalent to $$\sum_{i=1}^n (\sum_{k=1}^n c_{ij}^k T_k)\alpha_i = 0 \text{ for } j=1,\ldots,n.$$ Define the $n\times n$-matrix $A$ by $A(i,j) = \sum_{k=1}^n c_{ij}^k T_k$. Then $\dim K^f = n -{\mathrm{\mathop{rank}}}(A)$. So the dimension of $K^f$ is minimal if and only if the rank of $A$ is maximal. Now the rank of $A$ is not maximal if and only if certain polynomial expressions in the $T_k$ (i.e., determinants of certain minors of $A$) vanish. Therefore, if the $T_k$ are chosen randomly and uniformly from a sufficiently large set, then with high probability the rank of $A$ will be maximal.
Here we consider the case where $K=C_{\mathfrak{g}}(e)$, where $e$ is a nilpotent element of the simple Lie algebra ${\mathfrak{g}}$. Then by Vinberg’s inequality we have that ${\mathrm{\mathop{ind}}}(K)$ is at least the rank of ${\mathfrak{g}}$. So if we find an $f$ such that $\dim (K^f) = {\mathrm{\mathop{rank}}}({\mathfrak{g}})$, then we have proved that ${\mathrm{\mathop{ind}}}(K)={\mathrm{\mathop{rank}}}({\mathfrak{g}})$. Moreover, the above discussion shows that we will quickly find such an $f$ by randomly choosing the $T_k$.
With the help of an implementation of this algorithm in [GAP]{}4, we have checked Elashvili’s conjecture for the exceptional types (which, except $G_2$, are the remaining open cases). As a result we can conclude that Elashvili’s conjecture holds for all simple Lie algebras.
Centralizers in centralizers {#sec:dblcen}
============================
Let $e\in{\mathfrak{g}}$ be a nilpotent element. Let $C_e= C_{\mathfrak{g}}(e)$ be the centralizer of $e$ in ${\mathfrak{g}}$. Let $x\in C_e$ and consider the centralizer $C_{e,x}$ of $x$ in $C_e$ (i.e., $C_{e,x}$ is the set of all elements of ${\mathfrak{g}}$ commuting with both $e$ and $x$). From [@richardson] it follows that $C_{e,x}$ contains a commutative subalgebra of dimension equal to ${\mathrm{\mathop{rank}}}({\mathfrak{g}})$. Hence the dimension of $C_{e,x}$ is at least the rank of ${\mathfrak{g}}$. In [@sekiguchi] the following question is considered: given $e$ does there exist $x\in C_e$ such that the dimension of $C_{e,x}$ equals the rank of ${\mathfrak{g}}$? The main result of that paper is a counter example to the question for the case where ${\mathfrak{g}}$ is of type $F_4$.
With the lists of representatives of the nilpotent orbits we can easily tackle this question in all simple Lie algebras of exceptional type. Let $e\in{\mathfrak{g}}$ be a nilpotent element, and let $x_1,\ldots,x_m$ be a basis of $C_e$. Set $x=T_1x_1+\cdots +T_mx_m$. Then the centralizer of $x$ in $C_e$ is equal to the kernel of ${\mathrm{\mathop{ad}}}x$ (restricted to $C_e$). So the dimension of $C_{e,x}$ is minimal if the rank of the matrix ${\mathrm{\mathop{ad}}}x$ is maximal. Now the entries of this matrix are linear polynomials in the $T_i$. It follows that for a random choice of the $T_i$, with very high probability, the rank of ${\mathrm{\mathop{ad}}}x$ is maximal. So this gives a probabilistic algorithm for determining the minimal dimension of $C_{e,x}$ (recall that we are varying $x$, and keeping $e$ fixed). Once the minimal dimension is found with this algorithm we can prove it rigorously as follows. Let $x$ be an element such that $\dim C_{e,x}$ is (hypothetically) minimal, as produced by the algorithm. If $\dim C_{e,x} = {\mathrm{\mathop{rank}}}({\mathfrak{g}})$ then we have proved that the minimal dimension of a $C_{e,x}$ is ${\mathrm{\mathop{rank}}}({\mathfrak{g}})$, as it cannot be smaller. Secondly, if the dimension that we find happens to be bigger, then we compute the rank of the matrix ${\mathrm{\mathop{ad}}}x$, where $x=T_1x_1+\cdots +T_mx_m$ and we let the $T_i$ be generators of a rational function field. The rank of that matrix will equal the maximal rank of any ${\mathrm{\mathop{ad}}}x$ for $x\in C_e$.
Using this algorithm we arrive at the following result.
Let ${\mathfrak{g}}$ be a simple Lie algebra of exceptional type, and $e\in{\mathfrak{g}}$ nilpotent. Then the minimal dimension of a $C_{e,x}$ is equal to ${\mathrm{\mathop{rank}}}({\mathfrak{g}})$, except in three cases, which are listed in the following table:
[|l|l|l|]{}
type of ${\mathfrak{g}}$ & label of $e$ & dimension of minimal $C_{e,x}$\
$G_2$ & $A_1+\widetilde{A}_1$ & $3$\
$F_4$ & $\widetilde{A_2}+A_2$ & $6$\
$E_8$ & $A_5+A_2+A_1$ & $12$\
In all three cases it turns out that a minimal $C_{e,x}$ is abelian. Furthermore, in each case it is possible to choose the element $x\in C_e$ such that it is homogeneous of degree $-1$ with respect to the grading of ${\mathfrak{g}}$ defined by the ${\mathfrak{\mathop{sl}}}_2$-triple containing $e$.
In relation to [@sekiguchi] we remark the following. In [@sekiguchi] it is wrongly stated that in $G_2$ all minimal $C_{e,x}$ have dimension equal to ${\mathrm{\mathop{rank}}}({\mathfrak{g}})$. The result for $F_4$ is the same as in [@sekiguchi]. Finally, the problem for $E_8$ is left open in [@sekiguchi].
Also, as a straightforward corollary of the proposition, it follows that in the exceptional types a minimal $C_{e,x}$ is always abelian.
Representatives of nilpotent orbits {#sec:tables}
===================================
In the tables below we list the nilpotent orbits in the Lie algebras of exceptional type. For each orbit we have given a label, the weighted Dynkin diagram, and the Dynkin diagram of a representative. We remark the following. If more than one label was possible, we have chosen the simplest one that we could find. This means that we have preferred a label of the form $X_n$ over a label of the form $X_n(a_i)$. Furthermore, we have preferred labels such that the Dynkin diagram of a corresponding representative has as few lines as possible. In the Dynkin diagram a black node means that the corresponding root is long. Finally, the labels corresponding to each node refer to the basis elements of the simple Lie algebras as present in [GAP]{}4. In Appendix \[sec:appB\] we list the positive roots of each root system of exceptional type, in the order in which they are used by [GAP]{}4. Now, if in the tables in this section a Dynkin diagram of a representative has labels $i_1,\ldots,i_k$, then the corresponding representative is the sum of the root vectors corresponding to the $i_j$-th positive root for $1\leq j\leq k$.
[|l|c|l|]{}
\
label & diagram & representative\
&
(10,7) (-8,0) (22,0) (-7,-3)[(1,0)[28]{}]{} (-5,0)[(1,0)[24]{}]{} (-7,3)[(1,0)[28]{}]{}
&\
$A_1$ & 1 0 &
(5,15) (3,0) (5,5)
\
$\widetilde{A}_1$ & 0 1 &
(5,15) (3,0) (5,5)
\
$A_1+\widetilde{A}_1$ & 2 0 &
(5,15) (3,0) (5,5) (23,0) (25,5)
\
$G_2$ & 2 2 &
(10,17) (0,0) (0,5) (30,0) (32,5) (1,-3)[(1,0)[28]{}]{} (3,0)[(1,0)[24]{}]{} (1,3)[(1,0)[28]{}]{}
\
[|l|c|l|]{}
\
label & diagram & representative\
&
(20,7) (-20,0) (0,0) (20,0) (40,0) (-17,0)[(1,0)[14]{}]{} (2,2)[(1,0)[16]{}]{} (2,-2)[(1,0)[16]{}]{} (23,0)[(1,0)[14]{}]{}
&\
$A_1$ & 1 0 0 0 &
(5,15) (3,0) (5,5)
\
$\widetilde{A}_1$ & 0 0 0 1 &
(5,15) (3,0) (5,5)
\
$A_1+\widetilde{A}_1$ & 0 1 0 0 &
(5,15) (3,0) (5,5) (23,0) (25,5)
\
$A_2$ & 2 0 0 0 &
(5,15) (3,0) (5,5) (23,0) (25,5) (5,0)[(1,0)[17]{}]{}
\
$\widetilde{A}_2$ & 0 0 0 2 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{}
\
$\widetilde{A}_1+A_2$ & 0 0 1 0 &
(5,15) (0,0) (0,5) (18,0) (20,5) (38,0) (40,5) (21,0)[(1,0)[14]{}]{}
\
$B_2$ & 2 0 0 1 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{}
\
$\widetilde{A}_2+A_1$ & 0 1 0 1 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (40,0) (43,5)
\
$B_2+A_1$ & 1 0 1 0 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{} (40,0) (43,5)
\
$\widetilde{A}_2+A_2$ & 0 2 0 0 &
(35,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (40,0) (43,5) (63,0) (65,5) (43,0)[(1,0)[17]{}]{}
\
$B_3$ & 2 2 0 0 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{} (40,0) (43,5) (26,0)[(1,0)[11]{}]{}
\
$C_3$ & 1 0 1 2 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{} (40,0) (43,5) (26,0)[(1,0)[11]{}]{}
\
$C_3+A_1$ & 0 2 0 2 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{} (40,0) (43,5) (26,0)[(1,0)[11]{}]{} (63,0) (65,5)
\
$B_4$ & 2 2 0 2 &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,-1)[(1,0)[14]{}]{} (6,1)[(1,0)[14]{}]{} (40,0) (43,5) (26,0)[(1,0)[11]{}]{} (57,0) (60,5) (43,0)[(1,0)[11]{}]{}
\
$F_4$ & 2 2 2 2 &
(20,15) (0,0) (3,5) (20,0) (23,5) (40,0) (43,5) (60,0) (63,5) (3,0)[(1,0)[14]{}]{} (22,2)[(1,0)[16]{}]{} (22,-2)[(1,0)[16]{}]{} (43,0)[(1,0)[14]{}]{}
\
[|l|c|l|]{}
\
label & diagram & representative\
&
(40,25) (-20,0) (0,0) (20,0) (40,0) (60,0) (20,20) (-17,0)[(1,0)[14]{}]{} (3,0)[(1,0)[14]{}]{} (23,0)[(1,0)[14]{}]{} (43,0)[(1,0)[14]{}]{} (20,3)[(0,1)[14]{}]{}
&\
$A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0$ &
(5,15) (3,0) (5,5)
\
$2A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5)
\
$3A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5)
\
$A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{}
\
$A_2+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5)
\
$2A_2$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5)
\
$2A_1+A_2$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5)
\
$A_3$ & $1~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5)
\
$A_1+2A_2$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1$ &
(100,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (63,0) (66,0)[(1,0)[14]{}]{} (65,5) (83,0) (85,5)
\
$A_3+A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5)
\
$A_3+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5) (83,0) (85,5)
\
$A_4$ & $2~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5)
\
$D_4$ & $0~~~~0~~~~\overset{\text{\normalsize 2}}{2}~~~~0~~~~0$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30)
\
$A_4+A_1$ & $1~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (83,0) (85,5)
\
$A_5$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{}
\
$D_5(a_1)$ & $1~~~~1~~~~\overset{\text{\normalsize 2}}{0}~~~~1~~~~1$ &
(25,40) (33,0) (22,-2) (53,0) (56,-2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,23) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25)
\
$A_5+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_5$ & $2~~~~0~~~~\overset{\text{\normalsize 2}}{2}~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21)
\
$E_6(a_1)$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{0}~~~~2~~~~2$ &
(25,40) (33,0) (22,-2) (53,0) (56,-2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,23) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (73,20) (56,20)[(1,0)[14]{}]{} (75,23)
\
$E_6$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{2}~~~~2~~~~2$ &
(40,30) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{}
\
[|l|c|l|]{}
\
label & diagram & representative\
&
(100,25) (0,0) (20,0) (40,0) (60,0) (80,0) (100,0) (40,20) (3,0)[(1,0)[14]{}]{} (23,0)[(1,0)[14]{}]{} (43,0)[(1,0)[14]{}]{} (63,0)[(1,0)[14]{}]{} (83,0)[(1,0)[14]{}]{} (40,3)[(0,1)[14]{}]{}
&\
$A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5)
\
$2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5)
\
$(3A_1)''$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5)
\
$(3A_1)'$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5)
\
$A_2$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{}
\
$4A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5) (63,0) (65,5)
\
$A_2+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5)
\
$A_2+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5)
\
$A_3$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5)
\
$2A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5)
\
$A_2+3A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~0~~~~0$ &
(100,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5) (83,0) (85,5)
\
$(A_3+A_1)''$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5)
\
$2A_2+A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$(A_3+A_1)'$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5)
\
$(A_3+2A_1)'$ & $0~~~~2~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$(A_3+2A_1)''$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$D_4$ & $2~~~~2~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30)
\
$A_3+3A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (103,0) (105,5)
\
$A_3+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{}
\
$A_4$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{}
\
$A_3+A_2+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~2~~~~0~~~~0$ &
(110,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$(A_5)''$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{}
\
$D_4+A_1$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11)
\
$A_4+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5)
\
$D_4+2A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11) (83,6) (85,11)
\
$A_4+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$(A_5)'$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{}
\
$(A_5+A_1)''$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_5(a_1)+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~2~~~~0~~~~0$ &
(25,40) (33,0) (20,-2) (53,0) (56,-2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,24) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (80,0) (82,5)
\
$D_6(a_2)$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0~~~~2$ &
(25,40) (33,0) (20,-2) (53,0) (56,2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,25) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (82,0) (56,0)[(1,0)[23]{}]{} (85,3)
\
$(A_5+A_1)'$ & $0~~~~2~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_5$ & $2~~~~2~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21)
\
$A_5+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (103,0) (105,5) (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$A_6$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$D_5+A_1$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21) (83,0) (85,5)
\
$D_6(a_1)$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25)
\
$D_6(a_1)+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (99,0) (102,5)
\
$D_6$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~2~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (63,20) (63,3)[(0,1)[14]{}]{} (65,21)
\
$A_7$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~0$ &
(135,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$E_6$ & $2~~~~2~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~0$ &
(40,30) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{}
\
$D_6+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (63,20) (63,3)[(0,1)[14]{}]{} (65,21) (103,0) (105,5)
\
$E_7(a_2)$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{0}~~~~2~~~~0~~~~2$ &
(25,40) (33,0) (20,-2) (53,0) (56,2) (33,3)(33,17) (36,0)[(1,0)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,25) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (82,0) (56,0)[(1,0)[23]{}]{} (85,3) (82,20) (56,20)[(1,0)[23]{}]{} (85,23)
\
$E_7(a_1)$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{0}~~~~2~~~~2~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (93,20) (76,20)[(1,0)[14]{}]{} (95,25)
\
$E_7$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{2}~~~~2~~~~2~~~~2$ &
(100,40) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (86,0)[(1,0)[14]{}]{} (105,5)
\
[|l|c|l|]{}
\
label & diagram & representative\
&
(120,25) (0,0) (20,0) (40,0) (60,0) (80,0) (100,0) (120,0) (40,20) (3,0)[(1,0)[14]{}]{} (23,0)[(1,0)[14]{}]{} (43,0)[(1,0)[14]{}]{} (63,0)[(1,0)[14]{}]{} (83,0)[(1,0)[14]{}]{} (40,3)[(0,1)[14]{}]{} (103,0)[(1,0)[14]{}]{}
&\
$A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5)
\
$2A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5)
\
$3A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5)
\
$A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{}
\
$4A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~0~~~~0$ &
(50,15) (3,0) (5,5) (23,0) (25,5) (43,0) (45,5) (63,0) (65,5)
\
$A_2+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5)
\
$A_2+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5)
\
$A_3$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5)
\
$A_2+3A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (65,5) (83,0) (85,5)
\
$2A_2$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5)
\
$2A_2+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~1~~~~0$ &
(100,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (45,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$A_3+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5)
\
$(A_3+2A_1)'$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$D_4$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2~~~~2$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30)
\
$2A_2+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0~~~~0~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (46,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (103,0) (105,5)
\
$(A_3+2A_1)''$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5)
\
$A_3+3A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (103,0) (105,5)
\
$A_3+A_2$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{}
\
$A_4$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~0~~~~2$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{}
\
$A_3+A_2+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_4+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1~~~~2$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11)
\
$A_3+A_2+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (123,0) (125,5)
\
$A_4+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5)
\
$2A_3$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$D_4+2A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~2$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11) (83,6) (85,11)
\
$A_4+2A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (103,0) (105,5)
\
$A_4+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0~~~~0$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$A_5$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~1$ &
(5,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{}
\
$D_5(a_1)+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0~~~~2$ &
(25,40) (33,0) (20,-2) (53,0) (56,-2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,24) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (80,0) (82,5)
\
$A_4+A_2+A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~0$ &
(130,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5)
\
$D_4+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 2}}{0}~~~~0~~~~0~~~~0~~~~2$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11) (83,6) (85,11) (66,6)[(1,0)[14]{}]{}
\
$(A_5+A_1)''$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_5$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~0~~~~2~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21)
\
$A_4+A_3$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~1~~~~0$ &
(130,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$(A_5+A_1)'$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~0~~~~0~~~~1$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5)
\
$D_5(a_1)+A_2$ & $0~~~~1~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~1~~~~0~~~~1$ &
(25,40) (33,0) (20,-2) (53,0) (56,-2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,24) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (80,0) (82,5) (100,0) (102,5) (83,0)[(1,0)[14]{}]{}
\
$D_4+A_3$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1~~~~0$ &
(45,43) (3,6) (1,11) (23,6) (23,-4) (6,6)[(1,0)[14]{}]{} (26,6)[(1,0)[14]{}]{} (43,6) (45,11) (23,26) (23,9)[(0,1)[14]{}]{} (25,30) (63,6) (65,11) (83,6) (85,11) (66,6)[(1,0)[14]{}]{} (103,6) (86,6)[(1,0)[14]{}]{} (105,11)
\
$A_5+2A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0~~~~1~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (123,0) (125,5)
\
$A_5+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$D_5+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~1~~~~0~~~~1~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21) (83,0) (85,5)
\
$A_5+A_2+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~2~~~~0~~~~0~~~~0$ &
(150,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (123,0) (125,5) (106,0)[(1,0)[14]{}]{} (143,0) (145,5)
\
$A_6$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$D_5+2A_1$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21) (83,0) (85,5) (103,0) (105,5)
\
$A_6+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5)
\
$D_6(a_1)+A_1$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (99,0) (102,5)
\
$(A_7)''$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~0~~~~2$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$D_5+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~2~~~~0~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21) (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{}
\
$D_6$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~0~~~~0~~~~1~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (63,20) (63,3)[(0,1)[14]{}]{} (65,21)
\
$E_6$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~0~~~~2~~~~2~~~~2$ &
(40,30) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{}
\
$D_5+A_3$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~1$ &
(50,35) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (43,0) (45,5) (63,0) (46,0)[(1,0)[14]{}]{} (65,5) (43,20) (43,3)[(0,1)[14]{}]{} (45,21) (83,0) (85,5) (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (106,0)[(1,0)[14]{}]{} (125,5)
\
$(A_7)'$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~1~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$A_7+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~2$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{} (143,0) (145,5)
\
$D_6+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (63,20) (63,3)[(0,1)[14]{}]{} (65,21) (103,0) (105,5)
\
$D_8(a_3)$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~0~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,5) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (93,0) (95,5) (113,0) (115,5) (76,0)[(1,0)[14]{}]{} (96,0)[(1,0)[14]{}]{}
\
$D_6+2A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{0}~~~~2~~~~0~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (63,20) (63,3)[(0,1)[14]{}]{} (65,21) (103,0) (105,5) (123,0) (125,5)
\
$E_6+A_1$ & $1~~~~0~~~~\overset{\text{\normalsize 0}}{1}~~~~0~~~~1~~~~2~~~~2$ &
(40,30) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (105,5)
\
$E_7(a_2)$ & $0~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0~~~~2~~~~2$ &
(25,40) (33,0) (20,-2) (53,0) (56,2) (36,0)(50,0) (33,3)[(0,1)[14]{}]{} (53,3)[(0,1)[14]{}]{} (36,20)[(1,0)[14]{}]{} (33,20) (24,25) (53,20) (55,23) (4,20) (7,20)[(1,0)[23]{}]{} (3,25) (82,0) (56,0)[(1,0)[23]{}]{} (85,3) (82,20) (56,20)[(1,0)[23]{}]{} (85,23)
\
$A_8$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2~~~~0$ &
(120,15) (3,0) (5,5) (23,0) (25,5) (6,0)[(1,0)[14]{}]{} (43,0) (26,0)[(1,0)[14]{}]{} (45,5) (63,0) (65,5) (46,0)[(1,0)[14]{}]{} (83,0) (85,5) (66,0)[(1,0)[14]{}]{} (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{} (143,0) (145,5) (126,0)[(1,0)[14]{}]{}
\
$D_7$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~1~~~~0~~~~1$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (83,20) (83,3)[(0,1)[14]{}]{} (67,21) (103,0) (105,0) (86,0)[(1,0)[14]{}]{}
\
$E_6+A_2$ & $0~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2~~~~2$ &
(40,30) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (105,5) (123,0) (106,0)[(1,0)[14]{}]{} (125,5)
\
$E_7(a_1)$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~0~~~~2~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (93,20) (76,20)[(1,0)[14]{}]{} (95,25)
\
$D_8(a_1)$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2~~~~0$ &
(100,40) (13,20) (13,25) (16,20)[(1,0)[14]{}]{} (93,0) (80,-2) (113,0) (115,-2) (96,0)(110,0) (93,3)[(0,1)[14]{}]{} (113,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (55,24) (73,20) (75,23) (33,20) (36,20)[(1,0)[14]{}]{} (35,25) (93,20) (113,20) (76,20)[(1,0)[14]{}]{} (96,20)[(1,0)[14]{}]{} (95,25) (115,25)
\
$E_7(a_1)+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~0~~~~2~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (53,0) (40,-2) (73,0) (76,-2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,24) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (93,20) (76,20)[(1,0)[14]{}]{} (95,25) (113,0) (115,0)
\
$E_7$ & $2~~~~1~~~~\overset{\text{\normalsize 1}}{0}~~~~1~~~~2~~~~2~~~~2$ &
(100,40) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (86,0)[(1,0)[14]{}]{} (105,5)
\
$D_8$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~0~~~~2$ &
(50,35) (3,0) (5,5) (6,0)[(1,0)[14]{}]{} (23,0) (25,5) (43,0) (45,5) (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (63,0) (65,5) (83,0) (66,0)[(1,0)[14]{}]{} (85,5) (103,20) (103,3)[(0,1)[14]{}]{} (105,21) (103,0) (105,5) (86,0)[(1,0)[14]{}]{} (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
$E_7+A_1$ & $2~~~~0~~~~\overset{\text{\normalsize 0}}{2}~~~~0~~~~2~~~~2~~~~2$ &
(100,40) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (86,0)[(1,0)[14]{}]{} (105,5) (123,0) (125,5)
\
$E_8(a_2)$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{0}~~~~2~~~~0~~~~2~~~~2$ &
(25,40) (4,20) (7,20)[(1,0)[14]{}]{} (6,25) (53,0) (40,-2) (73,0) (76,2) (56,0)(70,0) (53,3)[(0,1)[14]{}]{} (73,3)[(0,1)[14]{}]{} (56,20)[(1,0)[14]{}]{} (53,20) (44,25) (73,20) (75,23) (24,20) (27,20)[(1,0)[23]{}]{} (23,25) (102,0) (76,0)[(1,0)[23]{}]{} (105,3) (102,20) (76,20)[(1,0)[23]{}]{} (105,23)
\
$E_8(a_1)$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{0}~~~~2~~~~2~~~~2~~~~2$ &
(100,40) (4,20) (4,25) (7,20)[(1,0)[14]{}]{} (24,20) (24,25) (27,20)[(1,0)[14]{}]{} (73,0) (60,-2) (93,0) (96,-2) (76,0)(90,0) (73,3)[(0,1)[14]{}]{} (93,3)[(0,1)[14]{}]{} (76,20)[(1,0)[14]{}]{} (73,20) (64,24) (93,20) (95,23) (44,20) (47,20)[(1,0)[23]{}]{} (43,25) (113,20) (96,20)[(1,0)[14]{}]{} (115,25)
\
$E_8$ & $2~~~~2~~~~\overset{\text{\normalsize 2}}{2}~~~~2~~~~2~~~~2~~~~2$ &
(100,40) (3,0) (5,4) (23,0) (25,4) (43,0) (45,4) (63,0) (65,4) (83,0) (85,4) (43,20) (47,21) (6,0)[(1,0)[14]{}]{} (26,0)[(1,0)[14]{}]{} (46,0)[(1,0)[14]{}]{} (66,0)[(1,0)[14]{}]{} (43,3)[(0,1)[14]{}]{} (103,0) (86,0)[(1,0)[14]{}]{} (105,5) (123,0) (125,5) (106,0)[(1,0)[14]{}]{}
\
The exceptional root systems in [GAP]{} {#sec:appB}
=======================================
In this appendix we list the positive roots of the exceptional root systems, in the order in which they appear in [GAP]{}4. The tables have to be read from left to right, and from top to bottom. So the first root is the one top left, the second root is the second one on the first line, and the last root is the one bottom right. For each root its coefficients with respect to a basis of simple roots are given.
The roots for $F_4$ in Table \[tab:F4rt\] may seem slightly strange. This is due to the fact that in [GAP]{}4 the positive roots are ordered differently than usual. In this table the coefficients of each root with respect to the “usual” ordering of a basis of simple roots is given (i.e., as in [@bou4]). However, the roots are listed in the same order as they are in [GAP]{}4.
[|c|c|c|c|c|c|]{}
\
10 & 01 & 11 & 21 & 31 & 32\
[|c|c|c|c|c|c|c|c|]{}
\
0001 & 1000 & 0010 & 0100 & 0011 & 1100 & 0110 & 0111\
1110 & 0120 & 1111 & 0121 & 1120 & 1121 & 0122 & 1220\
1221 & 1122 & 1231 & 1222 & 1232 & 1242 & 1342 & 2342\
[|c|c|c|c|c|c|c|c|c|]{}
\
$10\overset{\text{\normalsize 0}}{0}00$ & $00\overset{\text{\normalsize 1}}{0}00$ & $01\overset{\text{\normalsize 0}}{0}00$ & $00\overset{\text{\normalsize 0}}{1}00$ & $00\overset{\text{\normalsize 0}}{0}10$ & $00\overset{\text{\normalsize 0}}{0}01$ & $11\overset{\text{\normalsize 0}}{0}00$ & $00\overset{\text{\normalsize 1}}{1}00$ & $01\overset{\text{\normalsize 0}}{1}00$\
$00\overset{\text{\normalsize 0}}{1}10$ & $00\overset{\text{\normalsize 0}}{0}11$ & $11\overset{\text{\normalsize 0}}{1}00$ & $01\overset{\text{\normalsize 1}}{1}00$ & $00\overset{\text{\normalsize 1}}{1}10$ & $01\overset{\text{\normalsize 0}}{1}10$ & $00\overset{\text{\normalsize 0}}{1}11$ & $11\overset{\text{\normalsize 1}}{1}00$ & $11\overset{\text{\normalsize 0}}{1}10$\
$01\overset{\text{\normalsize 1}}{1}10$ & $00\overset{\text{\normalsize 1}}{1}11$ & $01\overset{\text{\normalsize 0}}{1}11$ & $11\overset{\text{\normalsize 1}}{1}10$ & $11\overset{\text{\normalsize 0}}{1}11$ & $01\overset{\text{\normalsize 1}}{2}10$ & $01\overset{\text{\normalsize 1}}{1}11$ & $11\overset{\text{\normalsize 1}}{2}10$ & $11\overset{\text{\normalsize 1}}{1}11$\
$01\overset{\text{\normalsize 1}}{2}11$ & $12\overset{\text{\normalsize 1}}{2}10$ & $11\overset{\text{\normalsize 1}}{2}11$ & $01\overset{\text{\normalsize 1}}{2}21$ & $12\overset{\text{\normalsize 1}}{2}11$ & $11\overset{\text{\normalsize 1}}{2}21$ & $12\overset{\text{\normalsize 1}}{2}21$ & $12\overset{\text{\normalsize 1}}{3}21$ & $12\overset{\text{\normalsize 2}}{3}21$\
[|c|c|c|c|c|c|c|]{}
\
$10\overset{\text{\normalsize 0}}{0}000$ & $00\overset{\text{\normalsize 1}}{0}000$ & $01\overset{\text{\normalsize 0}}{0}000$ & $00\overset{\text{\normalsize 0}}{1}000$ & $00\overset{\text{\normalsize 0}}{0}100$ & $00\overset{\text{\normalsize 0}}{0}010$ & $00\overset{\text{\normalsize 0}}{0}001$\
$11\overset{\text{\normalsize 0}}{0}000$ & $00\overset{\text{\normalsize 1}}{1}000$ & $01\overset{\text{\normalsize 0}}{1}000$ & $00\overset{\text{\normalsize 0}}{1}100$ & $00\overset{\text{\normalsize 0}}{0}110$ & $00\overset{\text{\normalsize 0}}{0}011$ & $11\overset{\text{\normalsize 0}}{1}000$\
$01\overset{\text{\normalsize 1}}{1}000$ & $00\overset{\text{\normalsize 1}}{1}100$ & $01\overset{\text{\normalsize 0}}{1}100$ & $00\overset{\text{\normalsize 0}}{1}110$ & $00\overset{\text{\normalsize 0}}{0}111$ & $11\overset{\text{\normalsize 1}}{1}000$ & $11\overset{\text{\normalsize 0}}{1}100$\
$01\overset{\text{\normalsize 1}}{1}100$ & $00\overset{\text{\normalsize 1}}{1}110$ & $01\overset{\text{\normalsize 0}}{1}110$ & $00\overset{\text{\normalsize 0}}{1}111$ & $11\overset{\text{\normalsize 1}}{1}100$ & $11\overset{\text{\normalsize 0}}{1}110$ & $01\overset{\text{\normalsize 1}}{2}100$\
$01\overset{\text{\normalsize 1}}{1}110$ & $00\overset{\text{\normalsize 1}}{1}111$ & $01\overset{\text{\normalsize 0}}{1}111$ & $11\overset{\text{\normalsize 1}}{2}100$ & $11\overset{\text{\normalsize 1}}{1}110$ & $11\overset{\text{\normalsize 0}}{1}111$ & $01\overset{\text{\normalsize 1}}{2}110$\
$01\overset{\text{\normalsize 1}}{1}111$ & $12\overset{\text{\normalsize 1}}{2}100$ & $11\overset{\text{\normalsize 1}}{2}110$ & $11\overset{\text{\normalsize 1}}{1}111$ & $01\overset{\text{\normalsize 1}}{2}210$ & $01\overset{\text{\normalsize 1}}{2}111$ & $12\overset{\text{\normalsize 1}}{2}110$\
$11\overset{\text{\normalsize 1}}{2}210$ & $11\overset{\text{\normalsize 1}}{2}111$ & $01\overset{\text{\normalsize 1}}{2}211$ & $12\overset{\text{\normalsize 1}}{2}210$ & $12\overset{\text{\normalsize 1}}{2}111$ & $11\overset{\text{\normalsize 1}}{2}211$ & $01\overset{\text{\normalsize 1}}{2}221$\
$12\overset{\text{\normalsize 1}}{3}210$ & $12\overset{\text{\normalsize 1}}{2}211$ & $11\overset{\text{\normalsize 1}}{2}221$ & $12\overset{\text{\normalsize 2}}{3}210$ & $12\overset{\text{\normalsize 1}}{3}211$ & $12\overset{\text{\normalsize 1}}{2}221$ & $12\overset{\text{\normalsize 2}}{3}211$\
$12\overset{\text{\normalsize 1}}{3}221$ & $12\overset{\text{\normalsize 2}}{3}221$ & $12\overset{\text{\normalsize 1}}{3}321$ & $12\overset{\text{\normalsize 2}}{3}321$ & $12\overset{\text{\normalsize 2}}{4}321$ & $13\overset{\text{\normalsize 2}}{4}321$ & $23\overset{\text{\normalsize 2}}{4}321$\
[|c|c|c|c|c|c|]{}
\
$10\overset{\text{\normalsize 0}}{0}0000$ & $00\overset{\text{\normalsize 1}}{0}0000$ & $01\overset{\text{\normalsize 0}}{0}0000$ & $00\overset{\text{\normalsize 0}}{1}0000$ & $00\overset{\text{\normalsize 0}}{0}1000$ & $00\overset{\text{\normalsize 0}}{0}0100$\
$00\overset{\text{\normalsize 0}}{0}0010$ & $00\overset{\text{\normalsize 0}}{0}0001$ & $11\overset{\text{\normalsize 0}}{0}0000$ & $00\overset{\text{\normalsize 1}}{1}0000$ & $01\overset{\text{\normalsize 0}}{1}0000$ & $00\overset{\text{\normalsize 0}}{1}1000$\
$00\overset{\text{\normalsize 0}}{0}1100$ & $00\overset{\text{\normalsize 0}}{0}0110$ & $00\overset{\text{\normalsize 0}}{0}0011$ & $11\overset{\text{\normalsize 0}}{1}0000$ & $01\overset{\text{\normalsize 1}}{1}0000$ & $00\overset{\text{\normalsize 1}}{1}1000$\
$01\overset{\text{\normalsize 0}}{1}1000$ & $00\overset{\text{\normalsize 0}}{1}1100$ & $00\overset{\text{\normalsize 0}}{0}1110$ & $00\overset{\text{\normalsize 0}}{0}0111$ & $11\overset{\text{\normalsize 1}}{1}0000$ & $11\overset{\text{\normalsize 0}}{1}1000$\
$01\overset{\text{\normalsize 1}}{1}1000$ & $00\overset{\text{\normalsize 1}}{1}1100$ & $01\overset{\text{\normalsize 0}}{1}1100$ & $00\overset{\text{\normalsize 0}}{1}1110$ & $00\overset{\text{\normalsize 0}}{0}1111$ & $11\overset{\text{\normalsize 1}}{1}1000$\
$11\overset{\text{\normalsize 0}}{1}1100$ & $01\overset{\text{\normalsize 1}}{2}1000$ & $01\overset{\text{\normalsize 1}}{1}1100$ & $00\overset{\text{\normalsize 1}}{1}1110$ & $01\overset{\text{\normalsize 0}}{1}1110$ & $00\overset{\text{\normalsize 0}}{1}1111$\
$11\overset{\text{\normalsize 1}}{2}1000$ & $11\overset{\text{\normalsize 1}}{1}1100$ & $11\overset{\text{\normalsize 0}}{1}1110$ & $01\overset{\text{\normalsize 1}}{2}1100$ & $01\overset{\text{\normalsize 1}}{1}1110$ & $00\overset{\text{\normalsize 1}}{1}1111$\
$01\overset{\text{\normalsize 0}}{1}1111$ & $12\overset{\text{\normalsize 1}}{2}1000$ & $11\overset{\text{\normalsize 1}}{2}1100$ & $11\overset{\text{\normalsize 1}}{1}1110$ & $11\overset{\text{\normalsize 0}}{1}1111$ & $01\overset{\text{\normalsize 1}}{2}2100$\
$01\overset{\text{\normalsize 1}}{2}1110$ & $01\overset{\text{\normalsize 1}}{1}1111$ & $12\overset{\text{\normalsize 1}}{2}1100$ & $11\overset{\text{\normalsize 1}}{2}2100$ & $11\overset{\text{\normalsize 1}}{2}1110$ & $11\overset{\text{\normalsize 1}}{1}1111$\
$01\overset{\text{\normalsize 1}}{2}2110$ & $01\overset{\text{\normalsize 1}}{2}1111$ & $12\overset{\text{\normalsize 1}}{2}2100$ & $12\overset{\text{\normalsize 1}}{2}1110$ & $11\overset{\text{\normalsize 1}}{2}2110$ & $11\overset{\text{\normalsize 1}}{2}1111$\
$01\overset{\text{\normalsize 1}}{2}2210$ & $01\overset{\text{\normalsize 1}}{2}2111$ & $12\overset{\text{\normalsize 1}}{3}2100$ & $12\overset{\text{\normalsize 1}}{2}2110$ & $12\overset{\text{\normalsize 1}}{2}1111$ & $11\overset{\text{\normalsize 1}}{2}2210$\
$11\overset{\text{\normalsize 1}}{2}2111$ & $01\overset{\text{\normalsize 1}}{2}2211$ & $12\overset{\text{\normalsize 2}}{3}2100$ & $12\overset{\text{\normalsize 1}}{3}2110$ & $12\overset{\text{\normalsize 1}}{2}2210$ & $12\overset{\text{\normalsize 1}}{2}2111$\
$11\overset{\text{\normalsize 1}}{2}2211$ & $01\overset{\text{\normalsize 1}}{2}2221$ & $12\overset{\text{\normalsize 2}}{3}2110$ & $12\overset{\text{\normalsize 1}}{3}2210$ & $12\overset{\text{\normalsize 1}}{3}2111$ & $12\overset{\text{\normalsize 1}}{2}2211$\
$11\overset{\text{\normalsize 1}}{2}2221$ & $12\overset{\text{\normalsize 2}}{3}2210$ & $12\overset{\text{\normalsize 2}}{3}2111$ & $12\overset{\text{\normalsize 1}}{3}3210$ & $12\overset{\text{\normalsize 1}}{3}2211$ & $12\overset{\text{\normalsize 1}}{2}2221$\
$12\overset{\text{\normalsize 2}}{3}3210$ & $12\overset{\text{\normalsize 2}}{3}2211$ & $12\overset{\text{\normalsize 1}}{3}3211$ & $12\overset{\text{\normalsize 1}}{3}2221$ & $12\overset{\text{\normalsize 2}}{4}3210$ & $12\overset{\text{\normalsize 2}}{3}3211$\
$12\overset{\text{\normalsize 2}}{3}2221$ & $12\overset{\text{\normalsize 1}}{3}3221$ & $13\overset{\text{\normalsize 2}}{4}3210$ & $12\overset{\text{\normalsize 2}}{4}3211$ & $12\overset{\text{\normalsize 2}}{3}3221$ & $12\overset{\text{\normalsize 1}}{3}3321$\
$23\overset{\text{\normalsize 2}}{4}3210$ & $13\overset{\text{\normalsize 2}}{4}3211$ & $12\overset{\text{\normalsize 2}}{4}3221$ & $12\overset{\text{\normalsize 2}}{3}3321$ & $23\overset{\text{\normalsize 2}}{4}3211$ & $13\overset{\text{\normalsize 2}}{4}3221$\
$12\overset{\text{\normalsize 2}}{4}3321$ & $23\overset{\text{\normalsize 2}}{4}3221$ & $13\overset{\text{\normalsize 2}}{4}3321$ & $12\overset{\text{\normalsize 2}}{4}4321$ & $23\overset{\text{\normalsize 2}}{4}3321$ & $13\overset{\text{\normalsize 2}}{4}4321$\
$23\overset{\text{\normalsize 2}}{4}4321$ & $13\overset{\text{\normalsize 2}}{5}4321$ & $23\overset{\text{\normalsize 2}}{5}4321$ & $13\overset{\text{\normalsize 3}}{5}4321$ & $23\overset{\text{\normalsize 3}}{5}4321$ & $24\overset{\text{\normalsize 2}}{5}4321$\
$24\overset{\text{\normalsize 3}}{5}4321$ & $24\overset{\text{\normalsize 3}}{6}4321$ & $24\overset{\text{\normalsize 3}}{6}5321$ & $24\overset{\text{\normalsize 3}}{6}5421$ & $24\overset{\text{\normalsize 3}}{6}5431$ & $24\overset{\text{\normalsize 3}}{6}5432$\
[10]{}
N. Bourbaki. . Hermann, Paris, 1968.
R. W. Carter. . John Wiley & Sons, Chichester, 1985.
David H. Collingwood and William M. McGovern. . Van Nostrand Reinhold Mathematics Series. Van Nostrand Reinhold Co., New York, 1993.
J. Dixmier. . Gauthier-Villars, Paris, Bruxelles, Montréal, 1974.
D. Panyushev, A. Premet, and O. Yakimova. On symmetric invariants of centralisers in reductive [L]{}ie algebras. , 313(1):343–391, 2007.
Dmitri I. Panyushev. The index of a [L]{}ie algebra, the centralizer of a nilpotent element, and the normalizer of the centralizer. , 134(1):41–59, 2003.
V. L. Popov. The cone of [H]{}ilbert null forms. , 241(Teor. Chisel, Algebra i Algebr. Geom.):192–209, 2003.
R. W. Richardson. Commuting varieties of semisimple [L]{}ie algebras and algebraic groups. , 38(3):311–327, 1979.
Jir[ō]{} Sekiguchi. A counterexample to a problem on commuting matrices. , 59(9):425–426, 1983.
E. B. Vinberg and O. S. Yakimova. Complete families of commuting functions for coisotropic [H]{}amiltonian actions. arXiv:math/0511498v2 \[math.SG\], 2007.
O. S. Yakimova. The index of centralizers of elements in classical [L]{}ie algebras. , 40(1):52–64, 96, 2006.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we describe a new approach for mm-VLBI calibration that provides bona-fide astrometric alignment of the mm-wavelength images from a single source, for the measurement of frequency dependent effects, such as ‘core-shifts’ near the black hole of AGN jets. We achieve our astrometric alignment by solving firstly for the ionospheric (dispersive) contributions using wide-band cm-wavelength observations. [ Secondly we solve for the tropospheric (non-dispersive) contributions by using fast frequency-switching at the target mm-wavelengths. These solutions can be scaled and transferred]{} from the low frequency to the high frequency. To complete the calibration chain one additional step was required to remove a residual constant phase offset on each antenna. The result is an astrometric calibration and the measurement of the core-shift between 22 and 43 GHz for the jet in BL Lacertae to be -8$\pm$5, 20$\pm$6 $\mu$as, in RA and Declination, respectively. By comparison to conventional phase referencing at cm-wavelengths we are able to show that this core shift at mm-wavelengths is significantly less than what would be predicted by extrapolating the low frequency result, which closely followed the predictions of the Blandford & Königl conical jet model. As such it would be the first demonstration for the association of the VLBI core with a recollimation shock, normally hidden at low frequencies due to the optical depth, which could be responsible for the $\gamma$-ray production in blazar jets.'
author:
- 'Richard <span style="font-variant:small-caps;">Dodson</span>, María J. <span style="font-variant:small-caps;">Rioja</span>, Sol N. <span style="font-variant:small-caps;">Molina</span>, José L. <span style="font-variant:small-caps;">Gómez</span>'
title: 'High-precision Astrometric Millimeter Very Long Baseline Interferometry Using a New Method for Multi-Frequency Calibration'
---
Introduction {#sec:intr}
============
[ Results of over eight years of monthly monitoring of a sample of blazars (the most luminous and variable BL Lac objects and flat-spectrum radio quasars) with the VLBA at 7 mm by the Boston University blazar group [@JorstadMarscher2016][^1] show that most $\gamma$-ray flares are simultaneous (within errors) with the appearance of a new superluminal component or a major outburst in the VLBI core of the jet, defined as the bright, compact feature at the upstream end of the jet [see @marscher_08; @marscher_10; @2013ApJ...773..147J; @2015ApJ...808..162C; @2015ApJ...813...51C]. A burst in particle and magnetic energy density is therefore required when jet disturbances cross the radio core in order to produce $\gamma$-ray flares, which can naturally be explained by identifying the radio core with a recollimation shock [e.g., @daly_88; @gomez_95; @gomez_97; @Marscher:2009vs; @Marscher:2012gh; @2015ApJ...809...38M; @Marti2016].]{}
On the other hand, the standard Blandford & Königl conical jet model hypothesizes that the core is not a physical feature in the jet, but corresponds to the location at which the jet becomes optically thin, and therefore its position shifts with observing frequency [@bk_79; @konigl_81; @lob_98]. This is conventionally referred to as the ‘core-shift’. In this case the separation from the black hole is $r$=$r_0 \nu^\kappa$, where $\nu$ is the frequency and $\kappa$ is a value close to -1 [@lob_98 additionally there can be an offset from the nominal reference position]. Multi-frequency VLBI observations at centimeter wavelengths have measured this core frequency shift in multiple sources, albeit without phase-referencing [e.g., @kovalev_08; @osullivan_09; @sokolovsky_11; @fromm_15]. [**Nevertheless phase-referenced VLBI observations have confirmed that the cm-wavelength radio core indeed is consistent with the optically thick-thin transition, in a smaller number of targets, such as 3C395, 4C39.25, 1038+528, 3C390.1, M81, M87 and 3C454.3 [@lara_94; @guirado_95; @rioja_98; @ros_01; @martividal_11; @hada_11; @kutkin_14 respectively].**]{}
We have therefore two sets of results, one suggesting that the radio core corresponds to a recollimation shock while the other implies that it marks the transition between the optically thick-thin jet regimes. A possible solution to reconcile these apparently contradicting observational results is to consider that the core is located parsecs away from the central black hole [e.g., @marscher_02; @chatterjee_11; @fromm_15] and consists of a recollimation shock that leads to $\gamma$-ray flares as new perturbations in the jet flow cross its position [e.g., @JorstadMarscher2016]. At this distance from the black hole the core is optically thin at mm-wavelengths, while at longer wavelengths the core becomes optically thick, leading to the observed Blandford & Königl core frequency shift.
We have performed numerical simulations to test this proposed model, using the finite-volume code [*Ratpenat*]{}, which solves the equations of relativistic hydrodynamics [@perucho_10 and references therein]. The jet is launched with an initial over-pressure of 1.5 times that of the external medium in order to obtain a recollimation shock that can be identified with the core. Using the hydrodynamical results as input, we have then computed the synchrotron emission at different observing frequencies [for details of the numerical model used see @gomez_95; @gomez_97; @2003ApJ...585L.109A]. This is illustrated by Fig. \[fig:cs\_mod\], which shows the sequence of total intensity images at the different frequencies, as well as the evolution of the core position with frequency. Full details of the simulations will be published elsewhere; here we just summarise the not-unexpected conclusions, which are that, at cm-wavelengths (5 to 22 GHz) the simulations reproduce the opacity core-shift of a Blandford & Königl conical jet model, while at mm-wavebands (43 and 86 GHz) the core position clearly departs from this behavior, revealing the recollimation shock at a fixed jet location.
![[*Left:*]{} Results from the numerical simulations. From top to bottom, a sequence of synchrotron total intensity images computed at 86, 43, 22, 15, 12, 8, and 5 GHz, respectively, using a relativistic hydrodynamical model of a jet with a recollimation shock. [*Right:*]{} Position of the core (peak emission) as a function of frequency. The red curve indicates the best fit to the core positions between 5 and 22 GHz, which follows the expected opacity core-shift of a conical Blandford & Königl jet model. The 43 and 86 GHz simulations clearly deviate from the opacity core-shift curve, revealing the fixed location of the recollimation shock associated with the core. \[fig:cs\_mod\]](JC01.eps "fig:"){width="32.00000%"}![[*Left:*]{} Results from the numerical simulations. From top to bottom, a sequence of synchrotron total intensity images computed at 86, 43, 22, 15, 12, 8, and 5 GHz, respectively, using a relativistic hydrodynamical model of a jet with a recollimation shock. [*Right:*]{} Position of the core (peak emission) as a function of frequency. The red curve indicates the best fit to the core positions between 5 and 22 GHz, which follows the expected opacity core-shift of a conical Blandford & Königl jet model. The 43 and 86 GHz simulations clearly deviate from the opacity core-shift curve, revealing the fixed location of the recollimation shock associated with the core. \[fig:cs\_mod\]](core_shift.eps "fig:"){width="66.00000%"}
Testing observationally whether the VLBI core at millimeter wavelengths is indeed consistent with the presence of a recollimation shock requires therefore bona-fide astrometric measurements of the core-shift spanning a wide range of centimeter and millimeter wavelengths. Confirmation of this model will support the hypothesis that the majority of the $\gamma$-ray flares in AGN jets are produced by the passing of new superluminal features through a pattern of recollimation shocks in the innermost jet regions, which would also include the VLBI core [@2013ApJ...773..147J; @2015ApJ...813...51C; @gomez_16]. If our model is correct, we expect to see the Blandford & Königl core-shift at cm-wavebands (5, 8.4, 15, and 22 GHz), with deviation from this behavior at mm-wavebands (43 and 86 GHz). [ It should be also noted that at the innermost jet regions probed at mm-wavebands, other effects, such as radiative cooling or the parabolic jet shape found in M 87 by @asada_12, may also lead to a departure from the opacity core-shifts of the conical Blandford & Königl jet model. However none of these effects appear to affect the measurements performed by [@hada_11], perhaps due to the progressive increase in the Doppler boosting as the jet accelerates in the innermost jet regions [@asada_14].]{} [ Additionally, recent space VLBI [*RadioAstron*]{} observations of BL Lac at a record angular resolution of 21 $\mu$as have found evidence for the association of the radio core with a recollimation shock [@gomez_16], providing extra motivation for this work.]{}
Conventional phase referencing (PR) [@alef88; @beasley] is the best approach for this analysis in cm-wavelength observations. The source/frequency phase referencing (SFPR) method works well for mm-wavelength observations, and has been demonstrated with frequencies as high as 130GHz (2mm) [@rioja_15]. However SFPR requires a second calibrator source with-in about 10$^o$ of the target. The density of calibrators at 86GHz, or even at 43GHz, is not sufficient to guarantee that a suitable source will be within this range. Indeed, this was the case for the source discussed in this paper: [[BLLac]{}]{}. Therefore we have developed a method built on the Frequency Phase Transfer (FPT) approach that under-pins SFPR, but does not require a second source. The description and validation of this method is the focus of this paper.
Observations {#sec:obs}
============
The observations presented here were carried out on 2013 July 5 with the Very Long Baseline Array (VLBA) targeting the jet in BL Lac, and are part of a series of similar experiments on a sample of blazar sources aimed to test the correspondence of the mm-VLBI core with a recollimation shock. The analysis here provides the demonstration and explanation of a new technique we have developed for single source astrometric $\lambda$-astrometry mm-VLBI, where the high frequency images are astrometrically registered to a lower frequency image.
[ All observations were made at 2Gbps with 32MHz Intermediate Frequency (IF) bands]{}. Prime calibrators, 3C345 and 3C84, were observed for all frequency bands. Ten blocks of conventional phase referencing of [[BLLac]{}]{} at 5, 8, 15 and 22 GHz were performed, with [J2153+4322]{} and J2218+4146 as the reference sources and a cycle time of 80 seconds. Following each phase referencing block, we had fast frequency-switching observations, just on [[BLLac]{}]{}, between 22–43GHz and 22–86 GHz, with 30 seconds per scan over a 15 minute block. These were bracketed by ionospheric calibration blocks, which consist of observations switching between the L-band receiver range (16IFs between 1.4–1.7GHz), the wide band C-band receiver (16IFs between 3.9–7.9GHz) and the K-band receiver (16IFs between 21.8–22GHz), with 40 seconds of observing time at each band. [ For the switching observations the on-source time was 1.3 hours for each frequency in the pair.]{} [ The data reduction was in AIPS, following the standard path of correction for SEFD amplitudes, correction for Earth Orientation Parameters and correction for the Ionosphere, based on GPS models.]{}
Multi-Frequency Phase Referencing {#sec:meth}
=================================
The Source Frequency Phase Referencing method, which has been described in detail elsewhere [@vlba_31; @rioja_11a; @rioja_11b; @rioja_14; @rioja_15; @dodson_14], consists of two calibration steps. In a first step, the observations at the higher frequency bands are calibrated using near-simultaneous [@vlba_31; @rioja_11a; @rioja_14] or simultaneous [@rioja_14; @rioja_15; @dodson_14] observations at a lower frequency band, for each source. This is done for all frequency pairs which have an integer[^2] frequency ratio, by which the low frequency calibration phase solutions are scaled. This dual frequency calibration step eliminates the common non-dispersive residual errors (e.g. tropospheric propagation effects and geometric errors) in the complex visibility output of the correlator, providing an increased signal coherence at the higher frequency. The second step of the calibration removes the remaining dispersive residual errors (i.e. instrumental and ionospheric propagation effects) using the interleaved observations of another source. This two-step calibration retains the astrometric signature of any source position shifts between the two frequencies in the interferometric phase observable. The Fourier transformation of the SFPR dataset is the SFPR map, which conveys a bona-fide astrometric measurement of the relative separation or shift between the reference points in the images at the two frequencies, for the two sources. Results from SFPR analysis are to be found in the cited papers.
SFPR relies on the observation of a second [or multiple, as in the analysis of @rioja_15] calibrator. This can be some distance from the target, with successful demonstrations with separations as large as 11[@vlba_31; @rioja_15], but there are sources for which one still struggles to find a suitable calibrator at the highest frequencies. This is the case for [[BLLac]{}]{} where no mm-wavelength calibrator, for direct or reverse SFPR, could be found within 10. Therefore we have attempted to achieve phase referencing in a similar fashion to SFPR, but without the second source. Our approach in this experiment is to calibrate all frequencies against a well-known source with precise astrometric position, then solve for the residual delays for the target, across a wide frequency span. This allows us to measure the residual TEC in the target direction, which is used to produce an ionosphere-free dataset for all frequencies. This method we dub Multi-Frequency Phase Referencing (MFPR), as now our calibration scheme allows relative astrometry between the (mm) frequency bands corrected by observations at multiple (cm) frequencies of the target.
Methods {#sec:method}
=======
The prime calibration was against 3C345, except for 86 GHz, where we had many missed scans and the data quality was very poor. For this frequency 3C84 was the prime calibrator and [ prime calibration]{} could only be performed for the BR, KP, LA, OV and PT antennas. Prime calibration removes the instrumental terms, plus all the atmospheric contributions in the direction of the calibrator, at the time of the observations. We handled the structural contribution from the source by hybrid mapping the data before astrometric calibration, to produce a reference image which is used in the analysis. The conventional phase referencing at cm-wavebands was calibrated following standard procedures and the detailed interpretation will be reported elsewhere [@molina_bllac].
For the ionospheric correction blocks we use the delay (only) from each IF to measure the Total Electron Content (TEC) contribution on the line of sight towards [[BLLac]{}]{}, as a function of time. [ This is a measurement of the residual ionospheric contribution, after correction with the GPS data and the subtraction of the TEC in the direction of the prime calibrator, at the time of that scan.]{} We fitted a linear slope in [ $\nu^{-2}$]{} to the semi-simultaneous delay measurements (i.e. such that $\tau(\nu)=\tau_{\rm trop}+\tau_{\rm iono} \nu_{GHz}^{-2}$) and calculate the residual TEC contribution as a function of time, for that line of sight, using $\Delta$TEC=0.75$\tau_{\rm iono}$. Here $\tau(\nu)$ are the measured delays as a function of frequency for one block of ionospheric calibration observations, $\tau_{\rm trop}$ is the non-dispersive delay [ (from both clock and tropospheric contributions)]{}, $\tau_{\rm iono}$ is the ionospheric delay (at 1 GHz), $\nu_{GHz}$ is the frequency in GHz and $\Delta$TEC is the deduced residual ionospheric contribution in Total Electron Count Units (TECU), to the line of sight of the target. [ We are assuming, as we have corrected the instrumental terms ($\tau_{\rm inst}$) using the prime calibration, that these are constant during the experiment and therefore can be ignored in this analysis.]{} The derived $\Delta$TEC is then used to calculate the ionospheric contribution for each IF, at each time interval. The frequency-dependent delay can directly be calculated, but note that the sign of the phase has to be reversed, as the ionospheric contribution is a group delay not a phase delay. Finally we solved for the (ionosphere-free) delay, rate and phase on the (ionosphere-corrected) 22 GHz data and applied these solutions, suitably scaled by the frequency ratio, to the ionosphere-corrected 43 and 86 GHz data. [ As this calibration scheme uses low frequency calibration blocks, either side of the high frequency science observations, and these are to correct for the ionospheric contributions to the atmosphere, we dubbed these IonospheriC Excision blocks (ICE-blocks). ]{} This recognises the commonality with the so called Geodetic Block calibration schemes that dealt with the static tropospheric contributions [@brunthaler_05; @honma_08; @reid_micro]. Our expectations were that, having corrected for the ionospheric and the tropospheric contributions, we should be left with the high frequency datasets astrometrically aligned to the 22 GHz dataset. However, as will be discussed in the results, we found that our initial calibration scheme was inadequate. In our observations of 3C345 we have a gap of about five minutes between the prime calibration observations at 22 and 43 GHz, which was used for a pointing correction. An even longer gap existed between the 22 and 86 GHz prime-calibrator scans. Therefore we could not align the phase of these scans to a common point in time. This caused an unknown phase offset to be introduced between the data at 22 GHz and higher frequencies, in the observation of the target source at the target frequencies. Therefore prime-calibration should have included a fast frequency-switching scan on the calibrator as well as the target; this will be included in future observations. [ How we resolved this problem for these observations is discussed in Section \[sec:res:off\].]{}
Results {#sec:res}
=======
Ionospheric measurements
------------------------
Each ICE-Block consists of multiband observations at 1.4–1.7, 3.9–7.9 and 22 GHz, that is 21 to 18 cm, 8 to 4 cm, and 1.3 cm. These bracket the fast frequency-switching observations of the target source, which either consist of 22–43 or 22–86 GHz blocks. The ICE-blocks are calibrated following standard procedures and the measured delays for each IF fitted to derive $\Delta$TEC. Figure \[fig:tec\_mu\] shows the delays as a function of frequency for one scan on one antenna. Figure \[fig:tec\_t\] shows the derived $\Delta$TEC for all antennas during this experiment. [ The fitting errors can be used to estimate the measurement precision of the $\Delta$TEC residuals. The error level was $\sim$0.1TECU, if one excluded MK and SC (which were not included in the final MFPR imaging, due to the poor quality of the data) and 0.2TECU if all antennas are included. 0.1TECU would contribute 2.2$^\circ$ of phase at 22 GHz, which would be scaled up to 3.3$^\circ$ and 8.2$^\circ$ of phase noise at 43 and 86 GHz, respectively [@rioja_11a].]{}
![A typical delay $\tau(\nu)$ for one particular antenna (BR) and 2 minute ICE-block scan (UT 05:56), as a function of wavelength. It shows the curvature (following $\nu^{-2}$) that directly measures the $\Delta$TEC residual (in this case -4.4 TECU, indicated with the solid line), for that solution interval and for that antenna, in the line of sight of the target. \[fig:tec\_mu\]](delay_fit_1_5.eps){width="10cm"}
![The $\Delta$TEC residuals for all antennas across the duration of the experiment. The values for most antennas range between $\pm$5TECU, as would be expected for data that has been corrected with the default TEC maps, using TECOR. The antennas MK and SC show the largest deviations. Errors in individual measurements are typically 0.1TECU. \[fig:tec\_t\]](tec_sols2.eps){width="10cm"}
Constant Phase Offset measurement and Astrometric Results {#sec:res:off}
---------------------------------------------------------
After subtraction of the calculated ionospheric delays and phases from the 22, 43, and 86 GHz datasets, we self-calibrated the 22 GHz [[BLLac]{}]{} data against a hybrid map of the source, scaled the solutions by the frequency ratio and use these to correct the higher frequencies. Thus the ionospheric and tropospheric contributions are removed. However, because the initial prime-calibrations at different frequencies were not at close points in time, there is an introduced constant, but unknown, phase at each station. This is shown in Fig. \[fig:fpt\] for 43 and 86 GHz. The Fourier inversion of this data, after excluding antennas MK and SC plus the low elevation data (the first and last hour), gives us our initial image. However this image did not recover the source structure, because of the error in the initial calibration chain.
![Ionosphere and Troposphere corrected visibilities on [[BLLac]{}]{} showing the dominant constant phase offset (indicated with light grey lines) arising from the non-simultaneous observation of the prime calibrator, for [*left*]{} 43 GHz with baselines to Los Alamos (LA) only and [*right*]{} 86 GHz with all baselines that could be calibrated. \[fig:fpt\]](bllac-resid-ph-2.eps "fig:"){width="8cm"} ![Ionosphere and Troposphere corrected visibilities on [[BLLac]{}]{} showing the dominant constant phase offset (indicated with light grey lines) arising from the non-simultaneous observation of the prime calibrator, for [*left*]{} 43 GHz with baselines to Los Alamos (LA) only and [*right*]{} 86 GHz with all baselines that could be calibrated. \[fig:fpt\]](bllac-resid-ph-w-2.eps "fig:"){width="8cm"}
To remove a constant phase yet not lose the astrometric signal we trialed all possible models (point sources on a 10 $\mu$as grid within 0.5 mas of the phase centre), generating a single phase correction for each antenna for the whole dataset, using self-calibration. Using this calibration table we repeated the imaging and inspected the peak flux, the residual RMS, and the dynamic range as a function of the model. The peak in the dynamic range, plotted in Fig. \[fig:quality\_dr2\], gives an astrometric alignment of -18$\pm$10,+12$\pm$10 $\mu$as, based on the 95% confidence limits of a 2D second order polynomimal fit to the surface. [ Note that the dynamic ranges achieved (a maximum of $\sim$20) for a single point source fit are much less than those for the self-calibrated hybrid images, for which we had a dynamic range of $\sim$200.]{} The most precise test, however, is the alignment of the location of the peak flux and that of the trial model. The minimum absolute offset in the alignment gives an astrometric result of -8$\pm$5,20$\pm$6 $\mu$as, based on the 95% confidence limits of two orthogonal 1D second order polynomial fits to the surface, as shown in Figure \[fig:quality\_align\]. The precision of the alignment is approximately double that of the precision from the peak in the dynamic range so we adopt this value as our astrometric result. The image of [[BLLac]{}]{} made with the phase corrections for an offset of -8$\pm$5,20$\pm$6 $\mu$as is shown in Fig. \[fig:qband\_mfpr\]. [ Unfortunately the 86 GHz data was of too poor quality to produce useful results]{}. In a subsequent paper we will extend our analysis to 86 GHz using a better dataset from our sample of blazars included in this project.
![The dynamic range as a function of the $\Delta \alpha$,$\Delta \delta$ of the input model. Also marked is the peak value, at -18, 12 $\mu$as, and the error bounds of $\pm$10 $\mu$as. \[fig:quality\_dr2\]](DynamRange.eps){width="8cm"}
![The alignment of the requested model and the resultant peak of emission, as a function of the $\Delta \alpha$,$\Delta \delta$ of the input model. Also marked is the best alignment at -8, 20 $\mu$as and the error bounds of $\pm$5, 6 $\mu$as. \[fig:quality\_align\]](Align_k.eps){width="8cm"}
![Image of [[BLLac]{}]{} at 43 GHz, Multi-frequency Phase Referenced to the 22 GHz data. The offset from centre (-8,+20 $\mu$as) is the bona-fide astrometric core shift between 22 and 43 GHz. Contours are 60, 70, 80, 95 and 99% of the peak flux (3.3 Jy/beam). The restoring beam size is 0.25$\times$0.34, PA -22$^o$. \[fig:qband\_mfpr\]](bllac-43ghz.eps){width="8cm"}
To test the physicality of our results we compared the derived 22–43 GHz core-shift measurement to the preliminary results from the cm-wavelength phase referencing. We performed conventional phase referencing for the 4.8, 8.4, 15, and 22 GHz (see Fig. \[fig:conv\_pr\]) data against the compact calibrator [J2153+4322]{} (Fig. \[fig:conv\_pr3\]), which is 2$^\circ$ from [[BLLac]{}]{}. The images are similar and in agreement with published observations. We restored the [[BLLac]{}]{} images to the common beamsize of the 22 GHz image used as the reference for the FPT to avoid blending issues that otherwise shift the apparent positions. We measured the relative positions of the peak of emission (compared to that of the referenced source) at all frequencies and found a systematic shift across the sky between 4.8 and 22 GHz, with $r_0$ of 5.3 masGHz$^{\kappa}$, $\kappa$ of -0.99 and an offset of 45 $\mu$as. These parameters would lead to a prediction of 120$\pm$30 $\mu$as for the core-shift between 22 and 43 GHz. This is much greater than found. Additionally we fitted for all frequencies, including 43 GHz. The best fit was obtained with $r_0$ of 7.7 masGHz$^{\kappa}$, $\kappa$ of -1.32 and an offset of 190 $\mu$as. This fit, although valid, [ would imply a significant departure from the expected value of $\kappa \simeq$-1, for the case of equipartition between jet particle and magnetic field energy densities [e.g., @lob_98], as found at cm-wavebands in BL Lac and other multiple sources [i.e., @osullivan_09; @sokolovsky_11; @hada_11]. It would also imply a large]{} error in the astrometric position of [[BLLac]{}]{} and/or [J2153+4322]{}, which is unlikely. Figure \[fig:mm2cm\_cs\] shows the location of the peaks of emission, along with predictions from the two models.
![Overlaid astrometrical-registered images of [[BLLac]{}]{} for the nominal phase centre of 22:02:43.291 and 42:16:39.980 in RA and Declination, derived using conventional methods and phase referenced to [J2153+4322]{}. Four frequencies are plotted: 4.8, 8.4, 15.2 and 21.9 GHz in red, green, blue and black respectively. Contours are 60, 70, 80, 95 and 99% of the peak flux for each frequency, which are 1.6, 3.7, 2.1 and 1.6 Jy/beam. \[fig:conv\_pr\]](bllac-pr2.eps){width="8cm"}
![Overlaid self-calibrated images of [J2153+4322]{} for the nominal phase centre of 21:53:50.959 and 43:22:54.500 in RA and Declination. All images are restored with the 22 GHz beam parameters of the image (0.57$\times$0.72, PA -26$^o$) used to reference the 43 GHz data. Four frequencies are plotted: 4.8, 8.4, 15.2 and 21.9 GHz in red, green, blue and black respectively. Contours are 60, 70, 80, 95 and 99% of the peak flux for each frequency, which are 0.25, 0.28, 0.12 and 0.08 Jy/beam. \[fig:conv\_pr3\]](J2153+4322-pr2.eps){width="8cm"}
![Astrometric core-shifts of [[BLLac]{}]{}, derived using conventional phase referencing between 4.8 and 22 GHz, plotted as a function of frequency. All frequencies were restored with the beamsize at 22 GHz. Additionally we plot the position of the 43 GHz peak emission, derived using our MFPR method between 22/43GHz, and added to the position of peak emission of the 22 GHz. Error bars show estimates for the accuracy in conventional phase referencing (beam-width over dynamic range), except for the MFPR result, where the errors are from the measured precision given in the text. Overlaid in blue is the model from the fitting of the cm-wavelength data (where $\kappa$ is -0.99 and $r_0$ is 5.3 masGHz$^{\kappa}$) and in red dashes the fit for all data (in which case $\kappa$ is -1.32 and $r_0$ is 7.7 masGHz$^{\kappa}$) \[fig:mm2cm\_cs\]](bk_model_vs_data_abs2.eps){width="8cm"}
Discussions {#sec:disc}
===========
[ We have demonstrated that the effects of the atmosphere can be calibrated in a step-wise manner, by decomposing them into dispersive and non-dispersive contributions.]{} We have included for the first time a wide-band measurement of the residual ionospheric contributions to VLBI data, for the line of sight of the target. Our measurement of the ionosphere, derived from the group delay curvature, is a development of the standard geodetic two point fit (at 2.4 and 8.4-GHz) approach [@sxfeeds; @new_geodetic], whereas we fit the data over multiple frequencies. We performed simple simulations with typical values for $\Delta$TEC and $\Delta\ell$, the residual TEC and path length, and measurement errors of 0.1nsec (1$^\circ$ phase error across 16 MHz). We explored the required frequency span to be able to predict accurate values for $\Delta$TEC, and found that the lower frequencies are the most crucial. At least 2.4 GHz would be required, and 1.4 GHz would be preferred. [ Using our data we investigated]{} the achievable [ reliability]{} using just the new VLBA wideband C-band and K-band data, and found the accuracy in the determined $\Delta$TEC to be $\sim$1TECU. [ Such levels are those expected (i.e. by scaling with $(\nu_1/\nu_2)^2$) and would be sufficient mitigate the ionospheric contribution to the errors in the astrometry of methanol masers at 6.7GHz.]{}
The Frequency Phase Transfer has been thoroughly demonstrated previously [@rioja_15 and references therein], therefore we would expect the approaches taken here to be successful. The conditions for which we may expect this to break down are: where instrumental terms are not stable, introducing a time variable non-dispersive term that is not a function of [ $\nu^{-2}$]{}; and where there are baseline dependent terms introduced, by for example having a poor model of the prime calibrator. For our analysis we do not believe the first of these are issues, as the VLBA has extremely stable instrumental terms.
[ We addressed the issue of the baseline dependent phase terms introduced by the prime-calibrator by using a hybrid model in the initial fringe fitting stage.]{} This worked well for the core-dominated source 3C345, but failed for the more complex structure of 3C84, where there were insufficient data at 86 GHz to constrain the model. We note that any core-shift in the prime calibrator will appear as a constant phase introduced between the two frequencies, as it will arise from a single scan. [ Therefore it will be absorbed into the constant phase corrections discussed next.]{}
To remove the phase introduced between the two frequencies from the prime-calibration we performed a grid search to find the best constant phase that matched the data. Our approach of testing for model stability is similar to super resolution. In super-resolution one adjusts the model to the uv-data to locate the best fitting location using a minimisation method, thereby producing a positional accuracy greater than the resolution. In our case we are testing the fit of a model to the uv-data, with a cycle of self-calibration, in a stepwise fashion, over the parameter space. [**This aspect of our analysis is perhaps the most innovative, but also suffers from the limitations in our minimisation methods. We have no mechanism to measure the co-variance between the fitted parameters of source position and constant station-based phase offset. If the experiment had been a ‘snapshot’ these two would be degenerate, however as the observations spanned 8 hours this degeneracy is broken. We estimated our errors from the fitting to the 2D surface of the results. The use of independent measures for these surfaces reduces the possibility of degeneracy. That the maximum dynamic range (and peak flux and minimum residual RMS) align with those of the minimum absolute offset in the alignment gives confidence in our results.**]{}
Our conclusions would be significantly strengthened if we were able to produce a joint analysis of the 22 to 43 and 86 GHz data. However this was not possible, because of the poor quality of the observations at 86 GHz, due to technical issues. These data could not even be self-calibrated to produce acceptable images. Therefore we believe the residuals probably arise from baseline-based (rather than station based) contamination from the prime-calibrator scan, which can not be corrected for by our procedures.
Ideally we would compare our new method with results from conventional phase referencing for the same source. However if conventional phase referencing was possible at these frequencies we would not have needed to develop these new methods. [ We used the measurement of the core-shifts made at lower frequencies with conventional phase referencing, and extrapolated these results to compare with our MFPR measurement made at mm-wavelengths.]{} However these can provide only the Blandford & Königl core-shift; our expectation is that the higher frequencies would have smaller than predicted core-shifts. This is consistent with what we discover, as shown in Fig. \[fig:mm2cm\_cs\]. Consequently we can not use this approach to validate our method.
We attempted to derive the core shift following the approach of aligning optically thin features, but we could not get sufficiently accurate results from our own data. [ Optically thin features tend to be of lower surface brightness and therefore it is difficult to accurately determine the centroids [@hovatta_14].]{} Nevertheless cross-correlation of VLBA images firstly by @osullivan_09, and secondly by @gomez_16 (with observations made a few months after ours in November 2013) have determined a core-shift of 30$\pm$20 $\mu$as and 21 $\mu$as, respectively, between 22 and 43 GHz. This is in close agreement with the value we find from the MFPR analysis, which provides extra support for the reliability of our new method. The alignment reported by @osullivan_09 is based on the cross correlation of images at frequencies between 5 and 43 GHz. They find an average core-shift of the 22GHz position, referenced to 43 GHz, of 40$\pm$20$\mu$as, whereas we have compared with only the measurement between 22-43 GHz. All measurements are consistent within the errors. They fitted their cross-correlation alignments to obtain a value of $\kappa$ (following our definition) of -1.01 between 5 and 43GHz, which would a first sight appear to be in contradiction with our results. However the functional expression ($(\nu_2-\nu_1)/(\nu_1 \nu_2)$) that they fit is not very sensitive to deviations between 22 and 43 GHz. [ Replacing all of their 43GHz measurements with the 22GHz values (i.e. inserting a zero core-shift between 22 and 43GHz) and repeating the fitting reproduces the published result, within errors.]{} Furthermore, in our analysis, we found that blending of components within the uniform weighted beam at the low frequencies distorted our results, which is why we restored with the super resolved beam of the 22 GHz. This important issue was not addressed in the @osullivan_09 analysis.
[ This multi-frequency approach is one of two methods we are currently developing to improve astrometric calibration of the ionospheric contributions, the other being to use multiple calibrators around the target and to solve for the spatial structure of the atmosphere, which we call [*MultiView*]{} [@rioja_09; @rioja_16]. This method, which requires more calibrators rather than less, simultaneously solves for a 2D Ionospheric and Tropospheric phase screen over the array, and will be extremely suitable for VLBI stations with multiple beams, such as SKA and ASKAP.]{}
Conclusions {#sec:conc}
===========
We have presented a development of the SFPR method, which we call Multi-Frequency Phase Referencing (MFPR). This has been used to measure the core-shift (or the $\lambda$-astrometry) for [[BLLac]{}]{}, between 22 and 43 GHz. The MFPR method presented here involved a measurement of the $\Delta$TEC on the line of sight of the target, with precision of 0.1TECU. Once the data is corrected for the ionosphere the tropospheric correction is measured at the lower frequency and applied to the higher frequency. The high frequency data should then be registered to the low frequency data using Frequency Phase Transfer. Our data required an additional step, to minimise over a constant phase offset introduced by the prime calibration, which can be avoided in future observations with improved scheduling. After these steps the astrometric offset between the two frequencies is the bona-fide core-shift, which has been derived from observations of a single source by careful calibration of the atmospheric contributions.
This new method opens up a large number of possibilities for astrometric analysis in mm-VLBI. Conventional phase referencing is not possible for mm-VLBI, and for a significant number of sources, such as [[BLLac]{}]{}, a suitable calibrator as required for SFPR can not be found. The method of alignment of optically thin components [ suffers from both questionable validity, and a requirement of high sensitivity]{} to detect low surface brightness features. MFPR bypasses all of these issues, requiring only a detectable target source at the lower frequencies. With simultaneous observations this method will be applicable to mm and sub-mm wavelengths.
The results from the MFPR measures a core-shift for [[BLLac]{}]{} of -8$\pm$5,20$\pm$6 $\mu$as between 22 and 43 GHz. This is significantly less than the prediction from measurements of the core-shift at cm-wavelengths, but in line with both the theoretical expectations and other work.
Further analysis, to be published in @molina_bllac, will improve the initial measurement of the core-shift at cm-wavelengths, allowing us to deduce if we are truly uncovering deviations from the Blandford & Königl model. If so this would be the first detection of the predicted association of the mm-VLBI core with a recollimation shock responsible for the $\gamma$-ray emission in blazar [ jets, in agreement with the findings of [@gomez_16].]{} The fact that the calibration of the 86 GHz showed promise but was defeated by an unusually large antenna failure rate gives us confidence that we will be able to perform MFPR at 86 GHz in future demonstrations.
[**Acknowledgements**]{}
We would like to especially acknowledge the numerous conversations with Richard Porcas about the best approaches for the analysis, and the referee for comments that improved the paper. The work at IAA-CSIC is supported by the Spanish Ministerio de Economía y Competitividad grant AYA2013-40825-P. The VLBA is operated by National Radio Astronomy Observatory and is a facility of the National Science Foundation operated under cooperative agreement with Associated Universities Inc.
[51]{} natexlab\#1[\#1]{}
, W. 1988, in IAU Symposium, Vol. 129, The Impact of VLBI on Astrophysics and Geophysics, ed. M. J. [Reid]{} & J. M. [Moran]{}, 523
, M.-[Á]{}., [Mart[í]{}]{}, J.-M., [G[ó]{}mez]{}, J.-L., [et al.]{} 2003, , 585, L109
, K., & [Nakamura]{}, M. 2012, , 745, L28
, K., [Nakamura]{}, M., [Doi]{}, A., [Nagai]{}, H., & [Inoue]{}, M. 2014, , 781, L2
, A. J., & [Conway]{}, J. E. 1995, in Astronomical Society of the Pacific Conference Series, Vol. 82, Very Long Baseline Interferometry and the VLBA, ed. J. A. [Zensus]{}, P. J. [Diamond]{}, & P. J. [Napier]{}, 327
, R. D., & [K[ö]{}nigl]{}, A. 1979, , 232, 34
, A., [Reid]{}, M. J., [Falcke]{}, H., [Greenhill]{}, L. J., & [Henkel]{}, C. 2005, Science, 307, 1440
, C., [G[ó]{}mez]{}, J. L., [Jorstad]{}, S. G., [et al.]{} 2015, , 813, 51
Casadio, C., G[ó]{}mez, J. L., Grandi, P., [et al.]{} 2015, ApJ, 808, 162
, R., [Marscher]{}, A. P., [Jorstad]{}, S. G., [et al.]{} 2011, , 734, 43
, R. A., & [Marscher]{}, A. P. 1988, , 334, 539
, R., & [Rioja]{}, M. J. 2009, [VLBA Scientific Memorandum n. 31: Astrometric calibration of mm-VLBI using ”Source/Frequency Phase Referenced” observations]{}, Tech. rep., NRAO
, R., [Rioja]{}, M. J., [Jung]{}, T.-H., [et al.]{} 2014, , 148, 97
, C. M., [Perucho]{}, M., [Ros]{}, E., [Savolainen]{}, T., & [Zensus]{}, J. A. 2015, , 576, A43
, J. L., [Mart[í]{}]{}, J. M., [Marscher]{}, A. P., [Ib[á]{}[ñ]{}ez]{}, J. M., & [Alberdi]{}, A. 1997, , 482, L33
, J. L., [Marti]{}, J. M. A., [Marscher]{}, A. P., [Ibanez]{}, J. M. A., & [Marcaide]{}, J. M. 1995, , 449, L19
, J. L., [Lobanov]{}, A. P., [Bruni]{}, G., [et al.]{} 2016, , 817, 96
, J. C., [Marcaide]{}, J. M., [Alberdi]{}, A., [et al.]{} 1995, , 110, 2586
, K., [Doi]{}, A., [Kino]{}, M., [et al.]{} 2011, Nat, 477, 185
, M., [Tamura]{}, Y., & [Reid]{}, M. J. 2008, , 60, 951
, T., [Aller]{}, M. F., [Aller]{}, H. D., [et al.]{} 2014, , 147, 143
, S., & [Marscher]{}, A. 2016, Galaxies, 4, 47
, S. G., [Marscher]{}, A. P., [Smith]{}, P. S., [et al.]{} 2013, , 773, 147
, A. 1981, , 243, 700
, Y. Y., [Lobanov]{}, A. P., [Pushkarev]{}, A. B., & [Zensus]{}, J. A. 2008, , 483, 759
, A. M., [Sokolovsky]{}, K. V., [Lisakov]{}, M. M., [et al.]{} 2014, , 437, 3396
, L., [Alberdi]{}, A., [Marcaide]{}, J. M., & [Muxlow]{}, T. W. B. 1994, , 285, 393
, A. P. 1998, , 330, 79
Marscher, A. P. 2009, Approaching Micro-Arcsecond Resolution with VSOP-2: Astrophysics and Technologies ASP Conference Series, 402, 194
—. 2012, Int. J. Mod. Phys. Conf. Ser., 08, 151
, A. P., [Jorstad]{}, S. G., [G[ó]{}mez]{}, J.-L., [et al.]{} 2002, , 417, 625
, A. P., [Jorstad]{}, S. G., [D’Arcangelo]{}, F. D., [et al.]{} 2008, , 452, 966
, A. P., [Jorstad]{}, S. G., [Larionov]{}, V. M., [et al.]{} 2010, , 710, L126
, J. M., [Perucho]{}, M., & [G[ó]{}mez]{}, J. L. 2016, ArXiv e-prints
, I., [Marcaide]{}, J. M., [Alberdi]{}, A., [et al.]{} 2011, , 533, A111
, Y., [G[ó]{}mez]{}, J. L., [Nishikawa]{}, K.-I., [et al.]{} 2015, , 809, 38
Molina, S. 2016,
, S. P., & [Gabuzda]{}, D. C. 2009, , 400, 26
, M., [Bosch-Ramon]{}, V., & [Khangulyan]{}, D. 2010, , 512, L4
, M. J., & [Honma]{}, M. 2014, , 52, 339
, M., & [Dodson]{}, R. 2011, AJ, 141, 114
, M., [Dodson]{}, R., [Malarecki]{}, J., & [Asaki]{}, Y. 2011, AJ, 142, 157
, M., [Dodson]{}, R., [Orosz]{}, G., [Imai]{}, H., & [Frey]{}, S. 2016, AJ
, M., [Dodson]{}, R., [Porcas]{}, R. W., [et al.]{} 2009, in 8th International e-VLBI Workshop, 14
, M. J., [Dodson]{}, R., [Jung]{}, T., & [Sohn]{}, B. W. 2015, , 150, 202
, M. J., & [Porcas]{}, R. W. 1998, in Astronomical Society of the Pacific Conference Series, Vol. 144, IAU Colloq. 164: Radio Emission from Galactic and Extragalactic Compact Sources, ed. J. A. [Zensus]{}, G. B. [Taylor]{}, & J. M. [Wrobel]{}, 95
, M. J., [Dodson]{}, R., [Jung]{}, T., [et al.]{} 2014, , 148, 84
, E., [Marcaide]{}, J. M., [Guirado]{}, J. C., & [P[é]{}rez-Torres]{}, M. A. 2001, , 376, 1090
, K. V., [Kovalev]{}, Y. Y., [Pushkarev]{}, A. B., [Mimica]{}, P., & [Perucho]{}, M. 2011, , 535, A24
Sovers, O. J., Fanselow, J. L., & Jacobs, C. S. 1998, Rev. Mod. Phys., 70, 1393
Williams, W., Nixon, D., Reilly, H., Withington, J., & Bathker, D. 1979, DSN Progress Report 42, 52, 51
[^1]: The VLBA-BU-BLAZAR monitoring program; see http://www.bu.edu/blazars/VLBAproject.html
[^2]: for non-integer ratios see the analysis and discussions in @dodson_14
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We measure Young’s modulus of a free polyelectrolyte multilayer film by studying osmotically induced swelling of polyelectrolyte multilayer microcapsules filled with the polyelectrolyte solution. Different filling techniques and core templates were used for the capsule preparation. Varying the concentration of the polyelectrolyte inside the capsule, its radius and the shell thickness yielded an estimate of an upper limit for Young’s modulus of the order of 100 MPa. This corresponds to an elastomer and reflects strong interactions between polyanions and polycations in the multilayer.'
author:
- 'O. I. Vinogradova'
- 'D. Andrienko'
- 'V. V. Lulevich'
- 'S. Nordschild'
- 'G. B. Sukhorukov'
title: 'Young’s modulus of polyelectrolyte multilayers from microcapsule swelling '
---
Introduction
============
Molecularly thin polyelectrolyte multilayer films are composed of alternating layers of oppositely charged polyions and are important for a variety of potential applications [@1; @2]. Supported multilayer films are normally produced via layer-by-layer (LbL) adsorption of polyanions and polycations on a planar [@1] or spherical [@3] charged solid surface. In the later case, the colloidal template can be dissolved to give so-called polyelectrolyte microcapsules [@4]. The shell of such microcapsules is nothing more than a free standing multilayer film. The free thin film geometry allows investigation of properties not accessible in the bulk or in supported films and helps to gain a better understanding of polyelectrolytes in general.
There have been a number of recent experimental studies of mechanical properties of thin multilayer films. First, by studying osmotically induced buckling of “hollow” (water inside) capsules immersed in a polyelectrolyte solution [@5], the Young’s modulus was found to be above 1000 MPa, close to the elasticity of the *bulk plastics* [@6]. The second, more recent, approach is based on measuring the deformation of microcapsules under applied load using an atomic force microscope (AFM) [@7; @8; @9]. This method yielded an estimate of the limit for Young’s modulus of the order of 1-10 MPa which corresponds to an *elastomer* [@6]. The reason for such a discrepancy between two approaches might be hidden in the assumptions of the models used to fit experimental data. The first model assumes that the shell is highly permeable for water, even on short timescales of the relaxation of buckling deformation. The second approach treats the shell as impermeable for water on short timescales, relying on the conservation of capsule volume for small deformations. The accuracy of existing experimental results does not allow to conclude which approach is more realistic.
In this paper we propose an alternative way to probe the elastic properties of a polyelectrolyte multilayer. The method is based on studying the swelling of microcapsules filled with a solution of a strong polyelectrolyte. The size of a swollen capsule depends on the Young’s modulus of the capsule shell. Fitting the experimental data to the prediction of a simple model yields Young’s modulus of the polyelectrolyte multilayer of the order of 100 MPa. We argue that our results give the limit of Young’s modulus. Taken together with recent AFM observations [@7], this result leads to the conclusion that the mechanical behavior of a polyelectrolyte multilayer is that of an *elastomer.*
Theoretical model
=================
We consider a “filled” (polyelectrolyte solution inside) capsule immersed in a low molecular weight solvent (water in our case). The capsule swells due to excess osmotic pressure of the inner solution. The osmotic pressure of polyelectrolyte solutions is the sum of polymer and counterion contributions. However, in a salt-free solution the latter exceeds the osmotic pressure due to polymer itself by several orders of magnitude [@10; @11; @12]. The capsule shell is permeable to the solvent (on time scales larger than the characteristic diffusion time), but impermeable to the encapsulated polymer of high molecular weight. Then the solvent diffuses into the capsule until the elastic force of the stretched shell balances the osmotic pressure. We assume that the inner polyelectrolyte solution remains electroneutral, i.e. all counter ions due to polyelectrolyte dissociation remain in the capsule interior. A similar assumption was used in Ref. [@5].
We also assume that, for small relative deformations, the response of the capsule shell is elastic. If the capsule swells from the initial radius $r_0
$ to a final radius $r$ the energy of stretching of the shell is given by the elastic theory of membranes [@13]
$$\label{eq1}
G = 4 \pi \frac{E}{1 - \nu }h\left( {r - r_0 } \right)^2,$$
where $E$ is Young’s modulus, $\nu $ is Poisson’s ratio ($\nu \approx 1 /
2)$, and $h$ is the shell thickness ($h < < r_0 )$.
For a dilute solution of the inner polyelectrolyte, or, alternatively, small concentration $n = N / V$ of counterions in the capsule, the osmotic pressure induced by the counterions reads
$$\label{eq2}
\Pi = \varphi \frac{N}{V}k_B T.$$
Here $N$ is the number of counterions and $\varphi \le 1$ is the osmotic coefficient, defined as the ratio of experimentally measured osmotic pressure $\Pi $ to the ideal osmotic pressure of all counterions. The difference between these two values is due to a fraction of condensed counterions being bound to the polyelectrolyte chain and not contributing to the osmotic pressure [@14; @15; @16].
The work done by the osmotic pressure $\Pi $ to swell the capsule from radius $r_0 $ to $r$ then reads
$$\label{eq3}
A = \int_{V_0 }^V {\Pi dV} = - 3\varphi Nk_B T\ln \frac{r}{r_0 }.$$
The equilibrium radius of the capsule is given by the minimum of the total energy $F = G + A$, where $\partial F / \partial r = 0$, giving
$$\label{eq4}
r = \frac{1}{2}r_0 \left( {1 + \sqrt {1 + \frac{3}{2}\frac{(1 - \nu )\varphi
Nk_B T}{Ehr_0^2 }} } \right).$$
To relate the number of counterions to the concentration of the polymer in the solution we need to know the degree of dissociation, i.e. the number of counterions per monomer. For a strong polyelectrolyte this number could be taken as 1, i.e. one counterion per monomer. Then
$$\label{eq5}
N = N_A cV_0 = \frac{4}{3}\pi r_0^3 cN_A ,$$
where $c$ is the concentration of the polymer solution in the capsule before it swells, $N_A $ is the Avogadro number. Substituting Eq. (\[eq5\]) into Eq. (\[eq4\]) one obtains
$$\label{eq6}
r = \frac{1}{2}r_0 \left( {1 + \sqrt {1 + \frac{2\pi r_0 }{h}\frac{(1 - \nu
)}{E}\varphi cRT} } \right),$$
where $R = k_B N_A $ is the universal gas constant. Equation (\[eq6\]) relates the size of the swollen capsule to the concentration of the inner solution, thickness of the capsule shell, and Young’s modulus.
Our model allows design of a swelling experiment. To determine the Young’s modulus, one can measure the deformation of the capsule as a function of concentration $c$ and the shell thickness $h$ and then fit the experimental data to Eq. (\[eq6\]).
Experimental
============
For an initial application of our approach we have chosen to study two types of capsules, characterized by different methods of encapsulation and templated on different cores. The first type are “filled” capsules prepared on manganese carbonate templates by a controlled precipitation [@8; @17], i.e. by an assembly of the inner layer of polyelectrolyte shell by means of multivalent ions with the subsequent extraction of these ions and polymer release into the capsule interior. The second type of “filled” capsules were made from the “hollow” ones, templated on melamine formaldehyde particles, by regulating their permeability for high molecular weight polymers [@9; @18].
As a polyelectrolyte for encapsulation we have chosen sodium polystyrene sulfonate (PSS). The behavior of this highly charged flexible polyanion in salt-free conditions has been studied both theoretically [@15] and experimentally [@10; @11; @19; @20]. The value of $\varphi
$ for PSS was found to range from 0.2 to 1 and suggests some condensation of counter-ions [@15; @17].
Materials and Methods
---------------------
### Materials
The fluorescent dye Rhodamine B isothiocyanate (RBITC), monomers allylamide and 4-styrenesulfonic acid sodium salt hydrate (SS), shell-forming polyelectrolytes poly(sodium 4-styrenesulfonate) (PSS; $\rm M_{W} \sim
70,000 \rm g/mol$ ) and poly(allylamine hydrochloride) (PAH; $ \rm M_{W}
\sim 70,000 \rm g/mol$), and ethylenediaminetetraacetic acid (EDTA) were purchased from Sigma-Aldrich Chemie GmbH, Germany. Hydrochloric acid (HCl), acetone and sodium chloride (NaCl) were purchased from Riedel-de Haën; Germany. The ionic initiator potassium peroxodisulfate (K$_{2}$S$_{2}$O$_{8})$ for the radical polymerisation and Y(NO$_{3})_{3}$ were obtained from Merck GmbH; Germany. All chemicals were of analytical purity or higher quality and were used without further purification.
To produce labelled PSS for encapsulation we used a modification of a method published in Ref. [@21]. First, labelled allylamide was made which was afterwards mixed with 4-styrenesulfonic acid sodium salt hydrate and then copolymerised radically. The allylamide was mixed with RBITC solved in ethanol. The mixture was stirred for four hours at room temperature. Afterwards SS was added in an amount corresponding to a label grade of about 200 monomer units each. Then K$_{2}$S$_{2}$O$_{8}$ was added to this solution as an ionic initiator for the radical polymerisation. The mixture was heated up to $80^0$C and was stirred for four hours in a nitrogen atmosphere. Afterwards it was centrifuged with a membrane which filters molecules with a molecular weight of 100,000 g/mol or more, and polydispersity was found to be M$_{W}$/M$_{N} \sim $1.8. The remaining PSS was chopped into small parts and washed with ethanol until no more colour could be observed in the filtrate.
Suspensions of monodispersed weakly cross-linked melamine formaldehyde particles (MF-particles) with a radius of $r_0 = 2.0\pm 0.1 \quad \mu $m were purchased from Microparticles GmbH (Berlin, Germany).
The manganese carbonate template (MnCO$_{3})$ was prepared by a mixing method described in Ref. [@22]. Briefly, acidic manganese sulfate solution ($9\times 10^{ - 3}$ M, pH=4.2 adjusted by sulfuric acid) was added at 1:1 volume ratio to $2.25\times 10^{ - 3}$ M NH$_{4}$HCO$_{3}$. Then the stirred mixture was aged at $50^{0}$C for 16 hours. The resulting MnCO$_{3}$ particles had a spherical shape with a radius of 1.85$\pm $0.2 $\mu $m and $2.5\pm 0.2\;\mu $m .
Water used for all experiments was purified by a commercial Milli-Q Gradient A10 system containing ion exchange and charcoal stages, and had a resistivity higher than 18M$\Omega $/cm.
### Methods
### Capsule Preparation {#capsule-preparation .unnumbered}
Two different approaches have been exploited for preparation of filled capsules.
![Scheme illustrating the preparation of “filled” microcapsules of the first type. []{data-label="fig1"}](paper1.eps){width="8cm"}
![Scheme illustrating the preparation of “filled” microcapsules of the second type.[]{data-label="fig2"}](paper2.eps){width="8cm"}
The preparation of filled capsules of the first type consisted of several steps (Fig.1). The first step (Fig.1a) was surface controlled precipitation of labeled PSS (by complex formation with Y$^{3 + }$ ions) on the surface of MnCO$_{3}$ [@8; @23; @24]. By varying the number of precipitated layers we were able to tune the surface density for adsorbed layers, and, therefore, the number of PSS molecules precipitated on MnCO$_{3
}$ particles. In our experiments we prepared samples with 20-80 layers of adsorbed labeled PSS, which should lead to a surface density, $\rho $, in the range from $5\times 10^{ - 5}$ to $2\times 10^{ - 4}$mol/m$^{2}$. These values were calculated by assuming that the surface density for a monolayer is constant, and that the amount of adsorbed PSS grows linearly with the number of deposited layers [@17; @24]. As a result, each capsule contained up to a few pg of PSS, depending on the template size and amount of assembled layers. With such a method, the concentration of encapsulated polyelectrolyte is approximately equal to $c = 3\rho / r_0 $. The MnCO$_{3}$ particles were dissolved in $ \sim $ 1mol/L HCl after assembling of 7 layers in order to facilitate the process of core removal since shells with a thickness of more then 10 layers prevent ion penetration [@25]. Core dissolution led to the formation of “double shell” structured capsules (Fig.1c). The inner shell formed by the PSS/Y$^{3 + }$ complex was not stable and was decomposed either by metal-ion complex agents (EDTA) or in salt solution. Yttrium ions were gradually expelled out of the outer stable shell formed by PSS/PAH while PSS molecules were released into the capsule interior (Fig.1d). Then the polyelectrolyte capsules were covered additionally with a number of layers varied in the interval from 1 to 13 to tune the final shell thickness, which, as a result, varied from 4 to 10 PSS/PAH bilayers.
The “filled” capsules of the second type were made from pre-formed “hollow” capsules [@26]. The original “hollow” capsules were produced by a standard LbL assembly of 4 PSS/PAH bilayers on MF particles. MF-particles coated with PSS/PAH multilayers were dissolved in $ \sim $ 1mol/L HCl and MF-oligomers were removed by washing, as described in [@27]. Then the encapsulation of polymer included several steps. (Fig.2). The original “hollow” capsules (Fig.2a) were exposed to acetone/water mixture (1:1) to make the polyelectrolyte multilayer permeable for high molecular weight polymer [@9], and the PSS molecules were added to the mixture. The permeable state of the capsule shell allows the polymers to penetrate inside (Fig.2b). During the encapsulation process the PSS concentration was increased gradually to avoid an osmotic collapse of the microcapsules [@5]. The initial concentration was 1g/L, and was doubled every hour. When the required concentration was reached, the mixture was diluted with water and the multilayer shells were assumed to return to an impermeable state (Fig.2c). After washing in pure water the capsules contain polymer solution (Fig.2d). With such a method, the concentration of encapsulated polyelectrolyte (before swelling) is approximately equal to the final concentration in the bulk.
### Confocal Laser Scanning Microscopy.
To scan the capsule shape and to measure the concentration of PSS inside the capsules we used a commercial confocal microscope manufactured by Olympus (Japan) consisting of the confocal laser scanning-unit Olympus FV 300 in combination with an inverted microscope Olympus IX70 equipped with a high resolution 100 ~~~~ oil immersion objective. The excitation wavelength was chosen according to the label Rhodamine (543 nm). The z-position scanning was done in steps of 0.2 $\mu $m. The diameters of the swollen capsules were determined optically with an accuracy of 0.4 $\mu $m. Concentration measurements were performed via the fluorescence intensity coming from the interior of the PSS containing capsules. In this case we assumed that fluorescence is directly proportional to PSS concentration and used a calibration curve of fluorescence intensity of free polymer in the bulk solution. The measured concentration was then recalculated to the initial concentration inside unswollen capsules.
Results and discussion
======================
The 3D confocal scanning showed that the capsules filled with PSS have spherical form. It should be noted that immediately after the preparation the capsule sizes were close to the size of the templates used for their preparation. The capsules swell for at least several days before reaching their equilibrium size, so that all the measurements of the radius of the swollen capsules were performed 2 weeks after filling with PSS. Fig.3 (top) shows a typical confocal fluorescence image of the swollen capsules. The fluorescence intensity suggests a uniform concentration in the capsule’s interior. The bright interior of the capsules did not change with time, and there was no fluorescence signal from water. This proves that the capsules are in the impermeable state. Typical fluorescence intensity profiles along the diameter of the capsules are presented in Fig.3 (bottom). We note that the level of fluorescence from the wall is higher than from the interior, which could be connected with some adsorption of the inner polyelectrolyte [@17].
![Confocal images of polyelectrolyte microcapsules filled with PSS (top) and typical fluorescence profile along the diameter of the capsule (bottom)[]{data-label="fig3"}](paper3.eps){width="6cm"}
![Confocal images of polyelectrolyte microcapsules filled with PSS (top) and typical fluorescence profile along the diameter of the capsule (bottom)[]{data-label="fig3"}](paper4.eps){width="8cm"}
The size of the swollen capsules was determined as an average of 6-10 capsules. The variability in size of the similarly prepared capsules was always within the error of optical measurements. We found that the “filled” capsules are always larger than the original colloidal template, and that their radius depends on the size of the original template, the number of the bilayers in the shell, and the amount of encapsulated polyelectrolyte. These observations are consistent with our theoretical model.
Fig.4 illustrates the typical dependence of equilibrium radius on the shell thickness. Here, the results for the first type of “filled” capsules (MnCO$_{3 }$ template) filled with PSS solutions of different concentrations, and made on the particles of the same size $r_0 = 1.85\pm 0.2 \quad \mu $m are given. It was previously found [@3; @28] that the thickness of a PSS/PAH bilayer is in the range of 3-5 nm. Here to evaluate the thickness of the shell with a known number of PSS/PAH bilayers we use the average value of 4 nm, as before [@5; @7]. The radius of the swollen capsules decreases with the shell thickness and is larger for the capsules with higher concentration of the inner polyelectrolyte solution. We fitted these experimental results to Eq. (\[eq6\]) taking the combination of Young’s modulus and the osmotic coefficient $E / \varphi $ as a fitting parameter and obtained the value of $E / \varphi \sim 200$ MPa for both curves presented in Fig.4. One can see that the predictions of the model are indeed confirmed by experiment, and the continuum mechanics approach is applicable for a molecularly thin multilayer film.
![Radius of the swollen capsule made on the MnCO$_{3}$ template of radius 1.85$\mu $m as a function of the shell thickness. Two concentrations are shown: $0.16$ mol/L (squares) and $0.32$mol/L (circles). Fitting (solid curves) corresponds to $E / \varphi \sim 200\;\mbox{MPa}$ []{data-label="fig5"}](paper5.eps){width="8cm"}
![Radius of the swollen capsule as a function of monomer concentration in the inner solution. Fitting (solid curves) $E / \varphi \sim 300$ MPa (MnCO$_{3 }$template, circles) and $E / \varphi \sim 200$ MPa (MF template, squares)[]{data-label="fig6"}](paper6.eps){width="8cm"}
The dependence of the equilibrium radius on the concentration of the inner polyelectrolyte, both the fitting curves and the experimental data, are shown in Fig.5. Here we present data both for the capsules of the first (MnCO$_{3 }$ template) and the second (MF template) types. The capsules are made on the templates of different size $r_0 = 2.5\pm 0.2\mu $m and $r_0 =
2.0\pm 0.1 \quad \mu $m, correspondingly. The shell was always composed by four PSS/PAH bilayers. From the fit of experimental data we have obtained $E /
\varphi \sim 300$ MPa for the first type, and $E / \varphi \sim 200$ MPa for the second type of capsules.
One can conclude that, taking into account realistic values of the osmotic coefficients [@15; @17] for PSS solutions, Young’s modulus $E$ found in our swelling experiment is of the order of 100 Mpa. This value is confined between the values found in the osmotic buckling experiment [@5] and recent AFM results [@7] and requires further comments.
We remark and stress that current study gives the possible value of Young’s modulus. In reality, the excess osmotic pressure could be smaller than estimated from the known concentration of polyelectrolyte chains. One reason for such a decrease could be connected with a partial dissociation of the PSS, while our model assumed that it is fully dissociated. One also cannot exclude that some of the counter-ions might condense on the polyelectrolyte shell. Besides that, a portion of counter-ions could escape from the inner polyelectrolyte solution to the outer solvent, reducing the osmotic pressure difference. Such a possibility was not included in our model. All these effects (also ignored in Ref. [@5]) will effectively lead to smaller values of the Young’s modulus.
Thus, our results strongly support the results of previous AFM studies [@7] implying that we are dealing with an *elastomer* [@6]. In other words, the value of Young’s modulus of the polyelectrolyte multilayer falls in the range characteristic for cross-linked rubbers. Such a mechanical behavior is probably due to strong interaction between polyanions and polycations in the multilayer. Our model also shows that the final permeability of the shell to water is very important for the stabilization of the microcapsules, increasing the threshold of the buckling transition [@5].
We also note that the current method includes only a few assumptions, compared to the buckling transition and AFM measurements. We assume that the capsule deformations are small and elastic. These assumptions are also present in the theory of buckling and in the model used to describe capsules deformed in the AFM experiment. The theory of buckling transition assumes that the capsule is highly permeable to water even on a short time scale of the buckling deformation. The AFM experiment-based model postulates a priori a spherical shape of the capsule (except in the contact regions), conservation of capsule volume, and neglects bending deformations. There is no need for these assumptions in describing the capsule swelling, since the capsule shape is always spherical and the process of swelling is a slow process, i.e. the diffusion of water through the capsule shell takes place and does not affect the capsule swelling.
Conclusion
==========
We have provided a theoretical model, which relates Young’s modulus, shell thickness, capsule radius and concentration of the inner polyelectrolyte solution. The validity of the model was experimentally verified, and confirms the applicability of the macroscopic continuum mechanics approach to polyelectrolyte multilayer microcapsules. Young’s modulus of the molecularly thin polyelectrolyte multilayer was found to be of the order of 100 MPa. This represents the possible value for Young’s modulus and is close to the elasticity of an *elastomer*. Such a value reflects a high degree of local interactions between polyanions and polycations. Hence, mechanically at small deformations the polyelectrolyte multilayer resembles a cross-linked rubber material.
Acknowledgements
================
We acknowledge the support of the Alexander von Humboldt Foundation through a research fellowship (DA) and a Sofja Kovalevskaja award (GBS). We are grateful to I.L.Radtchenko for help with polyelectrolyte encapsulation, and to Y.Fedutik for MnCO$_{3}$ synthesis. M.Deserno, C.Holm, and V.Khrenov are thanked for helpful discussions. We also thank R.G.Horn for valuable remarks on the manuscript.
[33]{}
Decher, G. *Science* **1997**, $277$, 1232-1237.
Yoo, D.; Shiratori, S.; Rubner, M. *Macromolecules* **1998**, $31$, 4309-4318.
Sukhorukov, G. B.; Donath, E.; Lichtenfeld , H.; Knippel, E.; Budde, A.; Mohwald, H. *Colloids Surfaces A* **1998**, $137$, 253-266.
Donath, E.; Sukhorukov, G. B.; Caruso, F.; Davis, S.; Möhwald, H. *Angew. Chem.* **1998**, $37$, 2202-2205.
Gao, C.; Donath, E.; Moya, S.; Dudnik, V.; Mohwald, H. *Eur. Phys. J. E* **2001**, $5$, 21-27.
Shackelford, J. F.; William, A.; Juns, P. *Materials Science and Engineering Handbook*, 2nd ed.; CRC press, 1994.
Lulevich, V. V.; Andrienko, D.; Vinogradova, O. I., submitted (cond-mat/0306737).
Lulevich, V. V.; Radtchenko, I. L.; Sukhorukov, G. B.; Vinogradova, O. I. *J. Phys. Chem. B* **2003**, $107$, 2735-2740.
Lulevich, V. V.; Radtchenko, I. L.; Sukhorukov, G. B.; Vinogradova, O. I. *Macromolecules* **2003**, $36$, 2832-2837.
Takahashi, A.; Kato, N.; Nagasawa, M. *J. Phys. Chem.* **1970**, $74$, 944-946.
Wang, L.; Bloomfield, V. A. *Macromolecules* **1990**, $23$, 804-809.
Barrat, J. L.; Joanny, J. F. *Adv. Chem. Phys.* **1996**, $94$, 1-66.
Landau, L. D.; Lifshitz. *Theory of elasticity. Course of theoretical physics*; Butterworth-Heinemann: Oxford, 1995; Vol. 7.
Stevens, M. J.; Kremer, K. *J. Chem. Phys.* **1995**, $103$, 1669-1690.
Micka, U.; Holm, C.; Kremer, K. *Langmuir* **1999**, $15$, 4033-4044.
Liao, Q.; Dobrynin, A. V.; Rubinstein, M. *Macromolecules* **2003**, $36$, 3399-3410.
Radtchenko, I. L.; Sukhorukov, G. B.; Leporatti, S.; Khomutov, G. B.; Donath, E.; Mohwald, H. *J. Colloid Interface Sci.* **2000**, $230$, 272-280.
Sukhorukov, G. B.; Antipov, A. A.; Voigt, A.; Donath, E.; Möhwald, H. *Macromol. Rapid Commun.* **2001**, $22$, 44-46.
Essafi, W.; Lafuma, F.; Williams, C. E. *J. Phys. II France* **1995**, $5$, 1267-1275.
Essafi, W.; Lafuma, F.; Williams, C. E. *Eur. Phys. J. B* **1999**, $9$, 261-266.
Dähne, L.; Leporatti, S.; Donath, E.; Möhwald, H. *J. Am. Chem. Soc.* **2001**, $123$, 5431-5436.
Gaponik, N.; Radtchenko, I. L.; Gerstenberger, M. R.; Fedutik, Y. A.; Sukhorukov, G. B.; Rogach, A. L. *Nano Letters* **2003**, $3$, 369-371.
Dudnik, V.; Sukhorukov, G. B.; Radtchenko, I. L.; Möhwald, H. *Macromolecules* **2001**, $34$, 2329-2334.
Radtchenko, I. L.; Sukhorukov, G. B.; Möhwald, H. *Colloids Surfaces A* **2002**, $202$, 127-133.
Petrov, A. I.; Gavryushkin, A. V.; Sukhorukov, G. B. *J. Phys. Chem. B* **2003**, $107$, 868-875.
Lvov, Y.; Antipov, A. A.; Mamedov, A.; Möhwald, H.; Sukhorukov, G. B. *Nano Letters* **2001**, $1$, 125-128.
Sukhorukov, G. B.; Donath, E.; Davis, S.; Lichtenfeld, H.; Caruso, F.; Popov, V. I.; Möhwald, H. *Polym. Adv. Tech.* **1998**, $9$, 759-767.
Estrela-Lopis, I.; Leporatti, S.; Moya, S.; Brandt, A.; Donath, E.; Mohwald, H. *Langmuir* **2002**, $18$, 7861-7866.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract:
- |
We develop a distribution regression model under endogenous sample selection. This model is a semiparametric generalization of the Heckman selection model that accommodates much richer patterns of heterogeneity in the selection process and effect of the covariates. The model applies to continuous, discrete and mixed outcomes. We study the identification of the model, and develop a computationally attractive two-step method to estimate the model parameters, where the first step is a probit regression for the selection equation and the second step consists of multiple distribution regressions with selection corrections for the outcome equation. We construct estimators of functionals of interest such as actual and counterfactual distributions of latent and observed outcomes via plug-in rule. We derive functional central limit theorems for all the estimators and show the validity of multiplier bootstrap to carry out functional inference. We apply the methods to wage decompositions in the UK using new data. Here we decompose the difference between the male and female wage distributions into four effects: composition, wage structure, selection structure and selection sorting. After controlling for endogenous employment selection, we still find substantial gender wage gap – ranging from 21% to 40% throughout the (latent) offered wage distribution that is not explained by observable labor market characteristics. We also uncover positive sorting for single men and negative sorting for married women that accounts for a substantive fraction of the gender wage gap at the top of the distribution. These findings can be interpreted as evidence of assortative matching in the marriage market and glass-ceiling in the labor market.
[: Sample selection, distribution regression, quantile, heterogeneity, uniform inference, gender wage gap, assortative matching, glass ceiling]{}
- 'The supplementary material contains additional empirical results. These results include the estimates of the coefficients of the employment (selection) equation, estimates and 95% confidence intervals for the decomposition of the employment rate, estimates and 95% confidence bands for some coefficients of the wage equations not reported in the main text, estimates and 95% confidence bands for the components of the wage decomposition of observed wages in the specification (1), and all the estimates and 95% confidence bands for the offered and observed wages and their decompositions for the specifications (2)–(4).'
author:
- Victor Chernozhukov
- 'Iván Fernández-Val'
- Siyi Luo
bibliography:
- '../bibliography.bib'
title: 'Distribution Regression with Sample Selection, with an Application to Wage Decompositions in the UK '
---
introduction {#sec:intro}
============
Sample selection is ubiquitous in empirical economics. For example, it arises naturally in the estimation of wage equations because we do not observe wages of individuals who do not work [@gronau74; @heckman74], and product demands because we do not observed quantities purchased by consumers who do not have access to the product. Sample selection biases the estimation of causal or predictive effects when the reasons for not observing the data are related to the outcome of interest. For example, there is sample selection bias in the estimation of a wage equation whenever the employment status and offered wage depend on common unobserved variables such as ability, motivation or skills. The most popular solution to the sample selection bias is the Heckman selection model (HSM) introduced in [heckman74]{}. This classical model offers a convenient and parsimonious way to account for sample selection by making parametric assumptions about the outcome and selection processes. Our development is motivated by the observation that, in addition to the parametric structure, the classical model imposes strong homogeneity assumptions on how covariates affect the outcome and selection processes and how the selection process operates itself. We develop a generalization of the HSM that relaxes all these three homogeneity restrictions. The resulting model is a semiparametric model, where key parameters are function-valued, thereby considerably generalizing the classical selection model.
Following the literature, we model sample selection using two latent variables for the selection and outcome processes and relate the distribution of these variables with the distribution of the corresponding observed variables. Here we find convenient to work with a local Gaussian representation (LGR) of the joint distribution of the latent variables, which we introduce in the paper. This representation is unique for any joint distribution and might be of independent interest in other settings. The identification analysis is very transparent with the LGR. Thus, we show that the parameters of the LGR are partially identified in the presence of endogenous sample selection because there are only two free probabilities to identify three parameters. We rely on exclusion restrictions to point-identify the three parameters nonparametrically. These conditions require of a binary covariate that does not affect the distribution of the latent outcome and dependence between the latent selection and outcome variables.
Once we have established nonparametric identification with the exclusion restrictions, we introduce a flexible semiparametric distribution regression (DR) model with covariates for the LGR. This model generalizes the HSM by adding multiple sources of heterogeneity to the selection and outcome processes. Thus, it allows for observed and unobserved heterogeneity in selection sorting, together with unobserved heterogeneity in the effect of the covariates on the selection sorting and outcome. In the case of the wage equation, the model can capture the presence of heterogenous returns to schooling across the wage distribution, or positive sorting at the top of the wage distribution and negative sorting at the bottom. The model is semiparametric because its parameters are function-valued and can be applied without modification to continuous, discrete and continuous-discrete outcomes. We show how to construct interesting functionals of the model parameters such as actual and counterfactual distributions of latent and observed outcomes, which can be applied to policy evaluation, treatment effects, wage decompositions and discrimination analysis accounting for sample selection. In the case of wage decompositions, we show how to identify two new effects: a selection sorting effect and a selection structure effect. Selection sorting is determined by whether the employed individuals have higher or lower offered or latent wages than unemployed individuals with the same characteristics. Selection structure is determined by the proportion of employed individuals and how they are selected based on observed characteristics.
We develop a two-step estimator for the model parameters. The first step consists of a probit regression for the selection equation, which is identical to the first step in the Heckman two-step method [@heckman79]. The second step estimates multiple DRs with sample selection correction. The difference between these DRs and the standard DRs without sample selection is that we run bivariate probits instead of univariate probits [@foresiperacchi95; @chernozhukov+13inference]. We estimate functionals of the parameters using the plug-in method. We derive functional central limit theorems for all the estimators and show how to use these results to perform uniform inference on function-valued parameters. This type of inference is useful to construct confidence bands and test hypotheses such as whether a coefficient or effect is uniformly zero, constant or positive. We implement the inference methods using Kolmogorov-Smirnov type statistics where the critical values are obtained via multiplier bootstrap [@gz84] applied to the estimated influence functions (as in [ch06]{}, [ks12]{} and [cck13]{}). This bootstrap scheme is convenient in our setting because it avoids repeated computation of estimators in constructing the bootstrap draws of the statistic. We prove the validity of multiplier bootstrap by deriving bootstrap functional central limit theorems for all the estimators.
We apply our methods to study the relationship between wage and employment in the U.K. using updated data from 1978 to 2013. To this end we estimate wage equations for men and women and carry out several wage decompositions accounting for endogenous selection into employment. Here, we uncover positive sorting among single men and negative sorting among married women. This difference in selection sorting is consistent with assortative matching in the marriage market. It also explains a substantive proportion (but not all) of the gender wage gap at the top of the distribution, which is consistent with recent explanations based on glass ceiling theory. We still find the strong evidence of the gender wage gap – most of the gender wage gap in offered wages (as well as in observed wages) is accounted by differences in the wage structure that are often associated with gender discrimination in the labor market. The effect of education is positive and increases along the distribution. All the heterogeneity that we find is inconsistent with the restrictions of the classical selection model, lending support for the use of the generalized HSM.
#### **Literature review**
The sample selection problem has a long history in statistics and econometrics. Classical references can be found in [gronau74]{}, [heckman74]{}, [lee82]{}, [goldberger83]{}, \[Section 10.7\][amemiya85]{}, \[Section 9.4\][maddala86]{}, [manski89]{}, [manski94]{}, and [vella98]{}. A popular solution to the problem is the HSM developed by Heckman in a sequence of papers [@heckman74; @heckman76; @heckman79; @heckman90]. This model has been extended in several dimensions. [lee83]{}, [prieger02]{} and [smith03]{} replaced the bivariate standard normal copula with other parametric copulas, and [marchenko12]{} replaced the bivariate normal by a bivariate t-distribution to apply the HSM to heavy tailed data. [ahn93]{}, [powell94]{}, [andrews98]{}, and [newey99]{} developed semiparametric versions of the HSM and [das03]{} a nonparametric version, all focusing on location effect versions with homogeneous effects. None of the models considered in these extensions accommodates all the sources of heterogeneity allowed by our model.
[ab17]{} proposed another extension of the HSM, which like our model allows for multiple sources of heterogeneity.[^1] Their method relies on quantile regression to model the marginal distribution of the latent outcome coupled with a parametric model for the copula of the latent selection and outcome variables. They estimate the model parameters using a three-step method where the first step is the same as in our method, but the second and third steps involve an iterative procedure that alternates between quantile regressions to estimate the outcome equation and nonlinear GMM to estimate the parameters of the copula. They also rely on numerical simulation to estimate functionals of the parameters such as actual and counterfactual distributions of the latent and observed outcomes. Compared to our method, they model the covariates effects as direct on the conditional quantile of the latent distribution, whereas we model the covariate effects as direct on the conditional latent distribution – hence in our framework covariates affect the conditional quantiles indirectly. Further, their modeling approach imposes homogeneity on the copula function, which rules out forms of copula heterogeneity across the distribution of the latent outcome, which are permitted in our approach. Moreover, their quantile regression model requires the latent outcome to be continuous, whereas our distribution regression model can deal with any type of outcome and is therefore more widely applicable. Our method is computationally simpler as it does not involve any iteration between methods in the second step, nor does rely on numerical simulation to estimate functionals. The identification assumptions are also different and not nested: we impose more structure on the dependence between the outcome and selection processes, whereas they require more variation on the excluded covariates. We provide a more detailed comparison of the identifying assumptions in Appendix \[app:ab17\].[^2]
#### **Outline**
Section \[sec:id\] examines the identification problem under sample selection using a new representation of a joint distribution. Section \[sec:model\] introduces the DR model with selection and associated functionals, estimators of the model parameters and functionals, and a multiplier bootstrap method to perform functional inference. Section \[sec:theory\] provides asymptotic theory for the estimation and inference methods, and Section \[sec:empirics\] reports the results of the empirical application. Appendices \[app:id\]–\[app:scores\] give the proofs of the main results and additional technical results. The online Supplemental Material (SM) contains additional empirical results.
Another View of the Sample Selection Problem {#sec:id}
============================================
Local Gaussian Representation of a Joint Distribution
-----------------------------------------------------
We start by characterizing a local Gaussian representation (LGR) of the joint distribution of two random variables that is convenient to provide a new view of the identification problem with sample selection and motivate our modeling choices later.
Let $Y^*$ and $D^*$ be two random variables with joint cumulative distribution function (CDF) $F_{Y^*,D^*}$ and marginal CDFs $F_{Y^*}$ and $F_{D^*}$. We label these variables with asterisks because they will be latent variables when we introduce sample selection. Our first result shows that $F_{Y^*,D^*}$ can be represented via a standard bivariate normal distribution at a point and with a correlation parameter that depend on the evaluation point $(y,d)$.
\[lemma:lgr\] Let $F_{Y^*,D^*}$ be a joint CDF, then, for any $(y,d) \in \mathbb{R}^2$, $$F_{Y^*,D^*}(y,d) = \Phi_2(\mu(y), \nu(d); \rho(y,d)),$$ where $\mu(y) \in \mathbb{R}$, $\nu(d) \in \mathbb{R}$, $\rho(y,d) \in [-1,1]$, and $\Phi_2(\cdot, \cdot; \rho)$ is the joint CDF of a standard bivariate normal random variable with parameter $\rho$. Moreover, the values of $\mu(y)$, $\nu(d)$ and $\rho(y,d)$ are uniquely determined by $\mu(y) = \Phi^{-1}(F_{Y^*}(y))$, $\nu(d) = \Phi^{-1}(F_{D^*}(d))$, and the solution in $\rho$ of $$F_{Y^*,D^*}(y,d) = \Phi_2(\Phi^{-1}(F_{Y^*}(y)), \Phi^{-1}(F_{D^*}(d)); \rho),$$ where $\Phi$ is the standard normal CDF. Hence, the representation is unique.
Lemma \[lemma:lgr\] establishes that any joint CDF admits a unique representation as a sequence of standard bivariate normal distributions. This result is stronger than the comprehensive property of the Gaussian copula that establishes that this copula includes the two Frechet bounds and independent copula by suitable choice of the correlation parameter, e.g., [smith03]{}. Lemma \[lemma:lgr\] easily extends to CDFs conditional on covariates $Z$ by making all the parameters dependent on the value of $Z$.
The parameter $\rho(y,d)$ can be interpreted as a measure of local dependence.[^3] Thus, when $\rho(y,d) = 0$, the distribution $F_{Y^*,D^*}$ factorizes at $(y,d)$: $$F_{Y^*,D^*}(y,d) = \Phi_2(\Phi^{-1}(F_{Y^*}(y)), \Phi^{-1}(F_{D^*}(d)); 0) = F_{Y^*}(y) F_{D^*}(d),$$ that is, the events $\{Y^*{\leqslant}y\}$ and $\{D^*{\leqslant}d\}$ are independent. Hence we can say that $Y^*$ and $D^*$ are “locally independent" at $(y,d)$.[^4] In general, the discrepancy $$|\Phi_2(\Phi^{-1}(F_{Y^*}(y)), \Phi^{-1}(F_{D^*}(d)); \rho(y,d))- \Phi_2(\Phi^{-1}(F_{Y^*}(y)), \Phi^{-1}(F_{D^*}(d)); 0)|$$ measures deviation away from independent factorization, thereby giving meaning to $\rho(y,d)$ as local dependence parameter.
In the LGR, the marginal CDFs of $Y^*$ and $D^*$ are represented by local Gaussian links $$F_{Y^*}(y) = \Phi(\mu(y)), \quad F_{D^*}(d) = \Phi(\nu(d)),$$ and the copula of $Y^*$ and $D^*$ is represented by a local Gaussian copula $$\label{LGC}
\begin{array}{l}
C_{Y^*,D^*}(u,v) = \Phi_2(\Phi^{-1}(u), \Phi^{-1}(v); \rho(y_u,d_v)), \\
\forall (u,v) \in [0,1] : \exists y_u \in \mathbb{R}: F_{Y^*}(y_u) = u, \quad \exists d_v \in \mathbb{R}: F_{D^*}(d_v) = v.
\end{array}$$ The LGR is convenient because it separates $\mu(y)$ and $\nu(d)$ as two parameters determining the marginals of $Y^*$ and $D^*$ from $\rho(y,d)$ as a parameter determining the dependence between $Y^*$ and $D^*$.[^5]
[kolev06]{} developed a closely related result to (\[LGC\]) for the copula. They established that the copula of any bivariate distribution can be represented by the bivariate Gaussian copula with a local correlation parameter. The LGR is different from the local Gaussian approximation of [tjostheim13]{}, which approximates the distribution of a continuous bivariate random variable in a neighborhood of the point of interest by a bivariate normal distribution with local vector of means and variance-covariance matrix, see also [hjort96]{}. As [tjostheim13]{} noticed, unlike the LGR, a local Gaussian approximation that intersects with the joint distribution at the point of interest is not unique.
Identification of Sample Selection Model
----------------------------------------
We consider now the sample selection problem where we observe two random variables $D$ and $Y$, which can be defined in terms of the latent variables $D^*$ and $Y^*$ as $$\begin{aligned}
D &=& 1(D^* {\leqslant}0),\\
Y &=& Y^* \ \ \text{if} \ \ D=1,
\end{aligned}$$ i.e., $D$ is an indicator for $D^* {\leqslant}0$ and $Y^*$ is only observed when $D = 1$. The goal is to identify features of the joint distribution of the latent variables from the joint distribution of the observed variables.
The joint CDF of the observed variables can be written in terms of the LGR of $F_{Y^*,D^*}$ as $$F_{Y,D}(y,d) = \Phi_2(\mu(y), \nu(0); \rho(y,0))1(d {\geqslant}1) + [1 - \Phi(\nu(0))]1(d {\geqslant}0).$$ As shown below, the parameters of the LGR are partially identified because $F_{Y^*,D^*}$ is only partially identified. We proceed by characterizing the identified set for these parameters and provide exclusion restrictions to achieve point identification. Since there is a one-to-one relationship between $F_{Y^*,D^*}$ and its LGR, the identified set for the parameters of the LGR determine the identified set for $F_{Y^*,D^*}$. In what follows, we simplify the notation to $$\nu := \nu(0), \quad \rho(y) := \rho(y,0).$$ We can only hope to identify $\nu(d)$ and $\rho(y,d)$ at $d=0$ since we only observe whether $D^* {\leqslant}0$.
To understand the source of the partial identification in terms of the LGR, note that in the presence of sample selection there are two free probabilities, $
{{\mathrm{P}}}(D = 1)$ and ${{\mathrm{P}}}(Y {\leqslant}y \mid D = 1)$, to identify three parameters, $\mu(y)$, $\nu$ and $\rho(y)$. The parameter $\nu$ is pinned down by the selection probability as $$\nu = \Phi^{-1}({{\mathrm{P}}}(D = 1)).$$ The parameters $\mu(y)$ and $\rho(y)$ are partially identified as the solutions in $(\mu,\rho)$ to $$F_{Y,D}(y,1) - {{\mathrm{P}}}(D = 0) = \Phi_2(\mu,\Phi^{-1}({{\mathrm{P}}}(D = 1)); \rho).$$ These solutions form a one-dimensional manifold in $\mathbb{R} \times [-1,1]$ because $
\partial \Phi_2(\mu, \cdot; \rho)/\partial \mu > 0
$, $
\partial \Phi_2(\cdot, \cdot; \rho)/\partial \rho > 0
$, and $
\partial^2 \Phi_2(\cdot, \cdot; \rho)/\partial \mu \partial \rho > 0
$ [@spivak65; @munkres91]. The identified set of $(\mu(y),\rho(y))$ can be shrunk using additional information such as that $\rho(y)$ is known to lie in a subinterval of $[-1,1]$, e.g. $|\rho(y)| < 0.2$.
We use exclusion restrictions to achieve point identification of the parameters of the LGR. To state these restrictions in terms of the LGR, we start by extending the LGR to CDFs conditional on covariates. Let $Z$ be a random variable and $F_{Y^*,D^* \mid Z}$ be the joint CDF of $Y^*$ and $D^*$ conditional on $Z$. Then, $F_{Y^*,D^* \mid Z}$ admits the LGR: $$F_{Y^*,D^* \mid Z}(y, d \mid z) = \Phi_2(\mu(y \mid z), \nu(d \mid z); \rho(y,d \mid z)),$$ where $\mu(y \mid z) \in \mathbb{R}$, $\nu(d \mid z) \in \mathbb{R}$, and $\rho(y,d \mid z) \in [-1,1]$. This representation can be characterized using the same argument as in Lemma \[lemma:lgr\] after fixing the value of the covariate $Z$ and letting all the parameters of the LGR to depend on this value. The exclusion restrictions are:
\[ass:er\] There is a binary random variable $Z$ that satisfies:
1. Non-Degeneracy: $0 < {{\mathrm{P}}}(D = 1) < 1$ and $0 < {{\mathrm{P}}}(Z=1 \mid D = 1) <1$.
2. Relevance: $
{{\mathrm{P}}}(D = 1 \mid Z = 0) < {{\mathrm{P}}}(D = 1 \mid Z = 1) < 1.
$
3. Outcome exclusion: $\mu(y \mid z) = \mu(y)$.
4. Selection exclusion: $\rho(y,0 \mid z) = \rho(y,0)$.
The condition that $Z$ is binary is made to emphasize that our identification strategy does not rely on large variation of $Z$. If $Z$ is not binary we only require that Assumption \[ass:er\] be satisfied for two values of $Z$. Part (1) requires that there is sample selection and that $Z$ has variation in the selected population. It is used to guarantee that all the probabilities employed in the identification analysis are well-defined. Part (2) requires that $Z$ affects the probability of selection and rules out corner cases. The condition ${{\mathrm{P}}}(D = 1 \mid Z = 1) < 1$ precludes identification at infinity (see Remark \[remark:idi\]). The sign of the first inequality can be reversed by relabelling the values of $Z$. Part (3) is a standard exclusion restriction, which is not sufficient for point identification in the presence of sample selection [@manski94; @manski03]. It holds when $Y^*$ is independent of $Z$.[^6] Part (4) is an exclusion restriction in the local dependence between $Y^*$ and $D^*$, which holds if selection sorting is independent of $Z$. We explain this condition more in detail below with an example and compare it with the identification condition in [ab17]{} in Remark \[remark:ab17\].
We can get some intuition about the outcome and selection exclusion restrictions with an example. Consider a heteroskedastic bivariate normal model for the latent variables, i.e., $$(Y^*,D^*) \mid Z = z \sim \mathcal{N}_2\left( \left[\begin{array}{c} \mu_{Y^*}(z) \\ \mu_{D^*}(z)\end{array}\right], \left[\begin{array}{cc} \sigma_{Y^*}(z)^2 & \sigma_{Y^*,D^*}(z) \\ \sigma_{Y^*,D^*}(z) & \sigma_{D^*}(z)^2 \end{array}\right] \right).$$ The outcome exclusion imposes that $$\frac{y-\mu_{Y^*}(z)}{\sigma_{Y^*}(z)} = \mu(y),$$ whereas the selection exclusion imposes that $$\frac{\sigma_{Y^*,D^*}(z)}{\sigma_{Y^*}(z)\sigma_{D^*}(z)} = \rho.$$ If $Y^*$ is independent of $Z$, the outcome exclusion holds and the selection exclusion boils down to the condition that $\sigma_{Y^*,D^*}(z)/\sigma_{D^*}(z)$ does not depend on $z$. In other words, the slopes of the linear regressions of $Y^*$ on $D^*$ conditional on $Z$ are the same when $Z=0$ and $Z=1$.
We now show how the presence of an exclusion restriction helps identify the parameters of the conditional LGR. Under the exclusion restriction the conditional LGR at $d=0$ simplifies to $$\label{eq:clgr}
F_{Y^*,D^* \mid Z}(y, 0 \mid z) = \Phi_2(\mu(y), \nu(z); \rho(y)), \ \ z \in \{0,1\},$$ where $\nu(z) := \nu(0 \mid z)$ and $\rho(y) := \rho(y,0)$. The CDF of the observed variables conditional on $Z$ can be related to this conditional LGR as $$F_{Y,D \mid Z}(y,d \mid z) = \Phi_2(\mu(y), \nu(z); \rho(y)) 1(d {\geqslant}1) + [1 - \Phi(\nu(z))] 1(d {\geqslant}0) , \ \ z \in \{0,1\}.$$ As before, $\nu(z)$ is identified from the conditional selection probability as $$\label{eq:nu}
\nu(z) = \Phi^{-1}\left( {{\mathrm{P}}}(D=1 \mid Z = z)\right), \quad z \in \{0,1\}.$$ Moreover, $\mu(y)$ and $\rho(y)$ are identified as the solution in $(\mu, \rho)$ to $$\label{eq:murho}
F_{Y,D \mid Z}(y,1 \mid z) - {{\mathrm{P}}}(D=0 \mid Z=z) = \Phi_2(\mu, \Phi^{-1}\left( {{\mathrm{P}}}(D=1 \mid Z = z)\right); \rho), \ \ z \in \{0,1\}.$$ This is a nonlinear system of two equations in two unknowns that has unique solution because the Jacobian is a P-matrix for all $\mu \in \mathbb{R}$ and $\rho \in (-1,1)$ by Theorem 4 of [gale65]{}.
The following theorem provides a detailed identification analysis of the parameters of the conditional LGR in . It includes boundary cases where the parameters $\mu(y)$ and $\rho(y)$ can be either point or partially identified.
\[theorem:id\] Assume that Assumption \[ass:er\] holds. Then, $\nu(z)$ is identified by and there are the following cases for the identification of $\mu(y)$ and $\rho(y)$:
1. If $ F_{Y,D \mid Z} (y,1 \mid 1) - {{\mathrm{P}}}(D=0 \mid Z=1) = F_{Y,D \mid Z} (y,1 \mid 0) - {{\mathrm{P}}}(D=0 \mid Z=0)> 0$, $$\rho(y) = 1, \ \ \mu(y) = \Phi^{-1}\left(F_{Y,D \mid Z} (y,1 \mid 1) - {{\mathrm{P}}}(D=0 \mid Z=1) \right).$$
2. If $ F_{Y,D \mid Z} (y,1 \mid 1) < 1$ and $F_{Y,D \mid Z} (y,1 \mid 0) = 1$, $$\rho(y) = 1, \ \ \mu(y) = \Phi^{-1}\left(F_{Y,D \mid Z} (y,1 \mid 1) - {{\mathrm{P}}}(D=0 \mid Z=1) \right).
$$
3. If $ F_{Y,D \mid Z} (y,1 \mid 1) = F_{Y,D \mid Z} (y,1 \mid 0) = 1$, $$\rho(y) = 1, \ \ \mu(y) \in [\Phi^{-1}({{\mathrm{P}}}(D=1 \mid Z=1)), + \infty).$$
4. If $F_{Y,D \mid Z} (y,1 \mid 1) > {{\mathrm{P}}}(D=0 \mid Z=1)$ and $F_{Y,D \mid Z} (y,1 \mid 0) = {{\mathrm{P}}}(D=0 \mid Z=0)$, $$\rho(y) = - 1, \ \mu(y) = \Phi^{-1}\left(F_{Y,D \mid Z} (y,1 \mid 1) \right).$$
5. If $F_{Y,D \mid Z} (y,1 \mid 1) = F_{Y,D \mid Z} (y,1 \mid 0) < 1$, $$\begin{aligned}
& \rho(y) = - 1, \ \mu(y) = \Phi^{-1}\left(F_{Y,D \mid Z} (y,1 \mid 1) \right).\end{aligned}$$
6. If $F_{Y,D \mid Z} (y,1 \mid z) = {{\mathrm{P}}}(D=0 \mid Z=z)$, $z \in \{0,1\}$, $$\rho(y) = - 1, \ \mu(y) \in (-\infty, \Phi^{-1}\left(F_{Y,D \mid Z} (y,1 \mid 1)\right)].$$
7. Otherwise, $\mu(y)$ and $\rho(y)$ are point identified as the solution in $(\mu, \rho)$ to . This solution exists and is unique.
The boundary cases in Theorem \[theorem:id\] are easy to detect. In practice, partial identification usually occurs at extreme values of $y$. For example, case (3) arises for values of $y$ such that $Y > y$ a.s., and case (6) for values of $y$ such that $Y < y$ a.s.
\[remark:idi\] When ${{\mathrm{P}}}(D = 1 \mid Z = 1) = 1$ and $\mu(y \mid z) = \mu(y)$, the conditional LGR at $z=1$ gives $F_{Y,D \mid Z}(y,1 \mid 1) = \lim_{\nu \nearrow + \infty} \Phi_2(\mu(y), \nu; \rho(y \mid 1)) = \Phi(\mu(y))$, which identifies $\mu(y)$ by $$\mu(y) = \Phi^{-1}(F_{Y,D \mid Z}(y,1 \mid 1)),$$ without the selection exclusion restriction. This result is analogous to the identification at infinity of [chamberlain86]{} where $Z$ is continuous with unbounded support and $$\lim_{z \nearrow +\infty} {{\mathrm{P}}}(D = 1 \mid Z = z) = 1.$$ Note that $\rho(y \mid z)$ is not point identified without further restrictions.
\[remark:ab17\] Assumption \[ass:er\] is not nested with the conditions that AB17 used to show nonparametric identification of their model. We impose stronger restrictions in the dependence of the latent selection and outcome variables, but require less variation in the excluded covariate $Z$. We provide a more detailed comparison in Appendix \[app:ab17\]
Distribution Regression Model with Sample Selection {#sec:model}
===================================================
The Model
---------
We consider a semiparametric version of the LGR with covariates: $$\label{eq:drm}
F_{Y^*,D^*}(y,0 \mid Z=z) = \Phi_2(-x^\prime\beta(y), -z^\prime \pi;\rho(x'\delta(y))),$$ where $Y^*$ is the latent outcome of interest, which can be continuous, discrete or mixed continuous-discrete; $D^*$ is a latent variable that determines sample selection; $X$ is a vector of covariates; $Z=(Z_1,X)$; and $Z_1$ are excluded covariates, i.e., observed covariates that satisfy the exclusion restrictions. The excluded covariates avoid reliance on functional form assumptions to achieve identification. The model for the LGR consists of three indexes. We shall refer to $-x^\prime\beta(y)$ as the outcome equation, to $-z^\prime \pi$ as the selection equation, and to $\rho(x'\delta(y))$ as the selection sorting equation. We observe the selection indicator $D = 1(D^* > 0)$ and the outcome $Y=Y^*$ when $D=1$.[^7] In the empirical application that we consider below, $Y^*$ is offered wage, $D^*$ is the difference between offered wage and reservation wage, $D$ is an employment indicator, $Y$ is the observed wage, $X$ includes labor market characteristics such as education, age, number of children and marital status, and $Z_1$ includes measures of out-of-work income. We shall discuss the validity of these measures as excluded covariates in Section \[sec:empirics\].
The model is semiparametric because $y\mapsto \beta(y)$ and $y\mapsto \delta(y)$ are unknown functions, i.e. infinite dimensional parameters in general. This flexibility allows the effect of $X$ on the outcome and selection sorting to vary across the distribution. For example, it allows the return to education to vary across the distribution, the selection sorting to be different for high and low educated individuals, or to have positive selection sorting at the upper tail and negative at the bottom tail or vice versa. The function $u \mapsto \rho(u)$ is a known link with range $[-1,1]$, e.g. the Fisher transformation [@fisher15], $\rho(u) = \tanh(u)$. The corresponding distribution of $Y^*$ conditional on $Z$ is $$F_{Y^*}(y\mid Z=z)= \lim_{\nu \nearrow + \infty} F_{Y^*,D^*}(y,v \mid Z=z)=\Phi(-x^\prime \beta(y)), \ \ z = (x,z_1).$$ The selection bias arises because this distribution is different from the distribution of the observed outcome $Y$, i.e. $$F_{Y^*}(y\mid Z=z)\neq F_{Y}(y\mid Z=z,D=1) = \frac{\Phi_2(-x^\prime\beta(y), z^\prime \pi;-\rho(x'\delta(y)))}{\Phi(z^\prime \pi)}.$$
\[ex:hsm\] Consider the [heckman74]{} sample selection model (HSM): $$\begin{aligned}
D^* &=& Z^\prime\pi + V, \\
Y^* &=& X^\prime \beta+\sigma U, \end{aligned}$$ where $(U,V)$ is independent of $Z$ and has standard bivariate normal distribution with parameter $\rho$, such that $$F_{Y^*,D^*}(y,0 \mid Z=z)
= \Phi_2\left(\frac{y-x^\prime \beta}{\sigma}, -z^\prime \pi; \rho\right).$$ This is a special case of model with $$\beta_1(y)=(\beta_1-y)/\sigma, \ \ \beta_{-1}(y)= \beta_{-1}/\sigma, \ \ \rho(x'\delta(y)) = \rho.$$ The HSM therefore imposes strong homogeneity restrictions in the selection process and effect of the covariates on the outcome and selection sorting. Thus, only the intercept of $\beta(y)$ varies with $y$, and $\rho(x'\delta(y))$ is invariant to both $x$ and $y$.
The model has multiple data generating process representations as nonseparable systems. One example is $$\begin{aligned}
D^* &=& Z^\prime\pi + V, \ \ V\mid Z \sim \mathcal{N}(0,1), \\
0 &=& X^\prime \beta(Y^*) + \rho(X'\delta(Y^*))V+\sqrt{1-\rho(X'\delta(Y^*))^2} U, \ \ U\mid Z \sim \mathcal{N}(0,1),
\end{aligned}$$ where $U$ and $V$ are independent. For example, in the wage application $V$ can be interpreted as unobserved net benefit of working, and $U$ as unobserved skills or innate ability net of $V$. This representation is similar to the HSM in Example \[ex:hsm\] with the difference that the equation for $Y^*$ is nonseparable.[^8]
Functionals
-----------
There are several functionals of the parameters of the model that can be of interest. One is the marginal distribution of the latent outcome $Y^*$ $$F_{Y^*}(y) = \int F_{Y^*}(y\mid Z=z)dF_Z(z)=\int \Phi(-x^\prime \beta(y))dF_X(x),$$ where $F_Z$ and $F_X$ are the marginal distributions of $Z$ and $X$, respectively. In the case of the wage application, $F_{Y^*}$ corresponds to the distribution of the offered wage, which is a potential or latent outcome free of selection. We can also construct counterfactual distributions by combining coefficients $\beta(y)$ and distributions $F_X$ from different populations or groups. These distributions are useful to decompose the distribution of offered wages between females and males or between blacks and whites, which can be used to uncover discrimination in the labor market. Another functional is the probability of selection $${{\mathrm{P}}}(Y = 1) = \int {{\mathrm{P}}}(Y=1 \mid Z=z) dF_Z(z) = \int \Phi(z^\prime \pi)dF_Z(z),$$ which can be use to decompose differences in employment rates between employment structure effects, $\pi$, and composition effects, $F_Z$.
We can also use the model to construct distributions for the observed outcome using that $$\begin{aligned}
F_Y(y)
&=
\int \frac{\Phi_2\left( -x^\prime\beta(y), z^\prime \pi; -\rho(x'\delta(y)) \right)}{\Phi(z^\prime \pi)}dF_Z(z\mid D=1) \\
&= \frac{\int \Phi_2\left( -x^\prime\beta(y), z^\prime \pi; -\rho(x'\delta(y)) \right) dF_Z(z)}{\int \Phi(z^\prime \pi) dF_Z(z)},
\end{aligned}$$ where the second equality follows from the Bayes rule. We can again construct counterfactual distributions by changing $\beta(y)$, $\pi$, $\delta(y)$ and $F_Z$. In the wage application, we will decompose the differences in the wage distribution between genders or across time into changes in the worker composition $F_Z$, wage structure $\beta(y)$, selection structure $\pi$, and selection sorting $\delta(y)$. Both selection effects are new to this model.
\[remark:se\] To interpret the selection effects, it is useful to consider a simplified version of the model without covariates where $ F_Y(y;\pi,\rho)
= \Phi_2\left( -\beta, \pi; -\rho \right)/\Phi (\pi).
$ Here we drop the dependence of $\beta$ and $\rho$ on $y$ to lighten the notation, and make explicit the dependence of $F_Y$ on the selection parameters $\pi$ and $\rho$ to carry out comparative statics with respect to them. Then, by the properties of the normal distribution $$\frac{\partial F_Y(y;\pi,\rho)}{\partial \rho} = - \frac{\phi_2(-\beta,\pi; -\rho)}{ \Phi(\pi)} < 0,$$ and $$\frac{\partial F_Y(y;\pi,\rho)}{\partial \pi} \propto \Phi\left(\frac{-\beta + \rho \pi}{\sqrt{1-\rho^2}} \right) \Phi(\pi) - \int_{-\infty}^{\pi} \Phi\left(\frac{-\beta + \rho x}{\sqrt{1-\rho^2}} \right) \phi(x) dx \left\{\begin{array}{c} < 0 \text{ if } \rho < 0, \\ = 0 \text{ if } \rho = 0, \\ > 0 \text{ if } \rho > 0, \end{array}\right.$$ where $\Phi$ and $\phi$ are the standard normal CDF and probability density function (PDF), and $\phi_2(\cdot, \cdot; \rho)$ be the joint PDF of a standard bivariate normal random variable with parameter $\rho$.[^9] Increasing $\rho$ therefore shifts the distribution to the right (increases quantiles) because it makes selection sorting more positive while the size of the selected population is fixed. The effect of increasing $\pi$ is more nuanced and depends on the sign of $\rho$. Intuitively, $\pi$ affects the size of the selected population and the relative importance of observables and unobservables in the selection. For example, when selection sorting is negative, increasing the size of the selected population by increasing $\pi$ shifts the distribution of the right (increases quantiles) because the newly selected individuals have smaller (more negative) selection unobservables that correspond to larger (more positive) outcome unobservables. In other words, the newly selected individuals are relatively less adversely selected.
The sign of the selection effects might be different in the presence of covariates if the variation in the parameters changes the composition of the selected population. Consider the following extreme example with only one covariate based on the wage application. Let the covariate be an indicator for high skills. Assume that high-skilled workers are relatively more likely to participate than low-skilled workers, there is no selection sorting on unobservables, which corresponds to $\rho(x'\delta(y)) =0$ in the model, and the distribution of offered wages for high-skilled workers first-order stochastically dominates the same distribution for low-skilled workers. In this case increasing the probability of participation for high-skilled workers, which corresponds to increasing the component of $\pi$ associated to the high skill indicator in the model, both increases the overall probability of participation and shifts the distribution of observed wages to the right (increases quantiles), despite the lack of selection sorting. Intuitively, the distribution of observed wages is a mixture of the distribution of wages for employed high-skilled and low-skilled workers, and we are increasing the relative proportion of employed high-skilled workers. The opposite holds if the distribution of offered wages for high-skilled workers is first-order stochastically dominated by the same distribution for low-skilled workers.
Quantiles and other functionals of the distributions of latent and observed outcomes can be constructed by applying the appropriate operator. For example, the $\tau$-quantile of the latent outcome is $Q_{Y^*}(\tau) = \mathbf{Q}_{\tau}(F_{Y^*})$, where $\mathbf{Q}_{\tau}(F) := \inf\{y \in \mathbb{R} : F(y) {\geqslant}\tau\}$ is the quantile or left-inverse operator.
Estimation
----------
To estimate the model parameters and functionals of interest, we assume that we have a random sample of size $n$ from $(D,DY,Z)$, $\{(D_i, D_i Y_i,Z_i)\}_{i=1}^n$, where we use $D Y$ to indicate that we only observe $Y$ when $D=1$.
Before describing the estimators, it is convenient to introduce some notation. Let $\mathcal{Y}$ be the region of interest of $Y$, and denote $\theta_y := (\beta(y), \delta(y))$, where we replace the arguments in $y$ by subscripts to lighten the notation.[^10]
The estimation relies on the relationship between conditional distributions and binary regressions. Thus, the CDF of $Y$ at a point $y$ conditional on $X$ is the expectation that an indicator that $Y$ is less than $y$ conditional on $X$, $$F_{Y \mid X}(y \mid x) = {{\mathrm{E}}}[1(Y {\leqslant}y) \mid X = x].$$ To implement this idea, we construct the set of indicators for the selected observations $$I_{yi} = 1(Y_i{\leqslant}y) \text{ if } D_i = 1,$$ for each $y \in \mathcal{Y}$. In the presence of sample selection, we cannot just run a probit binary regression of $I_{yi}$ on $X_i$ to estimate the parameter $\beta(y)$ as in [foresiperacchi95]{} and [chernozhukov+13inference]{}. The problem is similar to running least squares in the HSM. Instead, we use that $$\begin{aligned}
\ell_i(\pi,\theta_y)=\left[1-\Phi(Z_i^\prime\pi)\right]^{1-D_i}&\times \Phi_2(-X_i^\prime\beta(y),Z_i^\prime\pi;-\rho(X_i'\delta(y)))^{D_iI_{yi}} \\ &\times\Phi_2(X_i^\prime\beta(y),Z_i^\prime\pi;\rho(X_i'\delta(y)))^{D_i(1-I_{yi})}
\end{aligned}$$ is the likelihood of $(D_i, I_{yi})$ conditional on $Z_i$. This likelihood is the same as the likelihood of a bivariate probit model or more precisely a probit model with sample selection [@zellner65; @poirier80; @vandeven81].
We estimate the model parameters using a computationally attractive two-step method to maximize the average log-likelihood, similar to the Heckman two-step method. The first step is a probit regression for the probability of selection to estimate $\pi$, which is identical to the first step in the Heckman two-step method. The second step consists of multiple distribution regressions (DRs) with sample selection corrections to estimate $\beta(y)$ and $\delta(y)$ for each value of $y \in \mathcal{Y}$. These steps are summarized in the following algorithm:
\[alg:tsdr\] (1) Run a probit for the selection equation to estimate $\pi$: $${\widehat}{\pi} = \operatorname*{arg\,max}_{c \in \mathbb{R}^{d_{\pi}} }L_1(c) = \frac{1}{n} \sum_{i=1}^n \left[D_{i}\log{\Phi(Z_i^\prime c)}+(1-D_{i})\log{\Phi(-Z_i^\prime c)}\right], \ \ d_{\pi} := \dim \pi.$$ (2) Run multiple DRs with sample selection correction to estimate $\theta_y$: for each $y \in \mathcal{Y}$ $$\begin{aligned}
{\widehat}{\theta}_y = \operatorname*{arg\,max}_{t=(b,d) \in \Theta} L_2(t, {\widehat}\pi) = \frac{1}{n} \sum_{i=1}^n D_{i} \left[ \right.&I_{yi} \log{\Phi_2\left(-X_i^\prime b,Z_i^\prime {\widehat}{\pi}; -\rho(X_i'd)\right)} \\
+ & \left.(1-I_{yi}) \log{\Phi_2\left(X_i^\prime b,Z_i^\prime {\widehat}{\pi}; \rho(X_i'd) \right)} \right],\end{aligned}$$ where $\Theta \in \mathbb{R}^{d_{\theta}}$ is a compact parameter set, and $$d_{\theta} := \dim \theta_u, \ \ \rho(u) = \tanh(u) = \frac{e^u - e^{-u}}{e^u + e^{-u}} \in [-1,1], \quad \frac{\partial \rho(u)}{\partial u} > 0.$$
In practice we replace the set $\mathcal{Y}$ by a finite grid $\bar{\mathcal{Y}}$ if $\mathcal{Y}$ contains many values. The estimators of the functionals of interested are constructed from the estimators of the parameters using the plug-in method. For example, the estimator of the distribution of the latent outcome is $$\label{eq:ldisthat}
{\widehat}{F}_{Y^*}(y)=\frac{1}{n}\sum_{i=1}^n\Phi(-X_i^\prime {\widehat}{\beta}(y)),$$ the estimator of the probability of selection is $${\widehat}{{{\mathrm{P}}}}(Y = 1) = \frac{1}{n}\sum_{i=1}^n \Phi(z^\prime {\widehat}\pi),$$ and the estimators of the counterfactual distributions of the observed outcome are constructed from $$\label{eq:odisthat}
{\widehat}{F}_Y(y \mid D=1)= \frac{\sum_{i=1}^n \Phi_2(-X_i^\prime {\widehat}{\beta}(y),Z_i^\prime {\widehat}{\pi};-\rho(X_i'{\widehat}{\delta}(y)))}{\sum_{i=1}^n \Phi(Z_i^\prime {\widehat}{\pi})},$$ by choosing the estimators of $ {\widehat}{\beta}(y)$, ${\widehat}{\pi}$, and ${\widehat}{\delta}(y)$ and the sample values of $Z$ appropriately. Estimators of quantiles and other functionals of these distributions are obtained by applying the operators that define the functionals to the estimator of the distribution. For example, the estimator of the $\tau$-quantile of the latent outcome is ${\widehat}Q_{Y^*}(\tau) = \mathbf{Q}_{\tau} ({\widehat}F_{Y^*})$.[^11]
Uniform Inference
-----------------
The model parameters and functionals of interest are generally function-valued. We show how to construct confidence bands for them that can be used to test functional hypotheses such as the entire function be zero, non-negative or constant. To explain the construction consider the case where the functional of interest is a linear combination of the model parameter $\theta_y$, that is the function $y\mapsto c'\theta_y,$ $y\in\mathcal{Y}$, where $c \in \mathbb{R}^{d_{\theta}}$. The set $CB_p(c'\theta_y)$ is an asymptotic $p$-confidence band for $c'\theta_y$ if it satisfies $${{\mathrm{P}}}\left[c'\theta_y \in CB_p(c'\theta_y), \forall y \in\mathcal{Y}\right] \to p, \text{ as } n \to \infty.$$ We form $CB_p(c'\theta_y)$ as $$CB_p(c'\theta_y)=c'{\widehat}{\theta}_y\pm cv(p)SE(c'{\widehat}{\theta}_y),$$ where ${\widehat}{\theta}_y$ is the estimator of $\theta_y$ defined in Algorithm \[alg:tsdr\], $SE(c'{\widehat}{\theta}_y)$ is the standard error of $c'{\widehat}{\theta}_y$, and $cv(p)$ is a critical value, i.e. a consistent estimator of the $p$-quantile of the statistic $$t_{\mathcal{Y}} = \sup_{y\in \mathcal{Y}} \frac{|c'{\widehat}{\theta}_y-c'\theta_y|}{SE(c'{\widehat}{\theta}_y)}.$$ We obtain the standard error and critical value from the limit distribution of ${\widehat}{\theta}_y$ derived in Section \[sec:theory\].
In practice, it is convenient to estimate the critical value using resampling methods. Multiplier bootstrap is computationally attractive in our setting because it does not require parameter re-estimation and therefore avoids the nonlinear optimization in both steps of Algorithm \[alg:tsdr\]. The multiplier bootstrap is implemented using the following algorithm:
\[alg:mb\] (i) For $b \in 1,\ldots, B$ and the finite grid $\bar{\mathcal{Y}} \subseteq \mathcal{Y}$, repeat the steps: (1) Draw the bootstrap multipliers $\{\omega_{i}^b : 1 {\leqslant}i {\leqslant}n \}$ independently from the data and normalized them to have zero mean, $$\omega_i^b = \tilde \omega_i^b - \sum_{i =1}^n \tilde \omega_i^b/n, \ \ \tilde \omega_i^b \sim \text{ i.i.d. } \mathcal{N}(0,1).$$ (2) Obtain the bootstrap estimator of the model parameter $${\widehat}\theta^b_y = {\widehat}\theta_y + n^{-1} \sum_{i = 1}^n \omega_i^b \ \widehat \psi_i({\widehat}\theta_y, {\widehat}\pi),$$ where $\widehat \psi_i({\widehat}\theta_y, {\widehat}\pi)$ is an estimators of the influence function of ${\widehat}{\theta}_y$ given in . (2) Construct bootstrap realization of maximal t-statistic $t_{\mathcal{Y}}$ for the functional of interest, $$t_{\mathcal{Y}}^b = \max_{y \in \bar{\mathcal{Y}}} \frac{|c'{\widehat}\theta^b_y - c'\widehat \theta_y|}{SE(c'{\widehat}{\theta}_y)}, \ \ SE(c'{\widehat}{\theta}_y) = \sqrt{c' \widehat \Sigma_{\theta_y\theta_y} c},$$ where $\widehat \Sigma_{\theta_y\theta_y}$ is an estimator of the asymptotic variance-covariance matrix of ${\widehat}{\theta}_y$ given in . (ii) Compute the critical value $cv(p)$ as the simulation $p$-quantile of $t_{\bar{\mathcal{Y}}}^b$, $$cv(p) = p-\text{quantile of } \{t_{\mathcal{Y}}^b : 1 {\leqslant}b {\leqslant}B\}$$
The centering of the multipliers in step (i1) of the algorithm is a finite sample adjustment. Confidence bands for other functionals of the model parameter can be constructed using a similar bootstrap method.
Asymptotic Theory {#sec:theory}
=================
We derive asymptotic theory for the estimators of the model parameters and functionals of interest.
Limit distributions
-------------------
We first introduce some notation that is useful to state the assumptions that we make to derive the limit distribution of the estimators. Let $\widetilde S_1 := \partial_{\pi} L_1(\pi)$ and $\widetilde S_{2y} := \partial_{\theta_y} L_2(\theta_y, \pi)$ be the scores of the first and second steps in Algorithm \[alg:tsdr\] evaluated at the true parameter values, and $H_1 := {{\mathrm{E}}}\left[ \partial_{\pi \pi'} L_1(\pi) \right]$ and $H_{2y} := {{\mathrm{E}}}\left[ \partial_{\theta_y \theta_y} L_2(\theta_y, \pi) \right]$ be the corresponding expected Hessians. Let $$\label{eq:cf}
\Sigma_{\theta_y\theta_{\tilde y}} := H_{2y}^{-1} \left\{n {{\mathrm{E}}}\left[ \widetilde S_{2y} \widetilde S_{2\tilde y}' \right] - J_{21y} H_1^{-1} J_{21\tilde y}' \right\} H_{2\tilde y}^{-1},$$ where $J_{21y} := {{\mathrm{E}}}\left[ \partial_{\theta_y \pi'} L_2(\theta_y, \pi) \right]$, $d_{\pi} := \dim \pi$, and $d_{\theta} := \dim \theta_y$.
\[ass:fclt\] (1) Random sampling: $\{(D^*_i, Y^*_i,Z_i)\}_{i=1}^n$ is a sequence of independent and identically distributed copies of $(D^*,Y^*,Z)$. We observe $D = 1(D^* > 0)$ and $Y = Y^*$ if $D=1$. (2) Model: the distribution of $(D^*,Y^*)$ conditional on $Z$ follows the DR model . (3) The support of $Z$, $\mathcal{Z}$, is a compact set. (4) The support of $Y$ is either finite or a bounded interval. In the second case, the density function of $Y$ conditional on $X$ and $D=1$, $f_{Y \mid X,D}(y \mid x, 1)$, exists, is uniformly bounded above, and is uniformly continuous in $(y,x)$ on the support of $(Y,X)$ conditional on $D=1$.(5) Identification and non-degeneracy: the equations ${{\mathrm{E}}}[\partial_{\pi} L_1(\tilde \pi)] = 0$ and ${{\mathrm{E}}}[\partial_{\theta_y} L_2(\tilde \theta_y, \tilde \pi)] = 0$ posses a unique solution at $(\tilde \pi, \tilde \theta_y) = (\pi,\theta_y)$ that lies in the interior of a compact set $\Pi \times \Theta \subset \mathbb{R}^{d_{\pi} + d_{\theta}}$ for all $y \in \mathcal{Y}$; and the matrices $H_1$, $H_{2y}$ and $\Sigma_{\theta_y\theta_{y}}$ are nonsingular for each $y \in \mathcal{Y}$.
Part (1) is a standard condition about the sampling and selection process, which is designed for cross sectional data. Part (2) imposes the semiparametric DR model on the LGR of the conditional distribution of $(D^*,Y^*)$ at $d=0$. Part (3) imposes some compactness conditions, which can be generalized at the cost of more complicated proofs. Part (4) covers continuous, discrete and mixed continuous-discrete outcomes. Part (5) imposes directly identification and that the variance-covariance matrix of the first-step estimator and the covariance function of the second-step estimator are well-behaved. Note that $H_1$, $H_{2y}$ and $J_{21y}$ are finite by Part (3). More primitive conditions for part (5) can be found in the conditional maximum likelihood literature, e.g., [newey94]{}.
The main result of this section is a functional central limit theorem for ${\widehat}{\theta}_y$. Let $\ell^{\infty}(\mathcal{Y})$ be the set of bounded functions on $\mathcal{Y}$, and $\leadsto$ denote weak convergence (in distribution).
\[thm:fclt\] Under Assumption \[ass:fclt\], $$\sqrt{n} ({\widehat}\pi - \pi) = -H_1^{-1} \widetilde{S}_1 + o_P(1) \leadsto Z_{\pi} \sim \mathcal{N}(0, -H_1^{-1}), \text{ in } \mathbb{R}^{d_{\pi}}$$ and $$\sqrt{n} ({\widehat}\theta_y - \theta_y)= - H_{2y}^{-1} \sqrt{n} \left( \widetilde S_{2y} - J_{21y} H_1^{-1} \widetilde S_1 \right) + o_P(1) \leadsto Z_{\theta_y} \text{ in } \ell^{\infty}(\mathcal{Y})^{d_{\theta}},$$ where $y \mapsto Z_{\theta_y}$ is a zero-mean Gaussian process with uniformly continuous sample paths and covariance function $
\Sigma_{\theta_y\theta_{\tilde y}}$, $y,\tilde y \in \mathcal{Y}$, defined in .
The first order term in the limit of $\sqrt{n} ({\widehat}\theta_y - \theta_y)$ is the sample average of the influence function of ${\widehat}\theta_y$. We construct an estimator of the covariance function $
\Sigma_{\theta_y\theta_{\tilde y}}$ based on this function. Thus, we form $$\label{eq:se}
\widehat \Sigma_{\theta_y\theta_{\tilde y}} = n^{-2} \sum_{i=1}^{n} \widehat \psi_{i}({\widehat}\theta_y, {\widehat}\pi) \widehat \psi_{i}({\widehat}\theta_{\tilde y}, {\widehat}\pi)'.$$ Here, $\widehat \psi_{i}$ is an estimator of the influence function of ${\widehat}\theta_y$, $$\label{eq:if}
\widehat \psi_{i}(t,c) = - \widehat H_{2y}(t,c)^{-1} \left( S_{2yi}(t,c) - \widehat J_{21y}(t,c) \widehat H_1(c)^{-1} S_{1i}(c) \right),$$ where $S_{1i}(c)$ and $S_{2iy}(t,c)$ are the individual scores of the first and second steps of Algorithm \[alg:tsdr\], $$\begin{aligned}
S_{1i}(c) &:=& \partial_{c} L_{1i}(c), \ L_{1i}(c) := D_{i}\log{\Phi(Z_i^\prime c)}+(1-D_{i})\log{\Phi(-Z_i^\prime c)}, \\ S_{2yi}(t,c) &:=& \partial_{t} L_{2yi}(t,c), \ t=(b,d)\\
L_{2yi}(t,c) &:= & D_{i} \left[ I_{yi} \log{\Phi_2\left(-X_i^\prime b,Z_i^\prime c; -\rho(x'd)\right)}
+ (1-I_{yi}) \log{\Phi_2\left(X_i^\prime b,Z_i^\prime c; \rho(x'd) \right)}\right], \end{aligned}$$ and $$\begin{aligned}
\widehat H_1(c) := \partial_{c c'} L_1(c), \ \widehat H_{2y}(t,c) := \partial_{t t'} L_2(t, c), \ \widehat J_{21y}(t,c) := \partial_{t c'} L_2(t, c),
$$ are estimators of $H_1$, $H_{2y}$, and $J_{21y}$ when evaluated at $c = {\widehat}\pi$ and $t = {\widehat}\theta_y$.
We now establish a functional central limit theorem for the estimators of functionals of the model parameters. This result is based on expressing the functional as a suitable operator of the model parameters and using the functional delta method [@vdV-W Chapter 3.9]. To present the result in a concise manner, we consider a generic functional $$u \mapsto \Delta_u = \varphi_u(\pi, \theta_\cdot, F_Z),$$ where $u \in \mathcal{U}$, a totally bounded metric space, and $\varphi_u$ is an operator that maps $\mathbb{D}_{\Delta}$ to the set $\ell^\infty(\mathcal{U})$, where $\Delta_\cdot$ takes values. Here $\mathbb{D}_{\Delta}$ denotes the space for the parameter tuple $(\pi, \theta_\cdot, F_Z)$; this space is not stated here explicitly, but is restricted by the regularity conditions of the previous section. Here we identify $F_Z$ with an integral operator $f \mapsto \int f(z) d F_Z(z)$ taking values in $\ell^\infty(\mathcal{F})$ that acts on a Donsker set of bounded measurable functions $\mathcal{F}$, which includes indicators of rectangular sets; see [chernozhukov+13inference]{} and examples below. The parameter space $\mathbb{D}_{\Delta}$ is a subset of a normed space $\mathbb{D}:=\mathbb{R}^{d_{\pi}} \times \ell^{\infty}(\mathcal{Y})^{d_{\theta}} \times \ell^{\infty}(\mathcal{F}) $. In this notation, the plug-in estimator of the functional $\Delta_u$ is $${\widehat}\Delta_u = \varphi_u({\widehat}\pi, {\widehat}\theta_y, {\widehat}F_Z),$$ where ${\widehat}\pi$ and ${\widehat}\theta_y$ are the estimators of the parameters defined in Algorithm \[alg:tsdr\] and ${\widehat}F_Z$ is the empirical distribution of $Z$.
We provide some examples. The distribution of the latent outcome is given by: $$F_{Y^*}(y) = \varphi_y(\pi, \theta_y, F_Z) = \int \Phi(-x'\beta_y) dF_Z(z),$$ $\mathcal{F}$ contains $\{\Phi(-\cdot'\beta_y) : y \in \mathcal{Y} \}$ as well as the indicators of all rectangles in $\overline{\mathbb{R}}^{d_z}$, $\overline{\mathbb{R}} := \mathbb{R} \cup \{-\infty, +\infty\}$, $d_z = \dim Z$, and $\mathcal{U} = \mathcal{Y}$. The quantile function of the latent outcome is $$Q_{Y^*}(\tau) = \varphi_{\tau}(\pi, \theta_y, F_Z) = \mathbf{Q}_{\tau} \mathbf{R} F_{Y^*},$$ $\mathcal{F}$ is the same as for the distribution of the latent outcome, $\mathcal{U}$ is a closed subset of $(0,1)$ including the quantile indexes of interest, $\mathbf{R}$ is the non-decreasing rearrangement operator, and $\mathbf{Q}_{\tau}$ is the left-inverse (quantile) operator. The distribution of the observed outcome is given by: $$F_Y(y \mid D=1) = \varphi_y(\pi, \theta_y, F_Z) = \frac{\int \Phi_2\left( -x^\prime\beta(y), z^\prime \pi; -\rho(x'\delta(y)) \right) dF_Z(z)}{\int \Phi(z^\prime \pi) dF_Z(z)},$$ $\mathcal{F}$ contains $\{\Phi_2\left( -\cdot^\prime\beta(y), \cdot^\prime \pi; -\rho(\cdot'\delta(y)) \right) : y \in \mathcal{Y} \}$ as well as the indicators of all rectangles in $\overline{\mathbb{R}}^{d_z}$, and $\mathcal{U} = \mathcal{Y}$.
The following result is a corollary of Theorem \[thm:fclt\] by the functional delta method. Let $UC(\mathcal{Y},\xi)$ be the set of functions on $\mathcal{Y}$ that are uniformly continuous with respect to $\xi$, a standard metric on $\mathbb{R}$, and $UC(\mathcal{F},\lambda)$ be the set of functionals on $\mathcal{F}$ that are uniformly continuous with respect to $\lambda$, where $\lambda(f,\tilde f) = [{{\mathrm{P}}}(f - \tilde f)^2]^{1/2}$ for any $f, \tilde f \in \mathcal{F}$.
\[cor:fclt\] Suppose that Assumption \[ass:fclt\] holds, and $(p, t_y, F) \mapsto \varphi_\cdot(p, t_y, F)$, from $\mathbb{D}_{\Delta} \subset \mathbb{D}$ to $ \ell^\infty(\mathcal{U})$ is Hadamard differentiable at $(\pi, \theta_y, F_Z)$, tangentially to $\mathbb{R}^{d_{\pi}} \times UC(\mathcal{Y},\xi)^{d_{\theta}} \times UC(\mathcal{F},\lambda)$ with derivative $(p, t_y, F) \mapsto \varphi_\cdot'(p, t_y, F)$ that is defined and continuous on $\mathbb{R}^{d_{\pi}} \times \ell^{\infty}(\mathcal{Y})^{d_{\theta}} \times \ell^{\infty}(\mathcal{F})$. Then, $$\sqrt{n}({\widehat}\Delta_u - \Delta_u) \leadsto Z_{\Delta_u} := \varphi_u'(Z_{\pi}, Z_{\theta_y}, Z_{F}) \text{ in } \ell^{\infty}(\mathcal{U}),$$ where $Z_{\pi}$ and $Z_{\theta_y}$ are the random limits in Theorem \[thm:fclt\], $Z_{F}$ is a tight $F_Z$-Brownian bridge, and $u \mapsto Z_{\Delta_u}$ is a tight zero-mean Gaussian process.
\[remark:hd\] The distributions of the latent and observed outcome together with counterfactual distributions constructed thereof are examples of Hadamard differentiable functions. In the case of the latent outcome, the result follows from the Hadamard differentiability of the counterfactual operator in [chernozhukov+13inference]{}. In the case of the observed outcome, the result follows from the differentiability of the counterfactual operator and the composition rule for Hadamard derivatives applied to the ratio of two functions. Quantile (left-inverse) functionals of these distributions are Hadamard differentiable under additional conditions that guarantee that the quantile operator is Hadamard differentiable. These include that the outcome variable be continuous with density bounded above and away from zero [@cfg10]. Then the Hadamard differentiability of the quantile function follows from the composition rule for Hadamard derivatives.
\[remark:qinfer\] There are two alternatives to construct confidence bands for quantile functions. The first approach is the standard method based on characterizing the limit distribution of the estimator of the quantile function using the delta method, which relies on the Hadamard differentiability of the inverse operator. As we mention in Remark \[remark:hd\], this differentiability requires of additional conditions including that the outcome variable be continuous. The second approach applies to any type of outcome variable. It is based on the generic method of [cfmw16]{} that inverts confidence bands for distribution functions into confidence bands of quantile function. This method does not rely on the delta method and is therefore more robust to modeling assumptions and widely applicable. It has the shortcoming, however, that the bands might not be centered at the point estimate of the quantile function. We apply the second method to obtain most of the results in the empirical application.
Multiplier Bootstrap
--------------------
We make the following assumption about the bootstrap multipliers of Algorithm \[alg:mb\]:
\[ass:mb\] The multipliers $(\omega_{1},...,\omega_{n})$ are i.i.d. draws from a random variable $\omega \sim \mathcal{N}(0,1)$, and are independent of $\{(D^*_i, Y^*_i,Z_i)\}_{i=1}^n$ for all $n$.
Let $${\widehat}{\theta}_y^b = {\widehat}\theta_y + n^{-1} \sum_{i = 1}^n \omega_i \ \widehat \psi_i({\widehat}\theta_y, {\widehat}\pi)$$ be the multiplier bootstrap version of ${\widehat}{\theta}_y$. We establish a functional central limit theorem for the bootstrap for ${\widehat}{\theta}_y$. Here we use $\leadsto_{{{\mathrm{P}}}}$ to denote bootstrap consistency, i.e. weak convergence conditional on the data in probability, which is formally defined in Appendix \[app:notation\].
\[thm:bfclt\] Under the conditions of Theorem \[thm:fclt\] and Assumption \[ass:mb\], $$\sqrt{n} ({\widehat}\theta_y^b - {\widehat}\theta_y) \leadsto_{{{\mathrm{P}}}} Z_{\theta_y} \text{ in } \ell^{\infty}(\mathcal{Y})^{d_{\theta}},$$ where $y \mapsto Z_{\theta_y}$ is the same Gaussian process as in Theorem \[thm:fclt\].
The following result is a corollary of Theorem \[thm:bfclt\] by the functional delta method for the bootstrap [@vdV-W Chapter 3.9]. Let ${\widehat}\Delta^b_u = \varphi_u({\widehat}\pi^b, {\widehat}\theta^b_y, {\widehat}F^b_Z)$, be the multiplier bootstrap version of ${\widehat}\Delta_u$ where $${\widehat}\pi^b = {\widehat}\pi - n^{-1} \sum_{i=1}^n \omega_i \ {\widehat}{H}_1({\widehat}\pi)^{-1} S_{1i}({\widehat}\pi),$$ and ${\widehat}F^b_Z$ is the weighted empirical distribution of $Z$ that uses $(1+\omega_1, \ldots, 1+\omega_n)$ as sampling weights.
Suppose that the conditions of Corollary \[cor:fclt\] and Assumption \[ass:mb\] hold. Then, $$\sqrt{n}({\widehat}\Delta^b_u - {\widehat}\Delta_u) \leadsto_{{{\mathrm{P}}}} Z_{\Delta_u} \text{ in } \ell^{\infty}(\mathcal{U}),$$ where $Z_{\Delta_u}$ is the same process as in Corollary \[cor:fclt\].
Wage Decompositions in the UK {#sec:empirics}
=============================
We apply the DR model with sample selection to carry out wage decompositions accounting for endogenous employment participation using data from the United Kingdom.
Data
----
The data come from the U.K. Family Expenditure Survey (FES) for the years 1978 to 2001, Expenditure and Food Survey (EFS) for the years 2002 to 2007, and Living Costs and Food Survey (LCFS) for the years 2008 to 2013. Despite the differences in the name, these surveys contain comparable information. Indeed, the FES was combined to the National Food Survey to form the EFS, which was renamed LCFS when it became a module of the Integrated Household Survey. The data from the FES has been previously used by [gosling00]{}, [brs03]{}, [bgim07]{} and [ab17]{} to study wage equations in the U.K. labor market. We are not aware of any previous use of the data from the EFS and LCFS for this purpose.[^12] The three surveys contain repeated cross-sectional observations for women and men. The selection of the sample is similar to the previous work that used the FES. Thus, we keep individuals with ages between 23 to 59 years, and drop full-time students, self-employed workers, those married with spouse absent, and those with missing education or employees whose wages are missing. This leaves a sample of 258,900 observations, 139,504 of them correspond to women and 119,765 to men. The sample size per survey year and gender ranges from 2,197 to 4,545.
The outcome of interest, $Y$, is the logarithm of real hourly wage rate. We construct this variable as the ratio of the weekly usual gross main nominal earning to the weekly usual working hours, deflated by the U.K. quarterly retail price index. The selection variable, $D$, is an indicator for being employed.[^13] The covariates, $X$, include 5 indicators for age when ceasing school (${\leqslant}$15, 16, 17–18, 19–20, 21–22 and ${\geqslant}$ 23), a quartic polynomial in age, an indicator of being married or cohabiting, 6 variables with the number of kids by age categories (1, 2, 3–4, 5–10, 11–16, and 17-18), 36 survey year indicators, and 11 region indicators (Northern 5.48%, Yorkshire 9.56%, North Western 10.20%, East Midlands 7.36%, West Midlands 9.13%, East Anglia 5.31%, Greater London 10.06%, South Eastern 16.82%, South Western 7.94%, Wales 4.99%, Scotland 8.92%, and Northern Ireland 4.23%).[^14]
The excluded covariate, $Z_1$, is a potential out-of-work income benefit interacted with the marital status indicator used before in [brs03]{} and [bgim07]{}. This benefit is constructed with the Institute for Fiscal Studies (IFS) tax and welfare-benefit model (TAXBEN). TAXBEN is a static tax and benefit micro-simulation model of taxes on personal incomes, local taxes, expenditure taxes, and entitlement to benefits and tax credits that operates on large-scale, representative, household surveys [@b09]. It is designed to calculate the income of a tax unit if the individual considered were out of work.[^15] It is composed of eligible unemployment and housing benefits, which are determined by the demographic composition of the tax unit and the housing costs that the tax unit faces. These costs vary by region and over time due to numerous policy changes that have occurred over time. There is no consensus in the literature about the validity of this variable as excluded covariate. In our case the outcome and exclusion restrictions imply that, conditional on the observed covariates, the offered wage and dependence between offered wage and net reservation wage do not depend on the level of the benefit. We shall assume that the exclusion restrictions are satisfied and refer to [brs03]{} and [bgim07]{} for a discussion on the plausibility of the outcome restriction. In Appendix \[app:wage\], we provide an example of a model for offered and reservation wages that satisfies the exclusion restrictions.
Table \[table:summstat\] reports means and standard deviations of all the variables used in the analysis. We report these statistics for the entire sample, and by employment status and gender. The overall employment rate is 74%. Women are 17% less likely to be employed than men, and the unconditional gender wage gap is 33%. Overall, women and men are similar in terms of covariates. Both working men and women are relatively more highly educated, younger, and more likely to be married than their non-working counterparts. Having young children and high out-of-work benefits is negatively associated with employment for women but not for men.
\#1[\^[\#1]{}$^{#1}$]{}
[l\*[6]{}[c]{}]{} & & &\
& All & Employed & All & Employed & All & Employed\
Log Hourly Wage & & 2.38 & & 2.54 & & 2.21\
& & (0.54) & & (0.51) & & (0.52)\
Employed & 0.74 & & 0.83 & & 0.66 &\
& (0.44) & & (0.38) & & (0.47) &\
\
${\leqslant}$ 15 & 0.33 & 0.30 & 0.33 & 0.31 & 0.33 & 0.29\
& (0.47) & (0.46) & (0.47) & (0.46) & (0.47) & (0.45)\
16 & 0.31 & 0.30 & 0.32 & 0.32 & 0.30 & 0.29\
& (0.46) & (0.46) & (0.47) & (0.47) & (0.46) & (0.45)\
17-18 & 0.18 & 0.19 & 0.16 & 0.17 & 0.20 & 0.22\
& (0.38) & (0.39) & (0.37) & (0.37) & (0.40) & (0.41)\
19-20 & 0.04 & 0.05 & 0.04 & 0.04 & 0.04 & 0.05\
& (0.20) & (0.21) & (0.20) & (0.20) & (0.21) & (0.22)\
21-22 & 0.09 & 0.11 & 0.09 & 0.10 & 0.09 & 0.12\
& (0.29) & (0.31) & (0.29) & (0.30) & (0.29) & (0.32)\
${\geqslant}$23 & 0.05 & 0.05 & 0.06 & 0.06 & 0.04 & 0.04\
& (0.21) & (0.22) & (0.23) & (0.24) & (0.19) & (0.20)\
Age & 40.13 & 39.84 & 40.22 & 39.76 & 40.06 & 39.92\
& (10.43) & (10.10) & (10.40) & (10.11) & (10.45) & (10.08)\
Married & 0.76 & 0.79 & 0.78 & 0.81 & 0.75 & 0.76\
& (0.43) & (0.41) & (0.42) & (0.39) & (0.43) & (0.43)\
\
0–1 & 0.06 & 0.05 & 0.06 & 0.06 & 0.06 & 0.03\
& (0.24) & (0.22) & (0.24) & (0.25) & (0.24) & (0.18)\
2 & 0.05 & 0.04 & 0.05 & 0.05 & 0.05 & 0.03\
& (0.23) & (0.21) & (0.23) & (0.23) & (0.23) & (0.18)\
3–4 & 0.10 & 0.09 & 0.10 & 0.10 & 0.11 & 0.07\
& (0.32) & (0.29) & (0.31) & (0.32) & (0.32) & (0.27)\
5–10 & 0.32 & 0.29 & 0.30 & 0.30 & 0.33 & 0.28\
& (0.64) & (0.61) & (0.62) & (0.62) & (0.65) & (0.59)\
11–16 & 0.30 & 0.30 & 0.28 & 0.29 & 0.32 & 0.32\
& (0.63) & (0.62) & (0.61) & (0.61) & (0.64) & (0.63)\
17–18 & 0.03 & 0.04 & 0.03 & 0.03 & 0.04 & 0.04\
& (0.19) & (0.19) & (0.18) & (0.18) & (0.19) & (0.20)\
Benefit Income & 5.44 & 5.50 & 5.25 & 5.29 & 5.60 & 5.73\
& (0.74) & (0.78) & (0.70) & (0.72) & (0.73) & (0.78)\
Observations & 258,900 & 190,765 & 119,396 & 98,764 & 139,504 & 92,001\
\
\
Figure \[fig:trends\] provides some background on the U.K. labor market using our data. The left panel shows that over 36 years the average wages of working men and women have continuously grown and the unconditional gender wage gap has progressively narrowed from $46\%$ to $24\%$. The middle panel indicates that the growth of average wage has come together with an increase in wage inequality for both working men and women until 2000. The positive trend in wage inequality has continued for men after 2000, but not for women. The right panel shows opposite trends in the employment rate for men and women, where the gender employment gap has steadily and sharply reduced from 34% to 8%.
![Trends in U.K. labor market 1978-2013 by gender: left panel reports the average of the log wage rate, the middle panel reports the 90-10 percentile spread of the log wage rate, and the right panel reports the employment rate[]{data-label="fig:trends"}](Inequality.pdf "fig:"){width="32.00000%"} ![Trends in U.K. labor market 1978-2013 by gender: left panel reports the average of the log wage rate, the middle panel reports the 90-10 percentile spread of the log wage rate, and the right panel reports the employment rate[]{data-label="fig:trends"}](Inequality.pdf "fig:"){width="32.00000%"} ![Trends in U.K. labor market 1978-2013 by gender: left panel reports the average of the log wage rate, the middle panel reports the 90-10 percentile spread of the log wage rate, and the right panel reports the employment rate[]{data-label="fig:trends"}](Inequality.pdf "fig:"){width="32.00000%"}
Empirical Specifications
------------------------
We estimate the DR model for different samples and carry out several wage decompositions where we compare the distributions of men and women, or the distributions over time within genders. The specifications of the selection and outcome equations include all the covariates described above except for the excluded covariates in the outcome equation. The parameter of the selection sorting function is notoriously more difficult to estimate than the parameters of the selection and outcome equations. We consider four simplified specifications of the sorting function where the covariates included in the index $X'\delta(y)$ are:
- Specification 1: a constant.
- Specification 2: a constant and the marital status indicator.
- Specification 3: a constant and a linear trend on the year of the survey.
- Specification 4: a constant and a linear trend on the year of the survey interacted with the marital status indicator.
We also experimented with other specifications that include the education indicators, indicators of survey year, or age. We do not report these results because they do not show any clear pattern mainly due to imprecision in the estimation of the parameter $\delta(y)$.[^16]
Model Parameters
----------------
We report point estimates and 95% confidence bands for the coefficients of the education and marital status indicators in the outcome equation and the correlation function in the selection sorting. Estimates and 95% confidence bands for the coefficients of the selection equation, coefficients of the fertility indicators in the outcome equation and coefficients in the selection sorting function are given in the SM. The estimates are obtained with Algorithm \[alg:tsdr\] replacing $\mathcal{Y}$ by a finite grid containing the sample quantiles of log real hourly wage with indexes $\{0.10, 0.11, \ldots, 0.90 \}$ in the pooled sample of men and women. We report all the estimates as a function of the quantile index. The confidence bands are constructed by Algorithm \[alg:mb\] with $B=500$ bootstrap repetitions and the same finite grid as for the estimates. We also report estimates from the HSM of Example \[ex:hsm\] with dash lines as a benchmark of comparison.[^17]
The estimates of the coefficients of the education and marital status indicators in the outcome equation are reported in Figure \[fig:coef\_men\_spec1\] for men and Figure \[fig:coef\_women\_spec1\] for women. These estimates correspond to specification 1. Estimates for specifications 2–4 are given in the SM. For all the specifications and genders, we find that the returns to education are heterogenous across the distribution and broadly increasing in the years of education (age leaving school). In contrast, the classical HSM model does not allow the heterogeneity and estimates constant coefficients. This restriction is rejected empirically, lending support for the generalized HSM. The coefficient of the marital status indicator is uniformly positive for men, whereas is negative but mainly statistically not different from zero for women. We cannot reject that this coefficient is homogeneous across the distribution for both men and women.[^18]
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 1 for men[]{data-label="fig:coef_men_spec1"}](Estimates_11378_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 1 for women[]{data-label="fig:coef_women_spec1"}](Estimates_11378_500.pdf){height="0.75\textheight" width="85.00000%"}
Figures \[fig:rho\_men\_spec1\]–\[fig:rho\_men\_spec4\] display the estimates of the sorting effect functions for specifications 1–4, respectively. The estimates of the coefficients of these functions for specifications 3 and 4 are given in the SM. Figure \[fig:rho\_men\_spec1\] shows positive selection sorting for men and negative selection sorting for women. In both cases we cannot reject that the sorting is constant across the distribution. This finding is refined in Figure \[fig:rho\_men\_spec2\], where we uncover that the positive male sorting comes mainly from bachelors, whereas the negative female sorting comes from married women. This pattern is consistent with a marriage market where there is assortative matching in offered wages given observable characteristics, where women with high potential wages are married to highly paid working men and decide not to work [@neal04]. Figure \[fig:rho\_men\_spec3\] shows that the sorting homogeneity found in the pooled sample hides some heterogeneity across time. Thus, we find that the male sorting is heterogeneous in the early years, negative at the bottom and positive at the top of the distribution, and progressively becomes homogenous. The female sorting is more homogenous over time, but also displays a positive trend, specially at the bottom of the distribution. Figure \[fig:rho\_men\_spec4\] shows that the trends in sorting are driven by married individuals at the bottom of the distribution and single individuals at the top of the distribution.[^19]
![Estimates and 95% confidence bands for the selection sorting function: specification 1[]{data-label="fig:rho_men_spec1"}](Estimates_11378_500.pdf "fig:"){height="50.00000%" width="49.00000%"} ![Estimates and 95% confidence bands for the selection sorting function: specification 1[]{data-label="fig:rho_men_spec1"}](Estimates_11378_500.pdf "fig:"){height="50.00000%" width="49.00000%"}
![Estimates and 95% confidence bands for the selection sorting function: specification 2[]{data-label="fig:rho_men_spec2"}](Estimates_11378_couple_500.pdf "fig:"){height=".8\textwidth" width="40.00000%"} ![Estimates and 95% confidence bands for the selection sorting function: specification 2[]{data-label="fig:rho_men_spec2"}](Estimates_11378_couple_500.pdf "fig:"){height=".8\textwidth" width="40.00000%"}
![Estimates and 95% confidence bands for the selection sorting function: specification 3[]{data-label="fig:rho_men_spec3"}](Estimates_11378_lintime_500.pdf "fig:"){height="50.00000%" width="49.00000%"} ![Estimates and 95% confidence bands for the selection sorting function: specification 3[]{data-label="fig:rho_men_spec3"}](Estimates_11378_lintime_500.pdf "fig:"){height="50.00000%" width="49.00000%"}
![Estimates and 95% confidence bands for the selection sorting function: specification 4[]{data-label="fig:rho_men_spec4"}](Estimates_11378_lintimexcouple_500.pdf "fig:"){height=".8\textwidth" width="40.00000%"} ![Estimates and 95% confidence bands for the selection sorting function: specification 4[]{data-label="fig:rho_men_spec4"}](Estimates_11378_lintimexcouple_500.pdf "fig:"){height=".8\textwidth" width="40.00000%"}
Distributions of Offered and Observed Wages, and Wage Decompositions
--------------------------------------------------------------------
Figure \[fig:latent\] shows point estimates of the quantiles of offered and observed wages for men and women based on specification 1. Estimates for the other specifications and confidence bands for all the specifications are given in the SM. The offered wage is a latent variable defined for all the individuals that is free of sample selection. As we showed in Section \[sec:model\], the distributions of both types of wages can be expressed as functionals of the model parameters, and estimated using the plug-in estimators and .[^20] We find opposite signs in the sample selection bias for men and women. The quantiles of the observed wages are below the quantiles of latent wages for men, but the opposite holds for women. This pattern is consistent with the sign of the estimates of the selection sorting function, where we found positive sorting for men and negative sorting for women.
![Estimates of the quantiles of observed and offered (latent) wages: specification 1[]{data-label="fig:latent"}](Comp_Dist.pdf "fig:"){height="49.00000%" width="49.00000%"} ![Estimates of the quantiles of observed and offered (latent) wages: specification 1[]{data-label="fig:latent"}](Comp_Dist.pdf "fig:"){height="49.00000%" width="49.00000%"}
Figure \[fig:latent2\] compares the quantile function of offered wages between men and women and carries out a gender discrimination analysis based on specification 1. The estimates and 95% confidence bands for the other specifications are reported in the SM. The gender discrimination analysis is based on the counterfactual distributions $$F_{Y^* \langle j, k \rangle}(y) = \int \Phi(-x'\beta_j(y)) dF_{X_k}(x),$$ where $\beta_j(y)$ is the coefficient of the wage equation in group $j$, $F_{X_k}$ is the distribution of the characteristics in group $k$, and $j$ and $k$ are group indices for women and men. $F_{Y^* \langle j, k \rangle}$ corresponds to the distribution of offered wages that we would observed when the wage structure is as in group $j$ and the distribution of characteristics is as in group $k$. We decompose the difference in the quantile functions of the latent wages between women (group 1) and men (group 0) using the counterfactual distributions as $$F_{Y^* \langle 1, 1 \rangle} - F_{Y^* \langle 0, 0\rangle} = [F_{Y^* \langle 1, 1 \rangle} - F_{Y^* \langle 0, 1 \rangle}] + [F_{Y^* \langle 0, 1 \rangle} - F_{Y^* \langle 0, 0 \rangle}],$$ where the first term is the wage structure or discrimination effect and the second term is the composition effect. We obtain estimates of the counterfactual distributions and quantile functions using the plug-in estimator in and the operators $ \mathbf{Q}_{\tau}$ and $\mathbf{R}$. We find that the wages offered to women are between $21$ and $40\%$ lower than the wages offered to men at the same quantile index. The majority of this difference is explained by differences in the wage structure, $\beta(y)$, whereas differences in composition, $F_Z$, have very little explanatory power. This result can be interpreted as evidence of gender discrimination in the labor market.
![Estimates and 95% confidence bands for the quantiles of offered (latent) wages and decomposition between women and men: specification 1[]{data-label="fig:latent2"}](DFQF_11378_500.pdf "fig:"){height="49.00000%" width="49.00000%"} ![Estimates and 95% confidence bands for the quantiles of offered (latent) wages and decomposition between women and men: specification 1[]{data-label="fig:latent2"}](DFQF_11378_500.pdf "fig:"){height="49.00000%" width="49.00000%"}
We next use the DR model to decompose changes in the distribution of the observed wage between women and men, and between the first and second halves of the sample period for each gender. We extract four components that correspond to different inputs of the DR model:
1. Selection (employment) sorting: $\delta(y)$.
2. Selection (employment) structure: $\pi$.
3. Outcome (wage) structure: $\beta(y)$.
4. Composition: $F_Z$.
To define the effects of these components, let $F_{Y\langle t, s, r, k\rangle}$ be the counterfactual distribution of wages when the sorting is as in group $t$, the employment structure is as in group $s$, the wage structure is as in group $r$, and the composition of the population is as in group $k$. The actual distribution in group $t$ therefore corresponds to $F_{Y\langle t, t, t, t\rangle}$. We assume that there are two groups indexed by $0$ and $1$ that correspond to demographic populations such as men and women, or time periods such as the first and second halves of the sample years. Then, we can decompose the distribution of observed wage between group $1$ and group $0$ as: $$\begin{gathered}
F_{Y\langle 1, 1, 1, 1 \rangle} - F_{Y\langle 0, 0, 0, 0 \rangle} = [F_{Y\langle 1, 1, 1, 1 \rangle} - F_{Y\langle 0, 1, 1, 1 \rangle}] + [F_{Y\langle 0, 1, 1, 1 \rangle} - F_{Y\langle 0, 0, 1, 1 \rangle}] \\ + [F_{Y\langle 0, 0, 1, 1 \rangle} - F_{Y\langle 0, 0, 0, 1 \rangle}] + [F_{Y\langle 0, 0, 0, 1 \rangle} - F_{Y\langle 0, 0, 0, 0 \rangle}],\end{gathered}$$ where the first term in square brackets of the right hand side is a sorting effect, the second an employment structure effect, the third a wage structure effect, and the forth a composition effect. This is a distributional version of the classical Oaxaca-Blinder decomposition that accounts for sample selection [@kitagawa55; @oaxaca73; @blinder73]. It is well-known that the order of extraction of the components in this type of decompositions might matter. As a robustness check, we estimate the decomposition changing the ordering of the components. In results not reported, we find that the main findings are not sensitive to the change of ordering.
In terms of the DR model, the counterfactual distribution can be expressed as the functional $$F_{Y\langle t, s, r, k\rangle}(y) = \frac{\int \Phi_2\left( -x^\prime\beta_r(y), z^\prime \pi_s; -\rho(x'\delta_t(y)) \right) dF_{Z_k}(z)}{\int \Phi(z^\prime \pi_s) dF_{Z_k}(z)},$$ where $\delta_t$ is the coefficient of the sorting function in group $t$, $\pi_s$ is the coefficient of the employment equation in group $s$, $\beta_r$ is the coefficient of the wage equation in group $r$, and $F_{Z_k}$ is the distribution of characteristics in group $k$. Given random samples for groups $0$ and $1$, we construct a plug-in estimator of $F_{Y\langle t, s, r, k\rangle}$ by suitably combining the estimators of the model parameters and distribution of covariates from the two groups.
Figure \[fig:dec\_spec1\] reports estimates of the quantile functions of observed wages for men and women, together with the relative contributions of each component to the decomposition between men (group $0$) and women (group $1$) based on specification 1. The bands for the contributions are joint for all the components and rely on the delta method; see Remark \[remark:qinfer\]. Estimates of the components of the decomposition and the analysis based on specifications 2–4 are given in the SM. The distribution for men first order stochastically dominates the distribution for women. Most of this gender wage gap is explained by differences in the wage structure, i.e. differences in the returns to observed characteristics that might be associated to gender discrimination. However, differences in sorting and employment structure also account for an important percentage of the gap, specially at the top of the distribution. Thus, we uncover that the negative female sorting explains about 30–40% of the gap at the top of the distribution. A possible explanation is that women with very high potential wages decide not to work because there are no high-paid jobs available to them due to glass ceiling [@abv03]. The negative contribution of the employment structure can be explained by the order of the decomposition where we are applying the male employment structure to the female distribution with positive male sorting. In this case we are increasing the proportion of employed women, where the added women come from a pool with lower positive selection, and this negative effect is not reversed by a change in the composition of the working women; see Remark \[remark:se\] for more details. The aggregate selection effect, defined as the sum of the selection sorting and selection structure effects, is positive and statistically significant at the top of the distribution; see Figure \[fig:adec\_spec1\] in the SM. Differences in the composition of the characteristics contribute very little to explain the gender gap. Finally, the estimates from the HSM in dash lines pick up the average contributions of the components, but miss all the heterogeneity across the distribution.
![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between men and women in specification 1[]{data-label="fig:dec_spec1"}](CondDFQF_11378_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"} ![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between men and women in specification 1[]{data-label="fig:dec_spec1"}](CondDFQF_11378_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"}
Figures \[fig:dec2\_women\_spec1\] and \[fig:dec2\_men\_spec1\] report estimates of the quantile functions of observed wages for the first and second halves of the sample period, together with the relative contributions of each component to the decomposition between second half (group $0$) and first half (group $1$) based on specification 1 for women and men, respectively. Estimates of the components of the decompositions are given in the SM. The distribution for the second half first order stochastically dominates the distribution for the first half in both cases. For women, the most important components are the wage structure and composition effects in this order. The importance of the wage structure is decreasing along the distribution, whereas the importance of the composition is increasing. Composition and wage structure are also the most important components for men. The small contributions of the selection sorting component to the change in the distribution of wages between the two time period for both genders seem to contradict the linear time trends that we found in the coefficient of the sorting selection function. This might be explained by the inability of a coarse partition of the sample into two halves to capture the gradual increase in selection sorting, together with the changes in the composition.
![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between first and second half of the sample period for women in specification 1[]{data-label="fig:dec2_women_spec1"}](CondDFQF_fhalf_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"} ![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between first and second half of the sample period for women in specification 1[]{data-label="fig:dec2_women_spec1"}](CondDFQF_fhalf_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"}
![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between first and second half of the sample period for men in specification 1[]{data-label="fig:dec2_men_spec1"}](CondDFQF_mhalf_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"} ![Estimates and 95% confidence bands for the quantiles of observed wages and decomposition between first and second half of the sample period for men in specification 1[]{data-label="fig:dec2_men_spec1"}](CondDFQF_mhalf_500.pdf "fig:"){height=".49\textwidth" width=".49\textwidth"}
Discussion
----------
The main findings can be summarized as: (1) heterogeneous positive effect of education and homogeneous effect of being married on offered wages for both genders; (2) positive sorting for men and negative sorting for women driven by single men and married women, which is consistent with assortative matching in the marriage market; (3) heterogeneity in selection sorting decreases gradually over time; (4) differences in returns to characteristics in the wage equation, which might be associated to gender discrimination in the labor market, account for most of the gender wage gap; (5) selection sorting on unobservables explains up to 39% of the gender wage gap at the top of the distribution, which can be taken as evidence of glass ceiling; and (6) changes in the structure of the wage equation and composition of the characteristics account for most of the differences in the wage distribution between the two halves of the sample period within each gender.
We compare and contrast these findings with previous results from the literature that studied similar issues. These results were obtained from different data and/or using different methodology. [bgim07]{} applied a bound approach that does not require of exclusion restrictions to study the evolution of wage inequality using the FES data for the period 1978–2000. They assumed positive sorting for men and women in some of their estimates to make the bounds more informative. Interestingly, they mentioned the possibility that the assumption is violated for married women due to assortative matching in the marriage market.[^21] They also found evidence against the validity of out-of-work benefit income as a valid excluded covariate for men. [ab17]{} using the same data from the FES, also found positive sorting for men, stronger for single than for married men, using an alternative methodology that combines quantile regression for the marginal distributions with a parametric model for the copula. Contrary to our findings, they also found positive selection for women, which is statistically significant only for married women. [mr08]{} estimated a HSM using data from the US-CPS for the periods 1975-1979 and 1995-1999. They found that the selection sorting for women shifted from negative to positive between the two periods. We also find for the UK that the sorting for most women has a positive trend over time, but remains negative even in 2013 for most of the distribution. [mw18]{} applied the methodology of [ab17]{} to data from the US-CPS for the period 1976–2014. They also found negative sorting for women at the beginning of the sample period that became positive during the 90s, and positive sorting for men throughout the entire period. [bertrand17]{} pointed out multiple possible explanations for the glass ceiling based on the field of education, psychological attributes or preferences for job flexibility that are compatible with our finding on the importance of sorting on unobservables at the top of the distribution. None of the previous papers distinguished between the selection sorting and selection structure effects.
One limitation of our dataset is that it does not contain a direct measure of work experience. As a final robustness check, we find that the results are not sensitive to the exclusion of college graduates from the sample by redoing the analysis excluding all the individuals who cease school after age 18. This is a relevant exclusion because work experience is a more relevant determinant of wage for highly educated workers.[^22]
Monte Carlo Simulation
======================
We conduct a Monte Carlo simulation calibrated to the empirical application to study the properties of the estimation and inference methods in small samples. The data generating process is the HSM of Example \[ex:hsm\] with the values of the covariates and parameters calibrated to the data for women in the last ten years of the sample (2004–2013). We do not use the entire dataset to speed up computation. We generate 500 artificial datasets and estimate the DR-model with the same specifications for the selection and outcome equations as in the empirical application and specification 1 for the selection sorting function, i.e. $\rho(x'\delta(y)) = \rho(y)$.
Figures \[figure:Sim\_col\], \[figure:Sim\_married\] and \[figure:Sim\_rho\] report the biases, standard deviations and root mean square errors for the estimators of the coefficients of the college (age when ceasing school 21–22) and marital status indicators in the outcome equation, and $\rho(y)$ in the selection sorting function, as a function of the quantile indexes of the values of log real hourly wage in the data used in the calibration.[^23] Although these coefficients are constant in the HSM, we do not impose this condition in the estimation. The estimates are obtained with Algorithm \[alg:tsdr\] replacing $\mathcal{Y}$ by a finite grid containing the sample quantiles of log real hourly wage with indexes $\{0.10, 0.11, \ldots, 0.90 \}$ in the original subsample of women in the last ten years of the sample. All the results are in percentage of the true value of the parameter. As predicted by the asymptotic theory, the biases are all small relative to the standard deviations and root mean squared errors. The estimation error increases for all the coefficients as we move away from the median towards tail values of the outcome.
![Bias, SD and RMSE for the coefficient of the college indicator in the outcome equation[]{data-label="figure:Sim_col"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for the coefficient of the college indicator in the outcome equation[]{data-label="figure:Sim_col"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for the coefficient of the college indicator in the outcome equation[]{data-label="figure:Sim_col"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"}
![Bias, SD and RMSE for the coefficient of the marital status indicator in the outcome equation[]{data-label="figure:Sim_married"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for the coefficient of the marital status indicator in the outcome equation[]{data-label="figure:Sim_married"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for the coefficient of the marital status indicator in the outcome equation[]{data-label="figure:Sim_married"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"}
![Bias, SD and RMSE for coefficient $\rho(y)$ in the selection sorting equation[]{data-label="figure:Sim_rho"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for coefficient $\rho(y)$ in the selection sorting equation[]{data-label="figure:Sim_rho"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"} ![Bias, SD and RMSE for coefficient $\rho(y)$ in the selection sorting equation[]{data-label="figure:Sim_rho"}](Sim_f_113104_coef.pdf "fig:"){width="32.00000%"}
Table \[table:simcb\] shows results on the finite sample properties of 95% confidence bands for the coefficients of the indicators of college and marital status in the outcome equation and $\rho(y)$ of the selection sorting function. The confidence bands are constructed by Algorithm \[alg:mb\] with $B=200$ bootstrap repetitions and the same grid of values $\bar{\mathcal{Y}}$ as for the estimators. We report the average length of the confidence bands integrated over threshold values, the average value of the estimated critical values, and the empirical coverages of the confidence bands. For comparison, we also report the coverage of pointwise confidence bands using the normal distribution, i.e. with critical value equal to 1.96. The last row computes the ratio of the standard error averaged across simulations to the simulation standard deviation, integrated over threshold values. We find that the bands have coverages close to the nominal level. As expected, pointwise bands severely undercover the entire functions. The standard errors based on the asymptotic distribution provide a fair approximation to the sampling variability of the estimator.
\[table:simcb\]
[lccc]{} & College & Married & $\rho(y)$\
Average Length & 0.38 & 0.16 & 0.35\
Average Critical Value & 2.91 & 2.89 & 2.88\
Coverage uniform band (%) & 96 & 98 & 96\
Coverage pointwise band (%) & 68 & 64 & 67\
Average SE/SD & 1.04 & 1.05 & 1.07\
Conclusion
==========
We develop a distribution regression model with sample selection that accommodates rich patterns of heterogeneity in the effects of covariates on outcomes and selection. The model is semiparametric in nature, as it has function-valued parameters, and is able to considerably generalize the classical selection model of [heckman74]{}. Furthermore, the model allows for richer covariate effects than the previous semiparametric generalizations which allowed the location effects for covariates. We propose to estimate the model by a process of probit regressions, indexed by threshold-dependent parameters. We show that the resulting estimators of the function-valued parameters are approximately Gaussian and concentrate in a $1/\sqrt{n}$ neigborhood of the true values. We present an extensive wage decomposition analysis for the U.K. using new data, generating both new findings and demonstrating the power of the method.
Proofs of Section \[sec:id\] {#app:id}
============================
Proof of Lemma \[lemma:lgr\]
----------------------------
By standard properties of the bivariate normal distribution, the marginals corresponding to the LGR are $\Phi(\mu(y))$ and $\Phi(\nu(d))$. Equalizing these marginals to the marginals of $F_{Y^*,D^*}(y,d)$ yields $$F_{Y^*}(y) = \Phi(\mu(y)), \quad F_{D^*}(d) = \Phi(\nu(d)),$$ which uniquely determine $\mu(y)$ and $\nu(x)$ as $\mu(y) = \Phi^{-1}(F_{Y^*}(y))$ and $\nu(d) = \Phi^{-1}(F_{D^*}(d))$. Plugging these expressions in the LGR gives $$F_{Y^*,D^*}(y,d) = \Phi_2(\Phi^{-1}(F_{Y^*}(y)), \Phi^{-1}(F_{D^*}(d)); \rho(y,d)).$$ Let $\phi_2(\cdot, \cdot; \rho)$ be the joint probability density function (PDF) of a standard bivariate normal random variable with parameter $\rho$. The previous equation uniquely determines $\rho(y,d)$ by the following properties of the standard bivariate normal distribution:
1. $\rho \mapsto \Phi_2(\cdot, \cdot; \rho)$ is continuously differentiable and $
\partial \Phi_2(\cdot, \cdot; \rho)/\partial \rho = \phi_2(\cdot, \cdot; \rho) > 0
$ [@sibuya59; @sungur90];
2. $
\lim_{\rho \nearrow 1} \Phi_2(x, y; \rho) = \min[\Phi(x),\Phi(y)];
$
3. $
\lim_{\rho \searrow -1} \Phi_2(x, y; \rho) = \max[ \Phi(x) + \Phi(y) -1, 0];
$
together with the Frechet-Hoeffding bounds $$\max[ \Phi(\mu(y)) + \Phi(\nu(d)) - 1, 0] {\leqslant}F_{Y^*,D^*}(y,d) {\leqslant}\min[\Phi(\mu(y)),\Phi(\nu(d))].$$
Proof of Theorem \[theorem:id\]
-------------------------------
The identification of $\nu(z)$ follows from equalizing the marginals with respect to $D^*$ of $F_{Y^*,D^* \mid Z}$ and the conditional LGR at $D^*=0$. Since $\nu(z)$ is identified, we shall use $\Phi(\nu(z))$ in place of ${{\mathrm{P}}}(D=1 \mid Z=z)$ and $\bar{\Phi}(\nu(z))$ in place of ${{\mathrm{P}}}(D=0 \mid Z=z)$ in the rest of the proof to lighten the notation.
Cases (1)–(3) correspond to $\rho(y) = 1$. This boundary case is identified because $\rho(y) = 1$ if and only if $F_{Y,D \mid Z} (y,1 \mid 1) - \bar{\Phi}(\nu(1))= F_{Y,D \mid Z} (y,1 \mid 0) - \bar{\Phi}(\nu(0)) > 0$ or $F_{Y,D \mid Z} (y,1 \mid 0) = 1$. The if part follows from the Frechet-Hoeffding bounds $$\label{eq:fh1}
F_{Y,D \mid Z} (y,1 \mid z) - \bar{\Phi}(\nu(z)) = F_{Y^*,D^* \mid Z} (y,0 \mid z) = \min[\Phi(\nu(z)),F_{Y^*}(y)], \ \ z \in \{0,1\},$$ and Assumption \[ass:er\](2). For the case $F_{Y,D \mid Z} (y,1 \mid 1) - \bar{\Phi}(\nu(1))= F_{Y,D \mid Z} (y,1 \mid 0) - \bar{\Phi}(\nu(0)) > 0$, the only if part follows because $\nu \mapsto \Phi_2(\cdot, \nu; \rho)$ is strictly monotonic when $\rho \in (-1,1)$ and $\nu(1) > \nu(0)$ by Assumption \[ass:er\](2). This shows that $\rho \not\in (-1,1)$. Moreover, this case is ruled out when $\rho = -1$ by the Frechet-Hoeffding bounds $$\label{eq:fh2}
F_{Y,D \mid Z} (y,1 \mid z) - \bar{\Phi}(\nu(z)) = F_{Y^*,D^* \mid Z} (y,0 \mid z) = \max[\Phi(\nu(z)) + F_{Y^*}(y) - 1, 0], \ \ z \in \{0,1\},$$ and Assumption \[ass:er\](2). The case $F_{Y,D \mid Z} (y,1 \mid 0) = 1$ implies that $\Phi_2(\mu(y), \nu(0); \rho(y)) = \Phi(\nu(0))$, which is only possible when $\rho(y) = 1$.
Now, we can analyze the identification of $\mu(y)$ using with $F_{Y^*}(y)= \Phi(\mu(y))$. Case (1) corresponds to $F_{Y,D \mid Z} (y,1 \mid 1) - \bar{\Phi}(\nu(1))= \Phi(\mu(y))$, which identifies $\mu(y)$. Case (2) corresponds to $F_{Y,D \mid Z} (y,1 \mid 0) = 1$ and $F_{Y,D \mid Z} (y,1 \mid 1) - \bar{\Phi}(\nu(1))= \Phi(\mu(y))$. The second equation identifies $\mu(y)$. Case (3) corresponds to $F_{Y,D \mid Z} (y,1 \mid z) = 1$, $z \in \{0,1\},$ which partially identify the parameter from $\Phi(\mu(y)) {\geqslant}\max[\Phi(\nu(0)), \Phi(\nu(1))] = \Phi(\nu(1))$ by Assumption \[ass:er\](2).
Cases (4)–(6) correspond to $\rho(y) = -1$. This boundary case is identified because $\rho(y) = -1$ if and only if $F_{Y,D \mid Z} (y,1 \mid 0) = \bar{\Phi}(\nu(0))$ or $F_{Y,D \mid Z} (y,1 \mid 0) = F_{Y,D \mid Z} (y,1 \mid 1)$. Symmetrically to $\rho(y) = 1$, the if part follows from the Frechet-Hoeffding bounds and Assumption \[ass:er\](2), whereas the only if part for the case $F_{Y,D \mid Z} (y,1 \mid 0) = F_{Y,D \mid Z} (y,1 \mid 1)$ follows from the Frechet-Hoeffding bounds and Assumption \[ass:er\](2). The only if part for $F_{Y,D \mid Z} (y,1 \mid 0) = \bar{\Phi}(\nu(0))$ follows because this case implies that $\Phi_2(\mu(y), \nu(0); \rho(y)) = 0$, which is only possible when $\rho(y) = -1$.
Now, we can analyze the identification of $\mu(y)$ using with $F_{Y^*}(y)= \Phi(\mu(y))$. Case (4) corresponds to $F_{Y,D \mid Z} (y,1 \mid 1) = \Phi(\mu(y))$, which identifies $\mu(y)$. Case (5) corresponds to $F_{Y,D \mid Z} (y,1 \mid 0) = \Phi(\mu(y))$ and $F_{Y,D \mid Z} (y,1 \mid 1) = \Phi(\mu(y))$. Both of these equations have the same solution that identifies $\mu(y)$. Case (6) corresponds to $F_{Y,D \mid Z} (y,1 \mid z) = \bar{\Phi}(\nu(z))$, $z \in \{0,1\}$, which partially identify the parameter from $\Phi(\mu(y)) {\leqslant}\min[\bar{\Phi}(\nu(0)), \bar{\Phi}(\nu(1))] = \bar{\Phi}(\nu(1)) = F_{Y,D \mid Z} (y,1 \mid z)$ by Assumption \[ass:er\](2).
Consider now the non-boundary case (7) where $\rho(y) \in (-1,1)$. The parameters $\mu(y)$ and $\rho(y)$ are identified as the solution in $(\mu,\rho)$ to . This nonlinear system of 2 equations has unique solution under Assumption \[ass:er\](2). This result follows from Theorem 4 of [gale65]{}, after showing that the Jacobian of the system is a P-matrix when $\rho(y) \in (-1,1)$.
Let $\partial_{\mu}\Phi_2(\mu,\nu;\rho) = \partial \Phi_2(\mu,\nu;\rho)/\partial \mu $ and $\partial_{\rho}\Phi_2(\mu,\nu;\rho) = \partial \Phi_2(\mu,\nu;\rho)/\partial \rho $. The Jacobian matrix of the system, $$J(\mu(y),\rho(y)) = \left(\begin{array}{cc}\partial_{\mu} \Phi_2(\mu(y), \nu(1); \rho(y)) & \partial_{\rho} \Phi_2(\mu(y), \nu(1); \rho(y)) \\
\partial_{\mu} \Phi_2(\mu(y), \nu(0); \rho(y)) & \partial_{\rho} \Phi_2(\mu(y), \nu(0); \rho(y)) \end{array}\right),$$ is a P-matrix for all $\mu(y) \in \mathbb{R}$ and $\rho(y) \in (-1,1)$ because by the properties of the bivariate normal CDF: $$\partial_{\mu} \Phi_2(\mu(y), \nu(1); \rho(y)) = \Phi\left( \frac{\nu(1) - \rho(y) \mu(y)}{\sqrt{1-\rho(y)^2}}\right) \phi(\mu(y)) > 0,$$ $$\partial_{\rho} \Phi_2(\mu(y), \nu(0); \rho(y)) = \phi_2(\mu(y), \nu(0); \rho(y)) > 0,$$ and $$\det(J(\mu(y),\rho(y))) = \phi(\mu(y))^2 \left[\Phi\left( \tilde \nu(1,y) \right) \phi\left( \tilde \nu(0,y)\right) - \Phi\left( \tilde \nu(0,y)\right) \phi\left( \tilde \nu(1,y)\right) \right] > 0,$$ where $\tilde{\nu}(0,y) = [\nu(0) - \rho(y) \mu(y)]/\sqrt{1-\rho(y)^2}$ and $\tilde{\nu}(1,y) = [\nu(1) - \rho(y) \mu(y)]/\sqrt{1-\rho(y)^2}$. In the last result we use that, by the properties of the normal distribution, $$\phi_2(\mu, \nu; \rho) = \phi\left( [\nu - \rho \mu]/\sqrt{1-\rho^2}\right) \phi(\mu)$$ and the inverse Mills ratio $\nu \mapsto \lambda(\nu) := \phi(\nu)/\Phi(\nu)$ is strictly decreasing in $\mathbb{R}$, so that $$\Phi\left( \tilde \nu(1,y)\right) \phi\left( \tilde \nu(0,y)\right) - \Phi\left( \tilde \nu(0,y)\right) \phi\left( \tilde \nu(1,y)\right) > 0,$$ since $ \tilde \nu(0,y) < \tilde \nu(1,y)$.
Detailed Comparison with AB17 {#app:ab17}
=============================
We need to introduce some notation to state the conditions of AB17. Let $p(z) = {{\mathrm{P}}}(D=1 \mid Z=z)$ and $V = F_{D^*\mid Z}(D^* \mid Z) $ such that $V \mid Z \sim U(0,1)$.[^24] AB17 assumed that (i) $(Y^*,V)$ are independent of $Z$, (ii) $v \mapsto C_{Y^*,V}(\cdot,v)$ is real analytic on the unit interval, where $C_{Y^*,V}$ is the copula of $(Y^*,V)$, and (iii) the support of $p(Z)$ contains an open interval. The condition (iii) requires $Z$ to have continuous variation and is therefore more restrictive than our assumption that $Z$ can be binary. We now show that our selection exclusion neither implies nor is implied by conditions (i) and (ii). Selection exclusion implies that for any $u \in [0,1]$ that satisfies $F_{Y^*}(y_u) = u$ for some $y_u$, $$C_{Y^*,V \mid Z}(u,p(z) \mid z) = C_{Y^*,D^* \mid Z}(u,p(z) \mid z) = \Phi_2(\Phi^{-1}(u), \Phi^{-1}(p(z)); \rho(y_u,0)) = C_{Y^*,V}(u,p(z)),$$ since $p(z) = F_{D^* \mid Z}(0 \mid z)$. This implication is weaker than condition (i) but it suffices for the identification argument in AB17. However, it only guarantees that $v \mapsto C_{Y^*,V}(\cdot,v)$ is real analytic on the support of $p(Z)$.[^25] Therefore, we conclude that selection exclusion implies conditions (i) and (ii) only if the support of $p(Z)$ is the unit interval. To verify that the converse is also not true, note that the LGR of $(Y^*,V)$ conditional on $Z$ under condition (i) is $$F_{Y^*,V \mid Z}(y,v \mid z) = \Phi_2(\tilde \mu(y), \tilde \nu(v); \tilde \rho(y,v)).$$ This, together with $\tilde \mu(y) = \mu(y)$ and $\tilde \nu(p(z)) = \nu(z)$, imply that $$F_{Y^*,D^* \mid Z}(y,0 \mid z) = \Phi_2(\tilde \mu(y), \tilde \nu(p(z)); \tilde \rho(y,p(z))) = \Phi_2(\mu(y), \nu(z); \tilde \rho(y,p(z))),$$ which satisfies the selection exclusion only if $\tilde \rho(y,v) = \tilde \rho(y)$ for all $v$ in the support of $p(Z)$, i.e. the local dependence between $Y^*$ and $V$ does not vary with the value of $V$ in this region. We finally note that condition (i) together with $\tilde \rho(y,v) = \tilde \rho(y)$ for all $v$ in the unit interval imply condition (ii) because $$C_{Y^*,V}(\cdot,v) = \Phi_2(\cdot, \Phi^{-1}(v); \tilde \rho(\cdot))$$ is a real analytic function with respect to $v$ in the unit interval. Alternatively, condition (ii) is equivalent to $v \mapsto \tilde \rho(\cdot,v)$ being real analytic, which is weaker than $\tilde \rho(y,v) = \tilde \rho(y)$.
Model for Offered and Reservation Wages {#app:wage}
=======================================
We provide a simple model for offered and reservation wages that satisfies the exclusion restrictions of Assumption \[ass:er\]. Let $Y^*$ and $R^*$ denote the offered wage and reservation wage, respectively. The latent employment variable is $D^* = Y^* - R^*$. We partition the vector covariates $Z$ as $Z=(X,Z_1)$, where $Z_1$ is the subset of covariates that will satisfy some exclusion restrictions.
Assume that $$(Y^*,R^*) \mid Z = z \sim \mathcal{N}_2\left( \left[\begin{array}{c} \mu_{Y^*}(x) \\ \mu_{R^*}(z)\end{array}\right], \left[\begin{array}{cc} \sigma_{Y^*}(x)^2 & \sigma_{Y^*,R^*}(x) \\ \sigma_{Y^*,R^*}(x) & \sigma_{R^*}(x)^2 \end{array}\right] \right),$$ where $z = (x,z_1)$. This assumption, in addition to joint normality, imposes that $Y^*$ is independent of $Z_1$, $Y^*$ and $R^*$ are covariance-independent of $Z_1$, and $R^*$ is variance-independent of $Z_1$, all conditional on $X$.[^26] Then, by the properties of the normal distribution $$(Y^*,D^*) \mid Z = z \sim \mathcal{N}_2\left( \left[\begin{array}{c} \mu_{Y^*}(x) \\ \mu_{D^*}(z)\end{array}\right], \left[\begin{array}{cc} \sigma_{Y^*}(x)^2 & \sigma_{Y^*,D^*}(x) \\ \sigma_{Y^*,D^*}(x) & \sigma_{D^*}(x)^2 \end{array}\right] \right),$$ where $\mu_{D^*}(z) = \mu_{Y^*}(x) - \mu_{R^*}(z)$, $\sigma_{Y^*,D^*}(x) = \sigma_{Y^*}(x)^2 - \sigma_{Y^*,R^*}(x)$, and $\sigma_{D^*}(x)^2 = \sigma_{Y^*}(x)^2 + \sigma_{R^*}(x)^2 - 2\sigma_{Y^*,R^*}(x)$. This distribution satisfies the exclusion restrictions of Assumption \[ass:er\] because it has the LGR: $$F_{Y^*,D^* \mid Z}(y, d \mid z) = \Phi_2\left(\frac{y - \mu_{Y^*}(x)}{\sigma_{Y^*}(x)}, \frac{d - \mu_{D^*}(z)}{\sigma_{D^*}(x)}; \frac{\sigma_{Y^*,D^*}(x)}{\sigma_{Y^*}(x)\sigma_{D^*}(x)}\right),$$ where the first and last argument do not depend on the value of the excluded covariates $Z_1$.
Notation {#app:notation}
========
We adopt the standard notation in the empirical process literature, e.g. [vdV-W]{}, $${\mathbb{E}_n}[f]={\mathbb{E}_n}[f(A)]=n^{-1}\sum_{i=1}^{n}f(A_{i}),$$ and $$\mathbb{G}_n[f]=\mathbb{G}_n[f(A)]=n^{-1/2}\sum_{i=1}^{n}(f(A_{i})-{{\mathrm{E}}}[f(A)]).$$ When the function $\widehat{f}$ is estimated, the notation should interpreted as: $$\mathbb{G}_n[\widehat{f}\ ]=\mathbb{G}_n[f]\mid_{f=\widehat{f}}\text{\ and \
}{{\mathrm{E}}}[\widehat{f}\ ]={{\mathrm{E}}}[f]\mid_{f=\widehat{f}}.$$ We also follow the notation and definitions in [vdV-W]{} of bootstrap consistency. Let $D_{n}$ denote the data vector and $E_{n}$ be the vector of bootstrap weights. Consider the random element $Z_{n}^{b}=Z_{n}(D_{n},E_{n})$ in a normed space $\mathbb{Z}$. We say that the bootstrap law of $Z_{n}^{b}$ consistently estimates the law of some tight random element $Z$ and write $Z_{n}^{b}\rightsquigarrow_{{{\mathrm{P}}}}Z$ in $\mathbb{Z}$ if $$\begin{array}{r}
\sup_{h\in\text{BL}_{1}(\mathbb{Z})}\left|{{\mathrm{E}}}^{b}h\left(Z_{n}^{b}\right)- {{\mathrm{E}}}h(Z)\right|\rightarrow_{{{\mathrm{P}}}^{*}}0,\end{array}
\label{boot1}$$ where $\text{BL}_{1}(\mathbb{Z})$ denotes the space of functions with Lipschitz norm at most 1, ${{{\mathrm{E}}}}^{b}$ denotes the conditional expectation with respect to $E_{n}$ given the data $D_{n}$, and $\rightarrow_{{{\mathrm{P}}}^{*}}$ denotes convergence in (outer) probability.
Proofs of Section \[sec:theory\]
================================
We use the $Z$-process framework described in Appendix E.1 of [chernozhukov+13inference]{}. To set-up the problem in terms of this framework, we need to introduce some notation. Let $W := (Z,D,YD)$ denote all the observed variables and $\xi_y := (\pi',\theta_y')'$ be a vector with the model parameters of the first and second steps. Let $$\varphi_{y,\xi}(W) :=
\left[
\begin{array}{c}
S_{1,\xi}(W) \\
S_{2y,\xi}(W)
\end{array}
\right] = \left[
\begin{array}{c}
\frac{\partial \ell_{1,\xi}(W)}{\partial \pi} \\
\frac{\partial \ell_{2y,\xi}(W)}{\partial \theta_y}
\end{array}
\right]$$ where $$\begin{aligned}
\ell_{1,\xi}(W) &:=& D\log{\Phi(Z^\prime \pi)}+(1-D)\log{\Phi(-Z^\prime \pi)}, \\
\ell_{2y,\xi}(W) &:=& D [I_{y} \log{\Phi_2\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)} + (1-I_{y}) \log{\Phi_2\left(X^\prime \beta(y),Z^\prime \pi; \rho(X'\delta(y)) \right)} ], \end{aligned}$$ be the scores of the first and second steps; and $$\label{eq:hess}
J(y) = {{\mathrm{E}}}\left[ \frac{\partial \varphi_{y,\xi}(W)}{\partial \xi'}\right] = \left[
\begin{array}{cc}
H_1 & 0 \\
J_{21y} & H_{2y}
\end{array}
\right]$$ be the expected Hessian evaluated at the true value of $\xi_y$. We provide more explicit expressions for the score and expected Hessian in Appendix \[app:scores\]. Note that $$\label{eq:invhess}
J^{-1}(y) =
\left[
\begin{array}{cc}
H_1^{-1} & 0 \\
- H_1^{-1} J_{21y} H_{2y}^{-1} & H_{2y}^{-1}
\end{array}
\right]$$ by the inverse of the partitioned inverse formula, and $$\label{eq:vscore}
{{\mathrm{E}}}[\varphi_{y,\xi}(W) \varphi_{\tilde y,\xi}(W)'] = \left[
\begin{array}{cc}
{{\mathrm{E}}}[S_{1,\xi}(W) S_{1,\xi}(W)'] & 0 \\
0 & {{\mathrm{E}}}[S_{2y,\xi}(W) S_{2\tilde y,\xi}(W)']
\end{array}
\right]$$ because ${{\mathrm{E}}}[S_{1,\xi}(W) S_{2y,\xi}(W)'] = 0$ for all $y \in \mathcal{Y}$.
Auxiliary Results
-----------------
We start by providing sufficient conditions that are useful to verify Condition Z in [chernozhukov+13inference]{}. They are an alternative to Lemma E.1 of [chernozhukov+13inference]{}, where we replace the requirement that the function $\xi \mapsto \Psi(\xi,y) := {{\mathrm{E}}}[\varphi_{y,\xi}(W)]$ is the gradient of a convex function by compactness of the parameter space for $\xi_y$ and an identification condition.[^27]
\[cont inverse\] Suppose that $\Xi$ is a compact subset of $\mathbb{R}^{d_{\xi}}$, and $\mathcal{Y}$ is a compact interval in $\mathbb{R}$. Let $\mathcal{I}$ be an open set containing $\mathcal{Y}$. Suppose that (a) $\Psi :\Xi \times \mathcal{I}
\mapsto \mathbb{R}^{d_{\xi}}$ is continuous, and $\xi \mapsto \Psi (\xi ,y) $ possesses a unique zero at $\xi_y$ that is in the interior of $\Xi$ for each $y\in \mathcal{Y}$, (b) for each $y\in \mathcal{Y}$, $\Psi (\xi_y,y)=0$, (c) $\frac{\partial }{\partial (\xi ^{\prime },y)}\Psi (\xi ,y)$ exists at $(\xi_y,y)$ and is continuous at $(\xi_y,y)$ for each $y\in
\mathcal{Y}$, and $\dot{\Psi}_{\xi_y,y}:=\frac{\partial }{\partial
\xi ^{\prime }}\Psi (\xi ,y)|_{\xi_y}$ obeys $\inf_{y\in
\mathcal{Y}}\inf_{\Vert h\Vert =1}\Vert \dot{\Psi}_{\xi_y,y}h\Vert
>c_{0}>0$. Then Condition Z of [chernozhukov+13inference]{} holds and $y \mapsto \xi_y$ is continuously differentiable.
**Proof of Lemma \[cont inverse\].** We restate the statement of Condition Z of [chernozhukov+13inference]{} with our notation for the reader’s reference.
<span style="font-variant:small-caps;"> Condition Z.</span> * Let $\mathcal{Y}$ be a compact set of some metric space, and $\Xi $ be an arbitrary subset of $\mathbb{R}^{d_{\xi}}$. Assume (i) for each $y\in \mathcal{Y}$, $\Psi (\cdot ,y):\Xi \mapsto
\mathbb{R}^{d_{\xi}}$ possesses a unique zero at $\xi_y$, and, for some $\delta >0$, $\mathcal{N}:=\cup _{y\in \mathcal{Y}}B_{\delta }(\xi_y)
$ is a compact subset of $\mathbb{R}^{d_{\xi}}$ contained in $\Xi $, (ii) the inverse of $\Psi (\cdot ,y)$ defined as $\Psi ^{-1}(x ,y) := \{ \xi \in
\Xi: \Psi(\xi,y) =x \}$ is continuous at $x=0$ uniformly in $y\in
\mathcal{Y}$ with respect to the Hausdorff distance, (iii) there exists $\dot{\Psi}_{\xi_y,y}$ such that $\lim_{t \searrow 0}\sup_{y\in {\mathcal{Y}},\Vert h\Vert=1}|t^{-1}[\Psi(\xi_y+th,y) -
\Psi(\xi_y,y)] -\dot{\Psi}_{\xi_y,y}h|=0$, where $\inf_{y\in \mathcal{Y}}\inf_{\Vert h\Vert =1}$ $\Vert \dot{\Psi}_{\xi_y,y}h\Vert >0$, and (iv) the maps $y \mapsto \xi_y$ and $y
\mapsto \dot{\Psi}_{\xi_y,y}$ are continuous.*
The first part of Z(i) follows immediately from condition (a). The verifications of the second part of Z(i), Z(iii) and Z(iv) are omitted because they follow by the same argument as in the proof of Lemma E.1 of [chernozhukov+13inference]{}. To show Condition Z(ii), we need to verify that for any $x_t \to 0$ such that $x_t \in \Psi(\Xi,y)$, $d_H(\Psi^{-1}(x_t,y) , \Psi^{-1}(0, y)) \to
0 $, where $d_H$ is the Hausdorff distance, uniformly in $y \in \mathcal{Y}$. Suppose by contradiction that this is not true, then there is $(x_t, y_t)$ with $x_t \to 0$ and $y_t \in \mathcal{Y}$ such that $d_H(\Psi^{-1}(x_t,y_t)
, \Psi^{-1}(0, y_t)) \not \to 0.$ By compactness of $\mathcal{Y}$, we can select a further subsequence $(x_k,y_k)$ such that $y_k \to y$, where $y \in
\mathcal{Y}$. We have that $\Psi^{-1}(0, y)=\xi_y$ is continuous in $y
\in \mathcal{Y}$, so we must have $d_H(\Psi^{-1}(x_k,y_k) , \Psi^{-1}(0, y))
\not \to 0.$ Hence, by compactness of $\Xi$, there is a further subsequence $u_l \in
\Psi^{-1}(x_l,y_l) $ with $u_l \to u$ in $\Xi$, such that $u \neq \Psi^{-1}(0, y)=\xi_y$, and such that $x_l= \Psi(u_l,y_l)
\to 0$. But, by continuity $\Psi(u_{l}, y_{l}) \to
\Psi(u,y) \neq 0$ since $u \neq \Psi^{-1}(0, y)$, yielding a contradiction.
Proof of Theorem \[thm:fclt\]
-----------------------------
We only consider the case where $\mathcal{Y}$ is a compact interval of $\mathbb{R}$. The case where $\mathcal{Y}$ is simpler. The proof follows the same steps as the proof of Theorem 5.2 of [chernozhukov+13inference]{} for the DR-estimator without sample selection using Lemma \[cont inverse\] in place of Lemma E.1 of [chernozhukov+13inference]{}. Let $\Psi(\xi,y) = P[\varphi_{y,\xi}]$ and $\widehat \Psi(\xi,y) = P_n[\varphi_{y,\xi}]$, where $P_n$ is the empirical measure and $P$ is the corresponding probability measure. From the first order conditions, the two-step estimator obeys $\widehat \xi_y = \phi(\widehat \Psi(\cdot,y),0)$ for each $y \in \mathcal{Y}$, where $\phi$ is the $Z$-map defined in Appendix E.1 of [chernozhukov+13inference]{}. The random vector $\widehat \xi_y$ is the estimator of $\xi_y = \phi(\Psi(\cdot,y),0)$ in the notation of this framework. Then, by step \[step1\] below, $$\sqrt{n} (\widehat \Psi - \Psi) \leadsto Z_{\Psi} \text{ in } \ell^{\infty}(\mathcal{Y} \times \mathbb{R}^{d_{\xi}} )^{d_{\xi}},\ Z_{\Psi}(y,\xi) = \mathbb{G} \varphi_{y,\xi},$$ where $d_{\xi} := \dim \xi_y$, $\mathbb{G} $ is a $P$-Brownian bridge, and $Z_{\Psi}$ has continuous paths a.s. Step \[step2\] verifies the conditions of Lemma \[cont inverse\] for $\dot{\Psi}(\xi_y,y) = J(y),$ the Hessian matrix defined in , which also implies that $y \mapsto \xi_y$ is continuously differentiable in the interval $\mathcal{Y}$. Then, by Lemma E.2 of [chernozhukov+13inference]{}, the map $\phi$ is Hadamard differentiable with derivative map $(\psi,0) \mapsto -J^{-1} \psi$ at $(\Psi,0)$. Therefore, we can conclude by the functional delta method that $$\label{eq:fclt}
\sqrt{n}(\widehat \xi_y - \xi_y) \leadsto Z_{\xi_y}:=- J^{-1}(y)Z_{\Psi}(y,\xi_y) \text{ in } \ell^{\infty}(\mathcal{Y})^{d_{\xi}},$$ where $y \mapsto Z_{\xi_y}$ has continuous paths a.s.
\[step1\] We verify that $\mathcal{G} = \{\varphi_{y,\xi}(W) : (y,\xi) \in \mathcal{Y} \times \mathbb{R}^{d_{\xi}} \}$ is $P$-Donsker with a square-integrable envelope. By inspection of the expression of $\varphi_{y,\xi}(W) = [ S_{1,\xi}(W)', S_{2y,\xi}(W)' ]'$ in Appendix \[app:scores\], $\varphi_{y,\xi}(W)$ is a Lipschitz transformation of VC functions with Lipschitz coefficient bounded by $c \| Z \|$ for some constant $c$ and envelope function $c \| Z \|$, which is square-integrable. Hence $\mathcal{G}$ is $P$-Donsker by Example 19.9 in [vdV]{}.
\[step2\] Conditions (a) and (b) are immediate by Assumption \[ass:fclt\]. To verify (c), note that for $(\tilde \xi,\tilde y)$ in the neighborhood of $(\xi_y,y)$, $$\frac{\partial \Psi(\tilde \xi,\tilde y)}{\partial (\tilde \xi', \tilde y)} = [J(\tilde \xi,\tilde y), R(\tilde \xi,\tilde y)],$$ where $$R(\tilde \xi,\tilde y) = - {{\mathrm{E}}}\left\{
\begin{array}{ccc}
0 \\
f_{Y \mid Z, D}(\tilde y \mid Z,1) \Phi_{\pi}(Z) \Phi_{\tilde \pi}(Z) \left[
\begin{array}{ccc}
G_{2,\tilde \xi}(Z) \\
G_{3,\tilde \xi}(Z)
\end{array}
\right] \otimes X
\end{array}
\right\},$$ for $\tilde \xi = (\tilde \pi', \tilde \beta', \tilde \rho')'$, and $$J(\tilde \xi,\tilde y) =
\left[
\begin{array}{ccc}
J_{11}(\tilde \xi,\tilde y) & J_{12}(\tilde \xi,\tilde y) \\
J_{21}(\tilde \xi,\tilde y) & J_{22}(\tilde \xi,\tilde y)
\end{array}
\right],$$ for $$J_{11}(\tilde \xi,\tilde y) = {{\mathrm{E}}}\left[\{g_1(Z'\tilde \pi)(D - \Phi_{\tilde \pi}(Z)) - G_1(Z'\tilde \pi) \phi(Z'\tilde \pi) \} ZZ' \right],$$ with $g_1(u) = d G_1(u)/ du$; $
J_{12}(\tilde \xi,\tilde y) = 0;
$ $$\begin{gathered}
J_{21}(\tilde \xi,\tilde y) = {{\mathrm{E}}}\left\{ [\Phi_{\pi}(Z) \Phi_{2,\tilde \xi}^{\nu}(Z) - \phi(Z'\pi)\Phi_{2,\xi_{\tilde y}}(Z) ] \left[
\begin{array}{ccc}
G_{2,\tilde \xi}(Z)\\
G_{3,\tilde \xi}(Z)
\end{array}
\right] \otimes XZ'\right\} \\
+ {{\mathrm{E}}}\left\{ (\Phi_{\pi}(Z)\Phi_{2,\tilde \xi}(Z) - \Phi_{\tilde \pi}(Z)\Phi_{2, \xi_{\tilde y}}(Z))
\left[
\begin{array}{ccc}
G_{2,\tilde \xi}^{\nu}(Z) \\
\rho'(X'\tilde \delta) G_{3,\tilde \xi}^{\nu}(Z)
\end{array}
\right] \otimes XZ'\right\},\end{gathered}$$ with $G_{j,\tilde \xi}^{\nu}(Z) := G_j^{\nu}\left(-X^\prime \tilde \beta,Z^\prime \tilde \pi; -\rho(X'\tilde \delta)\right)$ and $G_j^{\nu}(\mu,\nu;\rho) = \partial G_j(\mu,\nu;\rho)/\partial \nu$ for $j \in \{2,3\}$; and $$\begin{gathered}
J_{22}(\tilde \xi,\tilde y) = - {{\mathrm{E}}}\left\{ \Phi_{\pi}(Z) \left[
\begin{array}{ccc}
\Phi_{2,\tilde \xi}^{\mu}(Z) G_{2,\tilde \xi}(Z) & \Phi_{2,\tilde \xi}^{\rho}(Z) G_{2,\tilde \xi}(Z)\\
\Phi_{2,\tilde \xi}^{\mu}(Z) \rho'(X'\tilde \delta) G_{3,\tilde \xi}(Z) & \Phi_{2,\tilde \xi}^{\rho}(Z) \rho'(X'\tilde \delta) G_{3,\tilde \xi}(Z)
\end{array}
\right] \otimes XX' \right\} \\
+ {{\mathrm{E}}}\left\{ (\Phi_{\pi}(Z)\Phi_{2,\tilde \xi}(Z) - \Phi_{\tilde \pi}(Z)\Phi_{2, \xi_{\tilde y}}(Z))
\left[
\begin{array}{ccc}
G_{2,\tilde \xi}^{\mu}(Z) & G_{2,\tilde \xi}^{\rho}(Z) \\
\rho'(X'\tilde \delta) G_{3,\tilde \xi}^{\mu}(Z) & \rho'(X'\tilde \delta)^2 G_{3,\tilde \xi}^{\rho}(Z) + \rho''(X'\tilde \delta) G_{3,\tilde \xi}(Z)
\end{array}
\right] \otimes XX' \right\},\end{gathered}$$ with $ G_{j,\tilde \xi}^{a}(Z) := G_j^{a}\left(-X^\prime \tilde \beta,Z^\prime \tilde \pi; -\rho(X'\tilde \delta)\right) $ and $G_j^{a}(\mu,\nu;\rho) = \partial G_j(\mu,\nu;\rho)/\partial a$ for $j \in \{2,3\}$ and $a \in \{\mu,\rho\}$. In the previous expressions we use some notation defined in Appendix \[app:scores\].
Both $(\tilde \xi,\tilde y) \mapsto R(\tilde \xi,\tilde y)$ and $(\tilde \xi,\tilde y) \mapsto J(\tilde \xi,\tilde y)$ are continuous at $(\xi_y,y)$ for each $y \in \mathcal{Y}$. The computation above as well as the verification of the continuity follow from using the expressions of $\varphi_{y,\xi}$ in Appendix \[app:scores\], the dominated convergence theorem, and the following ingredients: (i) a.s. continuity of the map $(\tilde \xi,\tilde y) \mapsto \partial \varphi_{\tilde y,\tilde \xi}(W)/\partial \tilde \xi'$, (ii) domination of $\|\partial \varphi_{y,\xi}(W)/\partial \xi' \|$ by a square-integrable function $\| c Z\|$ for some constant $c$, (iii) a.s. continuity and uniform boundedness of the conditional density function $y \mapsto f_{Y \mid X,D}(y \mid X,1)$ by Assumption \[ass:fclt\], and (iv) $G_1(Z'\tilde \pi)$, $G_{2,\tilde \xi}(Z)$ and $G_{3,\tilde \xi}(Z)$ being bounded uniformly on $\tilde \xi \in \mathbb{R}^{d_{\xi}}$, a.s. By assumption, $J(y) = J(\xi_y,y)$ is positive-definite uniformly in $y \in \mathcal{Y}$.
The expressions of the limit processes given in the theorem follow by partitioning $Z_{\xi_y} = (Z_{\pi}', Z_{\theta_y}')'$ and using the expressions of $J^{-1}(y)$ and ${{\mathrm{E}}}[\varphi_{y,\xi}(W) \varphi_{\tilde y,\xi}(W)'] $ given in and .
Proof of Theorem \[thm:bfclt\]
------------------------------
Let $\widehat \xi_y^b := (\widehat \pi^{b'}, \widehat \theta_y^{b'})'$. By definition of the multiplier bootstrap draw of the estimator $$\sqrt{n}(\widehat \xi_y^b - \widehat \xi_y) = \mathbb{G}_n \omega^b \varphi_{y,\widehat \xi} = \mathbb{G}_n \omega^b \varphi_{y, \xi} + r_y,$$ where $\omega^b \sim N(0,1)$ independently of the data and $r_y := \mathbb{G}_n \omega^b (\varphi_{y,\widehat \xi} - \varphi_{y, \xi})$. Then the result follows from $\mathbb{G}_n \omega^b \varphi_{y, \xi} \leadsto_{{{\mathrm{P}}}} Z_{\xi_y}$ in step \[step1b\] and $r_y \leadsto_{{{\mathrm{P}}}} 0$ in step \[step2b\].
\[step1b\] Recall that $\varphi_{y,\xi}$ is $P$-Donsker by step \[step1\] of the proof of Theorem \[thm:fclt\]. Then, by ${{\mathrm{E}}}\omega^b = 0$, ${{\mathrm{E}}}(\omega^b)^2 = 1$ and the Conditional Multiplier Functional Central Limit Theorem [@vdV-W Theorem 2.9.6], $$\mathbb{G}_n \omega^b \varphi_{y, \xi} \leadsto_{{{\mathrm{P}}}} Z_{\xi_y},$$ where $Z_{\xi_y}$ is the same limit process as in .
\[step2b\] Note that $r_y \leadsto 0$ because $\varphi_{y,\xi}$ is $P$-Donsker and $\sqrt{n}(\widehat \xi_y - \xi_y) = O_{{{\mathrm{P}}}}(1)$ uniformly in $y \in \mathcal{Y}$ by Theorem \[thm:fclt\]. To show that $r_y \leadsto_{{{\mathrm{P}}}} 0$, we use that this statement means that for any $\epsilon > 0$, ${{\mathrm{E}}}^b1(\| r_y\|_2 > \epsilon) = o_{{{\mathrm{P}}}}(1)$ uniformly in $y \in \mathcal{Y}$. Then, the result follows by the Markov inequality and $${{\mathrm{E}}}{{\mathrm{E}}}^b1(\| r_y\|_2 > \epsilon) = {{\mathrm{P}}}(\| r_y\|_2 > \epsilon) = o(1),$$ uniformly in $y \in \mathcal{Y}$, where the latter holds by the Law of Iterated Expectations and $r_y \leadsto 0$.
Expressions of the Score and Expected Hessian {#app:scores}
=============================================
Score
-----
Let $\Phi_{\pi}(Z) := \Phi(Z'\pi)$ and $\Phi_{2,\xi_y}(Z) := \Phi_2\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$. Note that by the properties of the standard bivariate normal distribution $\Phi_2\left(X^\prime \beta(y),Z^\prime \pi; \rho(X'\delta(y))\right) = \Phi_{\pi}(Z) - \Phi_{2,\xi_y}(Z)$. Then, straighforward calculations yield $$S_{1,\xi}(W) = \frac{\partial \ell_{1,\xi}(W)}{\partial \pi} = G_1(Z'\pi) [D - \Phi_{\pi}(Z)] Z,$$ where $G_1(u) = \phi(u)/[\Phi(u)\Phi(-u)]$, and $$S_{2y,\xi}(W) = \frac{\partial \ell_{2y,\xi}(W)}{\partial \theta_y} = D(\Phi_{2,\xi_y}(Z) - \Phi_{\pi}(Z) I_y)
\left[
\begin{array}{ccc}
G_{2,\xi_y}(Z) \\
\rho'(X'\delta(y)) G_{3,\xi_y}(Z)
\end{array}
\right] \otimes X,$$ where $G_{2,\xi_y}(Z) := G_2\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$ and $G_{3,\xi_y}(Z) := G_3\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$ with $$G_2(\mu,\nu; \rho) = \frac{\Phi_2^{\mu}(\mu,\nu;\rho)}{\Phi_2(\mu,\nu;\rho)[\Phi(\nu) - \Phi_2(\mu,\nu;\rho)]}, \ \
G_3(\mu,\nu; \rho) = \frac{\Phi_2^{\rho}(\mu,\nu;\rho)}{\Phi_2(\mu,\nu;\rho)[\Phi(\nu) - \Phi_2(\mu,\nu;\rho)]},$$ for $$\label{eq:phi2mu}
\Phi_2^{\mu}(\mu,\nu;\rho) = \frac{\partial \Phi_2(\mu,\nu;\rho)}{\partial \mu} = \Phi\left(\frac{\nu - \rho \mu}{\sqrt{1 - \rho^2}} \right) \phi(\mu),$$ and $$\label{eq:phi2rho}
\Phi_2^{\rho}(\mu,\nu;\rho) = \frac{\partial \Phi_2(\mu,\nu;\rho)}{\partial \rho} = \phi_2(\mu,\nu;\rho).$$
To show and , start from the factorization $$\Phi_2(\mu,\nu;\rho) = \int_{-\infty}^{\mu} \Phi\left(\frac{\nu-\rho v}{\sqrt{1-\rho^2}}\right)\phi(v) dv.$$ Then, follows from taking the partial derivative with respect to $\mu$ using the Leibniz integral rule. Taking the partial derivative with respect to $\rho$ yields $$\begin{aligned}
\frac{\partial \Phi_2(\mu,\nu;\rho)}{\partial \rho} &= \int_{-\infty}^{\mu} \phi\left(\frac{\nu-\rho v}{\sqrt{1-\rho^2}}\right)\frac{\rho \nu - v}{(1-\rho^2)^\frac{3}{2}}\phi(v)dv \\
&= \int_{-\infty}^\mu \frac{1}{\sqrt{2\pi}}\exp{\left[-\frac{(\nu-\rho v)^2}{2(1-\rho^2)}\right]} \frac{1}{\sqrt{2\pi}}\exp{\left[-\frac{v^2}{2}\right]} \frac{\rho \nu - v}{(1-\rho^2)^\frac{3}{2}} dv \\
&= \int_{-\infty}^\mu \frac{\rho \nu -v}{2\pi(1-\rho^2)^\frac{3}{2}} \exp{\left[-\frac{\nu^2-2\rho v\nu +v^2}{2(1-\rho^2)}\right]}dv \\
&= \frac{1}{2\pi\sqrt{1-\rho^2}}\exp{\left[-\frac{\nu^2-2\rho \mu\nu +\mu^2}{2(1-\rho^2)}\right]} = \phi_2(\mu,\nu;\rho)
\end{aligned}$$
Expected Hessian
----------------
Straighforward calculations yield $$H_1 = {{\mathrm{E}}}\left[ \frac{\partial \ell_{1,\xi}(W)}{\partial \pi \partial \pi'} \right] = - {{\mathrm{E}}}\left[ G_1(Z'\pi) \phi(Z'\pi) ZZ'\right], \ \ {{\mathrm{E}}}\left[ \frac{\partial \ell_{1,\xi}(W)}{\partial \pi \partial \theta_y'} \right] = 0,$$ $$J_{21y} = \frac{\partial \ell_{2y,\xi}(W)}{\partial \theta_y \partial \pi'} = {{\mathrm{E}}}\left\{ [\Phi_{\pi}(Z) \Phi_{2,\xi_y}^{\nu}(Z) - \phi(Z'\pi)\Phi_{2,\xi_y}(Z) ] \left[
\begin{array}{ccc}
G_{2,\xi_y}(Z) \\
\rho'(X'\delta(y)) G_{3,\xi_y}(Z)
\end{array}
\right] \otimes XZ'\right\},$$ where $\Phi_{2,\xi_y}^{\nu}(Z) = \Phi_2^{\nu}\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$ with $$\Phi_2^{\nu}(\mu,\nu;\rho) = \frac{\partial \Phi_2(\mu,\nu;\rho)}{\partial \nu} = \Phi\left(\frac{\mu - \rho \nu}{\sqrt{1 - \rho^2}} \right) \phi(\nu),$$ by a symmetric argument to , and $$H_{2y} = \frac{\partial \ell_{2y,\xi}(W)}{\partial \theta_y \partial \theta_y'} = - {{\mathrm{E}}}\left\{ \Phi_{\pi}(Z) \left[
\begin{array}{cc}
\Phi_{2,\xi_y}^{\mu}(Z) G_{2,\xi_y}(Z) & \Phi_{2,\xi_y}^{\rho}(Z) G_{2,\xi_y}(Z) \\
\Phi_{2,\xi_y}^{\mu}(Z) \rho'(X'\delta(y)) G_{3,\xi_y}(Z) & \Phi_{2,\xi_y}^{\rho}(Z) \rho'(X'\delta(y)) G_{3,\xi_y}(Z)
\end{array}
\right] \otimes XX' \right\},$$ where $\Phi_{2,\xi_y}^{\mu}(Z) := \Phi_{2}^{\mu}\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$ and $\Phi_{2,\xi_y}^{\rho}(Z) := \Phi_{2}^{\rho}\left(-X^\prime \beta(y),Z^\prime \pi; -\rho(X'\delta(y))\right)$.
[lcc|lcc]{} Variable & Male & Female & Variable & Male & Female\
educ16 & 0.25 & 0.06 & numch34 & -0.18 & -0.63\
& (0.01) & (0.01) & & (0.02) & (0.01)\
educ1718 & 0.46 & 0.20 & numch510 & -0.18 & -0.33\
& (0.02) & (0.01) & & (0.01) & (0.01)\
educ1920 & 0.42 & 0.16 & numch1116 & -0.16 & -0.15\
& (0.03) & (0.02) & & (0.01) & (0.01)\
educ2122 & 0.74 & 0.28 & numch1718 & -0.02 & -0.11\
& (0.02) & (0.02) & & (0.03) & (0.02)\
educ23 & 0.51 & 0.15 & benefit & -0.35 & -0.42\
& (0.02) & (0.02) & & (0.01) & (0.01)\
couple & -4.02 & -8.14 & benefit$\times$couple & 0.87 & 1.40\
& (0.09) & (0.08) & & (0.02) & (0.02)\
numch1 & -0.16 & -0.90 & constant & 2.50 & 2.75\
& (0.02) & (0.02) & & (0.08) & (0.07)\
numch2 & -0.18 & -0.77 &\
& (0.02) & (0.02) &\
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 2 for men[]{data-label="fig:coef_men_spec2"}](Estimates_11378_couple_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 2 for women[]{data-label="fig:coef_women_spec2"}](Estimates_11378_couple_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 3 for men[]{data-label="fig:coef_men_spec3"}](Estimates_11378_lintime_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 3 for women[]{data-label="fig:coef_women_spec3"}](Estimates_11378_lintime_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 4 for men[]{data-label="fig:coef_men_spec4"}](Estimates_11378_lintimexcouple_500.pdf){height="0.75\textheight" width="85.00000%"}
![Estimates and 95% confidence bands for coefficients of education and marital status in the outcome equation: specification 4 for women[]{data-label="fig:coef_women_spec4"}](Estimates_11378_lintimexcouple_500.pdf){height="0.75\textheight" width="85.00000%"}
{height="0.75\textheight" width="85.00000%"}
{height="0.75\textheight" width="85.00000%"}
{height=".85\textwidth" width="49.00000%"} {height=".85\textwidth" width="49.00000%"}
{height=".75\textwidth" width="49.00000%"} {height=".75\textwidth" width="49.00000%"}
{height="49.00000%" width="49.00000%"} {height="49.00000%" width="49.00000%"}
{height="49.00000%" width="49.00000%"} {height="49.00000%" width="49.00000%"}
{height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"}
{height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"}
{height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"} {height="33.00000%" width="32.00000%"}
[c c | c c]{}\
& &\
& & Male & Female\
& & 83 & 59\
& & (82, 83) & (59, 59)\
& & 83 & 66\
& & (83, 83) & (66, 66)\
![Estimates and 95% confidence bands for decomposition between men and women with aggregated selection effects in specification 1[]{data-label="fig:adec_spec1"}](CondDFQF_11378_500.pdf){height=".49\textwidth" width=".49\textwidth"}
{height=".32\textwidth" width=".32\textwidth"} {height=".32\textwidth" width=".32\textwidth"} {height=".32\textwidth" width=".32\textwidth"}
{height=".32\textwidth" width=".32\textwidth"} {height=".32\textwidth" width=".32\textwidth"} {height=".32\textwidth" width=".32\textwidth"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}\
{height="45.00000%" width="45.00000%"} {height="45.00000%" width="45.00000%"}
[^1]: See [arellano17]{} for a recent survey on sample selection in quantile models.
[^2]: Finally, from a technical point of view, [ab17]{} only derived pointwise limit theory for the estimators of the model parameters, whereas we derive functional limit theory for the estimators of the model parameters and related functionals.
[^3]: See [survey18]{} for a recent survey on measures of local dependence.
[^4]: This concept is different from the local independence of [doksum94]{}, which is local in only one of the variables. Thus, for example, our concept is symmetric in $Y^*$ and $D^*$ whereas the concept in [doksum94]{} is not.
[^5]: Note that the marginals of $F_{Y^*,D^*}$ do not identify separately the mean and variances of the local Gaussian representation.
[^6]: [kitagawa2010]{} developed a test for the outcome exclusion.
[^7]: The minus signs in are included to take into account that the selection is defined by $D^* > 0$ instead of $D^* {\leqslant}0$. We use this definition to facilitate the interpretation of the parameters and the comparison with the classical Heckman selection model; see Example \[ex:hsm\].
[^8]: Note that in Example \[ex:hsm\] the equation for $Y^*$ can be written as $0= (X'\beta - Y^*)/\sigma + \rho V+\sqrt{1-\rho^2} \widetilde{U}$, where $\widetilde{U}$ is standard normally distributed and independent of $V$ and $Z$.
[^9]: To obtain the derivative we use that $\Phi_2\left( -\beta, \pi; -\rho \right) = \int_{-\infty}^{\pi} \Phi\left( \frac{-\beta + \rho x}{\sqrt{1-\rho^2}}\right) \phi(x) dx$.
[^10]: If the support of $Y$ is finite, $\mathcal{Y}$ can be the entire support, otherwise $\mathcal{Y}$ should be a subset of the support excluding low density areas such as the tails.
[^11]: If the function $y \mapsto {\widehat}F_{Y^*}(y)$ is not increasing, we monotonize it before taking the left-inverse. See Section \[sec:theory\] for details.
[^12]: See [rv18]{} for another recent application of the data to the analysis of female labour force participation.
[^13]: For data before 1990, $D=0$ if the individual is in one of the following status: seeking work, sick but seeking work, sick but not seeking work, retired and unoccupied. For those in and after 1990, $D=0$ if the individual is seeking work and available, waiting to start work, sick or injured, retired or unoccupied.
[^14]: In the rest of the paper we shall refer to an individual being married or cohabiting as married.
[^15]: Our definition of the out-of-work benefit income is slightly different from the definition of [brs03]{} and [bgim07]{}. They calculated it as the income of a tax unit if all the individuals within the tax unit were out of work. In our view our definition might better reflect the opportunity cost or outside value option of working that the individual faces.
[^16]: The main results on the coefficients of the wage equation and wage decompositions presented below are not sensitive to the specification of the sorting equation.
[^17]: We report estimates of $\beta/\sigma$ in the outcome equation of the HSM for comparability; see Example \[ex:hsm\].
[^18]: We find more heterogeneity in the coefficient of the marital status indicator in the specifications 2 and 4 that include marital status in the selection sorting function.
[^19]: We do not report confidence bands for specifications 3 and 4 to avoid cluttering. The confidence bands for the coefficients of the selection sorting function in the SM show that the results on the trends are statistically significant.
[^20]: The model-based estimator of the observed distribution in produces almost identical estimates to the empirical distribution of the observed wages.
[^21]: In results not reported, we find that the negative sorting for married women is robust to the definition of the out-of-work benefit income variable. Thus, we find similar estimates using the income of a tax unit if all the individuals within the tax unit were out of work as the excluded covariate, as in [bgim07]{}.
[^22]: These results are available from the authors upon request.
[^23]: We find similar results for the other coefficients of the outcome equation. We do not report these reports for the sake of brevity.
[^24]: We assume that $D^*$ is absolutely continuous with strictly increasing distribution. This assumption is without loss of generality because the distribution of $D^*$ is only identified at $D^*=0$.
[^25]: Note that $v \mapsto \Phi_2(\cdot, \Phi^{-1}(v); \rho(\cdot,0))$ is a real analytic function.
[^26]: Under joint normality of $Y^*$ and $R^*$, the identification conditions of [ab17]{} require the same independence restrictions.
[^27]: We adapt the notation of [chernozhukov+13inference]{} to our problem by using $y$, $\mathcal{Y}$, $\xi_y$, $d_{\xi}$ and $\Xi$ in place of $u$, $\mathcal{U}$, $\theta_0(u)$, $p$, and $\Theta$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the metallicities and kinematics of nearby stars known to have planetary-mass companions in the general context of the overall properties of the local Galactic Disk. We have used Strömgren photometry to determine abundances for both the extrasolar-planet host stars and for a volume-limited sample of 486 F, G and K stars selected from the Hipparcos catalogue. The latter data show that the Sun lies near the modal abundance of the disk, with over 45% of local stars having super-solar metallicities. Twenty of the latter stars (4.1%) are known to have planetary-mass companions. Using that ratio to scale data for the complete sample of planetary host stars, we find that the fraction of stars with extrasolar planets rises sharply with increasing abundance, confirming previous results. However, the frequency remains at the 3-4% level for stars within $\pm0.15$ dex of solar abundance, and falls to $\sim1\%$ only for stars with abundances less than half solar. Given the present observational constraints, both in velocity precision and in the available time baseline, these numbers represent a lower limit to the frequency of extrasolar planetary systems. A comparison between the kinematics of the planetary host stars and a representative sample of disk stars suggests that the former have an average age which is $\sim60\%$ of the latter.'
author:
- 'I. Neill Reid'
title: On the Nature of Stars with Planets
---
Introduction
============
The discovery of the first extrasolar planetary system stands as one of the key scientific and philosophical advances of the twentieth century. While the existence of other planetary systems had been postulated for several centuries (Dick, 1998), and could even be regarded as likely, particularly following the detection of circumstellar disks around young stars (see Sargent & Beckwith, 1993), Mayor & Queloz’ (1995) radial velocity measurements of 51 Pegasi marked a definitive transition from speculation to observation. The relatively short time interval which has elapsed since that initial discovery has seen the identification of a plethora of additional systems, notably by Marcy, Butler and collaborators. Taken together, those systems provide sufficient numbers for a statistical comparison of the characteristics of stars with planetary-mass companions against the overall distribution of properties of the local Galactic Disk. The results of such a study have obvious implications for estimating the likely frequency of extrasolar planets (ESPs), particularly potentially habitable systems.
Comparative studies of this type must pay due regard to several important caveats. First, it is clear that most of the ESP systems discovered to date bear little resemblance to our own Solar System: 51 Pegasi-like systems feature ‘hot jupiters’, Jovian-mass planets in sub-Mercurian orbits, while over half of the current ESP catalogue have orbital eccentricities comparable to, or exceeding, that of Mercury and Pluto. Those circumstances, however, may at least partly reflect observational selection; these systems have relatively short periods and relatively high velocity amplitudes, and are therefore the easiest to detect. All of the ‘hot jupiter’ ESPs have reflex motions of tens of ms$^{-1}$, and it seems likely that we have a fairly complete census of these objects. However, it is only now that observations are achieving both the requisite velocity precision and the decade-plus time baselines which are required for the detection of Jovian analogues, and systems bearing a closer resemblance to the Solar System are starting to emerge amongst the most recent discoveries (Vogt [*et al.*]{}, 2001). Thus, it is possible that the properties of the current ESP catalogue may reflect extreme, rather than characteristics, systems.
By the same token, it seems likely that the present catalogue includes only a subset of extrasolar planetary systems in the Solar Neighbourhood. Studies estimate that between 3 and 5% of F, G-type stars have currently-detectable ESP systems (Marcy & Butler, 2000). Tabachnik & Tremaine (2001), in particular, have used maximum-likelihood analysis to estimate that current observations indicate a planetary frequency of 3% amongst solar-type stars, but that the frequency might be as high as 15% if the companion mass function is extrapolated to terrestial-mass systems. Thus, the observed detection frequency may well underestimate the true frequency of solar-type stars with planetary systems, and possibly provides a biased sampling of their characteristics.
Nonetheless, the current dataset offers a first cut at determining the conditions required for the formation of planetary systems. How are the ESP primaries distinguished from the average local field star? Studies to date have focused on chemical abundance, with strong indications that stars known to have planets tend to have solar or super-solar metallicity (Gonzalez, 1998; Santos et al, 2001). While this may indicate a requirement on the initial conditions at formation, there have also been suggestions that these higher abundances are a consequence of planet formation (Lin [*et al.*]{}, 1996), reflecting pollution of the stellar atmosphere by migrating gas giants (Gonzalez, 1997; Laughlin, 2000).
Placing this result in the broadest context requires consideration of both correlations which might exist with other properties of the planetary host stars, and comparison against data for a reliable reference sample of representative disk stars. The latter criterion is not met in some recent analyses. In this paper we re-examine the abundance distribution of the ESP hosts, matched against similar data for an Hipparcos-based, volume-limited sample of FGK stars. We also compare the kinematics of ESP hosts against the velocity distribution of local disk stars. The paper is organised as follows: the following section presents basic data for the ESP host stars; section 3 discusses abundance calibration and the metallicity distribution; section 4 examines the kinematics of the sample; and section 5 summarises our main conclusions.
The planetary hosts in the HR diagram
=====================================
The sample
----------
Table 1 lists basic photometric and parallax data for stars currently known to possess at least one planetary-mass companion. We shall refer to those stars as ESP host stars. In compiling this list, we follow the Geneva convention ([*http://obswww.unige.ch/ udry/planet/*]{}) of setting an upper mass limit of M$_2\sin{i} = 17$M$_J$, where M$_J$ is the mass of Jupiter. There are only four systems where M$_2\sin{i}$ exceeds 10M$_J$. The parameters listed for the planetary systems are taken from the Extrasolar Planets Encyclopedia maintained by J. Schneider at [*http://cfa-www.harvard.edu/planets/*]{}.
Since we only measure M$_2\sin{i}$ for most of these systems, there is clearly potential for the inclusion of higher-mass companions on low-inclination orbits, either low-mass stars or brown dwarfs. Indeed, there may well be an overlap between the upper mass range of planets and the lower mass-range of brown dwarfs[^1], leading to an inherent ambiguity in interpretation. Since those two classes of objects may have different intrinsic properties, it is important to conside the likely level of cross-contamination.
The degree of contamination depends on the prevalence of brown dwarfs as close companions to solar-type stars. Few unequivocal examples of such systems have been detected, leading both to the postulation of a ‘brown dwarf desert’ at $a < 10$ a.u. (Marcy & Butler, 2000), and support for the hypothesis that the low mass (M$_2\sin{i}<10$M$_J$) companions that [*are*]{} detected are not a simple extension of the companion mass function at higher masses. On the other hand, there have been counter suggestions. Both Heacox (1999) and Stepinski & Black (2000) have pointed out the similarity between the orbital properties of planetary-mass systems and stellar binaries, although that might reflect similar dynamical evolution rather than similar origins. More directly, Han [*et al.*]{} (2000) and Gatewood [*et al.*]{} (2000) have analysed radial velocity data in tandem with Hipparcos Intermediate Astrometric Data, and claim that in many cases the best-fit orbits have low inclination, and correspondingly high true masses in the brown dwarf or even stellar régime.
The Han [*et al.*]{} result has been scrutinised intently and generally found wanting. Statistically, the scarcity of systems with M$_2\sin{i} > 15$ M$_J$ demands a rather unlikely observational conspiracy, with orbital axes aligned within a few degrees of the line of sight. Under a random distribution of inclinations, one would expect several hundred systems with brown-dwarf mass companions for each planetary-mass system; not only are the latter systems not observed, but there are not sufficient G dwarfs in the Solar Neighbourhood to meet the numerical requirements. The amplitudes of the derived astrometric orbits are comparable with the uncertainties in the Hipparcos IAD measurements, and Pourbaix (2001) has shown that the low inclinations found by Han [*et al.*]{} are largely an artefact of the fitting technique used. Pourbaix & Arenou (2001) conclude that $\rho$ CrB is the only system where the data merit interpretation as a near face-on orbit, but HST astrometry by McGrath [*et al.*]{} (2001) sets an upper limit on the semi-major axis at $a\sin{i} < 0.3$ milliarcseconds (mas) corresponding to M$_2 < 30 $M$_J$, rather than the 1.5 mas and $0.14\pm0.05$M$_\odot$ suggested by Gatewood [*et al.*]{} (2000).
Current estimates of the companion mass function are generally in good agreement, finding approximately equal numbers per decade in mass (${dN \over dM} \propto M^{-1}$) for $M < 10$M$_J$, with a sharp drop at higher masses (Jorissen [*et al.*]{}, 2001; Zucker & Mazeh, 2001; Tabachnik & Tremaine, 2001). Indeed, astrometric observations have shown that many of the candidate brown dwarf companions are, in fact, low-mass M dwarfs (Halbwachs [*et al.*]{}, 2000), further enhancing the brown dwarf desert. Based on these results, we expect little contamination ($<$5%) in the ESP sample listed in Table 1. Nonetheless, one star deserves comment. As discussed later, HD 114762 is one of the most metal-poor stars in the sample. The measured stellar rotational velocity suggests that the star is being viewed close to pole-on, implying a low inclination for an equatorial orbit. In that case, the companion, detected by Latham [*et al.*]{} (1989) is a candidate for the first brown dwarf identification.
The (M$_V$, (B-V)) colour-magnitude diagram
-------------------------------------------
All of the stars except BD -10 3666 were observed by Hipparcos, and most have trigonometric parallaxes measured to an accuracy of better than 5%. The BV photometry listed in Table 1 is also taken from the Hipparcos catalogue (ESA, 1997), although we use the literature data cited therein in preference to Tycho photometry, given the systematic offset (and occasional large random errors) between the latter system and standard Johnson data (Bessell, 2000). BD -10 3666 has UBV photometry by Ryan (1992) but no measured trigonometric parallax, and the absolute magnitude listed in Table 1 is based on an estimated photometric parallax. Since both this star and HD 4203 were selected for observation based on the known high metallicity, neither plays a role in the statistical comparison discussed in the following sections.
The photometry and astrometry listed in Table 1 allow a detailed assessment of the distribution of the ESP host stars in the HR diagram. Figure 1 makes that comparison, where the reference main sequence is provided by data for stars within 25 parsecs which have accurate photometry and trigonometric parallaxes measured to a precision of ${\sigma_\pi \over \pi} < 10\%$. All of the latter stars are members of the Galactic Disk. Three features are immediately noticeable:
1. G dwarfs contribute the majority of planet detections, with Gl 876 still the only M dwarf known to have planetary-mass companions. To a large extent, the distribution likely reflects the continuing observational focus on solar-type stars.
2. The current sample includes a significant number of evolved stars. At least six stars lie at the base of the subgiant branch, while four have evolved a considerable distance from the main sequence. These stars are identified in Table 1.
3. The ESP host stars are distributed throughout the full width of the main sequence. This is important since chemical abundance is the dominant factor which governs the location of single stars on the main sequence - at a given colour, metal-rich stars are more luminous than metal-poor stars. Thus, the observed distribution points to a range of metallicity amongst stars with planets which is at least comparable to the abundance distribution amongst the underlying local disk (thin+thick) population.
The last point is particularly pertinent given the recent emphasis laid on the high metal abundance measured for at least some of the ESP primaries (Gonzalez [*et al.*]{}, 1999, 2001; Santos [*et al.*]{}, 2001). Many of those abundances are significantly higher than the value usually taken as the median for the Galactic Disk. Most previous studies, however, treat the metallicity distributions of the ESP hosts and of the disk as separate entities. The following section places the former in the context of the latter, and considers how the high chemical abundances are reconciled with the distribution evident in Figure 1.
Chemical abundances
===================
Measuring stellar metallicity
-----------------------------
The metal content of stellar atmospheres can be measured using a wide variety of techniques. In general, the accuracy of the final measurement is at least inversely proportional to the difficulty of the observation. Analyses based on high-resolution spectroscopy are usually more reliable than those which utilise broadband colours, but photometric data are obtained much more readily than echelle spectra. Thus, any statistical analysis requiring a dataset of even modest dimensions must balance two factors - availability and accuracy.
As a further complication, comparative studies must ensure that data drawn from different analyses are tied to a consistent system. Different measurement techniques not only have different random uncertainties, but can also exhibit systematic discrepanciesin scale and/or zeropoint, as discussed in the context of globular cluster distance determination by Gratton [*et al.*]{} 1997) and Reid (1998). Apart from differences in the choice of standard stars, we note that the solar iron abundance was re-calibrated relatively recently, revised downward from A(Fe)=7.67 to A(Fe)=7.54 (Biemont [*et al.*]{}, 1991), where A(Fe) is the logarithmic abundance on a scale where A(H)=12. While most abundances are measured differentially, this re-calibration might lead to a systematic offset depending on how and when the abundances of the standard stars were determined.
This issue is a concern since, while nearly all of the ESP host stars have recent high-resolution spectroscopic abundance measurements, most estimates of the underlying field-star metallicity distribution rest on lower resolution techniques. Santos [*et al.*]{} (2001) have addressed this problem to some extent by providing high-resolution spectral analyses for 43 G dwarfs in 42 systems drawn from the volume-limited sample of stars with (B-V)$< 1.1$ and $d < 17$ parsecs. That sample is scarcely sufficient in size, however, to provide an adequate mapping of the distribution of disk properties. We adopt an alternative strategy in this paper.
Strömgren uvby photometry provides a relatively simple means of determining abundances for F, G and early K-type stars (Strömgren, 1966). Data are available for most solar-type stars brighter than 10th magnitude. Indeed, both Giménez (2000) and Laughlin (2000) have applied those measurements to studying the ESP host stars. Metallicities are determined by measuring the differential line-blanketting via the m$_1$ and c$_1$ indices, where $$m_1 \ = \ (v-b) \ - \ (b-y)$$ and $$c_1 \ = \ (u-v) \ - \ (v-b)$$ The latter index is also gravity sensitive, allowing discrimination between dwarfs and subgiants.
The literature contains several calibrations of the Strömgren indices against metallicity. Giménez adopts that given by Olsen (1984); we follow Laughlin in using the more recent calibrations derived by Schuster & Nissen (1989). They provide two calibrating relations, $$\begin{aligned}
[Fe/H]_{uvby} & = & 1.052 - 73.21m_1 + 280.9m_1(b-y) + 333.95m_1^2(b-y) - 595.5m_1(b-y)^2 \\
& & + [5.486 - 41.61m_1 - 7.963(b-y)] \times \log{(m_1 - [0.6322 - 3.58(b-y) + 5.20 (b-y)^2])}\end{aligned}$$ for F stars, ($0.22 \le (b-y) < 0.375$, $0.03 \le m_1 \le 0.21$, $0.17 \le c_1 \le 0.58$ and $-3.5 \le [Fe/H] \le 0.2$), and $$\begin{aligned}
[Fe/H]_{uvby} & = & -2.0695 - 22.45m_1 - 53.8m_1^2 - 62.04m_1(b-y) + 145.5m_1^2(b-y) \\
& & (85.1m_1 - 13.8c_1 - 137.2m_1^2)c_1 \end{aligned}$$ for G stars ($0.375 \le (b-y) \le 0.59$, $0.03 \le m_1 \le 0.57$, $0.10 \le c_1 \le 0.47$ and $-2.6 \le [Fe/H] \le 0.4$).
Apart from BD -10 1366 and Gl 876, where Strömgren data offer little useful information, only HD 177830 (HIP 93746) lacks uvby photometry. Table 2 lists (b-y), m$_1$ and c$_1$ colour indices, taken from Hauck & Mermilliod’s (1998) catalogue, and the resulting \[Fe/H\]$_{uvby}$ for the remaining stars, together with the metallicities derived from high-resolution spectroscopy. As expected, there is a systematic offset to lower abundances in the Strömgren calibration. The results are compared in Figure 2, where the upper panels plot the full dataset (see also Figure 1 in Laughlin, 2000). We note that residuals tend to increase among early-type K stars, (b-y)$>$0.5, where both methods become more problematic.
We have quantified the comparison using the datasets from Santos [*et al.*]{} (2001) and Gonzalez and collaborators, which provide two internally consistent datasets of moderate size. Based on the twenty-two ESP host stars observed by Santos [*et al.*]{} (2001), we derive $$\delta [Fe/H] \ = \ [Fe/H]_{uvby} - [Fe/H]_{sp} \ = \ -0.117\pm0.095$$ where the uncertainties quoted are the rms dispersion about the mean. The formal standard error of the mean, $\sigma_\mu$, is 0.020 dex and the median offset is -0.13 dex. The twenty-two stars in the Gonzalez dataset give an almost identical result, $$\delta [Fe/H] \ = \ -0.118\pm0.113, \quad \sigma_\mu \ = \ 0.024$$ The median offset is -0.10 dex. Combining both datasets with forty field stars from the Santos [*et al.*]{} reference sample gives $$\delta [Fe/H] \ = \ -0.083\pm0.118, \quad \sigma_\mu \ = \ 0.013$$ The median offset is -0.07 dex. This is in excellent agreement with Gratton [*et al.’s*]{} independent analysis of 152 stars spanning a much larger range of metallicity, where they derive $$\delta [Fe/H] \ = \ -0.102\pm0.151, \quad \sigma_\mu \ = \ 0.012$$ Given uncertainties of $\pm0.06$ dex in the Santos [*et al.*]{} and Gonzalez spectroscopic measurements, the measured dispersion indicates that the Strömgren data have typical uncertainties of $\pm0.1$ dex at near-solar abundances, sufficient accuracy for present purposes.
Based on this comparison, we conclude that the Strömgren abundance scale is offset by -0.1 dex from the most recent calibrations. We note that the offset is intriguingly similar to the re-calibration of the solar abundance, although that similarity may be coincidental. Rather than attempt to correct the metallicity measurements, we base our analysis on the [*uvby*]{} scale; in effect, we adopt the somewhat paradoxical definition $$[Fe/H]_{uvby} (\odot) \ = \ -0.1$$ Since we are considering the comparative distributions of field star and ESP host star abundances, consistency is more important than the numerical value chosen for the fiducial zeropoint.
The reference sample
--------------------
Defining a suitable reference sample is crucial to assessing how the properties of the ESP host stars map onto the overall field star distribution. Complete, volume-limited datasets offer the most reliable comparison, but have generally not been available to previous studies. Murray [*et al.*]{} (2001) have attempted to turn Hipparcos data to this end, with a reference sample defined by selecting HD stars with $\pi > 10$ mas and ${\sigma_\pi \over \pi} < 10\%$. However, the colour-magnitude diagram for this dataset (Figure 1 in Murray & Chaboyer, 2001) is clearly biased strongly towards F-type and early G-type dwarfs, partly reflecting sampling in the Hipparcos catalogue at $V > 9$th magnitude, and partly reflecting the fact that the HD catalogue was selected from blue photographic plates. The resulting dataset therefore provides a biased subset of nearby disk stars. Murray [*et al.’s*]{} metallicities are taken from the Cayrel de Strobel (1997) catalogue, and therefore represent an amalgam of heterogenous sources. Thus, this dataset is not suitable as a local reference.
Several other analyses (e.g. Gonzalez, 1999; Butler [*et al.*]{}, 2000) have relied on the nearby-star sample defined by Favata [*et al.*]{} (1997: F97) to represent the abundance distribution of local disk stars. That sample, however, is severely flawed in several important respects, as illustrated in Figure 3. The F97 abundances are derived from high-resolution spectra, and a comparison with Strömgren abundances gives $$\delta [Fe/H] \ = \ -0.05\pm0.15, \ {\rm 69 \ stars}$$ somewhat less than the offsets derived in the previous section. The full sample of 90 stars includes both components of several binaries, giving those systems double weight in the abundance distribution, and spans a substantially larger colour range than ESP host stars (excluding Gl 876). Finally, and most significantly, the sample is neither complete nor volume-limited. Favata [*et al.*]{} (1996) constructed the original sample by taking a randomly-selected subset of 200 stars with $0.5 < (B-V) < 1.4$ from the second Catalogue of Nearby Stars (Gliese, 1969; Gliese & Jahrei[ß]{}, 1979: CNS2). Ninety-four of those stars were observed spectroscopically. Unfortunately, while the CNS2 has a nominal distance limit of 25 parsecs, subsequent Hipparcos astrometry has shown that a significant fraction of the stars lie at much larger distances. Figure 3 shows that at least 40% of the F97 dataset lies beyond 25 parsecs. Our analysis demands a more reliable reference dataset.
Figure 4 shows the (M$_V$, (B-V)) colour-magnitude outlined by the 1549 stars in the Hipparcos catalogue with formal trigonometric parallax measurements exceeding 40 mas ($d \le 25$ parsecs, (m-M)$\le 1.99$). As in Figure 1, we use the literature BV photometry in the catalogue in preference to Tycho data. We have not applied any cut based on parallax precision, but use different symbols to identify the 1477 stars with parallaxes measured to a precision of better than 20%. Nearly all of the stars lying below the main-sequence in this figure have inaccurate parallax measurements.
The box superimposed on Figure 4 isolates 488 stars with $0.5 \le {\rm (B-V)} \le 1.0$ and $2.0 \ge {\rm M}_V \ge 7.0$, matching the colour/magnitude range of the bulk of the ESP host stars. The limiting magnitude for completeness in the Hipparcos catalogue is $$V \ = \ 7.9 \ + \ 1.1 sin|b|$$ so the 25-parsec sample is effectively complete over the whole sky for M$_V \le 5.9$ (the dotted line in Figure 4). The full catalogue is $\sim25\%$ incomplete for stars with $8 \le {\rm V} < 9$, but should be significantly more complete for nearby stars, since all stars suspected of being within 25 parsecs were included in the input catalogue - as emphasised by the large numbers of M dwarfs in Figure 4. Indeed, Jahrei[ß]{} & Wielen (1997) argue that the Hipparcos catalogue is essentially complete to M$_V=8.5$ for stars within 25 parsecs of the Sun. Thus, the F, G and K stars isolated in Figure 4 effectively represent a complete, volume-limited sample. We identify these stars as the FGK25 Hipparcos dataset. Note that the sample includes 20 known ESP hosts from Table 1.
As noted in the previous section, Strömgren photometry is now available for a substantial number of bright F, G and K stars, and is readily accessible through the catalogue compiled by Hauck & Mermilliod (1998). We have cross-referenced the FGK25 Hipparcos dataset against that catalogue and located photometry for 419 of the 486 stars - 86% of the sample. Those stars are identified as solid squares in Figure 4. The sample includes members of binary systems, but only one star per system. It is clear that almost every star with (B-V)$<$0.8 has photometry, while the late-G and K dwarfs which lack Strömgren measurements are distributed over the full width of the main sequence, interspersed with stars which have uvby photometry. We conclude that the abundance distribution deduced from these data is characteristic of the parent population(s) of the ESP host stars.
The abundance distribution of local field stars
-----------------------------------------------
The overwhelming majority of the stars in the FGK25 sample are members of the Galactic Disk. Five stars, however, lie significantly below the main sequence. These are HIP 57939, 62951, 67655, 79537 and 79979. Two of these stars, HIP 62951 and HIP 79979, are in binary systems where the companion has affected the parallax determination; Fabricius & Makarov (2000) have reanalysed the Hipparcos data for HIP 62951 and find $\pi=2.4$ mas. These are the only two stars in the FGK25 sample with ${\sigma_\pi \over \pi} > 0.2$ (there are only 6 other stars with ${\sigma_\pi \over \pi} > 0.1$), and we exclude both from our sample. The remaining three subluminous stars are [*bona-fide*]{} metal-poor subdwarfs. HIP 57939 is HD 103095, or Groombridge 1830, the well-known intermediate-abundance (\[Fe/H\]$_{uvby}$=-1.4) subdwarf; HIP 67655, or HD 120559, has \[Fe/H\]$_{uvby}$=-0.94; and HIP 79537 is HD 145417, \[Fe/H\]$_{uvby}$=-1.25. The presence of three such stars in this volume-limited sample suggests a somewhat higher density normalisation ($(\sim0.6\pm0.35)\%$ relative to the disk population) than usually adopted for the local Galactic halo[^2].
Before comparing these results against other recent analyses, we should emphasise the limited nature of the present study. The question addressed here can be stated as follows:\
Based on current statistics, and given a sample of stars drawn from the Galactic mid-Plane near the Sun, what is the frequency of ESP systems as a function of metallicity.
Our goal in constructing the reference sample of field stars, therefore, is not an unbiased estimate of the present-day metallicity distribution of the Galactic Disk - a parameter used to constrain Disk star formation histories. That undertaking requires limiting analysis to stars with main-sequence lifetimes older than the age of the Disk, avoiding possible bias through a disproportionate contribution from recent star formation episodes. Moreover, given a potential correlation between metallicity and velocity, one should weight each star’s contribution by its W-velocity to allow for the residence time in the mid-Plane and convert volume density to surface density. The latter is not an option for the present sample, since over 25% of the stars lack radial velocities. Moreover, our (B-V) limits, modelled on the known ESP systems, include early-type G and late-F stars, whose main-sequence lifetimes are shorter than 10 Gyrs. Thus, our field star sample is tailored to provide a local snapshot of the present-day abundance distribution in the mid-Plane, rather than an integrated history of star formation in the Disk.
We have compared the abundance distribution of the FGK25 dataset against results from two other studies: the Favata [*et al.*]{} (1997) analysis, described above, and the recent study by Haywood (2001: H2001). In both cases, we consider volume-limited samples (i.e. the distribution is not weighted by W velocity as in Figure 3 of F97). Haywood’s analysis [*is*]{} aimed at determining an unbiased estimate of the Disk abundance distribution, so while his initial sample is selected to have M$V < 8.5$, (B-V)$>$0.25 and $\pi > 40$ mas (based on Hipparcos data), the final analysis is limited to 328 stars with $M_V > 4.5$; that is, main-sequence dwarfs with lifetimes longer than the age of the Disk. Abundances are derived primarily from Geneva photometry, supplemented by Strömgren data, with the metallicity scale effectively adjusted to the high-resolution (Santos/Gonzalez) system.
Figure 5 compares the abundance distributions derived in those studies against our own results. For consistency, we have adjusted all of the metallicity scales to match \[Fe/H\]$_{uvby}$. All three distributions peak at values close to the solar abundance, with a substantial fraction of the sample (45% in the FGK25 dataset) having super-solar metallicities. As discussed by Haywood, this represents a significant revision of previous analyses, and may reflect a bias against metal-rich stars in samples selected based on spectral type rather than distance/colour. We note that only a small fraction of the local Disk have abundances of less than $1 \over 3$ solar: only 25 stars (6%) in the FGK25 sample. If $\approx10\%$ of the local stars are members of the thick disk, as suggested by kinematic analyses (Reid, Hawley & Gizis, 1995), then the mean abundance of that sub-population lies much closer to the solar metallicity than the value of \[Fe/H\]=-0.6 adopted in some Galactic models.
Both the F97 and H2001 samples show a more extended distribution towards higher abundances than the FGK25 dataset. This is somewhat surprising, since the latter sample, extending to stars brighter than M$_V=4$, should include a higher proportion of younger stars which are likely to be more metal-rich. The discrepancy may originate from the abundance calibrations. Figure 6 shows the metallicity distribution as a function of (B-V) colour for the H2001 and FGK25 datasets. The former shows a clear trend of increasing metallicity at redder colours, suggesting a possible systematic bias in the abundance calibration of Geneva photometry. Further observations are required to verify this hypothesis.
For present purposes, the most significant point is that the metallicities discussed here for both field stars and ESP host stars are derived from a single source - Strömgren photometry, as calibrated by Schuster & Nissen (1989). Thus, the comparison between the two abundance distributions described in the following section is internally fully self-consistent.
The abundance distribution and the frequency of giant planets
--------------------------------------------------------------
Before comparing metallicities, Figure 7 matches the (m$_1$, (b-y)) and (c$_1$, (b-y)) distributions of the ESP host stars and the FGK25 Hipparcos sample. We have distinguished between main sequence stars and potential subgiants in the former sample. As previously noted by Giménez (2000), a significant fraction of the ESP host stars have high c$_1$ values, suggesting low gravities and a mildly evolved status. In some cases, however, the colours reflect high metallicity rather than low gravity; thus, both HIP 43587 (55 Cnc) and 79248 (HD 145675) lie well above the ((b-y), c$_1$) sequence at (b-y)$\sim$0.42, but their location in the (M$_V$, (B-V)) plane demands that both are main-sequence dwarfs. Spectroscopy indicates that both are super-metal rich (Table 2). Nonetheless, the fraction of subgiants amongst known ESP hosts ($\sim15\%$) is at least a factor of three higher than that in the volume-limited sample. This may reflect an observational selection effect, since evolved stars are intrinsically more luminous, and therefore more likely to be included in the radial velocity monitoring programs.
The reddest star in the sample is the K2 subgiant, HD 27442 (HIP 19921). This lies beyond the formal limits of the Schuster & Nissen (1989) abundance calibration, at (b-y)=0.65; however, the derived metallicity, \[Fe/H\]$_{uvby}$=0.26, is not inconsistent with the spectroscopic determination of \[Fe/H\]=+0.22, so we have retained the star in the sample. We have excluded both BD -10 1366 and HD 4203 from the statistical comparison. As noted above, those stars were added to the Keck/Lick radial velocity program because they were known to be extremely metal-rich. The M dwarfs, Gl 876, is also excluded, but we include HD 177830, adjusting its abundance from \[Fe/H\]=0.36 to \[Fe/H\]$_{uvby}$=0.26.
The sample of ESP host stars is not volume-limited, particularly given the fact that the original target list was constructed before the availability of Hipparcos data, and is therefore subject to the type of distance errors illustrated in Figure 3. There may therefore be underlying biases reflecting the initial selection of which stars to monitor. Those effects can be quantified once the full dataset is available. Nonetheless, it is not unreasonable to hope that the current catalogue of ESP host stars provides a representative subset of stars with currently-detectable planetary systems; that is, stars with relatively massive (super-Jovian) companions on relatively short-period ($<$ few years) orbits.
Figure 8 plots the abundance distribution of both the ESP host stars and the reference FGK25 Hipparcos dataset. The subgiant contribution to the former distribution is shown as the shaded histogram and is consistent with the overall distribution. As emphasised in §3.1, these data are all on the Strömgren system, \[Fe/H\]$_{uvby}$, placing the Sun as -0.1 dex, at the mode of the local abundance distribution.
Visual comparison clearly confirms previous suggestions that the abundance distribution of the ESP host stars is weighted more heavily towards super-solar metallicity than the field distribution. To quantify that comparison we have combined the distributions by scaling the upper distribution to match the observed fraction of ESP hosts amongst the volume-complete sample. As noted above, 20 of the 486 stars in the FGK25 sample, or 4.1%, are known to have planetary-mass companions. This fraction is broadly consistent with previous estimates (e.g. Marcy & Butler, 2000). We have used this factor to scale the metallicity distribution in the uppermost panel of Figure 8, and the two lowest panels in Figure 8 show the fraction of ESP host stars, $$f_{ESP} \ = \ {N_{ESP} \over N_{tot}}$$ as a function of metallicity. There is an obvious trend with abundance, with f$_{ESP}$ rising to near unity at the highest abundances; both of the stars in the highest-metallicity bin of the FGK25 sample are known to have planetary-mass companions. However, even at an abundance of $2\over5$th solar, $\sim1\%$ of F, G and early-K stars are predicted to have Jovian-mass planetary companions. HD 114762b, the likely brown dwarf, contributes the spike at \[Fe/H\]$_{uvby}$=-0.7 dex.
Finally, we have compared the properties of the individual planetary systems against the abundances derived from the Strömgren data. Figure 9 shows the results, where we identify separately systems with multiple components (including the Sun, represented by Jupiter and Saturn). There is no obvious correlation between \[Fe/H\]$_{uvby}$ and any of the observed characteristics.
Discussion
----------
Is the correlation with metallicity evident in Figure 8 a selection effect? Metal-rich stars are more luminous than their metal-poor counterparts, and therefore, like subgiants, might be expected to be better represented in a target list which is partly magnitude limited. However, it seems unlikely that this type of bias could account for the smooth trend evident in the observations, particularly given the identification of planetary companions to BD -10 1366 and HD 4203, stars specifically added to the Keck program because they were known to be super-metal-rich. Thus, the simplest interpretation of Figure 8 is that the correlation represents a real physical phenomenon. The explanation for this phenomenon is somewhat less clear.
Perhaps appropriately, the two mechanisms proposed to account for the observed correlation mirror the classic nature [*versus*]{} nurture debates of biological behavioural sciences. Under the first hypothesis, planetary systems (at least those with giant planets) form more readily in the dustier environment likely to be present in high-metallicity circumstellar disks. Under the alternative hypothesis, gas giants migrate inwards due to dynamical friction with residual disk material and are absorbed into the stellar envelope. The enhanced metal content of the planet (solar system giants are likely to have Z$> 0.1$) enriches the metallicity of the outer convective envelope, leading to a higher measured chemical abundance.
One question mark hanging over the planetary pollution hypothesis centres on the details of the enhanced metal content of Jovian planets. In astronomical terms, ‘metals’ encompass all elements except hydrogen and helium - but not all metals are created equal. Metallicity measurements for F, G and early K-type stars are primarily measuring blanketting due to heavy elements, notably iron. The giant planets are known to have non-cosmic inter-element abundance ratios, but if the additional ‘metals’ are ices (C, N, O) rather than minerals (Fe, Si, Ni), as suggested by the possible absence of a rocky core in Jupiter (Guillot, 1999), then planetary pollution will have little effect on the apparent metallicity of the stellar envelope.
These two competing scenarios are discussed extensively by, amongst others, Gonzalez (1997, 1999), Laughlin (2000), Murray [*et al.*]{} (2001) and Santos [*et al.*]{} (2001). A major prediction of the pollution hypothesis is that the degree of metallicity enhancement should increase with increasing mass of the parent star. This follows from the corresponding decrease in mass of the convective envelope; adding high-Z material gives a proportionately larger increase in metallicity. Both Laughlin and Murray & Chaboyer (2001) have argued that this effect is present, although the latter authors note that a similar trend is present in their reference sample, and both analyses are based on a subset of the current catalogue of ESP hosts. Santos [*et al.*]{} (2001), in contrast, arrive at the opposite conclusion based on analysis of more than 60 systems. In their analysis, they rightly place more emphasis on the location of the upper envelope of the abundance distribution as a function of mass, rather than the mean metallicity.
All three of these analyses use theoretical tracks to deduce masses for individual ESP host stars. Figure 10 shows an alternative, more empirical approach. We have separated the current sample into main-sequence stars and subgiants based on location in Figure 1, and plot the abundance as a function of (b-y). For the main-sequence sample, (b-y) effectively traces mass, and the absence of any strong trend in the location of the high-metallicity boundary, in particular, a decrease in \[Fe/H\]$_{max}$ at redder (b-y) colours, supports the conclusion reached by Santos [*et al.*]{} Moreover, the latter authors point out that as stars evolve onto the subgiant branch, the convective envelope increases in size, diluting the effect of any planetary pollution. It is clear from Figure 10 that the evolved stars can be as metal-rich as the main-sequence dwarfs; indeed, the K2 subgiant HD 24427 (HIP 19921) is amongst the most metal-rich stars in the sample. Thus, these results suggest that planetary systems are born metal-rich, rather than having high metallicity thrust upon them.
Kinematics
==========
Velocity dispersion and ages
----------------------------
The age distribution of the ESP host stars is clearly an important parameter for understanding the Galactic origins of these systems. Two techniques have been used to estimate ages for individual stars: isochrone fitting; and the level of chromospheric activity, as measured through emission at the Ca II H & K lines. Both methods can be applied to individuals, but both have limitations. Isochrone fitting provides reliable ages for relatively F and early-G stars, but becomes less accurate for longer-lived, later-type stars. Chromospheric ages, quantified using the $R'_{HK}$ index (Soderblom [*et al.*]{}, 1991), are more readily derived, but are also less reliable since there is a considerable dispersion in activity amongst individual stars with similar ages (see Figure 10 in Soderblom [*et al.*]{}). Moreover, variability is an issue; as Henry [*et al.*]{} (1996) point out, the Sun’s age could be estimated as anywhere between 2.2 and 8 Gyrs depending on when the observations are taken during the Solar cycle. Finally, both of these techniques become significantly less reliable at ages exceeding $\sim2$ Gyrs (as evidenced by continuing uncertainties in the Galactic star formation history).
Nonetheless, both of these methods provide useful insight into the age distribution, and both have been applied by Gonzalez and co-workers (Gonzalez & Laws, 1998; Gonzalez, 1999; Gonzalez [*et al.*]{}, 2001) to estimate ages for 33 of the systems listed in Table 1. A comparison between the different estimates emphasises the inherent uncertainties, most dramatically for HD 217107 and HD 222582, where both have chromospheric age estimates of 5.6 Gyrs, but isochrone estimates of 1.2 and 11 Gyrs, respectively. Averaging the results for all 33 stars, we derive a mean age of 5.6 Gyrs ($\sigma=3.6$ Gyrs).
Space motions cannot be used to provide age estimates for individual stars. However, stellar kinematics offer an alternative means of comparing the [*average*]{} properties of diverse groups of stars. Velocity dispersion increases with age, probably through the mechanisms of orbital diffusion (Wielen, 1977) and scattering due to molecular clouds (Spitzer & Schwarzschild, 1953). A comparison between the velocity distributions of the ESP host stars and the local disk can test whether there is a significant difference in the mean age of the two samples.
We have calculated space motions for the ESP hosts using astrometric data from the Hipparcos catalogue and the available radial velocity measurements. Table 3 lists those data and the resulting (U, V, W) motions, where U is positive toward the Galactic Centre, V positive in the direction of rotation, and W directed toward the NGP.
All of the FGK25 stars have accurate proper motions and parallaes from Hipparcos, but only 60% have published radial velocities, rendering the sample unsuitable as a reference. However, the volume-complete M-dwarf sample from the PMSU survey of nearby stars (Reid [*et al.*]{}, 1995: PMSU1) gives a ready alternative, providing an unbiased representation of the kinematics of local disk stars. Reid [*et al.*]{} (2002: PMSU4) have revised the original dataset to incorporate more recent astrometric data, notably from Hipparcos, besides including higher-accuracy radial velocities from echelle observations summarised by Gizis [*et al.*]{} (2001: PMSU3). The final sample is comparable in size to the FGK25 dataset, with 436 systems lying within M$_v$-dependent distance limits ranging from 10 to 20 parsecs.
Figure 11 compares the velocity distributions of the two datasets. The left-hand panels plot the two-component velocity distributions; the right-hand panels show probability plots of the (U, V, W) distributions. As originally discussed by Lutz & Upgren (1980), these diagrams plot the cumulative distribution of a sample, C(x), against the difference with respect to the mean value, $\bar x$, in units of the standard deviation. A normal distribution, $f(x) \ = \ {1 \over \sqrt{2 \pi \sigma}} . e^{-{{(x-\bar x)}^2 \over \sigma^2}}$, gives a straight line, slope $\sigma$, in this plane. Figure 11 plots three empirical velocity distributions: data for the ESP host stars; for the full PMSU M-dwarf sample; and for the PMSU dMe dwarfs, with H$\alpha$ emission exceeding 1Å equivalent width. As discussed by Hawley [*et al.*]{} (1996; PMSU2), chromospheric emission is an age-dependent phenomenon, so the last dataset is characteristic of a moderately young stellar population, $\langle \tau \rangle \approx 1-2$Gyrs.
It is clear from Figure 11 that the velocity distribution of the ESP host stars is more closely matched to the full M dwarf sample than to the dMe sample. Quantitatively, linear fits to the central regions of the probability plots ($-1.9 < rms < 1.9$) give $$\begin{aligned}
(\sigma_U & = & 35.2; \sigma_V = 22.9; \sigma_W = 17.3: U_\odot = -6.3;
V_\odot = -24.2; W_\odot = -7.7) \\
& & \sigma_{tot} = 45.4 \ {\rm km s^{-1}; 63 \ systems}\end{aligned}$$ where $\sigma_{tot}$ is the overall velocity dispersion. Applying the same technique to the M dwarf samples gives $$\begin{aligned}
(\sigma_U & = & 21.2; \sigma_V = 14.1; \sigma_W = 13.0: U_\odot = -14.5;
V_\odot = -11.5; W_\odot = -8.1) \\
& &\sigma_{tot} = 28.6 \ {\rm km s^{-1}; 69 \ systems}\end{aligned}$$ for the emission line dwarfs and $$\begin{aligned}
(\sigma_U & = & 39.1; \sigma_V = 38.8; \sigma_W = 23.6: U_\odot = -5.7;
V_\odot = -9.6; W_\odot = -3.3) \\
& & \sigma_{tot} = 59.9 \ {\rm km s^{-1}; 404 \ systems}\end{aligned}$$ for the full sample.
Based on this comparison, we conclude that the current sample of F, G and K-type ESP hosts is younger, on average, than the overall disk population, but includes stars significantly older than typical of the dMe sample. Quantitatively, if we assume diffusion with $\sigma \propto \tau^{1 \over 2}$, then $$\langle \tau_{ESP} \rangle \ \sim \ 0.6 \langle \tau_{dM} \rangle \ \sim 2.5 \langle \tau_{dMe} \rangle$$ suggesting an average age of 3-4 Gyrs for ESP host stars for an approximately uniform star-formation rate in a 10-Gyrs-old disk. This younger mean age is not unexpected, given the higher proportion of metal-rich stars amongst the ESP sample. The average metal abundance of the Galactic disk is expected to increase with time, as successive generations of star formation contribute additional nucleosynthetic debris to the interstellar medium, so a sample biased toward high metallicities is also likely to be biased towards stars that are younger than average.
Kinematics and metallicity
--------------------------
The previous section considered the overall distribution of velocities of the ESP host stars. We can also look for correlations using velocities for the individual stars, correcting the observed heliocentric data for the solar motion with respect to the Local Standard of Rest (LSR). For the latter parameter, we use the values derived by Dehnen & Binney (1998), $$( U_\odot, \ V_\odot, W_\odot; 10.0, \ 5.3, \ 7.2)$$ where these values give the motion of the Sun with respect to the LSR. Thus, the Sun is moving towards the Galactic Centre, towards the direction of rotation and towards the NGP, and the observed velocities must be corrected accordingly. We denote the corrected velocities as (U’, V’, W’).
Figure 12 plots velocities for the ESP host stars as a function of abundance. We also indicate the location of the Sun on these diagrams. There is no obvious correlation between metallicity and either the (U’, V’, W’) component velocities or the total motion with respect to the LSR, V$_{tot}$. The highest metallicity stars in the sample span essentially the same range of velocities as the solar-abundance and sub-solar abundance ESP host stars.
Several previous studies have commented on the relatively low velocity ($\sim 13$ kms$^{-1}$) of the Sun with respect to the LSR. Gonzalex (1999), in particular, has invoked the Weak Anthropic Principle (Barrow & Tipler, 1988) in conjunction with this property, arguing that the small offset from co-rotation minimises excursions into the potentially dangerous environment (supernovae, gravitational interactions) of spiral arms, therefore providing the long term quiescence which may be necessary for advanced life forms to develop. We can make two points in this context:
- first, it is clear from Figure 12 than $\sim10\%$ of the known ESP host stars have velocities, V$_{tot}$, within a few kms$^{-1}$ of that of the Sun. Indeed, the transiting system, HD 209458, has a space motion with respect to the LSR which is almost identical with that of the Sun, while HD 114783 has a relative motion of only 6.2 kms$^{-1}$. Gonzalez [*et al.*]{} (2001) derive age estimates of 3 Gyrs (isochrones) and 4.3 Gyrs (activity) for the former star. HD 114783 is too red to allow reliable an isochrone-based age estimate, but Vogt [*et al.*]{} (2001) note that HD 114783 is chromospherically inactive, $\log{R'_{HK}}$=-4.96, or $\sim4.8$ Gyrs for the Donahue (1993) calibration. Both stars are therefore likely to have ages similar to that of the Sun.
- second, the Weak Anthropic Principle (WAP) can be expressed in two ways: as a positive concept, in that the planetary environment must permit the development of advanced lifeforms; or as a less restrictive, negative concept, in that the environment should not be inimical to the development of advanced lifeforms. Whether one chooses to express the WAP as a positive or a negative concept depends on other issues, notably belief in the likelihood of life developing elsewhere in the Universe. In either case, with a current sample of one known inhabited planet, the WAP should be given the same scientific weight as its converse, the Copernican Principle (“we’re not special”). Both are interesting philosophical concepts, which may have explanatory power; neither carries any evidentiary weight in the present context.
Finally, we have compared the distribution of properties of the extrasolar planetary systems against the systemic velocities to search for possible trends or correlations. The only potentially significant result is shown in the uppermost panel of Figure 13, plotting M$_2 \sin{i}$ against velocity perpendicular to the Plane. The data suggest that, with the exception of HD 114762, higher-mass companions tend to be found in systems with low W velocities. The result is statistically marginal, but might indicate a correlation with the mass of the parent circumstellar disk. Clearly more data are required to confirm whether this effect is real.
Summary and conclusions
=======================
Over sixty stars with planetary-mass companions are now known. While these stars neither constitute a volume-limited sample nor, probably, a complete sampling of the full range of planetary systems, they provide sufficient numbers for a preliminary investigation of the characteristics of the parent stars. In this paper we have compared the chemical abundance distribution and kinematics of those stars against data for representative samples of the local disk. Our metallicities are based on Strömgren photometry, using the calibration derived by Schuster & Nissen (1989). We have shown that the resulting metallicity scale is offset to lower abundances with respect to recent high-resolution spectroscopic measurements. This discrepancy is not important for our purposes, since Strömgren photometry is available for 86% of our reference sample - an Hipparcos-selected sample of 486 F, G and K stars within 25 parsecs of the Sun. The abundance distributions derived for both datasets are therefore internally consistent, although the solar abundance on this scale is $[Fe/H]_{uvby} \sim -0.1$ dex.
Comparing the abundance distributions of the two datasets, it is clear that, as noted in previous studies, systems currently known to have extrasolar planets are heavily weighted to high metallicities. We have used the fraction of known ESP systems in the volume-complete sample (20/486, or 4.1%) to set the two distributions on a common scaling, and compute the observed frequency as a function of chemical abundance. The results show a strong trend with abundance, with effectively 100% frequency at $[Fe/H]_{uvby} > 0.3$. However, even at abundances of less than $1 \over 2$ solar, 1 to 2% of stars are likely to have planetary-mass companions in the mass/semi-major axis/eccentricity range detectable using current techniques. Clearly, these statistics represent a lower limit to the actual frequency of extrasolar planetary systems.
How rare are solar-abundance F, G, and K stars with planets? Note that while the frequency of ESP hosts increases with \[Fe/H\], the absolute number of systems declines rapidly at high abundances. Thus, planetary systems with parent stars of near-solar abundance contribute a significant fraction of the total current sample. Based on the full FGK25 Hipparcos dataset, the local number density of stars with metallicities within $\pm0.15$ dex of the solar abundance is 0.0044 stars pc$^{-3}$. The corresponding number density of ESP host stars, based on the data plotted in Figure 7, is 0.00018 stars pc$^{-3}$. Consider an annulus centred on the Solar Radius, $R_\odot = 8$ kpc., diameter 50 parsecs. Extrapolating from the local sample, we expect $\sim17,500$ ESP host stars within this very limited subset of the Galactic Disk. Casting the net wider, consider a wedge, thickness (perpendicular to the Plane) 50 parsecs, between Galactic radii of 7 and 9 kiloparsecs, a range which encompasses relatively minor changes in mean abundance and stellar number density. Based on our calculations, we would expect a total of over 900,000 solar-type stars with Jovian-mass planetary companions.
We have also compared the kinematics of the ESP host stars against the local Galactic disk via observations of a volume-limited sample of M dwarfs. The planetary hosts exhibit a velocity distribution which is relatively well matched to a Gaussian in each component, but with lower dispersions than in the field-star sample. This suggests that the average age is only $\sim60\%$ that of a representative subset of the disk. This may reflect the higher proportion of metal-rich stars in the ESP host sample. Individual stars, however, span a wide range of motions, with velocities of up to 50-60 kms$^{-1}$ with respect to the Local Standard of Rest, and no obvious correlation between kinematics and abundance.
I would like to thank Geoff Marcy for providing radial velocity measurements for several stars in advance of publication; David Trilling, for useful comments; and David Koerner, for interesting discussion and sparking my initial interest in this topic. The research for this paper made extensive use of the SIMBAD database, maintained by Strasberg Observatory, and of Jean Schneider’s ‘Extrasolar Planets Encyclopedia’.
Barrow, J.D., Tipler, F.J. 1988, [*The Anthropic Cosmological Principle*]{}, Oxford Univ. press (Oxford)
Bessell, M.S. 2000, PASP, 112, 961
Biemont, E., Baudoux, M., Kurucz, R.L, Ansbacher, W., Pinnington, E.H. 1991, , 249, 539
Butler, R.P., Vogt, S.S., Marcy, G.W., Fischer, D.A., Henry, G.W., Apps, K. 2000, , 545, 504
Cayrel de Stobel, G., Soubiran, C., Friel, E.D., Ralite, N., Francois, P. 1997, , 124, 299
Davis Philip, A.G., Egret, D. 1980, , 40, 199
Dehnen, W., Binney, J.J. 1998, , 298, 387
Dick, S.J., 1998, [*Life on other worlds*]{}, Cambridge University Press, (Cambridge)
Donahue, R.A., 1993, Ph.d. thesis, New Mexico State University
Duflot, M., Figon, P. Meyssonier, N. 1995, , 114, 269
ESA, 1997, The Hipparcos Catalogue
Fabricius, C., Makarov, V.V. 2000, , 144, 45
Favata, F., Micela, G., Sciortino, S. 1996, , 311, 951
Favata, F., Micela, G., Sciortino, S. 1997, , 323, 809 (F97)
Fouts, G., Sandage, A. 1986, , 91, 1189
Gatewood, G., Han, I., Black, D.C. 2001, , 548, L61
Giménez, A. 2000, , 356, 213
Gizis, J.E., Reid, I.N., Hawley, S.L. 2001, , subm. (PMSU3)
Gliese, W. 1969, Catalogue of Nearby Stars, Veroff. Astr. Rechen-Instituts, Heidelberg, Nr. 22
Gliese, W., Jahrei[ß]{}, H. 1979, A&AS 38, 423
Gonzalez, G. 1997, , 285, 403
Gonzalez, G. 1998, , 334, 221
Gonzalez, G. 1999, , 308, 447
Gonzalez, G., Vanture, A.D. 1998, , 339, L29
Gonzalez, G., Wallerstein, G., Saar, S. 1999, , 511, L111
Gonzalez, G., Laws, C. 2000, , 119, 390
Gonzalez, G., Laws, C., Tyagi, S., Reddy, B.E. 2001, , 121, 432
Gratton, R.G., Carretta, E., Clementini, G., Sneden, C., 1997, in [*Hipparcos Venice ’97*]{}, ed B Battrick (ESA), p 339
Griffin, R.F. 1972, , 155, 449
Guillot, T. 1999, Science, 286, 272
Halbwachs, J.L., Arenou, F., Mayor, M., Udry, S., Queloz, D. 2000, , 355, 581
Han, I., Black, D.C., Gatewood, G. 2001, , 548, L57
Hauck, B., Mermilliod, M. 1998, , 129, 431
Hawley, S.L., Gizis, J.E., Reid, I.N. 1996, , 112, 2799 \[PMSU2\]
Haywood, M. 2001, , 325, 1365
Heacox, W.D. 1999, , 526, 928
Henry, T.J., Soderblom, D.R., Donahue, R.A., Baliunas, S.L. 1996, , 111, 439
Jahrei[ß]{}, H., Wielen, R. 1997, Proc. ESA Stmp. 402, [*Hipparcos - Venice ’97*]{}, ESA Publications, Noordwijk, p. 675
Jorissen, A, Mayor, M., Udry, S. 2001, , 379, 992
Latham, D.W., Mazeh, T., Stefanik, R.P., Mayor, M., Burki, G. 1989, Nature, 339. 38
Laughlin, G. 2000, , 545, 1064
Laws, C., Gonzalez, G. 2001, , in press
Lin, D.N.C., Bodenheimer, P., Richardson, D.C. 1996, Nature, 380, 606
Lutz, T.E., Upgren A.R. 1980, , 85, 573
Mcgrath, M.A. [*et al.*]{} 2001, DPS, 33, 6001
Marcy, G.W., Benitz, K.J. 1989, , 344, 441
Marcy, G.W., Butler, R.P. 2000, , 112, 137
Mayor, M., Queloz, D. 1995, Nature, 378, 355
Murray, N., Chaboyer, B., Arras, P., Hansen, B., Noyes, R.W. 2001, , 555, 810
Murray, N., Chaboyer, B. 2001, , subm.
Naef, D., Latham, D., Mayor, M. [*et al.*]{}, 2001, , in press
Nidever, D., Marcy, G.W., Butler, R.P., Fischer, D.A., Vogt, S.S. 2002, , subm.
Olsen, E.H. 1984, , 57, 443
Oppenheimer, B.R., Hambly, N.C., Digby, A.P., Hodgkin, S.T., Saumon, D. 2001, Science, 292, 698
Pourbaix, D. 2001, , 369, L22
Pourbaix, D., Arenou, F. 2001, , 372, 935
Randich, S., Gratton, R., Pallavicini, R., Pasquini, L., Carretta, E. 1999, , 348, 487
Reid, I.N., Hawley, S.L., Gizis, J.E. 1995, , 110, 1838 \[PMSU1\]
Reid, I.N. 1998 , 115, 204
Reid, I.N., Hawley, S.L., Gizis, J.E. 2002, , subm. \[PMSU4\]
Reid, I.N., Sahu, K.C., Hawley, S.L. 2001, , 559, 942
Ryan, S.G. 1992, , 104, 1144
Santos, N.C., Israelian, G., Mayor, M. 2000, , 363, 228
Santos, N.C., Israelian, G., Mayor, M. 2001, , 373, 1019
Sargent, A.I., Beckwith, S.V.W. 1993, Physics Today, 46, 22
Schaller, G., Schaerer, D., Meynet, G., Maeder, A. 1992, , 96, 269
Schaerer, D., Charbonnel, C., Meynet, G., Maeder, A., Schaller, G. 1993, , 102, 339
Schuster, W.J., Nissen, P.F. 1989, A&A, 221, 65
Soderblom, D.R., Duncan, D.K., Johnson, D.R.H. 1991, , 375, 722
Spitzer, L., Schwarzschild, M. 1953, , 118, 106
Stepinksi, T.F., Black, D.C. 2000, , 356, 903
Strömgren, B. 1966, Ann. Rev. Astr. Ap., 4, 433
Tabachnik, S., Tremaine, S. 2001, , in press
Tinney, C.G., Butler, R.P., Marcy, G.W., Jones, H.R.A., Penny, A.J., Vogt, S.S., Apps, K., Henry, G.W. 2001. , 551, 507
Vogt, S.S., Butler, R.P, Marcy, G.W., Fischer, D.A., Pourbaix, D., Apps, K., Laughlin, G. 2001a, , subm.
Vogt, S.S., Butler, R.P, Marcy, G.W., Apps, K. 2001b, , in press
Wielen, R. 1977, , 60, 263
Zucker, S., Mazeh, T. 2001, , 562, 1038
S. Zucker, D. Naef, D.W. Latham, M. Mayor, T. Mazeh, [*et al.*]{} 2001, , in press
[rrrrcrrrrl]{} & BD -10 3166 & 5.5:& 0.84& 12$\pm$ 4& 0.48& 0.05& 3.49& 0.00& 1\
522 & HD 142 & 3.65& 0.52& 39.00$\pm$ 0.64& 1.36& 0.98& 338.00& 0.37&\
1292 & HD 1237 & 5.36& 0.75& 56.76$\pm$ 0.53& 3.31& 0.49& 133.82& 0.51&\
3479 & HD 4208 & 5.21& 0.67& 30.58$\pm$ 1.08& 0.80& 1.70& 829.00& 0.04&\
3502 & HD 4203 & 4.22& 0.73& 12.85$\pm$ 1.27& 1.64& 1.09& 406.00& 0.53& 1, subgiant\
5054 & HD 6434 & 4.69& 0.60& 24.80$\pm$ 0.89& 0.48& 0.15& 22.09& 0.30&\
6643 & HD 8574 & 3.90& 0.58& 22.65$\pm$ 0.82& 2.23& 0.76& 228.80& 0.40&\
7513 & HD 9826 & 3.44& 0.54& 74.25$\pm$ 0.72& 0.71& 0.06& 4.62& 0.03& $\upsilon$ And, 2\
8159 & HD 10697 & 3.73& 0.67& 30.71$\pm$ 0.81& 6.59& 2.00& 1083.00& 0.12& subgiant\
9683 & HD 12661 & 4.59& 0.72& 26.91$\pm$ 0.83& 2.83& 0.78& 264.50& 0.33&\
10138 & HD 13445 & 5.98& 0.77& 91.63$\pm$ 0.61& 4.00& 0.11& 15.78& 0.05&\
12048 & HD 16141 & 4.00& 0.71& 27.85$\pm$ 1.39& 0.21& 0.35& 75.82& 0.28& subgiant\
12653 & HD 17051 & 4.22& 0.57& 58.00$\pm$ 0.55& 2.26& 0.92& 320.10& 0.16& $\iota$ Hor\
14954 & HD 19994 & 3.31& 0.57& 44.69$\pm$ 0.75& 2.00& 1.30& 454.00& 0.20&\
16537 & HD 22049 & 6.19& 0.88& 310.74$\pm$ 0.85& 0.86& 3.30& 2502.10& 0.61& $\epsilon$ Eri\
17096 & HD 23079 & 4.42& 0.58& 28.90$\pm$ 0.56& 2.54& 1.48& 627.30& 0.06&\
19921 & HD 27442 & 3.14& 1.08& 54.84$\pm$ 0.50& 1.43& 1.18& 423.00& 0.02& subgiant\
20723 & HD 28185 & 4.82& 0.71& 25.28$\pm$ 1.08& 5.60& 1.00& 385.00& 0.06&\
24205 & HD 33636 & 4.77& 0.58& 34.85$\pm$ 1.33& 7.70& 2.60& 1553.00& 0.39&\
26381 & HD 37124 & 5.07& 0.67& 30.08$\pm$ 1.15& 1.04& 0.58& 155.00& 0.19&\
26394 & HD 39091 & 4.35& 0.60& 54.92$\pm$ 0.45& 10.37& 3.34& 2115.30& 0.62&\
27253 & HD 38529 & 2.80& 0.74& 23.57$\pm$ 0.92& 0.81& 0.13& 14.41& 0.28& subgiant\
31246 & HD 46375 & 5.22& 0.86& 29.93$\pm$ 1.07& 0.25& 0.04& 3.02& 0.02&\
33212 & HD 50554 & 4.40& 0.53& 32.23$\pm$ 1.01& 4.90& 2.38& 1279.00& 0.42&\
33719 & HD 52265 & 4.06& 0.54& 35.63$\pm$ 0.84& 1.13& 0.49& 118.96& 0.29&\
40687 & HD 68988 & 4.36& 0.62& 17.00$\pm$ 0.96& 1.90& 0.07& 6.28& 0.14&\
42723 & HD 74156 & 3.57& 0.54& 15.49$\pm$ 1.01& 1.56& 0.28& 51.61& 0.65& 3\
43177 & HD 75289 & 4.05& 0.58& 34.55$\pm$ 0.56& 0.42& 0.05& 3.51& 0.00&\
43587 & HD 75732 & 5.46& 0.87& 79.80$\pm$ 0.84& 0.84& 0.11& 14.65& 0.05& 55 Cnc\
45982 & HD 80606 & 5.10& 0.72& 17.13$\pm$ 5.77& 3.41& 0.44& 111.78& 0.93&\
47007 & HD 82943 & 4.35& 0.59& 36.42$\pm$ 0.84& 0.88& 0.73& 221.60& 0.54& 4\
47202 & HD 83443 & 5.05& 0.79& 22.97$\pm$ 0.90& 0.35& 0.04& 2.99& 0.08& 5\
50786 & HD 89744 & 2.79& 0.49& 25.65$\pm$ 0.70& 7.20& 0.88& 256.00& 0.70&\
52409 & HD 92788 & 4.76& 0.69& 30.94$\pm$ 0.99& 3.80& 0.94& 340.00& 0.36&\
53721 & HD 95128 & 4.36& 0.56& 71.04$\pm$ 0.66& 2.41& 2.10& 1096.00& 0.10& 47 UMa, 6\
59610 & HD 106252 & 4.49& 0.64& 26.71$\pm$ 0.91& 6.81& 2.61& 1500.00& 0.54&\
60644 & HD 108147 & 4.07& 0.50& 25.93$\pm$ 0.69& 0.34& 0.10& 10.88& 0.56&\
64426 & HD 114762 & 4.26& 0.52& 24.65$\pm$ 1.44& 11.00& 0.30& 84.03& 0.33& 7\
64457 & HD 114783 & 6.02& 0.91& 48.95$\pm$ 1.06& 0.99& 1.20& 501.00& 0.10&\
65721 & HD 117176 & 3.71& 0.69& 55.22$\pm$ 0.73& 6.60& 0.43& 116.60& 0.40& 70 Vir, subgiant\
67275 & HD 120136 & 3.53& 0.48& 64.12$\pm$ 0.70& 3.87& 0.05& 3.31& 0.02& $\tau$ Boo\
68162 & HD 121504 & 4.30& 0.59& 22.54$\pm$ 0.91& 0.89& 0.32& 64.60& 0.13&\
72339 & HD 130332 & 5.68& 0.75& 33.60$\pm$ 1.51& 1.08& 0.09& 10.72& 0.05&\
74500 & HD 134987 & 4.40& 0.70& 38.98$\pm$ 0.98& 1.58& 0.78& 260.00& 0.25&\
77740 & HD 141937 & 4.63& 0.60& 29.89$\pm$ 1.08& 9.70& 1.49& 658.80& 0.40&\
78459 & HD 143761 & 4.19& 0.61& 57.38$\pm$ 0.71& 1.10& 0.23& 39.47& 0.03& $\rho$ CrB\
79248 & HD 145675 & 5.38& 0.90& 55.11$\pm$ 0.59& 3.30& 2.50& 1619.00& 0.35& 14 Her\
86796 & HD 160691 & 4.23& 0.70& 65.46$\pm$ 0.80& 1.97& 1.65& 743.00& 0.62& subgiant\
87330 & HD 162020 & 6.63& 0.96& 31.99$\pm$ 1.48& 13.73& 0.08& 8.43& 0.28&\
89844 & HD 168443 & 4.03& 0.70& 26.40$\pm$ 0.85& 7.20& 0.29& 57.90& 0.54& 8, subgiant\
90004 & HD 168746 & 4.78& 0.69& 23.19$\pm$ 0.96& 0.24& 0.07& 6.41& 0.00&\
90485 & HD 169830 & 3.11& 0.48& 27.53$\pm$ 0.91& 2.96& 0.82& 230.40& 0.34&\
93746 & HD 177830 & 3.32& 1.09& 16.94$\pm$ 0.76& 1.28& 1.00& 391.00& 0.43& subgiant\
94076 & HD 178911 & 3.29& 0.63& 20.42$\pm$ 1.57& 6.29& 0.33& 71.49& 0.12& subgiant\
94645 & HD 179949 & 4.09& 0.51& 36.97$\pm$ 0.80& 0.84& 0.05& 3.09& 0.05&\
96901 & HD 186427 & 4.55& 0.66& 46.70$\pm$ 0.52& 1.50& 1.70& 804.00& 0.67& 16 CygB\
97336 & HD 187123 & 4.46& 0.61& 20.87$\pm$ 0.71& 0.52& 0.04& 3.10& 0.03&\
98714 & HD 190228 & 3.34& 0.75& 16.10$\pm$ 0.81& 4.99& 2.31& 1127.00& 0.43& subgiant\
99711 & HD 192263 & 6.30& 0.94& 50.27$\pm$ 1.23& 0.76& 0.15& 23.87& 0.03&\
100970 & HD 195019 & 4.05& 0.64& 26.77$\pm$ 0.89& 3.43& 0.14& 18.30& 0.05&\
104903 & HD 202206 & 4.75& 0.71& 21.58$\pm$ 1.14& 14.68& 0.80& 258.97& 0.42&\
108859 & HD 209458 & 4.29& 0.53& 21.24$\pm$ 1.00& 0.69& 0.05& 3.53& 0.00&\
109378 & HD 210277 & 4.90& 0.77& 46.97$\pm$ 0.79& 1.28& 1.10& 437.00& 0.45&\
111143 & HD 213240 & 3.75& 0.61& 24.54$\pm$ 0.81& 3.70& 1.60& 759.00& 0.31&\
113020 & Gl 876 & 11.81& 1.60& 212.69$\pm$ 2.10& 1.98& 0.21& 61.02& 0.27& 9\
113357 & HD 217014 & 4.56& 0.66& 65.10$\pm$ 0.76& 0.47& 0.05& 4.23& 0.00& 51 Peg\
113421 & HD 217107 & 4.71& 0.72& 50.71$\pm$ 0.75& 1.28& 0.07& 7.11& 0.14&\
116906 & HD 222582 & 4.59& 0.60& 23.84$\pm$ 1.11& 5.40& 1.35& 576.00& 0.71&\
[rrrrrrrr]{} 522& 0.04& 10& 0.332& 0.168& 0.416& -0.06& -0.10\
1292& 0.11& 1& 0.459& 0.289& 0.300& -0.06& -0.17\
3479& -0.24& 9& 0.413& 0.213& 0.285& -0.22& 0.02\
3502& 0.22& 9& 0.467& 0.288& 0.392& 0.22& 0.00\
5054& -0.55& 1& 0.384& 0.159& 0.274& -0.54& 0.01\
6643& -0.09& 11& 0.362& 0.169& 0.378& -0.22& -0.13\
7513& 0.12& 2& 0.346& 0.176& 0.415& -0.02& -0.14\
8159& 0.16& 3& 0.440& 0.238& 0.379& 0.08& -0.08\
9683& 0.35& 3& 0.448& 0.267& 0.398& 0.25& -0.10\
10138& -0.20& 1& 0.484& 0.337& 0.287& -0.14& 0.06\
12048& 0.15& 1& 0.422& 0.213& 0.378& -0.02& -0.17\
12653& 0.25& 1& 0.357& 0.188& 0.364& 0.08& -0.17\
14594& 0.26& 1& 0.361& 0.185& 0.422& 0.00& -0.26\
16537& -0.07& 1& 0.504& 0.430& 0.263& -0.28& -0.21\
17096&& & 0.369& 0.179& 0.330& -0.14&\
19921& 0.22& 12& 0.651& 0.513& 0.406& 0.26& 0.04\
20723& 0.24& 1& 0.443& 0.264& 0.352& 0.15& -0.09\
24205& -0.13& 9& 0.378& 0.177& 0.324& -0.20& -0.07\
26381& -0.41& 3& 0.421& 0.202& 0.280& -0.37& 0.04\
26394&& & 0.371& 0.193& 0.363& 0.09&\
27253& 0.39& 1& 0.471& 0.278& 0.437& 0.23& -0.16\
31246& 0.21& 3& 0.502& 0.401& 0.337& 0.01& -0.20\
33212& 0.02& 11& 0.366& 0.179& 0.347& -0.12& -0.14\
33719& 0.24& 1& 0.360& 0.190& 0.404& 0.08& -0.16\
40687& 0.24& 9& 0.405& 0.244& 0.387& 0.36& 0.12\
42723& 0.13& 11& 0.375& 0.181& 0.390& -0.06& -0.19\
43177& 0.27& 1& 0.360& 0.191& 0.405& 0.10& -0.17\
43587& 0.45& 4& 0.536& 0.357& 0.415& 0.10& -0.35\
45982& 0.43& 5& 0.470& 0.312& 0.361& 0.20& -0.23\
47007& 0.33& 1& 0.386& 0.217& 0.390& 0.27& -0.06\
47202& 0.39& 1& 0.488& 0.349& 0.368& 0.18& -0.21\
50786& 0.30& 3& 0.338& 0.184& 0.451& 0.14& -0.16\
52409& 0.31& 3& 0.433& 0.253& 0.376& 0.22& -0.09\
53721& 0.01& 6& 0.391& 0.202& 0.343& 0.02& 0.01\
59610& -0.16& 11& 0.390& 0.187& 0.341& -0.13& 0.03\
60644& 0.20& 1& 0.346& 0.177& 0.391& -0.01& -0.21\
64426& -0.60& 6& 0.365& 0.125& 0.297& -0.74& -0.14\
64457& 0.33& 9& 0.521& 0.458& 0.309& -0.17& -0.50\
65721& -0.01& 6& 0.446& 0.232& 0.351& -0.07& -0.06\
67275& 0.32& 7& 0.318& 0.177& 0.439& 0.13& -0.19\
68162& 0.17& 1& 0.381& 0.189& 0.361& -0.02& -0.19\
72339& 0.05& 3& 0.475& 0.305& 0.316& -0.02& -0.07\
74500& 0.32& 3& 0.435& 0.256& 0.374& 0.22& -0.10\
77740& 0.16& 11& 0.388& 0.225& 0.346& 0.24& 0.08\
78459& -0.29& 6& 0.394& 0.178& 0.337& -0.27& 0.02\
79248& 0.50& 8& 0.537& 0.366& 0.438& 0.13& -0.37\
86796& 0.28& 10& 0.432& 0.244& 0.393& 0.20& -0.08\
87330& 0.01& 1& 0.579& 0.534& 0.244& 0.11& 0.10\
89844& 0.10& 3& 0.455& 0.233& 0.377& -0.06& -0.16\
90004& -0.06& 1& 0.435& 0.223& 0.342& -0.09& -0.03\
90485& 0.22& 1& 0.328& 0.177& 0.446& 0.09& -0.13\
94076& 0.28& 13& 0.403& 0.219& 0.378& 0.16& -0.12\
94645& 0.22& 14& 0.346& 0.183& 0.384& 0.08& -0.14\
96901& 0.07& 7& 0.416& 0.226& 0.354& 0.09& 0.02\
97336& 0.16& 8& 0.405& 0.224& 0.365& 0.17& 0.01\
98714& -0.24& 1& 0.482& 0.264& 0.306& -0.27& -0.03\
99711& -0.03& 3& 0.541& 0.493& 0.275& -0.20& -0.17\
100970& 0.16& 15& 0.419& 0.204& 0.362& -0.11& -0.27\
104903& 0.37& 1& 0.435& 0.253& 0.390& 0.24& -0.13\
108859& 0.04& 3& 0.361& 0.174& 0.362& -0.15& -0.19\
109378& 0.23& 1& 0.466& 0.285& 0.369& 0.16& -0.07\
111143& 0.16& 1& 0.387& 0.190& 0.399& -0.02& -0.18\
113357& 0.21& 3& 0.416& 0.233& 0.371& 0.18& -0.03\
113421& 0.39& 1& 0.456& 0.299& 0.376& 0.28& -0.11\
116906& -0.01& 16& 0.406& 0.202& 0.345& -0.08& -0.07\
[rrrrcrrrrrrl]{} 522& 39.00 $\pm$ 0.64& 575.2& -39.9& 2.6& 2& -58.2& -37.2& -12.1& 58.0&\
1292& 56.76 $\pm$ 0.53& 433.9& -57.9& -5.8& 1& -33.0& -16.6& 2.7& 27.5&\
3479& 30.58 $\pm$ 1.08& 313.5& 150.0& 55.4& 2& -53.1& -5.2& -55.9& 65.0&\
3502& 12.85 $\pm$ 1.27& 125.2& -124.0& -14.1& 7& -16.5& -59.3& -25.2& 57.3&\
5054& 24.80 $\pm$ 0.89& -169.0& -527.7& 23.0& 1& 85.4& -66.6& -3.4& 113.5&\
6643& 22.65 $\pm$ 0.82& 252.6& -158.6& 18.9& 1& -44.3& -37.0& -30.3& 52.1&\
7513& 74.25 $\pm$ 0.72& -172.6& -381.0& -28.3& 2& 28.5& -22.1& -14.6& 42.7&\
8159& 30.71 $\pm$ 0.81& -45.0& -105.4& -43.5& 1& 35.6& -26.7& 15.0& 55.0&\
9683& 26.91 $\pm$ 0.83& -107.8& -175.3& -52.2& 6& 55.1& -31.7& -0.1& 70.6&\
10138& 91.63 $\pm$ 0.61& 2092.8& 654.5& 56.6& 1& -97.5& -75.9& -28.5& 114.4&\
12048& 27.85 $\pm$ 1.39& -156.9& -437.1& -53.0& 2& 85.8& -41.0& 2.4& 102.7&\
12653& 58.00 $\pm$ 0.55& 333.7& 219.2& 15.5& 2& -31.2& -16.7& -7.3& 24.1&\
14594& 44.69 $\pm$ 0.75& -209.6& -69.2& 18.3& 2& 2.9& 9.0& -28.2& 28.5&\
16537& 310.74 $\pm$ 0.85& -976.4& 18.0& 15.5& 2& -3.3& 7.2& -20.0& 19.1&\
17096& 28.90 $\pm$ 0.56& -193.6& -91.9& -22.2& 8& 29.1& 29.6& 1.4& 53.2&\
19921& 54.84 $\pm$ 0.50& -48.0& -167.8& 29.3& 2& 15.1& -22.1& -19.3& 32.5&\
20723& 25.28 $\pm$ 1.08& 80.8& -31.1& 50.3& 1& -49.2& -15.2& -11.7& 40.7&\
24205& 34.85 $\pm$ 1.33& 180.8& -137.3& -1.0& 6& 5.8& -28.2& 11.1& 33.3&\
26381& 30.08 $\pm$ 1.15& -79.8& -420.0& -12.0& 2& 21.6& -47.4& -44.4& 64.4&\
26394& 54.92 $\pm$ 0.45& 312.0& 1050.2& 9.4& 2& -82.9& -46.4& 0.5& 84.1&\
27253& 23.57 $\pm$ 0.92& -80.1& -141.8& 28.9& 2& -12.6& -24.8& -33.7& 33.0&\
31246& 29.93 $\pm$ 1.07& 114.2& -96.8& 4.0& 2& 6.0& -21.5& 8.8& 27.9&\
33212& 32.23 $\pm$ 1.01& -37.3& -96.4& -3.9& 1& 3.6& -10.0& -11.5& 15.0&\
33719& 35.63 $\pm$ 0.84& -115.8& 80.3& 53.8& 1& -52.2& -21.0& -8.7& 45.1&\
40687& 17.00 $\pm$ 0.96& 128.3& 31.7& -69.5& 7& 75.1& -21.0& -10.6& 86.6&\
42723& 15.49 $\pm$ 1.01& 25.0& -200.5& 3.8& 1& 28.8& -51.7& -18.4& 61.5&\
43177& 34.55 $\pm$ 0.56& -20.5& -227.7& 9.3& 1& 20.9& -12.5& -21.8& 34.9&\
43587& 79.80 $\pm$ 0.84& -485.5& -234.4& 26.6& 2& -36.5& -18.2& -8.1& 29.5&\
45982& 17.13 $\pm$ 5.77& 47.0& 6.9& 3.8& 1& 6.9& 2.9& 11.4& 26.5&\
47007& 36.42 $\pm$ 0.84& 2.4& -174.1& 8.1& 1& 10.3& -19.8& -8.9& 25.0&\
47202& 22.97 $\pm$ 0.90& 22.4& -120.8& 28.8& 1& 20.0& -30.4& -12.1& 39.4&\
50786& 25.65 $\pm$ 0.70& -120.2& -138.6& -6.5& 2& -10.5& -29.7& -14.1& 25.4&\
52409& 30.94 $\pm$ 0.99& -12.6& -222.8& -4.5& 1& 16.1& -22.2& -20.9& 34.0&\
53721& 71.04 $\pm$ 0.66& -315.9& 55.2& 12.6& 2& -24.6& -2.6& 2.1& 17.5&\
59610& 26.71 $\pm$ 0.91& 23.8& -279.4& 15.5& 1& 28.8& -43.4& 0.3& 54.9&\
60644& 25.93 $\pm$ 0.69& -181.6& -60.8& -5.1& 1& -30.4& -11.6& -13.9& 22.4&\
64426& 24.65 $\pm$ 1.44& -582.7& -2.0& 49.9& 2& -81.8& -69.9& 59.0& 117.0&\
64457& 48.95 $\pm$ 1.06& -138.1& 9.6& -12.0& 7& -15.5& -2.8& -8.7& 6.2&\
65721& 55.22 $\pm$ 0.73& -234.8& -576.2& 4.9& 2& 13.3& -51.8& -4.0& 52.1&\
67275& 64.12 $\pm$ 0.70& -480.3& 54.2& -15.6& 2& -33.5& -19.0& -6.3& 27.2&\
68162& 22.54 $\pm$ 0.91& -250.6& -84.0& 19.5& 1& -27.6& -52.0& -1.4& 50.3&\
72339& 33.60 $\pm$ 1.51& -129.6& -140.8& -12.5& 1& -9.3& -26.2& -10.7& 21.2&\
74500& 38.98 $\pm$ 0.98& -399.0& -75.1& 3.4& 2& -21.6& -39.6& 20.3& 45.5&\
77740& 29.89 $\pm$ 1.08& 97.1& 24.0& -3.0& 1& 2.8& 13.3& -8.7& 22.6&\
78459& 57.38 $\pm$ 0.71& -196.9& -773.0& 18.4& 2& 54.7& -35.4& 20.8& 76.7&\
79248& 55.11 $\pm$ 0.59& 132.5& -298.4& -13.9& 1& 23.8& -12.2& -16.3& 35.6&\
86796& 65.46 $\pm$ 0.80& -15.1& -191.2& 9.0& 2& 2.9& -14.4& -7.6& 15.8&\
87330& 31.99 $\pm$ 1.48& 21.0& -25.2& -27.5& 1& -27.7& 2.7& -1.0& 20.4&\
89844& 26.40 $\pm$ 0.85& -92.2& -224.2& -48.7& 1& -29.7& -57.9& -6.1& 56.2&\
90004& 23.19 $\pm$ 0.96& -22.1& -69.2& -25.6& 1& -19.4& -22.3& -2.9& 19.8&\
90485& 27.53 $\pm$ 0.91& -0.8& 15.2& -17.2& 1& -16.9& 1.1& 4.0& 14.6&\
93746& 16.94 $\pm$ 0.76& -40.7& -51.8& -74.0& 2& -23.8& -72.1& -7.0& 68.2&\
94076& 20.42 $\pm$ 1.57& 47.1& 194.5& -40.4& 1& -58.3& -19.7& 1.4& 51.2&\
94645& 36.97 $\pm$ 0.80& 114.8& -101.8& -24.7& 7& -26.6& -12.9& -11.0& 18.7&\
96901& 46.70 $\pm$ 0.52& -135.1& -163.5& -27.1& 2& 17.8& -29.6& -1.7& 37.4&\
97336& 20.87 $\pm$ 0.71& 143.1& -123.2& -17.5& 3& 2.3& -16.0& -43.4& 39.7&\
98714& 16.10 $\pm$ 0.81& 104.9& -69.8& -50.2& 1& -20.0& -47.3& -35.6& 51.6&\
99711& 50.27 $\pm$ 1.23& -63.4& 262.3& -10.8& 1& -16.4& 10.1& 19.8& 31.8&\
100970& 26.77 $\pm$ 0.89& 349.5& -56.9& -92.7& 2& -72.3& -77.3& -36.5& 99.6&\
104903& 21.58 $\pm$ 1.14& -38.2& -119.8& 14.6& 1& 22.5& -19.2& -10.0& 35.5&\
108859& 21.24 $\pm$ 1.00& 28.9& -18.4& -14.8& 1& -5.6& -15.6& 0.6& 13.7&\
109378& 46.97 $\pm$ 0.79& 85.5& -449.8& -20.9& 3& 4.3& -50.2& -6.2& 47.1&\
111143& 24.54 $\pm$ 0.81& -135.2& -194.1& -0.5& 1& 25.6& -29.9& 23.1& 52.9&\
113020& 212.69 $\pm$ 2.10& 960.3& -675.6& -1.8& 5& -12.5& -20.0& -11.5& 15.5&\
113357& 65.10 $\pm$ 0.76& 208.1& 61.0& -31.2& 2& -14.9& -28.0& 14.7& 31.9&\
113421& 50.71 $\pm$ 0.75& -6.1& -16.0& -12.1& 4& -1.1& -7.8& 9.3& 19.0&\
116906& 23.84 $\pm$ 1.11& -145.4& -111.1& 12.1& 7& 36.6& -0.3& -11.5& 46.4&\
[^1]: In making this statement, we define planets as forming in a circumstellar disk, while brown dwarfs form as independent accreting cores in the parent molecular cloud. Following common usage, we also require planets to be in orbit around a more massive central body. Other definitions of these terms are possible.
[^2]: This has implications for the interpretation of the nature of the cool white dwarfs identified in recent proper motion surveys (Oppenheimer [*et al.*]{}, 2001). An increased local density of the [*stellar*]{} halo easily accounts for the observed numbers of high velocity stars, as suggested by Reid [*et al.*]{}, 2001).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the static quantum potential for a theory of anti-symmetric tensor fields that results from the condensation of topological defects, within the framework of the gauge-invariant but path-dependent variables formalism. Our calculations show that the interaction energy is the sum of a Yukawa and a linear potentials, leading to the confinement of static probe charges.'
address: |
$^1$Departamento de Física, Universidad Técnica F. Santa María, Casilla 110-V, Valparaíso, Chile.\
$^2$Instituto de Física, Universidade Federal do Rio de Janeiro, 21945, Rio de Janeiro, Brazil.
author:
- 'P. Gaete$^{1}$[^1] and C. Wotzasek $^{2}$[^2]'
title: Brane Condensation and Confinement
---
INTRODUCTION
============
One of the fundamental issues of theoretical physics is that of the confinement for the fundamental constituents of matter. In fact distinction between the apparently related phenomena of screening and confinement is of considerable importance in our present understanding of gauge theories. Field theories that yield the linear potential are important to particle physics, since those theories may be used to describe the confinement of quarks and gluons and be considered as effective theories of quantum chromodynamics. We study the confinement versus screening properties of some theories of massless antisymmetric tensors, magnetically and electrically coupled to topological defects that eventually condense, as a consequence of the Julia–Toulouse mechanism (JTM)[@JT]. This mechanism is the dual to the Higgs mechanism and has been shown to lead to a concrete massive antisymmetric theory with a jump of rank. We show that in the presence of two tensor fields the condensation induces not only a mass term and a jump of rank but also a BF coupling which will be responsible for the change from the screening to the confining phase of the theory.
An important issue here is the nature of the phase transition in the presence of a finite condensate of topological defects. It is this aspect, in $D=d+1$ dimensions for generic antisymmetric tensors theories, that is of importance for us. This issue was discussed long time ago [@JT] in the framework of ordered solid-state media and more recently in the relativistic context [@QT]. The basic idea in Ref.[@JT] was to consider models with non-trivial homotopy group able to support stable topological defects characterized by a length scale $r=1/M$, where the mass parameter $M$ is a cut-off for the low-energy effective field theory. The long wavelength fluctuations of the continuous distribution of topological defects are the new hydrodynamical modes for the effective theory that appear when topological defects condense. In [@JT] there is an algorithm to identify these modes in the framework of ordered solid-state media. However, due to the presence of non-linear terms, the lack of relativistic invariance and the need to introduce dissipation terms it becomes difficult to write down an action for the phase with a condensate of topological defects. In the relativistic context none of the above problems is present. In [@QT] an explicit form for the action in the finite condensate phase, for generic compact antisymmetric field theories was found. In this context the JTM is the natural generalization of the confinement phase for a vector gauge field.
In this paper we make use of the JTM, as presented in [@QT], to study the low-energy field theory of a pair of massless anti-symmetric tensor fields, say $A_p$ and $B_q$ with $p+q+2=D$, coupled electrically and magnetically to a large set of $(q-1)$-branes, characterized by charge $e$ and a Chern-Kernel $\Lambda_{p+1}$ [@HL], that eventually condense. It is shown that the effective theory that results displays the confinement property by computing explicitly the effective potential for a pair of static, very massive point probes. Basically, we are interested in studying the JTM in model field theories involving $B_q$ and $A_p$ coupled to a $(q-1)$-brane, according to the following action $$\begin{aligned}
\label{R10}
{\cal S} &=& \int \frac 12 \frac{(-1)^q}{(q+1)!}
\left[H_{q+1}(B_q) \right]^2 + e\, B_q J^{q}(\Lambda) + \frac 12
\frac{(-1)^p}{(p+1)!} \left[F_{p+1}(A_p) - e
\Lambda_{p+1}\right]^2,\end{aligned}$$ and consider the condensation phenomenon when $\Lambda_{p+1}$ becomes the new massive mode of the effective theory. Our compact notation here goes as follows. The field strength reads $F_{p+1}\left(A_p\right)= F_{\mu_1 \mu_2 \ldots
\mu_{p+1}}=\partial_{[\mu_1}A_{\mu_2\cdots\mu_{p+1}]}$ and $\Lambda_{p+1}=\Lambda_{\mu_1\cdots\mu_{p+1}}$ is a totally anti-symmetric object of rank ($p+1$). The conserved current $J^q(\Lambda)$ is given by a delta-function over the world-volume of the ($q-1$)-brane [@Kleinert:kx]. This conserved current may be rewritten in terms of the kernel $\Lambda_{p+1}$ as $$J^q(\Lambda) = \frac
1{(p+1)!} \epsilon^{q,\alpha ,p+1}\partial_\alpha \Lambda_{p+1}\;
,$$ and $\epsilon^{q,\alpha ,p+1} =
\epsilon^{\mu_1\ldots\mu_q,\alpha ,\nu_1\ldots\nu_{p+1}}$. This notation will be used in the discussion of the JTM in the next section as long as no chance of confusion occurs.
The Julia–Toulouse Mechanism and the Action in the Condensed Phase
==================================================================
Although the JT algorithm becomes problematic in the ordered solid state media, Quevedo and Trugenberg have shown that it leads to simple demands in the study of compact antisymmetric tensor, where it produces naturally the effective action for the new phase. They observed that when the $(d-h-1)$-branes condense this generates a new scale $\Delta$ related to the average density $\rho $ of intersection points of the $(d-h)$-dimensional world-hypersurfaces of the condensed branes with any $(h+1)$-dimensional hyperplane. The four requirements to describe effectively the dense phase are: (i) an action built up to two derivatives in the new field possessing (ii) gauge invariance, (iii) relativistic invariance and, most important, (iv) the need to recover the original model in the limit $\Delta
\to 0$. One is therefore led to consider the action for the condensate as $$\begin{aligned}
\label{RM15} {\cal S}_{\Omega}=\int
\frac {(-1)^h }{2 \Delta^2(h+1)!} \left[{F_{h+1}(\Omega_h)}
\right]^2 -\frac {(-1)^h\, h!}{2\, e^2} \left[\Omega_h - H_h
(\phi_{h-1})\right]^2,\end{aligned}$$ where $H_{\mu_1\cdots\mu_h} =
%\frac 1{(h+1)!}
\partial_{[\mu_1}\phi_{\mu_2\cdots\mu_{h}]}$ and the underlying gauge invariance is manifest by the simultaneous transformations $\Omega_{\mu_1\cdots\mu_h} \to
\Omega_{\mu_1\cdots\mu_h} +
\partial_{[\mu_1}\psi_{\mu_2\cdots\mu_h]}$ and $\phi_{\mu_1\cdots\mu_{h-1}} \to \phi _{\mu_1\cdots\mu_{h-1}}+
\psi_{\mu_1\cdots\mu_{h-1}}$. Upon fixing this invariance one can drop all considerations over $\phi_{h-1}$ after absorbing $H_h(\phi_{h-1})$ into $\Omega_h$, so that the action describes the exact number of degrees of freedom of a massive field whose mass parameter reads $m=\Delta/e$. This process, named as JTM, is dual to the Higgs mechanism. Here on the other hand, the new modes generated by the condensation of magnetic topological defects absorbs the original variables of the effective theory, thereby acquiring a mass while in the Higgs mechanism it is the original field that incorporates the degrees of freedom of the electric condensate to acquire mass. This difference explains the change of rank in the JT mechanism that is not present in the Higgs process. In the limit $\Delta \to 0$ the only relevant field configurations are those that satisfy the constraint $F_{h+1}(\Omega_h) =0$ whose solution reads $\Omega_{\mu_1\cdots\mu_h} =
%\frac 1{(h+1)!}
\partial_{[\mu_1}\psi_{\mu_2\cdots\mu_{h}]}$ where $\psi_{h-1}$ is an $(h-1)$-anti-symmetric tensor field. The field $\psi_{h-1} $ can then be absorbed into $\phi_{h-1} $ this way recovering the original low-energy effective action before condensation.
The distinctive feature of the JT mechanism is that after condensation $\Lambda_{p+1}$ is elevated to the condition of propagating field. The new degree of freedom absorbs the degrees of freedom of the tensor $A_p$ this way completing its longitudinal sector. The new mode is therefore explicitly massive. Since $A_p\to \Lambda_{p+1}$ there is a change of rank with dramatic consequences. The last term in (\[R10\]), displaying the magnetic coupling between the field-tensor $F_{p+1}(A_p)$ and the $(q-1)$-brane, becomes the mass term for the new effective theory in terms of the tensor field $\Lambda_{p+1}$ and a new dynamical term is induced by the condensation. The minimal coupling of the $B_q$ tensor becomes responsible for another contribution for the mass, this time of topological nature. Indeed the second term (\[R10\]) becomes a “$B\wedge F(\Lambda)$" term between the remaining propagating modes, inducing topological mass, in addition to the induced condensed mass, $$\begin{aligned}
\label{RM10} {\cal S}_{cond} &=& \int \frac{(-1)^q}{2(q+1)!}
\left[H_{q+1}(B_q) \right]^2 + e\, B_q
\epsilon^{q,\alpha,p+1}\partial_\alpha \Lambda_{p+1} + \nonumber\\
&+& \int \frac {(-1)^{p+1}}{2 (p+2)!} \left[F_{p+2}(\Lambda_{p+1})
\right]^2 -\frac {(-1)^{p+1}\, (p+1)!}2 m^2 \Lambda_{p+1}^2.\end{aligned}$$ Recall that the theory before condensation displayed two independent fields coupled to a $(q-1)$–brane. The nature of the two couplings were however different. The $A_p$ tensor, that was magnetically coupled to the brane, was then absorbed by the condensate after phase transition. On the other hand, the electric coupling, displayed by the $B_q$ tensor, became a “$B\wedge
F(\Lambda)$" topological term after condensation. There has been a drastic change in the physical scenario. We want next to obtain an effective action for the $B_q$ tensor. To this end we shall next integrate out the field $\Lambda$ describing the condensate to obtain, our final effective theory as $$\begin{aligned}
\label{acaoeffetiva} {\cal S}_{eff} =\int \frac{(-1)^{q+1}}{2\,
(q+1)!} H_{q+1}(B_q) \left(1 +\frac{e^2}{\Delta^2 + m^2}
\right)H^{q+1}(B_q).\end{aligned}$$
Interaction Energy
==================
Next we examine the screening versus confinement issue. We shall consider a specific example involving two Maxwell tensors coupled electrically and magnetically to a point-charge such that after condensation we end up with a Maxwell and a massive Kalb-Ramond field (the condensate) coupled topologically to each other. We shall calculate the interaction energy for the effective theory between external probe sources by computing the expectation value of the energy operator $H$ in the physical state $\left| \Phi
\right\rangle$ describing the sources, denote by $ \left\langle H
\right\rangle _\Phi$. The Kalb-Ramond field $\Lambda_{\mu\nu}$ carrying the degrees of freedom of the condensate is integrate out leading to $${\cal L} = - \frac{1}{4}F_{\mu \nu } \left( {1 + \frac{{e^2
}}{{\triangle^2 + m^2 }}} \right)F^{\mu \nu } - A_0 J^0 ,
\label{KR15}$$ where $J^0$ is an external current. We observe that the limits $e\to 0$ or $m\to 0$ are well defined and lead to a pure Maxwell theory or to a (topologically) massive model. Since the probe charges only couple to the Maxwell fields, the Kalb- Ramond condensate will not contribute to their interaction energy in the first case because in the limit where the parameter $e\to 0$ the Maxwell field and the condensate decouple. The second limit means that we are back to the non-condensed phase. As so the confinement of the probe charges are expected to disappear being taken over by an screening phase controlled by the parameter $e$ playing the role of topological mass.
Once this is done, the canonical quantization of this theory from the Hamiltonian point of view follows straightforwardly. The canonical momenta read $\Pi ^\mu = - \left( {1 + \frac{{e^2
}}{{\Delta ^2 + m^2 }}} \right)F^{0\mu }$ with the only nonvanishing canonical Poisson brackets being $\left\{ {A_\mu
\left( {t,x} \right),\Pi ^\nu \left( {t,y} \right)} \right\} =
\delta _\mu ^\nu \delta \left( {x - y} \right)$. Since $\Pi_0$ vanishes we have the usual primary constraint $\Pi_0=0$, and $\Pi
^i = \left( {1 + \frac{{e^2 }}{{\Delta ^2 + m^2 }}}
\right)F^{i0}$. The canonical Hamiltonian is thus $$H_C = \int {d^3 } x\left\{ { - \frac{1}{2}\Pi ^i \left( {1 +
\frac{{e^2 }}{{\Delta ^2 + m^2 }}} \right)^{ - 1} \Pi _i + \Pi
^i \partial _i A_0 + \frac{1}{4}F_{ij} \left( {1 + \frac{{e^2
}}{{\Delta ^2 + m^2 }}} \right)F^{ij} + A_0 J^0 }.
\right\}.\label{KR25}$$ Time conservation of the primary constraint $ \Pi _0$ leads to the secondary Gauss-law constraint $\Gamma _1 \left( x \right) \equiv
\partial _i \Pi ^i - J^0 = 0$. The preservation of $\Gamma_1$ for all times does not give rise to any further constraints. The theory is thus seen to possess only two constraints, which are first class, therefore the theory described by $(\ref{KR15})$ is a gauge-invariant one. The extended Hamiltonian that generates translations in time then reads $H = H_C + \int {d^3 } x\left(
{c_0 \left( x \right)\Pi _0 \left( x \right) + c_1 \left( x
\right)\Gamma _1 \left( x \right)} \right)$, where $c_0 \left( x
\right)$ and $c_1 \left( x \right)$ are the Lagrange multiplier fields. Moreover, it is straightforward to see that $\dot{A}_0
\left( x \right)= \left[ {A_0 \left( x \right),H} \right] = c_0
\left( x \right)$, which is an arbitrary function. Since $ \Pi^0 =
0$ always, neither $ A^0 $ nor $ \Pi^0 $ are of interest in describing the system and may be discarded from the theory. The quantization of the theory requires the removal of nonphysical variables, which is done by imposing a gauge condition such that the full set of constraints becomes second class. A convenient choice is found to be [@Pato] $\Gamma _2 \left( x \right)
\equiv \int\limits_{C_{\xi x} } {dz^\nu } A_\nu \left( z \right)
\equiv \int\limits_0^1 {d\lambda x^i } A_i \left( {\lambda x}
\right) = 0$, where $\lambda$ $(0\leq \lambda\leq1)$ is the parameter describing the spacelike straight path $ x^i = \xi ^i +
\lambda \left( {x - \xi } \right)^i $, and $ \xi $ is a fixed point (reference point). There is no essential loss of generality if we restrict our considerations to $ \xi ^i=0 $. In this case, the only nonvanishing equal-time Dirac bracket is $$\left\{ {A_i \left( x \right),\Pi ^j \left( y \right)} \right\}^ *
=\delta{ _i^j} \delta ^{\left( 3 \right)} \left( {x - y} \right) -
\partial _i^x \int\limits_0^1 {d\lambda x^j } \delta ^{\left( 3
\right)} \left( {\lambda x - y} \right). \label{KR45}$$ We now turn to the problem of obtaining the interaction energy between pointlike sources in the model under consideration. The state $\left| \Phi \right\rangle$ representing the sources is obtained by operating over the vacuum with creation/annihilation operators. We want to stress that, by construction, such states are gauge invariant. In the case at hand we consider the gauge-invariant stringy $\left|{\overline \Psi \left( \bf y
\right)\Psi \left( {\bf y^ \prime } \right)} \right\rangle$, where a fermion is localized at ${\bf y}\prime$ and an antifermion at $
{\bf y}$ as follows [@Dirac2], $$\left| \Phi \right\rangle \equiv \left| {\overline \Psi \left(
\bf y \right)\Psi \left( {\bf y}\prime \right)} \right\rangle =
\overline \psi \left( \bf y \right)\exp \left(
{iq\int\limits_{{\bf y}\prime}^{\bf y} {dz^i } A_i \left( z
\right)} \right)\psi \left({\bf y}\prime \right)\left| 0
\right\rangle, \label{KR60}$$ where $\left| 0 \right\rangle$ is the physical vacuum state and the line integral appearing in the above expression is along a spacelike path starting at ${\bf y}\prime$ and ending $\bf y$, on a fixed time slice. It is worth noting here that the strings between fermions have been introduced in order to have a gauge-invariant function $\left| \Phi \right\rangle $. In other terms, each of these states represents a fermion-antifermion pair surrounded by a cloud of gauge fields sufficient to maintain gauge invariance. As we have already indicated, the fermions are taken to be infinitely massive (static).
From our above discussion, we see that $\left\langle H
\right\rangle _\Phi$ reads $$\left\langle H \right\rangle _\Phi = \left\langle \Phi
\right|\int {d^3 } \left\{ { - \frac{1}{2}\Pi _i \left( {1 -
\frac{{e^2 }}{{\nabla ^2 - m^2 }}} \right)^{ - 1} \Pi ^i }
\right\}\left| \Phi \right\rangle, \label{KR65}$$ where, in this static case, $\Delta ^2 = - \nabla ^2$. Observe that when $e=0$ we obtain the pure Maxwell theory, as mentioned after (\[KR15\]). From now on we will suppose $e\neq 0$.
Next, from the foregoing Hamiltonian analysis, $\left\langle H
\right\rangle _\Phi$ becomes $\left\langle H \right\rangle _\Phi =
\left\langle H \right\rangle _0 + V^{\left( 1 \right)} +
V^{\left( 2 \right)}$, where $\left\langle H \right\rangle _0 =
\left\langle 0 \right|H\left| 0 \right\rangle$. The $V^{\left( 1
\right)}$ and $V^{\left( 2 \right)}$ terms are given by: $$V^{\left( 1 \right)} = - \frac{{q^2 }}{2}\int {d^3 x} \int_{\bf
y}^{\bf y^\prime} {dz^\prime_i } \delta ^{\left( 3 \right)} \left(
{x - z^\prime} \right)\frac{1}{{\nabla _x^2 - M^2 }}\nabla _x^2
\int_{\bf y}^{\bf y^\prime} {dz^i } \delta ^{\left( 3 \right)}
\left( {x - z} \right), \label{KR80}$$ and $$V^{\left( 2 \right)} = \frac{{q^2 m^2}}{2}\int {d^3 x}
\int_{\bf y}^{\bf y^\prime} {dz^\prime_i } \delta ^{\left( 3
\right)} \left( {x - z^\prime} \right)\frac{1}{{\nabla _x^2 - M^2
}} \int_{\bf y}^{\bf y^\prime} {dz^i } \delta ^{\left( 3 \right)}
\left( {x - z} \right), \label{KR85}$$ where $ M^2\equiv{m^2+ e^2} $ and the integrals over $z^i$ and $z^\prime_i$ are zero except on the contour of integration.
The $V^{\left( 1 \right)}$ term may look peculiar, but it is nothing but the familiar Yukawa interaction plus self-energy terms. In effect, as was explained in Ref. [@GG2], the expression (\[KR80\]) can also be written as $$V^{\left( 1 \right)} = \frac{{e^2 }}{2}\int_{\bf y}^{{\bf
y}^{\prime} } {dz_i^{\prime}}\partial _i^{z^{\prime}} \int_{\bf
y}^{{\bf y}^{\prime}} {dz^i }\partial _z^i G\left( {{\bf
z}^{\prime},{\bf z}} \right)= - \frac{{q^2 }}{{4\pi }}\frac{{e^{
- M|{\bf y} - {\bf y}^ {\prime}| } }}{{|{\bf y} - {\bf
y}^{\prime}|}}, \label{KR90}$$ where we used that the Green function $G({\bf z}^{\prime},{\bf z})
= \frac{1}{{4\pi }}\frac{{e^{ - M|{\bf z}^{\prime} - {\bf z}|}
}}{{|{\bf z}^{\prime} - {\bf z}|}}$ and remembered that the integrals over $z^i$ and $z_i^{\prime}$ are zero except on the contour of integration. The expression then reduces to the Yukawa-type potential after subtracting the self-energy terms. We now turn our attention to the calculation of the $V^{\left( 2
\right)}$ term, which is given by $$V^{\left( 2 \right)} = \frac{{q^2 m^2 }}{2}\int_{\bf y}^{{\bf
y}^{\prime} } {dz^{{\prime} i} } \int_{\bf y}^{{\bf y}^{\prime} }
{dz^i } G({\bf z}^{\prime} ,{\bf z}). \label{KR105}$$ It is appropriate to observe here that the above term is similar to the one found for the system consisting of a gauge field interacting with a massive axion field [@GG2]. Notwithstanding, in order to put our discussion into context it is useful to summarize the relevant aspects of the calculation described previously [@GG2]. In effect, as was explained in Ref. [@GG2], by using the Green function in momentum space, that is, $\frac{1}{{4\pi }}\frac{{e^{ - M|{\bf z}^{\prime} -
{\bf z}|} }}{{|{\bf z}^{\prime} - {\bf z}|}} = \int {\frac{{d^3
k}}{{\left( {2\pi } \right)^3 }}\frac{{e^{i{\bf k} \cdot \left(
{{\bf z}^{\prime}- {\bf z}} \right)} }}{{{\bf k}^2 + M^2 }}}$, the expression (\[KR105\]) can also be written as $$V^{\left( 2 \right)} = q^2 m^2 \int {\frac{{d^3 k}}{{\left(
{2\pi } \right)^3 }}} \left[ {1 - \cos \left( {{\bf k} \cdot {\bf
r}} \right)} \right]\frac{1}{{({\bf k}^2 + M^2)
}}\frac{1}{{\left( {{\bf {\hat n}} \cdot {\bf k}} \right)^2 }},
\label{KR115}$$ where ${\bf {\hat n}}\equiv \frac{{{\bf y} - {\bf
y}^{\prime}}}{{|{\bf y} - {\bf y}^{\prime}| }}$ is a unit vector and ${\bf r}={\bf y}-{\bf y^{\prime}}$ is the relative vector between the quark and antiquark. Since ${\bf {\hat n}}$ and ${\bf
r}$ are parallel, we get accordingly $V^{\left( 2 \right)} =
\frac{{q^2 m^2 }}{{8\pi ^3 }}\int\limits_{ - \infty }^\infty
{\frac{{dk_r }}{{k_r^2 }}} \left[ {1 - \cos \left( {k_r r}
\right)} \right]\int\limits_0^\infty {d^2 k_T \frac{1}{{(k_r^2 +
k_T^2 + M^2) }}}$, where $k_T$ denotes the momentum component perpendicular to ${\bf r}$. Integration over $k_T$ yields $
V^{\left( 2 \right)} = \frac{{q^2 m^2 }}{{8\pi ^2 }}\int\limits_{
- \infty }^\infty {\frac{{dk_r }}{{k_r^2 }}} \left[ {1 - \cos
\left( {k_r r} \right)} \right]\ln \left( {1 + \frac{{\Lambda ^2
}}{{k_r^2 + M^2 }}} \right)$, where $\Lambda$ is an ultraviolet cutoff. We also observe at this stage that similar integral was obtained independently in Ref.[@Suganuma] in the context of the dual Ginzburg-Landau theory by an entirely different approach.
We now proceed to compute the previous integral. For this purpose we introduce a new auxiliary parameter $\varepsilon$ by making in the denominator of the previous integral the substitution $k_r^2\rightarrow k_r^2+\varepsilon^2$. Thus it follows that $
V^{\left( 2 \right)} \equiv \lim _ {\varepsilon \to 0}
{\widetilde V}^{\left( 2 \right)}= \lim _{\varepsilon \to
0}\frac{{q^2 m^2 }}{{8\pi ^2 }}\int\limits_{ - \infty }^\infty
{\frac{{dk_r }}{{(k_r^2 + \varepsilon ^2) }}} \left[ {1 - \cos
\left( {k_r r} \right)} \right]\ln \left( {1 + \frac{{\Lambda ^2
}}{{k_r^2 + M^2 }}} \right)$. We further note that the integration on the $k_r$-complex plane yields ${\widetilde
V}^{\left( 2 \right)} = \frac{{q^2 m^2 }}{{8\pi }}\left( {\frac{{1
- e^{ - \varepsilon |{\bf y} - {{\bf y}^\prime}| } }}{\varepsilon
}} \right)\ln \left( {1 + \frac{{\Lambda ^2 }}{{M^2 - \varepsilon
^2 }}} \right)$. Taking the limit $\varepsilon \to 0$, this expression then becomes $ V^{\left( 2 \right)} = \frac{{q^2 m^2
}}{{8\pi }}|{\bf y} - {{\bf y}^\prime}| \ln \left( {1 +
\frac{{\Lambda ^2 }}{{M^2 }}} \right)$.
This, together with Eq.(\[KR90\]), immediately shows that the potential for two opposite charges located at ${\bf y}$ and ${\bf y^\prime}$ is given by $$V(L) = - \frac{{q^2 }}{{4\pi }}\frac{{e^{ - ML} }}{L} +
\frac{{q^2 m^2 }}{{8\pi }}L\ln \left( {1 + \frac{{\Lambda ^2
}}{{M^2 }}} \right), \label{KR145}$$ where $L\equiv|{\bf y}-{\bf {y^\prime}}|$.
Final Remarks
=============
We have studied the confinement versus screening issue for a pair of antisymmetric tensors coupled to topological defects that eventually condense, giving a specific realization of the Julia–Toulouse phenomenon. We have seen that the Julia–Toulouse mechanism for a couple of massless antisymmetric tensors is responsible for the appearance of mass and the jump of rank in the magnetic sector while the electric sector becomes a BF–type coupling. The condensate absorbs and replaces one of the tensors and becomes the new massive propagating mode but does not couple directly to the probe charges. The effects of the condensation are however felt through the BF coupling with the remaining massless tensor. It is therefore not surprising that they become manifest in the interaction energy for the effective theory. We have obtained the effective theory for the condensed phase in general and computed the interaction energy between two static probe charges, in a specific example, in order to test the confinement versus screening properties of the effective model. Our results show that the interaction energy in fact contains a linear confining term and an Yukawa type potential. It can be observed that confinement completely disappears in the limit $m\to
0$ while the screening takes over controlled by the topological mass parameter instead. Although we have considered the case where the effective model consists of the BF–coupling between a Kalb-Ramond field (that represents the condensate) and a Maxwell field, our results seem to be quite general. A direct calculation for tensors of arbitrary rank in the present approach is however a quite challenging problem that we hope to be able to report in the future.
B. Julia and G. Toulouse, J. Physique Lett. [**40**]{} L395 (1979).
F. Quevedo, C. A. Trugenberger, Nucl.Phys.[**B501**]{} 143 (1997).
\[7\] F.R. Harvey and H.B. Lawson Jr., A theory of characteristic currents associated with a singular connection, Asterisque 213 (1993), Sociètè Mathèmatique de France.
H. Kleinert, “Gauge Fields In Condensed Matter. Vol. 1: Superflow And Vortex Lines, Disorder Fields, Phase Transitions,” World Scientific Publishing Co., Singapore 1989.
A.M. Polyakov, “Gauge Fields and Strings", Harwood Academic Publishers (1987).
P. Orland, Nucl. Phys. [**B205**]{} 107 (1982). S. Deguchi and Y. Kokubo, Mod. Phys. Lett. [**A17**]{}, 503 (2002).
P. Gaete, Z. Phys. [**C76**]{}, 355 (1997); Phys. Lett. [**B515**]{}, 382 (2001); Phys. Lett. [**B582**]{}, 270 (2004).
P. A. M. Dirac, [*The Principles of Quantum Mechanics*]{} (Oxford University Press, Oxford, 1958); Can. J. Phys. [**33**]{}, 650 (1955).
P. Gaete and E. I. Guendelman, “Confinement in the presence of external fields and axions", hep-th/0404040.
H. Suganuma, S. Sasaki and H. Toki, Nucl. Phys. [**B435**]{}, 207 (1995).
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the interference of two photons with different colors in the context of a Hong-Ou-Mandel experiment, in which single photons enter each of the input ports of a beam splitter, and exit in the same, albeit undetermined, output port. Such interference is possible if one uses an active (energy-non-conserving) beam splitter. We find scenarios in which one “red” and one “blue” photon enter the beam splitter, and either two red or two blue photons exit, but never one of each color. We show how the precise form of the active beam-splitter transformation determines in what way the spectral degrees of freedom of the input photons should be related to each other for perfect destructive interference of the different-color components in the output. We discuss two examples of active beam splitters: one is a gedanken experiment involving a moving mirror and the other is a more realistic example involving four-wave mixing in an optical fiber.'
author:
- 'M. G. Raymer,$^1$ S. J. van Enk,$^1$ C. J. McKinstrie$^2$ and H. J. McGuinness$^1$'
title: Interference of two photons of different color
---
Introduction
============
Bosons, in contrast to fermions, tend to occupy the same state when given the opportunity. This tendency arises from the fundamental quantum commutation relations of bosonic operators, and is responsible, for example, for the creation of laser light and the Bose-Einstein condensate (BEC). In the case of coherent matter waves, such as those in a BEC, stimulated scattering of bosonic atoms enhances the probability for the number of atoms in an already occupied atomic mode (in particular, the ground state of their external motion) to increase with time, and to grow to macroscopic proportions.
At the level of a few individual photons (as opposed to stimulated emission in lasers, in which the presence of many photons is crucial), the propensity of two photons to bunch together was first observed experimentally by Hong, Ou and Mandel (HOM) [@hon87]. Specifically, when two separate photons are incident simultaneously upon a 50/50 beam splitter from opposite sides, the two output modes that emerge from the beam splitter each contain either two or zero photons, but never one photon each.
{width="1.8in"}
Let us illustrate this effect in a simple case, which is illustrated in Fig. 1. (The complete theory is given in subsequent sections.) Two input modes (1 and 2) contain one photon each with the same center frequency, temporal width and polarization, which arrive at the beam splitter at the same time. The initial two-photon state is denoted by $$|\Psi\rangle_{{\rm in}}=a_{1{\rm in}}^\d a_{2{\rm in}}^\d|{\rm vac}\rangle, \label{1.1}$$ where $a_{1{\rm in}}^\d$ and $a_{2{\rm in}}^\d$ are creation operators for input modes 1 and 2, respectively. (These modes need not be monochromatic, but can be wave packets [@blow90; @lou00].) The relation between the input and output mode operators can be written as $$\begin{aligned}
a_{1{\rm in}}^\dagger&=&\tau a_{1{\rm out}}^\dagger - \rho a_{2{\rm out}}^\dagger, \nonumber\\
a_{2{\rm in}}^\dagger&=&\rho a_{1{\rm out}}^\dagger + \tau a_{2{\rm out}}^\dagger, \label{1.2}\end{aligned}$$ where the coefficients $\tau$ (transmissivity) and $\rho$ (reflectivity) were assumed to be non-negative, without loss of generality, and satisfy the auxiliary equation $\tau^2 + \rho^2 = 1$. (Usually, one would give the output operators in terms of the input operators, as in the Heisenberg picture, but for our purposes the inverse transformations are more useful.) The input and output operators satisfy the commutation relations $[a_k,a_k^\d] = 1$ for k = 1 or 2. All other commutators are zero. By combining Eqs. (\[1.1\]) and (\[1.2\]), and using the fact that the vacuum is invariant under the action of the beam splitter, one finds that the output state is $$\begin{aligned}
|\Psi\rangle_{{\rm out}}&=&
(\tau a_{1{\rm out}}^\dagger-\rho a_{2{\rm out}}^\dagger)
(\rho a_{1{\rm out}}^\dagger+\tau a_{2{\rm out}}^\dagger)
|{\rm vac}\rangle\nonumber\\
&=&\{\tau\rho[(a_{1{\rm out}}^\dagger)^2-(a_{2{\rm out}}^\dagger)^2]+(\tau^2-\rho^2)a_{1{\rm out}}^\dagger a_{2{\rm out}}^\dagger\}|{\rm vac}\rangle. \label{1.3}\end{aligned}$$ In the case of exactly 50% beam splitting, for which $\tau=\rho=1/\sqrt{2}$, Eq. (\[1.3\]) reduces to $$|\Psi\rangle_{{\rm out}} =
(|2\rangle|0\rangle-|0\rangle|2\rangle)/\sqrt{2}, \label{1.4}$$ where $|n\>$ denotes an $n$-photon Fock state. (All the other modes are still vacuum modes.) In this case the part of the state with one photon in each output port is zero. The physical interpretation of this cancelation is that there are two paths leading to the same final state component $|1\rangle|1\rangle$: Either both photons transmit through the beam splitter or both photons reflect. The probability amplitudes for these two events have opposite signs \[$\tau^2$ and $-\rho^2$ in Eq. (\[1.3\])\], so they cancel. The resulting two-photon bunching is a basic property of nature and is at the heart of quantum information processing schemes using linear optics (beam splitters, polarizers, etc.) [@klm].
Notice that this type of HOM interference does not depend on any relative phase between the input states: Unlike coherent states or classical light waves (whose interference properties at a beam splitter certainly depend on the relative phase of the complex amplitudes), Fock states do not possess any physically relevant phase on which interference could depend.
In the original setting [@hon87], for the HOM effect to work, the photons must be in states that are identical in their polarization and spectral degrees of freedom (but their input states differ in their propagation directions). The question arises: Must two photons (or other bosons) have identical properties before they come together for this type of interference to occur? For example, assuming they have identical polarization states, must two photons initially be in identical energy (spectral) states in order to interfere?
In this paper, we show that the answer is negative: By using an active, frequency-shifting beam splitter, one can observe two-photon interference between photons of different color. The active process can be either reflection by a moving mirror (for small frequency shifts), or three-wave mixing in an optical crystal [@lou61; @van04; @alb04; @rou04] or four-wave mixing in an optical fiber (for large frequency shifts) [@mck04; @mck05; @gna06; @mec06]. The HOM effect is due to destructive interference in the output state, not the input state. We derive conditions on the forms of the input states that are required for high-visibility two-photon interference. These phenomena, in addition to demonstrating a fundamental property of light, will have applications in quantum information experiments involving photons of different color [@rod04; @fer04; @pfi04; @mck08].
Passive Beam Splitters
======================
Henceforth, we will use a simplified notation: Instead of writing relations that define the input operators in terms of output operators, as in Eq. (\[1.2\]), we will simply write $a_1^\dagger \mapsto \tau a_1^\dagger - \rho a_2^\dagger$ etc., to reflect how the input state changes to the output state in the most straightforward way. The operators $a_1$ and $a_2$ appearing on the right-hand side are thus meant to be output operators.
If the beamsplitter is a passive device that preserves the energy and spectrum of each photon separately, then the unitary transformation between the input and output channels can be written as $$\begin{aligned}
a_1^\dagger(\omega) &\mapsto &\tau a_1^\dagger(\omega) - \rho a_2^\dagger(\omega), \nonumber \\
a_2^\dagger(\omega) &\mapsto &\rho a_1^\dagger(\omega) + \tau a_2^\dagger(\omega). \label{2.1}\end{aligned}$$ Here we use creation operators $a_1^\dagger(\omega)$ for monochromatic modes with frequency $\omega$ impinging on one input port of the beam splitter and $a_2^\dagger(\omega)$ for modes impinging on the other input port. (We exploit a quasi one-dimensional picture of propagating single-photon wave packets. This picture is valid in the paraxial limit and when the radiation is sufficiently narrowband [@blow90; @lou00].) These operators satisfy the commutation relations $[a_k(\omega),a_k^\d(\omega')] = \delta(\omega - \omega')$ for $k = 1$ or 2. All other commutators are zero. Suppose now we have an input state containing two photons, one impinging on each input port, described by spectral amplitudes $\phi_1(\omega_1)$ and $\phi_2(\omega_2)$, respectively, which satisfy the normalization conditions $$\int d\omega_k|\phi_k(\omega_k)|^2=1. \label{2.2}$$ Then the input state is $$|\Psi\>_{{\rm in}} = \int d\omega_1 \phi_1(\omega_1)
a_1^\dagger(\omega_1)
\int d\omega_2
\phi_2(\omega_2)
a_2^\dagger(\omega_2)|{\rm vac}{\rangle}. \label{2.3}$$ This input state is mapped by the beam-splitter transformation onto the output state $$\begin{aligned}
|\Psi\>_{{\rm out}} = \int\int \hspace{-0.1in}&&d\omega_1 d\omega_2 \phi_1(\omega_1)\phi_2(\omega_2)[\tau\rho a_1^\dagger(\omega_1)a_1^\dagger(\omega_2) - \tau\rho a_2^\dagger(\omega_1)a_2^\dagger(\omega_2) \nonumber \\
&&+\ \tau^2a_1^\dagger(\omega_1)a_2^\dagger(\omega_2) - \rho^2 a_1^\dagger(\omega_2)a_2^\dagger(\omega_1)
]|{\rm vac}{\rangle}. \label{2.4}\end{aligned}$$ The last two terms can be made to cancel each other under the right conditions, and that complete destructive interference would correspond to HOM interference. The interference is perfect whenever $\tau^2 = \rho^2$ and $$\phi_1(\omega_1)\phi_2(\omega_2)=\phi_1(\omega_2)\phi_2(\omega_1) \label{2.5}$$ for [*all*]{} frequencies $\omega_1$ and $\omega_2$. This condition can be rewritten as $$\phi_1(\omega_1)/\phi_2(\omega_1) = \phi_1(\omega_2)/\phi_2(\omega_2). \label{2.6}$$ The left-hand and right-hand sides of Eq. (\[2.6\]) are functions of different variables, implying that they cannot actually depend on those variables. Hence, $$\phi_1(\omega_1)=C\phi_2(\omega_2), \label{2.7}$$ where $C$ is a constant. Normalization of the quantum states gives $|C|=1$, so $C$ is just the phase factor $\exp(i\theta)$ for some phase $\theta$. Equivalently, in the time domain, the Fourier transforms of the spectral density functions must satisfy $$\tilde{\phi}_1(t)=C\tilde{\phi}_2(t) \label{2.8}$$ for destructive interference.
From these conditions one may get the impression that it is important to have indistinguishable photons at the input. But that is a little misleading; what counts is indistinguishability at the output, as we will show. The same output state can be reached by two different paths, and it is the interference between the two paths that matters (as [*always*]{} in quantum mechanics).
Active Beam Splitters
=====================
Moving semi-transparent mirror
------------------------------
Consider a semi-transparent mirror moving to the left in one dimension. In the mirror frame, two channels with the same carrier frequency (color), and opposite propagation directions (right and left) impinge on the mirror \[Fig. 2($a$)\]. The input and output channels are related by Eqs. (\[2.1\]). However, in the laboratory frame \[Fig. 2($b$)\], light initally propagating to the right will be blue-shifted upon reflection, whereas light initially propagating to the left will be red-shifted upon reflection (and transmitted light does not change frequency).
![Interaction of two waves ($r$ and $l$) at a semi-transparent mirror ($M$). In the mirror frame, the waves have the same frequency and opposite directions of propagation (right and left). In the laboratory frame, the waves have different frequencies and directions.[]{data-label="pump"}](tci_f2a.eps "fig:"){width="2.4in"} ![Interaction of two waves ($r$ and $l$) at a semi-transparent mirror ($M$). In the mirror frame, the waves have the same frequency and opposite directions of propagation (right and left). In the laboratory frame, the waves have different frequencies and directions.[]{data-label="pump"}](tci_f2b.eps "fig:"){width="2.4in"}
It is convenient to label the right-propagating channel “red” and the left-propagating channel “blue.” (Of course, in order to shift actual red light to actual blue light and [*vice versa*]{}, the speed $v$ would have to be of order $c$.) The moving mirror transforms mode operators in the following way: a\^\_R() &&a\^\_R() - (/)a\^\_B(/),\
a\^\_B() &&()a\^\_R() + a\^\_B(). \[3.1.1\] The constant $\alpha$ is determined by the Doppler effect, which acts twice: $$\alpha = (1 - \beta)/(1 + \beta) < 1, \label{3.1.2}$$ where the normalized mirror speed is $\beta=v/c$. The factors of $\alpha$ in Eqs. (\[3.1.1\]), which are absent from Eqs. (\[2.1\]), are required to conserve the number of photons, because red modes in the frequency interval $d\omega_R$ are coupled to blue modes in the interval $d\omega_B = d\omega_R/\alpha$. Transformation (\[3.1.1\]) is unitary, preserving the commutation relations for the transformed operators. For example, &&\[a\_B(),a\^\_B(’)\] = (-’),\
&&\[a\_B(),a\^\_R(’)\] = 0, \[3.1.3\] where we used $\delta[\alpha(\omega-\omega')] = \delta(\omega-\omega')/ \alpha.$ Although this transformation preserves the number of photons, it does not preserve their energy, thus earning the right to be called active.
Now suppose we have one photon in each input mode, described by wave packets centered around red and blue frequencies, as the input state: $$|\Psi{\rangle}_{{\rm in}} = \int d{\omega_R}\phi_R(\omega_R)a^\dagger_R(\omega_R)
\int d{\omega_B} \phi_B(\omega_B)a^\dagger_B(\omega_B)|{\rm vac}{\rangle}, \label{3.1.4}$$ where the spectral amplitudes satisfy the normalization conditions (\[2.2\]). This state is transformed by the moving mirror into |\_[[out]{}]{} = &&d[\_R]{}d[\_B]{} \_R(\_R)\_B(\_B)\[a\^\_R(\_R) - (/)a\^\_B(\_R/)\]\
&&|[vac]{}. \[3.1.5\] The interesting effect is that there can be destructive interference in the output terms where one red and one blue photon emerge. There are two amplitudes for reaching that final state with opposite signs; the (necessary and sufficient) condition for exact destructive interference for all frequencies $\omega_R$ and $\omega_B$ is $$\label{23}
\tau^2\phi_R(\omega_R)\phi_B(\omega_B)=
\rho^2\phi_R(\alpha\omega_B)\phi_B(\omega_R/\alpha). \label{3.1.6}$$ This condition implies that $\tau^2 = \rho^2$ and $$\label{24}
\phi_R(\omega_R)/
\phi_B(\omega_R/\alpha)=\phi_R(\alpha\omega_B)/\phi_B(\omega_B) \label{3.1.7}$$ for all frequencies $\omega_R$ and $\omega_B$, which in turn implies that \_R(\_R) &= &C\_B(\_R/),\
\_B(\_B) &= &C\^[-1]{}\_R(\_B), \[3.1.8\] with $C$ a constant independent of $\omega_R$ and $\omega_B$. These two conditions are really the same. Normalization gives $|C|^2=1/\alpha$, so we can write $C=\exp(i\theta)/\sqrt{\alpha},$ with $\theta$ some constant phase. In the time domain this condition gives $$\tilde{\phi}_B(t):=\int d\omega_B \phi_B(\omega_B)\exp(i\omega_B t) = C^{-1}
\int d(\omega_R/\alpha) \phi_R(\omega_R)\exp(i \omega_R t/\alpha), \label{3.1.9}$$ which means that $$\sqrt{\alpha}\exp(i\theta)\tilde{\phi}_B(t)=
\tilde{\phi}_R( t/\alpha). \label{3.1.10}$$ If the input wavepackets satisfy these relations, then the output state of the moving mirror will contain either two red photons moving to the right, or two blue photons moving to the left, but never one of each color.
Frequency translation in fiber, simple version
----------------------------------------------
Now consider four-wave mixing in a fiber. Among the many processes that may occur, we are interested in frequency translation (Bragg scattering), in which two strong, classical pump waves ($p$ and $q$) convert “signal” ($s$) photons into “idler” ($r$) photons, or [*vice versa*]{} [@mck04; @mck05; @gna06; @mec06]. The frequency-matching condition for this process, which is illustrated in Fig. 3, is \_q + \_r = \_s + \_p. \[3.2.1\] It follows from Eq. (\[3.2.1\]) that the difference between the signal and idler frequencies, $\omega_s - \omega_r = \Omega$, equals the difference between the pump frequencies, $\omega_q - \omega_p$. It is convenient to refer to the idler as the “red” sideband and the signal as the “blue” sideband. (Once again, we use these terms just to indicate two different colors.) If the pumps are monochromatic (CW), then each red frequency is coupled to one blue frequency. If the red frequencies all lie within the phase-matched bandwidth of the convertor, the red-to-blue conversion efficiency is frequency independent. If the pump powers and fiber length are chosen judiciously, the probability for frequency translation to occur for a given photon is 50%, with the remaining 50% probability assigned to the photon’s frequency staying the same.
![Two classical pumps ($p$ and $q$) can drive a four-wave mixing process in which a single signal ($s$) photon is frequency-converted to a single idler ($r$) photon, or [*vice versa*]{}.[]{data-label="pump"}](tci_f3ax.eps "fig:"){width="2.4in"} ![Two classical pumps ($p$ and $q$) can drive a four-wave mixing process in which a single signal ($s$) photon is frequency-converted to a single idler ($r$) photon, or [*vice versa*]{}.[]{data-label="pump"}](tci_f3bx.eps "fig:"){width="2.4in"}
The transformation between the red and blue modes is [@mck04; @mck05] a\_R\^(\_R) &&a\_R\^(\_R) - a\_B\^(\_R+),\
a\_B\^(\_B) &&a\_R\^(\_B-) + a\_B\^(\_B). \[3.2.2\] (An equivalent transformation governs frequency up-conversion in a crystal [@lou61].) This transformation is unitary. By following the steps between Eqs. (\[3.1.4\]) and (\[3.1.6\]), one finds that the HOM interference is perfect when $$\tau^2\phi_R(\omega_R)\phi_B(\omega_B)=
\rho^2\phi_R(\omega_B-\Omega)\phi_B(\omega_R+\Omega). \label{3.2.3}$$ Condition (\[3.2.3\]) implies that $\tau^2 = \rho^2$ and \_R(\_R) &= &C\_B(\_R+),\
\_B(\_B) &= &C\^[-1]{}\_R(\_B-), \[3.2.4\] with $C$ a constant, independent of $\omega_R$ and $\omega_B$. Once again, these two conditions are the same. Normalization gives $|C|=1$, and hence one can write $C=\exp(i\theta)$. In the time domain, $$\exp(i\theta-i\Omega t)\tilde{\phi}_B(t) = \tilde{\phi}_R(t), \label{3.2.5}$$ which means that the temporal shapes should be identical, with an overall frequency difference equal to $\Omega$. Equation (\[3.2.5\]) provides necessary and sufficient conditions for frequency translation to display HOM interference in this simple case: Two input photons, one red and one blue, traversing a nonlinear fiber with the appropriate amount of frequency translation, lead to an output state which contains either two red photons or two blue photons, but never one of each color.
Frequency translation in fiber, general version
-----------------------------------------------
If one of the pumps is pulsed, or both pumps are pulsed, each frequency component of the signal is coupled to many frequency components of the idler and [*vice versa*]{} [@mck09]. In either case, the transformation over all frequencies is of the form $$a^\dagger(\omega)\mapsto \int d\omega' G(\omega,\omega')a^\dagger(\omega'), \label{3.3.1}$$ where the kernel $G$ satisfies the unitarity condition $$\int d\omega' G(\omega,\omega') G^*(\omega'',\omega')=\delta(\omega-\omega''). \label{3.3.2}$$ Let us split the frequency interval into two non-overlapping parts, “red” and “blue,” and let us correspondingly write $a(\omega)=a_R(\omega_R)$ whenever $\omega$ is a “red” frequency, and $a(\omega)=a_B(\omega_B)$ whenever $\omega$ is a “blue” frequency. Then we can rewrite the kernel in the block-matrix form $$\left[ \begin{array}{cc} G_{RR}(\omega_R,\omega'_R) & G_{RB}(\omega_R,\omega'_B) \\
G_{BR}(\omega_B,\omega'_R)&
G_{BB}(\omega_B,\omega'_B) \end{array} \right], \label{3.3.3}$$ where the constituent (Green) kernels $G_{RR}$, $G_{RB}$, $G_{BR}$ and $G_{BB}$ describe red-red transmission, blue-red conversion, red-blue conversion and blue-blue transmission, respectively. The Green kernels satisfy the unitarity conditions $$\begin{aligned}
\label{3.3.4}
\int d\omega_R' G_{RR}(\omega_R,\omega'_R) G^*_{RR}(\omega_R'',\omega_R')
+
\int d\omega_B' G_{RB}(\omega_R,\omega'_B) G^*_{RB}(\omega_R'',\omega_B')
&=&\delta(\omega_R-\omega_R''), \nonumber \\
\int d\omega_R' G_{BR}(\omega_B,\omega'_R) G^*_{RR}(\omega_R'',\omega_R')
+
\int d\omega_B' G_{BB}(\omega_B,\omega'_B) G^*_{RB}(\omega_R'',\omega_B')
&=&0,
$$ and similar conditions in which $R \leftrightarrow B$. In terms of these kernels, the transformations between the red and blue modes are $$\begin{aligned}
a_R^\dagger(\omega_R) &\mapsto &\int d\omega'_R G_{RR}(\omega_R,\omega'_R)a_R^\dagger(\omega'_R) +
\int d\omega'_B G_{RB}(\omega_R,\omega'_B)a_B^\dagger(\omega'_B), \nonumber \\
a_B^\dagger(\omega_B) &\mapsto &\int d\omega'_R G_{BR}(\omega_B,\omega'_R)a_R^\dagger(\omega'_R) +
\int d\omega'_B G_{BB}(\omega_B,\omega'_B)a_B^\dagger(\omega'_B). \label{3.3.5}\end{aligned}$$ The interference of one red and one blue photon in the final state is perfectly destructive when $$\begin{aligned}
\int\int \hspace{-0.1in} &&d \omega_R d\omega_B \phi_R(\omega_R) \phi_B(\omega_B) [G_{RR}(\omega_R,\omega_R')
G_{BB}(\omega_B,\omega_B') \nonumber \\
&&+\ G_{RB}(\omega_R,\omega_B') G_{BR}(\omega_B,\omega_R')] = 0 \label{3.3.6}\end{aligned}$$ for all frequencies $\omega'_R$ and $\omega'_B$. By rearranging the terms in Eq. (\[3.3.6\]), so that functions of $\omega_R'$ and $\omega_B'$ appear on different sides, one obtains the interference conditions d\_R \_R(\_R)G\_[RR]{}(\_R,\_R’) &= &Cd\_B \_B(\_B)G\_[BR]{}(\_B,\_R’),\
d\_B \_B(\_B)G\_[BB]{}(\_B,\_B’) &= &-C\^[-1]{}d\_R \_R(\_R)G\_[RB]{}(\_R,\_B’), \[3.3.7\] where $C$ is a constant, independent of frequency. The first condition states that the spectral amplitude of red output photons arising from red input photons be equal (up to a constant) to the spectral amplitude of red output photons arising from blue input photons. The second condition is similar, but for blue output photons.
In general, though, these conditions will not be fulfilled with given pumps in a given fiber. One needs to design the four-wave mixing process in such a way that with 50% probability a red photon is either converted to a blue photon or remains a red photon with a possibly altered spectral wave function. The opposite transformations must hold for the incoming blue photon. The fiber dispersion, pump pulses, and the input red and blue wave packets must all be designed properly to ensure that the output red and blue wave packets are identical for either input channel. (This is work in progress.)
More precisely, if one uses the unitarity conditions (\[3.3.4\]) to rewrite and combine Eqs. (\[3.3.7\]), one obtains the interference conditions $$\begin{aligned}
\label{cond}
\phi_R(\omega_R)&=&C
\int d\omega_B
K(\omega_R,\omega_B)\phi_B(\omega_B), \nonumber \\
\phi_B(\omega_B)&=&C^{-1}
\int d\omega_R
K^*(\omega_R,\omega_B)\phi_R(\omega_R),\label{3.3.8}\end{aligned}$$ where the HOM kernel $K(\omega_R,\omega_B)$ is $$K(\omega_R,\omega_B)=\int d\omega'
[G^*_{RR}(\omega_R,\omega')G_{BR}(\omega_B,\omega') - G^*_{RB}(\omega_R,\omega')G_{BB}(\omega_B,\omega')]. \label{3.3.9}$$ Recall that every hermitian kernel can be decomposed in terms of its eigenvalues, which are real, and a single set of eigenfunctions, which are orthonormal. In a similar way, every complex kernel has the Schmidt (singular-value) decomposition [@eke95; @law00] $$K(\omega_R,\omega_B) =\sum_n \sigma_n R_n(\omega_R) B_n^*(\omega_B), \label{3.3.10}$$ where the coefficients (singular values) $\sigma_n$ are real and non-negative, and the mode functions $\{R_n(\omega_R)\}$ and $\{B_n(\omega_B)\}$ are self-orthonormal. Subroutines are available, which determine the mode functions numerically. By substituting decomposition (\[3.3.10\]) into Eqs. (\[3.3.8\]), one finds that the interference conditions can be fulfilled only if there is at least one unit coefficient among the Schmidt coefficients. If there is exactly one unit coefficient, say $\sigma_1=1$, then one must have $\phi_R(\omega_R) = \exp(i\theta) R_1(\omega_R)$ and $\phi_B(\omega_B) = B_1(\omega_B)$ in order for complete interference to occur. If there are multiple unit coefficients, one can take superpositions of the corresponding modes. If the kernel possesses [*only*]{} unit coefficients, then one can choose any red spectral amplitude $\phi_R(\omega_R)$ and find a corresponding blue spectral amplitude. The simple cases of the preceding subsections are, in fact, such degenerate cases, whenever $\tau^2 = \rho^2$. In the general case, which corresponds to pulsed pumps, one expects that only certain red and blue wavepackets, of finite bandwidth and duration, lead to perfect HOM interference. In all of these cases, $C$ must again be the phase factor $\exp(i\theta)$. Thus, the fiber and pump requirements for perfect HOM interference may be stated in a different and more succinct way: The HOM kernel must have at least one unit singular value.
One can analyze the general case in an alternative way, by giving the Schmidt decompositions of the Green kernels directly, rather than the HOM kernel $K$. By using the results of [@was], which are reviewed in the Appendix of this paper, one finds that $$\begin{aligned}
G_{RR}(\omega_R,\omega_R')
&=&\sum_n \tau_n V_n(\omega_R) v_n^*(\omega'_R), \nonumber \\
G_{RB}(\omega_R,\omega_B')
&=&-\sum_n \rho_n V_n(\omega_R) w_n^*(\omega'_B), \nonumber \\
G_{BR}(\omega_B,\omega_R')
&=&\sum_n \rho_n W_n(\omega_B) v_n^*(\omega'_R), \nonumber \\
G_{BB}(\omega_B,\omega_B')
&=&\sum_n \tau_n W_n(\omega_B) w_n^*(\omega'_B), \label{3.3.13}\end{aligned}$$ where the non-negative coefficients $\tau_n$ and $\rho_n$ satisfy the auxiliary equations $\tau_n^2 + \rho_n^2 = 1$, and the sets of mode functions $\{V_n\}$, $\{v_n\}$, $\{W_n\}$, and $\{w_n\}$ are self-orthonormal. By combining Eqs. (\[3.3.9\]) and (\[3.3.13\]), one obtains the alternative decomposition K(\_R,\_B) = 2\_n \_n\_n V\_n\^\*(\_R) W\_n(\_B). \[3.3.14\] It follows from Eqs. (\[3.3.10\]) and (\[3.3.14\]) that $\sigma_n = 2\rho_n\tau_n$, $R_n = V_n^*$ and $B_n = W_n^*$. Notice that $\sigma_n \le 1$, so the HOM condition $\sigma_n = 1$ represents a special (ideal) case.
The coefficients $\tau_n$ and $\rho_n$ can be interpreted as generalized beam-splitter coefficients, with $\tau_n$ playing the role of a transmission coefficient (not changing the color) and $\rho_n$ the corresponding reflection coefficient (changing the color), for each mode $n$. Specifically, the Schmidt-mode expansions a\_R\^()|\_[in]{} &= &\_n a\_n\^V\_n(),\
a\_B\^()|\_[in]{} &= &\_n b\_n\^W\_n(),\
a\_R\^()|\_[out]{} &= &\_n c\_n\^v\_n(),\
a\_B\^()|\_[out]{} &= &\_n d\_n\^w\_n(), \[3.3.15\] allow one to rewrite the continuous input and output operators for red and blue light in terms of the discrete operators $\{a_n,b_n,c_n,d_n\}$ [@blow90]. In terms of these specially constructed discrete modes, the output operators are related to the input operators in the simple, pair-wise manner $$\begin{aligned}
a_n^\d &= &\tau_nc_n^\d - \rho_nd_n^\d, \nonumber \\
b_n^\d &= &\rho_nc_n^\d + \tau_nd_n^\d. \label{3.3.16}\end{aligned}$$ Equations (\[3.3.16\]) are equivalent to the two-mode equations (\[1.2\]). As expected, a unit $\sigma_n$ corresponds to a 50/50 beam-splitter transformation with $\tau_n=\rho_n=1/\sqrt{2}$.
Moreover, description (\[3.3.13\]) allows us to quantify the effect of imperfections in frequency translation. That is, suppose that the fiber process we actually designed does not have any coefficient $\tau_n$ equal to $1/\sqrt{2}$. Then the best we can do is to use the mode that corresponds to the transmission coefficient that is closest to $1/\sqrt{2}$, say $\tau_1\approx 1/\sqrt{2}$. In fact, if we choose in that case $\phi_R(\omega_R)=V_1^*(\omega_R)$ and $\phi_B(\omega_B)=W_1^*(\omega_B)$, then the amplitude of the component of the output state containing one red and one blue photon is $$\langle {\rm vac}|a_R(\omega_R)a_B(\omega_B)|\Psi\rangle_{\rm out} = (\tau_1^2-\rho_1^2) V_1^*(\omega_R)W_1^*(\omega_B). \label{3.3.17}$$ This equation generalizes the corresponding term in Eq. (\[1.3\]). The deviation from perfect two-photon interference can be quantified as the probability to find one red and one blue photon (at any red-blue frequency pair) in the output, and (\[3.3.17\]) shows that this probability is $P_{RB}=(\tau_1^2-\rho_1^2)^2$.
Conclusions
===========
In this paper, we showed that one can observe interference between two photons of different color. These photons arise independently from separate sources before they come together and interfere. The important point is that both photons end up in the same final spectral state (or mode), even though they began in two spectrally distinct states. It is not necessary for them to be identical initially, except in the obvious sense that they are both photons, which are fundamentally identical particles (if one takes a particle view of photons). For example, if one “red” and one “blue” photon enter the active interferometer, then either two red or two blue photons exit, but never one of each color.
Observing such bosonic bunching of photons, whose colors are different initially, requires reversible and symmetric inelastic scattering processes. Examples of such processes include three-wave mixing (wavelength down-conversion) in a crystal [@van04; @alb04; @rou04] and four-wave mixing (Bragg scattering) in a fiber [@gna06; @mec06]. Both types of process can shift photon wavelengths by several hundred nanometers, with high photon-conversion efficiencies. Such two-photon interference should find applications in quantum information science, in schemes using entanglement between optical modes of different color.
MGR wishes to dedicate his contribution to this paper to the memory of Krzysztof Wodkiewicz — a friend for nearly three decades. He will always value the positive influence that Krzysztof had as a friend, teacher and collaborator. This work was supported in part by NSF grant ECCS-0802109.
Decomposition of a Unitary Transformation
=========================================
In this appendix, it is shown that a $2n$-mode unitary transformation is equivalent to $n$ 2-mode beam-splitter-like transformations. As an illustrative example, consider the effects of a rotator (circular phase-shifter) on a monochromatic wave [@shu62; @mck06]. Let $A_+$ and $A_-$ be the components of the $2 \times 1$ amplitude vector $A$ relative to the circularly-polarized basis vectors $E_+ = [1,i]^t/2^{1/2}$ and $E_- = [1,-i]^t/2^{1/2}$, respectively, and let $\theta$ be one-half of the differential phase shift. Then the input and output vectors are related by matrix equation A() = M()A(0), \[a1\] where the $2 \times 2$ transfer matrix M() = \[a2\] is diagonal, with complex-conjugate eigenvalues. Now let $A_x$ and $A_y$ be the amplitude components relative to the linearly-polarized basis vectors $E_x = [1,0]^t$ and $E_y = [0,1]^t$, respectively. Then the associated transfer matrix M() = \[a3\]. What appears as phase shifts relative to one basis appears as rotation relative to another. The second transfer matrix also describes the effects of a beam splitter on two input waves ($\tau = \cos\theta$ and $\rho = \sin\theta$).
Now consider the $2n \times 2n$ unitary matrix M = , \[a4\] where the $n \times n$ matrices $A$, $B$, $C$ and $D$ describe signal-signal transmission, idler-signal conversion, signal-idler conversion and idler-idler transmission, respectively. Our derivation of the beam-splitter decomposition of a unitary transformation is similar to the derivation of the Bloch–Messiah decomposition of a squeezing transformation [@blo62; @bra05; @mck09]. The unitarity conditions are &= &\
&= &. \[a5\] The complex matrix $A$ has the singular-value decomposition (SVD) A = U\_aD\_aV\_a\^, \[a6\] where $U_a$ and $V_a$ are unitary, and $D_a = \diag(\alpha_j)$. In the standard SVD, the $\alpha_j$ are non-negative. However, one can also allow them to be negative. Similar SVDs exist for the other transmission matrices, in which $D_b = \diag(\beta_j)$, $D_c = \diag(\gamma_j)$ and $D_d = \diag(\delta_j)$. The first diagonal condition in Eq. (\[a5\]) implies that U\_aD\_a\^2U\_a\^+ U\_bD\_b\^2U\_b\^= I, \[a7\] from which it follows that $U_a = U_b$ and $\alpha_j^2 + \beta_j^2 = 1$. The other diagonal conditions imply that $U_c = U_d$, $V_a = V_c$, $V_b = V_d$, $\alpha_j^2 = \delta_j^2$ and $\beta_j^2 = \gamma_j^2$. The off-diagonal conditions all imply that $\delta_j = \alpha_j$ and $\gamma_j = -\beta_j$. Let $E_{1j}$ be the $j$th column of $U_a$, $E_{2j}$ be the $j$th column of $U_d$, $F_{1j}$ be the $j$th column of $V_c$ and $F_{2j}$ be the $j$th column of $V_b$ (so $E_{1j}$ is an eigenvector of $AA^\d$, $E_{2j}$ is an eigenvector of $DD^\d$, $F_{1j}$ is an eigenvector of $C^\d C$ and $F_{2j}$ is an eigenvector of $B^\d B$). Then these vectors form self-orthonormal sets and M = \_j . \[a8\] Equation (\[a8\]) is the beam-splitter decomposition [@was]. $F_{1j}$ and $F_{2j}$ are the input eigenvectors, whereas $E_{1j}$ and $E_{2j}$ are the output eigenvectors. It follows from Eq. (\[a8\]) that M\^= \_j . \[a9\] Thus, if the forward transformation is determined by the parameters $(\alpha_j,\beta_j,-\beta_j,\alpha_j)$, the backward transformation is determined by the parameters $(\alpha_j,-\beta_j,\beta_j,\alpha_j)$, just like the two-mode transformation (\[a3\]). Furthermore, the roles of the input and output eigenvectors are reversed. Equations (\[a8\]) and (\[a9\]) remain valid if one replaces the aforementioned real parameters by the complex parameters $(\alpha_j,\beta_j,-\beta_j^*,\alpha_j^*)$ and $(\alpha_j^*,-\beta_j,\beta_j^*,\alpha_j)$ in the forward and backward transformations, respectively, where $|\alpha_j|^2 + |\beta_j|^2 = 1$.
The Green kernels of the text are related to the constituent matrices of this appendix by the relations $G_{rr} = A^t$, $G_{rb} = C^t$, $G_{br} = B^t$ and $G_{rr} = D^t$, from which it follows that the HOM kernel $K = A^\d B - C^\d D$. By combining this result with decomposition (\[a8\]), one finds that K = 2\_j \_j\^\*\_j F\_[1j]{}F\_[2j]{}\^. \[a10\] Equation (\[a10\]) is consistent with Eq. (\[3.3.14\]), which defines the mode functions of the text ($\phi_r$ and $\phi_b$) indirectly. These mode functions are defined directly by the interference conditions (\[3.3.8\]). In the matrix notation of this appendix, they are the eigenvectors of $KK^\d$ and $K^\d K$, respectively. By combining the formula for $K$ with conditions (\[a5\]), one finds that KK\^&= &4A\^AC\^C, \[a11\]\
K\^K &= &4B\^BD\^D. \[a12\] $A^\d A$ and $C^\d C$ have the eigenvalues $|\alpha_j|^2$ and $|\beta_j|^2$, respectively, and the (common) eigenvectors $F_{1j}$, whereas $B^\d B$ and $D^\d D$ have the eigenvalues $|\beta_j|^2$ and $|\alpha_j|^2$, respectively, and the (common) eigenvectors $F_{2j}$. Hence, the mode functions $\phi_{rj} = F_{1j}$ and $\phi_{bj} = F_{2j}$. The associated singular values $\sigma_j = 2|\alpha_j\beta_j|$ are the square roots of the (common) eigenvalues of $KK^\d$ and $K^\d K$. These results are consistent with Eq. (\[3.3.10\]).
[99]{}
C. K. Hong, Z. Y. Ou, and L. Mandel, [*Measurement of subpicosecond time intervals between two photons by interference*]{}, Phys. Rev. Lett. [**59**]{}, 2044–2043 (1987).
K. J. Blow, R. Loudon, S. J. D. Phoenix, and T. J. Shepherd, [*Continuum fields in quantum optics*]{}, Phys. Rev. A [**42**]{}, 4102–4114 (1990).
R. Loudon, [*The Quantum Theory of Light, 3rd Ed.*]{} (Oxford University Press, 2000).
E. Knill, R. Laflamme, and G. J. Milburn, [*A scheme for efficient quantum computation with linear optics*]{}, Nature [**409**]{}, 46–52 (2001).
W. H. Louisell, A. Yariv and A. E. Siegman, [*Quantum fluctuations and noise in parametric processes*]{}, Phys. Rev. [**124**]{}, 1646–1654 (1961).
A. P. VanDevender and P. G. Kwiat, [*High-efficiency single photon detection via frequency up-conversion*]{}, J. Mod. Opt. [**51**]{}, 1433–1445 (2004).
M. A. Albota and F. N. C. Wong, [*Efficient single-photon counting at 1.55$\mu$m by means of frequency upconversion*]{}, Opt. Lett. [**29**]{}, 1449–1451 (2004).
R. V. Roussev, C. Langrock, J. R. Kurz and M. M. Fejer, [*Periodically poled lithium niobate waveguide sum-frequency generator for efficient single-photon detection at telecommunication wavelengths*]{}, Opt. Lett. [**29**]{}, 1518–1520 (2004).
C. J. McKinstrie, S. Radic and M. G. Raymer, [*Quantum noise properties of parametric amplifiers driven by two pump waves*]{}, Opt. Express [**12**]{}, 5037–5066 (2004).
C. J. McKinstrie, J. D. Harvey, S. Radic and M. G. Raymer, [*Translation of quantum states by four-wave mixing in fibers*]{}, Opt. Express [**13**]{}, 9131–9142 (2005).
A. H. Gnauck, R. M. Jopson, C. J. McKinstrie, J. C. Centanni and S. Radic, [*Demonstration of low-noise frequency conversion by Bragg scattering in a fiber*]{}, Opt. Express [**14**]{}, 8989–8994 (2006).
D. Méchin, R. Provo, J. D. Harvey and C. J. McKinstrie, [*180-nm wavelength conversion based on Bragg scattering in a fiber*]{}, Opt. Express [**14**]{}, 8995–8999 (2006).
A. V. Rodionov and A. S. Chirkin, [*Entangled photon states in consecutive nonlinear optical interactions*]{}, JETP Lett. [**79**]{}, 253–256 and 582 (2004).
A. Ferraro, M. G. A. Paris, M. Bondani, A. Allevi, E. Puddu and A. Andreoni, [*Three-mode entanglement by interlinked nonlinear interactions in optical $\chi^{(2)}$ media*]{}, J. Opt. Soc. Am. B [**21**]{}, 1241–1249 (2004).
O. Pfister, S. Feng, G. Jennings, R. Pooser and D. Xie, [*Multipartite continuous-variable entanglement from concurrent nonlinearities*]{}, Phys. Rev. A [**70**]{}, 020302 (2004).
C. J. McKinstrie, S. J. van Enk, M. G. Raymer and S. Radic, [*Multicolor multipartite entanglement produced by vector four-wave mixing in a fiber*]{}, Opt. Express [**16**]{}, 2720–2739 (2008).
A. Ekert and P. L. Knight, [*Entangled quantum systems and the Schmidt decomposition*]{}, Am. J. Phys. [**63**]{}, 415–423 (1995).
C. K. Law, I. A. Walmsley and J. H. Eberly, [*Continuous frequency entanglement: Effective finite Hilbert space and entropy control*]{}, Phys. Rev. Lett. [**84**]{}, 5304–5307 (2000).
W. Wasilewski and M. G. Raymer, [*Pairwise entanglement and readout of atomic-ensemble and optical wave-packet modes in traveling-wave Raman interactions*]{}, Phys. Rev. A [**73**]{}, 063816 (2006).
W. A. Shurcliff, [*Polarized Light*]{} (Harvard University Press, 1962).
C. J. McKinstrie and H. Kogelnik, [*Nonlinear wave propagation in a rapidly-spun fiber*]{}, Opt. Express [**14**]{}, 8072–8087 (2006).
C. Bloch and A. Messiah, [*The canonical form of an asymmetric tensor and its application to the theory of superconductivity*]{}, Nucl. Phys. [**39**]{}, 95–106 (1962).
S. L. Braunstein, [*Squeezing as an irreducible resource*]{}, Phys. Rev. A [**71**]{}, 055801 (2005).
C. J. McKinstrie, [*Unitary and singular value decompositions of parametric processes in fibers*]{}, Opt. Commun. [**282**]{}, 583–593 (2009).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the data obtained from XMM-Newton, we show the gradual evolution of two periodicities of $\approx$4500 s and $\approx$2200 s in the decay phase of the flare observed in a solar analog EK Dra. The longer period evolves firstly for first 14 ks, while the shorter period evolves for next 10 ks in the decay phase. We find that these two periodicities are associated with the magnetoacoustic waves triggered in the flaring region. The flaring loop system shows cooling and thus it is subjected to the change in the scale height and the acoustic cut-off period. This serves to filter the longer period magnetoacoustic waves and enables the propagation of the shorter period waves in the later phase of the flare. We provide the first clues of the dynamic behaviour of EK Dra’s corona which affects the propagation of waves and causes their filtering.'
author:
- 'A. K. Srivastava'
- 'J. C. Pandey and Subhajeet Karmakar'
- 'Partha Chowdhury and Y.-J. Moon'
- Marcel Goossens
- 'P. Jelínek'
- 'M. Mathioudakis'
- 'J.G. Doyle'
- 'B.N. Dwivedi'
title: First Evidence of the Frequency Filtering of Magnetoacoustic Waves in the flaring star EK Dra
---
Introduction
============
Magnetohydrodynamic (MHD) waves and oscillation are important physical phenomena arose in the magnetized plasma that provide a significant diagnostics (e.g., magnetic field, density and magnetic scale heights, transport phenomena, etc) of the localized solar and stellar coronae [@2005LRSP....2....3N; @2009SSRv..149..119N]. The waves and oscillations are ubiquitous in the solar atmosphere, and are studied both observationally and theoretically in different magnetic structures. There are several pieces of evidence that they also exist in the atmosphere of similar type magnetically active stars . These waves or periodic magnetic reconnection may evolve there during powerful flaring energy release, and can modulate various emissions across the electromagnetic spectrum of star .
Likewise solar atmosphere, these MHD waves may most likely excite in the structured magnetic tubes of the stellar atmosphere. Solar atmosphere offers significant coupling, conversion, reflection of MHD modes due to its complex magnetic field and plasma structuring, therefore, imposes a constraint on propagation of these waves and transport of associated power (e.g., McAteer et al. 2003; Schrijver & Title, 2003; De Pontieu et al., 2004; Centeno et al., 2006; Vecchio et al., 2007; Khomenko & Cally, 2012, and references cited therein).
In the solar atmosphere, most powerful acoustic oscillations are associated with five min photospheric motions, which may convert into magnetoacoustic waves into the region where plasma beta becomes one (Fedun et al., 2009). However, the propagation of 5 min acoustic oscillations are highly debated, and depends upon certain properties of the plasma as well as magnetic field of the localized solar atmosphere (e.g., De Pontieu et al., 2004; Khomenko et al., 2008; Yuan et al., 2014). The structured solar atmosphere works as a wave-guide for various MHD modes, and also serve as a frequency filter transporting selective frequencies and associated energies (e.g., Fedun et al., 2011; Murawski et al., 2016). However, this fact has never been explored as far as the stellar coronae of similar stars are considered.
In this paper, we have investigated the quasi-periodic oscillation detected during the decay phase of a flaring event in an active star EK Dra. EK Dra is a young solar analog which provides an opportunity to study the magnetic activity of the infant Sun. Except the age, EK Dra has similar properties to that of the Sun. It is a G1.5V star with 2.75 d rotation period and located at 34 pc away from the earth. We study for the first time that the flaring atmosphere in EK Dra filters the evolved frequencies, and thus has the structured atmosphere likewise Sun. The observational data and reduction are described in section 2. Results are outlined in section 3 and Section 4 depicts the theoretical interpretations. Discussion along with the conclusions are given in the last section.
Observations
============
EK Dra had been observed for 54.9 ks by the XMM-Newton satellite using the European Photon Imaging Camera (EPIC) and Reflection Grating Spectrometer(RGS) at 14:01:58 UT on December 30, 2000. The EPIC is composed of three CCDs behind three X-ray telescopes (Jansen et al. 2001); the twin metal oxide semiconductor (MOS) CCDs, MOS1 and MOS2 (Turner et al. 2001), and one p-n junction CCD, PN (Str[ü]{}der et al. 2001). The RGS is composed of two identical grating spectrometers, RGS1, and RGS2, behind different mirrors (den Herder et al. 2001). The data were reduced using the standard Science Analysis System (SAS) of XMM-Newton version 16.0.0 with updated calibration files. X-ray light curves were generated from on-source counts obtained from circular regions with a radius of 40 arcsecs around the source. The background was chosen from source-free regions on the detectors surrounding the source. The detailed spectral analyses of EPIC and RGS data were performed by Scelsi et al. (2005).
Results
=======
X-ray light curve and loop parameters
-------------------------------------
Top panel of Fig. \[fig:lc\] shows the background subtracted X-ray light curves of EK Dra in the 0.3-10.0 keV energy band. A sudden enhancement is followed by a gradual decay indicating a flare during the observations. The intensity reached a maximum value at 13.4 ks from the start of the MOS observation. The flare lasted for $\sim 10$ ks. The continuous line shows the best fit exponential function and we found the e-folding decay time ($\tau_d$) of $4221 \pm 98~{\mathrm{seconds}}$. The detrended light curve during the decay phase is shown in the bottom panel of the Fig. \[fig:lc\].
Time-resolved spectral analysis of the observed flare by Scelsi et al. (2005) showed that the temperature and emission measure during the flare follow the behaviour of light curve. The temperature had a peak before the emission measure. The average maximum temperature was found to be $42^{+1.6}_{-1.0}$ MK. The maximum temperature $T_{max}$ was calculated from the observed maximum temperature using method of Reale (2007), and it is estimated to be $91\pm27$ MK. Scelsi et al. (2005) have also derived the equilibrium temperature of the flaring plasma as 8.4 MK at the end of the decay phase. According to the state-of-art hydrodynamic model of Reale et al. (1997), the loop length is estimated as
$$L ~(~cm~)~ = 5.27\times10^3 \frac{\tau_d ~ \sqrt{T_{max}}}{F(\zeta)} ~~~~~~ {\rm for ~0.35 <\zeta < 1.6}
\label{eq:loop}$$
{width="175mm"} {width="175mm"}
The dimensionless correction factor is $F(\zeta) = \frac{0.51}{\zeta-0.35}+1.36$; where $\zeta$ is the slope of the log $\sqrt{\rm EM} $ log $T$ diagram (Reale 2007). Scelsi et al. (2005) have derived a value of 1.0 for $\zeta$ using two data points during the decay of the flare. As temperature and emission measure are decaying during the post flare phase, therefore, by considering all four data point during this phase (see Table B.1 and Figure B.1 of of Scelsi 2005), we derived the value of $\zeta$ as $1.0\pm 0.5$. Using equation \[eq:loop\], we derive a loop length of $9.9 \pm 3.1 \times 10^{10}$ cm. Table \[tab:parameter\] summarizes the derived loop parameters.
The electron density of EK Dra was derived by using He-like triplets from O[vii]{}. The most intense He-like lines correspond to the transitions between the $n = 2$ shell and the $n = 1$ ground state shell. The excited state transitions to the ground state are called resonance ($r$), inter-combination ($i$) and forbidden ($f$) lines. In the X-ray spectra, the ratio of fluxes in forbidden and inter-combination lines ($R = f/i$) is sensitive to the electron density (Porquetd et al. 2001). We have used the CHIANTI atomic database version 8.0.2 (Dere et al. 1997; Delzanna et al. 2015) to derive the density from $R-$ratios. Using $R = 3.0\pm1.7$ and a temperature of 1.5 MK (see Table 5 of Scelsi et al. 2005 also), the density was estimated to be $2.5\times10^{10}$ cm$^{-3}$. For the fully ionized plasma the pressure p ($=2n_ekT_{max}$) was derived to be 628 dyne cm$^{-2}$. Using the relation, $p=B^2/8\pi$, the minimum magnetic field required to confine the plasma was derived to be 126 G.
Parameters Value
------------------------------- -------------------------------
Decay Time ($\tau_d$) $4234\pm98$s
$T_{max}$ $9.1\pm2.7 \times 10^7$ K
Loop Length(L) $9.9\pm3.1 \times 10^{10}$ cm
Electron Density ($n_e$) $2.5\times10^{10}$ cm$^{-3}$
Pressure ($p = 2n_ekT_{max}$) 628 dyne cm$^{-2}$
Magnetic Field (B)$^1$ 126 Gauss
: Loop parameters of EK Dra[]{data-label="tab:parameter"}
\
$^1$ minimum magnetic field to confine the plasma
{width="160mm"}
Power spectra and associated results
------------------------------------
We use the wavelet analysis code (Torrence & Compo 1998) to examine the statistically significant periodicities in the EK Dra time-series. Using wavelet analysis, the search for periodicities in XMM-Newton X-ray light curves is carried out by a time-localized function, which is continuous in both frequency and time. We use the Morlet wavelet function in the present time-series analysis which is defined as follows $$\psi_{t} (s) = \pi^{-1/4} e^{i \omega t} \ e^{ \left (\frac {-t^{2}}{2s^{2}} \right)}$$
where $t$, $s$, $\omega$ and $\pi^{-1/4}$, are time, wavelet scale, oscillation frequency, and the normalization constant, respectively. The Morlet wavelet is a sinusoidal function modified by a Gaussian envelop. In the Morlet wavelet, the Fourier period ($P$) is related to the wavelet scale ($s$) by the relation P = 1.03 s. The wavelet is convolved with the chosen observed time series to further determine the contribution of frequency within this time series, which matches the sinusoidal portion by varying the scale of the wavelet function. This method provides the power spectrum of the oscillations in different X-ray light curves of the flaring epoch in EK Dra as observed by XMM-Newton. The Morlet wavelet consists of an $"$edge effect$"$ , which is present in the analyses of time series data. However, this effect is significant only in regions lying within the cone-of-influence (COI) which demarks where possible estimated periods, very close to either the measurement interval or the maximum length of the time series, cannot be convincingly detected. Fig. 2 shows the wavelet spectrum of the 70 sec bin detrended X-ray light curve of the flaring epoch on EK Dra (see Fig. 1) in its decay phase with PN and MOS detectors. The light curves are detrended by fitting an exponential function to remove the long term trend of flare variations (Fig. 1, bottom panel). The intensity wavelets show the evolution of two periodicities at $\approx$4500 s and $\approx$2200 s respectively. It should be also noted that these two periodicities are detected in the X-ray light curves observed by two detectors (PN & MOS) simultaneously onboard XMM-Newton, therefore, they must be considered as reliable periods present in the decay phase of the flare. Moreover, a unique property of these two periodicities is that they are not evolving simultaneously. Instead, the long period of 4500 s evolves firstly up to a duration of $\approx$14 ks, and thereafter, the lower period of 2200 s evolves. The power associated with both periods is globally distributed as they repeat $>$3 cycles in the total time epoch of the observations (Fig. 2). Moreover, the evolved periodicities are consistent in the X-ray light curves observed at different bins in both the detectors PN & MOS.
Theoretical Interpretation
==========================
The quasi-periodicities of $\approx$4500 s and $\approx$2200 s evolve in the decay phase of the stellar flare, and the second periodicity of 2200 s is almost half of the first period 4500 s. The most likely explanation for these periodicities is that they are associated with the evolution of MHD waves. An investigation of the wavelets diagrams (Fig. 2) shows that both periods evolve gradually. The 4500 s period evolves firstly for 14 ks during the decay phase of the flare, while 2200 s period evolves thereafter for 10 ks. Both periods are distributed for $>$ 3 cycles in the intensity wavelet exhibiting their global nature and associated power distribution. Therefore, these periods are not associated with any transient reconnection process which may lead to the localized burst and enhancement of the intensity in real time series as well as power in the localized Fourier domain (e.g., Doyle et al., 2018).
We suggest that the 4500 s period may be associated with the evolution of magnetoacoustic waves in the flaring loop system of EK Dra. Interestingly, the longer period switched towards a comparatively shorter period in due course of time in the observational base-line, and they do not evolve simultaneously. This is the first most likely evidence that the magnetized atmosphere of a solar analogue acts like a frequency filter and it is well structured. We have analyzed how the flaring atmosphere in the atmosphere of EK Dra is filtering the wave frequencies.
Fig. 3 shows the variation of the temperature of the atmosphere in the decay phase of the flare. Assuming hydrostatic equilibrium, the scale height (H) of the atmosphere depends on temperature (T) as H=k$_{B}$T/$\mu$g, and changes as the temperature changes. Here, k$_{B}$ is the Boltzman constant, $\mu$ is the mean molecular weight in the terms of proton’s rest mass and g the surface gravity of a star. For EK Dra, we took g = 329 ms$^{-2}$ (Scelsi et al. 2005) and $\mu$ = 0.62 (for solar abundances). It should be noted that in any particular instance, the flaring loop system is considered to be an isothermal loop system, although it is a cooling as far as the elapse of time epoch is concerned. It is shown in Fig. 3 that both the scale height and temperature drop to their minimum value up to $\sim$14 ks after the flare peak, and remain almost at their minimum values during the course of the observations in the decay phase of the flare. This is the time when frequency filtering is evident in the wavelet diagrams (Fig. 2). Here, the average temperature was converted to maximum temperature as per the relation $T = 0.13 T_{obs}^{1.16}$ (see Reale 2007). The acoustic cutoff frequency of the atmosphere depends upon the sound speed and scale height as $\Omega_{ac}$=c$_{s}$/2H. Therefore, the ratio of the acoustic cut-off frequency near the flare peak ($\Omega_{ac1}$), and after 14 ks ($\Omega_{ac2}$) in the decay phase of the flare will be
$$\frac{{\Omega_{ac1}}}{{\Omega_{ac2}}}=\sqrt{\frac{T1}{T2}}\times\frac{H2}{H1}\,,$$
which becomes 0.6. For a maximum temperature of 91 MK at the flare peak, we have estimated the acoustic cut-off period of the flaring atmosphere as 5268 s. As the time elapsed, the loop system is cooled down and after 14 ks the acoustic cut-off period will be reduced to the value of 2820 s.
This clearly demonstrates that the broad-band (multiple) frequency waves are excited in the flaring loop system during the onset of the flare. In the beginning, the most dominant period evolves at the longer period, and the 4500 s periodicity is detected in the intensity wavelet. The higher acoustic cut-off period of 5268 s easily allows the lower periods having dominant power to propagate in the flaring loop system that is evident in the wavelet power spectrum (Fig. 2). As the flaring atmosphere cools down, plasma structuring changes cause the reduced cut-off period up to a value of 2820 s. Therefore, it eliminates the dominant longer period waves having periods more than the acoustic cut-off period, however, it allows the lower periods (here 2200 s) to propagate in the atmosphere. Once the dominant longer periods are filtered by the atmosphere, the lower periods having the significant power becomes evident in the intensity wavelet (Fig. 2).
Fig. 4 shows the fitted oscillatory (yellow) curves on the actual emissions in the decay phase of the flare. The fitted curve is the combination of two dominant signals with periods of 4500 s and 2200 s respectively. This curve is based on the variation of exponentially decaying harmonic function with the detrended light curve. The function is described as follows
$$I(t)=\sum_{n=1}^{2}{A_{n}Cos}\Bigg(\frac{2\pi}{P_{n}}.t+\phi_{n}\Bigg)e^{\frac{-\delta}{P_{n}}.t}\,,$$
where A$_{n}$ and $\phi_{n}$ are amplitudes and phases corresponding to the oscillatory periods, P$_{n}$ (4500s and 2200 s), respectively, and $\delta$ is the damping factor. The best fit of the function (cf., yellow curve in Fig. 4) gives $\delta=0.64\pm0.013$, which indicates a damping. This curve fits well the observed light curve by XMM-Newton detectors consisting of the signals corressponding to the detected periodicities (cf., Figs. 1-2). The synthesized curve shows the decay, which is similar to the observed decay when the initial amplitude of the oscillations reduces to its 1/e value in the observational base-line. It should be noted that the synthesized signal with the combination of two observed periodicities fits well the observed light curves up to 25 ks of the observations, and produces a consistent result. This indicates that multiple periodicities are excited in the flaring loop system episodically during the flare energy release, and these are related with the magnetoacoustic wave modes. However, the structured flaring atmosphere works as a natural filter for the excited waves. Initially the dominant long period waves (P$_{1}$=4500 s) are detected in the observational baseline, and thereafter, they get attenuated by the medium and only short period waves (P$_{1}$=2200 s) are allowed.
There may also be another possibility that these detected periods are associated with the multiple harmonics of the standing slow magneto-acoustic waves. The 4500 s periodicity may arise due to the fundamental mode of slow magnetoacoustic modes which can perturb the density and can also modulate the emissions of the flaring loops where they excite. While, the 2200 s periodicity may arise due to the first overtone evolved in the same flaring loop system. The loop length is estimated as 1$\times$10$^{11}$cm (cf., Table 1), therefore, the phase speed of the fundamental mode of the oscillations is V$_{ph}$=2L/P$\approx$440 km s$^{-1}$ that is indeed a slow mode while we compare its phase speed with the local sound speed at average flare temperature in its decay phase (cf., Fig. 3). However, these periods do not evolve simultaneously as a multiple wave harmonics. Instead, they evolve one by one (from longer to shorter period) in the decay phase of the flare when flaring loops are being cooled (Fig. 3). The most likely scenario is that due to the injection of flare energy, the pressure (and thus density) of the hot loops are perturbed causing the evolution of various slow magnetoacoustic waves. Since the loop’s environment is continuously changing in the decay phase of the flare, therefore, the associated acoustic cut-off frequency also changes, which allows certain frequencies to propagate in most predominant ways.
Discussion and Conclusions
==========================
In the solar coronal loops, due to the injection of flare energy, the evolution of slow magnetoacoustic waves are one of the most common physical scenario (e.g., Kumar et al., 2016; Nakariakov & Zimovets, 2011; Fang et al., 2015, and references cited therein). Apart from propagating waves, the standing slow waves are also well observed in cool and hot flaring loops in the solar atmosphere (e.g., Ofman & Wang, 2002; Wang & Solanki, 2004; Srivastava et al., 2010; Kumar et al., 2015, and references cited therein). These waves are also observed in the stellar coronae (Srivastava et al., 2013).
In the present paper, we firstly observe the existence of slow magnetoacoustic waves over a range of periods in the flaring region. The highly dynamic atmosphere of EK Dra enables selective propagation of these waves, while it attenuates other frequencies. Similar behaviour is well observed in the Sun’s atmosphere where complex structuring of its atmosphere causes filtering of the magnetoacoustic waves (Yuan et al., 2014; Murawski et al., 2016). The frequency filtering imposed above the localized flaring corona of EK Dra may indicate the behaviour of the high pass filter, and in fact such filter may not amplify incident powers. Therefore, both the periods were simultaneously present in first 14 ks of the flaring epoch in its decay phase which is seen in MOS signals (Fig. 1, bottom-panel), however, 4500 s period and related power were dominant in that phase. As the temperature minimum condition emerges in the flaring region due to the gradual cooling, the changing background atmosphere filters the longer period waves and allow the passage of comparatively smaller periods selectively. We provide the first evidence of such physical behaviour in the magnetized corona of the sun-like star EK Dra. Our present study provides a new insight that the EK Dra solar analog may have very similar atmosphere and wave activity as what we observe on the Sun. However, more investigations need to be carried out about this star using multiwavelength observations to understand magnetohydrodynamic (MHD) wave activity in its corona. This will also provide the clues that how its atmosphere is channeling the wave energy, and affecting the wave propagation properties.
Acknowledgment
==============
This work uses data obtained by XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and the USA (NASA). AKS and MM acknowledge the support of UKIERI Project to support their joint research. PJ acknowledges support from Grant 16-13277S of the Grant Agency of the Czech Republic. PC and YJM acknowledge the support by the BK21 plus program from the National Research Foundation of Korea to Kyung Hee University, Korea. JCP, AKS, and SK also acknowledge the DST-RFBR grant INT/RUS/RFBR/P-271.
[99]{} Aschwanden, M. J., & Schrijver, C. J. 2011, , 736, 102 Centeno, R., Collados, M., & Trujillo Bueno, J. 2006, , 640, 1153 Cho, I.-H., Cho, K.-S., Nakariakov, V. M., Kim, S., & Kumar, P. 2016, , 830, 110 De Pontieu, B., Erd[é]{}lyi, R., & James, S. P. 2004, , 430, 536 Del Zanna, G., Dere, K. P., Young, P. R., Landi, E., & Mason, H. E. 2015, , 582, A56 den Herder, J. W., Brinkman, A. C., Kahn, S. M., et al. 2001, , 365, L7 Doyle, J. G., Shetye, J., Antonova, A. E., et al. 2018, , 475, 2842 Fang, X., Yuan, D., Van Doorsselaere, T., Keppens, R., & Xia, C. 2015, , 813, 33 Fedun, V., Erd[é]{}lyi, R., & Shelyag, S. 2009, , 258, 219 Fedun, V., Verth, G., Jess, D. B., & Erd[é]{}lyi, R. 2011, , 740, L46 Goossens, M., Erd[é]{}lyi, R., & Ruderman, M. S. 2011, , 158, 289 Jansen, F., Lumb, D., Altieri, B., et al. 2001, , 365, L1 Jess, D. B., Mathioudakis, M., Erd[é]{}lyi, R., et al. 2009, Science, 323, 1582 Jel[í]{}nek, P., Karlick[ý]{}, M., & Murawski, K. 2015, , 812, 105 Jel[í]{}nek, P., Srivastava, A. K., Murawski, K., Kayshap, P., & Dwivedi, B. N. 2015, , 581, A131 Jel[í]{}nek, P., & Murawski, K. 2013, , 434, 2347 Karlick[ý]{}, M., & Jel[í]{}nek, P. 2016, , 590, A4 Khomenko, E., Centeno, R., Collados, M., & Trujillo Bueno, J. 2008, , 676, L85 K[ö]{}nig, B., Guenther, E. W., Woitas, J., & Hatzes, A. P. 2005, , 435, 215 Kumar, P., Nakariakov, V. M., & Cho, K.-S. 2015, , 804, 4 Kumar, P., Nakariakov, V. M., & Cho, K.-S. 2016, , 822, 7 Mathioudakis, M., Seiradakis, J. H., Williams, D. R., et al. 2003, , 403, 1101 Mathioudakis, M., Bloomfield, D. S., Jess, D. B., Dhillon, V. S., & Marsh, T. R. 2006, , 456, 323 McAteer, R. T. J., Gallagher, P. T., Williams, D. R., et al. 2003, , 587, 806 Mitra-Kraev, U., Harra, L. K., Williams, D. R., & Kraev, E. 2005, , 436, 1041 Murawski, K., Musielak, Z. E., Konkol, P., & Wi[ś]{}niewska, A. 2016, , 827, 37 Nakariakov, V. M., & Verwichte, E. 2005, Living Reviews in Solar Physics, 2, 3 Nakariakov, V. M., & Melnikov, V. F. 2009, , 149, 119 Nakariakov, V. M., & Zimovets, I. V. 2011, , 730, L27 Ofman, L., & Wang, T. 2002, , 580, L85 Pandey, J. C., & Singh, K. P. 2008, , 387, 1627 Pandey, J. C., & Srivastava, A. K. 2009, , 697, L153 Pascoe, D. J. 2014, Research in Astronomy and Astrophysics, 14, 805-830 Porquet, D., Mewe, R., Dubau, J., Raassen, A. J. J., & Kaastra, J. S. 2001, , 376, 1113 Pugh, C. E., Nakariakov, V. M., & Broomhall, A.-M. 2015, , 813, L5 Reale, F., Betta, R., Peres, G., Serio, S., & McTiernan, J. 1997, , 325, 782 Reale, F. 2007, , 471, 271 Scelsi, L., Maggio, A., Peres, G., & Pallavicini, R. 2005, , 432, 671 Schrijver, C. J., & Title, A. M. 2003, , 597, L165 Srivastava, A. K., Lalitha, S., & Pandey, J. C. 2013, , 778, L28 Srivastava, A. K., Zaqarashvili, T. V., Uddin, W., Dwivedi, B. N., & Kumar, P. 2008, , 388, 1899 Srivastava, A. K., & Dwivedi, B. N. 2010, , 15, 8 Srivastava, A. K., & Goossens, M. 2013, , 777, 17 Srivastava, A. K., Shetye, J., Murawski, K., et al. 2017, Scientific Reports, 7, 43147 Str[ü]{}der, L., Briel, U., Dennerl, K., et al. 2001, , 365, L1 Turner, M. J. L., Abbey, A., Arnaud, M., et al. 2001, , 365, L27 Yuan, D., Sych, R., Reznikova, V. E., & Nakariakov, V. M. 2014, , 561, A19 Vecchio, A., Cauzzi, G., Reardon, K. P., Janssen, K., & Rimmele, T. 2007, , 461, L1 Wang, T. J., & Solanki, S. K. 2004, , 421, L33 White, R. S., Verwichte, E., & Foullon, C. 2012, , 545, A129
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The effect of disorder on pinning and wetting models has attracted much attention in theoretical physics ([*e.g.*]{} [@cf:FLNO; @cf:DHV]). In particular, it has been predicted on the basis of the [*Harris criterion*]{} that disorder is [*relevant*]{} (annealed and quenched model have different critical points and critical exponents) if the return probability exponent ${\alpha}$, a positive number that characterizes the model, is larger than $1/2$. Weak disorder has been predicted to be [*irrelevant*]{} ([*i.e.*]{} coinciding critical points and exponents) if ${\alpha}<1/2$. Recent mathematical work (in particular [@cf:Ken; @cf:DGLT; @cf:GT_cmp; @cf:GT_prl]) has put these predictions on firm grounds. In [*renormalization group*]{} terms, the case ${\alpha}=1/2$ is a [*marginal*]{} case and there is no agreement in the literature as to whether one should expect disorder relevance [@cf:DHV] or irrelevance [@cf:FLNO] at marginality. The question is particularly intriguing also because the case ${\alpha}=1/2$ includes the classical models of two-dimensional wetting of a rough substrate, of pinning of directed polymers on a defect line in dimension $(3+1)$ or $(1+1)$ and of pinning of an heteropolymer by a point potential in three-dimensional space. Here we prove disorder relevance both for the general ${\alpha}=1/2$ pinning model and for the hierarchical version of the model proposed in [@cf:DHV], in the sense that we prove a shift of the quenched critical point with respect to the annealed one. In both cases we work with Gaussian disorder and we show that the shift is at least of order $\exp(-1/\beta^4)$ for $\beta$ small, if $\beta^2$ is the disorder variance.\
\
2000 *Mathematics Subject Classification: 82B44, 60K37, 60K05, 82B41*\
\
*Keywords: Pinning and Wetting Models, Hierarchical Models on Diamond Lattices, Quenched Disorder, Harris Criterion, Fractional Moment Estimates, Coarse Graining*
address:
- ' Universit[é]{} Paris Diderot (Paris 7) and Laboratoire de Probabilit[é]{}s et Modèles Aléatoires (CNRS U.M.R. 7599), U.F.R. Mathématiques, Case 7012 (Site Chevaleret), 75205 Paris cedex 13, France '
- ' Universit[é]{} Paris Diderot (Paris 7) and Laboratoire de Probabilit[é]{}s et Modèles Aléatoires (CNRS U.M.R. 7599), U.F.R. Mathématiques, Case 7012 (Site Chevaleret), 75205 Paris cedex 13, France '
- 'CNRS and ENS Lyon, Laboratoire de Physique, 46 Allée d’Italie, 69364 Lyon, France '
author:
- Giambattista Giacomin
- Hubert Lacoin
- Fabio Lucio Toninelli
title: |
Marginal relevance of disorder\
for pinning models
---
Introduction
============
Wetting and pinning on a defect line in $(1+1)$-dimensions {#sec:intro1}
----------------------------------------------------------
The intense activity aiming at understanding phenomena like wetting in two dimensions [@cf:Abraham] and pinning of polymers by a defect line [@cf:FLN] has led several people to focus on a class of simplified models based on random walks. In order to describe more realistic, spatially inhomogeneous situations, these models include disordered interactions. While a very substantial amount of work has been done, it is quite remarkable that some crucial issues are not only mathematically open (which is not surprising given the presence of disorder), but also controversial in the physics literature.
Let us start by introducing the most basic, and most studied, model in the class we consider (it is the case considered in [@cf:FLNO; @cf:DHV], but also in [@cf:BM; @cf:GN; @cf:GS1; @cf:GS2; @cf:SC; @cf:TangChate], up to some inessential details, although the notations used by the various authors are quite different). Let $S=\{S_0,S_1, \ldots\}$ be a simple symmetric random walk on ${{\ensuremath{\mathbb Z}} }$, [*i.e.*]{}, $S_0=0$ and $\{ S_n-S_{n-1}\}_{n \in
{{\ensuremath{\mathbb N}} }}$ is an IID sequence (with law ${{\ensuremath{\mathbf P}} }$) of random variables taking values $\pm 1$ with probability $1/2$. It is better to take a directed walk viewpoint, that is to consider the process $\{ (n, S_n)\}_{n=0,1, \ldots}$. This random walk is the [*free model*]{} and we want to understand the situation where the walk interacts with a substrate or with a defect line that provides [*disordered*]{} ([*e.g.*]{} random) rewards/penalties each time the walk hits it (see Fig. \[fig:SRWwetting\]). The walk may or may not be allowed to take negative values: we call [ *pinning on a defect line*]{} the first case and [*wetting of a substrate*]{} the second one. It is by now well understood that these two cases are equivalent and we briefly discuss the wetting case only in the caption of Figure \[fig:SRWwetting\]: the general model we will consider covers both wetting and pinning cases. The interaction is introduced via the Hamiltonian $$\label{eq:Haus}
H_{N,{\omega}}(S):=-\sum_{n=1}^{N}\left({\beta}{\omega}_n+h-\log {{\ensuremath{\mathbb E}} }(\exp({\beta}{\omega}_1))\right){\mathbf{1}}_{\{S_n=0\}},$$ where $N\in 2{\mathbb{N}}$ is the system size, $h$ (homogeneous pinning potential) is a real number, ${\omega}:=\{{\omega}_1,{\omega}_2,\ldots\}$ is a sequence of IID centered random variables with finite exponential moments (in this work, we will restrict to the Gaussian case), $\beta\ge0$ is the disorder strength and ${{\ensuremath{\mathbb E}} }$ denotes the average with respect to ${\omega}$. It will be soon clear what is the notational convenience in introducing the non-random term $\log
{{\ensuremath{\mathbb E}} }(\exp({\beta}{\omega}_1))$ (which could be absorbed into $h$ anyway).
=10.4 cm
The Gibbs measure ${{\ensuremath{\mathbf P}} }_{N,{\omega}}$ for the pinning model is then defined as $$\label{eq:Paus}
\frac{{\,\text{\rm d}}{{\ensuremath{\mathbf P}} }_{N,{\omega}}}{{\,\text{\rm d}}{{\ensuremath{\mathbf P}} }}(S)=\frac{e^{-H_{N,{\omega}}(S)}{\mathbf{1}}_{\{S_N=0\}}}{Z_{N,{\omega}}}$$ and of course $Z_{N,{\omega}}:={{\ensuremath{\mathbf E}} }[\exp(-H_{N,{\omega}}(S)){\mathbf{1}}_{\{S_N=0\}}]$, where ${{\ensuremath{\mathbf E}} }$ denotes expectation with respect to the simple random walk measure ${{\ensuremath{\mathbf P}} }$. Note that we imposed the boundary condition $S_N=0=S_0$ (just to be consistent with the rest of the paper). It is well known that the model undergoes a localization/delocalization transition as $h$ varies: if $h$ is larger than a certain threshold value $h_c(\beta)$ ([*quenched critical point*]{}) then, under the Gibbs measure, the system is [ *localized*]{}: the contact fraction, defined as $$\frac1N{{\ensuremath{\mathbf E}} }_{N,{\omega}}\left[\sum_{n=1}^N{\mathbf{1}}_{\{S_n=0\}}\right],$$ tends to a positive limit for $N\to\infty$. On the other hand, for $h<h_c({\beta})$ the system is [*delocalized*]{}, [*i.e.*]{}, the limit is zero.
The result we just stated is true also in absence of disorder (${\beta}=0$) and a remarkable fact for the homogeneous ([*i.e.*]{} non-disordered) model is that it is exactly solvable ([@cf:Fisher; @cf:Book] and references therein). In particular, we know that $h_c(0)=0$, [*i.e.*]{}, an arbitrarily small reward is necessary and sufficient for pinning, and that the free energy behaves quadratically close to criticality. If now we consider the [*annealed measure*]{} corresponding to , that is the model in which one replaces both $\exp(-H_{N,{\omega}}(S))$ and $Z_{N,{\omega}}$ by their averages with respect to ${\omega}$, one readily realizes that the annealed model is a homogeneous model, and precisely the one we obtain by setting ${\beta}=0$ in . Therefore one finds that the [*annealed critical point*]{} $h_c^a({\beta})$ equals $0$ for every $\beta$, and that the [*annealed free energy*]{} ${\textsc{f}}^a({\beta},h)$ behaves, for $h\searrow 0$, like ${\textsc{f}}^a({\beta},h)\sim const\times h^2$, while it is zero for $h\le 0$.
Very natural questions are: does $h_c({\beta})$ differ from $h_c^a({\beta})$? Are quenched and annealed critical exponents different? As we are going to explain, the first question finds contradictory answers in the literature, while no clear-cut statement can really be found about the second. Below we are going to argue that these two questions are intimately related, but first we make a short detour in order to define a more general class of models. It is in this more general context that the role of the disorder and the specificity of the simple random walk case can be best appreciated.
Reduction to renewal-based models
---------------------------------
As argued in the caption of Figure \[fig:SRWwetting\], the basic underlying process is the [*point process*]{} $\tau:=\{\tau_0, \tau_1,
\ldots\}$, which is a renewal process (that is $\{
\tau_{n}-\tau_{n-1}\}_{n \in {\mathbb{N}}}$ is an IID sequence of integer-valued random variables). We set $K(n):={{\ensuremath{\mathbf P}} }(\tau_1=n)$. It is well known that, for the simple random walk case, $\sum_{n\in{\mathbb{N}}} K(n)=1 $ (the walk is recurrent) and $K(n)\stackrel{n \to \infty} \sim 1/(\sqrt{4\pi}n^{3/2})$. This suggests the natural generalized framework of models based on discrete renewal processes such that $$\label{eq:Kintro}
\sum_{n\in {\mathbb{N}}} K(n) \le 1 \ \text { and } \ K(n)\stackrel{n \to \infty} \sim \frac{C_K}{n^{1+{\alpha}}},$$ with $C_K>0$ and ${\alpha}>0$. We are of course employing the standard notation $a_n \sim b_n$ for $\lim a_n/b_n =1$. The case $\sum_{n\in {\mathbb{N}}} K(n) < 1$ refers to transient (or [*terminating*]{}) renewals (of which the wetting case is an example), see also Remark \[rem:trans\] below. This framework includes for example the simple random walk in $d \ge 3$, for which $\sum_{n\in {\mathbb{N}}} K(n) < 1$ and ${\alpha}=(d/2)-1$, but it is of course much more general. We will come back with more details on this model, but let us just say now that the definition of the Gibbs measure is given in this case by -, with $S$ replaced by $\tau$ in the left-hand side and with the event $\{S_n=0\}$ replaced by the event $\{\text{there is } j \text{ such that } \tau_j=n\}$.
Harris criterion and disorder relevance: the state of the art
-------------------------------------------------------------
The questions mentioned at the end of Section \[sec:intro1\] are typical questions of disorder relevance, [*i.e.*]{}, of stability of critical properties with respect to (weak) disorder. In renormalization group language, one is asking whether or not disorder drives the system towards a new fixed point. A heuristic tool which was devised to give an answer to such questions is the [*Harris criterion*]{} [@cf:Harris], originally proposed for random ferromagnetic Ising models. The Harris criterion states that disorder is relevant if the specific heat exponent of the pure system is positive, and irrelevant if it is negative. In case such critical exponent is zero (this is called a [*marginal case*]{}), the Harris criterion gives no prediction and a case-by-case delicate analysis is needed. Now, it turns out that the random pinning model described above is a marginal case, and from this point of view it is not surprising that the question of disorder relevance is not solved yet, even on heuristic grounds: in particular, the authors of [@cf:FLNO] (and then also [@cf:GS1; @cf:GS2] and, very recently, [@cf:GN]) claimed that for small $\beta$ the quenched critical point coincides with the annealed one (with our conventions, this means that both are zero), while in [@cf:DHV] it was concluded that they differ for every $\beta>0$, and that their difference is of order $\exp(-const/\beta^2)$ for $\beta$ small (we mention [@cf:BM; @cf:SC; @cf:TangChate] which support this second possibility). Note that such a quantity is smaller than any power of $\beta$, and therefore vanishes at all orders in weak-disorder perturbation theory (this is also typical of marginal cases).
In an effort to reduce the problem to its core, beyond the difficulties connected to the random walk or renewal structure, a [ *hierarchical pinning model*]{}, defined on a diamond lattice, was introduced in [@cf:DHV]. In this case, the laws of the partition functions for the systems of size $N$ and $2N$ are linked by a simple recursion. The role of $\alpha$ is played here by a real parameter $B\in(1,2)$, which is related to the geometry of the hierarchical lattice. Also in this case, the Harris criterion predicts that disorder is relevant in a certain regime (here, $B<B_c:=\sqrt 2$) and irrelevant in another ($B>B_c$), while $B=B_c$ is the marginal case where the specific heat critical exponent of the pure model vanishes. Again, the authors of [@cf:DHV] predicted that disorder is marginally relevant for $B=B_c$, and that the difference between annealed and quenched critical point behaves like $\exp(-const/\beta^2)$ for $\beta$ small (they gave also numerical evidence that the critical exponent is modified by disorder).
Let us mention that hierarchical models based on diamond lattices have played an important role in elucidating the effect of disorder on various statistical mechanics models: we mention for instance [@cf:DG].
The mathematical comprehension of the question of disorder relevance in pinning models has witnessed remarkable progress lately. First of all, it was proven in [@cf:GT_cmp] that an arbitrarily weak (but extensive) disorder changes the critical exponent if ${\alpha}>1/2$ (the analogous result for the hierarchical model was proven in [@cf:LT]). Results concerning the critical points came later: in [@cf:Ken; @cf:T_cmp] it was proven that if ${\alpha}<1/2$ then $h_c({\beta})=0$ (and the quenched critical exponent coincides with the annealed one) for ${\beta}$ sufficiently small (the analogous result for the hierarchical model was given in [@cf:GLT]). Finally, the fact that $h_c({\beta})>0$ for every ${\beta}>0$ (together with the correct small-${\beta}$ behavior) in the regime where the Harris criterion predicts disorder relevance was proven in [@cf:GLT] in the hierarchical set-up, and then in [@cf:DGLT; @cf:AZ] in the non-hierarchical one. One can therefore safely say that the comprehension of the relevance question is by now rather solid, [ *except in the marginal case*]{} (of course some problems remain open, for instance the determination of the value of the quenched critical exponent in the relevant disorder regime, beyond the bounds proved in [@cf:GT_cmp]).
Marginal relevance of disorder
------------------------------
In this work, we solve the question of disorder relevance for the marginal case $\alpha=1/2$ (or $B=B_c$ in the hierarchical situation), showing that [*quenched and annealed critical points differ for every disorder strength ${\beta}>0$*]{}. We also give a quantitative bound, $h_c({\beta})\ge \exp(-const/\beta^4)$ for ${\beta}$ small, which is however presumably not optimal. The method we use is a non-trivial extension of the [*fractional moment – change of measure method*]{} which already allowed to prove disorder relevance for $B<B_c$ in [@cf:GLT] or for ${\alpha}>1/2$ in [@cf:DGLT]. A few words about the evolution of this method may be useful to the reader. The idea of estimating non-integer moments of the partition function of disordered systems is not new: consider for instance [@cf:BPP] in the context of directed polymers in random environment, or [@cf:AizM] in the context of Anderson localization (in the latter case one deals with non-integer moments of the propagator). However, the power of non-integer moments in pinning/wetting models was not appreciated until [@cf:T_AAP], where it was employed to prove, among other facts, that quenched and annealed critical points differ for large ${\beta}$, irrespective of the value of ${\alpha}\in(0,\infty)$. The new idea which was needed to treat the case of weak disorder (small ${\beta}$) was instead introduced in [@cf:GLT; @cf:DGLT], and it is a change-of-measure idea, coupled with an [*iteration procedure*]{}: one changes the law of the disorder ${\omega}$ in such a way that the new and the old laws are very close in a certain sense, but under the new one it is easier to prove that the fractional moments of the partition function are small. In the relevant disorder regime, ${\alpha}>1/2$ or $B<B_c$, it turns out that it is possible to choose the new law so that the ${\omega}_n$’s are still IID random variables, whose law is simply tilted with respect to the original one. This tilting procedure is bound to fail if applied for arbitrarily large volumes, but having such bounds for sufficiently large, but finite, system sizes is actually sufficient because of an iteration argument (which appears very cleanly in the hierarchical set-up).
In order to deal with the marginal case we will instead introduce a long-range anti-correlation structure for the ${\omega}$-variables. Such correlations are carefully chosen in order to reflect the structure of the two-point function of the annealed model and, in the non-hierarchical case, they are restricted, via a coarse-graining procedure inspired by [@cf:T_cg], only to suitable [ *disorder pockets*]{}.
We mention also that one of us [@cf:Hubert] proved recently that disorder is marginally relevant in a different version of the hierarchical pinning model. What simplifies the task in that case is that the Green function of the model is spatially inhomogeneous and one can take advantage of that by tilting the ${\omega}$-distributions in a inhomogeneous way (keeping the ${\omega}$’s independent). The Green function of the hierarchical model proposed in [@cf:DHV] is instead constant throughout the system and inhomogeneous tilting does not seem to be of help (as it does not seem to be of help in the non-hierarchical case, since it does not match with the coarse graining procedure).
The paper is organized as follows: the hierarchical (resp. non-hierarchical) pinning model is precisely defined in Section \[sec:Hmodel\] (resp. in Section \[sec:nHmodel\]), where we also state our result concerning marginal relevance of disorder. Such result is proven in Section \[sec:Hproof\] in the hierarchical case, and in Section \[sec:nHproof\] in the non-hierarchical one.
In order not to hide the novelty of the idea with technicalities, we restrict ourselves to Gaussian disorder and, in the case of the non-hierarchical model, we do not treat the natural generalization where $K(\cdot)$ is of the form $K(n)={L(n)}/{n^{3/2}}$ with $L(\cdot)$ a slowly varying function [@cf:Feller2 VIII.8]. We plan to come back to both issues in a forthcoming paper [@cf:kbodies].
The hierarchical model {#sec:Hmodel}
======================
Let $1<B<2$. We study the following iteration which transforms a vector $\{R_n^{(i)}\}_{i\in{\mathbb{N}}}\in ({\mathbb{R}}^+)^{\mathbb{N}}$ into a new vector $\{R_{n+1}^{(i)}\}_{i\in{\mathbb{N}}}\in({\mathbb{R}}^+)^{\mathbb{N}}$: $$\label{eq:Rn+1}
R^{(i)}_{n+1}\, =\, \frac{R_n^{(2i-1)} R_n^{(2i)}+(B-1)}B,$$ for $n\in{\mathbb{N}}\cup\{0\}$ and $i\in{\mathbb{N}}$.
In particular, we are interested in the case in which the initial condition is random and given by $R_0^{(i)}=e^{\beta{\omega}_i-\beta^2/2+h}$, with ${\omega}:=\{{\omega}_i\}_{i\in{\mathbb{N}}}$ a sequence of IID standard Gaussian random variables and $h\in{\mathbb{R}}, {\beta}\ge0$. We denote by ${{\ensuremath{\mathbb P}} }$ the law of ${\omega}$ and by ${{\ensuremath{\mathbb E}} }$ the corresponding average. In this case, it is immediate to realize that for every given $n$ the random variables $\{R_n^{(i)}\}_{i\in{\mathbb{N}}}$ are IID. We will study the behavior for large $n$ of $X_n:=R_n^{(1)}$.
It is easy to see that the average of $X_n$ satisfies the iteration $$\label{eq:homomap}
{{\ensuremath{\mathbb E}} }({X_{n+1}})\, =\, \frac{({{\ensuremath{\mathbb E}} }{X_n})^2+(B-1)}B,$$ with initial condition ${{\ensuremath{\mathbb E}} }({X_0})=e^h$. The map has two fixed points: a stable one, ${{\ensuremath{\mathbb E}} }X_n=(B-1)$, and an unstable one, ${{\ensuremath{\mathbb E}} }X_n=1$. This means that if $0\le{{\ensuremath{\mathbb E}} }X_0<1$ then ${{\ensuremath{\mathbb E}} }X_n$ tends to $(B-1)$ when $n\to\infty$, while if ${{\ensuremath{\mathbb E}} }X_0>1$ then ${{\ensuremath{\mathbb E}} }X_n$ tends to $+\infty$.
In [@cf:DHV] and [@cf:GLT], the model with $B>2$ was considered. However, the cases $B\in(1,2)$ and $B\in(2,\infty)$ are equivalent. Indeed, if $R_n^{(i)}$ satisfies with $B>2$, it is immediate to see that ${\widehat}R_n^{(i)}:= R_n^{(i)}/(B-1)$ satisfies the same iteration but with $B$ replaced by ${\widehat}B:=B/(B-1)\in(1,2)$. In this work, we prefer to work with $B\in(1,2)$ because things turn out to be notationally simpler ([*e.g.*]{}, the annealed critical point (defined in the next section) turns out to be $0$ rather than $\log (B-1)$). In the following, whenever we refer to results from [@cf:GLT] we give them for $B\in(1,2)$.
Quenched and annealed free energy and critical point
----------------------------------------------------
The random variable $X_n$ is interpreted as the partition function of the hierarchical random pinning model on a diamond lattice of generation $n$ (we refer to [@cf:DHV] for a clear discussion of this connection). The [*quenched free energy*]{} is then defined as $$\label{eq:F}
{\textsc{f}}(\beta,h):=\lim_{n\to\infty}\frac1{2^n}{{\ensuremath{\mathbb E}} }\log X_n.$$ In [@cf:GLT Th. 1.1] it was proven, among other facts, that for every $\beta\ge 0,h\in{\mathbb{R}}$ the limit exists and it is non-negative. Moreover, ${\textsc{f}}(\beta,\cdot)$ is convex and non-decreasing. On the other hand, the [*annealed free energy*]{} is by definition $$\label{eq:Fa}
{\textsc{f}}^a(\beta,h):=\lim_{n\to\infty}\frac1{2^n}\log {{\ensuremath{\mathbb E}} }X_n.$$ Since the initial condition of was normalized so that ${{\ensuremath{\mathbb E}} }X_0=e^h$, it is easy to see that the annealed free energy is nothing but the free energy of the non-disordered model: $$\label{eq:qz}
{\textsc{f}}^a(\beta,h)={\textsc{f}}(0,h).$$ Non-negativity of the free energy allows to define the [*quenched critical*]{} point in a natural way, as $$\label{eq:hc}
h_c(\beta):=\inf\{h\in{\mathbb{R}}: {\textsc{f}}(\beta,h)>0\},$$ and analogously one defines the [*annealed critical point*]{} $h_c^a(\beta)$. In view of observation , one sees that $h_c^a(\beta)=h_c(0)$. Monotonicity and convexity of ${\textsc{f}}(\beta,\cdot)$ imply that ${\textsc{f}}(\beta,h)=0$ for $h\le h_c(\beta)$.
For the annealed system, the critical point and the critical behavior of the free energy around it are known (see [@cf:DHV] or [@cf:GLT Th. 1.2]). What one finds is that for every $B\in(1,2)$ one has $h_c(0)=0$, and there exists $c:=c(B)>0$ such that for all $0\le h\le 1$ $$\label{eq:alpha-1}
c(B)^{-1}h^{1/\alpha}\le {\textsc{f}}(0,h)\le c(B) h^{1/\alpha},$$ where $$\label{eq:alpha}
\alpha:=\frac{\log(2/B)}{\log 2}\in (0,1).$$ Observe that $\alpha$ is decreasing as a function of $B$, and equals $1/2$ for $B=B_c:=\sqrt2$.
Disorder relevance or irrelevance
---------------------------------
The main question we are interested in is whether quenched and annealed critical points differ, and if yes how does their difference behave for small disorder. Jensen’s inequality, ${{\ensuremath{\mathbb E}} }\log X_n\le \log {{\ensuremath{\mathbb E}} }X_n$, implies in particular that ${\textsc{f}}(\beta,h)\le {\textsc{f}}(0,h)$ so that $h_c(\beta)\ge h_c(0)=0$. Is this inequality strict?
In [@cf:GLT] a quite complete picture was given, except in the marginal case $B=B_c$ which was left open:
[@cf:GLT Th. 1.4] If $1<B<B_c$, $h_c({\beta})>0$ for every ${\beta}>0$ and there exists $c_1>0$ such that for $0\le \beta\le 1$ $$\label{eq:relh}
c_1 \beta^{2\alpha/(2\alpha-1)}\le h_c(\beta)\le c_1^{-1} \beta^{2\alpha/(2\alpha-1)}.$$
If $B=B_c$ there exists $c_2>0$ such that for $0\le \beta\le 1$ $$\label{eq:margh}
h_c(\beta)\le \exp(-c_2/\beta^2).$$
If $B_c<B<2$ there exists $\beta_0>0$ such that $h_c(\beta)=0$ for every $0<\beta\le \beta_0$.
The main result of the present work is that in the marginal case, the two critical points [*do differ*]{} for every disorder strength:
\[th:main\] Let $B=B_c$. For every $0<\beta_0<\infty$ there exists a constant $0<c_3:=c_3(\beta_0)<\infty$ such that for every $0<\beta\le\beta_0$ $$\label{eq:main}
h_c(\beta)\ge \exp(-c_3/\beta^4).$$
The non-hierarchical model {#sec:nHmodel}
==========================
We let $\tau:=\{\tau_0,\tau_1,\ldots\}$ be a renewal process of law ${{\ensuremath{\mathbf P}} }$, with inter-arrival law $K(\cdot)$, [*i.e.*]{}, $\tau_0=0$ and $\{\tau_i-\tau_{i-1}\}_{i\in{\mathbb{N}}}$ is a sequence of IID integer-valued random variables such that $$\label{eq:K}
{{\ensuremath{\mathbf P}} }(\tau_1=n)= : K(n)\stackrel{n\to\infty}\sim \frac {C_K}{n^{1+\alpha}},$$ with $C_K>0$ and $\alpha>0$. We require that $K(\cdot)$ is a probability on ${\mathbb{N}}$, which amounts to assuming that the renewal process is recurrent. We require also that $K(n)>0$ for every $n\in{\mathbb{N}}$, but this is inessential and it is just meant to avoid making a certain number of remarks and small detours in the proofs to take care of this point.
As in Section \[sec:Hmodel\], ${\omega}:=\{{\omega}_1,{\omega}_2,\ldots\}$ denotes a sequence of IID standard Gaussian random variables. For a given system size $N\in{\mathbb{N}}$, coupling parameters $h\in {\mathbb{R}}$, $\beta\ge 0$ and a given disorder realization ${\omega}$ the partition function of the model is defined by $$\label{eq:Znh}
Z_{N,{\omega}}:={{\ensuremath{\mathbf E}} }\left[e^{\sum_{n=1}^N(\beta{\omega}_n+h-\beta^2/2)\delta_n}\delta_N\right],$$ where $\delta_n:={\mathbf{1}}_{\{n\in\tau\}}$, while the quenched free energy is $$\label{eq:F_nh}
{\textsc{f}}(\beta,h):=\lim_{N\to\infty}\frac1N{{\ensuremath{\mathbb E}} }\log Z_{N,{\omega}},$$ (we use the same notation as for the hierarchical model, since there is no risk of confusion). Like for the hierarchical model, the limit exists and is non-negative [@cf:Book Ch. 4], and one defines the critical point $h_c(\beta)$ for a given $\beta\ge0$ exactly as in . Again, one notices that the annealed free energy, [*i.e.*]{}, the limit of $(1/N)\log{{\ensuremath{\mathbb E}} }Z_{N,{\omega}}$, is nothing but ${\textsc{f}}(0,h)$, so that the annealed critical point is just $h_c(0)$.
With respect to most of the literature, our definition of the model is different (but of course completely equivalent) in that usually the partition function is defined as in with $h-\beta^2/2$ replaced simply by $h$.
The annealed (or pure) model can be exactly solved and in particular it is well known [@cf:Book Th. 2.1] that, if $\alpha\ne 1$, there exists a positive constant $c_K$ (which depends on $K(\cdot)$) such that $$\label{eq:annF}
{\textsc{f}}(0,h)\stackrel{h\searrow0}\sim c_K h^{\max(1,1/\alpha)}.$$ In the case $\alpha=1$, has to be modified in that the right-hand side becomes $\phi(1/h)h$ for some slowly-varying function $\phi(\cdot)$ which vanishes at infinity [@cf:Book Th. 2.1]. In particular, note that $h_c(0)=0$ so that $h_c(\beta)\ge0$ by Jensen’s inequality, exactly like for the hierarchical model.
\[rem:trans\] The assumption of recurrence for $\tau$, [*i.e.*]{}, $\sum_{n\in{\mathbb{N}}}K(n)=1$, is by no means a restriction. In fact, as it has been observed several times in the literature, if $\Sigma_K:=\sum_{n\in{\mathbb{N}}}K(n)<1$ one can define ${\widetilde}K(n):=K(n)/\Sigma_K$, and of course the renewal $\tau$ with law ${\widetilde}{{\ensuremath{\mathbf P}} }(\tau_1=n)={\widetilde}K(n)$ is recurrent. Then, it is immediate to realize from definition that $${\textsc{f}}(\beta,h)={\widetilde}{\textsc{f}}(\beta,h+\log \Sigma_K),$$ ${\widetilde}{\textsc{f}}$ being the free energy of the model defined as in - but with ${{\ensuremath{\mathbf P}} }$ replaced by ${\widetilde}{{\ensuremath{\mathbf P}} }$. In particular, $h^a_c(\beta)=-\log \Sigma_K$. This observation allows to apply Theorem \[th:main2\] below, for instance, to the case where $\tau$ is the set of returns to the origin of a symmetric, finite-variance random walk on ${\mathbb{Z}}^3$ (pinning of a directed polymer in dimension $(3+1)$): indeed, in this case holds with $\alpha=1/2$. For more details on this issue we refer to [@cf:Book Ch. 1].
Relevance or irrelevance of disorder
------------------------------------
Like for the hierarchical model, the question whether $h_c(\beta)$ coincides or not with $h_c(0)$ for $\beta$ small has been recently solved, [*except in the marginal case*]{} ${\alpha}=1/2$:
If $0<\alpha<1/2$, there exists $\beta_0>0$ such that $h_c(\beta)=0$ for every $0\le \beta\le \beta_0$. If ${\alpha}=1/2$, there exists a constant $c_4>0$ such that for $\beta\le 1$ $$h_c(\beta)\le \exp(-c_4/\beta^2).$$ If ${\alpha}>1/2$, $h_c({\beta})>0$ for every ${\beta}>0$ and, if in addition ${\alpha}\ne1$, there exists a constant $c_5>0$ such that if $\beta\le 1$ $$c_5\beta^{\max(2{\alpha}/(2{\alpha}-1),2)}\le h_c(\beta)\le c_5^{-1}\beta^{\max(2{\alpha}/(2{\alpha}-1),2)}.$$ If ${\alpha}=1$ there exist a constant $c_6>0$ and a slowly varying function $\psi(\cdot)$ vanishing at infinity such that for ${\beta}\le 1$ $$c_6\beta^{2}\psi(1/\beta)\le h_c(\beta)\le c_6^{-1}\beta^{2}\psi(1/\beta).$$
The results for ${\alpha}\le 1/2$, together with the critical point upper bounds for ${\alpha}>1/2$, have been proven in [@cf:Ken], and then in [@cf:T_cmp]; the lower bounds on the critical point for ${\alpha}>1/2$ have been proven in [@cf:DGLT] (the result in [@cf:DGLT] is slightly weaker than what we state here and the case ${\alpha}=1$ was not treated) and then in [@cf:AZ] (with the full result cited here).
The case $\alpha=0$ has also been considered, but in that case has to be replaced by $K(n)=L(n)/n$, with $L(\cdot)$ a function varying slowly at infinity and such that $\sum_{n\in {\mathbb{N}}}K(n)=1$. For instance, this corresponds to the case where $\tau$ is the set of returns to the origin of a symmetric random walk on ${\mathbb{Z}}^2$. In this case, it has been shown in [@cf:AZ2] that quenched and annealed critical points coincide for every value of $\beta\ge0$.
Let us recall also that it is proven in [@cf:GT_cmp] that, for every ${\alpha}>0$, we have $$\label{eq:smooth}
{\textsc{f}}(\beta,h)\, \le \, \frac{1+{\alpha}}{2{\beta}^2}\,(h-h_c({\beta}))^2,$$ for all ${\beta}>0,h>h_c({\beta})$: this means that when ${\alpha}>1/2$ disorder is relevant also in the sense that it changes the free-energy critical exponent ([*cf.*]{} ). The analogous result for the hierarchical model, with $(1+{\alpha})$ replaced by some constant $c(B)$ in , is proven in [@cf:LT].
In the present work we prove the following:
\[th:main2\] Assume that holds with $\alpha=1/2$. For every ${\beta}_0>0$ there exists a constant $0<c_7:=c_7({\beta}_0)<\infty$ such that for $\beta\le \beta_0$ $$h_c(\beta)\, \ge\, e^{-c_7/\beta^4}.$$
Marginal relevance of disorder: the hierarchical case {#sec:Hproof}
=====================================================
Preliminaries: a Galton-Watson representation for $X_n$ {#sec:trees}
-------------------------------------------------------
One can give an expression for $X_n$ which is analogous to that of the partition function of the non-hierarchical model, and which is more practical for our purposes. This involves a Galton-Watson tree [@cf:T-Harris] describing the successive offsprings of one individual. The offspring distribution concentrates on $0$ (with probability $(B-1)/B$) and on $2$ (with probability $1/B$). So, at a given generation, each individual that is present has either no descendant or two descendants, and this independently of any other individual of the generation. This branching procedure directly maps to a random tree (see Figure \[fig:treemarg\]): the law of such a branching process up to generation $n$ (the first individual is at generation $0$) or, analogously, the law of the random tree from the root (level $n$) up to the leaves (level $0$), is denoted by ${{\ensuremath{\mathbf P}} }_n$. The individuals that are present at the $n^{\textrm{th}}$ generation are a random subset ${\mathcal R}_n $ of $\{ 1, \ldots, 2^n\}$. We set ${\delta}_j:= {\mathbf{1}}_{j \in {\mathcal R}_n}$. Note that the mean offspring size is $2/B>1$, so that the Galton-Watson process is supercritical.
The following procedure on the standard binary graph ${{\ensuremath{\mathcal T}} }^{(n)}$ of depth $n+1$ (again, the root is at level $n$ and the leaves, numbered from $1$ to $2^n$, at level $0$) is going to be of help too. Given $\mathcal I\subset \{1,\ldots,2^n\}$, let $\mathcal T^{(n)}_{\mathcal
I}$ be the subtree obtained from $\mathcal T^{(n)}$ by deleting all edges except those which lead from leaves $j\in\mathcal I$ to the root. Note that, with the offspring distribution we consider, in general $\mathcal T^{(n)}_{\mathcal I}$ is not a realization of the $n$-generation Galton-Watson tree (some individuals may have just one descendant in $\mathcal T^{(n)}_{\mathcal I}$, see Figure \[fig:treemarg\]). Let $v(n,\mathcal I)$ be the number of nodes in $\mathcal T^{(n)}_{\mathcal I}$, with the convention that leaves are not counted as nodes, while the root is.
=10.7 cm \[r\][*level $0$*]{} \[r\][*(the leaves)*]{} \[r\][*(the root)*]{} \[r\][*level $1$*]{} \[r\][*level $2$*]{} \[r\][*level $4$*]{} \[r\][$\ldots\ \ $ ]{}
\[th:deltas\] For every $n\ge0$ we have $$\label{eq:Rn}
X_n\, = \,{{\ensuremath{\mathbf E}} }_n\left[
e^{\sum_{i=1}^{2^n}(\beta {\omega}_i+h-\beta^2/2)\delta_i}\right].$$ For every $n\ge 0$ and $\mathcal I\subset \{1,\ldots,2^n\}$, one has $$\label{eq:ffv1}
{{\ensuremath{\mathbf E}} }_n\bigg[\prod_{i\in\mathcal I}\delta_i\bigg]\, =\, B^{-v(n,\mathcal I)}.$$ In particular, ${{\ensuremath{\mathbf E}} }_n[\delta_i]=B^{-n}$ for every $i=1,\ldots,2^n$, [*i.e.*]{}, the Green function is constant throughout the system.
The right-hand side in for $n=0$ is equal to $\exp(h-{\beta}^2/2+ {\beta}{\omega}_1)$. Moreover, at the $(n+1)^{\textrm{th}}$ generation the branching process either contains only the initial individual (with probability $(B-1)/B$) or the initial individual has two children, which we may look at as initial individuals of two independent Galton-Watson trees containing $n$ new generations. We therefore have that the basic recursion is satisfied.
The second fact, , is a direct consequence of the definitions (see also the caption of Figure \[fig:treemarg\]).
The representation we have introduced in this section shows in particular that ${{\ensuremath{\mathbb E}} }X_n$ is just the generating function of $\vert \mathcal R_n\vert$ and the free energy ${\textsc{f}}(0, h)$ is therefore a natural quantity for the Galton-Watson process: and in fact $1/{\alpha}$ (${\alpha}$ given in ) appears in the original works on branching processes by T. E. Harris (of course not to be confused with A. B. Harris, who proposed the disorder relevance criterion on which we are focusing in this work).
The proof of Theorem \[th:main\]
--------------------------------
While the discussion of the previous section is valid for every $B\in(1,2)$, now we have to assume $B=B_c=\sqrt 2$. However some of the steps are still valid in general and we are going to replace $B$ with $B_c$ only when it is really needed. The proof is split into three subsections: the first introduces the fractional moment method and reduces the statement we want to prove, which is a statement on the limit $n \to \infty$ behavior of $X_n$, to finite-$n$ estimates. The estimates are provided in the second and third subsection.
### The fractional moment method
Let $U_n^{(i)}$ denote the quantity $[R_n^{(i)}-(B-1)]_+$ where $[x]_+=\max(x,0)$. Using the inequality $$[rs+r+s]_+\le [r]_+[s]_++[r]_++[s]_+,$$ which holds whenever $r,s\ge-1$, it is easy to check that implies
$$\label{eq:recU}
U^{(i)}_{n+1}\le \frac{U_n^{(2i-1)}U_n^{(2i)}+(U_n^{(2i-1)}+U_n^{(2i)})(B-1)}{B}.$$
Given $0<\gamma<1$, we define $A_n:={{\ensuremath{\mathbb E}} }([X_n-(B-1)]_+^\gamma)$. From above and by using the fractional inequality $$\label{eq:fracineq}
\left(\sum a_i\right)^{\gamma}\, \le\, \sum a_i^{\gamma},$$ which holds whenever $a_i\ge0$, we derive $$A_{n+1}\le \frac{A_n^2+2(B-1)^{\gamma}A_n}{B^{\gamma}}.$$ One readily sees now that, if there exists some integer $k$ such that $$\label{eq:k}
A_k\, < \, B^{\gamma}-2(B-1)^{\gamma},$$ then $A_n$ tends to zero as $n$ tends to infinity (this statement is easily obtain by studying the fixed points of the function $x\mapsto
{(x^2+2(B-1)^{\gamma}x)}/{B^{\gamma}}$). On the other hand, $$\label{eq:LB0411}
{{\ensuremath{\mathbb E}} }\left[ X_n ^\gamma \right] \, \le \, {{\ensuremath{\mathbb E}} }\left([X_n-(B-1)]_++(B-1)\right)^\gamma\le
(B-1)^\gamma + A_n,$$ and therefore implies that ${\textsc{f}}({\beta}, h)=0$ since, by Jensen inequality, we have $$\label{eq:JensAn}
\frac 1{2^n} {{\ensuremath{\mathbb E}} }\log X_n \, \le \, \frac 1{2^n\gamma} \log {{\ensuremath{\mathbb E}} }\left[ X_n ^\gamma \right].$$ Note that, to establish ${\textsc{f}}({\beta}, h)=0$, it suffices to prove that $\limsup_n 2^{-n}\log A_n \le 0$, hence our approach yields a substantially stronger piece of information, [*i.e.*]{} that the fractional moment $A_n$ does go to zero.
In order to find a $k$ such that holds we introduce a new probability measure ${\widetilde}{{\ensuremath{\mathbb P}} }$ (which is going to depend on $k$) such that ${\widetilde}{{\ensuremath{\mathbb P}} }$ and ${{\ensuremath{\mathbb P}} }$ are equivalent, that is mutually absolutely continuous. By Hölder’s inequality applied for $p=1/\gamma$ and $q=1/(1-\gamma)$ we have $$\begin{split}
\label{eq:hold}
A_k\, =\, {\widetilde}{{\ensuremath{\mathbb E}} }\left[ \frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\;[X_k-(B-1)]_+^{\gamma}\right] \le
\left({{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\frac{\gamma}{1-\gamma}}\right]\right)^{1-\gamma}
\left({\widetilde}{{\ensuremath{\mathbb E}} }\left[ [X_k-(B-1)]_+\right]\right)^{\gamma},
\end{split}$$ and a sufficient condition for is therefore that $$\label{eq:cond1}
{\widetilde}{{\ensuremath{\mathbb E}} }\left[ [X_k-(B-1)]_+\right]\, \le\, \left({{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\frac{\gamma}{1-\gamma}}\right]\right)^{1-\frac{1}{\gamma}}\left(B^{\gamma}-2(B-1)^{\gamma}\right)^{\frac{1}{\gamma}}.$$
Let $x^{(0)}_n$ be obtained applying $n$ times the annealed iteration $x\mapsto (x^2+(B-1))/B$ to the initial condition $x^{(0)}_0=0$. One has that $x_n^{(0)}$ approaches monotonically the stable fixed point $(B-1)$. Since the coefficients in the iteration are positive, one has for every $h,{\beta},{\omega}$ that $X_n
\ge x_n^{(0)} \stackrel{n\to\infty}\nearrow B-1$ (this is a deterministic bound) and therefore, for any given $\zeta>0$, one can find an integer $n_\zeta$ such that if $n\ge n_\zeta$ we have $${\widetilde}{{\ensuremath{\mathbb E}} }\left[ [X_n-(B-1)]_+\right]\, \le \, {\widetilde}{{\ensuremath{\mathbb E}} }\left[ X_n-(B-1)\right]+\frac \zeta 4.$$ Moreover, since $\left(B^{\gamma}-2(B-1)^{\gamma}\right)^{\frac{1}{\gamma}}- (2-B) \stackrel{\gamma
\nearrow 1}\sim
-c_B (1-\gamma)$ for some $c_B>0$, one can find $\gamma=\gamma_\zeta$ such that $\left(B^{\gamma}-2(B-1)^{\gamma}\right)^{\frac{1}{\gamma}}\ge 2-B-\zeta/4$. At this point, if $\gamma=\gamma_\zeta$, $k\ge n_\zeta$ and if ${\widetilde}{{\ensuremath{\mathbb P}} }$ is such that $$\left({{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\frac{\gamma}{1-\gamma}}\right]\right)^{1-\frac{1}{\gamma}}\ge 1-\frac {\zeta}4.$$ (recall that ${\widetilde}{{\ensuremath{\mathbb P}} }$ depends on $k$) and ${\widetilde}{{\ensuremath{\mathbb E}} }[X_k]\le 1-\zeta$ then is satisfied and ${\textsc{f}}({\beta}, h)=0$.
We sum up what we have obtained:
\[th:moment\]Let $\zeta>0$ and choose $\gamma(=\gamma_\zeta)$ and $n_\zeta$ as above. If there exists $k\ge n_\zeta$ and a probability measure ${\widetilde}{{\ensuremath{\mathbb P}} }$ (such that ${{\ensuremath{\mathbb P}} }$ and ${\widetilde}{{\ensuremath{\mathbb P}} }$ are equivalent probabilities) such that $$\label{eq:close}
\left({{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\frac{\gamma}{1-\gamma}}\right]\right)^{1-\frac{1}{\gamma}}\, \ge \, 1-\frac{\zeta}4 ,$$ and $${\widetilde}{{\ensuremath{\mathbb E}} }[X_k]\le 1-\zeta,$$ then the free energy is equal to zero.
### The change of measure
In order to use wisely the result of the previous section, we have to find a measure ${\widetilde}{{\ensuremath{\mathbb E}} }:={\widetilde}{{\ensuremath{\mathbb E}} }_n$ on the environment which is, in a sense, close to ${{\ensuremath{\mathbb E}} }$ ([*cf.*]{} ), and that lowers significantly the expectation of $X_n$. In [@cf:GLT] we introduced the idea of changing the mean of the ${\omega}$-variables, while keeping their IID character. This strategy was enough to prove disorder relevance for $B<B_c$, but it is not effective in the marginal case $B=B_c$ we are considering here. Here, instead, we choose to introduce [*weak, long range*]{} negative correlations between the different ${\omega}_i$ without changing the laws of the 1-dimensional marginals. As it will be clear, the covariance structure we choose reflects the hierarchical structure of the model we are considering.
In the sequel we take $h\ge h_c(0)=0$.
We define ${\widetilde}{{\ensuremath{\mathbb P}} }_n$ by stipulating that the variables ${\omega}_i,i>2^n,$ are still IID standard Gaussian independent of ${\omega}_1, \ldots, {\omega}_{2^n}$, while ${\omega}_1, \ldots, {\omega}_{2^n}$ are Gaussian, centered, and with covariance matrix $$\label{eq:covar}
C:=I-{\varepsilon}V,$$ where $I$ is the $2^n\times 2^n$ identity matrix, ${\varepsilon}>0$ and $V$ is a symmetric $2^n\times 2^n$ matrix with zero diagonal terms and with positive off-diagonal terms (${\varepsilon}$ and $V$ will be specified in a moment). The choice $V_{ii}=0$ implies of course $\text{Trace}(V)=0$, and we are also going to impose that the Hilbert-Schmidt norm of $V$ verifies $\Vert V \Vert ^2 := \sum_{i,j } V_{i,j}^2 =\text{Trace}(V^2)=1$. This in particular implies that $C$ is positive definite (so that ${\widetilde}{{\ensuremath{\mathbb P}} }_n$ exists!) as soon as ${\varepsilon}<1$: this is because $\Vert V \Vert$, being a matrix norm, dominates the spectral radius of $V$.
Now, still without choosing $V$ explicitly, we compute a lower bound for the left–hand side of . The mutual density of ${\widetilde}{{\ensuremath{\mathbb P}} }_n$ and ${{\ensuremath{\mathbb P}} }$ is $$\label{eq:densit}
\frac{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }_n}{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}({\omega})\, =
\,
\frac{e^{-1/2 ((C^{-1}-I){\omega},{\omega})}}{\sqrt{\det C}},$$ with the notation $(Av,v):=\sum_{1\le i,j\le 2^n}A_{ij}v_i v_j$, and therefore a straightforward Gaussian computation gives $$\label{eq:dets}
\left( {{\ensuremath{\mathbb E}} }\left[ \left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }_n}\right)^{\gamma/(1-\gamma)}\right]\right)^{1-1/\gamma}\, =\,
\frac{(\det[I-({\varepsilon}/(1-\gamma)) V])^{(1-\gamma)/(2\gamma)}}{(\det C)^{1/(2\gamma)}}.$$ If we want to prove a lower bound of the type , a necessary condition is of course that the numerator in is positive: this is ensured by requiring ${\varepsilon}< 1-\gamma$. For the next computation we are going to require also that $ {\varepsilon}/(1-\gamma)
\le 1/2$: we are going in fact to use that $\log (1+x) \ge x -x^2$ if $x \ge -1/2$, and $\text{Trace}(V)=0$ to obtain that $$\begin{gathered}
\label{eq:dano}
\det\left[I-({\varepsilon}/(1-\gamma)) V\right] \, =\,
\exp \left( \text{Trace}( \log (I-({\varepsilon}/(1-\gamma))V)) \right)\\
\ge \,
\exp\left( -\frac{{\varepsilon}^2}{(1-\gamma)^2} \Vert V \Vert ^2 \right),\end{gathered}$$ while $\log (1+x) \le x$ and the traceless character of $V$ directly imply $\det C \le 1$ so that finally $$\label{eq:dets-est}
\left( {{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }_n}\right)^{\gamma/(1-\gamma)}\right]\right)^{1-1/\gamma}\, \ge\,
\exp \left( -\frac{{\varepsilon}^2}{2\gamma(1-\gamma)} \right).$$
Next, we estimate the expected value of $X_n$ under the modified measure: from we see that $$\begin{split}
\label{eq:EZN}
{\widetilde}{{\ensuremath{\mathbb E}} }_n X_n&\, =\, {{\ensuremath{\mathbf E}} }_n\left[e^{(h-(\beta^2/2))\sum_{i=1}^{2^n}
\delta_i}\,{\widetilde}{{\ensuremath{\mathbb E}} }_n e^{\sum_{i=1}^{2^n}
{\beta}{\omega}_i\delta_i}\right]\\
&\, =\, {{\ensuremath{\mathbf E}} }_n\left[e^{-{\varepsilon}(\beta^2/2) (V\delta,\delta)+\sum_{i=1}^{2^n}
h\delta_i}\right]
\, \le\, e^{2^n h}\,{{\ensuremath{\mathbf E}} }_n\left[e^{-{\varepsilon}(\beta^2/2)
(V\delta,\delta)}\right].
\end{split}$$
Finally we choose $V$. From , it is not hard to guess that the most convenient choice, subject to the constraint $\Vert V\Vert^2=1$, is $$\label{eq:V}
V_{ij}\, =\, {{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]\bigg /\sqrt{\sum_{1\le i\ne j\le 2^n}({{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j])^2},$$ for $i\ne j$, while we recall that $V_{ii}=0$. The normalization in can be computed with the help of Proposition \[th:deltas\]: $$\label{eq:2n}
\sum_{1\le i\ne j\le 2^n}\left({{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]\right)^2=
2^n\sum_{1<j\le 2^n}\left({{\ensuremath{\mathbf E}} }_n[\delta_1\delta_j]\right)^2=
2^n\sum_{1\le a\le n}\frac{2^{a-1}}{B_c^{2(n+(a-1))}}=n.$$ In the second equality, we used the fact that there are $2^{a-1}$ values of $1<j\le 2^n$ such that the two branches of the tree $\mathcal T^{(n)}_{\{1,j\}}$ join at level $a$ ([*cf.*]{} the notations of Section \[sec:trees\]), and such tree contains $n+a-1$ nodes.
As a side remark, note that if $B_c<B<2$ (irrelevant disorder regime) the left-hand side of instead goes to zero with $n$, while for $1<B<B_c$ (relevant disorder regime) it diverges exponentially with $n$.
So, in the end, our choice for $V$ is: $$\label{eq:VV}
V_{ij}=\left\{
\begin{array}{lll}
{{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]}/{\sqrt{n}} &\mbox{if}& i\ne j\\
0&\mbox{if}& i=j.
\end{array}
\right.$$
### Checking the conditions of Lemma \[th:moment\]
To conclude the proof of Theorem \[th:main\] we have to show that if $\beta\le \beta_0$ and $h\le \exp(-c_3/\beta^4)$ (and provided that $c_3= c_3(\beta_0)$ is chosen large enough) the conditions of Lemma \[th:moment\] are satisfied. The main point is therefore to estimate the expectation of $X_n$ under ${\widetilde}{{\ensuremath{\mathbb P}} }_n$.
Recalling that ([*cf.*]{} ) $$\label{eq:q2}
{\widetilde}{{\ensuremath{\mathbb E}} }_n X_n\le {{\ensuremath{\mathbf E}} }_n \left[e^{-(\beta^2/2){\varepsilon}\sum_{1\le i\ne j\le 2^n}\delta_i\delta_j
\frac{{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]}{\sqrt{n}}}\right]e^{2^n h}
,$$ we define $$Y_n:=\sum_{1\le i\ne j\le 2^n}\delta_i\delta_j
\frac{{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]}{n}.$$ Thanks to , we know that ${{\ensuremath{\mathbf E}} }_n (Y_n)=1$, so that the Paley-Zygmund inequality gives $$\label{eq:PZ}
{{\ensuremath{\mathbf P}} }_n \left(Y_n\ge 1/2\right) \,=\,
{{\ensuremath{\mathbf P}} }_n \left(Y_n\ge (1/2){{\ensuremath{\mathbf E}} }_n(Y_n)\right)\, \ge
\, \frac{({{\ensuremath{\mathbf E}} }_n(Y_n))^2}{4\,{{\ensuremath{\mathbf E}} }_n(Y_n^2)}=
\frac{1}{4\,{{\ensuremath{\mathbf E}} }_n(Y_n^2)}.$$ We need therefore the following estimate, which will be proved at the end of the section:
\[th:secmom\] We have: $$(1\le )\;\mathcal K:=\sup_{n} {{\ensuremath{\mathbf E}} }_n[Y_n^2]<\infty.$$
Together with this implies $${{\ensuremath{\mathbf P}} }_n[Y_n\ge 1/2]\,\ge \, \frac{1}{4\mathcal K},$$ so that, for all $n\ge 0$, $$\label{eq:bd}
\begin{split}
{{\ensuremath{\mathbf E}} }_n \left[e^{-(\beta^2/2){\varepsilon}\sum_{1\le i\ne j\le 2^n}\delta_i\delta_j
\frac{{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]}{\sqrt{n}}}\right]\, &=\,
{{\ensuremath{\mathbf E}} }_n\left[e^{-\frac{\sqrt{n}{\beta}^2{\varepsilon}}2 Y_n}\right]\\
&\le \, 1-\frac{1}{4\mathcal K}\left(1-4
\mathcal K\exp\left(-\frac{\sqrt{n}{\beta}^2{\varepsilon}}{4}\right)\right).
\end{split}$$ We fix $\zeta:=1/(40\mathcal K)$ and we choose $\gamma=\gamma_\zeta$ ([*cf.*]{} Lemma \[th:moment\]) and ${\varepsilon}$ in small enough so that ([*cf.*]{} ) $$\label{eq:upb2}
\left[ {{\ensuremath{\mathbb E}} }\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }_n}\right)^{\gamma/(1-\gamma)}\right]^{1-1/\gamma}\, \ge\,
\exp \left( -\frac{{\varepsilon}^2}{2\gamma(1-\gamma)} \right)\ge 1-\frac{\zeta}{4}.$$ Then one can check with the help of that for $n\ge {50\mathcal K}/{({\beta}^4{\varepsilon}^2)}$, $$\label{eq:upb3}
{{\ensuremath{\mathbf E}} }_n \left[e^{-(\beta^2/2){\varepsilon}\sum_{1\le i\ne j\le 2^n}\delta_i\delta_j
\frac{{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]}{\sqrt{n}}}\right]\, \le \, 1-3\zeta.$$ We choose $n=n_{\beta}$ in $\left[\frac{50\mathcal K}{{\beta}^4{\varepsilon}^2},\frac{50\mathcal K}{{\beta}^4{\varepsilon}^2}+1\right)$ and $h=\zeta 2^{-n}$. If ${\varepsilon}$ has been chosen small enough above (how small, depending only on $\beta_0$), this guarantees that $n\ge n_\zeta$, where $n_\zeta$ was defined just before Lemma \[th:moment\]. Injecting in finally gives $${\widetilde}{{\ensuremath{\mathbb E}} }[X_n]\le (1-3\zeta)e^{\zeta}\le 1-\zeta.$$
The two conditions of Lemma \[th:moment\] are therefore verified, which ensures that the free energy is zero for this value of $h$. In conclusion, for every ${\beta}\le \beta_0$ we have proven that $$h_c({\beta})\,\ge
\, \zeta\, 2^{-n_{{\beta}}}\, \ge\, \frac 1{80\mathcal K }\exp\left(-\frac{50 \mathcal K \log 2}{{\beta}^4{\varepsilon}^2} \right),$$ for some ${\varepsilon}={\varepsilon}(\beta_0)$ sufficiently small but independent of ${\beta}$.
We have $$\label{eq:Y2}
{{\ensuremath{\mathbf E}} }_n(Y_n^2)=\frac 1{n^2}\sum_{1\le i\ne j\le 2^n}\sum_{1\le k\ne l\le 2^n}{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]{{\ensuremath{\mathbf E}} }_n[\delta_k\delta_l]
{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j\delta_k\delta_l].$$ We will consider only the contribution coming from the terms such that $i\ne k,l$ and $j\ne k,l$. The remaining terms can be treated similarly and their global contribution is easily seen to be exponentially small in $n$. (For instance, when $i=k$ and $j=l$ one gets $$\frac 1{n^2}\sum_{1\le i\ne j\le 2^n}{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]^3
\, \le\,
\frac1n{{\ensuremath{\mathbf E}} }_n(Y_n)\max_{1\le i<j\le 2^n}{{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j],$$ which is exponentially small in $n$, in view of Theorem \[th:deltas\].)
=12 cm \[r\][level $0$: the leaves]{} \[r\][ $v$]{} \[r\][level $2$]{} \[c\][level $n=4$: the root]{} \[c\][$\ldots$]{} \[c\][level $N$: the root]{}
From now on, therefore, we assume that $i,j,k,l$ are all distinct. Two cases can occur:
1. the tree $\mathcal T^{(n)}_{\{i,j,k,l\}}$ (it is better to view it here has the backbone tree, not as the Galton-Watson tree, see Figure \[fig:treemarg\]) has two branches, which themselves bifurcate into two sub-branches, [*cf.*]{} Fig. \[fig:2t\](a) for an example. We call $c$ the level at which the first bifurcation occurs ($c=n$ in the example of Fig. \[fig:2t\](a)), and $a,b$ the levels at which the two branches bifurcate. One has clearly $1\le a<c\le n$ and $1\le b<c\le n$. All trees of this form can be obtained as follows: first choose a leaf $f_1$, between $1$ and $ 2^n$. Then choose $f_2$ among the $2^{a-1}$ possible ones which join with $f_1$ at level $a$, $f_3$ among the $2^{c-1}$ which join with $f_1$ at level $c$ and finally $f_4$ among the $2^{b-1}$ which join with $f_3$ at level $b$. Clearly we are over-counting the trees (note for example that already in the choice of $f_1$ and $f_2$ we are over-counting by a factor $2$), but we are only after an [*upper bound*]{} for ${{\ensuremath{\mathbf E}} }_n(Y_n^2)$ (the same remark applies to case (2) below). We still have to specify how to identify $(f_1,f_2,f_3,f_4)$ with a permutation of $(i,j,k,l)$. When $(f_1,f_2,f_3,f_4)=(i,j,k,l)$ we get the following contribution to : $$\label{eq:caso11}
\frac1{n^2} \sum_{1\le a<c\le n}\sum_{1\le b<c\le n}\frac{2^{n+a+b+c-3}}{B_c^{n+a+b+c-3}B_c^{n+a-1} B_c^{n+b-1}},$$ where we used Theorem \[th:deltas\] to write, [*e.g.*]{}, ${{\ensuremath{\mathbf E}} }_n[\delta_i\delta_j]=B_c^{-n-a+1}$. Since $B_c=\sqrt 2$ we can rewrite as $$\frac1{\sqrt 2n^2}\sum_{1<c\le n}(c-1)^2 2^{-(n-c)/2},$$ which is clearly bounded as $n$ grows.
If instead $(f_1,f_2,f_3,f_4)=(i,k,j,l)$ or $(f_1,f_2,f_3,f_4)=(i,k,l,j)$, one gets $$\frac1{n^2} \sum_{1\le a<c\le n}\sum_{1\le b<c\le n}\frac{2^{n+a+b+c-3}}{B_c^{n+a+b+c-3}B_c^{n+c-1} B_c^{n+c-1}},$$ which is easily seen to be $O(1/n^2)$.
All the other permutations of $(i,j,k,l)$ give a contribution which equals, by symmetry, one of the three we just considered.
2. the tree $\mathcal T^{(n)}_{\{i,j,k,l\}}$ has two branches: one of them does not bifurcate, the other one bifurcates into two sub-branches, one of which bifurcates into two sub-sub-branches, [*cf.*]{} Figure \[fig:2t\](b). Let $a_1,a_2,a_3$ be the levels where the three bifurcations occur, ordered so that $1\le a_1<a_2<a_3\le n$. This time, we choose $f_1$ between $1$ and $2^n$ and then, for $i=1,2,3$, $f_{i+1}$ among the $2^{a_i-1}$ leaves which join with $f_1$ at level $a_i$. If $(f_1,f_2,f_3,f_4)=(i,j,k,l)$ one has in this case $$\begin{gathered}
\phantom{move}
\frac1{n^2} \sum_{1\le a_1<a_2<a_3\le n}\frac{2^{n+a_1+a_2+a_3-3}}{B_c^{n+a_1+a_2+a_3-3}B_c^{n+a_1-1} B_c^{n+a_3-1}}\,
=
\\
\frac1{\sqrt 2 n^2} \sum_{1\le a_1<a_2<a_3\le n}2^{-(n-a_2)/2},\end{gathered}$$ which is $O(1/n)$. Finally, when $(f_1,f_2,f_3,f_4)$ is equal to $(i,k,j,l)$ or to $(i,k,l,j)$ one gets $$\begin{gathered}
\phantom{move}
\frac1{n^2} \sum_{1\le a_1<a_2<a_3\le n}\frac{2^{n+a_1+a_2+a_3-3}}{B_c^{n+a_1+a_2+a_3-3}B_c^{n+a_2-1} B_c^{n+a_3-1}}\,=
\\
\frac1{\sqrt 2 n^2} \sum_{1\le a_1<a_2<a_3\le n}2^{-(n-a_1)/2},\end{gathered}$$ which is $O(1/n^2)$.
Marginal relevance of disorder: the non-hierarchical case {#sec:nHproof}
=========================================================
Here we prove Theorem \[th:main2\] and therefore we assume that holds with $\alpha=1/2$.
We choose and fix once and for all a $\gamma \in (2/3,1)$ and set for $h>0$ $$\label{eq:kh1}
k:=k(h):=\left\lfloor \frac 1h\right\rfloor.$$
\[rem:k\] In [@cf:DGLT] the choice $k(h)= \lfloor 1/{\textsc{f}}(0,h)\rfloor $ was made and it corresponds to choosing $k(h)$ equal to the correlation length of the annealed system. In our case $1/{\textsc{f}}(0,h) \stackrel{h\searrow 0}
\sim 1/(c_K h^2)$ ([*cf.*]{} ) and therefore may look surprising. However, there is nothing particularly deep behind: for ${\alpha}=1/2$, due to the fact that we have to prove delocalization for $h\le \exp(-c_7/{\beta}^4)$, choosing $k(h)$ that diverges for small $h$ like $1/h$ instead of $1/h^2$ just leads to choosing $c_7$ different by a factor $2$ (and we do not track the precise value of constants). We take this occasion to stress that it is practical to work always with sufficiently large values of $k(h)$, and this can be achieved by choosing $c_7$ sufficiently large.
We divide ${\mathbb{N}}$ into blocks $$\label{eq:blocks}
B_i:=\{(i-1)k+1,(i-1)k+2,\ldots,ik\}\;\mbox{with\;\;}i=1,2,\ldots .$$ From now on we assume that $(N/k)$ is integer, and of course it is also the number of blocks contained in the interval $\{1,\ldots,N\}$.
We define, in analogy with the hierarchical case, $$A_N\, :=\, {{\ensuremath{\mathbb E}} }\left(Z_{N,{\omega}}^\gamma\right),$$ and we note that, as in , Jensen’s inequality implies that a sufficient condition for ${\textsc{f}}(\beta,h)=0$ is that $A_N$ does not diverge when $N\to\infty$. Therefore, our task is to show that for every ${\beta}_0>0$ we can find $c_7>0$ such that for every $\beta\le \beta_0$ and $h$ such that $$\label{eq:assh}
0<h\le \exp(-c_7/\beta^4),$$ one has that $\sup_N A_N<\infty$.
Decomposition of $Z_{N,{\omega}}$ and change of measure
-------------------------------------------------------
The first step is a decomposition of the partition function similar to that used in [@cf:T_cg], which is a refinement of the strategy employed in [@cf:DGLT]. For $0<i\le j$ we let $Z_{i,j}:=Z_{(j-i),\theta^i{\omega}}$, with $(\theta^i{\omega})_a:={\omega}_{i+a},
a\in {\mathbb{N}}$, [*i.e.*]{}, $\theta^i{\omega}$ is the result of the application to ${\omega}$ of a shift by $i$ units to the left. We decompose $Z_{N,{\omega}}$ according to the value of the first point ($n_1$) of $\tau$ after $0$, the last point ($j_1$) of $\tau$ not exceeding $n_1+k-1$, then the first point ($n_2$) of $\tau$ after $j_1$, and so on. We call $i_r$ the index of the block in which $n_r$ falls, and $\ell:=\max\{r:n_r\le N\}$, see Figure \[fig:decomp\]. Due to the constraint $N\in\tau$, one has always $i_\ell=(N/k)$.
=12.4 cm \[c\][$0$]{} \[c\][$i_0=j_0=0$]{} \[c\][N]{} \[c\][$B_1$]{} \[c\][$B_3$]{} \[c\][$B_4$]{} \[c\][$B_9$]{} \[c\][$B_{10}$]{} \[c\][$B_{11}$]{} \[c\][$B_{14}$]{} \[c\][$Z_{n_1,j_1}$]{} \[c\][$Z_{n_2,j_2}$]{} \[c\][$Z_{n_3,j_3}$]{} \[c\][$Z_{n_4,N}$]{} \[c\][ $n_1$]{} \[c\][ $j_1$]{} \[c\][ $n_2$]{} \[c\][ $j_2$]{} \[c\][ $n_3$]{} \[c\][ $j_3$]{} \[c\][ $n_4$]{} \[c\][ $j_1$]{} \[c\][$k$]{} \[c\][$2k$]{} \[c\][ $n_1+k$]{} \[c\][ $n_2+k$]{} \[c\][ $n_3+k$]{}
In formulas: $$\label{eq:dec1}
Z_{N,{\omega}}=\sum_{\ell=1}^{N/k}\sum_{i_0:=0<i_1<\ldots<i_\ell=N/k}
{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)},$$ where $$\begin{gathered}
\label{eq:Zhat}
{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}\, :=\,
\sum_{n_1\in B_{i_1}}\sum_{j_1=n_1}^{n_1+k-1}{\sum_{\substack{n_2\in B_{i_2}: \\ n_2\ge
n_1+k}}}\sum_{j_2=n_2}^{n_2+k-1}\ldots {\sum_{\substack{n_{\ell-1}\in B_{i_{\ell-1}}: \\ n_{\ell-1}\ge
n_{\ell-2}+k}}}\sum_{j_{\ell-1}=n_{\ell-1}}^{n_{\ell-1}+k-1}{\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge n_{\ell-1}+k}}}\\
z_{n_1}K(n_1)Z_{n_1,j_1}z_{n_2}K(n_2-j_1)
Z_{n_2,j_2}\, \ldots \, z_{n_\ell}K(n_\ell-j_{\ell-1})Z_{n_\ell,N},\end{gathered}$$ and $z_n:=e^{\beta{\omega}_n+h-\beta^2/2}$.
Then, from inequality , we have $$\label{eq:A_nh}
A_N\,\le\, \sum_{\ell=1}^{N/k}\sum_{i_0:=0<i_1<\ldots<i_\ell=N/k}
{{\ensuremath{\mathbb E}} }\left[({\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)})^\gamma\right],$$ and, as in , we apply Hölder’s inequality to get $$\begin{gathered}
\label{eq:Hold}
{{\ensuremath{\mathbb E}} }\left[\left({\widehat}Z^{(i_1,\ldots,i_\ell)}_{\omega}\right)^\gamma\right]\,
=\\
{\widetilde}{{\ensuremath{\mathbb E}} }\left[\left({\widehat}Z^{(i_1,\ldots,i_\ell)}_{\omega}\right)^\gamma
\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}({\omega})\right]
\le \left(
{\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z^{(i_1,\ldots,i_\ell)}_{\omega}\right)^\gamma
\left({{\ensuremath{\mathbb E}} }\left[\left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\gamma/(1-\gamma)}\right]
\right)^{1-\gamma}.\end{gathered}$$ The new law ${\widetilde}{{\ensuremath{\mathbb P}} }:= {\widetilde}{{\ensuremath{\mathbb P}} }^{(i_1,\ldots,i_\ell)}$ will be taken to depend on the set $(i_1,\ldots,i_\ell)$. In order to define it, let first of all $$\label{eq:M}
M:=M(i_1,\ldots,i_\ell):=\{i_1,i_2,\ldots,i_\ell\}\cup \{i_1+1,i_2+1,\ldots,
i_{\ell-1}+1\}.$$ Then, we say that under ${\widetilde}{{\ensuremath{\mathbb P}} }$ the random vector ${\omega}$ is Gaussian, centered and with covariance matrix $$\label{eq:covarianze}
{\widetilde}{{\ensuremath{\mathbb E}} }({\omega}_i{\omega}_j)\, =\, {\mathbf{1}}_{i=j}-\mathcal C_{ij}\,:=\,
\begin{cases}
{\mathbf{1}}_{i=j}-H_{ij} & \text{ if there exists } u\in M \text{ such that } i,j \in B_u,
\\
{\mathbf{1}}_{i=j} & \text{ otherwise,}
\end{cases}
$$ and $$H_{ij}\,:=\,
\begin{cases}
(1-\gamma)/\sqrt{9\, k(\log k) \;|i-j|}&\text{ if } i\ne j,
\\
0& \text{ if } i=j.
\label{eq:H}
\end{cases}$$ Note that all the $\mathcal C_{ij}$’s are non-negative. It is immediate to check that the $k\times k$ symmetric matrix ${\widehat}H:=\{H_{ij}\}_{i,j=1}^k$ satisfies $$\label{eq:HS_H}
\Vert{\widehat}H\Vert\, :=\,
\sqrt{\sum_{i,j=1}^k H_{ij}^2}\, \le \frac{1-\gamma}2,$$ for $k$ sufficiently large. In words: $\omega_n$’s in different blocks are independent; in blocks $B_u$ with $u\notin M$ they are just IID standard Gaussian random variables, while if $u\in M$ then the random vector $\{{\omega}_n\}_{n\in B_u}$ has covariance matrix $I-{\widehat}H$, where $I$ is the $k\times k$ identity matrix. Note that, since $\Vert{\widehat}H\Vert$ dominates the spectral radius of ${\widehat}H$, guarantees that $I -{\widehat}H$ is positive definite (and also that $I -(1-\gamma)^{-1}{\widehat}H$ is positive definite, that will be needed just below).
The last factor in the right-hand side of is easily obtained recalling and independence of the ${\omega}_n$’s in different blocks, and one gets $$\label{eq:RN_nh}
\left({\widetilde}{{\ensuremath{\mathbb E}} }\left[ \left(\frac{{\,\text{\rm d}}{{\ensuremath{\mathbb P}} }}{{\,\text{\rm d}}{\widetilde}{{\ensuremath{\mathbb P}} }}\right)^{\gamma/(1-\gamma)}\right]
\right)^{1-\gamma}=\left(\frac{\det(I-{\widehat}H)}
{(\det(I-1/(1-\gamma){\widehat}H))^{1-\gamma}}
\right)^{|M|/2}.$$ Since ${\widehat}H$ has trace zero and its (Hilbert-Schmidt) norm satisfies , one can apply $\det(I-{\widehat}H)\le
\exp(-\mbox{Trace}({\widehat}H))=1$ and (with $V$ replaced by ${\widehat}H$ and ${\varepsilon}$ by $1$) to get that the right-hand side of is bounded above by $ \exp(|M|/2)$, which in turn is bounded by $\exp(\ell)$. Together with and , we conclude that $$A_N\le\sum_{\ell=1}^{N/k}\sum_{i_0:=0<i_1<\ldots<i_\ell=N/k}
e^{{\ell}} \left[{\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}
\right]^\gamma.
\label{eq:sqbr}$$
Reduction to a non-disordered model
------------------------------------
We wish to bound the right-hand side of with the partition function of a non-disordered pinning model in the delocalized phase, which goes to zero for large $N$. We start by claiming that $$\begin{gathered}
\label{eq:claimU}
{\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}\, \le\, \sum_{n_1\in B_{i_1}}\ldots {\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge
n_{\ell-1}+k}}} K(n_1)K(n_2-j_1)\ldots K(n_\ell-j_{\ell-1})\\
\times U(j_1-n_1)U(j_2-n_2)\ldots U(N-n_\ell),\end{gathered}$$ where $$\label{eq:U}
U(n)\, =\, c_{8}{{\ensuremath{\mathbf P}} }(n\in\tau){{\ensuremath{\mathbf E}} }\left[e^{-\beta^2\sum_{1\le i<j\le {n/2}}H_{ij}\delta_i
\delta_j}
\right],$$ and $c_8$ is a positive constant depending only on $K(\cdot)$. This is proven in Appendix \[sec:appprova\]. We are also going to make use of:
\[th:condlemma\] \[th:lemma\_cg\] There exists $C_2=C_2(K(\cdot))<\infty$ such that if, for some $\eta>0$, $$\label{eq:condlemma1}
\sum_{j=0}^{k-1}U(j)\le \eta {\sqrt k}$$ and $$\label{eq:condlemma2}
\sum_{j=0}^{k-1}\sum_{n\ge k}U(j)K(n-j)\le \eta,$$ then there exists $C_1=C_1(\eta, k,K(\cdot))$ such that the right-hand side of is bounded above by $$C_1 \eta^\ell C_2^{\ell}\prod_{r=1}^\ell\frac1{(i_r-i_{r-1})^{3/2}}.$$
It is important to note that $C_2$ does not depend on $\eta$.
Lemma \[th:condlemma\] is a small variation on [@cf:T_cg Lemma 3.1], but, both because the model we are considering is somewhat different and for sake of completeness, we give the details of the proof in Appendix \[sec:appprova\].
Now assume that conditions - are verified for some $\eta$. Collecting , and Lemma \[th:lemma\_cg\], we have then $$A_N\le C_1^\gamma\sum_{\ell=1}^{N/k}\sum_{i_0:=0<i_1<\ldots<i_\ell=N/k}
\left(\eta^\gamma\,C_2^{\gamma}e \right)^\ell\prod_{r=1}^\ell\frac1{(i_r-i_{r-1})^{(3/2)\gamma}}.
\label{eq:puremod}$$ In the right-hand side we recognize, apart from the irrelevant multiplicative constant $C_1^\gamma$, the partition function of a non-random $(\beta=0)$ pinning model with $N$ replaced by $N/k$, $K(\cdot)$ replaced by $${\widehat}K(n)\, =\, \frac1{n^{(3/2)\gamma}}\frac1{\sum_{i\ge1}i^{-(3/2)\gamma}},$$ and $h$ replaced by $$\label{eq:hbar}
{\widehat}h:=\log \left(\eta^\gamma\,C_2^{\gamma}e \sum_{n\in{\mathbb{N}}}\frac1{n^{(3/2)\gamma}}\right).$$ Note that ${\widehat}K(\cdot)$ is normalized to be a probability measure on ${\mathbb{N}}$, which is possible since (by assumption) $\gamma>2/3$, and that it has a power-law tail with exponent $(3/2)\gamma>1$. Thanks to Lemma \[th:puro\] below, one has that the right-hand side of tends to zero for $N\to\infty$ whenever $$\label{eq:condeta}
{\widehat}h<0.$$ Therefore, if $\eta$ is so small that holds, we can conclude that $A_N$ tends to zero for $N\to\infty$ and therefore ${\textsc{f}}(\beta,h)=0$.
The proof of Theorem \[th:main2\] is therefore concluded once we prove
\[th:stimeU\] Fix $\eta>0$ such that holds. For every ${\beta}_0>0$ there exists $0<c_7<\infty$ such that if $\beta\le {\beta}_0$ and $0<h\le
\exp(-c_7/\beta^4)$, conditions - are verified.
We need to show that the two hypotheses of Lemma \[th:condlemma\] hold and for this we are going to use the following result:
\[th:CE\] Under the law ${{\ensuremath{\mathbf P}} }$, the random variable $$W_L\, :=\, (\sqrt{L} \log L)^{-1}
\sum_{1\le i<j\le L} \delta_i
\delta_j /\sqrt{j-i},$$ converges in distribution, as $L$ tends to $\infty$, to $c \vert Z \vert$ ($Z\sim N(0,1)$ and $c$ a positive constant).
This lemma, the proof of which may be found just below (together with the explicit value of $c$), directly implies that, if we set $S(a,L):= {{\ensuremath{\mathbf E}} }\left[\exp\left(- a W_L\right)\right]$, we have $\lim_{a\to\infty} \lim_{L\to\infty}S(a,L)\, = \, 0$ and, by the monotonicity of $S (\cdot, L)$, we get $$\label{eq:small}
\lim_{a, L\to\infty} S(a,L)\, = \, 0.$$
Let us verify . Note first of all that ([*cf.*]{} and ) $$\begin{gathered}
U(n)\, =\, c_8 {{\ensuremath{\mathbf P}} }(n\in\tau)
S\left(\beta^2(1-\gamma)\sqrt{\log k}\,\sqrt{\frac {n/2}{9\, k}}\,
\frac{\log (n/2)}{
{\log k}}
, \frac n2\right)\\
=:\, c_8 {{\ensuremath{\mathbf P}} }(n\in\tau) s_{\beta}(k,n).\end{gathered}$$ We recall also that ([@cf:Doney Th. B]) $$\label{eq:doney}
{{\ensuremath{\mathbf P}} }(n\in\tau)\stackrel{n\to\infty}\sim \frac1{2\pi C_K \sqrt n},$$ and therefore there exists $c_9>0$ such that for every $n \in {\mathbb{N}}$ $$\label{eq:doney-bound}
{{\ensuremath{\mathbf P}} }(n\in\tau)\, \le \, \frac{c_9}{\sqrt {n}}.$$ Split the sum in according to whether $j \le {\delta}k$ or not (${\delta}={\delta}(\eta) \in (0,1)$ is going to be chosen below). By using $S(a,L)\le 1$ (in the case $j \le {\delta}k$) and we obtain $$\label{eq:interm4}
\sum_{j=0}^{k-1} U(j) \, \le \, c_8+c_8 c_9 \sum_{j=1}^{{\delta}k}
\frac1{\sqrt{j}} \, +\, c_8 c_9 \sum_{j= {\delta}k +1}^{k-1}
\frac1{\sqrt{j}}\, s_{\beta}(k,j).$$ Since if $c_7$ is chosen sufficiently large $$\beta^2\sqrt{\log k}\ge\sqrt{c_7-\beta^4\log 2}\ge \sqrt{c_7}/2,$$ and since $k$ may be made large by increasing $c_7$, we directly see that implies that $s_{\beta}(k,j)$ may be made smaller than (say) ${\delta}$ for every $\delta k <j<k$ by choosing $c_7$ sufficiently large. Therefore implies $$\label{eq:interm5}
\sum_{j=0}^{k-1} U(j) \, \le \, 4c_8 c_9 (\sqrt{{\delta}} + {\delta}) \sqrt{k}.$$ By choosing ${\delta}={\delta}(\eta) $ such that $4c_8 c_9 (\sqrt{{\delta}} + {\delta}) \le \eta$, we have . The proof of is absolutely analogous to the proof of and it is therefore omitted.
Proof of Lemma \[th:CE\]
------------------------
We introduce the notation $$Y_L^{(i)}\, :=\, \sum_{j=i+1}^{L} \frac{{\delta}_j}{\sqrt{j-i}}
,\text{\ so that \ }
W_L \, =\, \frac 1{\sqrt{L}\log L} \sum_{i=1}^{L-1}{\delta}_i Y_L^{(i)}.$$ Let us observe that, thanks to the renewal property of $\tau$, under ${{\ensuremath{\mathbf P}} }(\cdot \vert {\delta}_i=1)$, $Y_L^{(i)}$ is distributed like $Y_{L-i}:= Y_{L-i}^{(0)}$ (under ${{\ensuremath{\mathbf P}} }$). The first step in the proof is observing that, in view of , $$\begin{gathered}
{{\ensuremath{\mathbf E}} }\left[\frac 1{\sqrt{L}\log L} \sum_{i=(1-{\varepsilon})L }^{L-1}{\delta}_i Y_L^{(i)}\right]\,
=\\
\frac1{\log L \sqrt L}\sum_{i=(1-{\varepsilon})L}^{L-1}\sum_{j=i+1}^L\frac{
{{\ensuremath{\mathbf P}} }(i\in\tau){{\ensuremath{\mathbf P}} }(j-i\in\tau)}{\sqrt{j-i}}
\,
=
\, O({\varepsilon}),\end{gathered}$$ uniformly in $L$, so we can focus on studying $W_{L, {\varepsilon}}$, defined as $W_L$, but stopping the sum over $i$ at $(1-{\varepsilon}) L$. At this point we use that $$\label{eq:Chung-Erdos}
\lim_{L \to \infty} \frac{Y_L}{\log L} \, =\, \frac 1{2\pi C_K} \, =: {\widehat}c_K,$$ in $L^2({{\ensuremath{\mathbf P}} })$ (and hence in $L^1({{\ensuremath{\mathbf P}} })$). We postpone the proof of and observe that, thanks to the properties of the logarithm, it implies that for every ${\varepsilon}>0$ $$\label{eq:use}
\lim_{L \to \infty}
\sup_{q \in [{\varepsilon}, 1]} {{\ensuremath{\mathbf E}} }\left[ \left \vert
\frac 1{\log L} \sum_{j=1}^{qL} \frac {{\delta}_j}{\sqrt{j}} \, - {\widehat}c_K \right \vert \right]\, =\, 0.$$
Let us write $$R_L \, := W_{L, {\varepsilon}} \, -\, \frac{{\widehat}c_K} {\sqrt{L} } \sum_{i=1}^{(1-{\varepsilon}) L} {\delta}_i$$ and note that $ L^{-1/2} \sum_{i=1}^{(1-{\varepsilon}) L} {\delta}_i$ converges in law toward $\sqrt{(1-{\varepsilon})/(2\pi C_K^2)} \, \vert Z\vert$. This follows directly by using that the event $ \sum_{i=1}^{L} {\delta}_i \ge m$ is the event $\tau_m \le L$ ($\tau_m$ is of course the $m$-th point in $\tau$ after $0$) and by using the fact that $\tau_1$ is in the domain of attraction of the positive stable law of index $1/2$ [@cf:Feller2 VI.2 and XI.5]. It suffices therefore to show that ${{\ensuremath{\mathbf E}} }[\vert R_L\vert] $ tends to zero. We have $$\begin{gathered}
{{\ensuremath{\mathbf E}} }\left[\vert R_L\vert\right]\, \le\,
\frac {1}{\sqrt{L}} \sum_{i=1}^{(1-{\varepsilon}) L}
{{\ensuremath{\mathbf E}} }[ {\delta}_i ] {{\ensuremath{\mathbf E}} }\left [ \left.\bigg\vert \frac{Y^{(i)}_L}{\log L}-{\widehat}c_K \bigg \vert \;\right\vert
\delta_i=1 \right] \, =\,
\\
\frac {1}{\sqrt{L}} \sum_{i=1}^{(1-{\varepsilon}) L}
{{\ensuremath{\mathbf E}} }[ {\delta}_i ] {{\ensuremath{\mathbf E}} }\left[ \left\vert \frac{Y_{L-i}}{\log L}-{\widehat}c_K \right\vert
\right] \, =\, o(1), \end{gathered}$$ where in the last step we have used and .
Note that we have also proven that $c=(2\pi)^{-3/2} C_K^{-2}$ in the statement of Lemma \[th:CE\].
We are therefore left with the task of proving . This result has been already proven [@cf:chungerdos Th. 6] when $\tau$ is given by the successive returns to zero of a centered, aperiodic and irreducible random walk on ${{\ensuremath{\mathbb Z}} }$ with bounded variance of the increment variable. Note that, by well established local limit theorems, for such a class of random walks we have . Actually in [@cf:chungerdos] it is proven that holds almost surely as a consequence of $\text{var}_{{\ensuremath{\mathbf P}} }(Y_L)=O(\log L)$. What we are going to do is simply to re-obtain such a bound, by repeating the steps in [@cf:chungerdos] and using -, for the general renewal processes that we consider (as a side remark: also in our generalized set-up, almost sure convergence holds).
The proof goes as follows: by using it is straightforward to see that the limit as $L\to \infty$ of ${{\ensuremath{\mathbf E}} }[Y_L/\log L]$ is ${\widehat}c_K$, so that we are done if we show that $\text{var}_{{\ensuremath{\mathbf P}} }(Y_L/\log L)$ vanishes as $L\to \infty$. So we start by observing that $$\label{eq:CEst1}
\text{var}_{{\ensuremath{\mathbf P}} }(Y_L) \, =\, \sum_{i, j} \frac{{{\ensuremath{\mathbf E}} }[ {\delta}_i {\delta}_j]-
{{\ensuremath{\mathbf E}} }[ {\delta}_i] {{\ensuremath{\mathbf E}} }[ {\delta}_j]
}{\sqrt{ij}} \, =\, 2 \sum_{i=1}^{L-1}\sum_{j=i+1}^L
\frac{{{\ensuremath{\mathbf E}} }[ {\delta}_i {\delta}_j]-
{{\ensuremath{\mathbf E}} }[ {\delta}_i] {{\ensuremath{\mathbf E}} }[ {\delta}_j]
}{\sqrt{ij}}
+ O(1),$$ by . Now we compute $$\begin{split}
\sum_{i=1}^{L-1}\sum_{j=i+1}^L
\frac{{{\ensuremath{\mathbf E}} }[ {\delta}_i {\delta}_j]-
{{\ensuremath{\mathbf E}} }[ {\delta}_i] {{\ensuremath{\mathbf E}} }[ {\delta}_j]
}{\sqrt{ij}}
\, &=\,
\sum_{i=1}^{L-1} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_i]}{\sqrt{i}}
\left[\sum_{j=1}^{L-i} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_j]}{\sqrt{j+i}}-
\sum_{j=i+1}^{L} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_j]}{\sqrt{j}}
\right]
\\
&\le \,
\sum_{i=1}^{L-1} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_i]}{\sqrt{i}}
\left[\sum_{j=1}^{L-i} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_j]}{\sqrt{j+i}}-
\sum_{j=i+1}^{L} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_j]}{\sqrt{j+i}}
\right]
\\
&\le \,
\sum_{i=1}^{L-1} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_i]}{\sqrt{i}}
\sum_{j=1}^{i} \frac{{{\ensuremath{\mathbf E}} }[{\delta}_j]}{\sqrt{j+i}}
\\
\le \, \sum_{i=1}^{L-1} &\frac{{{\ensuremath{\mathbf E}} }[{\delta}_i]}{i}
\sum_{j=1}^{i} {{{\ensuremath{\mathbf E}} }[{\delta}_j]} \le c_9^2
\sum_{i=1}^{L-1} \frac{1}{i^{3/2}}
\sum_{j=1}^{i} \frac 1{j^{1/2}} = O(\log L),
\end{split}$$ where, in the last line, we have used . In view of , we have obtained $\text{var}_{{\ensuremath{\mathbf P}} }(Y_L)=O(\log L)$ and the proof , and therefore of Lemma \[th:CE\] is complete.
Some technical results and useful estimates
===========================================
Two results on renewal processes
--------------------------------
The first result concerns the non-disordered pinning model and is well known:
\[th:puro\] Let $ K(\cdot)$ be a probability on ${\mathbb{N}}$ which satisfies for some $\alpha>0$. If $ h<0$, we have that $$\label{eq:lemma1}
\sum_{\ell=1}^N \sum_{i_0:=0<i_1<\ldots<i_\ell=N}
e^{h\ell}\prod_{r=1}^\ell K(i_r-i_{r-1})\stackrel {N\to\infty}\longrightarrow0.$$
This is implied by [@cf:Book Th. 2.2], since the left-hand side of is nothing but the partition function of the homogeneous pinning model of length $N$, whose critical point is $h_c=0$ ([*cf.*]{} also ).
The second fact we need is
\[th:condiz\] There exists a positive constant $c$, which depends only on $K(\cdot)$, such that for every positive function $f_N(\tau)$ which depends only on $\tau\cap\{1,\ldots,N\}$ one has $$\label{eq:condiz}
\sup_{N>0} \frac{{{\ensuremath{\mathbf E}} }[f_N(\tau)|2N\in\tau]}{{{\ensuremath{\mathbf E}} }[f_N(\tau)]}\le c.$$
The statement follows by writing $f_N(\tau)$ as $f_N(\tau)\sum_{n=0}^N {\mathbf{1}}_{\{X_N=n\}}$, where $X_N$ is the last renewal epoch up to (and including) $N$, and using the bound $$\sup_N \max_{n=0,\ldots,N}\frac{
{{\ensuremath{\mathbf P}} }(X_N=n \vert 2N \in \tau)}{{{\ensuremath{\mathbf P}} }(X_N=n )} =:c<\infty ,$$ which is equation (A.15) in [@cf:DGLT] (this has been proven also in [@cf:T_cg], where the proof is repeated to show that $c$ can be chosen as a function of ${\alpha}$ only).
Proof of {#sec:appprova}
---------
Defining the event $$\label{eq:Omega}
\Omega_{\underline n,\underline j}:=\{N\in\tau\;\;\mbox{and}\;\;
\{j_{r-1},\ldots,n_r\}\cap \tau=\{j_{r-1},n_r\}\;\;\mbox{for all}\;\;
r=1,\ldots,\ell\},$$ with the convention that $j_0:=0$, we have $${\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}=\sum_{n_1\in B_{i_1}}\ldots {\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge
n_{\ell-1}+k}}}{{\ensuremath{\mathbf E}} }\left[e^{\sum_{n=1}^N(\beta{\omega}_n+h-\beta^2/2)\delta_n};\Omega_{\underline n,
\underline j}\right].$$ Since ${\widetilde}{{\ensuremath{\mathbb P}} }$ is a Gaussian measure and $\delta_i^2=\delta_i$ for every $i$, the computation of ${\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}$ is immediate: $${\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}=\sum_{n_1\in B_{i_1}}\ldots {\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge
n_{\ell-1}+k}}}{{\ensuremath{\mathbf E}} }\left[
e^{h\sum_{n=1}^N\delta_n-\beta^2/2\sum_{i,j=1}^N \mathcal C_{ij}\delta_i\delta_j}
;\Omega_{\underline n,
\underline j}\right].$$ In view of $\mathcal C_{ij}\ge 0$, we obtain an upper bound by neglecting in the exponent the terms such that $n_r\le i\le j_{r}$ and $n_{r'}\le j\le j_{r'}$ with $r\ne r'$. At that point, the ${{\ensuremath{\mathbf E}} }$ average may be factorized, by using the renewal property, and we obtain (recall that $\mathcal C_{ii}=0$) $$\begin{split}
{\widetilde}{{\ensuremath{\mathbb E}} }{\widehat}Z_{\omega}^{(i_1,\ldots,i_\ell)}\, \le\, & \sum_{n_1\in B_{i_1}}\ldots {\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge
n_{\ell-1}+k}}}K(n_1)\ldots K(n_\ell-j_{\ell-1})\\
&\times \prod_{r=1}^\ell{{\ensuremath{\mathbf E}} }\left[\left.e^{h\sum_{i=n_r}^{j_r}\delta_i-\beta^2\sum_{n_r\le i< j\le j_r}\mathcal C_{ij}
\delta_i\delta_j}{\mathbf{1}}_{\{j_{r}\in\tau\}}\right|n_r\in \tau\right],
\end{split}$$ with the convention that $j_\ell:=N$. We are left with the task of proving that $$\label{eq:auxU}
{{\ensuremath{\mathbf E}} }\left[\left.e^{h\sum_{i=n_r}^{j_r}\delta_i-\beta^2\sum_{n_r\le i<j\le j_r}\mathcal C_{ij}
\delta_i\delta_j}{\mathbf{1}}_{\{j_{r}\in\tau\}}\right|n_r\in \tau\right]\le U(j_r-n_r),$$ with $U(\cdot)$ satisfying . We remark first of all that the left-hand side of equals $${{\ensuremath{\mathbf P}} }(j_r-n_r\in\tau){{\ensuremath{\mathbf E}} }\left[\left.e^{h\sum_{i=n_r}^{j_r}\delta_i-\beta^2\sum_{n_r\le i<j\le j_r}\mathcal C_{ij}
\delta_i\delta_j}\right|n_r\in \tau,j_r\in\tau\right].$$ Since by construction $j_r-n_r<k(h)=\lfloor 1/h\rfloor$, one has $$e^{h\sum_{i=n_r}^{j_r}\delta_i}\, \le\, e.$$ As for the remaining average, assume without loss of generality that $|\{n_r,n_r+1,\ldots,j_r\}\cap B_{i_r}|\ge (j_r-n_r)/2$ (if this is not the case, the inequality clearly holds with $B_{i_r}$ replaced by $B_{i_r+1}$ and the arguments which follow are trivially modified). Then, $$\begin{gathered}
\label{eq:uffa}
{{\ensuremath{\mathbf E}} }\left[\left.e^{-\beta^2\sum_{n_r\le i< j\le j_r}\mathcal C_{ij}
\delta_i\delta_j
}\right|n_r\in \tau,j_r\in\tau
\right]\, \le
\\
{{\ensuremath{\mathbf E}} }\left[\left.
\exp\left(-\beta^2\sum_{0<i<j\le (j_r-n_r)/2}\delta_i\delta_j H_{ij}\right)
\right| j_r-n_r\in\tau \right].\end{gathered}$$ Finally, the conditioning in can be eliminated using Lemma \[th:condiz\], and is proved.
Proof of Lemma \[th:condlemma\]
--------------------------------
In this proof (and in the statement) two positive numbers $C_1$ and $C_2$ appear. $C_1$ is going to change along with the steps of the proof: it depends on $\eta$, $k$ and on $K(\cdot)$. $C_2$ instead is chosen once and for all below and it depends only on $K(\cdot)$. We start by giving a name to the right-hand side of : $$\begin{gathered}
\label{eq:restart}
Q\, :=\,
\sum_{n_1\in B_{i_1}}\sum_{j_1=n_1}^{n_1+k-1}{\sum_{\substack{n_2\in B_{i_2}: \\ n_2\ge
n_1+k}}}\sum_{j_2=n_2}^{n_2+k-1}\ldots {\sum_{\substack{n_{\ell-1}\in B_{i_{\ell-1}}: \\ n_{\ell-1}\ge
n_{\ell-2}+k}}}\sum_{j_{\ell-1}=n_{\ell-1}}^{n_{\ell-1}+k-1}{\sum_{\substack{n_\ell\in B_{N/k}: \\ n_\ell\ge n_{\ell-1}+k}}}\\
K(n_1)\ldots K(n_\ell-j_{\ell-1})
U(j_1-n_1)\ldots U(j_{\ell-1}-n_{\ell-1}) U(N-n_\ell)
.\end{gathered}$$ Since $N -n _\ell <k$, we can get rid of $U(N-n_\ell)\, (\le c_8
{{\ensuremath{\mathbf P}} }(N-n_\ell \in \tau) )$ and of the right-most sum (on $n_\ell$), replacing $n_\ell$ by $N$, by paying a price that depends on $k$ and $K(\cdot)$ (this price goes into $C_1$). Therefore we have $$\label{eq:lems1}
Q\, \le \,
C_1\, \sum_{n_1\in B_{i_1}} \ldots
\sum_{j_{\ell-1}=n_{\ell-1}}^{n_{\ell-1}+k-1}
K(n_1)\ldots K(n_\ell-j_{\ell-1})
U(j_1-n_1)\ldots U(j_{\ell-1}-n_{\ell-1})
,$$ where by convention from now on $n_\ell:=N$. Now we single out the long jumps. The set of long jump arrival points is defined as $$J\, =\, J( i_1,i_2, \ldots, i_\ell)\, :=\, \left\{r:\, 1 \le r \le \ell, \, i_r>i_{r-1}+2\right\},$$ and the definition guarantees that a long jump $\{j_{r-1},\ldots,n_r\}$ contains at least one whole block with no renewal point inside. For $r\in J$ we use the bound $$K(n_r-j_{r-1}) \, \le \, \frac{C_2}{(i_r-i_{r-1})^{3/2} k^{3/2}},$$ and we stress that we may and do choose $C_2$ depending only on $K(\cdot)$. For later use, we choose $C_2\ge 2^{3/2}$. This leads to $$\begin{gathered}
\label{eq:lems3}
Q\, \le \,
C_1\, {k^{-3|J| /2}}
\prod_{r \in J} \frac{C_2}{(i_r-i_{r-1})^{3/2}}
\\
\times \sum_{n_1\in B_{i_1}} \ldots
\sum_{j_{\ell-1}=n_{\ell-1}}^{n_{\ell-1}+k-1}
\left(
\prod_{r\in \{1,\ldots, \ell\}\setminus J}
K(n_r- j_{r-1})
\right)
\,
U(j_1-n_1)\ldots U(j_{\ell -1}-n_{\ell-1}).\end{gathered}$$ Now we perform the sums in and bound the outcome by using the assumptions and .
We first sum over $j_{r-1}$, $r \in J$, keeping of course into account the constraint $0\le j_{r-1}-n_{r-1}<k$. By using such sum yields at most $(\eta \sqrt{k})^{\vert J\vert}$ if $1 \notin J$. If $1 \in J$, for $r=1$ then $j_0=0$ and there is no summation: we can still bound the sum by $(\eta \sqrt{k})^{\vert J\vert}$, provided that we change the constant $C_1$.
Second, we sum over $j_{r-1},n_r$ for $r \in \{1,\ldots, \ell\}\setminus J$ and use . Once again we have to treat separately the case $r=1$, as above. But if $1 \notin \{1,\ldots, \ell\}\setminus J$ we directly see that the summation is bounded by $\eta^{\ell -\vert J \vert}$.
Finally, we have to sum over $n_r$, for $r\in J$. The summand does not depend on these variables anymore, so this gives at most $k^{\vert J \vert }$.
Putting these estimates together we obtain $$\label{eq:lems4}
Q\, \le \,
C_1\,
\frac{
(\eta \sqrt{k})^{\vert J\vert }
\eta^{\ell -\vert J \vert } k^{\vert J \vert}
}
{k^{3|J| /2}}
\prod_{r \in J} \frac{C_2}{(i_r-i_{r-1})^{3/2}}\, \le\,
C_1 \eta^\ell C_2^\ell
\prod_{r=1}^{\ell} \frac{1}{(i_r-i_{r-1})^{3/2}},$$ where, in the last step, we have used $C_2\ge 2^{3/2}$. The proof of Lemma \[th:condlemma\] is therefore complete.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are very grateful to Bernard Derrida for many enlightening discussions and to an anonymous referee for having observed the link between hierarchical pinning and Galton-Watson processes. The authors acknowledge the support of ANR, grant POLINTBIO. F.T. was partially supported also by ANR, grant LHMSHE.
[99]{}
Abraham, D. B. [Surface Structures and Phase Transitions, Exact Results]{}. *Phase transitions and critical phenomena* **10**, 1–74, Academic Press, London (UK), 1986.
Alexander, K. S. [The effect of disorder on polymer depinning transitions]{}. *Commun. Math. Phys.* **279** (2008), 117–146.
Alexander, K. S. and Zygouras, N. [Quenched and annealed critical points in polymer pinning models]{}. Preprint, arXiv:0805.1708 \[math.PR\].
Alexander, K. S. and Zygouras, N. [Equality of critical points for polymer depinning transitions with loop exponent one]{}. Preprint, arXiv:0811.1902 \[math.PR\].
Aizenman, M. and Molchanov, S. [ Localization at large disorder and at extreme energies: An elementary derivation]{}. *Commun. Math. Phys.* **157** (1993), 245–278.
Bhattacharjee, S. M. and Mukherji, S. [Directed polymers with random interaction: Marginal relevance and novel criticality]{}. *Phys. Rev. Lett.* **70** (1993), 49–52.
Buffet, E., Patrick, A. and Pulé, J. V. [Directed polymers on trees: a martingale approach]{}. *J. Phys. A* **26** (1993), 1823–1834.
Chung, K. L. and Erdös, P. [Probability limit theorems assuming only the first moment I]{}. *Mem. Am. Math. Soc.* **6** (1951), paper 3, 1–19. Derrida, B. and Gardner, E. [ Renormalization group study of a disordered model]{}. *J. Phys. A: Math. Gen.* **17** (1984), 3223–3236.
Derrida, B., Giacomin, G., Lacoin, H. and Toninelli, F. L. [Fractional moment bounds and disorder relevance for pinning models]{}. *Commun. Math. Phys.* **287** (2009), 867–887.
Derrida, B., Hakim, V. and Vannimenus, J. [Effect of disorder on two-dimensional wetting]{}. *J. Statist. Phys.* **66** (1992), 1189–1213.
Doney, R.A. [One-sided local large deviation and renewal theorems in the case of infinite mean]{}. *Probab. Theory Rel. Fields* **107** (1997), 451–465.
Feller, W. *An introduction to probability theory and its applications, Vol. II*. Second edition, John Wiley & Sons, 1971.
Fisher, M. E. [ Walks, walls, wetting, and melting]{}. J. Statist. Phys. **34** (1984), 667–729.
Forgacs, G., Lipowsky, R. and Nieuwenhuizen, Th. M. [The behavior of interfaces in ordered and disordered systems]{}. *Phase Transitions and Critical Phenomena* **14**, 135–363, Academic Press, London, 1991.
Forgacs, G., Luck, J. M., Nieuwenhuizen, Th. M. and Orland, H. [Wetting of a disordered substrate: exact critical behavior in two dimensions]{}. *Phys. Rev. Lett.* **57** (1986), 2184–2187.
Gangardt, D. M. and Nechaev, S. K. [ Wetting transition on a one-dimensional disorder]{}. *J. Statist. Phys.* **130** (2008), 483–502.
Giacomin, G. *Random polymer models*. Imperial College Press, World Scientific, 2007.
Giacomin, G., Lacoin, H. and Toninelli, F. L. [Hierarchical pinning models, quadratic maps and quenched disorder]{}. *Probab. Theory Rel. Fields*, in press.
Giacomin, G., Lacoin, H. and Toninelli, F. L. [Disorder relevance at marginality and critical point shift]{}. Peprint, arXiv:0906.1942 \[math-ph\].
Giacomin, G. and Toninelli, F. L. [Smoothing effect of quenched disorder on polymer depinning transitions]{}. *Commun. Math. Phys.* **266** (2006), 1–16.
Giacomin, G. and Toninelli, F. L. [Smoothing of depinning transitions for directed polymers with quenched disorder]{}. *Phys. Rev. Lett.* **96** (2006), 070602.
Grosberg, A. Y. and Shakhnovich, E. I. [An investigation of the configurational statistics of a polymer chain in an external field by the dynamical renormalization group method]{}. *Sov. Phys.-JETP* **64** (1986), 493–501.
Grosberg, A. Y. and Shakhnovich, E. I., [Theory of phase transitions of the coil-globule type in a heteropolymer chain with disordered sequence of links]{}, *Sov. Phys.-JETP*, **64** (1986), 1284–1290.
Harris, A. B. [Effect of random defects on the critical behaviour of Ising models]{}. *J. Phys. C* **7** (1974), 1671–1692.
Harris, T. E. *The theory of branching processes*. Springer-Verlag, Berlin, New York,1963.
Lacoin, H. [Hierarchical pinning model with site disorder: disorder is marginally relevant]{}. *Probab. Theory Rel. Fields*, in press.
Lacoin, H. and Toninelli, F. L. [ A smoothing inequality for hierarchical pinning models]{}. *Proceedings of the Summer School “Spin glasses", Paris, June 2007*, to appear.
Stepanow, S. and Chudnovskiy, A. L. [The Green’s function approach to adsorption of a random heteropolymer onto surfaces]{}. *J. Phys. A: Math. Gen.* **35** (2002), 4229–4238.
Tang, L.-H. and Chaté, H. [Rare-event induced binding transition of heteropolymers]{}. *Phys. Rev. Lett.* **86** (2001), 830–833.
Toninelli, F. L. [A replica-coupling approach to disordered pinning models]{}. *Commun. Math. Phys.* **280** (2008), 389–401.
Toninelli, F. L. [Disordered pinning models and copolymers: beyond annealed bounds]{}. *Ann. Appl. Probab.* **18** (2008), 1569–1587.
Toninelli, F. L. [ Coarse graining, fractional moments and the critical slope of random copolymers]{}. *Electron. Journal Probab.* **14** (2009), 531–547.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We model continuous-time information flows generated by a number of information sources that switch on and off at random times. By modulating a multi-dimensional information process over a random point field, we explicitly relate the discovery of relevant new information sources to jumps in conditional expectation processes of partially observed signals. We show that the underlying measure-valued process follows jump-diffusion dynamics, where the jumps are governed by information source switches. The dynamic representation gives rise to a set of martingales on random time intervals that capture evolving information states as well as to a state-dependent stochastic volatility evolution with jumps. The nature of information flows usually exhibits complex behaviour, however, we succeed in maintaining analytic tractability by introducing what we term the effective and complementary information processes. The effective information captures in a univariate process all information provided by the active sources, and the complementary information is a function-valued stochastic process which is constant between changes of states. One of the many potential applications is in finance: we associate the signal to a random payoff and price a financial derivative within the modulated information approach. It turns out that the price of a vanilla option can be expressed by a weighted sum of option values based on the possible state configurations at option expiry. This result may be viewed as an information-based analogue of Merton’s option price, but where jump-diffusion arises endogenously. The proposed information flow models also lend themselves to the quantification of informational advantage among competitive agents who have asymmetric information flows.'
author:
- |
Edward Hoyle$^1$, Andrea Macrina$^{2,\, 3}$ and Levent Ali Mengütürk$^{2}$\
\
$^{1}$AHL Partners LLP, Man Group plc\
London EC4R 3AD, U.K.\
$^{2}$Department of Mathematics, University College London\
London WC1E 6BT, U.K.\
$^{3}$African Institute of Financial Markets & Risk Management\
University of Cape Town, Rondebosch 7701, R.S.A.
date: 12 December 2018
title: Modulated Information Flows on Random Point Fields
---
[**Keywords:**]{} Filtration models, jump-diffusion dynamics, point processes, stochastic volatility, information-based modeling, asymmetric information.
Introduction
============
Information flows, especially in social systems, usually exhibit complex structures and behaviours. Through time, not only new information channels may suddenly appear, but also some of the existing information sources may cease—e.g. a news agency stops operating or a national statistical agency stops reporting particular economic data series—and some information may be temporarily interrupted over random periods of time. In financial markets, for example, exchanges halt the trading of securities, following large price moves, with *circuit breakers* for stocks and *limit-up* pricing rules for commodity futures. In general, if relevant information arrives gradually and continuously in time, then its impact on an observer’s inference is small over short time periods. However, if important news arrives sporadically or new information sources appear at discrete points in time (e.g. a relevant data vendor becomes available to the market), then it is reasonable to expect that the impact of new information on the observer’s inference is substantial. Such a view was taken in the work of [@19], where we find: [*“By its very nature, important information arrives only at discrete points in time. This component is modelled by a jump process reflecting the non-marginal impact of the information."*]{} Recent examples of the new information source phenomenon in the financial markets are Elon Musk’s tweet about taking Tesla private in 2018, including following announcements and developments[^1], and the fall in the Swatch stock price when the Swiss National Bank stopped defending the EUR-CHF currency floor in January 2015[^2]$^{,}$[^3]. Another incident of a similar nature is the effect on Snapchat’s share price following tweets by Kylie Jenner[^4]. The general situation we keep in mind in this paper is one where observers or agents in a given system, be they financial, security, regulatory, etc., assess a signal on the basis of varying information sets, e.g. various news channels/feeds, accessible to them at any point in time.
We shall give an example in financial markets on price formation, which is a complicated mechanism. Market agents will buy or sell a given asset at a given price if suitably incentivised. For market makers, the incentive may not be to profit from anticipated price changes or income, but instead to earn the bid-ask spread or commissions. Although investors often seek to profit from anticipated price changes and income, they may have other incentives to trade that are not directly linked to these, including risk management (to increase, decrease or hedge risk), an evolving mandate (such as a tightening of environmental, social and governance criteria), a change in risk aversion, and a change in wealth. However, it is commonly understood that the price of an asset should reflect the information that the market has about the asset. Consider the situation of a government considering the bailout of a bank or insurance company (the firm). The market may be aware of the firm being in financial difficulty, and this would then be reflected in the prices of the firm’s stock and bonds. The market may also anticipate a government bailout. However, due to insider trading legislation, any bailout negotiations between the government and the firm may be held behind closed doors. The market cannot be sure of the bailout until an official government announcement. We can view government’s communication of a bailout with the market as new source of information. The appearance of this new source may have certain stylized effects on prices. One such effect is an initial price correction as the market converts a bailout possibility into a certainty; a significant—and immediate—adjustment in the price. The correction is the market absorbing the totality of the private bailout negotiations in one gulp. Salient details for asset prices would include the bailout’s anticipated size and duration. A second effect is a series of small price adjustments as the government regularly reports on the progress of its bailout. This can be expected to continue until either the firm recovers and no longer requires government assistance (loans the government has made the firm are repaid, and equity the government has bought in the firm is sold), the firm is wholly taken over by the government, or the firm is allowed to fail. At this point the new source of market information on the firm ceases.
As a mathematical model, we develop a stochastic approach for information flows that can encapsulate these scenarios dynamically by allowing information sources to switch on and off at random times. We show that intelligence, produced on the basis of such information flows, features explicit dependence on the number and the specific information channels that are active or inactive over the course of the inference period. While the approach we develop can be applied more broadly in signal processing and related fields, our main aim is to systematically and explicitly relate randomly evolving information states to price dynamics. The mechanism whereby the jumps are induced is a result of how we construct the filtration with respect to which discounted conditional expectations are taken to compute asset prices. This is markedly different from conjecturing a continuous-time price process which is explicitly governed by jump-diffusion dynamics. Thus, a welcome consequence of our approach is an information-based [*endogenous*]{} jump-diffusion model, where such dynamics arise naturally from the proposed information system. In this sense, we provide an extension of the information-based framework, see [@4; @5; @6; @7; @16; @21], which take advantage of special properties of Brownian bridges, see [@12] and [@13]. In the same spirit, in order to explicitly model the information flow, we introduce a multivariate information process where each marginal is the sum of a random signal $X$, (e.g. a future payoff), and an independent Brownian bridge noise. To allow for random activation and deactivation of information sources, a random point field acts on the coordinates of the information process as a modulator of the information flow.
We shall highlight that we do not introduce jumps into the information flow model by embedding discontinuous noise into information processes. Instead, the jumps arise due to the discovery of *new* information processes with a *continuous* state-space. This has different and rather important implications on the dynamics of the best estimate of a random signal as opposed to including independent discontinuous noise with no information content on the signal. First, jumps are caused by random changes in the number of active information sources; hence, jumps carry information about $X$. Second, the continuous part of the price process is driven by different martingales on random time intervals representing the possible state-configurations of the information flow. Third, since the volatility process also jumps and is state-dependent, the framework offers a link to regime-switching models; see [@8; @14; @20]. Fourth, the undiscounted price process may remain constant for periods of time if all information sources are “lost”, which perhaps could be viewed as a model for certain features arising in illiquid markets or circuit breakers. Fifth, the proposed framework offers links to progressive enlargements of filtrations and stopped filtrations.
The paper structure is as follows. In Section 2, we introduce randomly modulated information flows and a multivariate time-changed information process. We construct the effective and complementary information processes, which provide valuable analytical tractability. The effective process is a one-dimensional aggregation of all active information sources and the complementary process reflects the information supplied by now inactive sources. We also derive the dynamics of the endogenous jump-diffusion process, provide a Feynman-Kać representation of the conditional expectation, and give an alternative expression for its jump-sizes. In Section 3, we price a vanilla option under the modulated information system, and highlight the similarity of its functional form with that obtained by [@19]. Our final application addresses information asymmetry between agents whose access to information sources differ. We here allude to a randomly evolving competition between agents, where some may have an informational advantage over a period of time and suddenly find that the wheel has turned with them having to come to terms with several of their information sources being disabled.
Modulation of information sources {#sec:framework}
=================================
Let $(\Omega, \mathcal{G}, (\mathcal{G}_t)_{t\geq0},\P)$ be a probability space equipped with a filtration $(\mathcal{G}_t)$. All considered filtrations are assumed to be right-continuous and complete. We introduce a random variable $X\in\mathcal{L}^2(\Omega,\mathcal{G},\P)$ with law $\nu$ and state-space $(\mathbb{X},\mathcal{B}(\mathbb{X}))$, where we assume $\mathbb{X}\subseteq\mathbb{R}$. We consider the time interval $t\in\mathbb{T}=[0,1]$, though it is straightforward to consider an arbitrary closed set $[0,T]$ for $T<\infty$, instead. Next we introduce a $(\mathcal{G}_t)$-adapted multivariate stochastic process $(\boldsymbol{\xi}_t)_{t\in\mathbb{T}}$, a so-called Brownian bridge information process, defined by $ \xi^{(i)}_t = t X + \s_i \b_t^{(i)}$, for $\s_i>0$ and $i=1,\ldots,n$, which takes values in $\R^n$. The process $(\beta_t^{(i)})_{0\le t\le 1}$ is a standard Brownian bridge. That is, it is a Gaussian process with mean function identically zero, and covariance function $(s,t)\mapsto \min[s,t] - st$, for $s,t\in\mathbb{T}$. We assume that $X$ and the Brownian bridges $\{(\b_t^{(i)})\}_{i=1,\ldots,n}$, are mutually independent. For $t\in(0,1)$, the random variable $\xi_t^{(i)}/t$ is equal to $X$ plus some independent Gaussian noise with variance $\s_i^2(1-t)/t$, and $\xi_1^{(i)}= X$, $i=1,\ldots,n$. If there were linear dependence (with non-singular covariation) between the Brownian bridges, then there would exist a linear transformation of $(\boldsymbol{\xi}_t)$ that would fit the proposed framework. Hence, allowing linear dependence between $\{(\b_t^{(i)})\}_{i=1,\ldots,n}$ does not enrich the model. For the rest of this paper, we let the mapping $h:\R\times\R\times[0,1)\times\R_+ \rightarrow \R_+$ be the Gaussian function given by $ h(x, y, t, \s) = \exp\left\{(x y-t x^2/2) / (\s^2(1-t))\right\}.$
Next we introduce a mechanism for the activation and deactivation of information sources. Let $\zeta$ be a random point field (a point process) on $\mathbb{T}^n$, independent of $(\boldsymbol{\xi}_t)$, generating a collection of times $\{\varpi^{i}_{1}, \ldots,\varpi^{i}_{k_i}\}$ for $i\in\{1,\ldots,n\}$ and finite $k_i\in\mathbb{N}_0$. For every $i$, we associate the random sequence $\{\varpi^{i}_{1}, \ldots,\varpi^{i}_{k_i}\}$ to a coordinate of a $(\mathcal{G}_t)$-adapted càdlàg jump process $(\boldsymbol{\tilde{J}}_t)$ with state space $\S=\{0,1\}^n$. We then define the $\R^{n\times n}$-valued process $(\boldsymbol{J}_t)$ by $J_t^{(i,j)} = \delta_{ij}\boldsymbol{\tilde{J}}_t$. Thus, $(\boldsymbol{J}_t)$ is a diagonal matrix-valued process indicating which coordinates of $(\boldsymbol{\xi}_t)$ are active through the modulated information process $(\boldsymbol{J_t}\,\boldsymbol{\xi}_t)$. When the $i$th information source is inactive, the $i$th coordinate of $(\boldsymbol{J}_t\,\boldsymbol{\xi}_t)$ is identically zero. We define a sub-algebra $\F_t\subseteq \mathcal{G}_t$ by $$\label{subF}
\F_t = \s\left( (\boldsymbol{J}_u\,\boldsymbol{\xi}_u)_{0\leq u \leq t}, (\boldsymbol{J}_u)_{0\leq u \leq t}, X\1_{\{t=1\}} \right), \nonumber$$ for $t\in\mathbb{T}$. We note that the switching process $(\boldsymbol{J}_t)$ is adapted to the filtration generated by $(\boldsymbol{J}_t\,\boldsymbol{\xi}_t)$, and so its appearance in the definition of $\F_t$ is superfluous. However, we might wish to generalise the model so that the coordinate $(\xi^{(i)}_t)$ is a jump process or such that $0< \P[\xi_t^{(i)}=0]<1$, for some $i\in\{1,\ldots, n\}$ and $t\in\mathbb{T}$. In these cases, it may not be possible to detect the (de)activation of a particular information coordinate if $(\boldsymbol{J}_t)$ is not $(\F_t)$-adapted. We also note that we add $X\1_{\{t=1\}}$ to the filtration to ensure that its value is revealed at $t=1$, even though all information sources may be inactive at $t=1$. This is not a mathematical requirement; if an envisaged application does not need $X$ to be (fully) observed at $t=1$, then $X\1_{\{t=1\}}$ may be excluded from the algebra $\mathcal{F}_t$. In the remainder of this paper, we shall omit it from the expressions for notational convenience unless necessary.
At time $t\in\mathbb{T}$, we denote the last time that the $i$th information source was active by $\t^{(i)}_t$. That is, we define $$\t^{(i)}_t = 0 \vee \sup\{ u : J_u^{(i,i)} = 1, u\in[0,t]\},$$ for $i=1,\ldots,n$, where we adopt the convention that $\sup \emptyset = -\infty$. Thus, the process $(\t^{(i)}_t)$ is increasing and progressively measurable with respect to $(\F_t)$, and if the $i$th process has never been active up until time $t$, then $\t^{(i)}_t=0$. Also, by definition, the initial condition is $\t^{(i)}_0=0$.
\[rem1\] The dynamics of $(\t^{(i)}_t)$ for $t\in\mathbb{T}$ are given by $$\dd \t^{(i)}_t = J_t^{(i,i)}\dd t + (t-\t^{(i)}_{t-}) \dd J^{(i,i)}_t.$$
See Appendix for the proof.
We now turn our attention to the processing of information about $X$, for which, we shall determine the conditional distribution of $X$ given $\F_t$.
\[lemmaMarkov\] Let $\xi^{(i)}(u)$ be the value of $(\xi^{(i)}_t)$ at $u\in\mathbb{T}$.
1. For any $A\in\mathcal{B}(\mathbb{X})$ and $t_i \in \mathbb{T}$, $$\P\left[X\in A \left|\, \{(\xi^{(i)}_s)_{0\le s\leq t_i}\}_{i=1,\ldots,n} \right.\right]
= \P\left[X\in A \left|\, \{\xi^{(i)}(t_i)\}_{i=1,\ldots,n} \right.\right]$$
2. For any $t\in\mathbb{T}$, the sigma-algebra $\F_t$ is equivalent to $$\F_t = \s\left( \{(\xi^{(i)}(\t^{(i)}_u))_{0\leq u \leq t}\}_{i=1,\ldots,n}, (\boldsymbol{J}_u)_{0\leq u \leq t} ,X\1_{\{t=1\}}\right)$$
For $t<1$, we follow the lead of the proof of Proposition 3.4 in [@15]. It is sufficient to show that $$\begin{aligned}
\P\left[ X \in \d x_0 \left|\, \xi^{(1)}_{t_{1,1}} = x_{1,1}, \ldots, \xi^{(1)}_{t_{1,k_1}} = x_{1,k_1}, \ldots, \xi^{(n)}_{t_{n,1}} = x_{n,1}, \ldots, \xi^{(n)}_{t_{n,k_n}} = x_{n,k_n}\right.\right]&
\\=\P\left[ X \in \d x_0 \left|\, \xi^{(1)}_{t_{1,k_1}} = x_{1,k_1}, \ldots, \xi^{(n)}_{t_{n,k_n}} = x_{n,k_n}\right.\right]&,\end{aligned}$$ for all $k_i \in \N_+$ where $i\in\{1,\ldots,n\}$, all $0 < t_{i,1} < \cdots < t_{i,k_i} < 1$, and all $(x_{i,1}, \ldots,x_{i,k_i})\in \mathbb{R}^n$. We have $$\begin{aligned}
&\P\left[ X \in \d x_0 \left|\, \xi^{(1)}_{t_{1,1}} = x_{1,1}, \ldots, \xi^{(1)}_{t_{1,k_1}} = x_{1,k_1}, \ldots, \xi^{(n)}_{t_{n,1}} = x_{n,1}, \ldots, \xi^{(n)}_{t_{n,k_n}} = x_{n,k_n}\right.\right]
\\
&=\frac{\P\left[ \left. \bigcap_{i=1}^{n} \bigcap_{j=1}^{k_i} \xi^{(i)}_{t_{i,j}} \in\dd x_{i,j} \,\right| X = x_0\right] \P[X\in \dd x_0]}
{\int_{\mathbb{X}} \P\left[ \left. \bigcap_{i=1}^{n} \bigcap_{j=1}^{k_i} \xi^{(i)}_{t_{i,j}} \in\dd x_{i,j} \,\right| X = x_0\right] \P[X\in \dd x_0]}&
\\
&=\frac{\prod_{i=1}^{n} \P\left[ \left. \bigcap_{j=1}^{k_i} \xi^{(i)}_{t_{i,j}} \in\dd x_{i,j} \,\right| X = x_0\right] \P[X\in \dd x_0]}
{\int_{\mathbb{X}} \prod_{i=1}^{n} \P\left[ \left. \bigcap_{j=1}^{k_i} \xi^{(i)}_{t_{i,j}} \in\dd x_{i,j} \,\right| X = x_0\right] \P[X\in \dd x_0]}.&\end{aligned}$$ Given $X$, each coordinate process $(\xi^{(i)}_t)$ is a drifting Brownian bridge. It follows that $$\P\left[ \left. \bigcap_{j=1}^{k_i} \xi^{(i)}_{t_{i,j}} \in\dd x_{i,j} \,\right| X = x_0\right]
= \frac{f_{1-t_{i,k_i}}(x_0-x_{i,k_i})}{f_1(x_0)} \prod_{j=1}^{k_i} f_{t_{i,j}-t_{i,j-1}}(x_{i,j}-x_{i,j-1})\dd x_{i,j},$$ where $f_t(x)$ is the marginal density function of standard Brownian motion. Hence, we have $$\begin{aligned}
&\P\left[ X \in \d x_0 \left|\, \xi^{(1)}_{t_{1,1}} = x_{1,1}, \ldots, \xi^{(1)}_{t_{1,k_1}} = x_{1,k_1}, \ldots, \xi^{(n)}_{t_{n,1}} = x_{n,1}, \ldots, \xi^{(n)}_{t_{n,k_n}} = x_{n,k_n}\right.\right]
\\
&\hspace{6cm}=\frac{\prod_{i=1}^n \frac{f_{1-t_{i,k_i}}(x_0-x_{i,k_i})}{f_1(x_0)}f_{t_{i,k_i}}(x_{i,k_i}) \, \nu(\dd x_0)}
{\int_{\mathbb{X}} \prod_{i=1}^n \frac{f_{1-t_{i,k_i}}(x_0-x_{i,k_i})}{f_1(x_0)}f_{t_{i,k_i}}(x_{i,k_i}) \, \nu(\dd x_0)}&
\\
&\hspace{6cm}=\frac{\prod_{i=1}^n \P\left[ \left. \xi^{(i)}_{t_{i,k_i}} \in\dd x_{i,k_i} \,\right| X = x_0\right] \P[X\in \dd x_0]}
{\int_{\mathbb{X}} \prod_{i=1}^n \P\left[ \left. \xi^{(i)}_{t_{i,k_i}} \in\dd x_{i,k_i} \,\right| X = x_0\right] \P[X\in \dd x_0]}&
\\
&\hspace{6cm}=\frac{\P\left[ \left. \bigcap_{i=1}^{n} \xi^{(i)}_{t_{i,k_i}} \in\dd x_{i,k_i} \,\right| X = x_0\right] \P[X\in \dd x_0]}
{\int_{\mathbb{X}} \P\left[ \left. \bigcap_{i=1}^{n}\xi^{(i)}_{t_{i,k_i}} \in\dd x_{i,k_i} \,\right| X = x_0\right] \P[X\in \dd x_0]}
\\
&\hspace{6cm}=\P\left[ X \in \d x \left|\, \xi^{(1)}_{t_{1,k_1}} = x_{1,k_1}, \ldots, \xi^{(n)}_{t_{n,k_n}} = x_{n,k_n}\right.\right],&
\end{aligned}$$ as required. The result is trivial for $t=1$ due to the bridge property. The second part of the lemma follows from the fact that the coordinates of the modulated process $(J_t^{(i,i)} \xi^{(i)}_t)$ and the time-changed process $(\xi^{(i)}(\t^{(i)}_t))$ differ only when the $i$th source is inactive; during these times the coordinate process is zero, and $(\xi^{(i)}(\t^{(i)}_t))$ takes the source’s last active value.
\[conditionaldist\] The $\F_t$-conditional distribution of $X$ is given by $$\begin{aligned}
\P[X \in \dd x \,|\, \F_t]
&= \frac{\prod_i h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)}
{\int_{\mathbb{X}} \prod_i h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)},
\end{aligned}$$ for $t \in [0,1)$.
First we note that, for $t_i\in[0,1)$ and $i=1,\ldots,n$, we have $$\begin{aligned}
\label{Bayes}
\P\left[X\in \dd x \left|\, \left\{\xi^{(i)}_{t_i}\right\}_{i=1,\ldots, n} \right.\right]
&= \frac{\prod^n_{i=1} \exp\left[-\frac{1}{2} \left( \frac{\xi^{(i)}_{t_i} - t_i x}{\s_i\sqrt{t_i(1-t_i)}} \right)^2 \right] \nu(\dd x)}
{\int_{\mathbb{X}} \prod^n_{i=1} \exp\left[-\frac{1}{2} \left( \frac{\xi^{(i)}_{t_i} - t_i x}{\s_i\sqrt{t_i(1-t_i)}} \right)^2 \right] \nu(\dd x)}\nonumber
\\ &= \frac{\prod^n_{i=1} \exp\left( \frac{x \xi^{(i)}_{t_i}-t_i x^2/2}{\s_i^2 (1-t_i)}\right) \nu(\dd x)}
{\int_{\mathbb{X}} \prod^n_{i=1} \exp\left(\frac{ x \xi^{(i)}_{t_i}-t_i x^2/2}{\s_i^2 (1-t_i)}\right) \nu(\dd x)}\nonumber
\\ &= \frac{\prod^n_{i=1} h\left(x, \xi^{(i)}_{t_i}, t_i, \s_i\right) \nu(\dd x)}{\int_{\mathbb{X}} \prod^n_{i=1} h\left(x, \xi^{(i)}_{t_i}, t_i, \s_i\right) \nu(\dd x)},
\end{aligned}$$ by the Bayes formula. For the computation of the conditional distribution $\P[X \in \dd x \,|\, \F_t]$, the first step is to use $\{\xi^{(i)}(\t^{(i)}_t)\}_{i=1,\ldots, n}$ to enlarge the information set we are conditioning on, and then to apply the tower property. We here refer to Lemma \[lemmaMarkov\]. $$\P[X \in \dd x \,|\, \F_t]
= \E \left[\left. \P\left[ X\in \dd x \left\vert\, \left\{(\xi^{(i)}_s)_{0\le s\leq \t^{(i)}_t}\right\}_{i=1,\ldots, n},\, \F_t \right.\right] \,\right| \F_t \right].$$ The $\sigma$-algebra $\F_t$ contains the history $(\boldsymbol{J}_s)_{0\le s\le t}$, which tells up to what time $t_i$ the information coordinates $\{\xi^{(i)}(\t^{(i)}_s)_{0\le s\le t_i}\}_{i=1,\dots,n}$ have been active. Thus once the stopping times $\{\t^{(i)}_t\}_{i=1,\dots,n}$ have occurred, one knows that $(\xi^{(i)}(\t^{(i)}_t))=(\xi^{(i)}_s)_{0\le s\le t_i}$ for $\tau^{(i)}_t=t_i\le t$. Therefore we may apply Lemma \[lemmaMarkov\] to obtain $$\E \left[\left. \P\left[ X\in \dd x \left\vert\, \left\{(\xi^{(i)}_s)_{0\le s\leq \t^{(i)}_t}\right\}_{i=1,\ldots, n},\, \F_t \right.\right] \,\right| \F_t \right]= \E \left[\left. \P\left[ X\in \dd x \left|\, \left\{\xi^{(i)}(\t^{(i)}_t) \right\}_{i=1,\ldots, n},\, \F_t \right.\right] \,\right| \F_t \right].$$ By use of Equation (\[Bayes\]) and $\F_t$-measurability, it follows that $$\begin{aligned}
\P[X \in \dd x \,|\, \F_t]&= \E \left[\left. \P\left[ X\in \dd x \left|\, \left\{\xi^{(i)}(\t^{(i)}_t) \right\}_{i=1,\ldots, n},\, \F_t \right.\right] \,\right| \F_t \right]
\\ &= \frac{\prod^n_{i=1} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)}
{\int_{\mathbb{X}} \prod^n_{i=1} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)},
\end{aligned}$$ for $t_i\in[0,1)$ and $i=1,\ldots,n$, which gives the statement.
Effective and complementary information
---------------------------------------
We shall now introduce the so-called *effective information process* and the *complementary information process*, which provide us with valuable analytical tractability. In particular, this useful parametrisation of the information flow allows one to reduce the multi-dimensional information system—corresponding to the number of active information sources—to a one-dimensional (effective) information process containing all information provided by the active sources. Although simple, this idea reduces the complexity of the information flow model substantially without compromising its effectiveness or reducing its richness.
Let the set-valued stochastic process $(\mathcal{J}_t)$ be given by $\mathcal{J}_t = \{i : J_t^{(i,i)} = 1\}$. For $\mathcal{J}_t$ non-empty, the effective information process $(\hat{\xi_t})_{t\in\mathbb{T}}$ is defined by $$\hat{\xi_t}=\hat{\s_{t}}^2 \sum_{i \in \mathcal{J}_t} \s^{-2}_i\xi^{(i)}_t$$ taking values in $\R$, with effective volatility parameter $ \hat{\s_{t}} = \left(\sum_{i \in \mathcal{J}_t} \s^{-2}_i\right)^{-1/2}$. For $\mathcal{J}_t = \emptyset$, set $\hat{\xi}_t = 0$ and $\hat{\s_{t}}=0$.
Note that adding elements to $\mathcal{J}_t$ decreases the size of $\hat{\s_{t}}$. The interpretation is that adding sources of information decreases the effective noise in the system. The dynamics of the effective information process are dependent on the current state $\boldsymbol{\tilde{J}}_t\in\S$. We can make this dependence more explicit by rewriting the effective information as follows: $$\hat{\xi}_t =
\left\{
\begin{aligned} &0 && \text{if $\boldsymbol{J}_t=\mathbf{0}$},
\\ &\frac{\boldsymbol{\xi}_t^{\tp} \boldsymbol{J}_t \, \boldsymbol{\rho}}{ \mathbf{1}^{\tp} \boldsymbol{J}_t \, \boldsymbol{\rho}} && \text{otherwise,}
\end{aligned}
\right.$$ where $\boldsymbol{\rho} =(\s_1^{-2},\ldots,\s_n^{-2})^{\tp}$ and $\mathbf{1}=(1,\ldots,1)^{\tp}$.
\[lemmaeffectiveinfo\] If $\mathcal{J}_t\neq \emptyset$, then the effective information process is given by $$\hat{\xi}_t = tX + \hat{\s_{t}} \hat{\b}_t,$$ where $(\hat{\b}_t)_{t\in\mathbb{T}}$, defined by $ \hat{\b}_t = \hat{\s_{t}} \sum_{i \in \mathcal{J}_t} \s^{-1}_i\b^{(i)}_t$, is a standard Brownian bridge between jumps.
See Appendix for the proof.
The effective information process jumps every time $(\boldsymbol{\tilde{J}}_t)$ changes state. The jump is caused by the change in the number of Brownian bridges defining the effective information process as well as the number of volatility control parameters defining the effective volatility process $(\hat{\s_{t}})$.
Let the set-valued complementary process $(\mathcal{J}_t^\complement)$ be given by $\mathcal{J}_t^\complement = \{i : J_t^{(i,i)} = 0\}$. The complementary information process $(\eta_t)_{t\in\mathbb{T}}$ is a function-valued process defined by $$\eta_t: x\mapsto \prod_{i\in\mathcal{J}_t^\complement} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right),$$ where $\eta_t:=1$ if $\mathcal{J}_t^\complement = \emptyset$.
The complementary information process is piecewise constant between state changes of $(\boldsymbol{\tilde{J}}_t)$. The next statement on the joint Markov property of effective and complementary information processes will be very useful for the rest of this work.
\[effectivecomplementary\] The measure $\nu_t(\dd x)=\P[X \in \dd x \,|\, \F_t]$ satisfies $$\begin{aligned}
\nu_t(\dd x) = \frac{h\left(x, \hat{\xi}_t, t, \hat{\s_{t}} \right) \eta_t(x) \, \nu(\dd x)}
{\int_{\mathbb{X}} h\left(x, \hat{\xi}_t, t, \hat{\s_{t}} \right) \eta_t(x) \,\nu(\dd x)},
\end{aligned}$$ and hence, $\nu_t(\dd x) = \P[X \in \dd x \,|\, \hat{\xi}_t, \eta_t]$ for $t \in [0,1)$.
Using Proposition \[conditionaldist\] and noting that $\t^{(i)}_t = t$ if $i\in\mathcal{J}_t$, we have $$\begin{aligned}
\P[X \in \dd x \,|\, \F_t]=& \frac{\prod_{i\in\mathcal{J}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i\in\mathcal{J}_t^\complement} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)}
{\int_{\mathbb{X}} \prod_{i\in\mathcal{J}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i\in\mathcal{J}_t^\complement} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right) \nu(\dd x)}
\\ =& \frac{\prod_{i\in\mathcal{J}_t} \exp\left\{\frac{x \xi^{(i)}(\t^{(i)}_t)-\t^{(i)}_t x^2/2}{\s_i^2(1-\t^{(i)}_t)}\right\}
\eta_t(x) \, \nu(\dd x)}
{\int_{\mathbb{X}} \prod_{i\in\mathcal{J}_t} \exp\left\{\frac{x \xi^{(i)}(\t^{(i)}_t)-\t^{(i)}_t x^2/2}{\s_i^2(1-\t^{(i)}_t)}\right\}
\eta_t(x) \, \nu(\dd x)},
\\ =& \frac{\exp\left\{\sum_{i\in\mathcal{J}_t}\frac{x \xi^{(i)}_t-t x^2/2}{\s_i^2(1-t)}\right\}
\eta_t(x) \, \nu(\dd x)}
{\int_{\mathbb{X}} \exp\left\{\sum_{i\in\mathcal{J}_t}\frac{x \xi^{(i)}_t-t x^2/2}{\s_i^2(1-t)}\right\}
\eta_t(x) \, \nu(\dd x)},
\end{aligned}$$ where we adopt the convention that a product with an empty index set is equal to one.
Endogenous jump-diffusion {#EJD}
-------------------------
Besides proposing stochastic information flow models, another aim of this paper is to derive a dynamical equation for the conditional expectation process, which we later use to model asset price dynamics, as an example. The dynamics of the conditional expectation process turns out to be jump-diffusion. The jumps arise from the activation of new sources of information generating the filtration used to compute the conditional expectation, which is unlike introducing a jump process (e.g. a Poisson process) in the dynamics in an [*ad hoc*]{} manner from the outset. To understand the evolution of the measure-valued process $(\nu_t)$, we first introduce two $(\F_t)$-adapted counting processes $(C_t)_{t\in\mathbb{T}}$ and $(N_t)_{t\in\mathbb{T}}$, given by $$\begin{aligned}
&C_t = \sum_{s\leq t} \1\{\boldsymbol{\tilde{J}}_{s}\ne\boldsymbol{\tilde{J}}_{s-}\},
&N_t = \sum_{s\leq t} \delta_s, \hspace{0.1in} \text{where} \hspace{0.1in} \delta_s = \1\{\mathcal{J}_s \backslash \mathcal{J}_{s-} \ne \emptyset \}.\end{aligned}$$ Hence, $C_t$ is the number of times $(\boldsymbol{\tilde{J}}_t)$ has changed state up to and including time $t$, and $N_t$ is the number of state changes in which at least one inactive information process becomes active. In view of Proposition \[measureSDE\], we define the following.
Let $\pi_j = \inf\{t:C_t=j\}$, with $\pi_0=0$. The process $(M_t)_{t\in\mathbb{T}}$ is defined by $$M_t= \sum_{j=1}^{C_t+1} M_t^{(j)} + \int_0^t (\hat{\xi}_s-\hat{\xi}_{s-}) \d N_s,$$ where $(M_t^{(j)})_{t\in\mathbb{T}}$ is given by $$M_t^{(j)} =
\left\{ \begin{aligned}
&0, && \text{for $t < \pi_{j-1}$,}
\\ &(\hat{\xi}_{t\wedge \pi_{j}-}-\hat{\xi}_{\pi_{j-1}}) - \int_{\pi_{j-1}}^{t\wedge \pi_{j}-}\frac{\E[ X \,|\, \hat{\xi}_s, \eta_s ] - \hat{\xi}_s}{ 1-s} \d s, &&\text{otherwise.}
\end{aligned} \right.$$
The following statement is important in order to assign a sufficient structure for the dynamics of the continuous part of $(M_t)$, so that the stochastic integrals with respect to the processes $\{(M_t^{(i)})\}_{i=1,\ldots,n}$ are well-defined for the subsequent proposition.
\[piecewisemartingale\] Let $t\in[\pi_{j-1},\pi_{j})$, where $\pi_{j-1}$ and $\pi_{j}$ are two arbitrarily chosen consecutive jump times of $(C_t)$. Then, $(M_t)_{\pi_{j-1} \leq t < \pi_{j}}$ is an $(\F_t)$-martingale.
The integrability condition $\E[\,|M_{t}|\,]<\infty$ for $t\in\mathbb{T}$ is satisfied. Next we show $\E[M_u\,\vert\,\F_t]=M_t$ for $u\geq t$ by following similar steps to [@4] while considering the random interval $t\in[\pi_{j-1},\pi_j)$ over which $(M_t)$ has no discontinuity. First, for $u\in[\pi_{j-1},\pi_j)$, we note that $$\begin{aligned}
\E[M_{u} \,|\, \F_t]&= \E[M_{u} - M_{t} \,|\, \F_t] + M_{t} \nonumber \\
&=M_{t}+\E\left[\left. \hat{\xi}_u-\hat{\xi}_t \,\right| \hat{\xi}_t, \eta_t \right] +\E\left[\left.\int^{u}_{t}\frac{\hat{\xi}_s}{1-s} \dd s \,\right| \hat{\xi}_t, \eta_t \right] -\E\left[\left. \int^{u}_{t}\frac{\E[X \,|\, \hat{\xi}_s, \eta_s]}{1-s} \dd s \,\right| \hat{\xi}_t, \eta_t \right]. \nonumber
\end{aligned}$$ Writing the terms explicitly and using the tower property, we have $$\begin{aligned}
\E[M_{u} \,|\, \F_t] = M_{t} &+ \E\left[\left. Xu+\hat{\s}_{u}\hat{\b}_u \,\right| \hat{\xi}_t, \eta_t \right]-\E\left[\left.Xt+\hat{\s}_{t}\hat{\b}_t \,\right| \hat{\xi}_t, \eta_t \right] \nonumber \\
&+\E\left[X \left|\, \hat{\xi}_t, \eta_t \right.\right]\int^{u}_{t}\frac{s}{1-s}\dd s +\E\left[\left. \int^{u}_{t}\frac{\hat{\s}_{s}\hat{\b}_s}{1-s}\dd s \,\right| \hat{\xi}_t, \eta_t \right] \nonumber \\
&-\E\left[X \left|\, \hat{\xi}_t, \eta_t \right.\right]\int^{u}_{t}\frac{1}{1-s}\dd s. \nonumber
\end{aligned}$$ Note that all the terms involving $X$ disappear, and we are left with $$\begin{aligned}
\E[M_{u} \,|\, \F_t] &= M_{t}+\E\left[\left. \hat{\s}_{u}\hat{\b}_u \,\right| \hat{\xi}_t, \eta_t \right]
-\E\left[\left. \hat{\s}_{t}\hat{\b}_t \,\right| \hat{\xi}_t, \eta_t \right]
+\int^{u}_{t}\frac{\E[\hat{\s}_{s}\hat{\b}_s \,|\, \hat{\xi}_t, \eta_t]}{1-s}\dd s \nonumber \\
&= M_{t} + \hat{\s_{t}}\left(\E\left[\left. \hat{\b}_u \,\right| \hat{\xi}_t, \eta_t \right]
-\E\left[\left. \hat{\b}_t \,\right| \hat{\xi}_t, \eta_t \right] +\int^{u}_{t}\frac{\E[\hat{\b}_s \,|\, \hat{\xi}_t, \eta_t]}{1-s}\dd s\right), \nonumber
\end{aligned}$$ where we used the fact that $(\hat{\s_{s}})_{t\leq s \leq u}=\hat{\s_{t}}$, since we reside in $[\pi_{j-1},\pi_j)$, and hence, $\hat{\s_t}$ remains constant. By recalling the mutual independence between $X$ and all the $\{(\b^{(i)}_t)\}_{i=1\ldots,n}$, and the tower property, we may write the following: $$\begin{aligned}
\E\left[\left. \hat{\b}_u \,\right| \hat{\xi}_t, \eta_t \right]=\E\left[\left. \E\left[\left. \hat{\b}_u \,\right| X,\hat{\b}_t,\eta_t\right] \,\right| \hat{\xi}_t, \eta_t \right]=\E\left[\left. \E\left[\left. \hat{\b}_u \,\right|\hat{\b}_t \right] \,\right| \hat{\xi}_t, \eta_t\right]=\frac{1-u}{1-t}\,\E\left[\left. \hat{\b}_t \,\right| \hat{\xi}_t, \eta_t\right].\nonumber
\end{aligned}$$ With the above expression at hand, we have $$\begin{aligned}
\E\left[\left. \hat{\b}_u \,\right| \hat{\xi}_t, \eta_t \right]
-\E\left[\left. \hat{\b}_t \,\right| \hat{\xi}_t, \eta_t \right]
+\int^{u}_{t}\frac{\E[\hat{\b}_s \,|\, \hat{\xi}_t, \eta_t]}{1-s}\dd s=0, \nonumber
\end{aligned}$$ which proves $\E[M_{u} \,|\, \F_t]= M_{t}$ for $(t,u)\in[\pi_{j-1},\pi_j)$.
\[measureSDE\] The measure-valued process $(\nu_t)$ satisfies $$\begin{aligned}
\nu_t(A) &= \nu(A) + \sum_{j=1}^{C_t}\int_{\pi_{j-1}}^{\pi_j}\int_A \frac{x-\E[X\,|\, \hat{\xi}_s, \eta_s]}{\hat{\s_{s}}^2 (1-s)} \nu_s(\dd x) \d M_s^{(j)} \\
&+\int_{\pi_{C_t}}^{t}\int_A\frac{x-\E[X\,|\, \hat{\xi}_s, \eta_s]}{\hat{\s_{s}}^2 (1-s)} \nu_s(\dd x) \d M_s^{(C_t+1)}
+ \sum_{s\leq t} (\nu_s(A) - \nu_{s-}(A))\delta_s,\end{aligned}$$ for $t \in [0,1)$ and for any $A\in\mathcal{B}(\mathbb{X})$.
See Appendix for the proof.
We emphasise that the dynamics of $(M_t)_{t\in\mathbb{T}}$ are state-dependent, adapting to the information configuration of the system at any given time. The measure-valued process $(\nu_t)$ evolves continuously when there is no information switch or when an active source becomes inactive. The dynamics exhibit jumps from one state-dependent martingale to another only when an information source becomes active. The process stays constant if all information processes are inactive over a random period of time.
\[ExpectationSDE\] Let $(W_t)$ be an $(\F_t)$-adapted process defined by $W_t=M_t/\hat{\s}_t$.
1. Let $t\in[\pi_{j-1},\pi_{j})$, where $\pi_{j-1}$ and $\pi_{j}$ are two arbitrarily chosen consecutive jump times of $(C_t)$. Then $(W_t)$ is an $(\F_t)$-Brownian motion.
2. Let $(W_t^{(j)})$ be the Brownian motion given above for $t\in[\pi_{j-1}, \pi_j \wedge 1)$. Then, the process $X_t:=\mathbb{E}[X \,|\, \F_t]$ for $t\in[0,1)$, satisfies $$\begin{aligned}
X_t &= \E[X] + \sum_{j=1}^{C_t}\int_{\pi_{j-1}}^{\pi_j} \frac{\Gamma_s}{\hat{\s_{s}} (1-s)} \d W_s^{(j)} +\int_{\pi_{C_t}}^{t} \frac{\Gamma_s}{\hat{\s_{s}} (1-s)} \d W_s^{(C_t+1)} + \sum_{s\leq t} (X_s - X_{s-})\delta_s.\end{aligned}$$ where $(\Gamma_t)_{t\in\mathbb{T}}$ given by $\Gamma_t =\var[X\,|\, \hat{\xi}_t, \eta_t]$ is an $(\F_t)$-supermartingale.
3. Let $X\in\mathcal{L}^k(\Omega,\mathcal{G},\P)$ for $k\geq2$. Then, the process $X_t^{(k)}:=\mathbb{E}[X^k \,|\, \F_t]$ for $t\in[0,1)$, satisfies $$\begin{aligned}
X_t^{(k)} &= \E[X^k] + \sum_{j=1}^{C_t}\int_{\pi_{j-1}}^{\pi_j} \frac{X_s^{(k+1)} - X_s^{(k)}X_s}{\hat{\s_{s}} (1-s)} \d W_s^{(j)} \notag \\
&+\int_{\pi_{C_t}}^{t} \frac{X_s^{(k+1)} - X_s^{(k)}X_s}{\hat{\s_{s}} (1-s)} \d W_s^{(C_t+1)} + \sum_{s\leq t} (X^{(k)}_s - X^{(k)}_{s-})\delta_s.\end{aligned}$$
See Appendix for the proof.
The diffusion coefficient of $(X_t)$ can be interpreted as a stochastic volatility process with jumps. This process arises naturally from the information flow system, which is a welcome consequence of the proposed framework without an [*a priori*]{} assumption on volatility dynamics. In applications, chiefly in asset pricing and financial risk management, the volatility process—and sometimes its dynamical equation—are of crucial importance. Accurate estimates of volatilities are important for measuring the risk of financial assets.
\[jumpdist\] Let $\mathcal{K}_t = \mathcal{J}_t \backslash \mathcal{J}_{t-}$. The jump size of $(X_t)$ at time $t$ is $X_t-X_{t-} = g(Z) - X_{t-}$, where the conditionally normal random variable $Z$ is given by $$Z = \sum_{i\in\mathcal{K}_t} \frac{ \xi_t^{(i)}}{\s_i^2(1-t)},$$ and the function $g:\R\rightarrow\R$, for $t \in [0,1)$, is given by $$g(z) = \frac{\int_{\mathbb{X}} x \exp(xz)\prod_{i \notin \mathcal{K}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i \in \mathcal{K}_t} \exp\left( \frac{-t x^2}{2\s_i^2(1-t)} \right) \nu(\dd x)}
{\int_{\mathbb{X}} \exp(xz)\prod_{i \notin \mathcal{K}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i \in \mathcal{K}_t} \exp\left( \frac{-t x^2}{2\s_i^2(1-t)} \right) \nu(\dd x)}.$$
For $i\in\mathcal{K}_t$, we have the following $$\left.\xi_t^{(i)} \right|_{X, \F_{t-}} \sim \mathcal{N}\left(\xi^{(i)}(\t^{(i)}_{t-})+(t-\t^{(i)}_{t-})X, \,\s_i^2(t-\tau^{(i)}_{t-})(1-t) \right).$$ Given $X$, the variables $\{\xi_t^{(i)}\}_{i=1,\ldots,n}$ are mutually independent. It follows that the conditional distribution of $Z$ is also Gaussian, that is, $$\left. Z \right|_{X, \F_{t-}} \sim \mathcal{N}\left(\sum_{i \in \mathcal{K}_t} \frac{\xi^{(i)}(\t^{(i)}_{t-})+(t-\t^{(i)}_{t-})X}{\s_i^2(1-t)}, \,
\sum_{i \in \mathcal{K}_t}\frac{t-\tau^{(i)}_{t-}}{\s_i^2(1-t)} \right).$$ Hence, the density of $Z$ is given by $$z \mapsto \frac{1}{\sqrt{2\pi V}}\int_{x\in\R} \exp\left( -\frac{(z-U(x))^2}{2V} \right) \nu_{t-}(\dd x),$$ where we have defined the following: $$\begin{aligned}
U(x) = \sum_{i \in \mathcal{K}_t} \frac{\xi^{(i)}(\t^{(i)}_{t-})+(t-\t^{(i)}_{t-})x}{\s_i^2(1-t)}, \hspace{0.1in}
V = \sum_{i \in \mathcal{K}_t}\frac{t-\tau^{(i)}_{t-}}{\s_i^2(1-t)}.
\end{aligned}$$ Note that one can decompose the conditional distribution of $X$ in terms of $\mathcal{K}_t$ and write $$\P[X \in \dd x \,|\, \F_t]
= \frac{\exp(xZ)\prod_{i \notin \mathcal{K}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i \in \mathcal{K}_t} \exp\left(\frac{-t x^2}{2\s_i^2(1-t)} \right) \nu(\dd x)}
{\int_{\mathbb{X}} \exp(xZ)\prod_{i \notin \mathcal{K}_t} h\left(x, \xi^{(i)}(\t^{(i)}_t), \t^{(i)}_t, \s_i\right)
\prod_{i \in \mathcal{K}_t} \exp\left(\frac{-t x^2}{2\s_i^2(1-t)} \right)\nu(\dd x)}.$$ Hence, the statement follows.
\[feynmankac\] Let $\mu(\hat{\xi}_t, \eta_t,t) =(\E[ X \,|\, \hat{\xi}_t, \eta_t ] - \hat{\xi}_t)/(1-t)$ and $(\lambda_t)$ be the intensity process of $(N_t)$. Let $\psi:\R\rightarrow\R$ and $\phi:\R\rightarrow\R$ be continuous bounded functions.
1. The conditional expectation $$v(\hat{\xi}_t, \eta_t,t) = \E\left[\left. e^{-\int_t^1 \psi(s)\dd s}\phi(X) \,\right| \hat{\xi}_t, \eta_t,X\1_{\{t=1\}}\right],$$ satisfies the partial differential equation $$\begin{gathered}
\frac{\partial v(\hat{\xi}_t, \eta_t,t)}{\partial t}
+ \mu(\hat{\xi}_t, \eta_t,t)\frac{\partial v(\hat{\xi}_t, \eta_t,t)}{\partial \hat{\xi}_t}
+ \hat{\s}_t^2\frac{\partial^2 v(\hat{\xi}_t, \eta_t,t)}{2\partial \hat{\xi}_t^2}
\\
- \psi(t)v(\hat{\xi}_t, \eta_t,t)
+ (v(\hat{\xi}_t, \eta_t,t)
- v(\hat{\xi}_{t-}, \eta_{t-},t-))\lambda_t = 0,\end{gathered}$$ with the boundary condition $v(\hat{\xi}_1, \eta_1,1) = \phi(X)$.
2. Choose the random point field $\zeta$ such that $\tilde{J}_t^{(i)}=1$ implies $\tilde{J}_u^{(i)}=1$ for $u\geq t$. Let there be at least one active source of information at $t=0$. Then, $$v(\hat{\xi}_t, t) = \E\left[\exp\left(-\int_t^1 \psi(s)\dd s\right)\phi(X) \,\bigg\vert\, \hat{\xi}_t\right],$$ satisfies the same partial differential equation of $v(\hat{\xi}_t, \eta_t,t)$ with the boundary condition $v(\hat{\xi}_1,1) = \phi(X)$, and where $\eta_t=1$ for all $t\in\mathbb{T}$.
See Appendix for the proof.
We conclude this section by producing an example of modulated information and associated endogenous jump-diffusion. In the figure below, we show a simulated path of the conditional expectation process $(X_t)$ given in Proposition \[ExpectationSDE\]. This jump-diffusion process jumps whenever an information coordinate is activated. At each point in time a Markov chain dictates which information coordinates are active. As expected, the effective information process has jump-diffusion dynamics, where the jumps occur whenever an information source is activated or deactivated. The interpretation is that when an information source (coordinate) comes online, the process $(X_t)$ instantaneously adjusts to account for the additional information contributed by the newly activated source. However, no jumps in $(X_t)$ are produced when an information source is deactivated. Because of the permanent nature of each “shock” given to $(X_t)$ whenever a information source is activated, we speak of a [*permanent impact*]{} on the value of $(X_t)$. The inclusion of a temporary impact on $(X_t)$ is a suggestion for future research, where the informational contribution of a particular information coordinate (source) “fades away” over time once the source is deactivated.
![Simulation of endogenous jump-diffusion system (one path). Here, an example is shown whereby the random point field is illustrated by a Markov chain. (a) the Markov chain, (b) the time-changed—or modulated—information process, (c) the effective information process, and (d) the endogenised jump-diffusion $X_t=\E[X_T\,\vert\,\F_t]$, for $0\le t <1$.](MarkovChain_Lambda20_C.png "fig:") ![Simulation of endogenous jump-diffusion system (one path). Here, an example is shown whereby the random point field is illustrated by a Markov chain. (a) the Markov chain, (b) the time-changed—or modulated—information process, (c) the effective information process, and (d) the endogenised jump-diffusion $X_t=\E[X_T\,\vert\,\F_t]$, for $0\le t <1$.](InfoEffective_Lambda20_C.png "fig:")\
![Simulation of endogenous jump-diffusion system (one path). Here, an example is shown whereby the random point field is illustrated by a Markov chain. (a) the Markov chain, (b) the time-changed—or modulated—information process, (c) the effective information process, and (d) the endogenised jump-diffusion $X_t=\E[X_T\,\vert\,\F_t]$, for $0\le t <1$.](InfoTau_Lambda20_C.png "fig:") ![Simulation of endogenous jump-diffusion system (one path). Here, an example is shown whereby the random point field is illustrated by a Markov chain. (a) the Markov chain, (b) the time-changed—or modulated—information process, (c) the effective information process, and (d) the endogenised jump-diffusion $X_t=\E[X_T\,\vert\,\F_t]$, for $0\le t <1$.](ExpectationProcess_Lambda20_C.png "fig:")
Multiple point fields
---------------------
In the spirit of [@4], we consider a system where $X$ is given by a function of multiple factors about which observers have differing information. That is, for some $m\in\mathbb{N}_+$, we introduce a vector $\boldsymbol{X}$ of mutually independent random variables $X^\alpha\in\mathcal{L}^2(\Omega,\mathcal{G},\P)$ with law $\nu^\alpha$ for $\alpha=1,\ldots,m$, respectively, each with the state-space $(\mathbb{X},\mathcal{B}(\mathbb{X}))$, where we choose $\mathbb{X}\subseteq\mathbb{R}$. We treat the case where $X=g(X^1,\ldots,X^m)$ for some bounded measurable function $g:\mathbb{X}^m\rightarrow\mathbb{X}$.
Accordingly, we introduce $\R^{n(\alpha)}$-valued $(\mathcal{G}_t)$-adapted multivariate stochastic processes $(\boldsymbol{\xi}^\alpha_t)_{t\in\mathbb{T}}$, for $\alpha=1,\ldots,m$, where $n(\alpha)$ highlights that the dimensions of $(\boldsymbol{\xi}^\alpha_t)$s may be different across $\alpha$. As before, we define information coordinates by $\xi^{\alpha,(i)}_t = t X^\alpha + \s_i^\alpha \b_t^{\alpha,(i)}$, for $\s_i^\alpha>0$ and $i=1,\ldots,n(\alpha)$. For the sake of parsimony, we assume that the standard Brownian bridges $\{\b_t^{\alpha,(i)}\}$s are mutually independent and of $X^\alpha$s across $i=1,\ldots,n(\alpha)$ and $\alpha=1,\ldots,m$.
Next we introduce a set of mutually independent random point fields $\zeta^\alpha$ for $\alpha=1,\ldots,m$ on $\mathbb{T}^{n(\alpha)}$, which are also independent of $(\boldsymbol{\xi}^\alpha_t)$s, each generating a collection of times $\{\varpi^{\alpha,i}_{1}, \ldots,\varpi^{\alpha,i}_{k_i}\}$ for $i=1,\ldots,n(\alpha)$ and finite $k_i\in\mathbb{N}_0$. For every $(\alpha,i)$, we associate the random sequence $\{\varpi^{\alpha,i}_{1}, \ldots,\varpi^{\alpha,i}_{k_i}\}$ to a coordinate of a $(\mathcal{G}_t)$-adapted càdlàg jump process $(\boldsymbol{\tilde{J}}^\alpha_t)$ with state space $\S=\{0,1\}^{n(\alpha)}$. We then define the $\R^{n(\alpha)\times n(\alpha)}$-valued process $(\boldsymbol{J}^\alpha_t)$ by $J_t^{\alpha,(i,j)} = \delta_{ij}\boldsymbol{\tilde{J}}^\alpha_t$. As before, $(\boldsymbol{J}^\alpha_t)$ is a diagonal matrix-valued process indicating which coordinates of $(\boldsymbol{\xi}^\alpha_t)$ are active through $(\boldsymbol{J_t}^\alpha\,\boldsymbol{\xi}^\alpha_t)$ for $\alpha=1,\ldots,m$. We define a sub-algebra $\H_t\subseteq \mathcal{G}_t$ by $$\label{subFextended}
\H_t = \s\left( (\boldsymbol{J}^\alpha_u\,\boldsymbol{\xi}^\alpha_u)_{0\leq u \leq t}, (\boldsymbol{J}^\alpha_u)_{0\leq u \leq t}, \boldsymbol{X}\1_{\{t=1\}}; \alpha=1,\ldots,m \right), \nonumber$$ for $t\in\mathbb{T}$. Keeping notations consistent and following our setup, we define $\R$-valued effective information processes $(\hat{\xi^\alpha_t})_{t\in\mathbb{T}}$ for $\alpha=1,\ldots,m$ by $$\begin{aligned}
\hat{\xi^\alpha_t}=(\hat{\s^\alpha_{t}})^2 \sum_{i \in \mathcal{J}^\alpha_t} (\s^\alpha_i)^{-2}\xi^{\alpha,(i)}_t, \hspace{0.1in} \text{where} \hspace{0.1in} \hat{\s^\alpha_{t}} = \left(\sum_{i \in \mathcal{J}^\alpha_t} (\s^\alpha_i)^{-2}\right)^{-1/2},\end{aligned}$$ where $\mathcal{J}^\alpha_t = \{i : J_t^{\alpha,(i,i)} = 1\}$. In a similar fashion, we define function-valued complementary information processes $(\eta^\alpha_t)_{t\in\mathbb{T}}$ for $\alpha=1,\ldots,m$ by $$\eta^\alpha_t: x^\alpha\mapsto \prod_{i\in\mathcal{J}_t^{\alpha \complement}} h\left(x^\alpha, \xi^{\alpha,(i)}(\t^{\alpha,(i)}_t), \t^{\alpha,(i)}_t, \s_i^\alpha\right),$$ where $\mathcal{J}_t^{\alpha \complement} = \{i : J_t^{\alpha,(i,i)} = 0\}$ and $\t^{\alpha,(i)}_t = 0 \vee \sup\{ u : J_u^{\alpha,(i,i)} = 1, u\in[0,t]\}$ for $i=1,\ldots,n(\alpha)$.
\[effectivecomplementaryextended\] The measure $\nu_t(\dd \boldsymbol{x})=\P[\boldsymbol{X} \in \dd \boldsymbol{x} \,|\, \H_t]$ satisfies $$\begin{aligned}
\nu_t(\dd \boldsymbol{x}) = \prod_{\alpha=1}^m \P[X^\alpha \in \dd x^\alpha \,|\, \hat{\xi}^\alpha_t, \eta^\alpha_t],\end{aligned}$$ for $t \in [0,1)$.
Using the independence properties given above, we have $$\begin{aligned}
\nu_t(\dd \boldsymbol{x}) &= \prod_{\alpha=1}^m \P[X^\alpha \in \dd x^\alpha \,|\, \s\left( (\boldsymbol{J}^\alpha_u\,\boldsymbol{\xi}^\alpha_u)_{0\leq u \leq t}, (\boldsymbol{J}^\alpha_u)_{0\leq u \leq t}, X^\alpha\1_{\{t=1\}}\right)] \nonumber \\
&= \prod_{\alpha=1}^m\frac{h\left(x^\alpha, \hat{\xi}^\alpha_t, t, \hat{\s_{t}^\alpha} \right) \eta^\alpha_t(x^\alpha) \, \nu^\alpha(\dd x^\alpha)}
{\int_{\mathbb{X}} h\left(x^\alpha, \hat{\xi}^\alpha_t, t, \hat{\s_{t}^\alpha} \right) \eta^\alpha_t(x^\alpha) \,\nu^\alpha(\dd x^\alpha)},\end{aligned}$$ by following similar steps as done in Proposition \[effectivecomplementary\].
We define $\pi^\alpha_j = \inf\{t:C^\alpha_t=j\}$, with $\pi^\alpha_0=0$, where $C^\alpha_t = \sum_{s\leq t} \1\{\boldsymbol{\tilde{J}}^\alpha_{s}\ne\boldsymbol{\tilde{J}}^\alpha_{s-}\}$, and also $N^\alpha_t = \sum_{s\leq t} \delta^\alpha_s$, where $\delta^\alpha_s = \1\{\mathcal{J}^\alpha_s \backslash \mathcal{J}^\alpha_{s-} \ne \emptyset \}$ for $\alpha=1,\ldots,m$.
\[ExpectationSDEextended\] Let $(W_t^{\alpha,(j)})$ be mutually independent $(\H_t)$-Brownian motions across $\alpha=1,\ldots,m$ for $t\in[\pi^\alpha_{j-1}, \pi^\alpha_j \wedge 1)$. Then, the process $X_t:=\mathbb{E}[X \,|\, \H_t]$ for $t\in[0,1)$ satisfies $$\begin{aligned}
X_t = \E[X] &+ \sum_{\alpha=1}^{m}\sum_{j=1}^{C^\alpha_t}\int_{\pi^\alpha_{j-1}}^{\pi^\alpha_j} \frac{\Theta^\alpha_s}{\hat{\s_{s}^\alpha} (1-s)} \d W_s^{\alpha,(j)} +\sum_{\alpha=1}^{m}\int_{\pi_{C^\alpha_t}}^{t} \frac{\Theta^\alpha_s}{\hat{\s_{s}^\alpha} (1-s)} \d W_s^{\alpha(C_t+1)} \nonumber \\
&+ \sum_{\alpha=1}^{m}\sum_{s\leq t} (X_s - X_{s-})\delta^\alpha_s,\end{aligned}$$ where $(\Theta^\alpha_t)_{t\in\mathbb{T}}$ is given by $\Theta^\alpha_t =\cov[X,X^\alpha\,|\,\H_t]$ for $\alpha=1,\ldots,m$.
Let $\nu_t^\alpha(\dd x^\alpha)=\P[X^\alpha \in \dd x^\alpha \,|\, \hat{\xi}^\alpha_t, \eta^\alpha_t]$ for $\alpha=1,\ldots,m$. Then using the independence properties given above, Proposition \[effectivecomplementaryextended\] and following similar steps as for Proposition \[measureSDE\], there exists a set of $(\H_t)$-Brownian motions $(W_t^{\alpha,(j)})$ over random periods $t\in[\pi^\alpha_{j-1}, \pi^\alpha_j \wedge 1)$, which are mutually independent from each other across $\alpha=1,\ldots,m$, where the brackets $\langle \d W_t^{a,(j)},\d W_t^{b,(j)}\rangle = 0$ for $a,b\in\{1,\ldots,m\}$, such that $(\nu_t)$ satisfies $$\begin{aligned}
\nu_t(\dd \boldsymbol{x}) &\law \nu_0(\dd \boldsymbol{x}) + \sum_{\alpha=1}^{m}\sum_{j=1}^{C_t^\alpha}\int_{\pi^\alpha_{j-1}}^{\pi^\alpha_j} \frac{x^\alpha-\E[X^\alpha\,|\, \hat{\xi}^\alpha_s, \eta^\alpha_s]}{\hat{\s_{s}^\alpha} (1-s)} \prod_{\alpha=1}^m\nu^\alpha_s(\dd x^\alpha) \d W_s^{\alpha,(j)} \nonumber \\
&+\sum_{\alpha=1}^{m}\int_{\pi_{C^\alpha_t}}^{t}\frac{x^\alpha-\E[X^\alpha\,|\, \hat{\xi}^\alpha_s, \eta^\alpha_s]}{\hat{\s_{s}^\alpha} (1-s)} \prod_{\alpha=1}^m\nu^\alpha_s(\dd x^\alpha) \d W_s^{\alpha,(C_t+1)}
+ \sum_{\alpha=1}^{m}\sum_{s\leq t} (\nu_s(\dd \boldsymbol{x}) - \nu_{s-}(\dd \boldsymbol{x}))\delta_s^\alpha,\end{aligned}$$ for $t \in [0,1)$. We also know that $\nu_t(\dd \boldsymbol{x}) = \prod_{\alpha=1}^m\nu^\alpha_t(\dd x^\alpha)$. Therefore, since $g:\mathbb{X}^m\rightarrow\mathbb{X}$ is a bounded measurable function, we have $$\begin{aligned}
X_t = \int_{\mathbb{X}^m}g(x^1,\ldots,x^m)\nu^1_t(\dd x^\alpha)\ldots\nu^m_t(\dd x^\alpha), \end{aligned}$$ and the statement follows via Lebesgue integration.
We notice that $(X_t)_{t\in\mathbb{T}}$ jumps whenever any of the inactive information sources on any of the factors $X^\alpha$ becomes active. In addition, the diffusion coefficient of $(X_t)$ takes the form of a stochastic covariance process, which itself jumps when any of the information sources in the system switches on. Over random time periods, when none of the information sources switch on, $(X_t)$ diffuses continuously.
\[feynmankacextended\] Let $\boldsymbol{\hat{\xi}}_t = [\hat{\xi}^1_t,\ldots,\hat{\xi}^m_t]$, and $\boldsymbol{\eta}_t = [\eta^1_t,\ldots,\eta^m_t]$. Also, let $\mu^\alpha(\hat{\xi}^\alpha_t, \eta^\alpha_t,t) =(\E[ X^\alpha \,|\, \hat{\xi}^\alpha_t, \eta^\alpha_t ] - \hat{\xi}^\alpha_t)/(1-t)$ and $(\lambda^\alpha_t)$ be the intensity process of $(N^\alpha_t)$. Let $\psi:\R\rightarrow\R$ and $\phi:\R\rightarrow\R$ be continuous bounded functions. Then the conditional expectation $$v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t) = \E\left[\left. e^{-\int_t^1 \psi(s)\dd s}\phi(X) \,\right| \boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,\boldsymbol{X}\1_{\{t=1\}}\right],$$ satisfies the partial differential equation $$\begin{gathered}
\frac{\partial v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t)}{\partial t}
+ \sum_{\alpha=1}^m\mu^\alpha(\hat{\xi}_t, \eta_t,t)\frac{\partial v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t)}{\partial \hat{\xi}^\alpha_t}
+ \sum_{\alpha=1}^m(\hat{\s}^\alpha_t)^2\frac{\partial^2 v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t)}{2(\partial \hat{\xi}^\alpha_t)^2}
\\
- \psi(t)v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t)
+ \sum_{\alpha=1}^m(v(\boldsymbol{\hat{\xi}}_t, \boldsymbol{\eta}_t,t)
- v(\boldsymbol{\hat{\xi}}_{t-}, \boldsymbol{\eta}_{t-},t-))\lambda^\alpha_t = 0,\end{gathered}$$ with the boundary condition $v(\boldsymbol{\hat{\xi}}_1, \boldsymbol{\eta}_1,1) = \phi(X)$.
We omit the proof of Proposition \[feynmankacextended\] to avoid repetition, which follows from similar steps as done in Proposition \[measureSDE\] and Proposition \[feynmankac\] by overlaying the independence properties given above, where we see that the brackets $\langle \d\hat{\xi}_t^{a,(j)},\d \hat{\xi}_t^{b,(j)}\rangle = 0$ for $a,b\in\{1,\ldots,m\}$.
Modulation as projection
------------------------
For any $t\in\mathbb{T}$ and for any fixed $\alpha=1,\ldots,m$, $\boldsymbol{J}^\alpha_t\,\boldsymbol{\xi}^\alpha_t$ defines an orthogonal projection from the space of all available information sources to an information subspace. That is, $(\boldsymbol{J}^\alpha_t)$ is a symmetric projection-valued stochastic process acting on $\boldsymbol{\xi}^\alpha_t$, whereas $\boldsymbol{A}^\alpha_t = \boldsymbol{I} - \boldsymbol{J}^\alpha_t$, with $\boldsymbol{I}$ the identity matrix, is an annihilator matrix determining a complementary information subspace. This motivates us to ask how far we can push this idea towards its logical limits for modelling information modulation. This brings forth one possible generalisation.
A *modulated Brownian bridge information system* is given by the probability space $(\Omega, \mathcal{H},\P)$ equipped with a right-continuous and complete filtration $(\mathcal{H}_t)_{t\in\mathbb{T}}$, where
1. $\H_t = \s\left( (\boldsymbol{P}^\alpha_u\,\boldsymbol{\xi}^\alpha_u)_{0\leq u \leq t}, (\boldsymbol{P}^\alpha_u)_{0\leq u \leq t}, \boldsymbol{X}\1_{\{t=T\}} ; \alpha=1,\ldots,m \right)$.
2. $(\boldsymbol{\xi}^\alpha_t)_{t\in\mathbb{T}}$ is an $\R^{n(\alpha)}$-valued Brownian bridge information process for $\alpha=1,\ldots,m$.
3. $(\boldsymbol{P}^\alpha_t)_{t\in\mathbb{T}}$ is an $n(\alpha)$-dimensional projection-valued stochastic process with real càdlàg paths for $\alpha=1,\ldots,m$.
4. $\boldsymbol{X}\in\mathcal{L}^2(\Omega,\mathcal{G},\P)$ is an $m$-dimensional vector of random variables with law $\boldsymbol{\nu}$ and state-space $(\mathbb{X}^m,\otimes^{m}_{\alpha=1}\mathcal{B}(\mathbb{X}))$, where $\mathbb{X}\subseteq\R$ and $\mathcal{B}(\mathbb{X})$ is the Borel $\sigma$-field.
5. $\mathbb{T}=[0,T]$ is a finite interval for some $T<\infty$.
This definition of the system does not explicitly rely on any random point field. However, one can assign a law to $(\boldsymbol{P}^\alpha_t)_{t\in\mathbb{T}}$ via associating it with a random point field $\zeta^\alpha$. Although we leave a detailed study of such a system for future research, we shall nonetheless provide a simple example to show how—what we call—*information mixing* arises.
Let $m=1$ and $(\boldsymbol{\xi}_t)_{t\in\mathbb{T}}$ be defined as above, and $p_t^{(ij)}$ denote the $(i,j)$th coordinate of an $n$-dimensional projection matrix $\boldsymbol{P}_t$. Then at any $t\in\mathbb{T}$, the $\sigma$-algebra $\H_t$ allows the mapping $\boldsymbol{P}_t\,\boldsymbol{\xi}_t \mapsto \boldsymbol{\psi}_t$, where $$\psi_t^{(i)} =
\left\{ \begin{aligned}
&0, && \text{if $p_t^{(ij)} =0$ for all $j\in\{1,\ldots,n\}$,}
\\ &tX + \hat{p}_t^{(i)}\sum_{j=1}^n p_t^{(ij)}\sigma_j\beta_t^{(j)}, &&\text{otherwise,}
\end{aligned} \right.$$ and where $\hat{p}_t^{(i)} = \left(\sum_{j=1}^n p_t^{(ij)}\right)^{-1}$. Thus, $\psi_t^{(i)}\law tX + \alpha^{(i)}_tB^{(i)}_t$ given that $(B^{(i)}_t)_{t\in\mathbb{T}}$ is a standard Brownian bridge and the bridge coefficient is $$\begin{aligned}
\alpha^{(i)}_t = \hat{p}_t^{(i)}\left(\sum_{j=1}^n (p_t^{(ij)})^2\sigma_j^2\right)^{\frac{1}{2}}.\end{aligned}$$ When $\boldsymbol{P}_t$ is diagonal (e.g. $\boldsymbol{P}_t = \boldsymbol{J}_t$), the observer of $\boldsymbol{P}_t\,\boldsymbol{\xi}_t$ also observes the active coordinates of $\boldsymbol{\xi}_t$, possibly scaled. However, in the more general case, the observer may not be able to decouple all the active coordinates of $\boldsymbol{\xi}_t$ from the observed $\boldsymbol{P}_t\,\boldsymbol{\xi}_t$, since $\boldsymbol{P}_t$ is non-invertible, unless $\boldsymbol{P}_t=\boldsymbol{I}$, for it is a projection. Hence, the original information provided from each source may get mixed with another.
Applications {#sec:application}
============
Thus far, we have kept the context of our approach fairly general without, for instance, specifying an interpretation of the signal $X$. Next, we apply the framework to, e.g., finance in the spirit of [@4]. We assume that $\P$ is the pricing measure, $(\F_t)$ is the market information, $X$ is the cash flow of a financial asset at $t=1$, and the price of the asset at time $t\in[0,1)$ is given by the discounted expected value of $X$ conditional on $\F_t$. We let the risk-free system be deterministic and denote the system of discount functions by $(P_{0t})_{0 \leq t <\infty}$. We assume that $P_{0t}$ is differentiable, strictly decreasing and that it satisfies $0 < P_{0t} \leq 1$ and $\lim_{t\rightarrow\infty}P_{0t}=0$. The no-arbitrage condition implies that $P_{t1}=P_{01}/P_{0t}$ for $t\leq 1$. We write $\overline{X}_t = P_{t1}X_t = P_{t1}\E[X\,|\,\F_t]$ for the price at time $t$. The price process $(\overline{X}_t)$ has jump-diffusion dynamics, see Section \[EJD\].
Merton-type jump-diffusion models for vanilla options
-----------------------------------------------------
We choose $\zeta$ such that $\tilde{J}_t^{(i)}=1$ implies $\tilde{J}_u^{(i)}=1$ for $u\geq t$. This removes any path-dependency in the system caused by randomly paused flows of information. We also let there be at least one active source of information at $t=0$, so that some information about $X$ is always available to market participants. As an example, we shall price a European-style call option with strike price $K$ and exercisable at a fixed time $t\in(0,1)$. The price at time zero is given by $C_{0}=P_{0t}\E\left[(\overline{X}_{t}-K)^{+}\right]$ where $0 < t < 1$. To make the derivation easier to follow, we represent processes explicitly as a function of the state $k\in \S$, e.g., we write $\hat{\xi}_t(k)$ for $\boldsymbol{\tilde{J}}_t=k$.
\[callprice\] The price at $t=0$ of the European-style call option is given by $$\begin{aligned}
C_{0}&=P_{0t}\sum_{k\in\S}\P(\boldsymbol{\tilde{J}}_{t} = k)\int_{\mathbb{X}}x\mathcal{N}\left(-\hat{z}_{t}(k)+\frac{x\sqrt{t}}{\hat{\sigma}_t^2(k)\sqrt{(1-t)}}\right)\nu(\dd x) \nonumber \\
&-P_{0t}\sum_{k\in\S}\P(\boldsymbol{\tilde{J}}_{t} = k)K\int_{\mathbb{X}}\mathcal{N}\left(-\hat{z}_{t}(k)+\frac{x\sqrt{t}}{\hat{\sigma}_t^2(k)\sqrt{(1-t)}}\right)\nu(\dd x),\nonumber\end{aligned}$$ for $t\in(0,1)$, where $\mathcal{N}(\cdot)$ is the standard normal distribution function and $\hat{z}_{t}(k)=\hat{\varsigma}(k)/\sqrt{t(1-t)}$, where $\hat{\varsigma}(k)$ is the unique solution to $$\begin{aligned}
\int_{\mathbb{X}}(P_{t1}x-K)\exp\left[\frac{1}{\hat{\sigma}_t^2(k)(1-t)}\left(x \hat{\varsigma}(k)-tx^2/2\right)\right]\nu(\dd x)=0. \nonumber\end{aligned}$$
See Appendix for the proof.
The option price can be represented as the weighted sum of Black-Scholes-Merton prices induced by the various combinations of active information processes. The functional form of the price is very similar to that presented in [@19] where jumps are driven by a Poisson process. In our framework, the jump-diffusion dynamics of the price process emerges from the nature of the market information (i.e. *endogenous*) and the jump distribution is not specified [*a priori*]{}. We refer to [@9] and [@10] for partial integro-differential equations and viscosity solutions that offer alternative techniques for option pricing with respect to jump-diffusion dynamics.
Information asymmetry and market competition
--------------------------------------------
We consider a setting in which there are two market agents who are unaware of each other’s actions and have differing access to a fixed universe of information sources. Let $(\boldsymbol{J}_t^{(1)})$ and $(\boldsymbol{J}_t^{(2)})$ be the diagonal matrices of $(\boldsymbol{\tilde{J}}_t^{(1)})\in\S$ and $(\boldsymbol{\tilde{J}}_t^{(2)})\in\S$, which are $\mathcal{G}_t$-adapted càdlàg jump processes of Agents 1 and 2, respectively. The jump processes are generated by two random point fields $\zeta^{(1)}$ and $\zeta^{(2)}$, both independent of $(\boldsymbol{\xi}_t)$. We define the sub-algebras $\F_t^{(j)}\subseteq \mathcal{G}_t$ by $$\F_t^{(j)} = \s\left( (\boldsymbol{J}_u^{(j)} \boldsymbol{\xi}_u)_{0\leq u \leq t}, (\boldsymbol{J}_u^{(j)})_{0\leq u \leq t}, X1_{\{t=T\}} \right),$$ for $j=\{1,2\}$. Then we introduce $\nu_t^{(j)}(A) = \P[X \in A \,|\, \F_t^{(j)}]$, where, due to Proposition \[effectivecomplementary\], we have $\nu_t^{(j)}(A) = \P[X \in A \,|\, \hat{\xi}_t^{(j)}, \eta_t^{(j)}]$. We note that $\P[\hat{\xi}_t^{(1)}=\hat{\xi}_t^{(2)}]<1$ and $\P[\eta_t^{(1)}=\eta_t^{(2)}]<1$, since the set-valued processes $(\mathcal{J}_t^{(j)})$, which define the effective and complementary information processes of the agents, are different. To model the information asymmetry in a dynamic competition between the agents, we shall utilise $f$-divergences, see [@3; @11]. We can define a symmetric $f$-divergence between equivalent probability measures $\P^{(1)}$ and $\P^{(2)}$ as $$\Delta_{f}\left[\P^{(1)}||\,\P^{(2)}\right]=\frac{1}{2}\left[\int_{\Omega}f\left(\frac{\dd \P^{(1)}(\omega)}{\dd \P^{(2)}(\omega)}\right)\dd \P^{(2)}(\omega) + \int_{\Omega}f\left(\frac{\dd \P^{(2)}(\omega)}{\dd \P^{(1)}(\omega)}\right)\dd \P^{(1)}(\omega)\right],\nonumber$$ for $\omega\in\Omega$, where $f$ is a convex function that satisfies $f(1)=0$, and where $\dd \P^{(1)}/\dd \P^{(2)}$ is the Radon-Nikodym derivative. Alternatively, a symmetric $f$-divergence can be defined in terms of probability densities, assuming that they exist. For convenience, we assume that $X$ has a density and write $\nu_t^{(j)}(\dd x) = p_t^{(j)}(x)\dd x$ such that $p_t^{(j)}(x)>0$ for $t\in[0,1)$ and $x\in\mathbb{X}$.
As an example, we shall utilise Kullback-Leibler divergence (relative entropy), which is not a distance metric on the space of probability distributions, but measures the information gain when moving from a prior distribution to a posterior distribution. For $t\in[0,1)$, $$\begin{aligned}
\Delta_{\text{KL}}\left[p_t^{(1)} ||\, p_t^{(2)}\right]&=\frac{1}{2}\int_{\mathbb{X}}\left[p_t^{(1)}(x)\log\left(\frac{p_t^{(1)}(x)}{p_t^{(2)}(x)}\right)+p_t^{(2)}(x)\log\left(\frac{p_t^{(2)}(x)}{p_t^{(1)}(x)}\right)\right]\dd x.\nonumber\end{aligned}$$ We represent the probability density functions of the agents explicitly as a function of the state and write $p_t^{(j)}(x;k_j)$ for $\boldsymbol{\tilde{J}}^{(j)}_t = k_j$, where $k_j\in\S$ for $j=\{1,2\}$. Thus, we have $p_t^{(j)}(x;k_j)\dd x = \P[X \in \dd x \,|\, \hat{\xi}_t(k_j),\eta_t^{(j)}(k_j)]$. Then, by defining the processes $(A_t(k_1,k_2))$ and $(B_t(k_1,k_2))$ by $$\begin{aligned}
&A_t(k_1,k_2)=\int_{\mathbb{X}}\log\left(\frac{p_t^{(1)}(x;k_1)}{p_t^{(2)}(x;k_2)}\right)\nu_t^{(j)}(\dd x;k_1),
&B_t(k_1,k_2)=\int_{\mathbb{X}}\log\left(\frac{p_t^{(2)}(x;k_2)}{p_t^{(1)}(x;k_1)}\right)\nu_t^{(2)}(\dd x;k_2), \nonumber\end{aligned}$$ respectively, we may write $$\begin{aligned}
2\Delta_{\text{KL}}\left[p_t^{(1)} || p_t^{(2)}\right]&=\sum_{k_1\in\S}\sum_{k_2\in\S}\left(A_t(k_1,k_2) +
B_t(k_1,k_2) \right)\1\{\boldsymbol{\tilde{J}}_{t}^{(1)}= k_1\}\1\{\boldsymbol{\tilde{J}}_{t}^{(2)}= k_2\}.\nonumber\end{aligned}$$ This is a useful representation to derive an SDE of the Kullback-Leibler asymmetry process via Proposition \[measureSDE\]. The process jumps every time one of the agents gains access to a new source of information. If both agents have the same information at a given time, that is, if $\nu_t^{(1)}(\dd x;k_1)=\nu_t^{(2)}(\dd x;k_2)$ for some $t\in[0,1)$, then the process is zero at that time.
The study of information asymmetry (see, for example [@1; @2; @18]) has benefited substantially from the theory of enlargements of filtrations; see [@17] and [@22].
Activation of new information sources at random times can be understood in terms of progressive enlargements of filtrations. Combining this with deactivation of information, we may view the proposed framework as a dynamic interplay between enlargements of filtrations and stopped filtrations.
We leave a formal treatment of this remark and a corresponding rigorous application of information asymmetry to market competition with multiple agents for future research.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank J. Akahori, G. W. Peters, J. Sekine, and participants in the Osaka-UCL Workshop on Stochastics, Numerics and Risk (March 2017), which was co-sponsored by The Daiwa Anglo-Japanese Foundation, and participants in the 6th International Conference, Mathematics in Finance, RSA (August 2017), for comments and suggestions.
Appendix {#sec:appendix .unnumbered}
========
Proposition \[rem1\] {#proposition-rem1 .unnumbered}
--------------------
Per coordinate $i$, the dynamics of $(\tau^{(i)}_t)$ can be decomposed into its continuous and discontinuous parts, where $\dd J^{(i,i)}_t=0$ or $\dd J^{(i,i)}_t=1$, respectively. Given that $(J^{(i,i)}_t)$ has state-space $\{0,1\}$, we have $\tau^{(i)}_t=t$ whenever $J^{(i,i)}_t=1$ and $\tau^{(i)}_t<t$ whenever $J^{(i,i)}_t=0$. Hence, for the continuous part, when $J^{(i)}_t=1$, $\dd \tau^{(i)}_t=\dd t$, and when $J^{(i,i)}_t=0$, $\dd \tau^{(i)}_t=0$. As for the discontinuous part, just before a jump, if $J^{(i,i)}_{t-}=1$, then $\dd \tau^{(i)}_t=\dd t$ since $\tau^{(i)}_{t-}=t$. If just before a jump, $J^{(i,i)}_{t-}=0$, then $\dd \tau^{(i)}_t=(t-t^*)$ given that $\tau^{(i)}_{t-}=t^*$ for some $t^*<t$ and $\tau^{(i)}_t=t$.
Lemma \[lemmaeffectiveinfo\] {#lemma-lemmaeffectiveinfo .unnumbered}
----------------------------
By definition, for $\mathcal{J}_t\neq\emptyset$, the effective information process is $\hat{\xi}_t = tX + \hat{\s_{t}} \sum_{i \in \mathcal{J}_t} \s^{-1}_i\b^{(i)}_t$. A linear combination of independent Brownian bridges is a Brownian bridge. Further, we have $$\begin{aligned}
\E\left[\hat{\s_{t}} \sum_{i \in \mathcal{J}_t} \s^{-1}_i\b^{(i)}_t\right] =0, \hspace{0.1in} \var\left[\hat{\s_{t}} \sum_{i \in \mathcal{J}_t} \s^{-1}_i\b^{(i)}_t\right] =t(1-t),\end{aligned}$$ hence, $(\hat{\b}_t)$ is a *standard* Brownian bridge between jumps.
Proposition \[measureSDE\] {#proposition-measuresde .unnumbered}
--------------------------
Due to the finite number of jumps generated by the random point field $\zeta$ on $\mathbb{T}^n$, the measure-valued $(\nu_t)$ can be represented by the sum of the continuous and the discontinuous components via the decomposition $\nu_t(A) = \nu_t^c(A) + \sum_{s\leq t} \Delta \nu_s(A)$. We first focus on the continuous part $\nu_t^c(\dd x)$, which can be written as $$\begin{aligned}
\nu_t^c(\dd x) = \frac{h\left(x, \hat{\xi}_t^c, t, \hat{\s}_{t}^c \right) \eta_t^c(x) \, \nu(\dd x)}
{\int_{\mathbb{X}} h\left(x, \hat{\xi}_t^c, t, \hat{\s}_{t}^c \right) \eta_t^c(x) \,\nu(\dd x)}. \nonumber\end{aligned}$$ The continuous part of the volatility parameter $\hat{\s}_{t}^c$ is constant between discontinuities. Then, we define a function of $\hat{\xi}_t^c$, $\eta_t^c$ and $t$ as $$g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right) = h\left(x,\hat{\xi}_t^c, t, \hat{\s}_{t}^c \right) \eta_t^c(x) \, \nu(\dd x),$$ and also $$G\left( \hat{\xi}_t^c, t, \eta_t^c ;\hat{\s}_{t}^c\right)=\int_{\mathbb{X}} g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right).$$ Using Ito’s lemma, we have $$\begin{aligned}
\dd g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right)
&= \frac{\partial g}{\partial t}\dd t + \frac{\partial g}{\partial \hat{\xi}_t^c} \dd \hat{\xi}_t^c + \frac{\partial^2 g}{2 \partial (\hat{\xi}_t^c)^2}\dd (\hat{\xi}_t^c)^2 \nonumber
\\
&\hspace{3.25cm}+ \frac{\partial g}{\partial \eta_t^c}\dd \eta_t^c + \frac{\partial^2 g}{2\partial (\eta_t^c)^2}\dd (\eta_t^c)^2 + \frac{\partial^2 g}{\partial \hat{\xi}_t^c \eta_t^c}\dd \hat{\xi}_t^c \dd \eta_t^c \nonumber
\\
&= \frac{\partial g}{\partial t}\dd t + \frac{\partial g}{\partial \hat{\xi}_t^c} \dd \hat{\xi}_t^c + \frac{\partial^2 g}{2 \partial (\hat{\xi}_t^c)^2} (\hat{\s}_{t}^c)^2 \dd t \nonumber
\\
&= g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right) \left(\frac{x\hat{\xi}_t^c}{(\hat{\s}_{t}^c)^2(1-t)^2}\dd t + \frac{x}{(\hat{\s}_{t}^c)^2(1-t)}\dd \hat{\xi}_t^c \right), \nonumber\end{aligned}$$ since the quadratic variation of $(\dd \hat{\b}_t^c)$ is $\dd t$ and $\eta_t^c$ (the continuous part of $\eta_t$) is constant between discontinuities. Then using Fubini’s theorem, we can write $$\begin{aligned}
\dd G\left( \hat{\xi}_t^c, t, \eta_t^c ;\hat{\s}_{t}^c\right) &= \int_{\mathbb{X}} g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right) \left(\frac{x\hat{\xi}_t^c}{(\hat{\s}_{t}^c)^2(1-t)^2}\dd t + \frac{x}{(\hat{\s}_{t}^c)^2(1-t)}\dd \hat{\xi}_t^c \right) \nonumber
\\
&= G\left( \hat{\xi}_t^c, t, \eta_t^c ;\hat{\s}_{t}^c\right) \left(\frac{\E[X\,|\, \hat{\xi}_t^c, \eta_t^c]\hat{\xi}_t^c}{(\hat{\s}_{t}^c)^2(1-t)^2}\dd t + \frac{\E[X\,|\, \hat{\xi}_t^c, \eta_t^c]}{(\hat{\s}_{t}^c)^2(1-t)}\dd \hat{\xi}_t^c \right). \nonumber\end{aligned}$$ Finally, for the bracket $\left\langle G,G \right\rangle$, we have $$\dd \left\langle G,G \right\rangle = G\left( \hat{\xi}_t^c, t, \eta_t^c ;\hat{\s}_{t}^c\right)^2\frac{\E[X\,|\, \hat{\xi}_t^c, \eta_t^c]^2}{(\hat{\s}_{t}^c)^2(1-t)^2}\dd t,$$ and the bracket $\left\langle g,G \right\rangle$ satisfies $$\dd \left\langle g,G \right\rangle = g\left( \hat{\xi}_t^c, t, \eta_t^c ;x,\hat{\s}_{t}^c,\dd x\right)G\left( \hat{\xi}_t^c, t, \eta_t^c ;\hat{\s}_{t}^c\right)\frac{x\E[X\,|\, \hat{\xi}_t^c, \eta_t^c]}{(\hat{\s}_{t}^c)^2(1-t)^2}\dd t.$$ Using the Ito quotient rule and rearranging terms, we have $$\dd \nu_t^c(\dd x) = \frac{x - \E[X\,|\, \hat{\xi}^c_t, \eta^c_t]}{(\hat{\s}_{t}^c)^2(1-t)}\nu_t^c(\dd x) \dd M_t^c,$$ for $t \in [0,1)$, where $$\dd M_t^c = \dd \hat{\xi}^c_t - \frac{\E[ X \,|\, \hat{\xi}^c_t, \eta^c_t ] - \hat{\xi}^c_t}{1-t} \d t.$$ The integral in the continuous part of $(M_t)$ is well-defined over $[0,t]$ with respect to the Lebesgue measure. That is, having $\pi_j\in(0,t)$ for $j=1,\ldots,m<\infty$ and denoting $u_0=\pi_0+=0$ and $u_{m+1}=\pi_{m+1}-=t$, $$\begin{aligned}
\int_0^t\frac{\E[ X \,|\, \hat{\xi}_s, \eta_s ] - \hat{\xi}_s}{ 1-s} \d s &=&\sum_j \lim_{u_{j-1} \rightarrow (\pi_{j-1})+} \hspace{0.1in} \lim_{u_j \rightarrow \pi_j-} \int_{u_{j-1}}^{u_{j}}\frac{\E[ X \,|\, \hat{\xi}_s, \eta_s ] - \hat{\xi}_s}{ 1-s} \d s \nonumber
\\
&+&\sum_j \lim_{u_j \rightarrow \pi_j+} \hspace{0.1in} \lim_{u_{j+1} \rightarrow (\pi_{j+1})-} \int_{u_{j}}^{u_{j+1}}\frac{\E[ X \,|\, \hat{\xi}_s, \eta_s ] - \hat{\xi}_s}{1-s} \d s. \nonumber\end{aligned}$$ Finally, the discontinuous part can be rewritten as $$\begin{aligned}
\sum_{s\leq t} \Delta \nu_s(\dd x)
= \sum_{s\leq t}\left(\nu_s(\dd x)-\nu_{s-}(\dd x)\right)\1\{\mathcal{J}_s \backslash \mathcal{J}_{s-} \ne \emptyset \},\end{aligned}$$ for $t \in [0,1)$. The statement follows by Lebesgue integration over any $A\in\mathcal{B}(\mathbb{X})$.
Proposition \[ExpectationSDE\] {#proposition-expectationsde .unnumbered}
------------------------------
Note that the random time $\pi_{j-1}$ is $(\F_t)$-measurable, $W_{\pi_{j-1}}=0$ and the subprocess $(\hat{\b}_t - \hat{\b}_{\pi_{j-1}})_{\pi_{j-1}\leq t \leq \pi_j}$ is an $(\F_t)$-Brownian bridge. Thus, the bracket $\dd\left\langle W_t, W_t \right\rangle$ for $t\in[\pi_{j-1},\pi_{j})$ is $\dd t$ given $(\F_t)$. Since the paths $(M_t)_{\pi_{j-1}\leq t < \pi_{j}}$ and $(\hat{\s}_t)_{\pi_{j-1}\leq t < \pi_{j}}$ are continuous, $(W_t)_{\pi_{j-1}\leq t < \pi_{j}}$ is an $(\F_t)$-Brownian motion by Lévy’s characterization. The dynamics given in the second part follows directly from Proposition \[measureSDE\], the $(\hat{\s}_t)$-standardization of $(M_t)$ and Lebesgue integration. For the $(\F_t)$-supermartingale property of $(\Gamma_t)$, define $(S_t)_{0\leq t < 1}$ by $S_t = X_t^2$. Using Ito’s lemma, $(S_t)$ is an $(\F_t)$-submartingale. Then from Doob-Meyer decomposition, $$\begin{aligned}
\E\left[ \E[X^2 \,|\, \F_t] - S_t \,|\, \F_s \right] = \E\left[ \E[X^2 \,|\, \F_t] - \left(Y_t - I_t\right) \,|\, \F_s \right] \leq \var[X\,|\, \F_s],\end{aligned}$$ where $(Y_t)$ is an $(\F_t)$-martingale and $(I_t)$ is an increasing predictable process. The last part is given by $$\begin{aligned}
X_t^{(k)} &= \E[X^k] + \sum_{j=1}^{C_t}\int_{\pi_{j-1}}^{\pi_j} \frac{\E[X^{k+1}\,|\, \hat{\xi}_s, \eta_s] - \E[X^{k}\,|\, \hat{\xi}_s, \eta_s]X_s}{\hat{\s_{s}} (1-s)} \d W_s^{(j)} \notag \\
&+\int_{\pi_{C_t}}^{t} \frac{\E[X^{k+1}\,|\, \hat{\xi}_s, \eta_s] - \E[X^{k}\,|\, \hat{\xi}_s, \eta_s]X_s}{\hat{\s_{s}} (1-s)} \d W_s^{(C_t+1)} + \sum_{s\leq t} (X^{(k)}_s - X^{(k)}_{s-})\delta_s.\end{aligned}$$ which follows from the first two parts and the Lebesgue integration from Proposition \[measureSDE\].
Proposition \[feynmankac\] {#proposition-feynmankac .unnumbered}
--------------------------
For part one, by use of the Doob-Meyer decomposition, we can write $N_t = \hat{N}_t + \Lambda_t$, where $(\hat{N}_t)$ is an $(\F_t)$-adapted martingale and $\Lambda_t = \int_0^t \lambda_s \dd s$ is the compensator process. Define $\hat{v}(\hat{\xi}_t, \eta_t,t) = e^{-\int_0^t \psi(s)\dd s} v(\hat{\xi}_t, \eta_t,t)$, which is a martingale. Then, by applying Ito’s lemma to $\hat{v}(\hat{\xi}_t, \eta_t,t)$ by decomposing it into continuous and discontinuous parts, we have $$\begin{aligned}
&\frac{\hat{v}(\hat{\xi}_t, \eta_t,t)}{\partial t}\dd t + \frac{\hat{v}(\hat{\xi}_t, \eta_t,t)}{\partial \hat{\xi}_t}\left(\dd M_t + \mu(\hat{\xi}_t, \eta_t,t)\dd t\right) \\
&+ \hat{\s}_t^2\frac{\partial^2 \hat{v}(\hat{\xi}_t, \eta_t,t)}{2\partial \hat{\xi}_t^2}\dd t + \left(\hat{v}(\hat{\xi}_t, \eta_t,t) - \hat{v}(\hat{\xi}_{t-}, \eta_{t-},t-) \right)(\dd\hat{N}_t + \dd\Lambda_t)\end{aligned}$$ Note that for the continuous part, we have $\dd \eta_t = 0$ and $\dd \left\langle \eta_t,\eta_t \right\rangle = 0$. Thus, once the conditional expectation is taken with respect to $\F_t$, the continuous part of $(M_t)$ and the discontinuous part $(\hat{N}_t)$ vanish, $$\begin{gathered}
e^{-\int_0^t \psi(s)\dd s}\left(\frac{\partial v(\hat{\xi}_t, \eta_t,t)}{\partial t} - \psi(t)v(\hat{\xi}_t, \eta_t,t) + \mu(\hat{\xi}_t, \eta_t,t)\frac{\partial v(\hat{\xi}_t, \eta_t,t)}{\partial \hat{\xi}_t} + \hat{\s}_t^2\frac{\partial^2 v(\hat{\xi}_t, \eta_t,t)}{2\partial \hat{\xi}_t^2}\right)\dd t \\
+ e^{-\int_0^t \psi(s)\dd s}(v(\hat{\xi}_t, \eta_t,t)
- e^{-\int_0^{t-} \psi(s)\dd s}v(\hat{\xi}_{t-}, \eta_{t-},t-))\lambda_t \dd t = 0,\end{gathered}$$ Once we divide both sides by $e^{-\int_0^t \psi(s)\dd s}\dd t$, the jump part remains to be $$\begin{aligned}
\left(v(\hat{\xi}_t, \eta_t,t) - e^{\int_{t-}^{t} \psi(s)\dd s}v(\hat{\xi}_{t-}, \eta_{t-},t-)\right)\lambda_t,\end{aligned}$$ However, $\int_{t-}^{t} \psi(s)\dd s=0$ due to continuity, and hence, $e^{\int_{t-}^{t} \psi(s)\dd s}=1$. Finally, $X\1_{\{t=1\}}$ ensures that the boundary condition is satisfied even if $\mathcal{J}_1=\emptyset$. For part two, active sources never deactivate. Thus for inactive states $i\in\mathcal{J}_t^c$, the complementary information is $\eta_t=1$ for all $t\in[0,1]$, since $\xi^{(i)}(0)=0$ and $\tau_t^{(i)}=0$ must hold for all $t\in[0,1]$. The boundary condition follows since at least one information source is active.
Proposition \[callprice\] {#proposition-callprice .unnumbered}
-------------------------
Let $\overline{X}_t(k) = P_{t1}\E[X\,|\,\hat{\xi}_t(k)]$, which is the price of the asset given that $\boldsymbol{\tilde{J}}_t$ is in state $k\in\S$. That is, since $\hat{\xi}_t$ has to be in one of the $2^n$ states at any time, $k$ should be understood as an identifier of the active and inactive coordinates of $\boldsymbol{\xi}_t$. Then, the price of the option is given by $$\begin{aligned}
C_{0}=\sum_{k\in\S}C_0(k)\P(\boldsymbol{\tilde{J}}_t=k), \nonumber\end{aligned}$$ for $t\in(0,1)$, where $C_0(k)=P_{0t}\E\left[(\overline{X}_t(k)-K)^{+}\right]$. This follows from the law of total expectation, where we have $$\begin{aligned}
\E\left[(\overline{X}_{t}-K)^{+}\right]=\E\left[\E\left[\left.(\overline{X}_{t}-K)^{+} \,\right|\, \boldsymbol{\tilde{J}}_{t}\right]\right]&=\sum_{k\in\S}\E\left[\left.(\overline{X}_{t}-K)^{+} \,\right|\, \boldsymbol{\tilde{J}}_{t}=k\right]\P(\boldsymbol{\tilde{J}}_{t} = k) \notag \\
&=\sum_{k\in\S}\E\left[(\overline{X}_{t}(k)-K)^{+}\right]\P(\boldsymbol{\tilde{J}}_{t}= k), \nonumber\end{aligned}$$ since $(\boldsymbol{\xi}_t)$ is Markovian and independent of the state process $(\boldsymbol{\tilde{J}}_t)$, and $\eta_t=1$ for all $t\in\mathbb{T}$. We can write the conditional distribution as an explicit function of $k\in\S$, $$\begin{aligned}
\nu_{t}(\dd x;k)=\frac{\exp\left[\frac{1}{(1-t)}\hat{\sigma}_t^{-2}(k)\left(x\hat{\xi}_{t}(k)-tx^2/2\right)\right]\nu(\dd x)}{\int_{\mathbb{X}}\exp\left[\frac{1}{(1-t)}\hat{\sigma}_t^{-2}(k)\left(x\hat{\xi}_{t}(k)-tx^2/2\right)\right]\nu(\dd x)}, \nonumber\end{aligned}$$ for $t\in(0,1)$. Following similar steps as in [@4], which we shall not repeat here, we have $$\begin{aligned}
C_{0}(k)=P_{0t}\E\left[\frac{1}{\Phi_{t}(k)}\left(\int_{\mathbb{X}}(P_{t1}x-K)\chi_{t}(x;k)\nu(\dd x)\right)^{+}\right], \nonumber\end{aligned}$$ where $\Phi_{t}(k)=\int_{\mathbb{X}}\chi_{t}(x;k)\nu(\dd x)$ and $$\chi_{t}(x;k)=\exp\left(\frac{x \hat{\xi}_{t}(k)-tx^2/2}{\hat{\sigma}_t^{2}(k)(1-t)}\right).$$ Then, for $t\in(0,1)$, we can define the measure $\mathbb{B}$ on $(\Omega,\mathcal{G},\{\mathcal{G}_{t}\})$ through a sequence of Radon-Nikodym derivatives $$\left\{\left.\frac{\d \mathbb{B}}{\d \mathbb{P}}\right | _{{\sigma((\hat{\xi}_{s}(k))_{0\leq s \leq t})}}\right\}_{k\in\S}=\left\{\frac{1}{\Phi_{t}(k)}\right\}_{k\in\S}, \nonumber$$ This follows because the process $(1/\Phi_{t}(k))$ is a $\P$-martingale; $\E\left[1/\Phi_{t}(k)|\hat{\xi}_{s}(k)\right]=1/\Phi_{s}(k)$ for $s<t$, and also $\Phi_{0}(k)=1$ and $\Phi_{t}(k)>0$. In particular, $$\begin{aligned}
(\Phi_{t}(k))^{-1}=\exp\left(-\int^{t}_{0}\frac{\E\left[X|\hat{\xi}_s(k)\right]}{\hat{\sigma}_s(k)(1-s)}\d W_{s}(k)-\frac{1}{2}\int^{t}_{0}\frac{\E\left[X|\hat{\xi}_{s}(k)\right]^2}{\hat{\sigma}_s^2(k)(1-s)^{2}}\dd s \right), \nonumber\end{aligned}$$ and the Novikov’s condition $$\begin{aligned}
\E\left[\exp\left(\frac{1}{2}\int^{t}_{0}\frac{\E\left[X|\hat{\xi}_{s}(k)\right]^2}{\hat{\sigma}^2_s(k)(1-s)^{2}}\dd s \right)\right]<\infty,\end{aligned}$$ is satisfied. Under the measure $\mathbb{B}$, the random variable $\hat{\xi}_{t}(k)$ is Gaussian and we have $$\begin{aligned}
C_{0}&=P_{0t}\sum_{k\in\S}\P(\boldsymbol{\tilde{J}}_{t} = k)\E^{\mathbb{B}}\left[\left(\int_{\mathbb{X}}(P_{t1}x-K)\chi_{t}(x;k)\nu(\dd x)\right)^{+}\right]. \nonumber\end{aligned}$$ The statement follows by computing the critical value $\hat{\varsigma}(k)$.
[99]{} Amendinger, J., Imkeller, P., Schweizer, M., *Additional Logarithmic Utility of an Insider*. Stochastic Processes and their Applications 75, 263-286 (1998) Ankirchner, S., Dereich, S., Imkeller, P., *The Shannon Information of Filtrations and the Additional Logarithmic Utility of Insiders*. The Annals of Probability 34(2), 743-778 (2006) Ali, M. S., Silvey, S.D., *A General Class of Coefficients of Divergence of One Distribution from Another*. Journal of Royal Statistical Society B 28, 131-142 (1966) Brody, D. C., Hughston, L. P., Macrina, A., *Information-Based Asset Pricing*. International Journal of Theoretical and Applied Finance 11, 107-142 (2008) Brody, D. C., Hughston, L. P., Macrina, A., *Dam Rain and Cumulative Gain*. Proceedings of Royal Society A 464, 1801-1822 (2008) Brody, D. C., Davis, M. H. A., Friedman, R. L., Hughston, L. P., *Informed Traders*. Proceedings of Royal Society A 465, 1103-1122 (2009) Brody, D. C., Hughston, L.P., Macrina, A., *Modelling Information Flows in Financial Markets*, Advanced Mathematical Methods for Finance, G. Di Nunno, B. Oksendal eds., Springer (2011) Bollen, N.P.B., Gray, S.F., Whalley, R.E., *Regime Switching in Foreign Exchange Rates: Evidence from Currency Option Prices*. Journal of Econometrics 94, 239-276 (2000) Cont, R., Tankov, P., *Financial Modelling with Jump Processes*. Chapman and Hall/CRC Financial Mathematics Series (2004) Crandall, M. G., Lions, P. L., *Viscosity Solutions of Hamilton-Jacobi Equations*. Transactions of the American Mathematical Society 277(1), 1-42 (1983) Csiszar, I., *Information-Type Measures of Difference of Probability Distributions and Indirect Observations*. Studia Sci. Math. 2, 299-318 (1967) Emery, M., Yor, M., *A Parallel between Brownian Bridges and Gamma Bridges*. Publications of the Research Institute for Mathematical Sciences 40, 669-688 (2004) Fitzsimmons, P. J., Pitman, J., Yor, M., *Markovian Bridges: Construction, Palm Interpretation, and Splicing*. Seminar on Stochastic Processes 33, 102-133 (1993) Hamilton, J. D., *A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle*, Econometrica 57, 357-384 (1989) Hoyle, E., Hughston, L.P., Macrina, A., *[[Lévy ]{}]{}Random Bridges and the Modelling of Financial Information*. Stochastic Processes and Their Applications 121(4), 856-884 (2011) Hughston, L.P., Macrina, A., *Information, Inflation, and Interest*. Banach Centre Publications, Institute of Mathematics, Polish Academy of Sciences, in L. Stettner, Editor, Advances in Mathematics of Finance 83, 117-138 (2008) Jacod, J., *Grossissement Initial, Hypothese (H’) et Theoreme de Girsanov*. In: Jeulin, T. and Yor, M. (eds) Grossissements de Filtrations: Exemples et Applications, Lecture Notes in Mathematics 1118, Springer, Berlin (1980) Leon, J.A., Navarro, R., Nualart, D., *An Anticipating Calculus Approach to the Utility Maximization of an Insider*, Math. Finance 13 (2003) Merton, R. C., *Option Pricing when Underlying Returns are Discontinuous*. Journal of Financial Economics 3(1-2), 125-144 (1976) Naik, V., *Option Valuation and Hedging Strategies with Jumps in the Volatility of Assets Returns*. Journal of Finance 48, 1969-1984 (1993) Rutkowski, M., Yu, N., *On the Brody-Hughston-Macrina Approach to Modelling of Defaultable Term Structure.* I. J. of Theoretical and Applied Finance 10, 557-589 (2007) Yor, M., *Grossisement de Filtration et Absolue Continuité de Noyaux.*. In: Jeulin, T. and Yor, M. (eds) Grossissements de Filtrations: Exemples et Applications, Lecture Notes in Mathematics 1118, Springer, Berlin, (1980)
[^1]: https://www.ft.com/content/708ec068-9a62-11e8-ab77-f854c65a4465
[^2]: https://www.ft.com/content/e6fb1c9e-9cb2-11e4-a730-00144feabdc0
[^3]: https://www.ft.com/content/3b4f6c14-9c9a-11e4-971b-00144feabdc0
[^4]: https://www.ft.com/content/161d8b5c-17f7-11e8-9e9c-25c814761640
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report on a recent study of finite-volume effects on the lowest-lying octet baryon masses using the covariant baryon chiral perturbation theory up to next-to-leading order by analysing the latest $n_f=2+1$ lattice QCD results from the NPLQCD Collaboration.'
address: |
$^1$School of Physics and Nuclear Energy Engineering, Beihang University,\
Beijing 100191, China\
$^2$Research Center for Nuclear Science and Technology, Beihang University,\
Beijing 100191, China\
$^3$State Key Laboratory of Nuclear Physics and Technology, School of Physics,\
Peking University, Beijing 100871, China\
$^4$Department of Physics, University of Stellenbosch, Stellenbosch, South Africa\
$^{*}$E-mail: [email protected]
author:
- 'Xiu-Lei Ren$^{1}$, Lisheng Geng$^{1,2,*}$ and Jie Meng$^{1,2,3,4}$'
title: 'Octet-baryon masses in finite space'
---
Introduction {#sec1}
============
Lattice QCD (LQCD) simulations have made remarkable progress in studies of strong-interaction physics (see, e.g., Refs. [@Bazavov:2009bb; @Hagler:2009ni; @Fodor:2012ll]). Recently, the ground state baryon spectrum has been calculated with 2+1 flavors and agreement with experimental data up to a few percent has been achived [@Loud:2009lt; @Aoki:2009; @Lin:2009; @Durr:2008; @Alexandrou:2009; @Aoki:2010; @Bietenholz:2010; @Beane:2011pc]. However,there are still a few obstacles in present lattice simulations [@Fodor:2012ll]. To obtain physical results, one needs to extrapolate the simulated results to the physical point, i.e., $m_{u/d}\rightarrow m_{u/d}({\rm phys.})$, $L(T)\rightarrow\infty$, $a\rightarrow 0$, where $m_{u/d}$ are the masses of $u$ and $d$ quarks, $L(T)$ is the spacial (temporal) lattice size, and $a$ is the lattice spacing.
Chiral perturbation theory (ChPT) [@Weinberg:1978kz; @Gasser:1983yg; @Gasser:1984gg; @Gasser:1987rb; @Bernard:1995dp; @Pich:1995bw; @Bernard:2007zu; @Scherer:2009bt] provides a model-independent way to perform the extrapolation in light quark masses–chiral extrapolation–and to study finite-volume effects. Due to the non-zero baryon mass in the chiral limit, baryon chiral perturbation theory (BChPT) has long suffered from the so-called power-counting-breaking problem [@Gasser:1987rb]. In order to deal with this problem, several recipes have been proposed. The most widely-used are the heavy baryon (HB) ChPT [@Jenkins:1991hb] , the infra-red (IR) BChPT [@Becher:1999ir] and the covariant baryon ChPT with extended-on-mass-shell (EOMS) scheme [@Gegelia:1999eo; @Fuchs:2003ms].
In the 2-flavor space, the finite-volume effects of nucleon masses have been studied using HBChPT and IR BChPT up to next-to-next-to-leading order (NNLO) [@AliKhan:2003cu]. The authors concluded that at NNLO relativistic ChPT can describe well the finite-volume corrections. However, a detailed study of finite-volume effects using three-flavor BChPT is still missing. In this talk we report on a first systemic study of finite-volume corrections to the masses of ground-state octet baryons using the EOMS BChPT by fitting the NPLQCD [@Beane:2011pc] LQCD data.
Theoretical Framework
=====================
Physically, finite-volume corrections can be easily understood: Because of the existence of space-time boundaries, the allowed momenta of virtual particles become discretized. In LQCD simulations of zero-temperature physics the temporal extent is generally larger than the spacial extent such that the integral in the temporal dimension can be treated as if it extends from $-\infty$ and $\infty$. As a result, only the integral in the spacial dimensions should be replaced by an infinite sum.
In a finite hypercube, the following differences are defined as the finite-volume corrections: $$\delta G_N = G_N(L)-G_N(\infty),\quad \delta G_D = G_D(L)-G_D(\infty).$$ where $G_{N/D}(L)$ and $G_{N/D}(\infty)$ denote the integrals calculated in a finite hypercube and in infinite space-time. Therefore, the octet-baryon masses at NLO EMOS BChPT in a finite box have the following form (see Ref. [@Geng:2011fvc] for the definitions of couplings $\xi^{(a,b,c)}_{\mathcal{B},\phi}$ and loop functions $H_{B}^{(b,c)}(m_{\phi})$): $$\begin{split}
M_{\mathcal{B}}&= M_0-\sum\limits_{\phi=\pi,K}\xi^{(a)}_{\mathcal{B},\phi}(b_0, b_D, b_F)\cdot m^2_{\phi} \\
&\quad+ \frac{1}{(4\pi F_0)^2}\sum_{\phi=\pi,K,\eta}
\left[~\xi_{\mathcal{B},\phi}^{(b)}(D, F)
\left(H_{B}^{(b)}(m_{\phi})+\delta G_N(L)\right)\right.\\
&\qquad\qquad \qquad \qquad \qquad \left.+\xi_{\mathcal{B},\phi}^{(c)}(\mathcal{C})
\left(H_{B}^{(c)}(m_{\phi})
+\delta G_D(L)\right)~\right].
\end{split}$$
Results and Discussions
=======================
The fitted results using both EOMS BChPT and HBChPT are shown in Fig. \[fig:res\]. Both methods provide a reasonable fit to the lattice data with similar quality ($\chi^2$/d.o.f.=1.6). The effects of the virtual decuplet baryons can be best seen by fitting the NPLQCD data with $C=0$. The corresponding results are shown by the solid (dashed) lines in Fig. \[fig:res\]. It is clear that in the fit the octet-decuplet transition plays a larger role in HB than in EOMS BChPT. In fact, in HBChPT virtual decuplet baryons play an even larger role than those of virtual octet baryons, which seems to be a bit unnatural (for a relevant discussion, see, e.g., Ref. [@Geng:2009hh]). In Ref. [@Beane:2011pc], it was concluded that the decuplet contributions must be taken into account. Our studies show that this is indeed the case, but more so in the HBChPT than in the covariant ChPT.
Using the LECs determined in the fit of the NPLQCD data, we also performed a chiral extrapolation. The EOMS extrapolations are in much better agreement with the experimental masses than the HB extrapolations (see Table III of Ref. [@Geng:2011fvc] ), which is consistent with the finding of Ref. [@MartinCamalich:2010fp].
Summary and conclusions
=======================
We have studied finite-volume corrections to the octet baryon masses by analyzing the latest $n_f=2+1$ NPLQCD data with EOMS BChPT and with HBChPT. It was shown that although both approaches can describe the lattice data reasonably well, the underlying physics is different: Decuplet contributions play a less important role in EOMS BChPT than in HBChPT at next-to-leading order because relativistic corrections enhance virtual octet contributions and reduce intermediate decuplet contributions.
Acknowledgements
================
L. S. Geng acknowledges support from the Fundamental Research Funds for the Central Universities and the National Natural Science Foundation of China (Grant No. 11005007).
[99]{}
A. Bazavov [*et al.*]{}, Rev. Mod. Phys. [**82**]{}, 1349 (2010).
Ph. Hägler, Phys. Rept. [**490**]{}, 49 (2010).
Z. Fodor, C. Hoelbling, Rev. Mod. Phys. [**84**]{}, 449 (2012).
A. Walker-Loud [*et al.*]{}, Phys. Rev. D [**79**]{}, 054502 (2009).
S. Aoki [*et al.*]{} (PACS-CS Coll.), Phys. Rev. D [**79**]{}, 034503 (2009).
H.W. Lin [*et al.*]{} (HSC Coll.), Phys. Rev. D [**79**]{}, 034502 (2009).
S. Durr [*et al.*]{} (BMW Coll.), Science 322, 1224 (2008).
C. Alexandrou [*et al.*]{} (ETM Coll.), Phys. Rev. D [**80**]{}, 114503 (2009).
S. Aoki [*et al.*]{} (PACS-CS Coll.), Phys. Rev. D [**81**]{}, 074503 (2010).
W. Bietenholz [*et al.*]{}, Phys. Lett. B [**690**]{}, 436 (2010).
S. R. Beane [*et al.*]{}, Phys. Rev. D [**84**]{}, 014507 (2011).
S. Weinberg, Physica A [**96**]{}, 327 (1979).
J. Gasser and H. Leutwyler, Ann. Phys. (N.Y.) [**158**]{}, 142 (1984).
J. Gasser and H. Leutwyler, Nucl. Phys. B [**250**]{}, 465 (1985).
J. Gasser, M. E. Sainio and A. Svarc, Nucl. Phys. B [**307**]{}, 779 (1988).
V. Bernard, N. Kaiser and U. G. Meissner, Int. J. Mod. Phys. E [**4**]{}, 193 (1995).
A. Pich, Rept. Prog. Phys. [**58**]{}, 563 (1995).
V. Bernard, Prog. Part. Nucl. Phys. [**60**]{}, 82 (2008).
S. Scherer, Prog. Part. Nucl. Phys. [**64**]{}, 1 (2010).
E. E. Jenkins and A.V. Manohar, Phys. Lett. B [**255**]{}, 558 (1991).
T. Becher and H. Leutwyler, Eur. Phys. J. C [**9**]{}, 643 (1999).
J. Gegelia and G. Japaridze, Phys. Rev. D [**60**]{}, 114038 (1999).
T. Fuchs, J. Gegelia, G. Japaridze, and S. Scherer, Phys. Rev. D [**68**]{}, 056005 (2003).
A. Ali Khan [*et al.*]{} (QCDSF-UKQCD Coll.), Nucl. Phys. B [**689**]{}, 175 (2004).
L.-S. Geng, X.-L. Ren, J. Martin-Camalich, and W. Weise, Phys. Rev. D [**84**]{}, 074024 (2011).
L. S. Geng, J. Martin Camalich and M. J. Vicente Vacas, Phys. Lett. B [**676**]{}, 63 (2009).
J. Martin Camalich, L. S. Geng and M. J. Vicente Vacas, Phys. Rev. D [**82**]{}, 074504 (2010).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The events of multiple neutron production under 2000g/cm$^2$ thick rock absorber were studied at the Tien Shan mountain cosmic ray station, at the altitude of 3340m above the sea level. From comparison of the experimental and Geant4 simulated neutron multiplicity spectra it follows that the great bulk of these events can be explained by interaction of cosmic ray muons with internal material of the neutron detector. In synchronous operation of the underground neutron monitor with the Tien Shan shower detector system it was found that the characteristics of the muonic component of extensive air showers which is seemingly responsible for generation of the neutron events underground do change noticeably within the energy range of the knee of primary cosmic ray spectrum. Some peculiar shower events were detected when the neutron signal reveals itself only $\sim$(100–1000)$\mu$s after the passage of the shower particles front which probably means an existence of corresponding delay of the muon flux in such events.'
address:
- 'P.N.Lebedev Physical Institute of the Russian Academy of Sciences (LPI), Leninsky pr., 53, Moscow, Russia'
- 'Institute of Ionosphere, Kamenskoye plato, Almaty, Kazakhstan'
- 'Al-Farabi Kazakh National University, Institute for Experimental and Theoretical Physics, al-Farabi pr., 71, Almaty, Kazakhstan'
- 'Satbayev University, Institute of Physics and Technology, 050032, Ibragimova str. 11, Almaty, Kazakhstan'
author:
- 'A.Shepetov'
- 'A.Chubenko'
- 'O.Kryakunova'
- 'O.Kalikulov'
- 'S.Mamina'
- 'K.Mukashev'
- 'R.Nam'
- 'V.Piscal'
- 'V.Ryabov'
- 'T.Sadykov'
- 'N.Saduev'
- 'N.Salikhov'
- 'E.Tautaev'
- 'L.Vildanova'
- 'Zh.Zhantayev'
- 'V.Zhukov'
title: 'Underground neutron events at Tien Shan and the properties of the $10^{14}-10^{17}$eV EAS muonic component'
---
=4
cosmic rays,extensive air shower,EAS ,muon
Introduction
============
Investigation of the events of multiple neutron production in underground detectors of the Tien Shan mountain cosmic ray station was started about a decade ago. Phenomenological results which have been obtained then on the properties of these events were reported in [@undgour1; @undgour2; @undgour2008icrcmexico]; in these publications it was stated that the original nature of neutron events observed underground still remains unclear. After completion of modification period of the complex detector system for the cosmic ray studies at Tien Shan station [@ontien-nim2016], and after systematic introduction of the Geant4 package based simulation methods for determination of the properties of its particle detectors it became possible to return to the problem of underground neutron events at new stage of experimental technique. In particular, the detection of neutron events in strict synchronization with the shower installation now permits to study precisely the neutron bearing properties and temporal characteristics of the penetrative component of extensive air showers (EAS). An overview of the results newly obtained in simultaneous operation of the Tien Shan shower detector system with the underground neutron detector is subject of the present message.
The underground neutron monitor
===============================
{width="60.00000%"} {width="40.00000%"}
The data discussed further on were obtained in the years 2012–2018 in a long term experiment with the neutron detector which is placed beneath a 2000g/cm$^2$ thick absorber, at altitude of 3340m above the sea level in the underground room of the Tien-Shan mountain cosmic ray station. Presently, this detector consists of a pair of separated units—*UPPER* and *LOWER*, both of which were made resembling the standard NM64 type neutron supermonitor [@carmichel_supermonitor]. In the underground room these detector units, which thus can be referenced as the the *underground monitor*, are placed one above the other as it is shown in the left picture frame of figure \[figiundgeffici\]. Both units include the layers of heavy target absorber where penetrative particles of cosmic radiation can experience nuclear interaction with lead nuclei. Evaporation neutrons which originate as a result of this interaction can be detected by the big $\diameter 150\times 2000$ mm$^2$ neutron counters with enriched $^{10}$BF$_3$ gas filling, so the detection of low-energy neutrons is possible there due to the reaction $n(^{10}$B,$^7$Li$)\alpha$. Before detection, the neutrons loose their initial MeV-order kinetic energy down to thermal level in multiple interactions with light nuclei within the sheets of internal moderator material which consists of the wooden boxes surrounding the counters. Another sheets of hydrogen enriched rubber (C$_2$H$_2$) which cover all the unit from outside play the role of external shielding to prevent the influence of environmental low energy neutrons background on the measurement of the cosmic ray connected neutron signal.
Hence, the response of the neutron monitor unit to interaction of a nuclear-active cosmic ray particle is connected with a number of electric pulse signals obtained from its counters during some fixed *time gate* period after this interaction. Hereafter, sum number of such signals will be designated as the neutron *multiplicity* $M$. In [@nmn2003] it was shown experimentally that for a monitor unit of considered construction the multiplicity $M$ is nearly proportional to the square root of energy deposit in primary interaction (in the GeV order energy range), and the typical duration of the gate time can be of the order of a few milliseconds. More precisely, the energy dependence $M(E)$ can be defined through the detailed simulation of primary interaction and subsequent moderation and diffusion processes of originating neutrons within the monitor material which can be made on the basis of the modern Geant4 toolkit [@yanke2011]. The results of such calculations for the hadronic (neutron) and muonic type primaries are presented in the right plot of figure \[figiundgeffici\].
{width="\textwidth"}
{width="\textwidth"}
Same Geant4 simulation model can be of use to answer the question which was put in [@undgour1] on nature of the underground neutron events. Taking into consideration the spectrum of energy deposits from penetrative cosmic ray particles which has been measured earlier in the same underground room of Tien Shan station in experiments with the ionization calorimeter [@ontienmuons_erlykin1973], and using it as the input for the Geant4 simulation series with $\mu^\pm$ type primaries one can obtain the expected neutron multiplicity spectrum of underground events. A comparison of such simulated and experimentally measured multiplicity spectra is made in the plot of figure \[figiundgsimispe\] where it is seen that both spectra do agree rather well with each other. From this fact a conclusion can be drawn that in the case of underground monitor we deal mostly with the products of nuclear interaction caused by cosmic ray muons, so this monitor as a whole can be used as a specific detector of the cosmic ray muonic component, and particularly of the muons which accompany the passage of extensive air showers.
Underground neutron events and extensive air showers
====================================================
{width="49.00000%"} {width="49.00000%"}
During operation periods of the Tien Shan shower installation [@ontien-nim2016] in the years 2015–2018 the data taking process at the underground neutron monitor was fulfilled under the control of external trigger signal which was generated in the moments when an extensive air shower had been detected above the surface of Tien Shan station. The trigger generation algorithm ensured a nearly (90–95)% selection probability of the EAS with primary energy above $E\gtrsim 10^{15}$eV whose axes were coming through the central part of shower installation, at a distance of $R\leqslant(25-30)$m from the underground detector system. The detection of smaller EAS up to $E_0\approx 10^{14}$eV was possible as well at that time with somewhat reduced efficiency. The upper energy limit of detected EASs is defined mostly by the geometrical area of the shower detector installation, and by the total duration time of the measurements. For the dataset presented here its value is about $10^{17}$eV.
The multiplicity spectra of the neutron events of muonic origin which have been registered simultaneously with an EAS are presented in figure \[figiundgeasspe\], in comparison with analogous spectra calculated over the whole set of underground events. It is seen there that generally the intensity of EAS accompanied events occurs being 2–3 orders of magnitude below the total flux of the $M\lesssim100$ events, but tends to match with the latter in the range of extremely high multiplicities. Seemingly, this difference can be explained by particulars of the shower trigger elaboration threshold and their influence on registration probability of the EAS connected neutron events underground.
To clarify more precisely what are the favorable conditions for detection of the muonic EAS component by neutron signal from its nuclear interaction, a correlation plot can be built between the multiplicity of underground neutron events $M$, and the core distance $R$ and the size (total number of the charged particles) $N_e$ of accompanying showers. An example of such correlation is presented in the left plot of figure \[figiundgeascounts\]. The range of distance values here is limited from above with geometrical sizes of the central detector “carpet” of the Tien Shan shower installation ($R^{max}\simeq 36$ m, see [@ontien-nim2016]) which was used for precise location of the EAS axis in the measurement series when the considered experimental data were obtained. From correlation plots of figure \[figiundgeascounts\] it is seen that the events with non-zero multiplicity values were observed mostly in the cases when a shower axis had passed in relative vicinity ($R\lesssim 20-25$ m) to the projection of the neutron detector position to the surface of the ground, but the distance between the EAS core and the location of the monitor in such events can be somewhat bigger for the showers with $N_e\gtrsim 10^6$. As well, the events with the comparatively high values of detected neutron multiplicity ($M\gtrsim10-30$) were met only amongst the EAS with $N_e\gtrsim 10^6$.
The dependence between the relative amount of EAS events with non-zero multiplicity of detected underground neutron signals and the average size of accompanying shower $N_e$ is illustrated by the right plot of figure \[figiundgeascounts\]. In this plot the average number of shower events with $M\geqslant 1$ is normalized to the total statistics of registered EAS with given $N_e$, $\mathcal{N}_{EAS}$. As it follows from this picture, in the showers with $N_e\lesssim 10^6$ the share of $M\geqslant 1$ events remains at an approximately one and the same low level. The latter nearly fits zero, and virtually it can be explained by random coincidences of shower trigger with background neutron events in the monitor. Contrary to this, in the range of shower sizes $N_e\approx 10^6$ the relation $\mathcal{N}_{M\geqslant 1} / \mathcal{N}_{EAS}$ starts to grow, and evidently it can not be put here into any agreement with its previous behaviour.
The revealed sharp increase in relative share of the showers which were accompanied by the underground neutron events means corresponding rise in the average multiplicity, or in the energy, or in both of the EAS connected muons which constitute the original source for generation of these events. As it follows from figure \[figiundgeascounts\], such a change of the average shower characteristics resides somewhere in vicinity of the $N_e\approx 10^6$ EAS size value.
It should be noted that at altitude of the Tien Shan mountain station the range of EAS sizes $N_e\approx 10^6$ corresponds to the primary energy of a cosmic ray particle $E_0\approx 3\cdot 10^{15}$ eV [@ontien_icrc1987__e0_through_ne_ru], i.e. to position of the well-known knee in the energy spectrum of cosmic rays. Since the newly found deviation of the mean characteristics of EAS connected muonic component occurs just at the same point on energy scale, probably it could be added to a wide list of the various peculiar effects which have been met up to date within this energy range by many research groups, and particularly in the experiments which have been held previously at the Tien Shan cosmic ray station [@ontien-nim2016].
{width="49.00000%"} {width="49.00000%"}\
{width="49.00000%"} {width="49.00000%"}
Another peculiarity observed in behaviour of the underground neutron events concerns their temporal characteristics. Since the shower trigger provides a rather strict binding to the EAS passage it is possible to trace precisely the time distribution of subsequent neutron signals with regard to initial moment of nuclear interactions which take place within detector unit at the hitting moment of EAS muons. It is natural to expect that generally such distributions must be of an exponent type which results from diffusion of the newly born evaporation neutrons within the monitor material. Indeed, most part of detected events do have such exponential shape, like a sample event shown in the top left panel of figure \[figiundgeasdelays\]. At the same time, among all EAS connected events it was found a noticeable amount of cases when the beginning of exponential intensity decrease has been too late by a some considerable time $T$ in relation to the moment of shower trigger, so that $T$ can be of up to a few hundreds of microseconds order. An example of a such “delayed” event is presented in the top right panel of figure \[figiundgeasdelays\].
Of course, certain part of the uncommon delays seen underground can be explained by arbitrary overlapping between the moments of EAS trigger and the background neutron events in monitor. To elucidate the role of random coincidences in the discussed effect a distribution can be built over observed values of the delay times $T$. Such distribution (normalized to the total amount of registered EAS cases with a non-zero neutron accompaniment underground) is presented in two bottom plots of figure \[figiundgeasdelays\] which were drawn for the events with the detected neutron multiplicity $M\geqslant 3$ and $M\geqslant 5$ correspondingly. As it is seen in these plots, the probability to find an overtaking signal from neutron detector in the negative range of delay times $T<0$ is about $P_{min}\approx (2-4)\cdot 10^{-4}$ $\mu$s$^{-1}\cdot$event$^{-1}$. Evidently, such pulses must be causally independent of any succeeding EAS, and it is this $P_{min}$ level which can be accepted as minimum background of random coincidences between the extensive air showers and underground neutron events.
Nevertheless, an evident and statistically reliable excess of the detected events number above $P_{min}$ is seen in both plots of figure \[figiundgeasdelays\] in the range of delay times between $T=0$ and $T\approx 500-800$ $\mu$s. This means that some noticeable part of detected delays can not be explained completely by accidental coincidences but must have some physical reason in EAS interaction properties. Since the primary source of underground neutron events is interaction of penetrative muons, same conclusion relates as well to the flux of the muonic component in extensive air showers.
Conclusion
==========
Presently, the investigation results of neutron generation in the underground monitor at Tien Shan station can be summarized as the following.
- The comparison of the experimental and Geant4 simulated neutron multiplicity spectra has shown that the neutron events observed underground can mostly be explained by interaction of the cosmic ray muons with internal material of neutron detector.
- In synchronous operation of the underground neutron monitor with the Tien Shan shower installation it was found that the origination frequency of neutron events starts to grow significantly around the knee of primary cosmic ray spectrum. Consequently, same conclusion can be made on either the average energy of EAS connected muons which are original source of neutron generation, or on the mean muon multiplicity in the above-the-knee EASs, or on both these characteristics.
- Some peculiar EAS events were detected in which the neutron signal underground reveals itself only a few hundreds of microseconds after the trigger which is generally generated at the passage moment of an EAS front. This circumstance means an existence of corresponding delay of the EAS connected muon flux in relation to the main bulk of shower particles in these events.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the grants \#BR05236201 and \#BR05236494 of IRN Program “Fundamental and applied studies in related fields of the physics of terrestrial, near-Earth and atmospheric processes and their practical application”, and by the grant \#0118RK00800 of the Cosmic Program of Kazakhstan Republic.
References {#references .unnumbered}
==========
[1]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} Chubenko A P, Shepetov A L, Vildanova L I, et al 2007 [*[Bull. Lebedev Phys. Inst.]{}*]{} [**34**]{} 107–113
Chubenko A P, Shepetov A L, Oscomov V V and Vildanova L I 2008 The underground neutron events at [Tien-Shan]{} [*[Proceedings of the 30th ICRC]{}*]{} vol 4 (HE part 1) (México City, México) pp 3–6
Chubenko A P, Shepetov A L, Chubenko P A, et al 2008 The underground neutron calorimeter for registration of the neutron-bearing cosmic ray component at [Tien Shan]{} [*[Proceedings of the 30th ICRC]{}*]{} vol 4 (HE1) (México City, México) pp 97–100
Chubenko A P, Shepetov A L, Antonova V P, et al 2016 [*[Nucl. Instrum. Methods A]{}*]{} [ **832**]{} 158–178
Carmichael H and Hatton C J 1964 [*[Can. J. Phys.]{}*]{} [**42**]{} 2443
Chubenko A P, Shepetov A L, Antonova V P, et al 2003 Multiplicity spectrum of [NM64]{} neutron supermonitor and hadron energy spectrum at mountain level [*[Proceedings of the 28th ICRC]{}*]{} ([Tsukuba, Japan]{}) pp 789–792
Abunin A A, Pletnikov E V, Shchepetov A L and Yanke V G 2011 [*[Bull. Russ. Acad. Sci. Phys.]{}*]{} [**75**]{} 866–868
Erlykin A D, Kulichenko A K, Machavariani S K and Nikolsky S I 1973 Investigation of cascades, produced by high energy muons [*Proceedings of the 13th [ICRC]{}*]{} vol 3 ([Denver, Colorado, USA]{}: [NASA Conference Publication No. 2376]{}) p 1803 (*Preprint* )
Adamov D S, Afanasjev B N, Arabkin V V, et al 1987 Phenomenological characteristics of [EAS]{} with [$N_e=2\cdot 10^5-2\cdot 10^7$]{} obtained by the modern [Tien-Shan]{} installation [“Hadron”]{} [*Proceedings of the 20th [ICRC]{}*]{} vol 5 ([Moscow, USSR]{}) pp 460–463
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Целью настоящей работы является более детальное рассмотрение недавно обнаруженного необычного эффекта сияния цветового заряда и его возможных физических проявлений, связанных с рассмотрением ансамблей частиц с цветовым зарядом на классическом уровне. Исследуются способы появления данного эффекта в произвольных системах точечных массивных частиц, которые могут быть разбиты на элементарные конфигурации. Также рассматривается возможное влияние этого эффекта на динамику частиц (в частности, на функцию распределения глюонов). Для естественной регуляризации выражений рассматриваются столкновения с заданным прицельным параметром. Показано, что в случае разумных значений прицельного параметра столкновения могут проходить в “электродинамическом” режиме, при этом вклад “сияния” зарядов в напряженности полей оказывается подавленным по сравнению с электродинамической картиной. Исходя из анализа ситуации с “цветовым эхо” следует, что сделанный вывод оказывается справедливым и для более сложных конфигураций частиц, так как жесткие глюонные поля могут генерироваться только в результате непосредственного столкновения, а не за счет каких-либо эффектов типа “эха”.'
author:
-
title: Классические глюонные поля релятивистских цветовых зарядов
---
ВВЕДЕНИЕ
========
Классические решения неабелевых полевых уравнений и их свойства неоднократно вызывали интерес в связи с их нелинейным характером в отличие от уравнений классической электродинамики. Немаловажную роль в этом сыграло и явление конфайнмента кварков. Однако, в отличие от электродинамики, напряженности неабелевых полей не являются непосредственными наблюдаемыми величинами, и более адекватным становится описание на языке квантовой теории (КХД). Классические поля можно трактовать как средние значения соответствующих операторов по некоторому состоянию в пределе больших квантовых чисел.\
В рамках другого подхода, связанного с разложением по константе связи, справедливого при высоких энергиях, рассматриваются вопросы, касающиеся неабелевого излучения, например, аналога синхротронного в электродинамике [@tuchin; @dbeyssi].\
В последнее время интерес к классическим конфигурациям полей возрос в связи с необходимостью использования транспортных кодов (где эти конфигурации могут быть включены, см., например, модификацию транспортного кода HSD [@toneev], которая учитывает электромагнитные поля на классическом уровне) для описания динамики систем в релятивистских столкновениях тяжелых ионов. В задачах релятивистских столкновений широкой популярностью пользуется модель так называемого конденсата цветового стекла (CGC), в которой квазиклассические поля являются источником глюонных конфигураций в переменных светового конуса (см., например, [@mclerran] или обзор [@muller]).\
Недавно было обнаружено [@dipodip], что при лобовых столкновениях ультрарелятивистских частиц с открытым цветом глюонные конфигурации напоминают электродинамические, однако сигнал от момента встречи существенно отличается. Именно, область наблюдения порядка нескольких ферми на короткое время оказывается заполнена глюонными полями огромной амплитуды (так называемое сияние цветового заряда), напоминающими ударную волну. Этот эффект не улавливается, например, в модели CGC в связи с тем, что работа ведется на световом конусе. Соответствующие полевые конфигурации могут быть важны при анализе процессов, происходящих на ранней стадии формирования кварк-глюонной плазмы, когда рождается большое количество кварк-антикварковых пар, моделью которых может выступать цветовой диполь. В [@dipoles] исследовались поля Вайцзеккера – Вильямса в квазиклассическом приближении для ультрарелятивистской системы “частица-диполь”, после чего результат сравнивался с аналогичным из CGC. Было показано, что в случае ориентированного вдоль направления движения диполя предел скоростей, близких к скорости света ($v \to 1$), существует для усредненных по времени столкновения с диполем полевых конфигураций. При этом усредненные значения напряженностей полей оказываются несингулярными, и функция распределения качественно совпадает с результатом работы [@struct]. Однако в [@dipoles] не учитывался эффект сияния цветового заряда, так как он появляется в следующем порядке разложения по константе связи.\
В [@dipodip] исследовались напряженности глюонных полей для трех простейших базовых конфигураций: лобовые столкновения систем типа “частица-частица”, “частица-диполь”, “диполь-диполь”. Регуляризация выражений (в силу отсутствия прицельного расстояния) производилась путем обрезания решений на искусственном временном масштабе $t _{min}$ и последующей сшивкой. Обнаруживается так называемое “сияние цветового заряда”, обусловленное вращением изовекторов зарядов в изотопическом пространстве, которое усиливается при переходе к более сложным конфигурациям. Сигнал от этого эффекта длится достаточно малые промежутки времени порядка $10^{-4}$ Фс, но охватывает всю область наблюдения порядка нескольких ферми. В силу приближенности решения не ухватывается эффект “цветового эха”, который заключается в приходе к третьей частице сигнала о вращении с момента встречи двух частиц.\
Дабы несколько прояснить ситуацию, будем рассматривать задачу, аналогичную поставленной в [@dipodip], но с заданным прицельным параметром, выступающим в качестве естественного регуляризатора выражений.
Поле классического заряда в неабелевой теории
=============================================
Для начала напомним, как выглядит ситуация в абелевой теории - электродинамике.\
В электродинамике поле точечного заряда $e$, движущегося по траектории $\mathbf{r}(t)$ со скоростью $\mathbf{v}(t)$, выражается при помощи потенциалов Лиенара – Вихерта [@landau] $$\label{Lienar}
\varphi = \frac{1}{4\pi}\left [\frac{e}{R - \mathbf{v}\mathbf{R}} \right ]_{t'} , \ \ \mathbf{A} = \frac{1}{4\pi} \left [ \frac{e\mathbf{v}}{R - \mathbf{vR}} \right ]_{t'} \ .$$ Запаздывающее время $t'$ определяется из уравнения $$\label{delay}
R(t') = t-t' \ ,$$ где $R = | \mathbf{R} |$ – расстояние от точки наблюдения до частицы.\
Напряженности электрического и магнитного полей даются выражениями: $$\label{EH}
\mathbf{E} = -\frac{\partial \mathbf{A}}{\partial t} - \mathbf{\nabla} \varphi , \ \ \mathbf{H} = \mathbf{\nabla} \times \mathbf{A}.$$ Из (\[EH\]) с помощью (\[Lienar\]), (\[delay\]) получают: $$\begin{gathered}
\mathbf{E} = \frac{1}{4\pi} \left [ \frac{e}{R^2} \frac{(1-v^2) (\mathbf{n} - \mathbf{v})}{(1-\mathbf{v}\mathbf{n})^3} + \frac{e}{R} \frac{\mathbf{n} \times (\mathbf{n} - \mathbf{v}) \times \dot {\mathbf{v}} }{(1-\mathbf{v}\mathbf{n})^3} \right ]_{t'} \ ,\\
\mathbf{H} = \mathbf{n} \times \mathbf{E}. \nonumber\end{gathered}$$
Естественным образом встает вопрос о возможности подобного описания систем из релятивиcтских цветовых зарядов, т. е. использования аналогов потенциалов Лиенара – Вихерта в неабелевой теории. При этом обнадеживает то, что для одной частицы во всех порядках по константе взаимодействия $g$ из уравнений Янга – Миллса получается кулоновское решение. Для системы из двух частиц в нулевом приближении по $g$ имеют место аналогичные кулоновские решения. Для интересующих релятивистских систем эффективная константа связи $\alpha_g \sim 0.4$, что позволяет строить ряды теории возмущений, используя только несколько первых порядков.
Для теории с лагранжианом (для простоты рассматриваем группу $SU(2)$) $$\mathcal{L} = -\frac{1}{4} \widetilde{{G_{\mu \nu}}} \widetilde{{G^{\mu \nu}}} - \widetilde{{j^\mu}} \widetilde{{A_\mu}} \ ,$$ где $\widetilde{A}_\mu = \left (A_\mu ^ 1, A_\mu ^2, A_\mu ^3 \right )$, тензор поля $\widetilde{{G_{\mu \nu}}} = \partial _\mu \widetilde{A_\nu} - \partial _\nu \widetilde{A_\mu} + g \widetilde{A_\mu} \times \widetilde{A_\nu}$, ковариантная производная действует по правилу $\widetilde{D^\mu} \widetilde{f} = \partial ^\mu \widetilde{f} + g \widetilde{A^\mu} \times \widetilde{f}$. Уравнения поля $$\widetilde {D^\mu} \widetilde{G_{\mu \nu}} = \widetilde{j_\nu}$$ для системы из двух цветовых зарядов $\widetilde{P}, \widetilde{Q}$ факторизуются, и удается получить так называемые уравнения хромостатики [@khriplovich] $$\begin{gathered}
\label{chromostat}
\mathbf{D}\mathbf{D} \Phi = \delta, \\
\nabla \times \nabla \times \mathbf{a} = g \mathbf{j} \ , \mathbf{j} = \Phi J \mathbf{D} \Phi. \nonumber\end{gathered}$$ Здесь $\Phi = \varphi - \star{\varphi}$ – двухкомпонентный столбец, где $\varphi_1 = \varphi_1 (\mathbf{x}), \varphi_2 = \varphi_2(\mathbf{x}), \star{\varphi_1} = \varphi_1(\mathbf{x_2}), \star{\varphi_2} = \varphi_2(\mathbf{x_1})$, $\delta = || \delta (\mathbf{x} - \mathbf{x_1}) \ , \ \delta (\mathbf{x} - \mathbf{x_2}) || $, ковариантная производная определена как $\mathbf{D}_{kl} = \nabla \delta_{kl} + g \mathbf{a} C_{kl}, k, l = 1,2$, а $C$ и $J$ есть матрицы: $$C = \begin{pmatrix}
-(\widetilde{P} \widetilde{Q}) \ \ \ -(\widetilde{Q} \widetilde{Q}) \\
(\widetilde{P} \widetilde{P}) \ \ \ (\widetilde{P} \widetilde{Q})
\end{pmatrix}, \ \ J = \begin{pmatrix}
0 \ \ 1 \\
-1 \ \ 0
\end{pmatrix}.$$ При этом векторное поле $\mathbf{a}$ оказывается натянутым на вектор $\widetilde{P} \times \widetilde{Q}$ ($\widetilde{P}, \widetilde{Q}$ – изовекторы цветовых зарядов частиц), а скалярное $\phi$ – на векторы $\widetilde{P}$ и $\widetilde{Q}$. Физический смысл системы уравнений (\[chromostat\]) состоит в том, что порожденное системой из двух зарядов глюонное поле само является источником заряда. Более того, оно также является и источником “тока”. Из данной системы уравнений можно извлечь, что векторное поле $\mathbf{a}$, генерируемое скалярной компонентой $\Phi$, находится в следующем порядке по константе связи по сравнению с $\Phi$. Система уравнений (\[chromostat\]) была исследована как аналитически, так и численно [@2body], и было показано, что для интересующих констант связи $g^2/4\pi < \sqrt{2}$ решение хорошо аппроксимируется кулоновскими потенциалами $\varphi$, при этом векторное поле напоминает поле постоянного магнита.\
Первое уравнение (\[chromostat\]) можно преобразовать к виду: $$\mathbf{}\nabla \mathbf{E} = \delta - g \mathbf{a} C \mathbf{E},$$ где столбец хромоэлектрического поля определяется как $\mathbf{E} = \mathbf{D} \Phi$. Дополнительное к $\delta$ слагаемое в правой части можно трактовать как плотность заряда порожденного глюонного поля $$G = -g \mathbf{a} C \mathbf {E}.$$ Так как $\widetilde{P} = -\widetilde{Q}$, то $G_1 = G_2 = \mathbf{a} \mathbf{\nabla} (\Phi_2 - \Phi_1)$, $\widetilde{G} = G_1 \widetilde{P} + G_2 \widetilde{Q} = 0$. Следовательно диполь можно рассматривать как электродинамический. Даже в случае параллельных зарядов $\widetilde{P} = \widetilde{Q}$ получим, что $G_1 = -G_2 = \mathbf{a} \mathbf{\nabla} (\Phi_1 + \Phi_2)$, откуда снова получим $\widetilde{G} = 0$. В случае иных конфигураций зарядов выражение для $G$ является нетривиальным и требует аккуратного учета.\
В связи с рассмотренным выше в [@dipodip] предлагается использовать приближенное решение уравнений Янга – Миллса для двух и более частиц в виде суперпозиции “потенциалов Лиенара-Вихерта” для каждой частицы: $$\varphi = \frac{1}{4\pi}\left [\frac{\widetilde{C}}{R - \mathbf{v}\mathbf{R}} \right ]_{t'}, \ \ \mathbf{A} = \frac{1}{4\pi} \left [ \frac{\widetilde{C}\mathbf{v}}{R - \mathbf{vR}} \right ]_{t'},$$ $\widetilde{C}$ – изовектор цветового заряда частицы. В результате с помощью стандартной процедуры получаются выражения для хромоэлектрического и хромомагнитного полей движущейся частицы: $$\begin{gathered}
\label{EHNA}
\widetilde{\mathbf{E}} = \frac{1}{4\pi} \left [ \frac{\widetilde{C}}{R^2} \frac{(1-v^2) (\mathbf{n} - \mathbf{v})}{(1-\mathbf{v}\mathbf{n})^3} + \frac{\widetilde{C}}{R} \frac{\mathbf{n} \times (\mathbf{n} - \mathbf{v}) \times \dot {\mathbf{v}} }{(1-\mathbf{v}\mathbf{n})^3} + \frac{\widetilde{D}}{R} \frac{ (\mathbf{n} - \mathbf{v})}{(1-\mathbf{v}\mathbf{n})^2}\right ]_{t'}, \\
\widetilde{\mathbf{H}} = \mathbf{n} \times \widetilde{\mathbf{E}}. \nonumber\end{gathered}$$ В отличие от электродинамики здесь появляется дополнительный член следующего порядка по $g$, связанный с изменением направления изовектора заряда ${\widetilde{D}} = \partial \widetilde{C} / \partial t$ (вращения в изотопическом пространстве), следующего из закона сохранения тока (уравнений совместности): $$\label{condition}
\widetilde {D^\mu} \widetilde{j_\mu} = 0$$ (в отличие от электродинамики $\partial \mu j_\mu \neq 0$).\
В [@dipodip] были получены полуаналитические формулы для вращений изовекторов цветовых зарядов для простейших конфигураций, на которые, как представляется, возможно разложить любой сложный процесс столкновения тяжелых ионов, типа “частица-частица”, “частица-диполь” и “диполь-диполь”. Полученные графики напряженностей глюонных полей, генерируемых при лобовом столкновении данных конфигураций в целом проявляют электродинамические черты, однако в момент встречи картина существенно отличается от электродинамической – генерируемые поля имеют гигантскую амплитуду (так называемое “сияние цветового заряда”) вследствие формально бесконечно быстрого вращения изовекторов цветовых зарядов при нулевом расстоянии между точечными частицами в момент столкновения (для устранения бесконечности точка момента встречи окружалась окрестностью, на границах которой происходила сшивка решений). Также был упомянут эффект, который не учитывался при построении графиков – эффект “цветового эха”, являющийся проявлением коллективной динамики цветовых зарядов при наличии более чем двух частиц.\
Будем решать задачу, аналогичную поставленной в [@dipodip], но с учетом прицельного расстояния. Обозначим прицельное расстояние как $a$. Частицы будем считать точечными и массивными. Исследование неточечных объектов весьма затруднительно, так как в этом случае возникает рассогласование вращения заряда на самой частице.
Поле двух цветовых зарядов
==========================
Аналогично работе [@dipodip] будем считать, что взаимодействие ”начинается” с некоторого момента, когда частицы сближаются на расстояние $D \sim 1$ ферми. Такая постановка задачи является оправданной, поскольку на таких масштабах и при релятивистских скоростях константа связи невелика, что позволяет рассматривать задачу с помощью потенциалов Лиенара – Вихерта. Траектории частиц считаются фиксированными (так называемое эйкональное приближение), ускорением зарядов пренебрегается. В этом случае естественно рассматривать прицельные параметры, удовлетворяющие условию $a<D$. Выберем следующую лабораторную систему координат: положения частиц задаются как $x_1 = 0$, $x_2 = a$, $z_1 = v t$, $z_2 = -w t$, $v, w > 0$, а время отсчитывается таким образом, что в момент $t=0$ частицы сближаются в указанной системе координат на минимальное расстояние, равное $a$ (указанный момент сближения на минимальное расстояние далее для краткости условно будем называть моментом столкновения). Обозначим время сближения частиц на расстояние $D$ через $T$. Очевидно, выполняется соотношение: $$T = -\frac{\sqrt{D^2 - a^2}}{(v+w)}.$$ Далее, если не оговорено иначе, будем рассматривать два значения прицельного параметра $a$: $a = 0.5$, $a = 0.05$ (в единицах ферми). Соответственно, время будем измерять в единицах Фс. Далее рассмотрим время прихода сигнала о сближении на расстояние $D$ к частицам. Решая уравнение (\[delay\]) с нужными значениями $t'$ (в данном случае полагая $t'$ = T), можно получить: $$\begin{gathered}
t_1' = \frac{T + Tvw + \sqrt{T^2(v+w)^2 + a^2(1-v^2)}}{1-v^2}, \\
t_2' = \frac{T + Tvw + \sqrt{T^2(v+w)^2 + a^2(1-w^2)}}{1-w^2}. \nonumber\end{gathered}$$ $t_1'$, $t_2'$ – времена прихода сигнала о сближении на расстояние $D$ к первой и второй частицам соответственно. С этого момента частицы начинают взаимодействовать, т.е. векторы их зарядов начинают вращаться согласно уравненияю (\[condition\]). Далее нужно рассмотреть время прихода сигнала о начале вращения к частицам, так как это изменит вид уравнений. Обозначим соответствующий момент времени как $t''$ (он одинаков для обеих частиц): $$t^{''} = \frac{t_2' + t_2'vw + \sqrt{t_2'^2(v+w)^2 + a^2(1-v^2)}}{1-v^2}.$$ Таким образом, задача разбивается на три этапа (см. далее), что позволяет строить решение на каждом из этапов в отдельности, должным образом сшивая затем полученные решения. Приведем некоторые оценки для характерных времен в зависимости от прицельного параметра $a$ (скорости $v$, $w$ полагаем $v = w = 0.99$, как и всюду в дальнейшем):
$$\begin{gathered}
\label{estimates}
a=0.5 \ \mapsto \ T = -0.44, \ t' = 0.14, \ t'' = 28.17,\\
a = 0.05 \ \mapsto \ T = -0.50, \ t' = -0.004, \ t'' = 0.14. \nonumber\end{gathered}$$
Видно, что в случае $a=0.5$ сигнал о начале взаимодействия приходит после реальной встречи частиц, т.е. частицы начинают вращаться после столкновения (выходят на первый этап в обозначениях работы \[6\]). В обоих случаях выход на второй этап (согласованное вращение, с момента $t''$) осуществляется после столкновения, в первом случае ($a=0.5$), это происходит существенно позже момента $t=0$. Приведем также оценки для времени прихода сигнала от точки встречи (точки минимального расстояния, соответствующей $t=0$): $$t_1^0 = \frac{a}{\sqrt{1-v^2}}, \ t_2^0 = \frac{a}{\sqrt{1-w^2}},$$ $$\begin{gathered}
a=0.5 \ \mapsto \ t_1^0 = 2.51, \\
a=0.05 \ \mapsto \ t_1^0 = 0.25. \nonumber\end{gathered}$$ Таким образом, картина столкновения при, например, $a=0.5$ имеет следующий вид: частицы, имея постоянные векторы в изотопическом пространстве, сталкиваются, только после этого в момент времени $\sim~ t_1' = 0.14$ заряды начинают вращаться относительно постоянных векторов (первый этап). Затем приходит сигнал о их столкновении (минимальном расстоянии) в момент времени $\sim t_1^0 = 2.51$, что дает максимальную амплитуду $\widetilde{E_{NA}} \sim d \widetilde{C} / dt$, после чего уже через большой промежуток времени, когда сигнал о вращении гаснет, заряды выходят на второй этап. Однако в случае $a = 0.5$ выход на второй этап происходит в момент времени $t \sim 28$ Фс, что не представляет интереса, так как к этому времени частицы успевают разлететься на достаточно большое расстояние.\
Уравнение совместности (\[condition\]) для двух частиц, несущих цветовые заряды $\widetilde{P}$ и $\widetilde{Q}$ с заданным прицельным расстоянием $a$ запишется в виде (везде далее модули цветовых зарядов измеряем в единицах $g$, вынося константу связи в качестве множителя):\
$$\begin{gathered}
\label{rot}
\dot {\widetilde{P}} = \alpha_g \frac{1+vw}{\sqrt{a^2 + (t(v+w) - wt^{**}_{12})^2} + sgn(t) \ w \ \sqrt{(t^{**}_{12})^2 - a^2}} \ \widetilde{Q}(t-t^{**}_{12}) \times \widetilde{P}, \\
\dot {\widetilde{Q}} = \alpha_g \frac{1+vw}{\sqrt{a^2 + ((v+w) - vt^{*}_{21})^2} + sgn(t) \ v \ \sqrt{(t^*_{21})^2 - a^2}} \ \widetilde{P}(t-t^{*}_{21}) \times \widetilde{Q}. \nonumber\end{gathered}$$\
где времена запаздывания определяются из решения уравнения (\[delay\]). Данный вид уравнений не позволяет аналитически (хотя и в три этапа с сшивкой результатов) выписать выражения для положения зарядов в изотопическом пространстве, так как выражение, стоящее перед векторным произведением, является сложной функцией $t$. Более того, в случае не стремящегося к нулю прицельного расстояния на втором этапе ($\sim t''$) уже нельзя пренебрегать запаздыванием, что значительно усложняет получение аналитического ответа. Численное исследование задачи вследствие наличия запаздывания требует неоправданно больших затрат – запаздывающие конфигурации приходится аппроксимировать тем или иным способом, что в случае малой требуемой точности (и разумного времени выполнения для преследуемых целей) ведет к потере данных, и вращение зарядов учитывается неправильно. В результате этого сигнал с момента встречи частиц, обнаруженный в работе [@dipodip], сильно смазывается и становится трудночитаемым.\
Однако можно сделать некоторые приближения и попытаться привести задачу к виду, пригодному для получения ответа в явной форме. Например, для прицельного параметра $a=0.5$ (дальнейшее исследование будем вести с этим значением $a$) сам процесс столкновения происходит как чисто электродинамический, векторы зарядов сохраняют свое первоначальное положение. Вращение же начинается с момента $t \sim t_1' \sim 0.14$. Можно заметить, что на масштабах $t \sim t_1'$ выполняется неравенство $a^2 (1-v^2) \ll (v+w)^2 t^2$, что позволяет пренебречь прицельным расстоянием как в выражениях для времен запаздывания $t_{12}^{**}, t_{21}^*$, так и в выражении, стоящем перед векторным произведением в (\[rot\]).
Первый временной этап $t<t'$
----------------------------
До шкалы времени $t'$ векторы зарядов постоянны: $$\begin{gathered}
\widetilde{P} = \widetilde{P_T}, \ \ t<t_1', \nonumber \\
\widetilde{Q} = \widetilde{Q_T}, \ \ t<t_2'. \end{gathered}$$ Столкновение происходит при постоянных векторах, что напоминает электродинамическую картину.
Второй временной этап $t' \le t \le t''$
----------------------------------------
С учетом описанных выше упрощений уравнения совместности на данном этапе принимают вид (учтено, что $t>0$): $$\begin{gathered}
\label{rotazero}
\dot {\widetilde{P}} = \alpha_g \frac{1+vw}{t(v+w)} \ \widetilde{Q_T} \times \widetilde{P}, \\
\dot {\widetilde{Q}} = \alpha_g \frac{1+vw}{t(v+w)} \ \widetilde{P_T} \times \widetilde{Q}. \nonumber\end{gathered}$$ С помощью подстановки $\varphi = - ln \ |t|$, уравнения приводятся к виду: $$\begin{gathered}
\label{rotphi}
\widetilde{P'} = \omega \ \widetilde{Q_T} \times \widetilde{P}, \\
\widetilde{Q'} = \omega \ \widetilde{P_T} \times \widetilde{Q}. \nonumber\end{gathered}$$ Здесь штрих обозначает дифференцирование по переменной $\varphi'$, $$\omega = \alpha_g \frac{1+vw}{v+w} \ .$$ На данном этапе векторы $\widetilde{P}$ и $\widetilde{Q}$ вращаются вокруг постоянных векторов $\widetilde{P_T}$ и $\widetilde{Q_T}$. Для записи решения можно естественным образом выбрать базис для каждой из частиц. С вектором $\widetilde{P}$ связана тройка ортонормированных векторов $$\widetilde{Q_T} \ , \ \ \widetilde{n}_{P_T} = \frac{\widetilde{P_T} \times \widetilde{Q_T}}{\sin\theta} \ , \ \ \tilde{m}_{P_T} = \widetilde{Q_T} \times \widetilde{n}_{P_T},$$ где $\cos\theta = (\widetilde{P_T} \widetilde{Q_T})$. Аналогично для вектора $\widetilde{Q}$ выбираем $$\widetilde{P_T} \ , \ \ \widetilde{n}_{Q_T} = - \widetilde{n}_{P_T} \ , \ \ \tilde{m}_{P_T} = \widetilde{P_T} \times \widetilde{n}_{Q_T}.$$ Решение может быть представлено в форме $$\begin{gathered}
\widetilde{P} = \cos \theta \widetilde{Q_T} + \sin \theta \{ \cos \left[ \omega(\varphi - \varphi_1') \right] \tilde{m}_{P_T} - \sin \left[ \omega (\varphi - \varphi_1') \right] \tilde{n}_{P_T} \} , \\
\widetilde{Q} = \cos \theta \widetilde{P_T} + \sin \theta \{ \cos \left[ \omega(\varphi - \varphi_1') \right] \tilde{m}_{Q_T} - \sin \left[ \omega (\varphi - \varphi_1') \right] \tilde{n}_{Q_T} \}, \nonumber \end{gathered}$$ где использованы обозначения $\varphi_1' = - ln \ t_1', \varphi_2' = -ln \ t_2'$.
На рис. \[ch-ch\] представлены соответствующие напряженности полей, генерируемых в ходе столкновения в точке наблюдения с координатами $x = 2$ Фм, $z = 1$ Фм при значении прицельного параметра $a = 0.5$, скорости частиц $v = 0.98$, $w = 0.99$. Векторы зарядов частиц единичные и задаются с помощью направляющих углов в изотопическом пространстве. Для первой частицы имеем $\phi_1 = \pi/20, \theta_1 = 0$, для второй - $\phi_2 = -\pi/20, \theta_2 = -\pi/1.95$, $\widetilde{P} = \left( \cos\phi_1 \sin\theta_1, \sin\phi_1 \sin\theta_1, \cos\theta_1 \right)$, $\widetilde{Q} = \left( \cos\phi_2 \sin\theta_2, \sin\phi_2 \sin\theta_2, \cos\theta_2 \right)$ соответственно. Напряженности нормированы на $M_\pi ^2$, где $M_pi$ - масса пи-мезона. Для удобства сравнения с электродинамической картиной заряды соответствующих электродинамических частиц измеряются в единицах $g$.
Отметим главное отличие полученных результатов от аналогичных, полученных без учета прицельного параметра. Сигнал от вращения частиц (цветовое свечение зарядов) присутствует, но его амплитуда на порядки подавлена по сравнению с [@dipodip]. Значения напряженностей полей от свечения оказываются даже меньше типичных “электродинамических” (кулоноподобных). Указанная особенность объясняется ненулевым значением прицельного параметра (а значит, небесконечной частотой вращения зарядов), а также тем фактом, что вращение зарядов начинается после столкновения.
Поле цветового заряда и цветового диполя
========================================
Рассмотрим диполь с расстоянием $\delta$ между зарядами в системе отсчета покоя, движущийся со скоростью $w$ навстречу первой частице с прицельным расстоянием $a$. Заряды диполя обозначим как $\widetilde{Q_2}, \widetilde{Q_3}$. Будем считать диполь ориентированным вдоль направления движения. В лабораторной системе координат произойдет сильное лоренцево сокращение расстояния $\delta$: $\delta' = (1-w^2) \delta$. Момент столкновения первой и третьей частиц (под третьей частицей имеем в виду $\widetilde{Q_3}$) $t_3 = \delta' / (v+w)$. Для рассматриваемых прицельных расстояний приход сигналов к третьей частице описывается аналогично (\[estimates\]) со сдвигом вправо на величину $t_3$. Вращение зарядов начинается после столкновения с обеими частицами диполя. Соответственно уравнения совместности решаются аналогично уравнениям для двух цветовых зарядов. Полное выражение системы уравнений в силу громоздкости приводить не будем. Для упрощения построений можно заметить, что слагаемые, описывающие вращения векторов $\widetilde{Q_2}$, $\widetilde{Q_3}$ друг относительно друга, содержат в числителе множитель $(1-w^2)$, вследствие чего ими можно пренебречь и учитывать вращение зарядов диполя только относительно заряда $\widetilde{P}$. Результаты численного исследования задачи представлены на рис. \[ch-dip\]. Значения скоростей и углов зарядов в изотопичеcком пространстве аналогичны рассмотренным в случае столкновения двух цветовых зарядов. При этом в диполе заряды-компаньоны имеют противоположный знак.
Поле двух цветовых диполей
==========================
Для двух цветовых диполей с зарядами $\widetilde{P_1}$, $\widetilde{P_2}$, $\widetilde{Q_1}$, $\widetilde{Q_2}$, движущихся со скоростями $v$, $w$ навстречу друг другу и ориентированных по направлению движения, построение решений осуществляется аналогичным образом. Вращение зарядов в каждом из диполей рассматривается только относительно зарядов из другого диполя. Полное выражение не приводим в силу громоздкости. Численное исследование представлено на рис. \[dip-dip\]. Значения скоростей диполей и углов в изотопическом пространстве для зарядов аналогичны рассмотренным выше. Заряды-компаньоны в диполях имеют противоположный знак. В случае диполь-дипольного столкновения вклад от сияния превосходит “электродинамический”, что объясняется компенсацией “электродинамических” вкладов от зарядов-компаньонов диполей.
Некоторые возможные физические проявления
=========================================
Вообще говоря, из выражения (\[EHNA\]) видно, что новый член, зависящий от $\widetilde{D}$ = $d\widetilde{C}/dt$, убывает с расстоянием как $1/R$, перенося сигнал от столкновения на большие расстояния. Из-за зависимости типа $1/R$ он может рассматриваться в качестве возможного источника излучения. В этом разделе мы проанализируем этот вопрос.\
Нетрудно заметить, что по пространственной структуре он мало отличается от “электродинамического” (первого слагаемого в выражении (\[EHNA\]). А значит, сам по себе он излучать не может. Действительно, рассмотрим этот член (далее будем называть его $R$-членом, в силу его зависимости типа $1/R$), порождаемый одной частицей (для определенности положим $\widetilde{C} = \widetilde{P}$). Вычислим поток энергии через поверхность (учитывая, что $\mathbf{n}^2 = 1$): $$\begin{gathered}
\mathbf{S} = \widetilde { \mathbf { E } } \times \widetilde { \mathbf { H } } = \widetilde { \mathbf { E } } \times \mathbf{n} \times \widetilde { \mathbf { E } } = \mathbf{n} \widetilde {E} ^2 - \widetilde { \mathbf { E } } (\widetilde { \mathbf { E } } \mathbf { n} ), \\
(\mathbf{S}\mathbf{n}) = \widetilde { E } ^2 - (\widetilde { \mathbf { E } } \mathbf {n} )^2, \nonumber \\
\widetilde { E } ^2 = \left[ \frac{\widetilde { D } ^2}{R^2} \frac{1 - 2\mathbf {v} \mathbf {n} + v^2 }{(1-\mathbf {v} \mathbf {n})^2} \right] _{t'} , \nonumber \\
(\widetilde { \mathbf { E } } \mathbf {n} )^2 = \left[ \frac{\widetilde { D } ^2}{R^2} \frac{(1 - \mathbf {v} \mathbf {n})^2 }{(1-\mathbf {v} \mathbf {n})^2} \right]_{t'}, \nonumber \\
\widetilde { E } ^2 = (\widetilde { \mathbf { E } } \mathbf {n} )^2 \ ==> \ (\mathbf{S} \mathbf{n}) = 0. \nonumber\end{gathered}$$ Таким образом, дополнительный член, связанный с вращением одной частицы, не дает вклада в поток энергии. Учет перекрестных членов вида $\mathbf{S}^{12} = \widetilde{\mathbf{E}}^{(1)} \times \mathbf{n}^{(2)} \times \widetilde{\mathbf{E}}^{(2)}$ нетривиален, так как для этого необходимо решать задачу рассеяния частиц, но, по всей видимости, они также должны давать нулевой вклад в поток энергии в силу закона сохранения. Слагаемые, содержащие произведение $\widetilde{C} \widetilde{D}$, обладают “неправильным” знаменателем, в результате чего на больших расстояниях их можно не учитывать.\
Полученный результат не является необычным, так как вращение зарядов происходит в изотопическом пространстве и не требует затрат энергии. Однако $R$-член, являясь большим по амплитуде и убывая как $1/R$, может переносить информацию о столкновении на большие расстояния (в масштабах сильного взаимодействия), и заключенная в поле энергия может расходоваться на рождение кварк-антикварковых пар. Также он может оказывать влияние на ускорение частиц, причем более значительное по сравнению с электродинамическим (так как не подавлен в продольном направлении множителем $(1-v^2)$, однако нужно не забывать, что он – следующего порядка малости по $g$), но в силу вращения зарядов является переменным и эффективно может не учитываться.\
Помимо возможного излучения $R$-член может изменять функцию распределения глюонов (Вайцзеккера – Вильямса). Интересно сравнить функцию распределения, полученную в классическом приближении в порядке $g^3$ с аналогичной из модели CGC (по аналогии с [@dipoles]). В [@dipoles] были получены формулы: $$\begin{gathered}
\label{weizsacker}
\frac{dE}{dt} = \int \frac{d^2 \mathbf{k_\perp}}{(2\pi)^2} \frac{2g^4}{3m^2} E^2_\parallel ({k_\perp}) \ , \\
E_\parallel ({k_\perp}) = \int d \mathbf{x_\perp} e^{-i (\mathbf{k_\perp} \mathbf{x_\perp})} \langle E_\parallel \rangle \ , \end{gathered}$$ из которых можно было получить функцию распределения глюонов $$f({k_\perp}) = \frac{2g^4}{3m^2} \frac{4\pi g}{dk_\perp} \left[ 1 - \sqrt{\frac{2}{\pi}}(dk_\perp)^{1/2} K_{1/2} (dk_\perp) \right] \ .$$ Здесь прицельный параметр $a$ обозначен как $x_\perp$, усреднение ведется по времени столкновения частицы с диполем. Качественно функция совпадает с аналогичной из [@struct], однако в отличие от [@struct] не имеет инфракрасной расходимости. $R$-член, не отличающийся по своей структуре от электродинамического (а для двух частиц зависимость $1/R$ превращается в $1/R^2$, так как дополнительный множитель $1/R$ приходит из закона вращения зарядов), не вносит существенного изменения в интеграл (\[weizsacker\]). Более того, как мы видели, столкновение с диполем на временных масштабах при прицельных расстояниях $a \sim 0.5, 0.05 $ происходит вовсе без вращения зарядов. Учет вращения становится существенным при малых прицельных расстояниях, но одновременно происходит выход за рамки эйконального приближения, и формула (\[weizsacker\]) становится неприменимой.
Также стоит отметить, что если при столкновении классический радиус частицы $a = g^2/m$ меньше обратной частоты вращения ($\omega \sim 1/a$, где $\omega$ – частота вращения, $a$ – прицельное расстояния), то частица реагирует на усредненное по времени вращения (за него в качестве минимально возможного можно взять время первого этапа) поле. Усреднение такого поля связано с усреднением слагаемых типа скалярных произведений $(\widetilde{P}\widetilde{Q})$ и $\dot{\widetilde{P}} \widetilde{Q}$. В первом случае в $(\widetilde{P}\widetilde{Q})$, несмотря на вращение, всегда имеется постоянная компонента (в случае первого этапа вращения равная начальному углу между векторами зарядов), которая никогда не зануляется при усреднении. Во втором случае (вклад от члена $\dot{\widetilde{P}} \widetilde{Q}$, обусловленного нелинейностью) постоянной компоненты нет, и при усреднении стоит ожидать зануление вкладов данного типа. В частности, для кварков имеем такую ситуацию.
Цветовое эхо
============
Теперь немного подробнее исследуем появление цветового эха, являющегося некоторым проявлением коллективной динамики системы. Именно, зададимся целью выяснить, возможна ли генерация “сияния” вследствие эха.\
Прежде всего, заметим, что в случае $a = 0.5$, $a = 0.05$ эхо для системы типа “частица + диполь” не появляется вовсе, так как заряды начинают вращаться после столкновения. Только начиная с $a \sim 10^{-2}$ эффекты типа эха могут проявляться. Аккуратное изучение эффекта является сложной вычислительной процедурой, однако, сделав некоторые упрощения, можно получить результат, который, как кажется, будет характерным и более сложных случаях.\
Для анализа вновь воспользуемся выражением (\[rot\]). Будем исследовать вращение третьей частицы $\widetilde{Q_3}$ из диполя под влиянием сигнала, пришедшего от частицы с зарядом $\widetilde{P}$ с момента времени $t \to 0$, когда в случае малого $a$ и происходит максимально быстрое вращение $\widetilde{P}$. Из (\[rot\]) видно, что в окрестности точки встречи частиц ($t \approx 0$) можно упростить уравнения и привести их к виду, который позволяет написать явное аналитическое решение в удобной для дальнейшего исследования форме:\
$$\begin{gathered}
\label{rotanonzero}
\dot {\widetilde{P} }= \alpha_g \frac{(1+vw)}{a \ \sqrt{1-w^2}} \widetilde{Q}(t-t^{**}_{12}) \times \widetilde{P}, \\
\dot {\widetilde{Q}} = \alpha_g \frac{(1+vw)}{a \ \sqrt{1-v^2}} \widetilde{P}(t-t^{*}_{21}) \times \widetilde{Q}. \nonumber\end{gathered}$$\
Пренебрегая запаздыванием (если $a$ достаточно мало, оно сказывается только на фазе вращения) и разницей в скоростях $v$ и $w$, можно получить явные выражения для зарядов $$\begin{gathered}
\widetilde{P} = c \widetilde{\Omega} + s \{ \cos [\omega(t-t'')] \widetilde{m_{p^{''}}} - \sin [\omega(t-t'')] \widetilde{n_{p^{''}}} \}, \\
\widetilde{Q} =c \widetilde{\Omega} - s \{ \cos [\omega(t-t'')] \widetilde{m_{p^{''}}} - \sin [\omega(t-t'')] \widetilde{n_{p^{''}}} \}, \nonumber \end{gathered}$$ где $$\omega = \alpha_g \frac{(1+vw)}{a \ \sqrt{1-w^2}} \approx \alpha_g \ \frac{(1+vw)}{a \ \sqrt{1-v^2}} \\, c = \cos \langle \frac{(\widetilde{P^{''}} \widetilde{Q^{''}})}{2} \rangle, s = \sin \langle \frac{(\widetilde{P^{''}} \widetilde{Q^{''}})}{2} \rangle,$$ а опорные векторы определяются как $$\widetilde{\Omega} = \frac{\widetilde{P^{''}} + \widetilde{Q^{''}}}{|\widetilde{P^{''}} + \widetilde{Q^{''}}|} \ , \widetilde{n_{p^{''}}} = \frac{\widetilde{P^{''}} \times \widetilde{\Omega}}{|\widetilde{P^{''}} \times \widetilde{\Omega}|} \ , \widetilde{m_{p^{''}}} = \widetilde{\Omega} \times \widetilde{n_{p^{''}}} \ ,$$ $\widetilde{P^{''}}, \widetilde{Q^{''}}$ - векторы $\widetilde{P}$ и $\widetilde{Q}$ в момент времени $t^{''}$ (в случае малых $a$ частицы успевают выйти на этап $t \sim t''$ до столкновения). Стоит отметить, что введение ненулевого (хотя и стремящегося к нулю) прицельного параметра является естественной регуляризацией знаменателя и позволяет проводить аналитические расчеты без разрывов решения. (\[rotanonzero\]) Сигнал от вращения с огромной частотой вектора $\widetilde{P}$ со шкалы времени $t''$ приходит до встречи частиц $\widetilde{P}$ и $\widetilde{Q_3}$. При этом вращение вектора $\widetilde{Q_3}$ приближенно описывается уравнением $$\label{rotthree}
\dot{\widetilde{Q_3}} = \frac{\omega_0}{t_3} \widetilde{P} \times \widetilde{Q_3} \ ,$$ где $$\omega_0 = \alpha_g \frac{1 + vw}{v+w} \ ,$$ $t_3$ – время встречи первой и третьей частиц, а значение $\widetilde{P}$ берется со шкалы $t''$.\
Таким образом, мы имеем векторное дифференциальное уравнение вида $$\dot{\widetilde{Q_3}} (t) = A(t) \widetilde{Q_3} (t) \ ,$$ где $\widetilde{Q_3} = \left (Q_3^1, Q_3^2, Q_3^3 \right)$, а матрица $A$ имеет вид $$A = \frac{\omega}{t_3} \begin{pmatrix}
0 & -P^*_3 & P^*_2 \\
P^*_3 & 0 & -P^*_1 \\
-P^*_2 & P^*_1 & 0
\end{pmatrix} = A_1 + A_2 \cos{\Omega_0 t} + A_3 \sin{\Omega_0 t} \ ,$$ где $P^* = \left(c, \ s_1 \cos\Omega_0 t + s_2 \sin \Omega_0 t, \ -s_1 \sin \Omega_o t + s_2 \cos \Omega_o t \right)$, $$\Omega_0 = \alpha_g \frac{w}{(1 + \sqrt{1-w^2}) \ a} \frac{1+w}{1-v} \ .$$ Введенные обозначения $s_1, s_2$ определены как $$s_1 = s \cos \left[ \omega \left(t'' + \frac{\delta'}{1-v}\right) \right] \ , \ \ s_2 = s \left [\sin \omega \left(t'' + \frac{\delta'}{1-v}\right) \right].$$ Решение такого уравнения формально можно представить в виде $$\widetilde{Q_3} (t) = F(t, t_{e_{1}}) \widetilde{Q_3} (t_{e_1}) \ ,$$ где оператор перехода дается выражением $$F(t, t_{e_1}) = \int_{t_{e_1}} ^ t A(s) F(s, t_{e_1}) ds = \hat{1} + \int_{t_{e_1}} ^ t A(s_1) ds_1 + \int_{t_{e_1}} ^ t ds_1 \int_{t_{e_1}} ^ {s_1} A(s_1) A(s_2) ds_2 + ... \ ,$$ а время $t_{e_1}$ является временем прихода сигнала о вращении заряда $\widetilde{P}$ к заряду $\widetilde{Q_3}$, $t_{e_1} \sim t'$. В случае, когда $ t - t_{e_1} = n 2\pi / \Omega_0 $, все интегралы с $A_2$, $A_3$ обнуляются и мы получаем, что $$\widetilde{Q_3} (t) = e^{A_1 (t - t_{e_1})} \widetilde{Q_3} (t_{e_1}) \ .$$ В матричном виде $$\label{asympth}
\widetilde{Q_3} (t) = \begin{pmatrix}
Q_{01} \\
Q_{02} \cos{\omega_0 c \tilde{t}} - Q_{02} \sin{\omega_0 c \tilde{t}} \\
Q_{03} \cos{\omega_0 c \tilde{t}} + Q_{03} \sin{\omega_0 c \tilde{t}} \\
\end{pmatrix} \ ,$$ где $\widetilde{Q_0}$ – начальные данные к моменту прихода сигнала, $\tilde{t} = t - t_{e_1}$.\
В случае $ t - t_{e_1} = \tilde{t} \ne n 2\pi / \Omega_0 $ интегралы не обнуляются и удается получить лишь асимптотический (при $\Omega_0 \to \infty$) вид решения. Покажем, что решение является ограниченным и в пределе $\Omega_0 \to \infty$ также имеет вид (\[asympth\]). При $ \tilde{t} \ne n 2\pi / \Omega_0 $ имеем $$\Phi (t, 0) = \Phi \left(\tilde{t}, \frac{2\pi}{\Omega_0} \left\lfloor \frac{2\pi \tilde{t} }{\Omega_0} \right\rfloor \right) \Phi \left(\frac{2\pi}{\Omega_0} \left\lfloor \frac{2\pi \tilde{t} }{\Omega_0} \right\rfloor, 0\right) = \Phi \left(\tilde{t}, \frac{2\pi}{\Omega_0} \left\lfloor \frac{2\pi \tilde{t} }{\Omega_0} \right\rfloor \right) \exp{\frac{2\pi}{\Omega_0} \left\lfloor \frac{2\pi \tilde{t}}{\Omega_0} \right\rfloor A_1} \ .$$ Таким образом, для получения аналитического решения нужно знать матрицу перехода за время, меньшее периода колебаний. Так как получение точного выражения нас не интересует, мы можем качественно описать асимптотическое поведение решений. Так как $\|A(\tilde{t}) \| \le \delta$ для некоторого $\delta \ge 0$ при всех $\tilde{t}\geq0$, то $$\| \Phi \left(\tilde{t_2}, \tilde{t_1} \right) \| \leq \exp{\delta(\tilde{t_2} - \tilde{t_1})} \ , \forall \tilde{t_2} \leq \tilde{t_1} \leq 0 \ .$$ Следовательно, $$\begin{gathered}
\| \Phi \left(\tilde{t}, \frac{2\pi}{\Omega_0} \left\lfloor \frac{2\pi \tilde{t} }{\Omega_0} \right\rfloor \right) \| \leq \exp{\left( \frac{2\pi \delta}{\Omega_0} \right)} , \\
\| \widetilde{Q_3}(t) \| \leq \exp{\left( \frac{2\pi \delta}{\Omega_0} \right)} \| \widetilde{Q}(0) \| \ . \end{gathered}$$ Также производная является ограниченной при увеличении $\Omega_0$: $$\left \| \frac{d\widetilde{Q_3}}{dt} \right \| \leq \| A(t) \| \| \widetilde{Q_3}(t) \| \leq \delta \exp{\left( \frac{2\pi \delta}{\Omega_0} \right)} \| \widetilde{Q}(0) \| \ .$$ Сделанные оценки объясняют вид решения (\[asympth\]) при $\Omega_0 \to \infty$ и произвольном значении $\tilde{t}$. Соответствующие численные расчеты для значений $\omega = 0.1, \omega = 10$ представлены на рис. \[Echo1\].\
Понятно, что в случае прихода сигнала с более раннего этапа вращение также происходит с конечной частотой. Видно, что несмотря на вращение вектора $\widetilde{P}$ с некоторой (возможно большой) частотой, в целом решение является регулярным с относительно небольшой частотой порядка $\omega_0 c / t_3$, а значит, не может сопровождаться эффектом сияния, который обусловлен огромными частотами вращения (при условии $t_3 \gg a\sqrt{1-v^2}$). Мы рассмотрели только первую итерацию цветового эха, но из рассуждений видно, что дальнейшие итерации также не приведут к эффекту сияния. В рассмотренных в предыдущих разделах ситуации с прицельным параметром порядка $0.5$ Фм, уравнение, аналогичное (\[rotthree\]) описывает динамику вращения одной из частиц диполя, но на более поздних временных масштабах. Качественно понятно, что и в этом случае ситуация будет похожа на представленную выше.
Заключение
==========
Мы рассмотрели задачу столкновения элементарных конфигураций частиц с заданным прицельным параметром. Как было показано, в случае разумных значений прицельного параметра столкновения вовсе проходят в “электродинамическом” режиме, так как векторы зарядов не вращаются в изотопическом пространстве. Вращение начинается после столкновения, что ведет к появлению нового слагаемого в напряженностях полей. Однако при выбранном значении прицельного параметра $a = 0.5$ Фм вклад “свечения” зарядов в напряженности полей оказывается существенно меньше аналогичного, возникающего при лобовом столкновении. При меньших значениях прицельного параметра вклад “свечения” растет, однако разумно обрезать дальнейшее уменьшение прицельного параметра при достижении некоторого значения. Показано, что непосредственно на динамику эффект “свечения” не оказывает существенного влияния.\
Исходя из анализа ситуации с цветовым эхо, видно, что сделанный вывод оказывается справедливым и для более сложных конфигураций частиц, так как жесткие глюонные поля могут генерироваться только в результате непосредственного столкновения, а не за счет каких-либо эффектов типа эха.\
Практический вывод о возможности описания системы в приближении ”чистой” электродинамики с усиленной константой связи особенно подтверждается при наличии прицельного параметра порядка нескольких десятых долей ферми, как естественного регуляризатора. В случае ультрарелятивистских скоростей и на расстояниях порядка 1 Фм нелинейность, связанная со слагаемыми, пропорциональными $g$, не изменяет кулоноподобного поведения.\
Автор выражает благодарность и А. М. Снигиреву за замечания и полезные обсуждения.
[99]{}
K. Tuchin, Phys. Rev. C [**82**]{}, 034904 (2010). A. Dbeyssi, D. Al Dirani and H. Zaraket, Phys. Rev. D [**84**]{}, 105033 (2011). V. Voronyuk, V. D. Toneev, W. Cassing, E. L. Bratkovskaya, V. P. Konchalovski and S. A. Voloshin, Phys. Rev. C [**83**]{}, 053911 (2011). L. McLerran and R. Venugopalan, Phys. Rev. D [**49**]{}, 3352 (1994);\
L. McLerran and R. Venugopalan, Phys. Rev. D [**50**]{}, 2225 (1994). A.H. Mueller, hep-ph/9911289. W. Cassing, V. V. Goloviznin, S. V. Molodtsov, [*et al.*]{}, Phys.Rev. C [**88**]{}, 6, 064909 (2011);\
В. Воронюк, В. В. Головизнин, Г. М. Зиновьев и др., ЯФ [**78**]{}, №3-4, 338-363 (2015). S. V. Molodtsov, A. M. Snigirev, G. M. Zinovjev, Phys. Lett. [**B**]{} 443, 387 (1998);\
S.V. Molodtsov, A.M. Snigirev, G.M. Zinovjev, Phys. Rev. C [**59**]{}, 955 (1999). J. Jalilian-Marian, A. Kovner, L. McLerran and H. Weigert, Phys. Rev. D [**55**]{}, 5414 (1997). Л. Д. Ландау, Е. М. Лифшиц, [*Теория поля*]{}, 7-е изд., испр. -М.: Наука. Физ.-мат. лит., 1988. В. В. Головизнин, C. В. Молодцов, А. М. Снигирев, ЯФ [**56**]{}, №6, 123 (1993) \[Phys. Atom. Nucl. [**56**]{}, 782, (1993)\]. И. В. Хриплович, ЖЭТФ [**74**]{}, 37 (1978) \[Sov. Phys. JETP [**47**]{}, 18 (1978)\].
****
**A. S. Zadora**
The aim of present work is to consider in more details recently theoretically observed exotic “color charge glow” effect and possible physical effects related to ensembles of color charged particles on the classical level. We study ways of occurrence of “color charge glow” in arbitrary systems of point massive particles which could be split into elementary configurations like color charges and color dipoles. Also the impact of this effect on dynamics of such systems is studying. For the natural regularization of expressions we consider collisions of particles with given impact factor. We show that with reasonable values of impact factor collisions could be quite of “electrodynamical” (Coloumb-like) regime and the contribution of “glow” turns out to be suppressed in comparison with “electrodynamical” picture. It follows from the analysis of “color echo” effect that this result is still correct if we consider more complicated configurations of particles because hard gluon fields could arise only due to immediate collisions but not as a result of some effects like “color echo”.
1. Напряженности полей, генерируемых в результате рассеяния частицы на частице, а) - компоненты хромоэлектричсекого поля, б) - компоненты хромомагнитного поля.
2. Напряженности полей, генерируемых в результате рассеяния частицы на диполе, а) - компоненты хромоэлектричсекого поля, б) - компоненты хромомагнитного поля.
3. Напряженности полей, генерируемых в результате рассеяния диполя на диполе.
4. Компоненты вектора $Q_3$ для $\omega = 0.1$ (a) и $\omega = 10$ (б).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'J.M. Chadney'
- 'M. Galand'
- 'T.T. Koskinen'
- 'S. Miller'
- 'J. Sanz-Forcada'
- 'Y.C. Unruh'
- 'R.V. Yelle'
bibliography:
- 'references.bib'
title: 'EUV-driven ionospheres and electron transport on extrasolar giant planets orbiting active stars'
---
Introduction
============
The ionospheres of solar system planets have been observed and modelled in great depth [e.g. @Nagy2002; @Witasse2008]. Visiting spacecraft making in situ measurements or remote observations of emission in the infrared (IR) and ultraviolet (UV), in combination with detailed modelling, have enabled us to glean at least a basic understanding of this layer of atmosphere throughout the solar system. This apparatus has not yet been fully applied to exoplanetary ionospheres. Many recent advances in technology and technique have enabled the detection of ions, such as C$^+$ and Si$^{2+}$, which have been detected during transits of the hot Jupiter HD209458b [@Vidal-Madjar2004; @Linsky2010]. However, observations remain difficult and must be supplemented by models. A potential diagnostic tool would be the measurement of emissions in the IR from the H$_3^+$ ion. This has been performed for Jupiter, Saturn, and Uranus [e.g. @Drossart1989; @Baron1991; @Trafton1993; @Rego2000; @Stallard2008a; @Miller2010], not only allowing determination of H$_3^+$ densities and temperatures, but also providing valuable constraints for ionospheric models [e.g. @Moore2015]. However, H$_3^+$ emissions have never been detected from an exoplanetary atmosphere, and the predicted emission levels are too low to be detected from planets orbiting stars of similar type and age to the Sun [@Shkolnik2006; @Koskinen2007a].
The ionosphere can be important in regulating the stability of upper atmospheres on extrasolar giant planets (EGPs). For example, previous modelling suggests that IR-active species, such as the H$_3^+$ ion, can act as a thermostat in EGP atmospheres and prevent them from undergoing hydrodynamic escape. Hence there are two different regimes of atmospheric escape depending on orbital distance, stellar heating and dissociation of molecular coolants upon which the composition and structure of the upper atmosphere depends. At large orbital distances ($a > 0.2$ AU for a Jupiter-like planet orbiting a Sun-like star), the main thermal escape mechanism is Jeans escape and the atmosphere is in a stable state of hydrostatic equilibrium, whereas planets with small orbital distances ($a < 0.2$ AU around a Sun-like star) undergo hydrodynamic escape [@Koskinen2007a; @Koskinen2014]. In addition, electrodynamics in the ionosphere can modulate escape rates and influence the structure of the upper atmosphere through ion drag and resistive heating [@Koskinen2014a].
Ionisation in upper planetary atmospheres occurs through two main processes. Photo-ionisation (or primary ionisation) by stellar photons has been calculated in past EGP models [e.g. @Yelle2004; @GarciaMunoz2007; @Koskinen2010; @Koskinen2013a]. However, to our knowledge, a full description of secondary ionisation, by photo-electrons and their secondaries, has not been included in EGP models. In solar system gas giants, secondary ionisation has been shown to strongly affect the lower ionosphere as well as the main ionisation peak [e.g. @Kim1994; @Galand2009].
In addition, previous EGP studies have used solar XUV (X-ray and extreme ultraviolet) fluxes as substitutes for stellar fluxes. Inter-stellar extinction makes it very difficult (or even impossible at certain wavelengths) to measure spectra of other stars in the extreme ultraviolet (EUV). However, in a recent study, @Sanz-Forcada2011 calculated mass-loss rates for gas giants orbiting close to a number of different stars. This was achieved by determining synthetic XUV spectra of these stars using emission measure distributions (EMD).
In contrast to previous models that assumed a Sun-like host star, we focus on the ionospheres of EGPs around young and active K- and M-dwarf stars. A significant number of the known exoplanets orbit such stars, and the upper atmospheres of these planets can be very different from similar planets orbiting Sun-like stars [@Chadney2015]. For the first time, we use actual stellar XUV spectra, determined using a stellar coronal model for each star. In addition to photo-ionisation, we also consider electron-impact ionisation by using a rigorous model of the energy degradation of suprathermal electrons. This is particularly important for planets orbiting active stars that typically emit more X-rays and short-wavelength EUV radiation than the Sun. This radiation is capable of enhancing the ionisation rates by producing high-energy photo-electrons that extend the ionosphere to lower altitudes. Based on our calculations, we also provide predictions of the H$_3^+$ emission power for planets orbiting active stars to determine whether they are better targets for future observations than planets around Sun-like stars.
The work described in this paper builds on that from @Chadney2015, in which the neutral thermosphere and atmospheric escape in EGPs around stars of different type and activity level were studied. In that work a simple method of scaling the stellar EUV flux was developed, based upon the observed X-ray flux. It was also shown that the transition from slow Jeans escape to hydrodynamic escape occurs at significantly larger orbital distances around more active stars. In the current paper, we combine the stellar coronal model and EGP thermosphere model [@Chadney2015] with a newly constructed 1D ionosphere model, in particular to study the effect of secondary ionisation by photo-electrons in EGP ionospheres around active stars.
It is important to note that we focus here upon ionisation in the thermosphere (at pressures $p<10^{-6}$ bar), where stellar EUV and soft X-ray radiation is absorbed. While in solar system planets, the main ionospheric peak is located in the thermosphere, @Koskinen2014a showed that in close-orbiting EGPs, the photo-ionisation of metals, such as Na and K, creates a stronger peak at lower altitudes, in the stratosphere. The ratio of the stellar X-ray to EUV flux is higher in more active stars [@Chadney2015] and this can further enhance the electron densities below the EUV ionisation layer.
This paper is laid out as follows. Section \[sec:models\] provides descriptions of the various components of the model. The results of the model runs are presented in Sect. \[sec:results\], including a study of the sensitivity of our results to the resolution of the H$_2$ photo-absorption cross section (Sect. \[sec:sig\_res\]). Section \[sec:prodrates\] discusses primary and secondary ionisation rates. Density predictions are provided in Sects. \[sec:stable\] and \[sec:hydrodyn\], focussing on planets undergoing Jeans escape, and hydrodynamic escape respectively. In Sect. \[sec:vary\_a\], we discuss the influence of orbital distance on ionospheric densities. Section \[sec:scaled\_spec\] presents the effects of using the scaled solar XUV spectra from @Chadney2015 to represent stellar spectra from other low-mass stars. Finally, Sect. \[sec:H3Pemissions\] provides estimates of IR emissions from the H$_3^+$ ion for planets orbiting a variety of different stars, at various orbital distances.
Models {#sec:models}
======
We use a set of coupled models to describe the upper atmosphere of EGPs. The relationship between the different model elements is shown in the schematic in Fig. \[fig:model\]. The only source of incident energy considered is stellar XUV radiation. We incorporate accurate descriptions of both the high-energy stellar spectrum, by using a coronal model (see Sect. \[sec:coronalmodel\]) and an upper planetary atmospheric model that consists of coupled thermospheric (see Sect. \[sec:thermomodel\]), ionospheric, and suprathermal electron transport models (see Sect. \[sec:ionomodel\]). In all cases, we assume a rotational period for the planet of 10 hours, similar to that of Jupiter and Saturn. This is justified because we concentrate on planets orbiting at distances larger than 0.2 AU from their host stars. Taking 0.2 AU as the inner most orbital distance in this study allows us to sample planets in both escape regimes, whilst also ensuring that the assumption of a pure H$_2$/H/He neutral upper atmosphere is more likely to be correct.
![Schematic of model elements (dark grey boxes) and inputs/outputs (light grey boxes).[]{data-label="fig:model"}](model_schematic3){width=".3\textwidth"}
Stellar coronal model {#sec:coronalmodel}
---------------------
We include in this study the K-dwarf $\epsilon$ Eridani ($T_{\text{eff}}=4900$ K) and the M-dwarfs AD Leonis ($T_{\text{eff}}=3370$ K) and AU Microscopii ($T_{\text{eff}}=3720$ K). These are close stars for which a good number of observations are available from Chandra, XMM-Newton, ROSAT, EUVE, and FUSE, which is important since these observations are used to calibrate the coronal model. This choice of stars also makes up a coarse parameter grid to study stars of different age and spectral type, where the Sun is not a valid proxy. $\epsilon$ Eridani has commonly been used as an analogue of the hot Jupiter host star HD189733, the two stars being of similar type, metallicity, and age [e.g. @Moses2011; @Venot2012]. AD Leonis has been used in previous studies of habitable planets [e.g. @Tarter2007]. AU Microscopii is a well-known flare star that is very young and active [@Cully1993]. As a result of the lower contrast between star and planet, observations of planetary atmospheres of EGPs around M stars have recently received particular attention. Selecting two M stars with similar bolometric luminosity but very different X-ray and EUV luminosity allows us to gauge the likely range of EGP ionospheres for cool host stars with different activity levels.
Our knowledge of the stellar EUV spectra needed as input to our ionospheric model is patchy: the problem is that the high-energy stellar radiation that is absorbed in upper planetary atmospheres is also absorbed by the inter-stellar medium (ISM). Photons of $\lambda < 91.2$ nm are capable of ionising an atmosphere composed of hydrogen and helium. For almost all stars but the Sun, the wavelength region between about 40 and 91.2 nm is unobservable. Therefore, to obtain XUV (combination of X-ray and EUV) stellar irradiances, either a coronal model of the star or an appropriate scaling of the solar spectrum must be used, such as that described in @Chadney2015. We obtain accurate XUV spectra using a stellar coronal model [@Sanz-Forcada2002; @Sanz-Forcada2003; @Sanz-Forcada2011]. These spectra are obtained by constructing an emission measure distribution (EMD) of the corona, transition region and chromosphere of each star. A more detailed description of the construction of the XUV spectra for each of the three selected stars is provided in @Chadney2015.
Thermospheric model {#sec:thermomodel}
-------------------
To calculate the number density, velocity, and temperture profiles in the upper atmosphere of a planet, we used the one-dimensional thermospheric EGP model developed by @Koskinen2013a [@Koskinen2013b; @Koskinen2014]. In all simulations the planetary parameters of HD209458b were used (radius $R_p =$ 1.32 $R_{\text{Jupiter}}$, mass $M_p =$ 0.69 $M_{\text{Jupiter}}$). The model solves the vertical equations of motion from the 10$^{-6}$ bar level up to the exobase for a fluid composed of H, H$_2$, and He, and their associated ions H$^+$, H$_2^+$, H$_3^+$, He$^+$, and HeH$^+$. The model is driven by the stellar irradiance and determines profiles with altitude of neutral densities, bulk velocity, and neutral temperature. The profiles obtained are a global average, obtained by dividing the incoming stellar flux by a factor of 4.
It should be noted that the thermospheric model is not fully coupled to the ionospheric model. Considering ionospheric densities when calculating the neutral thermosphere is important, especially to properly treat IR cooling by the H$_3^+$ ion. Therefore the thermospheric model does include a calculation of ionospheric densities. However, this calculation is not as thorough as that included in the full ionospheric model (see Sect. \[sec:ionomodel\]), as electron-impact ionisation by supra-thermal electrons is not taken into account. Instead, heating of the thermosphere due to photo-electrons is included in the thermospheric model using a fixed heating efficiency of 93 %. @Koskinen2013a showed that using a fixed heating efficiency is appropriate for altitudes below $3R_p$. The value of 93 % that we used is valid for photo-electrons created by photons of energy up to 50 eV at an electron mixing ratio of 0.1 [@Cecchi-Pestellini2009], which is attained near the temperature peak in our models. The principal aim of this study is to compare the differences in composition of the ionosphere of planets orbiting active stars with different high-energy spectral energy distributions, and not to precisely model the neutral temperature and density profiles in these atmospheres.
Ionospheric model {#sec:ionomodel}
-----------------
The ionospheric model constructed for this work solves the one-dimensional coupled continuity equations for the ions H$^+$, H$_2^+$, H$_3^+$, and He$^+$ to provide the densities of these ions. The new ionosphere model is required because the previous models did not include secondary ionisation by photo-electrons. In spherical coordinates, where only radial transport is considered, the continuity equation for each ion $i$ reduces to $$\label{eqn:cont_vert}
\frac{\partial n_i}{\partial t} + \frac{1}{r^2}\frac{\partial}{\partial r}\left(r^2 \Phi_i\right) = P_i - L_i,$$ where $n_i$ is the number density of ion $i$, $P_i$ is the production rate, and $L_i$ is the loss rate of ion $i$, $\Phi_i=n_iv_i$ is the radial transport flux, $v_i$ being the drift velocity of ion $i$. Ion production and loss occur through the photo-chemical reactions listed in Table \[tab:reactionsiono\]. Neutral species are ionised through photo-ionisation and electron-impact ionisation. We included the latter by using a suprathermal electron transport model adapted to EGPs that is based on the solution to the Boltzmann equation with transport, angular scattering, and energy degradation of photo-electrons and their secondaries taken into account. For further information on the suprathermal electron transport model, see @Moore2008a and @Galand2009.
We assumed here that the source of incident energy is solar or stellar XUV radiation. No electron precipitation from the space environment was included. A number of different parameters are required as inputs to the ionospheric model; these are the solar or stellar spectrum, neutral temperature and densities, chemical reaction rates, and cross sections. We used the TIMED/SEE instrument for measurements of solar XUV irradiances and a coronal model (see Sect. \[sec:coronalmodel\]) to produce stellar XUV spectra. The neutral atmosphere was assumed to be a constant background and was obtained for a set of planetary parameters and stellar irradiances using a thermospheric model (see Sect. \[sec:thermomodel\]). The assumption that the neutral densities remain constant is valid as long as ion densities remain small. This is the case in the ionospheres of solar system bodies, and we have also verified that this assumption remains valid for all the cases considered in this study because of the relatively large orbital distances.
Chemical reaction rates and cross sections are provided in Table \[tab:reactionsiono\] and Fig. \[fig:cross-sec\], respectively. Cross sections of two different wavelength resolutions are used to describe the photo-absorption of H$_2$. At wavelengths greater than the ionisation threshold, H$_2$ absorbs in a large number of narrow lines, comprising the Lyman, Werner, and Rydberg bands. Typical ionosphere models use a low-resolution version of the H$_2$ cross section in these bands, including all of the previous exoplanet models. In Sect. \[sec:sig\_res\] we present the effects on EGP atmospheres of using either low- or high-resolution versions of these cross sections.
The reaction of H$^+$ with vibrationally excited H$_2$ (reaction 11 in Table \[tab:reactionsiono\]) is an important loss process for H$^+$, but its reaction rate is not well constrained; as such, it is a large source of uncertainty in ionospheric models applied to giant planets. @McElroy1973 was the first to note that this reaction would become exothermic for vibrationally excited H$_2$ and that there may well be enough excited H$_2$ in gas giant atmospheres to render this reaction significant. The rate coefficient of reaction 11 is known to be $1.0\times10^9$ cm$^3$s$^{-1}$ [@Huestis2008]. However, the proportion of vibrationally excited H$_2$ is unknown. We used the expression determined by @Yelle2004 from Jupiter observations: $$\frac{\left[\text{H}_2(\nu\geq4)\right]}{\left[\text{H}_2\right]} = \text{exp}(-2.19\times10^4/T).$$ Therefore, in the ionospheric model, we assumed the following reaction: $$\text{H}^+ + \text{H}_2 \rightarrow \text{H}_2^+ + \text{H} \,\,\,\, \text{(reaction 11b)}$$ (note H$_2$ is in the base state), with the following reaction rate: $$k_{11\text{b}} = 1.0\times10^9 \text{exp}(-2.19\times10^4/T)\,\text{cm}^3\text{s}^{-1}.$$
Diurnal variations in rotating planets are determined by calculating ion densities with altitude at a given latitude and varying the stellar zenith angle over the course of a day. We neglected horizontal circulation and assumed that vertical gradients are dominant over horizontal gradients.
We assumed that any species heavier than He is confined below the lower boundary of the model, located at a pressure of 1 $\mu$bar. This pressure level was therefore assumed to correspond to the homopause. Species are diffusively separated above this level, heavier species being confined to lower altitudes. This assumption is valid if the eddy diffusion coefficient $K_{zz}$ is lower than $\sim10^2$ – $10^3$ m$^2$ s$^{-1}$ at a pressure level of 1 $\mu$bar. The eddy diffusion coefficient is not well known in exoplanetary atmospheres, but its value is estimated to be about $10^2$ m$^2$ s$^{-1}$ at Jupiter [@Yelle1996]. $K_{zz}$ may well be higher in close-orbiting EGPs than at Jupiter [@Koskinen2010], in which case the homopause could be located at a lower pressure than we assumed, leading to the destruction of the IR coolant H$_3^+$ through reactions with hydrocarbons, for instance. There are currently no constraints on $K_{zz}$ in EGPs however, a situation that might be improved by the detection of IR emissions from the H$_3^+$ ion (see Sect. \[sec:H3Pemissions\]).
Atmospheric escape could also cause heavier species to be present above the 1 $\mu$bar level. In certain planets undergoing hydrodynamic escape, hydrocarbons could be dragged upwards into the thermosphere by escaping hydrogen. We estimated in which of the cases discussed in this paper hydrocarbons could be present in the thermosphere because of atmospheric escape: it only occurs in EGPs orbiting within 0.2 AU from AD Leo (see Sect. \[sec:vary\_a\]).
\# Reaction Reaction rate Reference
---- --------------------------------------------------------------------------- ------------------------------------------------ ----------------
Photo-ionisation:
1 $\text{H}_2 + h\nu \rightarrow \text{H}_2^+ + e^-$
2 $\text{H}_2 + h\nu \rightarrow \text{H}^+ + \text{H} + e^-$
3 $\text{H}_2 + h\nu \rightarrow 2\text{H}^+ + 2e^-$
4 $\text{H} + h\nu \rightarrow \text{H}^+ + e^-$
5 $\text{He} + h\nu \rightarrow \text{He}^+ + e^-$
Electron-impact ionisation:
6 $\text{H}_2 + e^- \rightarrow \text{H}_2^+ + e^- + e^-$
7 $\text{H}_2 + e^- \rightarrow \text{H}^+ + \text{H} + e^- + e^-$
8 $\text{H}_2 + e^- \rightarrow 2\text{H}^+ + 2e^- + e^-$
9 $\text{H} + e^- \rightarrow \text{H}^+ + e^- + e^-$
10 $\text{He} + e^- \rightarrow \text{He}^+ + e^- + e^-$
Charge exchange
or proton transfer:
11 $\text{H}^+ + \text{H}_2(\nu \geq 4) \rightarrow \text{H}_2^+ + \text{H}$
12 $\text{H}_2^+ + \text{H}_2 \rightarrow \text{H}_3^+ + \text{H}$ $2.0\times10^{-9}$ @Theard1974
13 $\text{H}^+ + \text{H}_2 + \text{M} \rightarrow \text{H}_3^+ + \text{M}$ $3.2\times10^{-29}$ @Kim1994
14 $\text{He}^+ + \text{H}_2 \rightarrow \text{H}^+ + \text{H} + \text{He}$ $1.0\times10^{-9}\text{exp}(-5.7\times10^3/T)$ @Moses2000
15 $\text{He}^+ + \text{H}_2 \rightarrow \text{H}_2^+ +\text{He}$ $9.35\times10^{-15}$ @Anicich1993
16 $\text{H}_3^+ + \text{H} \rightarrow \text{H}_2^+ +\text{H}_2$ $2.1\times10^{-9}\text{exp}(-2.0\times10^4/T)$ @Harada2010
17 $\text{H}_2^+ + \text{H} \rightarrow \text{H}^+ +\text{H}_2$ $6.4\times10^{-10}$ @Karpas1979
Recombination:
18 $\text{H}^+ + e^- \rightarrow \text{H} + h\nu$ $4.0\times10^{-12}(300/T_e)^{0.64}$ @Storey1995
19 $\text{H}_2^+ + e^- \rightarrow 2\text{H}$ $2.3\times10^{-7}(300/T_e)^{0.4}$ @Auerbach1977
20 $\text{He}^+ + e^- \rightarrow \text{He} + h\nu$ $4.6\times10^{-12}(300/T_e)^{0.64}$ @Storey1995
21 $\text{H}_3^+ + e^- \rightarrow \text{H}_2 + \text{H}$ $2.9\times10^{-8}(300/T_e)^{0.64}$ @Sundstrom1994
22 $\text{H}_3^+ + e^- \rightarrow 3\text{H}$ $8.6\times10^{-8}(300/T_e)^{0.64}$ @Datz1995
\[tab:reactionsiono\]
![Photo-absorption (thick black and thin grey lines) and photo-ionisation (thin black lines) cross sections used in the ionospheric model for the three neutral species considered and their associated ions. Panel (a) shows cross sections for H$_2$; in thick black we show the low-resolution (LR) photo-absorption cross section and in thin grey the high-resolution (HR) version (at 700 K) including structure beyond the ionisation threshold. The thin solid line plots the production of H$_2^+$, the dashed line the production of H$^+$+H, and the dashed-dotted line the production of 2H$^+$. Panel (b) shows the photo-absorption cross section of H as a thick black line, which is indistinguishable from the production of H$^+$ through photo-ionisation, which is shown as a thin solid line. Panel (c) shows the photo-absorption cross section of He as a thick black line, which is indistinguishable from the production of He$^+$, which is shown as a thin solid line, and the production He$^{2+}$ (the dashed line).[]{data-label="fig:cross-sec"}](cross-sections){width=".48\textwidth"}
H$_3^+$ emission model
----------------------
H$_3^+$ emissions from EGP ionospheres may be detectable in optimal circumstances with current and in particular with future ground-based telescopes [e.g. @Shkolnik2006]. This would be very interesting because, in addition to providing constraints on the ionosphere, they could be used to detect non-transiting planets. Therefore, we determined emission strengths for the various planets that we modelled (see Sect. \[sec:H3Pemissions\]). The H$_3^+$ ion has two vibrational modes, of which only the second, $\nu_2$, is active. It is centred on a wavelength of 3.9662 $\mu$m. The energy emitted in a given line under local thermodynamic equilibrium (LTE) was calculated using the following equation[@Miller2010]: $$I(\omega_{if},T) = \frac{1}{4\pi Q(T)}g_f (2J_f+1) hc\,\omega_{if} A_{if} \text{exp}\left(-\frac{hcE_f}{kT}\right),$$ where the indices $i$ indicate the lower state and $f$, the upper state, $\omega_{if}$ is the frequency of the transition (per cm), $A_{if}$ is the Einstein $A$ coefficient of the transition, $g_f$ is the nuclear spin degeneracy, $J_f$ is the angular momentum of the upper state $f$, $E_f$ is the energy of the upper state (per cm), $Q(T)$ is the partition function of H$_3^+$, $h$ is the Planck constant, $c$ is the speed of light, $k$ is the Boltzmann constant, and $T$ is the temperature. The factor $hc$ converts wavenumbers to SI units, meaning that $I$ has units of $\text{W}\,\text{molecule}^{-1}\,\text{sr}^{-1}$. Energy levels and Einstein $A$ coefficients were obtained from @Dinelli1992. The power emitted in a given line was obtained by multiplying $I$ by the number of H$_3^+$ molecules and considering emission over $2\pi$ sr. Partition functions are those described in @Miller2010, which are valid for high-temperature atmospheres.
H$_3^+$ levels are often sub-thermally populated in planetary atmospheres. @Miller2013 estimated that this is the case up to densities of $10^{11} - 10^{12}$ cm$^{-3}$, and that it is due to high Einstein coefficients, meaning that there is competition between radiative relaxation and collisional de-excitation at high densities. The densities we predict are lower than $10^{11}$ cm$^{-3}$ (see Sects. \[sec:stable\], \[sec:hydrodyn\], and \[sec:vary\_a\]), it is therefore important to consider non-LTE effects when determining H$_3^+$ emission strengths. These effects were taken into account by introducing a weighting factor, using the method of @Oka2004, as detailed in @Miller2013. This weighting factor is dependent on the density of H$_2$.
We determined emission in the Q and R branches, which show stronger emission than the P branch, and we also provide values for the total H$_3^+$ output power. The total output power per H$_3^+$ molecule can be expressed as: $$E(\text{H}_3^+,T) = \sum_{if} I(\omega_{if},T).$$ To calculate this quantity, we used the following parametrisation, determined by @Miller2013: $$\text{log}_e \left(E(\text{H}_3^+,T)\right) = \sum_n C_n T^n,$$ where the coefficients $C_n$ are provided in Table 5 of @Miller2013.
Ionisation in EGPs orbiting stars of different activity levels {#sec:results}
==============================================================
Effect of the photo-absorption cross-section resolution {#sec:sig_res}
-------------------------------------------------------
The photo-absorption cross section of H$_2$ is very structured beyond the ionisation threshold ($\lambda_{\text{th,H}_2}=80.4$ nm), as a result of absorption in the Lyman, Werner, and Rydberg bands. However, all EGP and most solar system ionospheric studies [e.g. @Galand2009; @Koskinen2010] have previously used the low-resolution measurements of @Backx1976, taken at 3 – 4 nm intervals. We evaluated the impact on EGP ionospheres of using a high-resolution H$_2$ photo-absorption cross section with $\Delta\lambda=10^{-3}$ nm (plotted in grey in Fig. \[fig:cross-sec\](a)), instead of the low-resolution @Backx1976 measurements (plotted as a thick black line in Fig. \[fig:cross-sec\](a)). The high-resolution cross section we used consists of the @Backx1976 measurements below $\lambda_{\text{th,H}_2}=80.4$ nm and high-resolution calculations by @Yelle1993 at longer wavelengths. In addition, the H$_2$ photo-dissociation cross section measured by @Dalgarno1969 was added to the @Yelle1993 calculations for wavelengths between 80.4 nm and 84.2 nm. Until now, these cross sections with a high spectral resolution have not been used in studies of EGP ionospheres. However, at Saturn [@Kim2014] and Jupiter [@Kim1994], using high-resolution H$_2$ photo-absorption cross sections led to the prediction of a larger and more extended layer of hydrocarbon ions in the lower ionosphere.
We find that high-resolution cross sections are not required when modelling pure H/H$_2$/He EGP atmospheres. The ratio of atomic hydrogen to molecular hydrogen column densities is significantly larger in the extended EGP atmosphere than at Saturn or Jupiter. This means that absorption by H dominates absorption by H$_2$, and as such, the resolution of $\sigma_{\text{H}_2}^{\text{abs}}$ has a much weaker effect for EGPs than for the solar system cases described by @Kim1994 and @Kim2014. The differences between using high- and low-resolution H$_2$ photo-absorption cross sections are minor (change lower than 2.5 % in ion densities), therefore we consider that the low-resolution cross sections are sufficient for the remainder of this work on EGP ionospheres.
Even though high-resolution cross sections are not required here, this may not always be the case. If photo-ionisation reactions are included that involve ionisation threshold wavelengths larger than that of H ($\lambda_{\text{th,H}}=91.2$ nm), such as those involving certain hydrocarbons, then the effects of absorption by H$_2$ in the Lyman, Werner and Rydberg bands may no longer be weak. At these wavelengths, there is no longer any absorption by atomic H, and accordingly, absorption over a wide range of pressures by molecular H$_2$ occurs. The inclusion of the high-resolution cross section may therefore allow for more efficient ionisation below the thermosphere.
Production rates: primary and secondary ionisation {#sec:prodrates}
--------------------------------------------------
We included ionisation through both primary (photo-) ionisation and secondary (electron-impact) ionisation processes (see Sect. \[sec:ionomodel\]). Electron-impact ionisation operates through collisions with photo-electrons and their secondaries. In this section, we describe the contribution of each of these processes to ionospheric densities. We chose to run the model for EGPs orbiting at 1 AU from the Sun (at minimum and maximum activity), $\epsilon$ Eri, AD Leo, and AU Mic, to study the contribution of the different stellar spectral energy distributions. At this orbital distance, our assumption of a pure H$_2$/H/He thermosphere is valid, as long as the eddy diffusion coefficient is lower than $\sim10^2$ – $10^3$ m$^2$ s$^{-1}$. Below (Sect. \[sec:vary\_a\]), we examine the effect of varying the orbital distance.
Figure \[fig:stars\_Pe\_ptau1\_spectra\](b) shows the total ion production rates for photo-ionisation (solid lines) and electron-impact reactions (dashed lines) for planets orbiting each star. We also show in this figure the stellar spectra used to drive the ionospheric model in panel (a) and the pressures of unity optical depth in panel (c). The pressures $p$ at which $\tau=1$, shown in panel (c), allow us to identify the emission lines in the stellar spectrum (panel (a)) that are the main contributors to the different peaks in the ionisation rate plots (panels (b)). Solar spectra were obtained from TIMED/SEE on 15 May 2008 (F10.7 = 70 solar flux units) for the case of solar minimum and on 14 January 2013 (F10.7 = 154 solar flux units) for solar maximum. Synthetic spectra from the coronal model were used for the other stars. We note that $p(\tau=1)$ is dependent on the neutral atmosphere, but since the neutral atmospheres of planets in the same escape regime are so similar, the curves of $p(\tau=1)$ are also very similar. We have plotted the mean of the four $p(\tau=1)$ curves for planets in hydrostatic equilibrium as a thick black line in panel (c): those of planets around the Sun at minimum and maximum, $\epsilon$ Eri and AD Leo. Had we plotted each case individually, the difference would barely have been noticeable because the curves are mostly superposed. The curve of $p(\tau=1)$ for the case of a planet orbiting AU Mic is slightly different from the others (plotted as a light red line in panel (c)), since this planet is in a regime of hydrodynamic escape (see Sects. \[sec:stable\] and \[sec:hydrodyn\]).
The optical depth is determined by the densities of atmospheric constituents, the incoming stellar flux, and the photo-absorption cross sections (plotted in Fig. \[fig:cross-sec\]). Since we used low-resolution photo-absorption cross sections (see Sect. \[sec:sig\_res\]), the pressure of unity optical depth decreases monotonically with photon wavelength (i.e. the altitude of unity optical depth increases monotonically with wavelength). This means that higher energy photons deposit their energy lower in the ionosphere. We also note that due to the higher gradient of the $p(\tau=1)$ curve at lower wavelengths, which is associated with a large change in $\sigma^{\text{abs}}$, high-energy (e.g. soft X-ray, $\sim 0.1 - 10$ nm) stellar emission lines affect a much more extended region of atmosphere than higher wavelength stellar emissions.
{width="12cm"}
Because the spectral shapes are similar, the ionisation rate profiles of planets orbiting the Sun at solar min and max are very similar. The main photo-ionisation peak at around 6 nbar is due to the He II lines at 25.6 nm and 30.4 nm in the solar spectrum. These lines are also responsible for the main peaks in the photo-ionisation profiles in planets orbiting $\epsilon$ Eri, AD Leo, and AU Mic. The main peak in electron-impact ionisation in planets irradiated by the solar spectrum occurs at around 10 – 20 nbar and is formed by the strong Fe emission lines at wavelengths just below 20 nm. There is a secondary peak in the photo-ionisation and electron-impact ionisation profiles of EGPs around a solar-like star, at a pressure of 200 – 300 nbar. These secondary peaks are due to solar soft X-ray emission just below 10 nm.
The spectral shape of high-energy emissions is quite different in stars that are more active than the Sun. There is a gradual rise in irradiance with decreasing wavelength below about 20 nm that is responsible for an increase in ionisation rates with pressure above about 100 nbar. It seems that the peak ionisation rate is not always reached in the pressure domain of the model, especially for planets around the most active star, AU Mic. So for planets around active stars, there is significant absorption of high-energy radiation and ionisation by associated secondary electrons below 1 $\mu$bar, in the stratosphere. @Koskinen2014a showed that in close-in EGPs, ionisation of alkali metals, such as Na, K, and Mg, by far-ultraviolet (FUV) photons creates a strong ionisation layer in the stratosphere. Our results here indicate that this ionisation layer in the lower ionosphere can be enhanced by the ionisation of hydrogen by X-rays around active stars.
Another interesting effect of different spectral energy distributions appears in the ionisation profiles of planets around $\epsilon$ Eri and AD Leo. As examined in @Chadney2015, irradiance in the X-ray and EUV wavebands evolves at different rates over the lifetime of the star [see also @Ayres1997; @Ribas2005; @Sanz-Forcada2011]. High-energy stellar emissions decrease as the star ages, but the highest energy emissions decrease at a faster rate. Indeed, @Chadney2015 showed that the ratio of fluxes at the stellar surface, $F_{\text{EUV}}/F_{\text{X}}$ increases as the star ages. $\epsilon$ Eri is an older star than AD Leo and so has lower emissions in the X-ray, but still has higher emissions in the EUV. This results in higher ionisation in the lower part of the ionosphere in a planet orbiting AD Leo than $\epsilon$ Eri, but lower ionisation at higher altitudes. We note that this is the case even though AD Leo is smaller in size than $\epsilon$ Eri, based on which lower emission might have been expected for the same activity level. AD Leo is much younger and hence more active, however, which is a stronger factor than size.
![Primary efficiency (ratio of secondary to primary electron production rate) versus pressure in an EGP orbiting at 1 AU from the Sun at solar minimum (green), the Sun at solar maximum (blue), the star $\epsilon$ Eri (yellow), AD Leo (purple), and AU Mic (red).[]{data-label="fig:primary_efficiency_stars"}](prim_efficiency){width="48.00000%"}
The primary efficiency is defined as the ratio of electron-impact ionisation rate to photo-ionisation rate. This is plotted in Fig. \[fig:primary\_efficiency\_stars\] for each of the five cases. As expected, since the pressure at which $\tau=1$ increases monotonically with decreasing wavelength (see Fig. \[fig:stars\_Pe\_ptau1\_spectra\](c)), higher energy photons are absorbed lower in the ionosphere. This means that the highest energy photo-electrons, which are responsible for most of the ionisation, are formed at high pressures. The pressure at which electron-impact ionisation overtakes photo-ionisation is about 10 nbar, which is just below the main production peak (see Fig. \[fig:stars\_Pe\_ptau1\_spectra\](b)). Thus electron-impact ionisation increases the electron densities significantly below the main EUV ionisation peak, allowing for the lower atmosphere to be ionised more efficiently than in models that ignore electron-impact ionisation.
Ionospheric densities: atmosphere in hydrostatic equilibrium {#sec:stable}
------------------------------------------------------------
{width="12cm"}
This section describes the ionospheric densities in planets with atmospheres in hydrostatic equilibrium. These are planets orbiting at larger distances from their stars, with atmospheres that undergo thermal Jeans escape. For the cases studied here, EGPs around the Sun, $\epsilon$ Eri, and AD Leo, at an orbital distance of 1 AU, are in hydrostatic equilibrium. EGPs at 1 AU from the more active star AU Mic have atmospheres that lose material according to an organised outflow. The ion densities in these atmospheres, which are in the hydrodynamic escape regime, are detailed in Sect. \[sec:hydrodyn\].
The variation of electron density, as well as $n_{\text{H}^+}$ and $n_{\text{H}_3^+}$ over the course of a day, is presented in the contour plots in Fig. \[fig:densiy\_contourplot\], at a latitude of 30$^{\circ}$ for a planet orbiting the Sun at 1 AU. We note that the electron densities in Fig. \[fig:densiy\_contourplot\](a) are plotted on a logarithmic scale, whereas the ion densities of H$^+$ (Fig. \[fig:densiy\_contourplot\](b)) and H$_3^+$ (Fig. \[fig:densiy\_contourplot\](c)), are plotted on a linear scale. This is so that the two distinct regions are clearly visible in the electron density plot. They correspond to an $\text{H}^+$-dominated region in the upper ionosphere at $p\lesssim200$ nbar, and to an $\text{H}_3^+$-dominated region below. The $\text{H}^+$-dominated region is relatively uniform over the diurnal cycle, with a peak spread out over a wide pressure range.
The $\text{H}_3^+$-dominated region at $p\gtrsim200$ nbar clearly displays a strong diurnal variation (see also Fig. \[fig:densiy\_contourplot\](c)), with a density peak near 12 LT. H$_3^+$ densities decrease gradually in the evening, as photo-ionisation of H$_2$ diminishes with the setting star and the remaining $\text{H}_3^+$ is destroyed through electron dissociative recombination reactions 21 and 22. At dawn, as photo-ionisation is switched back on, there is a more rapid build-up of $\text{H}_3^+$ ions.
The diurnal variations in the H$^+$-dominated region are less pronounced, but can be seen in Fig. \[fig:densiy\_contourplot\](b) that shows H$^+$ densities. Unlike the $\text{H}_3^+$ density peak, the H$^+$ peak is offset with the stellar illumination peak. A maximum H$^+$ density of $3.9\times 10^6$ cm$^{-3}$ occurs at about 17 LT, the smallest peak is at 7 LT, where $n_{\text{H}^+}=3.4\times 10^6$ cm$^{-3}$. This lag behind the stellar illumination is due to the fact that H$^+$ is a long-lived ion.
![Diurnal variation of peak H$^+$ (in black) and H$_3^+$ (in grey) densities in an EGP orbiting the Sun at solar minimum (thin solid lines), the Sun at solar maximum (thick solid lines), the star $\epsilon$ Eri (dashed lines), and the star AD Leo (dotted lines).[]{data-label="fig:peak_nHP_nH3P_vs_LT_stars"}](peak_nHP_nH3P_vs_LT_stars_bw){width="49.00000%"}
Figure \[fig:peak\_nHP\_nH3P\_vs\_LT\_stars\] presents the variation over the course of a day of the peak densities and timings of the two main ionospheric constituents, H$^+$ (in black) and H$_3^+$ (in grey), for planets with atmospheres in hydrostatic equilibrium, orbiting each star. The same diurnal variation is predicted as for a planet orbiting the Sun during solar maximum, namely a strong variation in densities of the short-lived ion H$_3^+$, with a peak that is almost symmetrical around 12 LT, and a significantly smaller, almost imperceptible diurnal variation in H$^+$ density, with a peak at around 17 LT that is offset from the peak in stellar illumination. Peak H$_3^+$ densities increase along with the stellar X-ray flux, whereas peak H$^+$ densities are a function of the stellar EUV flux level. Thus, higher H$_3^+$ densities occur in the planet around AD Leo than around $\epsilon$ Eri, but higher H$^+$ densities are found in the planet orbiting $\epsilon$ Eri. Values of peak H$_3^+$ and H$^+$ densities are provided in Table \[tab:peak\_ni\_1au\].
Electron-impact ionisation affects ion densities in the ionosphere differently at different times of day. Figure \[fig:neIIplusI/neI\] is a contour plot showing the ratio of electron densities between a full calculation taking both photo- and electron-impact ionisation into account ($n_e(P_{I+II})$) and electron densities calculated with only photo-ionisation ($n_e(P_{I})$). These are shown for an EGP orbiting a Sun-like star at 1 AU, but are qualitatively similar for EGPs orbiting the other stars. Only the lower ionosphere is plotted since above a few 10$^1$ nbar, essentially $n_e(P_{I+II})/n_e(P_{I})\sim 1$. Including secondary ionisation has the strongest effect on electron densities shortly after dawn, at around 7 LT, which corresponds to the local time when H$_3^+$ densities increase strongly. This strong increase in the ratio just after dawn occurs because the strong increase in electron-impact ionisation after sunrise is combined with a low loss rate for the ionospheric plasma below $10^2$ nbar. In this region H$_2^+$ is the main ion produced under stellar illumination and is rapidly converted to H$_3^+$ (reaction 12). The main loss of H$_3^+$ is through electron dissociative recombination (reactions 21 and 22). After a night without ionisation, the ionosphere is strongly depleted, and therefore the loss rate for the ionospheric plasma is low. In addition, near sunrise the electron-impact ionisation contributes up to $10^2$ nbar, while at noon it is confined towards lower altitudes (higher pressures). The reason is that photo-electrons that are energetic enough to cause significant ionisation are produced by solar photons with wavelength shorter than 20 nm. Figure \[fig:stars\_Pe\_ptau1\_spectra\](c) shows that these photons are absorbed primarily below $10^1$ nbar at noon. At larger stellar zenith angles, the altitude of deposition increases, which explains the extension towards lower pressures of the contribution of electron-impact ionisation at sunrise. Since there are no photo-electrons during the night, there is little electron-impact ionisation. There are some differences after dusk, while extra ions are destroyed that were built up during the day though secondary ionisation.
![Ratio of electron densities determined including both primary (photo-) ionisation and secondary (electron-impact) ionisation $n_e(P_{I+II})$ and those determined including just primary ionisation $n_e(P_{I})$, for an EGP orbiting a Sun-like star at 1 AU.[]{data-label="fig:neIIplusI/neI"}](neII_I_over_neI_contour_p_vs_LT_sunmax1au_R2013a){width="48.00000%"}
max($n_{\text{H}_3^+}$) \[cm$^{-3}$\] max($n_{\text{H}^+}$) \[cm$^{-3}$\]
------------------ --------------------------------------- -------------------------------------
Sun at solar min $1.9\times10^4$ $2.9\times10^6$
Sun at solar max $2.8\times10^4$ $3.9\times10^6$
$\epsilon$ Eri $8.0\times10^4$ $6.2\times10^6$
AD Leo $1.4\times10^5$ $5.5\times10^6$
\[tab:peak\_ni\_1au\]
Ionospheric densities: atmosphere in hydrodynamic escape regime {#sec:hydrodyn}
---------------------------------------------------------------
{width="12cm"}
AU Mic is the youngest and most active star included in this study, and as such has the strongest XUV emissions. As a consequence, the HD209458b-type planet that we modelled is in a regime of hydrodynamic escape, even at an orbital distance of 1 AU, whereas the same planet orbiting at the same distance from a Sun-like star, $\epsilon$ Eri, or AD Leo is in the Jeans escape regime [@Chadney2015]. The atmosphere is more extended than at planets in the stable regime and molecules, such as H$_2$, are fully dissociated at high pressures. In the case of an EGP orbiting AU Mic at 1 AU, H$_2$ is dissociated between 1 and 0.1 nbar.
Similarly as in planets in the stable regime, the ionosphere of EGPs undergoing hydrodynamic escape is composed of two distinct regions, characterised by the major ion: H$^+$ is dominant above the H$_2$/H dissociation front, whereas H$_3^+$ dominates below. H$_2^+$ and He$^+$ ions are always minor constituents.
Contour plots of the electron and major ion densities are shown in Fig. \[fig:densiy\_contourplot\_aumic1au\] at 30$^{\circ}$ latitude on an EGP orbiting AU Mic at 1 AU. As in the planets orbiting less active stars, H$_3^+$ ions reach a well-defined daily peak near noon and are depleted during night time, when there is no stellar illumination. This is due to their short-lived nature. Conversely, the long-lived H$^+$ ions reach their daily peak density between 16 and 17 LT and are present in large quantities throughout the night because they only recombine slowly with electrons (reaction 18 is slow).
In contrast to the planets in the stable regime, H$_3^+$ is confined to higher pressures in planets orbiting AU Mic at 1 AU, since it can only be formed in regions where H$_2$ is present. Its peak density is higher than in the planets discussed in Sect. \[sec:stable\], however, as a result of the enhanced X-ray flux: the H$_3^+$ peak density reaches $2.4\times10^5$ cm$^{-3}$ at 12.2 LT (compared to $1.4\times10^5$ cm$^{-3}$ for a planet at 1 AU from AD Leo, see Table \[tab:peak\_ni\_1au\]). Densities of H$^+$ are also higher, a peak of $1.1\times10^7$ cm$^{-3}$ is present at 16.3 LT, at a pressure of about 4 nbar (compared to a peak of $5.5\times10^6$ cm$^{-3}$ at 1 AU from AD Leo). We note that the H$^+$ contour has a double peak, one at 4 nbar and the other at 0.3 nbar (see Fig.\[fig:densiy\_contourplot\_aumic1au\](b)). This is due to the dissociation of H$_2$, which takes place between these two peaks. In particular, reaction 17 ($\text{H}_2^+ + \text{H} \rightarrow \text{H}^+ +\text{H}_2$) only contributes to the peak at 4 nbar. Additionally, below the H$_2$/H dissociation front, photo-ionisation reactions 2 and 3 (ionisation of H$_2$ to form H$^+$) contribute to the peak at 4 nbar. The smaller peak at 0.3 nbar is mainly formed from photo-ionisation of atomic H.
The cross-over mass equation from @Hunten1987, given in Eq. 3.2 in @Koskinen2014 provides a rough estimate of whether escaping heavy species from the stratosphere can be present in the thermosphere. The expression derived in @Koskinen2014 is independent of altitude for an isothermal atmosphere. We applied this expression here using temperatures that are consistent with our model results. Heavy species, including hydrocarbons and water, could destroy the H$_3^+$ layer. Applying the cross-over mass equation gives a limiting mass-loss rate of $1.5\times 10^5$ kg s$^{-1}$ required for C to be present at a pressure of 500 nbar (in the H$_3^+$ layer) in an EGP at 1 AU from AU Mic. The upwelling mass flux at this same pressure level is $5.6\times 10^4$ kg s$^{-1}$, which is thus insufficient to drag heavy species into this region. This gives us confidence that our assumption that species heavier than He are confined below a pressure of 1 $\mu$bar is correct for an EGP at 1 AU from AU Mic.
Variations in ion densities with orbital distance {#sec:vary_a}
-------------------------------------------------
![Major ion densities for an EGP orbiting AD Leo at different orbital distances, determined at 30$^{\circ}$ latitude and 12 LT.[]{data-label="fig:ni_vs_p_adleo_diffa"}](ni_vs_p_adleo_diff_a_12LT_bw){width="49.00000%"}
In this section, we discuss the changes in the ionosphere with orbital distance by considering an HD209458b-type planet orbiting the star AD Leo at 1 AU, 0.5 AU, and 0.2 AU. The critical orbital distance at which this EGP transitions from Jeans to hydrodynamic escape is located between 1 AU and 0.5 AU [@Chadney2015]. Peak temperatures in the thermosphere are about 2,000 K at 1 AU, 8,000 K at 0.5 AU and 10,500 K at 0.2 AU. Ion densities of the two major ions H$^+$ and H$_3^+$, determined at noon at 30$^{\circ}$ latitude, are plotted in Fig. \[fig:ni\_vs\_p\_adleo\_diffa\].
As the planet is moved closer to the star, overall ion densities increase with increased stellar radiation, but not in a linear fashion. Within 0.5 AU, the atmosphere escapes hydrodynamically. H$_2$ is fully dissociated below about 1 nbar at 0.5 AU and 3 nbar at 0.2 AU. H$_3^+$ cannot be formed below this dissociation pressure. At small orbital distances, the density of H$^+$ increases vastly in the lower ionosphere: at a pressure of 500 nbar, it increases from $2.7\times10^4$cm$^{-3}$ at 1 AU to $8.9\times10^5$cm$^{-3}$ at 0.5 AU and $7.1\times10^6$cm$^{-3}$ at 0.2 AU. The evolution of the peak density of H$_3^+$ with orbital distance is slightly more complicated. H$_3^+$ is effectively destroyed by electron-recombination (reactions 21 and 22, see Table \[tab:reactionsiono\]). Because much more H$^+$ is formed at high pressures, the increase in electron density results in a higher loss rate for H$_3^+$. Therefore, as the H$^+$ peak is pushed to higher pressures with increased stellar flux, the H$_3^+$ peak is reduced. At 0.2 AU, the H$_3^+$ peak is below the lower boundary of the ionospheric model. At higher pressures, the likelihood that H$_3^+$ will further be destroyed through reaction with heavy species is higher. At small orbital distances H$_3^+$ is therefore only likely to be a minor constituent of the ionosphere at all pressure levels.
We obtained a rough estimate of whether escaping heavy species from the stratosphere can be present in the thermosphere by applying the cross-over mass equation from @Hunten1987. For an EGP at 0.2 AU from AD Leo, the limiting mass-loss rate required for C to be present at a pressure of 100 nbar is $2.7\times 10^6$ kg s$^{-1}$; and at this same pressure level, the upwelling mass flux is $3.1\times 10^6$ kg s$^{-1}$. It is thus likely that carbon species are present at this pressure level, which will deplete the H$_3^+$ layer in planets orbiting within 0.2 AU from AD Leo. These results indicate that close-orbiting hot-Jupiter exoplanets are not good candidates for observations of H$_3^+$ emissions. The level of H$_3^+$ IR emissions from the EGP atmospheres modelled in this work is presented in Sect. \[sec:H3Pemissions\].
Scaled solar spectra versus synthetic spectra {#sec:scaled_spec}
---------------------------------------------
{width="98.00000%"}
@Chadney2015 derived a new scaling of the solar XUV spectrum to match the spectra of other low-mass stars, using different scaling factors for the X-ray and EUV bandpasses. These scaled spectra were tested against synthetic spectra derived from the coronal model in terms of the neutral atmosphere. It was found that a scaling of the solar spectrum, using two different scaling factors, was adequate to estimate neutral densities and planetary mass-loss rates. In this section, we undertake the same comparison (synthetic to scaled solar spectra) using the ionospheric model. All of the model runs in this section are for planets with an orbital distance of 1 AU. At this orbital distance, the upper atmosphere of EGPs around AU Mic undergo hydrodynamic escape, whereas those orbiting $\epsilon$ Eri and AD Leo are in the classical Jeans escape regime.
Ion densities for planets irradiated by either the synthetic spectra or the scaled solar spectra corresponding to each of the three stars are shown in Fig. \[fig:ni\_vs\_p\_ss\]. Ion densities determined using the synthetic spectra are plotted as solid lines and those determined using the scaled spectra are plotted as dashed lines. All of the densities shown in this figure are determined at noon at 30$^{\circ}$ latitude.
At high pressures, more ions are present in the planets irradiated by the scaled spectra. This results in a lowering of the peak in H$^+$ density in planets orbiting all three stars. There is roughly an order of magnitude more H$^+$ at 200 nbar in planets around $\epsilon$ Eri and AD Leo and at 300 nbar in planets around AU Mic. However, at high pressures, densities of H$_3^+$ are either the same or slightly lower when using the scaled spectra. The density of H$_3^+$ is determined by a competition between its production from H$_2^+$ through reaction 12 (see Table \[tab:reactionsiono\]) and its loss through electron-recombination (reactions 21 and 22). At high pressures (between about 50 and 500 nbar), there is a significant excess of H$^+$ formed through photo- and electron-impact ionisation with the scaled spectra when compared to using the synthetic spectra. Thus there are also significantly more electrons at these pressures and the increased rates of reactions 21 and 22 counterbalance the increased production of H$_3^+$.
At low pressures, the densities of the minor species H$_3^+$, H$_2^+$, and He$^+$ are higher in the planets irradiated by synthetic spectra than in those irradiated by the scaled spectra because of more photo- and impact-ionisation. Densities of the major ion H$^+$ above about 10 nbar are very similar whether using the synthetic or the scaled solar spectra.
The difference in terms of ion production and densities is greater between using the scaled solar spectra and the synthetic stellar spectra than in terms of neutral densities. H$^+$ and electron densities vary by about an order of magnitude in the lower ionosphere. However, the density of H$_3^+$ is relatively well regulated and does not vary dramatically at pressures where it is most abundant. This means that at least in the cases studied here, cooling through IR H$_3^+$ emissions is not significantly affected by using scaled solar spectra. Hence the regime of atmospheric escape (and associated mass-loss rates) are well determined using scaled solar spectra with the thermospheric model.
H$_3^+$ emissions {#sec:H3Pemissions}
=================
We determined the total output power from the H$_3^+$ ion and the output in specific spectral lines for EGPs around a Sun-like star (Table \[tab:H3p\_power\_sun\]), $\epsilon$ Eri (Table \[tab:H3p\_power\_eeri\]), AD Leo (Table \[tab:H3p\_power\_adleo\]), and AU Mic (Table \[tab:H3p\_power\_aumic\]). The total output power is plotted in Fig. \[fig:H3p\_emission\_power\] for EGPs at different orbital distances around each star. The values given here were calculated using globally averaged H$_2$ and H$_3^+$ densities and considering emission from the planet over $2\pi$ steradians.
![Total H$_3^+$ emission power from EGPs orbiting various stars at different orbital distances.[]{data-label="fig:H3p_emission_power"}](H3Pemission_power){width="49.00000%"}
As a result of enhanced ionisation and hence increased H$_3^+$ column densities, the output power is higher around more active stars at a given orbital distance. For example, the total output from H$_3^+$ for a planet at 0.1 AU from AD Leo is $1.4\times10^{16}$ W, whereas for an EGP at the same distance from the Sun, it would be $2.4\times10^{15}$ W. Likewise, for a planet orbiting a given star, smaller orbital distances mean higher H$_3^+$ emissions. This is because our model predicts higher temperatures and H$_3^+$ column densities close to the star.
@Koskinen2007a determined H$_3^+$ emission powers for an EGP at different orbital distances around a solar-like star (HD209458) and obtained very similar values to our results for a planet orbiting the Sun. @Yelle2004 calculated an output power of $9.9\times10^{16}$ W for HD209458b, an EGP orbiting at 0.047 AU, which is also consistent with our findings.
It would appear that the best targets for detecting H$_3^+$ emissions are planets orbiting close-in to active stars with high XUV irradiances (such as AD Leo or AU Mic). However, this may not be the case: in the closest planets orbiting the most active stars, we have found that the H$_3^+$ layer is pushed to higher pressures than for planets in the stable escape regime at larger orbital distances (see Sect. \[sec:vary\_a\]). At these higher pressures, it is likely that the presence of heavy species (not yet included in our models) will destroy H$_3^+$. This means that in particular for EGPs at small orbital distances from AU Mic and AD Leo, the values of output power from H$_3^+$ presented here are likely overestimated. Therefore we did not determine H$_3^+$ emissions from an EGP at 0.1 AU from AU Mic.
Past observational campaigns have failed to detect these emissions from hot Jupiters. @Shkolnik2006 observed six low-mass stars (of type F, G, K, and M) using the NASA Infrared Telescope Facility’s Cryogenic Near-IR Facility Spectrograph (CSHELL). They determined emission limits for each observed system. The lowest limit they found is $6.3\times10^{17}$ W for GJ436b, a hot Jupiter orbiting an M dwarf located at 10.2 pc from Earth. According to our calculations, the emission limits set by the observations made by @Shkolnik2006 are too high to detect H$_3^+$ emissions.
Measurements of planetary absorption line depths as low as $10^{-4}$ with respect to the stellar continuum are becoming possible with ground-based high-resolution spectrographs such as CRIRES on the VLT [see e.g. @DeKok2013; @Brogi2012]. Based on our model predictions for H$_3^+$ emissions, we estimate the planetary emission to stellar continuum contrast to be approximately $7\times 10^{-7}$ for a planet at 0.2 AU from a star similar to AD Leo, dropping to $3\times 10^{-8}$ at 1 AU from the same star. These estimates are based on the 3.534 $\mu$m R(3,3-) line, for a spectral resolution of about 100,000 and on the simplistic assumption that the planet’s continuum emission can be approximated to first order by a black body at the equilibrium temperature. Therefore, even accounting for gains by observing multiple lines, the detection of H$_3^+$ emissions from EGPs is not possible with current technologies. The difficulties are further compounded by the relatively small number of H$_3^+$ emission lines (of the order of tens rather than thousands as in the case of CO absorption lines) that does not lead to the same gain in the cross-correlations. Furthermore, simulations by @DeKok2013 suggested that signatures from emission lines might be more difficult to retrieve than from absorption lines (see their Fig. 5).
A number of recent studies (@Snellen2015 [@Kawahara2014]; see also @Sparks2002) have suggested different methods to combine high spatial and high spectral resolution imaging techniques, with simulations suggesting contrast improvements of between a factor of 100 and 1000 depending on the technique and planet-star configuration. If these are indeed realised, line contrasts as low as $10^{-7}$ will become detectable and bring direct detections of H$_3^+$ in planetary systems such as the ones modelled here within our reach.
The presence of strong aurora may also increase the detectability of IR emissions from EGPs. There are currently no constraints on exoplanetary magnetospheres, but it is possible that these could induce emissions similar to the powerful auroral emissions seen at Jupiter [e.g. @Radioti2013; @Miller2006].
------- -------------------- -------------------- -------------------- --------------------
Sun
0.1 AU 0.2 AU 0.5 AU 1 AU
Q(1) $1.4\times10^{13}$ $6.4\times10^{12}$ $1.5\times10^{12}$ $3.8\times10^{11}$
Q(3) $9.5\times10^{12}$ $4.9\times10^{12}$ $1.2\times10^{12}$ $3.0\times10^{11}$
R(1) $1.1\times10^{13}$ $5.2\times10^{12}$ $1.3\times10^{12}$ $3.2\times10^{11}$
R(3) $1.5\times10^{13}$ $6.9\times10^{12}$ $1.7\times10^{12}$ $4.1\times10^{11}$
Total $2.4\times10^{15}$ $7.0\times10^{14}$ $1.6\times10^{14}$ $2.5\times10^{13}$
------- -------------------- -------------------- -------------------- --------------------
: H$_3^+$ emission power for an EGP orbiting a Sun-like star. All units are Watts.
\[tab:H3p\_power\_sun\]
------- -------------------- -------------------- -------------------- --------------------
$\epsilon$ Eri
0.1 AU 0.2 AU 0.5 AU 1 AU
Q(1) $3.5\times10^{13}$ $1.5\times10^{13}$ $2.5\times10^{12}$ $7.5\times10^{11}$
Q(3) $2.4\times10^{13}$ $1.1\times10^{13}$ $2.0\times10^{12}$ $5.6\times10^{11}$
R(1) $2.7\times10^{13}$ $1.2\times10^{13}$ $2.1\times10^{12}$ $6.0\times10^{11}$
R(3) $3.9\times10^{13}$ $1.7\times10^{13}$ $2.7\times10^{12}$ $8.1\times10^{11}$
Total $6.5\times10^{15}$ $2.0\times10^{15}$ $2.7\times10^{14}$ $7.4\times10^{13}$
------- -------------------- -------------------- -------------------- --------------------
: H$_3^+$ emission power for an EGP orbiting $\epsilon$ Eri. All units are Watts.
\[tab:H3p\_power\_eeri\]
------- -------------------- -------------------- -------------------- --------------------
AD Leo
0.1 AU 0.2 AU 0.5 AU 1 AU
Q(1) $7.0\times10^{13}$ $2.8\times10^{13}$ $4.7\times10^{12}$ $9.9\times10^{11}$
Q(3) $4.6\times10^{13}$ $1.9\times10^{13}$ $3.7\times10^{12}$ $7.6\times10^{11}$
R(1) $5.3\times10^{13}$ $2.1\times10^{13}$ $3.9\times10^{12}$ $8.1\times10^{11}$
R(3) $7.8\times10^{13}$ $3.0\times10^{13}$ $5.1\times10^{12}$ $1.1\times10^{12}$
Total $1.4\times10^{16}$ $3.7\times10^{15}$ $4.5\times10^{14}$ $9.1\times10^{13}$
------- -------------------- -------------------- -------------------- --------------------
: H$_3^+$ emission power for an EGP orbiting AD Leo. All units are Watts.
\[tab:H3p\_power\_adleo\]
------- -------------------- -------------------- --------------------
AU Mic
0.2 AU 0.5 AU 1 AU
Q(1) $6.7\times10^{13}$ $1.4\times10^{13}$ $2.7\times10^{12}$
Q(3) $4.4\times10^{13}$ $1.0\times10^{13}$ $2.1\times10^{12}$
R(1) $5.0\times10^{13}$ $1.1\times10^{13}$ $2.2\times10^{12}$
R(3) $7.5\times10^{13}$ $1.5\times10^{13}$ $3.0\times10^{12}$
Total $1.2\times10^{16}$ $1.7\times10^{15}$ $2.9\times10^{14}$
------- -------------------- -------------------- --------------------
: H$_3^+$ emission power for an EGP orbiting AU Mic. All units are Watts.
\[tab:H3p\_power\_aumic\]
Conclusion
==========
This paper builds upon @Chadney2015, where the authors discussed the neutral atmosphere and escape regimes of EPGs around these stars. Here, we have studied the ionised region of EGP upper atmospheres by applying an ionospheric model to planets irradiated by XUV radiation from the Sun at solar minimum and maximum, as well as with the stars $\epsilon$ Eri, AD Leo and AU Mic. As in solar system gas giants [e.g. @Moore2004; @Moore2009], we found that the dominant ions in the EUV-driven part of the ionosphere on EGPs are H$^+$ and H$_3^+$. In planets orbiting at large orbital distances, the upper atmosphere is in the ‘stable’ regime (i.e. undergoing Jeans escape). In this situation, there is a region in the lower ionosphere where H$_3^+$ is the dominant ion species. Peak H$_3^+$ number densities are reached in layers where stellar soft X-ray fluxes are absorbed, whereas the peak in H$^+$ is located in the EUV heating layer. Thus the highest value of $n_{\text{H}_3^+}$ is determined by stellar soft X-ray flux levels and the highest value of $n_{\text{H}^+}$ is determined by stellar EUV flux levels. This is a noteworthy difference, since stellar fluxes in the X-ray and EUV bands scale differently with stellar activity [see @Chadney2015].
As the XUV stellar flux is increased, whether because of a more active star or a smaller orbital distance, atmospheric temperatures and flux levels become high enough for the planet to shift to a regime of hydrodynamic escape [see @Chadney2015]. This leads to a change in temperature profile (a very large temperature peak is formed) and neutral density structure (molecular hydrogen is fully dissociated above a certain pressure level within the ionosphere). In this situation, the H$_3^+$ peak is pushed to higher pressures, mainly as a result of the increased photo- and electron-impact ionisation of H$^+$ at higher pressure levels (H$_3^+$ is destroyed though recombination with the increased number of electrons). At high enough stellar flux levels, H$^+$ is the dominant ion throughout the modelled region of atmosphere. Any peak in H$_3^+$ would then be confined to pressures below the lower boundary of the model, where it is likely destroyed though reactions with heavy species (e.g. water, hydrocarbons).
We here calculated the photo-electron energy degradation to determine ionisation by photo-electrons. This has not been included before in EGP studies, and we showed that this considerably affects ionisation below 10 nbar, where it is the dominant form of ionisation, pushing the ionosphere to lower altitudes than previously assumed.
The H$_3^+$ IR emissions we predict from EGPs around active stars are higher than for planets orbiting at the same distance from a star of similar age to the Sun. However, the emissions may still be too low to be detected by the current technologies. The past focus on detecting these emissions from hot Jupiters might not be the optimum strategy. It relies on the assumption that closer orbiting EGPs will have higher column densities of H$_3^+$ and higher temperatures, leading to higher emissions. This assumption is valid in a pure H/H$_2$/He atmosphere, such as that considered in this study. However, given that for EGPs orbiting very close to their stars, we predict that H$_3^+$ is confined to a layer at the bottom of the ionosphere, possibly below the homopause, it is likely destroyed by heavy species. The location of the homopause is dependent on the value of the eddy diffusion coefficient, which is not well known in exoplanets. The predictions of H$_3^+$ densities in our simulations rely on the homopause being located at a pressure of 1 $\mu$bar, meaning an eddy diffusion coefficient $K_{zz}$ lower than $\sim10^2$ – $10^3$ m$^2$ s$^{-1}$. Higher values of $K_{zz}$ will mean that hydrocarbons could be present in the H$_3^+$ layer. Therefore a detection of H$_3^+$ emissions would also place constraints on $K_{zz}$.
Given the low levels of H$_3^+$ emission that we predict, these may be difficult to detect with current telescopes, even with planets around active stars. However, new techniques involving high spatial and high spectral resolution may provide sufficient planet-to-star contrast to allow a detection of H$_3^+$ emissions. Despite this, other diagnostics of EGP ionospheres could prove to be more promising in the near future, such as the detection of radio emissions from magnetosphere-ionosphere coupling [@Nichols2011]. Although these observations have yet to succeed for EGPs, radio emissions emanating from a brown dwarf have recently been detected [@Hallinan2015].
J.M.C., M.G., and Y.C.U. are partially funded by the UK Science & Technology Facilities Council (STFC) through the Consolidated Grant to Imperial College London. J.M.C. has also received support from SFTC through a postgraduate studentship, award number ST/J500616/1. T.T.K. acknowledges support from the National Science Foundation (NSF) grant AST 1211514. J.S.F. acknowledges support from the Spanish MINECO through grant AYA2011-30147-C03-03.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The resilience of Supervisory Control and Data Acquisition (SCADA) systems for electric power networks for certain cyber-attacks is considered. We analyze the vulnerability of the measurement system to false data attack on communicated measurements. The vulnerability analysis problem is shown to be NP-hard, meaning that unless $P = NP$ there is no polynomial time algorithm to analyze the vulnerability of the system. Nevertheless, we identify situations, such as the full measurement case, where it can be solved efficiently. In such cases, we show indeed that the problem can be cast as a generalization of the minimum cut problem involving costly nodes. We further show that it can be reformulated as a standard minimum cut problem (without costly nodes) on a modified graph of proportional size. An important consequence of this result is that our approach provides the first exact efficient algorithm for the vulnerability analysis problem under the full measurement assumption. Furthermore, our approach also provides an efficient heuristic algorithm for the general NP-hard problem. Our results are illustrated by numerical studies on benchmark systems including the IEEE 118-bus system.'
author:
- 'Julien M. Hendrickx, Karl Henrik Johansson, Raphael M. Jungers, Henrik Sandberg and Kin Cheong Sou [^1]'
bibliography:
- 'gmc\_tsg.bib'
- 'exact\_l1.bib'
- 'references\_all.bib'
title: '****'
---
Introduction
============
Our society depends heavily on the proper operation of cyber-physical systems, examples of which include, but not limited to, intelligent transport systems, industrial automation systems, health care systems, and electric power distribution and transmission systems. These cyber-physical systems are supervised and controlled through Supervisory Control And Data Acquisition (SCADA) systems. Through remote terminal units (RTUs), SCADA systems collect measurements and send them to the state estimator to estimate the system states. The estimated states are used for subsequent operations such as system health monitoring and control. Any malfunctioning of these operations can lead to significant social and economical consequences such as the northeast US blackout of 2003 [@US_blackout2003].
The technology and the use of the SCADA systems have evolved a lot since they were introduced. The SCADA systems now are interconnected to office LANs, and through them they are connected to the Internet. Hence, today there are more access points to the SCADA systems, and also more functionalities to tamper with. For example, the RTUs can be subjected to denial-of-service attacks. The communicated data can be subjected to false data attacks. Furthermore, the SCADA master itself can be attacked. In the context of secured cyber-physical systems in general, [@AminCardenasSastry-HSCC-2009; @kn:Gupta2010] have considered denial-of-service-like attacks and their impact. [@kn:Bruno09] has studied replay attacks on the sensor measurements and [@Smith-IFAC-2011; @kn:Pasqualetti2011] have considered false data attacks. This paper investigates the cyber security issues related to false data attacks with the special focus on the measurement systems of power networks. The negative effects of false data attacks on power networks have been exemplified by malware such as Stuxnet and Duqu. False data attacks have received a lot of attention in the literature (e.g., [@LRN09; @STJ_SCS2010; @DS_SGC2010; @BRWKNO10; @KJTT11; @SSJ_CDC2011_mincut; @GBMKP11; @KP11]). [@LRN09] was the first to point out that a coordinated intentional data attack can be staged without being detected by the state estimation bad data detection (BDD) algorithm, a standard part of today’s SCADA/EMS system. [@LRN09; @STJ_SCS2010; @DS_SGC2010; @KJTT11; @SSJ_CDC2011_mincut; @GBMKP11; @KP11] investigate the construction problem for such “unobservable” data attack, especially the sparse ones involving relatively few meters to compromise, under various assumptions of the network (e.g., DC power flow model [@Abur_Exposito_SEbook; @Monticelli_SEbook]). In particular, [@LRN09] poses the attack construction problem as a cardinality minimization problem to find the sparsest attack including a given set of target measurements. References [@STJ_SCS2010; @DS_SGC2010; @SSJ_CDC2011_mincut] set up similar optimization problems for the sparsest attack including a given measurement. References [@KJTT11; @KP11] seek the sparsest nonzero attack and [@GBMKP11] finds the sparsest attack including exactly two injection measurements and possibly more line power flow measurements, under the assumption that all power injections are measured. The solution information of the above optimization problems can help network operators identify the vulnerabilities in the network and strategically assign protection resources (e.g., encryption of measurements and secured PMUs) to their best effect (e.g., [@DS_SGC2010; @BRWKNO10; @KP11]). On the other hand, the unobservable data attack problem has its connection to another vital EMS functionality – observability analysis [@Abur_Exposito_SEbook; @Monticelli_SEbook]. In particular, solving the attack construction problem can also solve an observability analysis problem (this was explained in [@SSJ_exact_l1 Section II-C]). This connection was first reported in [@KJTT11], and was utilized in [@SSJ_ckt] to compute the sparsest critical $p$-tuples for some integer $p$. This is a generalization of critical measurements and critical sets [@Abur_Exposito_SEbook].
To perform the cyber-security analysis in a timely manner, it is important to solve the data attack construction problem efficiently. This effort has been discussed, for instance, in [@LRN09; @STJ_SCS2010; @DS_SGC2010; @KJTT11; @SSJ_CDC2011_mincut; @GBMKP11; @KP11; @SSJ_exact_l1]. The efficient solution to the attack construction problem in [@STJ_SCS2010] is the focus of this paper. The matching pursuit method [@Mallat93matchingpursuit] employed in [@LRN09] and the basis pursuit method [@CDS_Basis_Pursuit] ($l_1$ relaxation and its weighted variant) employed in [@KP11] are common efficient (i.e., polynomial time) approaches to suboptimally solve the attack construction problem. However, these methods do not guarantee exact optimal solutions, and they might not be sufficiently accurate. For instance, [@SSJ_CDC2011_mincut] describes a naive application of basis pursuit and its consequences. While [@KJTT11; @KP11] provide polynomial time solution procedures for their respective attack construction problems, the problems therein are different from the one in this paper in that the considered problem in this paper is not a special case of the ones in [@KJTT11; @KP11]. In particular, in [@KJTT11] the attack vector contains at least one nonzero entry. However, this nonzero entry cannot be given a priori. This means that the problem considered in this paper is more general than the one in [@KJTT11]. [@KP11] needs to restrict the number of nonzero injection measurements attacked, while there is no such constraint in the problem considered in this paper.
Other relevant previous work include [@SSJ_ckt; @SSJ_CDC2011_mincut; @SSJ_exact_l1], which also consider the data attack construction problem in this paper. In [@SSJ_ckt; @SSJ_CDC2011_mincut] the attack construction problem is formulated as a graph generalized minimum cut problem (to be defined in Section \[subsec:gmc\_costly\_nodes\]). However, it is not known in [@SSJ_ckt; @SSJ_CDC2011_mincut] whether the generalized minimum cut problem can be solved efficiently (i.e., in polynomial time) or not. Indeed, [@SSJ_ckt; @SSJ_CDC2011_mincut] only provide approximate solutions. Instead, the current work establishes that the generalized minimum cut problem is indeed *exactly solvable in polynomial time*. This work establishes the result by constructing a practical polynomial time algorithm. Regarding [@SSJ_exact_l1], one of the main distinctions is that [@SSJ_exact_l1] makes an assumption that no bus injections are metered. The current result requires a different assumption that the network is *fully measured* as in [@GBMKP11] (i.e., all bus injections and line power flows are metered). In addition, [@SSJ_exact_l1] considers a more general case where the constraint matrix is totally unimodular, whereas the focus of the current paper is a graph problem. The setup considered by this paper is specific to power network applications and thus it enables a more efficient solution algorithm. Finally, note that the notion of minimum cut problem has been explored also in other power network applications (e.g., [@LRDP06; @inhibitbisect; @ETEP:ETEP255]).
[**Outline:**]{} In the next section, we present the optimization problem of interest, namely the security index problem, and discuss its applications. Then Sections \[sec:complexity\], \[sec:si\_full\_measurement\] and \[section-reformulation\] present the technical contributions of this paper, focusing on a specialized version of the security index problem defined in (\[opt:security\_index\]) in the end of Section \[subsec:security\_index\]. In Section \[sec:complexity\] the complexity of the security index problem is analyzed. We show that the security index problem is NP-hard in general, but in Section \[sec:si\_full\_measurement\] we demonstrate that under some realistic assumptions it can be restated as a generalized minimum cut ([Min Cut]{}) problem. In Section \[section-reformulation\] we show that the generalized [[Min Cut]{} ]{}problem can be solved efficiently, by reformulating it as a classical [[Min Cut]{} ]{}problem. The specialized version considered in Sections \[sec:complexity\], \[sec:si\_full\_measurement\] and \[section-reformulation\] turns out to be not restrictive, as far as the application of the proposed results is concerned. This will be explained in Section \[sec:si\_edge\_only\]. In Section \[section-example\] a simple numerical example is first presented to illustrate that the proposed solution correctly solves the generalized [[Min Cut]{} ]{}problem, while previous methods cannot. Then the efficiency and accuracy of the proposed solution to the security index problem are demonstrated through a case study with large-scale benchmark systems. We also demonstrate that our method provides an efficient and high quality approximate solution to the general problem security index problem which is NP-hard.
The Security Index Problem
==========================
In Section \[subsec:model\] the mathematical model of the power networks considered is first described. Then in Section \[subsec:security\_index\] the security index of power networks is defined.
Power Network Model and State Estimation {#subsec:model}
----------------------------------------
A power network is modeled as a graph with $n+1$ nodes and $m$ edges. The nodes and edges model the buses and transmission lines in the power network, respectively. In the present text, the terms node and bus are used interchangeably, and the same is true for edges and transmission lines (or simply lines). The topology of the graph is described by a directed incidence matrix $A \in \mathbb{R}^{(n+1) \times m}$, in which the directions along the edges are arbitrarily specified [@SSJ_CDC2011_mincut]. The physical property of the network is described by a nonsingular diagonal matrix $D \in \mathbb{R}^{m \times m}$, whose nonzero entries are the reciprocals of the reactance of the transmission lines. In general, the reactance is positive (i.e., inductive) and hence the matrix $D$ is assumed to be positive definite throughout this paper.
In the sequel, the set of all nodes and the set of all directed edges of the power network graph are denoted $V^0$ and $E^0$, respectively. The edge directions are consistent with those in $A$. An element of $V^0$ is denoted by $v_i \in V^0$, and an element of $E^0$ is denoted by $(v_i, v_j) \in E^0$ for $v_i \in V^0$ and $v_j \in V^0$. The set of all neighbors of $v_i$ is denoted by $N(v_i)$. A node $v_j$ is a neighbor of $v_i$ if either $(v_i, v_j) \in E^0$ or $(v_j, v_i) \in E^0$.
The states of the network include bus voltage phase angles and bus voltage magnitudes, the latter of which are typically assumed to be constant (one in the per unit system). Therefore, the network states can be captured in a vector $\theta \in {[0,2\pi)}^{n+1}$. The state estimator estimates $\theta$ based on the measurements obtained from the network. In reality the model relating the states and the measurements is nonlinear. However, for state estimation data attack analysis [@LRN09; @STJ_SCS2010; @DS_SGC2010; @BRWKNO10; @KJTT11; @SSJ_CDC2011_mincut; @GBMKP11; @KP11; @SSJ_exact_l1] (and more traditionally bad data analysis [@Abur_Exposito_SEbook; @Monticelli_SEbook; @HSKF75]) it suffices to consider the DC power flow model [@Abur_Exposito_SEbook; @Monticelli_SEbook]. In the DC power flow model the measurement vector, denoted as $z$, is related to $\theta$ by $$\label{def:H_matrix}
z = H \theta + \Delta z, \quad {\rm where} \quad H \triangleq \begin{bmatrix}
P_1 D A^T \\ -P_2 D A^T \\ P_3 A D A^T
\end{bmatrix}.$$ In (\[def:H\_matrix\]), $\Delta z$ can either be a vector of random error or intentional additive data attack (e.g., [@LRN09]), and $P_1$, $P_2$ and $P_3$ consist of subsets of rows of identity matrices of appropriate dimensions, indicating which measurements are actually taken. The term $P_1 D A^T \theta$ contains line power flow measurements, measured at the outgoing ends of the lines. Similarly, $-P_2 D A^T \theta$ contains the line power flow measurements at the incoming ends of the lines. The term $P_3 A D A^T \theta$ contains bus power injection measurements, one entry for each measured bus.
Measurement redundancy is a common practice in power networks [@Abur_Exposito_SEbook; @Monticelli_SEbook]. Therefore, it is assumed in this paper that the measurement system described by $H$ is observable – meaning that if any column of $H$ is removed the remaining submatrix still has rank $n$ [@Abur_Exposito_SEbook; @Monticelli_SEbook]. Note that $H$ cannot have rank $n+1$ since the sum of all columns of $H$ is always a zero column vector (a property of any incidence matrix $A$). In the practice of power system state estimation, it is customary to designate an arbitrary node as the reference and set the corresponding entry of $\theta$ to zero. Without loss of generality, it is assumed that the first entry of $\theta$ is zero (i.e., $\theta(1) = 0$) and denote $\theta_{2:}$ as the rest of the entries of $\theta$. For convenience, let $H_{2:}$ denote $H$ with the first column removed. By definition, $H \theta = H_{2:} \theta_{2:}$ and $H_{2:}$ has full column rank ($=n$) since $H$ is observable. Given measurements $z$, the estimate of the network states is typically determined via the least squares approach [@Abur_Exposito_SEbook; @Monticelli_SEbook]: $$\label{def:theta_hat}
\hat{\theta}_{2:} = {({H_{2:}}^T W H_{2:})}^{-1} {H_{2:}}^T W z,$$ where $W$ is a given positive-definite diagonal matrix, whose nonzero entries are typically the reciprocals of the variances of the measurement noise. The state estimate $\hat{\theta} = {\begin{bmatrix} 0 & {\hat{\theta}_{2:}}^T \end{bmatrix}}^T$ is subsequently fed to other vital SCADA functionalities such as optimal power flow (OPF) calculation and contingency analysis (CA). Therefore, the accuracy and reliability of $\hat{\theta}$ is of paramount concern.
Security Index {#subsec:security_index}
--------------
To detect possible faults or data attacks in the measurements $z$, the BDD test is commonly performed ([@Abur_Exposito_SEbook; @Monticelli_SEbook]). In a typical strategy, if the norm of the residual $$\label{def:BDD_residual_norm}
{\rm residual} \triangleq z - H_{2:} \hat{\theta}_2 = (I - H_{2:}{({H_{2:}}^T W H_{2:})}^{-1} {H_{2:}}^T W) \Delta z$$ is too big, then the BDD alarm is triggered. The BDD test is in general sufficient to detect the presence of a random error $\Delta z$ [@Abur_Exposito_SEbook; @Monticelli_SEbook]. However, in face of a coordinated malicious attack the BDD test can fail. In particular, in [@LRN09] it was reported that an attack of the form $$\label{def:unobservable_attack}
\Delta z = H \Delta \theta$$ for an arbitrary $\Delta \theta \in \mathbb{R}^{n+1}$ would result in a zero residual in (\[def:BDD\_residual\_norm\]) since $H \Delta \theta = H_{2:} \Delta \theta_{2:}$ for some $\Delta \theta_{2:} \in \mathbb{R}^n$. Data attack in the form of (\[def:unobservable\_attack\]) is unobservable from the BDD perspective, and this was also experimentally verified in [@TDSJ11] in a realistic SCADA system testbed. Since [@LRN09], there has been a significant amount of literature studying the unobservable attack in (\[def:unobservable\_attack\]) and its consequences to state estimation data integrity (e.g., [@STJ_SCS2010; @DS_SGC2010; @BRWKNO10; @KJTT11; @GBMKP11; @KP11]). In particular, [@STJ_SCS2010] introduced the notion of security index $\alpha_k$ for a measurement $k$ as the optimal objective value of the following cardinality minimization problem: $$\label{opt:card_min_con}
\begin{array}{cccl}
\alpha_k & \triangleq & \mathop{\rm min}\limits_{\Delta \theta \in \mathbb{R}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
& & \textrm{subject to} & H(k,:) \Delta \theta = 1,
\end{array}$$ where ${\rm card}(\cdot)$ denotes the cardinality of its argument, $k$ is the label of the measurement for which the security index $\alpha_k$ is computed, and $H(k,:)$ denotes the $k^{\rm th}$ row of $H$. $\alpha_k$ is the minimum number of measurements an attacker needs to compromise in order to attack measurement $k$ without being detected. In particular, a small $\alpha_k$ implies that measurement $k$ is relatively easy to compromise in an unobservable attack. As a result, the knowledge of the security indices for all measurements allows the network operator to pinpoint the security vulnerabilities of the network, and to better protect the network with limited resource. For example, [@DS_SGC2010] proposed a method to optimally assign limited encryption protection resources to improve the security of the network based on its security indices.
It should be emphasized that the security index defined in (\[opt:card\_min\_con\]) can provide a security assessment that the standard power network BDD procedure [@Abur_Exposito_SEbook; @Monticelli_SEbook] might not be able to provide. As a concrete example [@STJ_SCS2010], consider the simple network whose $H_{2:}$ matrix is $$H_{2:} = \begin{pmatrix} -1 & -1 & 0 \\ -1 & 0 & 0 \\ 1 & 0 & 0 \\
1 & 0 & -1 \\ 0 & -1 & 0
\end{pmatrix}.
\label{eq:Hex}$$ From (\[def:theta\_hat\]), the “hat matrix” [@Abur_Exposito_SEbook; @Monticelli_SEbook], denoted $K$ is defined according to $$\hat{z} = H_{2:} \hat{\theta}_2 = H_{2:} {({H_{2:}}^T W H_{2:})}^{-1} {H_{2:}}^T W z \triangleq K z.$$ Assuming $W = I$, the $K$ matrix associated with $H_{2:}$ in (\[eq:Hex\]) is $$K = \begin{pmatrix} 0.6& 0.2 & -0.2& 0 & 0.4 \\
0.2 & 0.4 & -0.4 & 0 & -0.2 \\
-0.2 & -0.4 & 0.4 & 0 & 0.2 \\
0 & 0 & 0 & 1 & 0 \\
0.4 & -0.2 & 0.2 & 0 & 0.6
\end{pmatrix}.
\label{eq:Kex}$$ The hat matrix $K$ shows how the measurements $z$ are weighted together to form a power flow estimate $\hat{z}$. The rows of the hat matrix can be used to study the measurement redundancy in the system [@Abur_Exposito_SEbook; @Monticelli_SEbook]. Typically a large degree of redundancy (many non-zero entries in each row) is desirable to compensate for noisy or missing measurements. In (\[eq:Kex\]), it is seen that all measurements are redundant except the fourth which is called a *critical measurement*. Without the critical measurement observability is lost. From the hat matrix one is led to believe that the critical measurement is sensitive to attacks. This is indeed the case, but some other measurements can also be vulnerable to attacks. It can be shown - for example using the method that we develop, that the security indices $\alpha_k$, $k = 1,\ldots,5$, respectively, are 2, 3, 3, 1, 2. Therefore, the fourth measurement (critical measurement) has security index one, indicating that it is indeed vulnerable to unobservable attacks. However, the first and the last measurements also have relatively small security indices. This is not obvious from $K$ in (\[eq:Kex\]). Hence, we cannot rely on the hat matrix for vulnerability analysis of power networks.
For ease of exposition but without loss of generality, instead of (\[opt:card\_min\_con\]) the following version of the security index problem with a specialized constraint will be the focus of the parts of the paper where the main technical contributions are presented (i.e., Sections \[sec:complexity\], \[sec:si\_full\_measurement\] and \[section-reformulation\]): $$\label{opt:security_index}
\begin{array}{cl}
\mathop{\rm minimize}\limits_{\Delta \theta \in {\mathbb{R}}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
\textrm{subject to} & A(:,e)^T \Delta \theta \neq 0,
\end{array}$$ where $e \in \{1,2,\ldots,m\}$ is given. The restriction introduced in (\[opt:security\_index\]) is that it can only enforce constraints on edge flows but not on node injections as directly allowed by (\[opt:card\_min\_con\]). We will see however in Section \[sec:si\_edge\_only\] that all results obtained for (\[opt:security\_index\]) can be extended to the general case in (\[opt:card\_min\_con\]).
The Security Index Problem is NP-hard {#sec:complexity}
=====================================
Consider a variant of (\[opt:card\_min\_con\]) where $k$ is not fixed (i.e., one wishes to minimize ${\rm card} \left(H \Delta \theta\right)$ under the constraint that at least one entry of $H \Delta \theta$ is nonzero). This variant of (\[opt:card\_min\_con\]) is known to be the cospark of $H_{2:}$ in compressed sensing [@CT05]. The cospark of $H_{2:}$ is the same as the spark of $F$, where $F$ is a matrix of full row rank such that $F H_{2:} = 0$ [@CT05]. The spark of $F$ is defined as the minimum number of columns of $F$ which are linearly dependent [@Donoho04032003]. It is established that computing the spark of a general matrix $F$ is NP-hard [@spark_NP_hard; @McCormick_PhD]. Consequently, because of the equivalence between spark and cospark, unless $P=NP$ there is no efficient algorithm to solve the security index problem in (\[opt:card\_min\_con\]) if the $H$ matrix is not assumed to retain any special structure. In power network applications, the $H$ matrix in fact possesses special structure as defined in (\[def:H\_matrix\]). Nevertheless, the security index problem, even the specialized version in (\[opt:security\_index\]), is still computationally intractable as indicated by the following statement:
\[thm:NP\_hardness\] Unless $P=NP$, there is no polynomial time algorithm that solves the problem , with $H$ defined in , even if $D$ is the identity matrix and $P_2=0$.
Our proof proceeds by reduction from the positive one-in-three 3SAT problem [@gopalan2006connectivity]: *Given a set of $M$ triples of indices $C_j = (\alpha_j,\beta_j,\gamma_j)\in \{1,\dots,n\}^3$, does there exist a vector $\tilde x\in \{0,1\}^n$ such that for every $j$, exactly one among $\tilde x_{\alpha_j}, \tilde x_{\beta_j}, \tilde x_{\gamma_j}$ is 1 and the others 0.*
Consider an instance of the positive one-in-three 3SAT problem, and let us build an equivalent instance of (\[opt:card\_min\_con\]). We set $P_2$ to 0, and set $D$ as the identity matrix. As a result, non-trivially zero entries of $H{\Delta \theta}$ corresponding to edges $(i,j)$ will be of the form ${\Delta \theta}_i - {\Delta \theta}_j$, while those corresponding to a node $i$ will be of the form $\sum_{j:(i,j)\in E} ({\Delta \theta}_j-{\Delta \theta}_i)$. We remind that the entry of an edge is trivially zero if the corresponding entry in $P_1$ is 0, and that of a node is trivially zero if the corresponding entry of $P_3$ is 0.
We begin by defining a node 1 and a node 0 connected by an edge whose corresponding entry in $P_1$ is set to 1. We set $k$ such that the constraint $H(k,:){\Delta \theta}= 1$ in corresponds to this edge, so that their must hold ${\Delta \theta}_1-{\Delta \theta}_0 = 1$ for any solution of the problem. Since $H{\Delta \theta}$ is not modified when adding a constant to all entries of ${\Delta \theta}$, we assume without loss of generality that ${\Delta \theta}_1 =1$ and ${\Delta \theta}_0 =0$.
The goal of the first part of our construction is to represent the variables. For every $i=1,\dots,n$, we define a node $x_i$ that we connect to both 1 and 0. We set to 1 the entries of $P_1$ corresponding to the edges $(1,x_i)$ and $(0,x_i)$, and to 0 the entries of $P_3$ corresponding to $x_i$. Observe that the two entries of $H{\Delta \theta}$ corresponding to these two edges are $1-{\Delta \theta}_{x_i}$ and $0-{\Delta \theta}_{x_i}$, which cannot be simultaneously 0. Moreover, one of them is equal to zero if and only if ${\Delta \theta}_{x_i}$ is either $0$ or $1$.
Taking into account the fact the entry of $H{\Delta \theta}$ corresponding to the edge $(1,0)$ is by definition 1, we have thus proved that ${\rm card}\left(H{\Delta \theta}\right) \geq n+1$ for any ${\Delta \theta}$, independently of the rest of the construction. Moreover, ${\rm card}\left(H{\Delta \theta}\right)=n+1$ only if ${\Delta \theta}_{x_i}\in \{0,1\}$ for every $i$, and if the entries of $H{\Delta \theta}$ corresponding to all the edges and nodes introduced in the sequel are 0. The remainder of the construction, represented in Fig. \[fig:construction\_NP\], is designed to ensure that all these entries can be 0 only if the (binary) values ${\Delta \theta}_{x_i}$ solves the initial instance of the positive one-in-three 3SAT problem. We first generate a reference value at $\frac{1}{3}$ for every clause: We define two nodes indexed by $\frac{2}{3}$ and $\frac{1}{3}$, and add the connections $(1,\frac{2}{3}),(\frac{2}{3},\frac{1}{3}),(\frac{1}{3},0)$. The entries of $P_1$ corresponding to these connections are set to 0, but the entry of $P_3$ corresponding to the nodes $\frac{1}{3}$ and $\frac{2}{3}$ are set to 1. Besides, we define for every clause $j=1,\dots,M$ a clause node $c_j$ connected to $\frac{1}{3}$ by an edge whose corresponding entry in $P_1$ is set to one.
The entries of $H{\Delta \theta}$ corresponding to the edges between $\frac{1}{3}$ and $c_j$ are ${\Delta \theta}_{\frac{1}{3}}-{\Delta \theta}_{c_j}$, which are thus zero only when ${\Delta \theta}_{c_j}={\Delta \theta}_{\frac{1}{3}}$ for every $j$. Using these equalities, observe now that the entry of $H{\Delta \theta}$ corresponding to $\frac{1}{3}$ is $${\Delta \theta}_{\frac{2}{3}} + {\Delta \theta}_{0} + \sum_{j=1}^M {\Delta \theta}_{c_j} - (2+M) {\Delta \theta}_{\frac{1}{3}} = {\Delta \theta}_{\frac{2}{3}} + 0 - 2 {\Delta \theta}_\frac{1}{3},$$ while the entry corresponding to $\frac{2}{3}$ is $1+{\Delta \theta}_{\frac{1}{3}} - 2{\Delta \theta}_{\frac{2}{3}}$. These two entries are thus equal to zero if and only if ${\Delta \theta}_{\frac{2}{3}} = \frac{2}{3}$ and ${\Delta \theta}_{c_j} ={\Delta \theta}_{\frac{1}{3}}= \frac{1}{3}$ for every $j$, as intended.
We now represent the clauses. For each $j$, we connect the clause node $c_j$ to the nodes $x_{\alpha_j}$, $x_{\beta_j}$ and $x_{\gamma_j}$ of the three variables involved by edges whose corresponding entries in $P_1$ are zero. On the other hand, we set to 1 the entry of $P_3$ corresponding to $c_j$. The corresponding (non trivially zero) entries of $H{\Delta \theta}$ are then $${\Delta \theta}{x_{\alpha_j}}+ {\Delta \theta}{x_{\beta_j}}+ {\Delta \theta}{x_{\gamma_j}}-3{\Delta \theta}_{c_j} =
{\Delta \theta}{x_{\alpha_j}}+ {\Delta \theta}{x_{\beta_j}}+ {\Delta \theta}{x_{\gamma_j}}-1.$$ Remembering that ${\Delta \theta}_{x_i}$ is either 1 or 0 for any $i$, this latter expression can be zero only if exactly one among ${\Delta \theta}{x_{\alpha_j}}$, ${\Delta \theta}{x_{\beta_j}}$ and ${\Delta \theta}{x_{\gamma_j}}$ is 1. If that is the case, setting $\tilde x_i= {\Delta \theta}_{x_i}$ for every $i$ yields a vector $x$ that solves the instance of positive one-in-three 3SAT.
We have thus shown that there exists a ${\Delta \theta}$ for which ${\rm card}\left(H {\Delta \theta}\right)=n+1$ only if the ${\Delta \theta}_{x_i}$ are binary, and if the binary vector $\tilde x$ obtained by setting $\tilde x_i = {\Delta \theta}_{x_i}$ solves the instance of positive one-in-three 3SAT. Conversely, one can verify that if a binary vector $\tilde x$ solves the instance of the one-in-three 3SAT problem, then setting ${\Delta \theta}_{x_i} = \tilde x_i$ for every $i$, ${\Delta \theta}_{c_j}={\Delta \theta}_{\frac{1}{3}} = \frac{1}{3}$ for every $j$ and ${\Delta \theta}_{\frac{2}{3}} = \frac{2}{3}$ yields a cost ${\rm card}\left(H{\Delta \theta}\right)=n+1$. The latter cost can thus be obtained if and only if the initial positive one-in-three 3SAT problem is achievable. This achieves the proof because our construction clearly takes an amount of time that grows polynomially with the size of the instance $C$, and unless $P=NP$ there is no polynomial time algorithm that solves the positive one-in-three 3SAT [@gopalan2006connectivity].
![Representation of a part of the construction of the proof of Theorem \[thm:NP\_hardness\], including the reference values of ${\Delta \theta}$ and one clause $C_j$. Edges are represented by dashed line when they are measured and continuous lines otherwise. Nodes are represented by squares when they are measured and circles otherwise. If ${\rm card}\left(H{\Delta \theta}\right)=n+1$, ${\Delta \theta}$ takes only values 1 and 0 for the $x_i$ and all entries of $H{\Delta \theta}$ other than those corresponding to the nodes $x_i$ must be zero. As a result, a dashed edge transmits no current and enforces equality between the values of the nodes to which it is incident, and circle nodes enforce that the sum of the currents on the incident edges should be 0. These constraints can only be satisfied if ${\Delta \theta}_{c_j}=\frac{1}{3}$, and if exactly one of the nodes involved in each clause is at 1 and the others at 0.[]{data-label="fig:construction_NP"}](construction_NP.eps)
(\[opt:card\_min\_con\]) is also NP-hard since (\[opt:security\_index\]) is a special case of (\[opt:card\_min\_con\]).
Tractable Special Cases of the Security Index Problem {#sec:si_full_measurement}
=====================================================
In Section \[subsec:full\_measurement\] we show that, under the full measurement assumption, the security index problem can be solved by solving its restriction where decision variables take binary values. Section \[subsec:equiv\_general\_thm\] presents the proof of the statement which implies our finding in Section \[subsec:full\_measurement\]. Section \[subsec:equiv\_general\_thm\] also discusses the relationship between the security index problem and its binary restriction defined in Section \[subsec:full\_measurement\]. Section \[subsec:gmc\_costly\_nodes\] describes the consequences of Sections \[subsec:full\_measurement\] and \[subsec:equiv\_general\_thm\], explaining how the security index problem can be reformulated as a generalized minimum cut problem with costly nodes, a graph problem whose efficient solution will be discussed in Section \[section-reformulation\].
The Security Index Problem Under Full Measurement Assumption {#subsec:full_measurement}
------------------------------------------------------------
Even though in general the security index problem in (\[opt:security\_index\]) is NP-hard for $H$ defined in (\[def:H\_matrix\]), there exist interesting specializations that are solvable in polynomial time. One such case is the *full measurement* situation where $P_1 = I$, $P_2 = I$ and $P_3 = I$. In [@SSJ_CDC2011_mincut; @GBMKP11] the full measurement assumption is also considered, motivated by the situations where all power flows and injections are measured in future smart grid applications. The polynomial time complexity of (\[opt:security\_index\]) under the full measurement assumption can be established in three steps. Firstly, it can be shown that problem (\[opt:security\_index\]) can be solved by solving a restriction where the decision vector $\Delta \theta$ is a binary vector. Secondly, in Section \[subsec:gmc\_costly\_nodes\] it will be shown that the binary restriction of (\[opt:security\_index\]) can be expressed in a generalized [[Min Cut]{} ]{}problem with costly nodes. Finally, this generalized [[Min Cut]{} ]{}problem can be shown to be solvable in polynomial time. This is to be explained in Section \[section-reformulation\].
The first step is formalized in the following statement, whose preliminary version appeared in [@SSJ_CDC2011_mincut].
\[thm:si\_sibin\_equiv\] Let $H$ in (\[def:H\_matrix\]) satisfy the full measurement assumption that $P_1 = I$, $P_2 = I$, and $P_3 = I$. Consider the following restriction of problem (\[opt:security\_index\]) with 0-1 binary decision vector: $$\label{opt:security_index_binary}
\begin{array}{cl}
\mathop{\rm minimize}\limits_{\Delta \theta \in {\left\{0,1\right\}}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
\textrm{subject to} & {A(:,e)}^T \Delta \theta \neq 0.
\end{array}$$ It holds that every optimal solution of (\[opt:security\_index\_binary\]) is an optimal solution of (\[opt:security\_index\]) (i.e., the problem with the same formulation except that $\Delta \theta$ is not restricted to binary values).
Proposition \[thm:si\_sibin\_equiv\] is a corollary of the more general Theorem \[thm:si\_sibin\_equiv\_general\] to be described in Section \[subsec:equiv\_general\_thm\].
Since there cannot be any all zero column in any incidence matrix $A$, problem ([\[opt:security\_index\]]{}) and (\[opt:security\_index\_binary\]) are always feasible. Proposition \[thm:si\_sibin\_equiv\] states that, under the full measurement assumption, an optimal solution of ([\[opt:security\_index\]]{}) can always be obtained by solving (\[opt:security\_index\_binary\]). The later problem will be shown to be solvable in polynomial time.
The Security Index Problem with Binary Decision Vector {#subsec:equiv_general_thm}
------------------------------------------------------
In the sequel, let $c_{ij} \ge 0$ represent the cost of attacking the power flow measurements of a line $(v_i,v_j)$, and $p_i \ge 0$ the cost of attacking the injection measurement at bus $v_i$. Problems (\[opt:security\_index\]) can be reformulated in a more general way that also allows taking into account the fact that tempering with certain measurements may be more expensive than with some others: $$\label{opt:gsi}
\begin{array}{cl}
\mathop{\rm minimize}\limits_{\Delta \theta \in {\mathbb{R}}^{n+1}} & c^T g(D A^T \Delta \theta) + p^T g(A D A^T \Delta \theta) \vspace{1mm} \\
\textrm{subject to} & {A(:,e)}^T \Delta \theta \neq 0,
\end{array}$$ and the corresponding reformulation of (\[opt:security\_index\_binary\]) is defined by $$\label{opt:gsi_bin}
\begin{array}{cl}
\mathop{\rm minimize}\limits_{\Delta \theta \in {\left\{0,1\right\}}^{n+1}} & c^T g(D A^T \Delta \theta) + p^T g(A D A^T \Delta \theta) \vspace{1mm} \\
\textrm{subject to} & {A(:,e)}^T \Delta \theta \neq 0.
\end{array}$$ In (\[opt:gsi\]) and (\[opt:gsi\_bin\]), $g$ is a vector-valued indicator function such that for any vector $x$, $g_i(x) = 1$ if $x_i \neq 0$ and $g_i(x) = 0$ otherwise. It can be seen that if $c_{ij} \in \{0,1,2\}$ and $p_i \in \{0,1\}$, then (\[opt:security\_index\]) and (\[opt:security\_index\_binary\]) are recovered. Let ${\begin{bmatrix} D A^T \Delta \theta \end{bmatrix}}_{(v_i,v_j)}$ denote the entry of $D A^T \Delta \theta$ corresponding to edge $(v_i,v_j)$, and let ${\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i}$ denote the entry of $A D A^T \Delta \theta$ corresponding to node $v_i$. With a slight abuse of notation, the symbol $g\big({\begin{bmatrix} D A^T \Delta \theta\end{bmatrix}}_{(v_i,v_j)} \big)$ denotes the entry of $g(D A^T \Delta \theta)$ corresponding to $(v_i,v_j)$. In addition, $g\left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i} \right)$ is defined similarly.
The following theorem characterizes the relationship between the security index problem in (\[opt:security\_index\]) and its binary restriction in (\[opt:security\_index\_binary\]) by studying their respective generalizations of (\[opt:gsi\]) and (\[opt:gsi\_bin\]) for arbitrary nonnegative vectors $c$ and $p$.
\[thm:si\_sibin\_equiv\_general\] Let $J_c$ and $J_b$, respectively, denote the optimal objective values of (\[opt:gsi\]) and (\[opt:gsi\_bin\]) with $A$ and $D$ defined in (\[def:H\_matrix\]), $c \in \mathbb{R}^m_+$, $p \in \mathbb{R}^{n+1}_+$ and $e \in \{1,2,\ldots,m\}$ given. Then $$\label{eqn:si_sibin_error_bound}
0 \le J_b - J_c \le \sum\limits_{v_i \in V^0} \max\Big\{0, \max\limits_{v_j \in N(v_i)} \big\{p_i - c_{ij} \big\} \Big\}.$$
First note that both (\[opt:gsi\]) and (\[opt:gsi\_bin\]) are always feasible with finite optimal objective values attained by some optimal solutions. In addition, $0 \le J_b - J_c$ holds because (\[opt:gsi\_bin\]) is a restriction of (\[opt:gsi\]). To show the upper bound in (\[eqn:si\_sibin\_error\_bound\]) the main idea is that for each feasible solution $\Delta \theta$ of (\[opt:gsi\]) it is possible to construct a feasible solution $\Delta \phi$ of (\[opt:gsi\_bin\]), such that the objective value difference is bounded from above by $\sum\limits_{v_i \in V^0} \max\Big\{0, \max\limits_{v_j \in N(v_i)} \big\{p_i - c_{ij} \big\} \Big\}$. The construction is as follows. Let $\Delta \theta$ be a feasible solution of (\[opt:gsi\]), and let $\Delta \theta(v_i)$ be its entry corresponding to node $v_i \in V^0$. Since $\Delta \theta$ is feasible, the constraint $A(:,e)^T \Delta \theta \neq 0$ implies that there exist two nodes denoted $v_s$ and $v_t$ with $e$ corresponding to either $(v_s,v_t)$ or $(v_t,v_s)$ such that $\Delta \theta(v_s) \neq \Delta \theta(v_t)$. Without loss of generality, it is assumed that $\Delta \theta(v_s) > \Delta \theta(v_t)$. Define $\Delta \phi \in {\{0,1\}}^{n+1}$ by $$\label{def:delta_tilde_theta}
\Delta \phi(v_i) = \begin{cases}
1 & \text{if} \; \Delta \theta(v_i) > \Delta \theta(v_t) \\
0 & \text{if} \; \Delta \theta(v_i) \le \Delta \theta(v_t)
\end{cases}
\quad \forall \; v_i \in V^0.$$ Note that $\Delta \phi$ is feasible to (\[opt:gsi\_bin\]) since $\Delta \phi(v_s) \neq \Delta \phi(v_t)$ by construction. Also notice that for any two nodes $v_i$ and $v_j$ if $\Delta \theta(v_i) = \Delta \theta(v_j)$ then $\Delta \phi(v_i) = \Delta \phi(v_j)$. Hence, in the objective functions of (\[opt:gsi\]) and (\[opt:gsi\_bin\]) it holds that $$\label{eqn:pf_edge_cost_noincrease}
c_{ij} g\big({\begin{bmatrix} D A^T \Delta \theta\end{bmatrix}}_{(v_i,v_j)} \big) \ge c_{ij} g\big({\begin{bmatrix} D A^T \Delta \phi\end{bmatrix}}_{(v_i,v_j)} \big), \quad \forall \; (v_i,v_j) \in E^0.$$ In other words, for each edge the contribution to the objective function with the new solution $\Delta \phi$ is smaller than or equal to that with the initial one $\Delta \theta$. To finish the proof, the objective function contribution due to the node injections needs to be investigated. Let $V_b \subset V^0$ be defined such that $$\label{eqn:pf_Vb_def}
v_i \in V_b \;\; \iff \;\; g\left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i} \right) = 0, \; \; g\left({\begin{bmatrix} A D A^T \Delta \phi \end{bmatrix}}_{v_i}\right) = 1.$$ In essence, $V_b$ encompasses all causes for $J_b > J_c$. Consider $v_i \in V_b$, since $g\left({\begin{bmatrix} A D A^T \Delta \phi \end{bmatrix}}_{v_i}\right) = 1$, there exists $v_k \in N(v_i)$ such that $\Delta \theta(v_k) \neq \Delta \theta(v_i)$. Consequently, the fact that $g \left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i}\right) = 0$ implies that there exists $v_i^+ = {\rm argmax}_{v_k \in N(v_i)} \{\Delta \theta(v_k)\}$ such that $\Delta \theta(v_i^+) > \Delta \theta(v_i)$. Similarly, there exists $v_i^- = {\rm argmin}_{v_k \in N(v_i)} \{\Delta \theta(v_k)\}$ such that $\Delta \theta(v_i^-) < \Delta \theta(v_i)$. If $\Delta \theta(v_i) > \Delta \theta(v_t)$, then (\[def:delta\_tilde\_theta\]) implies that $\Delta \phi(v_i^+) = \Delta \phi(v_i) = 1$. In addition, it holds that if $v_i^+ \in V_b$ then $v_i \neq {\rm argmax}_{v_k \in N(v_i^+)} \{\Delta \theta(v_k)\}$. This is true because ${\rm argmax}_{v_k \in N(v_i^+)} \{\Delta \theta(v_k)\} > \Delta \theta(v_i^+) > \Delta \theta(v_i)$ if $v_i^+ \in V_b$. Conversely, if $\Delta \theta(v_i) \le \Delta \theta(v_t)$, then $\Delta \phi(v_i^-) = \Delta \phi(v_i) = 0$. Similar to the case with $v_i^+$, if $v_i^- \in V_b$, then $v_i \neq {\rm argmin}_{v_k \in N(v_i^-)} \{\Delta \theta(v_k)\}$. In summary, for each $v_i \in V_b$, there exists an edge $e_i \in E^0$ in one of the following forms $(v_i, v_i^+)$, $(v_i^+, v_i)$, $(v_i, v_i^-)$ or $(v_i^-, v_i)$ such that
$$\begin{aligned}
\label{eqn:pf_ei1} & g\left({\begin{bmatrix} D A^T \Delta \theta\end{bmatrix}}_{e_i} \right) = 1, \;\;
g\left({\begin{bmatrix} D A^T \Delta \phi\end{bmatrix}}_{e_i} \right) = 0 \\
\label{eqn:pf_ei2} & e_i \neq e_j \quad \forall \; v_i \neq v_j \\
\label{eqn:pf_ei3} & \Big| {\big\{e_i \; \vline \; v_i \in V_b \big\}} \Big| = |V_b|
\end{aligned}$$
Using the above argument, the inequality in (\[eqn:si\_sibin\_error\_bound\]) can be deduced as follows: For all feasible solutions $\Delta \theta$ of (\[opt:gsi\]), it holds that $$\label{eqn:ubpf1}
\begin{array}{cl}
& J_b - c^T g(D A^T \Delta \theta) - p^T g(A D A^T \Delta \theta) \vspace{2mm} \\
\le & c^T g(D A^T \Delta \phi) + p^T g(A D A^T \Delta \phi) \vspace{2mm} \\
& - c^T g(D A^T \Delta \theta) - p^T g(A D A^T \Delta \theta)
\end{array}$$ because $\Delta \phi$ is a feasible solution of (\[opt:gsi\_bin\]) and $J_b$ is the optimal objective value of (\[opt:gsi\_bin\]). Because of (\[eqn:pf\_ei2\]), $\big\{e_i \; \vline \; v_i \in V_b \big\}$ does not contain duplicated edges. Therefore, the right-hand-side of (\[eqn:ubpf1\]) is equal to $$\label{eqn:ubpf2}
\begin{array}{l}
\sum\limits_{v_i \in V_b} p_i \left( g\left({\begin{bmatrix} A D A^T \Delta \phi \end{bmatrix}}_{v_i} \right) - g\left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i} \right) \right) \vspace{2mm} \\
+ \sum\limits_{\big\{e_i \; \vline \; v_i \in V_b \big\}} c_{e_i} \Bigg( g\left({\begin{bmatrix} D A^T \Delta \phi \end{bmatrix}}_{e_i} \right) - g\left({\begin{bmatrix} D A^T \Delta \theta \end{bmatrix}}_{e_i} \right) \Bigg) \vspace{2mm} \\
+ \sum\limits_{v_i \in {V^0 \setminus V_b}} p_i \left( g\left({\begin{bmatrix} A D A^T \Delta \phi \end{bmatrix}}_{v_i} \right) - g\left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i} \right) \right) \vspace{2mm} \\
+ \sum\limits_{E^0 \setminus \big\{e_i \; \vline \; v_i \in V_b \big\}} c_{e_i} \Bigg( g\left({\begin{bmatrix} D A^T \Delta \phi \end{bmatrix}}_{e_i} \right) - g\left({\begin{bmatrix} D A^T \Delta \theta \end{bmatrix}}_{e_i} \right) \Bigg).
\end{array}$$ In addition, because of (\[eqn:pf\_Vb\_def\]), (\[eqn:pf\_ei1\]) and (\[eqn:pf\_ei3\]) the expression in (\[eqn:ubpf2\]) is equal to $$\label{eqn:ubpf3}
\begin{array}{l}
\sum\limits_{v_i \in V_b} \big( p_i - c_{e_i} \big) \vspace{1mm} \\
+ \sum\limits_{v_i \in {V^0 \setminus V_b}} p_i \left( g\left({\begin{bmatrix} A D A^T \Delta \phi \end{bmatrix}}_{v_i} \right) - g\left({\begin{bmatrix} A D A^T \Delta \theta \end{bmatrix}}_{v_i} \right) \right) \vspace{2mm} \\
+ \sum\limits_{E^0 \setminus \big\{e_i \; \vline \; v_i \in V_b \big\}} c_{e_i} \Bigg( g\left({\begin{bmatrix} D A^T \Delta \phi \end{bmatrix}}_{e_i} \right) - g\left({\begin{bmatrix} D A^T \Delta \theta \end{bmatrix}}_{e_i} \right) \Bigg).
\end{array}$$ Because of (\[eqn:pf\_Vb\_def\]) and (\[eqn:pf\_edge\_cost\_noincrease\]), the last two sums in (\[eqn:ubpf3\]) are nonpositive. Therefore, it holds that $$\label{eqn:ubpf4}
\begin{array}{cl}
& J_b - c^T g(D A^T \Delta \theta) - p^T g(A D A^T \Delta \theta) \vspace{2mm} \\
\le & \sum\limits_{v_i \in V_b} \big( p_i - c_{e_i} \big) \vspace{1mm} \\
\le & \sum\limits_{v_i \in V_b} \max\limits_{v_j \in N(v_i)} \big\{p_i - c_{ij} \big\} \vspace{1mm} \\
\le & \sum\limits_{v_i \in V^0} \max\Big\{0, \max\limits_{v_j \in N(v_i)} \big\{p_i - c_{ij} \big\} \Big\}.
\end{array}$$ Finally, since (\[eqn:ubpf4\]) applies to all feasible solutions $\Delta \theta$ of (\[opt:gsi\]), the upper bound in (\[eqn:si\_sibin\_error\_bound\]) follows.
The full measurement assumption in Proposition \[thm:si\_sibin\_equiv\] corresponds to a special case in Theorem \[thm:si\_sibin\_equiv\_general\] where $c_{ij} = 2$ for all $(v_i,v_j) \in E^0$ and $p_i = 1$ for all $v_i \in V^0$. The inequalities in (\[eqn:si\_sibin\_error\_bound\]) imply Proposition \[thm:si\_sibin\_equiv\].
\[rem:exact\_conditions\] Theorem \[thm:si\_sibin\_equiv\_general\] suggests other situations where (\[opt:security\_index\]) and (\[opt:security\_index\_binary\]) are equivalent. One example is when there is a meter on each edge and there is at most one meter in each node. In this case, ${\begin{bmatrix} {P_1}^T {P_2}^T \end{bmatrix}}^T$ does not have a zero column and $P_3$ consists of subsets of rows of an identity matrix. This corresponds to $c_{ij} = 1$ and $p_i \le 1$ for all $i,j$ in Theorem \[thm:si\_sibin\_equiv\_general\]. Another situation suggesting equivalence is as follows: if an edge is not metered, then its two terminal nodes are not metered either. This corresponds to a case when $p_i \le \min\limits_{v_j \in N(v_i)} c_{ij}$ for all $v_i \in V^0$, implying that $\max\limits_{v_j \in N(v_i)} \big\{p_i - c_{ij} \big\} = 0$.
\[rem:approximation\] Without the full measurement assumption or conditions such as those described in Remark \[rem:exact\_conditions\], solving (\[opt:security\_index\_binary\]) can lead to an approximate solution to (\[opt:security\_index\]) with an error upper bound provided by (\[eqn:si\_sibin\_error\_bound\]). This error bound, however, is rather conservative since the summation is over all nodes $v_i \in V^0$. As developed in the proof, the summation is in fact over a subset $V_b$ of $V^0$. However, in general it is difficult to characterize the $V_b$ which leads to the tightest possible upper bound without first solving (\[opt:security\_index\]) to optimality.
Reformulating the Security Index Problem into Generalized Min Cut Problem with Costly Nodes {#subsec:gmc_costly_nodes}
-------------------------------------------------------------------------------------------
The above discussion suggests that the (exact or approximate) solution to the security index problem is obtained by solving (\[opt:gsi\_bin\]), whose graph interpretation will be the focus of this subsection. In (\[opt:gsi\_bin\]) the choice of 0 or 1 for each entry of $\Delta \theta$ is a partitioning of the nodes into two parts. The constraint ${A(:,e)}^T \Delta \theta \neq 0$ enforces that the two end nodes of edge $e$, denoted as $v_s$ and $v_t$, must be in two different parts of the partition. In the objective function, the term $c^T g(D A^T \Delta \theta)$ is the sum of the edge weights of the edges whose two ends are in different parts (i.e., edges that are “cut”, in an undirected sense). In addition, since $\Delta \theta$ has binary entries, a row of $A D A^T \Delta \theta$ is zero if and only if the corresponding node and all its neighbors are in the same part of the partition (i.e., none of the incident edges are cut). Therefore, the term $p^T g(A D A^T \Delta \theta)$ in the objective function is the sum of the node weights of the nodes connected to at least one cut edge. In summary, (\[opt:gsi\_bin\]) can be reinterpreted as a generalized minimum cut problem on an undirected graph (i.e., the original power network graph with the edge direction ignored). The generalization is due to the presence of the node weights.
We now define formally the [[Min Cut with costly nodes]{} ]{}problem (on any given directed graph) of which (\[opt:gsi\_bin\]) is a special case. Let $G(V,E)$ be a directed graph (we will see that the problem can be particularized to undirected graphs), where $V$ denotes the set of nodes $\{v_1,\dots,\kcs{v_{n+1}}\}$, and $E$ the set of directed edges; and suppose that a cost $c_{ij}\geq 0$ is associated to each directed edge $(v_i,v_j)$ and a cost $p_i\geq 0$ is associated to each node $v_i$. We designate two special nodes: a source node $v_s$ and a sink node $v_t$. The problem is the following:
\[prob-mincutmodif\] $$\label{opt:min_cost_node_partition}
\begin{array}{l}
\textrm{\bf {The {{\tt Min Cut with costly nodes} }problem.}} \vspace{1mm} \\
\textrm{Find a partition of $V$, denoted as $P = \{S_s,S_t\}$, such that} \vspace{1mm} \\
S_s,S_t \subset V, \quad S_s \cap S_t = \emptyset, \quad S_s \cup S_t = V, \quad s \in S_s, \quad t \in S_t \ \vspace{1mm} \\
\textrm{which minimizes the cost} \vspace{1mm} \\
\begin{array}{ccl}
C(P) & = & \sum\limits_{(v_i,v_j) \in E: v_i \in S_s, v_j \in S_t} c_{ij} \vspace{1mm} \\ & & + \sum\limits_{v_i \in S_s : \exists (v_i,v_j) \in E: v_j \in S_t} p_i + \sum\limits_{v_j \in S_t : \exists (v_i,v_j) \in E: v_i \in S_s} p_j.
\end{array}
\end{array}$$
By convention, if $v_i\in S_s, v_j\in S_t,$ for two nodes $v_i,v_j,$ we will say that both these nodes, and the edge $(v_i,v_j),$ are *in the cut*, or that this edge is cut.
Notice that in a directed graph an edge $(v_i,v_j)$ is cut if $v_i \in S_s$ and $v_j \in S_t$ but not in the reverse case, where $v_i \in S_t$ and $v_j \in S_s$, and the cost $c_{ij}$ is not incurred in that latter case. This asymmetry disappears however in symmetric graphs, in which to each edge $(v_i,v_j)$ with weight $c_{ij}$ corresponds a symmetric edge $(v_j,v_i)$ with same weight. For these symmetric graphs, the cost $c_{ij}$ is incurred as soon as $v_i$ and $v_j$ are not in the same set. Indeed, exactly one among $(v_i,v_j)$ and $(v_j,v_i)$ is in the cut in that case. The cost (\[opt:min\_cost\_node\_partition\]) consists then of the sum of the $c_{ij}$ over all pairs of nodes $v_i,v_j$ that are in different sets, and consists of the sum of the $p_i$ over all nodes that are adjacent to nodes in a different set. This is problem (\[opt:gsi\_bin\]). In addition, by letting $c_{ij}=c_{ji}=2$ for every edge and $p_i=1$ for every node, one recovers problem (\[opt:security\_index\_binary\]) under the full measurement assumption. We will show in Section \[section-reformulation\] how to solve Problem \[prob-mincutmodif\], and therefore the problems in (\[opt:gsi\_bin\]), (\[opt:security\_index\_binary\]) and (\[opt:security\_index\]).
An efficient solution to the tractable cases of the security index problem {#section-reformulation}
==========================================================================
This section presents the efficient solution to the [[Min Cut with costly nodes]{} ]{}problem (i.e., Problem \[prob-mincutmodif\]) introduced in Section \[subsec:gmc\_costly\_nodes\]. The proposed solution method also solves the security index problem under the full measurement assumption, since this problem is a special case of the [[Min Cut with costly nodes]{} ]{}problem.
Construction of an Auxiliary Graph {#subsec:aux_graph_construction}
----------------------------------
Consider a [directed]{} graph $G(V,E),\, V=\{v_1,\dots,v_{n+1}\}$ with a set of nonnegative weights $c_{ij} \ge 0$, and $p_i \ge 0$ for each node $v_i\in V$, a source node $v_s$ and a sink node $v_t$. We build an auxiliary graph $\tilde G$ using the following algorithm, illustrated in Fig. \[fig:repres\_graph\] on an example:
1. Define the set $\tilde V=\{\tilde v_i,\tilde w_i, \tilde z_i: 1 \leq i\leq n+1\}$ of nodes of the auxiliary graph.
2. Designate $\tilde v_s$ and $\tilde v_t$ as source and sink nodes respectively.
3. For all $1\leq i \leq n+1$, add the two directed edges $(\tilde w_i,\tilde v_i)$ and $(\tilde v_i,\tilde z_i),$ both with cost $p_i.$
4. For all $1\leq i,j \leq n+1:\, (v_i, v_j) \in E$
- add the edge $(\tilde v_i,\tilde v_j)$ with cost $c_{ij}.$
- add the two edges $(\tilde v_i,\tilde w_j)$ and $(\tilde z_i,\tilde v_j),$ both with a cost $C>\max_i p_i.$
The intuition behind the construction of $\tilde G$ is the following: Suppose that one wants to cut the edge $(\tilde v_i,\tilde v_j)$, then one must also cut at least either $(\tilde v_i, \tilde z_i)$ or $(\tilde z_i, \tilde v_j)$ (see Fig. \[fig:repres\_graph\]). Because the latter has a higher cost $C$, one will naturally cut $(\tilde v_i, \tilde z_i)$, incurring a cost $p_i$. Moreover, since that edge does not depend on $j$, one just needs to cut it (and pay the associated cost) once, independently of the number of other edges $(\tilde v_i,\tilde v_k)$ that will be cut. A similar reasoning applies to the path $(\tilde v_i, \tilde w_j)$ or $(\tilde w_j, \tilde v_j)$. Therefore, the cost of a minimum cut on $\tilde G$ will consists of the sum of all $c_{ij}$ for all edges $(\tilde v_i, \tilde v_j)$ in the cut, and of the sum of all $p_i$ for nodes incident to one or several edges $(\tilde v_i, \tilde v_j)$ or $(\tilde v_j, \tilde v_i)$ in the cut, i.e., to the cost of the equivalent cut on the initial graph $G$, taking the costly nodes into account.
Equivalence with Min Cut on the Auxiliary Graph
-----------------------------------------------
We now show formally that solving the standard [[Min Cut]{} ]{}problem on this weighted graph provides a solution to Problem \[prob-mincutmodif\] on the initial graph, and that a solution is obtained by directly translating the partition of the $\tilde v_i$ into the equivalent partition of the $v_i$.
\[thm:equiv\] Consider a graph $G(V,E)$ with a set of weights $c_{ij}\geq 0$ for each edge $(v_i,v_j)\in E$, and $p_i\geq 0$ for each node $v_i\in V$, a source node $v_s$ and a sink node $v_t$. Let $\tilde G(\tilde V,\tilde E)$ be the modified graph obtained from $G$ by the procedure described above, and the partition $\tilde V = \{\tilde S_s, \tilde S_t\}$ be an optimal solution of the standard [[Min Cut]{} ]{}problem for $\tilde G$. Then the partition $\{S_s,S_t\}$ of $V$, obtained by letting $v_i\in S_s$ if and only if $\tilde v_i \in \tilde S_s$, is an optimal solution to Problem \[prob-mincutmodif\] on $G$.
Let us call respectively $c^*$ and $\tilde c^*$ the optimal cost of Problem \[prob-mincutmodif\] on the graph $G$ and [[Min Cut]{} ]{}problem on the graph $\tilde G.$ In the sequel, we always assume that the source and sink nodes belong to the appropriate set of the partition.
We first prove that $\tilde c^* \leq c^*,$ by showing that for any cut in $G$ with cost $c$ (i.e., the sum of the costs of the edges [**and**]{} the nodes in the cut is $c$), one can build a cut in $\tilde G$ whose cost is equal to $c$ in the following way: For any $1\leq i\leq n+1,$
1. If $v_i \in S_s$, and all the out-neighbors of $v_i$ are in $S_s,$ put $\tilde v_i,$ $\tilde w_i$ and $\tilde z_i$ in $\tilde S_{s}.$
2. if $v_i \in S_t$, and all the in-neighbors of $v_i$ are in $S_t,$ put $\tilde v_i,$ $\tilde w_i$ and $\tilde z_i$ in $\tilde S_{ t}.$
3. if $v_i \in S_s$, and at least one out-neighbor of $v_i$ is in $S_t,$ put $\tilde v_i,$ $\tilde w_i$ in $\tilde S_{s}$ and $\tilde z_i$ in $\tilde S_t.$
4. if $v_i \in S_t$, and at least one in-neighbor of $v_i$ is in $S_s,$ put $\tilde v_i,$ $\tilde z_i$ in $\tilde S_{ t}$ and $\tilde w_i$ in $\tilde S_{s}.$
One can verify that no edge with cost $C$ is in the cut, and that an edge $(\tilde v_i,\tilde v_j)$ is in the cut if and only if the corresponding edge $(v_i,v_j)$ (which has the same weight) is in the initial cut. Moreover, for every node $i$, the edge $(\tilde w_i,\tilde v_i)$, of weight $p_i$, will be in the cut if and only if at least one edge arriving at $v_i$ was in the initial cut. Similarly, the edge $(\tilde v_i,\tilde z_i)$ will be in the cut if and only if at least one edge leaving $v_i$ is in the initial cut. So, there will be a contribution $p_i$ to the total cost if at least an edge arriving at $v_i$ is in the cut or at least one edge leaving $v_i$ is in the cut (note that the two situations cannot happen simultaneously.)\
As a conclusion, the cost of the cut $\{S_s,S_t\}$ in $G$ (counting the weights of the nodes) is equal to the cost of the cut $\{\tilde S_s, \tilde S_t\}$ in $\tilde G.$\
Consider now an arbitrary cut in $\tilde G$, and the corresponding cut in $G$ obtained by putting $v_i$ in $S_s$ if and only if $\tilde v_i \in \tilde S_{s},$ as explained in the statement of this theorem. We show that the cut of $G$ obtained has a cost (taking the vertex costs $p_i$ into account) smaller than or equal to the cost of the initial cut. This will imply that $\tilde c^* \geq c^*.$
The cost of this new cut $\{S_{s},S_t\}$ consists indeed of all the $c_{ij}$ of edges $(v_i,v_j)$ in the cut, and all the $p_i$ of the nodes at which arrives, or from which leaves an edge in the cut.
Consider first an edge $(v_i,v_j)$ in the cut, i.e., $v_i\in S_s,v_j\in S_t$. By construction, this implies that $\tilde v_i\in \tilde S_{s}$ and $\tilde v_j \in \tilde S_{t}$ so that the edge $(\tilde v_i, \tilde v_j)$ was also in the cut in $\tilde G$, incurring a same cost $c_{ij}$.
Consider now a node $v_i$ from which leaves at least one edge in the cut, incurring thus a cost $p_i$. (A symmetric reasoning applies if an edge in the cut arrives at $v_i$, and no node has edges in the cut both leaving from and arriving at it.) Call $v_j$ the node at which arrives that edge. We have thus $v_i\in S_s$ and $v_j\in S_t$, and therefore $\tilde v_i\in \tilde S_{s}$, $\tilde v_j \in \tilde S_{t}$ in $\tilde G$. This implies that one edge of the path consisting of $(\tilde v_i,\tilde z_i)$ and $(\tilde z_i,\tilde v_j)$ is in the cut. These edges have respective costs $p_i$ and $C>p_i$, so that a cost at least $p_i$ will be incurred by the cut in $\tilde G$. Note moreover that none of these edges will appear when considering other nodes and be counted more than once.
We have thus shown that to each cost in the cut $\{S_s,S_t\}$ for Problem \[prob-mincutmodif\] corresponds a larger or equal cost in $\{\tilde S_{s},\tilde S_{t}\}$ for the [[Min Cut]{} ]{}problem, and thus that the former has a smaller cost.
Therefore, if one takes any cut of optimal cost $\tilde c^*$ for the [[Min Cut]{} ]{}problem on $\tilde G$, and applies the procedure described in the theorem, one obtains a cut of $G$ with a smaller or equal cost for Problem \[prob-mincutmodif\]. Since we have proved that the optimal cost of the latter problem is at least $\tilde c^*$, this implies that $\tilde c^* = c^*$ and that the cost obtained is optimal for Problem \[prob-mincutmodif\] on $G$.
There exist many efficient polynomial time algorithms solving the [[Min Cut]{} ]{}problem exactly when the weights are nonnegative (e.g. [@Stoer:1997:SMA:263867.263872; @FF_MAX-FLOW]). Theorem \[thm:equiv\] implies that the same algorithms can be used to solve efficiently Problem \[prob-mincutmodif\], and therefore problem (\[opt:security\_index\_binary\]), and problem (\[opt:security\_index\]) in the fully measured case. Moreover, observe that the size of this new graph $\tilde G$ is proportional to that of $G$, as it has $3n$ nodes and $3|E| + 2n$ edges. The order of the polynomial measuring the efficiency of the algorithms remains therefore unchanged. In particular, if the standard [[Min Cut]{} ]{}problem on the new graph $\tilde G$ is solved using the algorithm in [@Stoer:1997:SMA:263867.263872] whose complexity is $O(n|E|+n^2\log(n))$, our algorithm has the same complexity.
Finally, consider a slight generalization of Problem \[prob-mincutmodif\] in which each node contains two different weights (one for cutting outgoing edges and the other for cutting incoming edges). Then with a corresponding modification in the auxiliary graph construction procedure in Section \[subsec:aux\_graph\_construction\] (in the fourth bullet), the proposed method can still solve the generalization in polynomial time.
The Original Security Index Problem Targeting Edge and Node {#sec:si_edge_only}
===========================================================
The relationship between the original security index problem in (\[opt:card\_min\_con\]), the problem in (\[opt:security\_index\]) and its binary restriction in (\[opt:security\_index\_binary\]) is summarized as follows: In the case where $H(k,:)$ in (\[opt:card\_min\_con\]) corresponds to the row of $P_1 D A^T$ and $-P_2 D A^T$, (\[opt:card\_min\_con\]) can be restated as (\[opt:security\_index\]) with an appropriate choice of $e$. Consequently, solving (\[opt:security\_index\_binary\]) either exactly solves (\[opt:card\_min\_con\]) or approximately solves (\[opt:card\_min\_con\]) with an error bound provided by (\[eqn:si\_sibin\_error\_bound\]), depending upon whether the full measurement assumption or similar ones in Remark \[rem:exact\_conditions\] are satisfied or not.
Next, consider the case where $H(k,:)$ corresponds to a row of $P_3 A D A^T$. The constraint $H(k,:) {\Delta \theta}= 1$ means that the power injection at the target node, denoted $v_s$, is nonzero. This implies that at least one edge incident to $v_s$ should have nonzero edge flow. Let $e_i$ with $i = 1,2,\ldots$ denote the column indices of $A$ of the incident edges of $v_s$. For any given $k$, consider the following instances (parameterized by $e_i$) $$\label{opt:card_min_con_extra_con}
\begin{array}{cccl}
J^i_{(\ref{opt:card_min_con_extra_con})} & \triangleq & \mathop{\textrm{min}}\limits_{\Delta \theta \in \mathbb{R}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
& & \textrm{subject to} & H(k,:) \Delta \theta \neq 0 \vspace{1mm} \\
& & & {A(:,e_i)}^T \Delta \theta \neq 0.
\end{array}$$ The minimum of $J^i_{(\ref{opt:card_min_con_extra_con})}$, over all $e_i$, is the optimal objective value of (\[opt:card\_min\_con\]). In addition, consider a relaxation of (\[opt:card\_min\_con\_extra\_con\]) as $$\label{opt:security_index_ei}
\begin{array}{cccl}
J^i_{(\ref{opt:security_index_ei})} & \triangleq & \mathop{\textrm{min}}\limits_{\Delta \theta \in \mathbb{R}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
& & \textrm{subject to} & {A(:,e_i)}^T \Delta \theta \neq 0,
\end{array}$$ and its binary restriction $$\label{opt:security_index_binary_ei}
\begin{array}{cccl}
J^i_{(\ref{opt:security_index_binary_ei})} & \triangleq & \mathop{\textrm{min}}\limits_{\Delta \theta \in {\{0,1\}}^{n+1}} & {\rm card}\big( H \Delta \theta \big) \vspace{1mm} \\
& & \textrm{subject to} & {A(:,e_i)}^T \Delta \theta \neq 0.
\end{array}$$ (\[opt:security\_index\_ei\]) is an instance of (\[opt:security\_index\]), and the fact that (\[opt:security\_index\_ei\]) has one fewer constraint than (\[opt:card\_min\_con\_extra\_con\]) implies that $$\label{eqn:si_relax_lb}
J^i_{(\ref{opt:security_index_ei})} \le J^i_{(\ref{opt:card_min_con_extra_con})}, \quad \forall \; i.$$ On the other hand, (\[opt:security\_index\_binary\_ei\]) is an instance of (\[opt:security\_index\_binary\]), and $$\label{eqn:si_bin_ub}
J^i_{(\ref{opt:security_index_binary_ei})} \ge J^i_{(\ref{opt:card_min_con_extra_con})}, \quad \forall \; i,$$ because if $\Delta \theta \in {\{0,1\}}^{n+1}$ is feasible to (\[opt:security\_index\_binary\_ei\]), then it is also feasible to (\[opt:card\_min\_con\_extra\_con\]). Notice, however, that a feasible solution of (\[opt:security\_index\_ei\]) need not be feasible to (\[opt:card\_min\_con\_extra\_con\]). Let $i^{\star}$ be defined such that $J^{i^{\star}}_{(\ref{opt:card_min_con_extra_con})} = \min\limits_i J^i_{(\ref{opt:card_min_con_extra_con})}$. The full measurement assumption or similar ones in Remark \[rem:exact\_conditions\] implies that $J^{i^{\star}}_{(\ref{opt:security_index_binary_ei})} = J^{i^{\star}}_{(\ref{opt:security_index_ei})}$. This, together with (\[eqn:si\_relax\_lb\]) and (\[eqn:si\_bin\_ub\]), suggests that $$J^{i^{\star}}_{(\ref{opt:card_min_con_extra_con})} \le J^{i^{\star}}_{(\ref{opt:security_index_binary_ei})} = J^{i^{\star}}_{(\ref{opt:security_index_ei})} \le J^{i^{\star}}_{(\ref{opt:card_min_con_extra_con})}.$$ This implies that the equalities above hold throughout, and solving (\[opt:security\_index\_binary\]) (by solving (\[opt:security\_index\_binary\_ei\])) indeed solves the original security index problem in (\[opt:card\_min\_con\]) (by solving (\[opt:card\_min\_con\_extra\_con\])). On the other hand, if the full measurement assumption does not hold, then $$J^{i^{\star}}_{(\ref{opt:card_min_con_extra_con})} \le J^{i^{\star}}_{(\ref{opt:security_index_binary_ei})} \le J^{i^{\star}}_{(\ref{opt:security_index_ei})} + \Delta J \le J^{i^{\star}}_{(\ref{opt:card_min_con_extra_con})} + \Delta J,$$ where the error upper bound $\Delta J$ can be obtained from (\[eqn:si\_sibin\_error\_bound\]). In conclusion, all exact or approximate results pertaining to the case between (\[opt:security\_index\]) and (\[opt:security\_index\_binary\]) apply to the case between (\[opt:card\_min\_con\]) and (\[opt:security\_index\_binary\]). As discussed in Remark \[rem:approximation\], the above error bound might be conservative. The approximation quality in practice will be demonstrated in Section \[section-example\] containing some numerical examples on benchmark power networks.
Numerical Examples {#section-example}
==================
Simple Illustrative Example of Problem \[prob-mincutmodif\] {#subsec:simple_example}
-----------------------------------------------------------
To illustrate that the proposed method is exact while previous methods (e.g., [@SSJ_CDC2011_mincut; @SSJ_ckt]) are not, consider an instance of Problem \[prob-mincutmodif\] depicted in Fig. \[fig:example\]. Only two partitions need to be considered: $S_s = \{v_s\}$ and $S_s = \{v_s,v_1,v_2\}$, with the respective objective values being 8 and 9 (indeed, the choice $S_s = \{v_s,v_2\}$ is strictly worse than $\{v_s,v_1,v_2\}$). As a comparison, the methods in [@SSJ_CDC2011_mincut; @SSJ_ckt] are also attempted. In particular, both [@SSJ_CDC2011_mincut; @SSJ_ckt] solve standard [[Min Cut]{} ]{}problems with edge weights only. In [@SSJ_CDC2011_mincut] the node weights are simply ignored, while in [@SSJ_ckt] the node weights are indirectly accounted for by adding them to the weights of the incident edges. Table \[tab:cut\_cost\] summarizes the objective values of the source sets $\{v_s\}$ and $\{v_s,v_1,v_2\}$ for the three graph setups. The italic numbers indicate the optimal objective values in the respective methods, suggesting that both [@SSJ_CDC2011_mincut; @SSJ_ckt] incorrectly choose $S_s = \{v_s,v_1,v_2\}$, which is suboptimal to Problem \[prob-mincutmodif\] in the current paper.
$S_s$ cost in our method cost in [@SSJ_CDC2011_mincut] cost in [@SSJ_ckt]
------------------- -------------------- ------------------------------- --------------------
$\{v_s\}$ [*8*]{} 2 10
$\{v_s,v_1,v_2\}$ 9 [*1*]{} [*9*]{}
: The objective values of source sets $S_s = \{v_s\}$ and $S_s = \{v_s,v_1,v_2\}$ in the graph setups of the current paper, [@SSJ_CDC2011_mincut] and [@SSJ_ckt]. As one can see, only our method finds the optimal cut.
\[tab:cut\_cost\]
Constructing the auxiliary graph as described in Section \[section-reformulation\] and solving the corresponding standard [[Min Cut]{} ]{}problem leads to the node partitioning in Fig. \[fig:example\_aug\]. In the auxiliary graph the optimal source set is $\{v_s,w_1,w_2\}$, with the objective value being 8. According to the rule in Theorem \[thm:equiv\], $\{v_s\}$ is the source set returned by the proposed procedure in this paper. It correctly solves Problem \[prob-mincutmodif\].
The Security Index Problem on Benchmark Systems {#subsec:benchmark}
-----------------------------------------------
To demonstrate the effectiveness and accuracy of the proposed solution, the security index problem for two benchmark systems is considered (IEEE 118-bus [@IEEE118bus] and Polish 2383-bus [@MATPOWER]). See Fig. \[fig:118bus6\] for an illustration of the 118-bus system.
First, the full measurement case is considered. The security index problem in (\[opt:card\_min\_con\]) is solved for each measurement, using the proposed solution and the methods from [@SSJ_CDC2011_mincut; @SSJ_ckt]. The proposed method is guaranteed to provide the exact optimal solutions, as explained earlier in the paper. Both for the 118-bus and 2383-bus cases, the methods from [@SSJ_CDC2011_mincut; @SSJ_ckt] are experimentally found to provide the exact solutions (though this is not guaranteed theoretically). The computation times for the three methods are listed in Table \[tab:solve-time\_full\_meas\], indicating that the methods have similar efficiency. The guarantee of optimality provided by our approach is obtained at no additional computational cost. The computation was performed on a PC with 2.4GHz CPU and 2GB of RAM. The minimum cut problems are solved using the MATLAB Boost Graph Library [@Gleich06contentsmatlab; @2002:BGL:504206].
our method [@SSJ_CDC2011_mincut] [@SSJ_ckt]
-------------- ------------ ----------------------- ------------
IEEE 118-bus 0.18s 0.23s 0.24s
2383-bus 29s 29s 33s
: computation times for all security indices in the full measurement case for the IEEE 118-bus and Polish 2383-bus benchmarks.
\[tab:solve-time\_full\_meas\]
Next, (\[opt:card\_min\_con\]) is considered when the full measurement assumption is removed. That is, the matrices $P_1$, $P_2$ and $P_3$ in (\[def:H\_matrix\]) need not be identities. In this test, the 118-bus system is considered. In the measurement system about 50% of power injections and power flows are measured. The measurements are chosen randomly, and the measurement system is verified to be observable (i.e., the corresponding $H_{2:}$ has full column rank ($=n$)). Since (\[opt:card\_min\_con\]) is NP-hard in general, no efficient solution algorithm has been known. Enumerative algorithms include, for instance, enumeration on the support of $H {\Delta \theta}$, finding the maximum feasible subsystem for an appropriately constructed system of infeasible inequalities [@Jokar:2008:EAS:1461600.1461607], and the big $M$ method to be described. The authors’ implementations of the first two methods turn out to be too inefficient for the applications concerned. Therefore, the big $M$ method is used, which sets up and solves the following optimization problem: $$\label{opt:si_MILP}
\begin{array}{cccl}
\mathop{\rm minimize}\limits_{\Delta \theta, \; y} & \quad \sum\limits_j y(j) & & \\
\textrm{subject to} & H \Delta \theta & \le & M y \\
& -H \Delta \theta & \le & M y \\
& H(k,:) \Delta \theta & = & 1 \\
& y(j) & \in & \{0,1\} \quad \forall \; j.
\end{array}$$ In (\[opt:si\_MILP\]), $M$ is a user-defined constant. If $M \ge {\|H \Delta \theta^\star\|}_\infty$ for at least one optimal solution $\Delta \theta^\star$ of (\[opt:card\_min\_con\]), then (\[opt:si\_MILP\]) provides the exact solution to (\[opt:card\_min\_con\]). Otherwise, solving (\[opt:si\_MILP\]) yields a suboptimal solution, optimal among all solutions $\Delta \theta$ such that ${\|H \Delta \theta\|}_\infty \le M$. In principle a sufficiently large $M$ can be found to ensure that the big $M$ method indeed provides the optimal solution to (\[opt:card\_min\_con\]) [@Schrijver:1986:TLI:17634]. However, this choice of $M$ is typically too large to be practical. In the numerical example in this section, $M$ is simply chosen to be $10^4$. (\[opt:si\_MILP\]) can be solved as a mixed integer linear program [@BT97] using solvers such as CPLEX [@CPLEX]. The solutions by the big $M$ method are treated as references for accuracy for the rest of the case study. Fig. \[fig:si118\] shows the (big $M$) security indices for all chosen measurements.
Alternatively, as described in Section \[sec:si\_edge\_only\] a suboptimal solution to (\[opt:card\_min\_con\]) can be obtained by solving (\[opt:security\_index\_binary\]) exactly using the proposed method in Section \[section-reformulation\] or the ones from [@SSJ_CDC2011_mincut; @SSJ_ckt]. As explained earlier, (\[opt:security\_index\_binary\]) can be formulated as Problem \[prob-mincutmodif\] with $p_i = 1$ if and only if the injection measurement at bus $v_i$ is taken and $c_{ij} = c_{ji} \in \{0,1,2\}$ being the total number of line power flow meters on the line connecting buses $v_i$ and $v_j$. Fig. \[fig:si118\_mc\], Fig. \[fig:si118\_mc1\] and Fig. \[fig:si118\_gmc\], respectively, show the security index test results with the three [[Min Cut]{} ]{}based methods (i.e., [@SSJ_CDC2011_mincut; @SSJ_ckt] and the one proposed in this paper).
These figures show only the big $M$ security indices (in light blue, the heights of the crosses) and the overestimation (in red, the heights between the crosses and the circles) for the measurements where the [[Min Cut]{} ]{}based methods do not agree with big $M$. The case study indicates that, among the three [[Min Cut]{} ]{}based methods, the proposed method provides the most accurate suboptimal solutions to (\[opt:card\_min\_con\]). In terms of computation time, the proposed method is most efficient as suggested by Table \[tab:solve-time\_partial\_meas\].
our method (s) [@SSJ_CDC2011_mincut] [@SSJ_ckt] big $M$
-------------- ---------------- ----------------------- ------------ ---------
IEEE 118-bus 0.17s 4.4s 0.21s 118s
: computation times for all security indices in the partial measurement case for the IEEE 118-bus benchmark.
\[tab:solve-time\_partial\_meas\]
Conclusions {#section-concl}
===========
It has been assumed that the security index problem, formulated as a cardinality minimization problem, cannot be solved efficiently. This paper formally confirms this conjecture by showing that the security index problem is indeed NP-hard. Nevertheless, the security index problem can be shown to be reducible to a [[Min Cut with costly nodes]{} ]{}problem (Problem \[prob-mincutmodif\]) under the full measurement assumption. In this paper, we show that this problem is equivalent to a standard [[Min Cut]{} ]{}problem on an auxiliary graph of proportional size, and can therefore be solved exactly and efficiently using standard techniques for the [[Min Cut]{} ]{}problem. Under the full measurement assumption, this allows computing the minimal number of measurements with which one must tamper in order to feed incorrect information on the SCADA system without being detected by a BDD method. The knowledge of this number can help strategically assigning protection resources (e.g., [@DS_SGC2010; @VSDS_jsac11]). Our method also solves a mathematically equivalent problem of robustness of the observability properties of the system with respect to the failure of some measurements, assuming again full measurement. It remains to be determined if the solution could be efficiently approximated in the general (not fully measured) case. Indeed, even though our approach already provides an approximate solution to such general problems we do not know if this approximation comes with any guarantee of accuracy.
Another interesting issue is the design question: in view of the exact solution of the security index problem presented in this paper, could one build efficient design methods in order to optimize the security index under some natural constraints?
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Jack H. Lutz[^1]'
- 'Brad Shutters[^2]'
bibliography:
- '../../bibliography.bib'
title: |
**[Approximate Self-Assembly of the\
Sierpinski Triangle]{}[^3]**
---
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Jim Lathrop, Xiaoyang Gu, Scott Summers, Dave Doty, Matt Patitz, and Brian Patterson for useful discussions.
[^1]: Department of Computer Science, Iowa State University, [[email protected]]([email protected])
[^2]: Department of Computer Science, Iowa State University, [[email protected]]([email protected])
[^3]: This research was supported in part by NSF grants 0652569 and 0728806.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Ciprian Chelba, Mohammad Norouzi, Samy Bengio\
Google\
`{ciprianchelba,mnorouzi,bengio}@google.com`
bibliography:
- 'paper.bib'
title: 'N-gram Language Modeling using Recurrent Neural Network Estimation'
---
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Oriol Vinyals and Rafał Józefowicz for support with the baseline implementation of LSTM LMs for UPenn Treebank in [@ptb_lm_tutorial] and One Billion Words Benchmark in [@rafal:one-billion-wds], respectively. We would also like to thank Maxim Krikun for thorough code reviews and useful discussions.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Real-time system technology traditionally developed for safety critical systems, has now been extended to support multimedia systems and virtual reality. A large number of real-time application, related to multimedia and adaptive control system, require more flexibility than classical real-time theory usually permits. This paper proposes an efficient adaptive scheduling framework in real-time systems based on period adjustment. Under this model periodic tasks can change their execution rates based on their importance values to keep the system underloaded. We propose Period\_Adjust algorithm, which consider the tasks whose periods are bounded as well as the tasks whose periods are not bounded.'
author:
-
title: 'Adaptive Scheduling in Real-Time Systems Through Period Adjustment'
---
Introduction
============
The real-time scheduling paradigms, both static such as rate monotonic scheduling \[13\], and dynamic such as earliest deadline first scheduling, do not fit well the requirements of advanced real-time applications in dynamic environments. Real-time systems are being increasingly designed for complex systems. For these applications, it is sometime impractical or impossible to provide static guarantees to real-time computation. These motivations have led to the emergence of the adaptation and overload management as a major research issue in real-time systems.\
An overview of prior art in overload management and adaptive scheduling techniques for real-time systems is given in Lu et al. \[14\]. Mechanism for detecting and handling timing errors including overloads are discussed in Stewart and Khosla \[20\], with emphasis on a specific application-oriented operating systems. An interesting technique for overload management in hard real-time control applications is described in Ramanathan et al. \[17\]. The author presents a scheduling policy deterministically guaranteeing $m$ out of any $k$ periodic task activations, along with a methodology able to minimize the effects of missed control-law updates. This work provides a solid foundation to graceful degradation policies of periodic real-time tasks. However, unless the overload duration is very short, the application could be significantly impaired by the loss of periodic execution for a number of real-time tasks. Dynamic Window Constrained Scheduling algorithm is similar except that the window $k$ is fixed. Mok et al. \[16\] modified Dynamic Window Constraint Scheduling, which is primarily deadline based by using the concept of Pfairness to improve the success rate for tasks with unit size execution time. Other frameworks such as the imprecise computation model and reward based model can be applied in the situation where quality of service is proportional to the amount of workload completed.\
The need for adaptive management of the Quality of Service has been widely recognized in the domain of the distributed multimedia systems. A graceful degradation of the communication subsystem is obtained in Abdelzaher and Shin \[1\] by means of QoS contracts specifying degraded acceptable QoS levels. Significant research has also been devoted to schedulers providing some degree of adaptation to cope with the dynamic overload environment. The need for scheduling systems providing real-time guarantee to a subset of tasks within a general operating system has been emphasized in the Stankovic et al. \[19\]. In Lu et al. \[14\] the authors assume a flexibility in timing requirements. To address the dynamics of the environment, they proposed a modified EDF adaptive scheduling framework based on feedback control methods and use feedback control loops to maintain a satisfactory deadline miss ratio when task execution times change.\
Many real-time task models have been proposed to extend timing requirements beyond the hard and soft deadlines based on the observation that jobs can be dropped without severely affecting performance \[4\]. Despite the success of some models in alleviating overload situation, it is sometime more suitable to execute jobs less often instead of dropping them or allocating fewer cycles. The work in Kuo et al. \[12\] is among the first to address this type of requirement. Load-adjustable algorithms and value-based policies are the main techniques proposed for graceful recovery from overload. A load adjustment mechanism is proposed in \[12\] in order to handle periodic processes with varying temporal parameters. The aim of this work is to determine feasible time parameter configurations (execution time $C$ and period $T$) and thus modify the real-time computation for collections of tasks. The configuration selection problem is solved by a harmonic approach achieving the maximum exploitation of the computational resources under any time parameter configuration. While appealing, this approach does not lend itself to many real-time systems, where execution times, in spite of their variability, cannot be set or chosen by the designer.\
In \[18\] Seto et al. considered the problem of finding a feasible set of task period as a non-linear programming problem, which seeks to optimize specific form of control performance measure. Cervin et al. used optimization theory to solve the period selection problem online by adaptively adjusting task periods with focus on optimizing specific control performance \[9\]. Baruah et al. \[2\] proposed a scheduling algorithm maximizing the effective processor during overload, given a minimum slack factor for all tasks.\
Buttazo et al. \[5\] proposed a flexible framework known as elastic task model, where deadline misses are avoided by increasing task periods until some desirable utilization is achieved. The work in \[14\] extends the basic elastic task model to handle cases where the computation time is unknown. In elastic task model \[6\],\[7\], periodic computations are modeled as springs with given elastic coefficients and minimum lengths. Requested variations in task execution rates or overload conditions are managed by changing the rates based on the spring’s elastic coefficients. Generalized elastic scheduling proposed by Chantem et al. \[10\],\[11\], by generalizing elastic scheduling approach. Although the Elastic model is nice but it does not consider the cases where the task periods of soft real-time systems may be unbounded or loosely bounded. We develop in this paper an efficient adaptive scheduling scheme in real-time systems through period adjustment, which consider the tasks having bounded as well as unbounded periods.\
This paper is organized as follows. Section 2 describes problem definition and motivation. Section 3 presents our proposed task model and the Period\_Adjust algorithm and its features. In section 4, we present the experimental results. Finally, section 5 contains conclusion.
Problem Definition and Motivation
=================================
Many models have been proposed in real-time scheduling theory to deal with adaptive scheduling and overload management. Some of the proposed models are based on the observation that less important jobs can be dropped without severely affecting performance. But dropping of jobs may not always be the best option, because it is sometime more suitable to execute the jobs less often instead of dropping them even if they are less important. Elastic task model \[6\] uses flexible framework but it do not consider the case where some of the soft real-time task may be loosely bounded or unbounded. We propose a novel scheduling framework based on period adjustment. Our algorithm considers the tasks whose periods are tightly bounded as well as the tasks whose periods are loosely bounded. We feel that this is more general model and this model performs nicely even when all tasks are bounded.\
Many soft real-time applications require the execution of periodic activities, whose rate can usually be defined within a certain range. The higher the frequency, the better the performance. Depending on the application domain, some tasks are rigidly imposed by the environment whereas other activities can be more flexible, producing significant results when their rates are within a certain range. For example, in multimedia systems the activities such as voice sampling, image acquisition, data compression, and video playing are performed periodically, but their execution rates are not so rigid. Depending on the requested quality of service, tasks may increase or decrease their execution rate to accommodate the requirements of other concurrent activities. However this period range may be flexible also. Suppose a soft real-time task has period range $(a,b)$, then in some application it may be possible to increase few time units above $b$ and decrease few time units below $a$, if by doing so system become schedulable. It is sometime counter intuitive that a soft real-time application which is schedulable in range $(a,b)$ can not be schedulable for the range $(a,(b+1))$ or alike. There are many flexible applications in multimedia and control applications in which we may be able to vary few time units across bound (upper or lower) without severely affecting the performance. We feel that there should be a general scheduling framework which can consider the flexible applications whose periods are unbounded alongwith the bounded one.
Proposed Work
=============
Task Model
----------
We consider the system where each task $\tau_i$ is periodic and is characterized by the following tuple: ($C_i$,$T_{i_0}$,$T_{i_{\min}}$,$T_{i_{\max}}$,$w_i$) for $i=1,\ldots N$. Where $N$ is the number of tasks in the system, $C_i$ is the worst case execution time and $T_{i_0}$ is the initial period of $\tau_i$. $T_{i_{\min}}$ denotes the minimum possible period of $\tau_i$ as specified by the application, and $T_{i_{\max}}$ represents the maximum period beyond which the system performance is no longer acceptable. The weighting factor $w_i$, represents importance of task $\tau_i$, to changing it’s period in face of changes. The longer the weighting factor of a task, the more will be it’s contribution towards the overall utilization. Given a task set $\Gamma$, tasks are arranged in a nondecreasing order of deadline.\
Each task $\tau_i$ in task set $\Gamma$ is divided in to two parts. $\Gamma_h$ for hard real-time tasks and $\Gamma_s$ for soft real-time tasks such that $\Gamma = \left\{ \Gamma_h \cup \Gamma_s \right\}$. $N$ is the total number of tasks in the systems. $n_h$ is the number of hard real-time tasks such that $(n_h=\left| \Gamma_h \right|)$. $n_s$ is the number of soft real-time tasks such that $(n_s=\left| \Gamma_s \right|)$. $w_i$ is the weighting factor or importance value of each soft real-time tasks in $\Gamma_s$. $w_i$’s for soft real-time tasks are arranged in such a way that $\sum_{i =1}^{n_s} w_i=1$, in other words $w_i$ represents fractional importance value or percentage of importance value of each soft real-time tasks towards the whole system performance. Furthermore a task in $\Gamma_s$ may belongs to $\Gamma_{s\_{bound}}$ or $\Gamma_{sp}$ or $\Gamma_{s\_{unbound}}$, i.e. ${\Gamma_s = \{\Gamma_{s\_{bound}} \cup \Gamma_{sp} \cup \Gamma_{s\_{unbound}}\}}$ where $\Gamma_{s\_{bound}}$ consists of those soft real-time tasks for which an upper bound or lower bound or both are imposed on tasks periods prior to execution or during execution, and $\Gamma_{sp}$ consists of those soft real-time tasks which have fixed periods or which requests for fixed periods during run time, whereas task set $\Gamma_{s\_{unbound}}$ consists of those tasks which are unbounded. However as a matter of fact period $T_i$ can not be less than worst case computation time $C_i$ of a task. Our scheduling algorithm emphasize such soft real-time application which have more number of tasks in $\Gamma_{s\_{unbound}}$.\
In this task model, all the tasks $\tau_i$, which does not belongs to $\Gamma_{s\_{bound}}$ can have $T_{i_{min}}$ or $T_{i_{max}}$ equal to $\Phi$, which means that they are unbounded. For each $(\tau_i \in \Gamma_h)$, $w_i=h_{rt}=1$ which means that all hard real-time tasks must execute provided they are schedulable. $T_i$ denotes the actual period of task $\tau_i$, which is constrained to be in the range \[$T_{i_{\min}}$, $T_{i_{\max}}$\] for the case $(\tau_i \in \Gamma_{s\_{bound}})$, whereas $C_i$ denotes actual execution time considered to be known a priori. In the case of tasks with variable computation times, $C_i$ will denote the actual worst case execution time. Any period variation is always subject to an utilization guarantee and is accepted only if there exists a feasible schedule such that tasks are scheduled by earliest deadline first algorithm. Hence if $\sum (C_i / T_{i_0}) \leq 1$, all tasks can be created at the minimum period $T_{i_0}$, otherwise the algorithm is used to adapt the task’s period to $T_i$ such that $\sum (C_i / T_{i}) = U_d \leq 1$, where $C_i$ is the actual online execution estimate and $U_d$ is some desired utilization factor. System designer can set $w_i$ statically or dynamically depending upon requirements. In static method, all soft real-time tasks are assigned $w_i$’s prior to start of the task execution and these $w_i$’s remains fixed up to the end of the task completions. In dynamic method, assignment of $w_i$ is event based i.e., weighting factor $w_i$ may be reassigned during the occurrence of any event such as, a new task arrival or completion of a task.
Period\_Adjust Algorithm
------------------------
We propose a new scheduling framework namely Period\_Adjust algorithm which accepts set of tasks $\Gamma$ and desired utilization $U_d$ and return set of periods for soft real-time tasks so as to maximize quality of service. We may set $U_d$ equal to the maximum schedulable utilization of individual scheduling algorithm. We can set $U_d = 1$ for dynamic scheduling algorithm like EDF, or we can set $U_d (n) = n (2^{1 / n} - 1)$ for the static scheduling RM algorithm, where $n$ is the number of independent, preemptable periodic tasks with relative deadline equal to their respective periods. In this algorithm we assume that deadline is equal to the period. We also assume that the execution time $C_i$ of all the tasks is given prior alongwith the periods of hard real-time tasks. The total task set $\Gamma$ is divided in two groups, namely the set of hard real-time tasks $\Gamma_h$, and the set of soft real-time tasks $\Gamma_s$. Furthur the set of soft real-time tasks may consists of $\Gamma_{sp}$, in which soft real-time task request for fixed period, $\Gamma_{s\_{bound}}$ in which tasks are bounded by maximum and minimum periods.\
Our Period\_Adjust algorithm works as follows: The first for loop computes the utilization of hard real-time tasks, then algorithm computes the summation of all utilization of task set $\Gamma_h$ to check for its feasibility. In the second for loop it computes the utilization of those tasks which request for period change, if there is no such task $U_{sp}$ is set to zero, after that it again checks for the feasibility of the schedulable utilization. The third for loop computes the tasks periods of all soft real time tasks in accordance with their weighting factor or importance value. Next the algorithm checks whether the periods of unbounded tasks are less than their computation time. If period is less than computation time, it replaces period by computation time. Finally it checks whether these periods exceeds their bounds for the bounded tasks, if this is the case it replaces periods with their bounds.
$U_h = \sum U_i$ $U_s = U_d - U_h$ $infeasible$
$U_i = \frac{C_i}{T_{i_{sp}}}_{}$
$U_{sp} = \sum U_i$ $U_s = U_d - U_h - U_{sp}$ $infeasible$ $T_i = \frac{C_i}{\Big(w_i + \frac{\sum w_{{sp}_i}}{N-n_h-n_{sp}}\Big)(U_d - U_h - U_{sp})}$ $T_i$ $T_i=C_i$ $mod=0$ $T_i = T_{i_{\min}}$ $T_i = T_{i_{\max}}$ $\Gamma_{s\_bound} = \Gamma_{s\_bound} - \tau_i$ $\Gamma_{sp} = \Gamma_{sp} + \tau_i$ $mod=1$\
Period\_Adjust$(\Gamma, U_d)$ $feasible$
If computed period $T_i$ for a bounded task is less than the minimum period $T_{i_{min}}$, we can simply replace $T_i$ by $T_{i_{min}}$, because increasing the period leads to less overall utilization. However, if the computed period $T_i$ is greater than the maximum period $T_{i_{max}}$, we can not simply replace $T_i$ by $T_{i_{max}}$, because decreasing the period leads to increased utilization, which may exceeds the schedulable utilization. Therefore corresponding task is removed from bounded task set $\Gamma_{s\_{bound}}$ to fixed period task set $\Gamma_{sp}$ and Period\_Adjust algorithm is re-invoked. In this algorithm we assume that in soft real-time application there are many cases where either no bounds are available or no bounds are required for soft real-time tasks.
Experimental Results
====================
In this section we present the experimental results performed on our task model. We consider period selection with deadlines equal to periods. In all the following tables here onwards periods $(T_{i_0}, T_{i_{\min}}, T_{i_{\max}})$ and computation times $(C_i)$ are expressed in milliseconds(ms).
-------------------- ------- ------------- --------------- --------------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $w_i$
\[0.5ex\] $\tau_1$ 18 100 50 150 0.30
$\tau_2$ 18 100 50 150 0.30
$\tau_3$ 18 100 50 150 0.18
$\tau_4$ 18 100 50 150 0.12
$\tau_5$ 18 100 50 150 0.10
\[1ex\]
-------------------- ------- ------------- --------------- --------------- -------
: Task Set Parameters[]{data-label="table 1"}
To execute the Period\_Adjust algorithm, we first use the task set parameters given in Table 1. In this experiment, all tasks starts at time 0 with an initial period of 100 time units and the task set is schedulable under EDF. Here the required maximum utilization of the overall system is $\frac{18}{50} + \frac{18}{50} + \frac{18}{50} + \frac{18}{50} + \frac{18}{50} = 1.8$, whereas the required minimum utilization of the overall system is $\frac{18}{150} + \frac{18}{150} + \frac{18}{150} + \frac{18}{150} + \frac{18}{150} = 0.6$. Since the current utilization is $\frac{18}{100} + \frac{18}{100} + \frac{18}{100} + \frac{18}{100} + \frac{18}{100} = 0.90$, the task set is schedulable under EDF. Assume that, at the 10sec, $\tau_{1_{}}$ needs to reduce its period to 50 time units, due to some changes in system dynamics not experienced by other tasks. Since the new required utilization of the system is $\frac{18}{50}_{} + \frac{18}{100} + \frac{18}{100} + \frac{18}{100} + \frac{18}{100} = 1.08$. which is greater than 1, and therefore as such it is not schedulable under EDF. We can observe that the required minimum utilization of the system is $\frac{18}{50} + \frac{18}{150} + \frac{18}{150} + \frac{18}{150} + \frac{18}{150} = 0.84$, which is less than 1. Therefore to allow for $\tau_1$ to change its period, the period of tasks $\tau_2$, $\tau_3$, $\tau_4$ and $\tau_5$ must increase for the system to remain schedulable. At time 20sec, $\tau_1$ goes back to its original period state. Fig. 1 shows the cumulative number of executed instances for each task as its period changes over time. When we execute Period\_Adjust algorithm on the above task sets, it will return the feasible set of task periods$(T_1 = 50, T_2 = 80, T_3 = 110, T_4 = 138, T_5 = 150)$.
Now we consider the same task set parameters with some change. Here we assume that soft real-time tasks $\tau_4$ and $\tau_5$ are not bounded, i.e. although the preferable maximum period is 150, some flexibility is provided by the application to increase or decrease the bound. In this case assume that at 10sec $\tau_{1_{}}$ needs to reduce the its period to 50 time units and $\tau_{2_{}}$ needs to reduce the its period to 60 time units, as shown in Table 2.
-------------------- ------- ------------- --------------- --------------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $w_i$
\[0.5ex\] $\tau_1$ 18 50 50 150 0.30
$\tau_2$ 18 60 50 150 0.30
$\tau_3$ 18 100 50 150 0.18
$\tau_4$ 18 100 $\Phi$ $\Phi$ 0.12
$\tau_5$ 18 100 $\Phi$ $\Phi$ 0.10
\[1ex\]
-------------------- ------- ------------- --------------- --------------- -------
: Task Set Parameters[]{data-label="table 2"}
For these task set parameters Task\_compress algorithm \[5\] is infeasible, whereas Period\_Adjust algorithm is feasible. In fact when we execute the Period\_Adjust algorithm on the above task sets, the corresponding periods obtained for the tasks are shown in Fig. 2 $(T_1 = 50, T_2 = 60, T_3 = 147, T_4 = 155, T_5 = 175)$.
Now, we consider the task set parameters given in Table 3 for the case of admission control policy during dynamic task activation.
-------------------- ------- ------------- --------------- --------------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $w_i$
\[0.5ex\] $\tau_1$ 30 100 50 350 0.20
$\tau_2$ 50 200 50 350 0.20
$\tau_3$ 70 300 50 350 0.20
$\tau_4$ 10 100 50 350 0.20
$\tau_5$ 10 70 50 350 0.20
\[1ex\]
-------------------- ------- ------------- --------------- --------------- -------
: Task Set Parameters[]{data-label="table 3"}
In this experiment $\tau_1$, $\tau_2$ and $\tau_3$ starts at time 0. They have the current utilization $\frac{30}{100} + \frac{50}{200} +
\frac{70}{300} = 0.78$ and therefore schedulable by EDF. At time 10sec two tasks $\tau_4$ and $\tau_5$ arrives which makes the total utilization $\frac{30}{100} + \frac{50}{200} + \frac{70}{300} + \frac{10}{100} + \frac{10}{70} = 1.03$. In order to allow the tasks $\tau_4$ and $\tau_5$ for execution, the tasks $\tau_1$, $\tau_2$ and $\tau_3$ can increase their period. Since both tasks $\tau_4$ and $\tau_5$ are of 10sec duration, after 20sec tasks $\tau_1$, $\tau_2$ and $\tau_3$ returns to their previous periods, as shown in the Fig. 3(Dynamic task activation). Now we consider the above task set parameters with some modification. In this case $\tau_4$ and $\tau_5$ arrives at 10sec having the computation times 30ms and 20ms respectively as shown in Table 4. Here task $\tau_3$ is loosely bounded (period of task $\tau_3$ should be preferably between 50 and 350 but not necessarily). In this case total utilization is $U = \frac{30}{100} + \frac{50}{200} + \frac{70}{300} + \frac{30}{100} + \frac{20}{70} = 1.37.$ Obviously task sets are not schedulable. Task set parameters alongwith importance values are given in the follwing table.
-------------------- ------- ------------- --------------- --------------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $w_i$
\[0.5ex\] $\tau_1$ 30 100 50 350 0.20
$\tau_2$ 50 200 50 350 0.20
$\tau_3$ 70 300 $\Phi$ $\Phi$ 0.20
$\tau_4$ 30 100 50 350 0.20
$\tau_5$ 20 70 50 350 0.20
\[1ex\]
-------------------- ------- ------------- --------------- --------------- -------
: Task Set Parameters[]{data-label="table 4"}
In this case also Task\_compress algorithm is infeasible. While Period\_Adjust algorithm is feasible. On execution periods returned by the Period\_Adjust algorithm are $(T_1 = 150, T_2 = 250, T_3 = 355, T_4 = 150, T_5 = 200)$.
{width="12cm" height="9cm"}
{width="12cm" height="9cm"}
{width="12cm" height="9cm"}
For the comparison purpose, here we use the task set parameters in \[7\], and we show that Period\_Adjust works nicely in these cases also.
-------------------- ------- ------------- --------------- --------------- ------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $E_i$ $w_i$
\[0.5ex\] $\tau_1$ 24 100 30 500 1 0.30
$\tau_2$ 24 100 30 500 1 0.30
$\tau_3$ 24 100 30 500 1.5 0.25
$\tau_4$ 24 100 30 500 2 0.15
\[1ex\]
-------------------- ------- ------------- --------------- --------------- ------- -------
: Task Set Parameters[]{data-label="table 5"}
Task set parameters are shown in Table 5. In this experiment four periodic tasks are created at time $t=0$. All the tasks start executing at their initial period, at $t = 10$sec $\tau_1$ decreases its period from 100ms to 33ms. At $t = 20$ms $\tau_1$ returns to its initial period. The result of the application of Period\_Adjust algorithm and Task\_compress algorithm on the above task sets is shown in the Fig. 4. It shows the actual number of instances executed by each task as a function of time. Next experiment consider the case of admission control policy during dynamic task activation (Table 6). Three tasks starts executing at the time $t = 0$ at their initial period. An other task $\tau_4$ arrives at time $t = 10$sec. Since tasks are not schedulable when $\tau_4$ is started, Period\_Adjust algorithm is invoked which increases the periods of other tasks to make the request of task $\tau_4$ fulfilled.
-------------------- ------- ------------- --------------- --------------- ------- -------
Task $C_i$ $T_{i_{0}}$ $T_{i_{min}}$ $T_{i_{max}}$ $E_i$ $w_i$
\[0.5ex\] $\tau_1$ 30 100 30 500 1 0.25
$\tau_2$ 60 200 30 500 1 0.25
$\tau_3$ 90 300 30 500 1 0.25
$\tau_4$ 24 50 30 500 1 0.25
\[1ex\]
-------------------- ------- ------------- --------------- --------------- ------- -------
: Task Set Parameters[]{data-label="table 6"}
Fig. 5 shows the actual number of instances executed by each task as a function of time during the execution of the Period\_Adjust algorithm and Task\_compress Algorithm.
{width="12cm" height="9cm"}
{width="12cm" height="9cm"}
Conclusions and Future Work
===========================
In this paper we have suggested Period\_Adjust algorithm for scheduling of tasks in which periods of soft real-time tasks are flexible. In this framework, periodic tasks can change their importance value to provide different quality of service. Importance value or weighting factor of soft real-time tasks are arranged in such a manner to keep the system underloaded. What makes Period\_Adjust more interesting is that it consider those soft real-time tasks whose periods are unbounded. The Period\_Adjust model is useful for supporting both multimedia systems and control applications in which the execution rates of some computational activities can not be properly predicted and they have to be dynamically tuned as a function of the current system state.\
We feel that Period\_Adjust model is a general model which can be applied in many applications. This framework can be extended to support the cases where deadline is less than period and computation time is variable.
[23]{}
Abdelzaher, T.F., Shin, K.G., “End-host architecture for QoS-adaptive communication,” In Proc. IEEE Real-Time Technology and Application Symposium, 1998.
Baruah, S.K., Haritsa J.R., “Scheduling for overload in Real-Time Systems,” IEEE Transactions on Computers, 1997.
Beccari, G., Caselli, S., Zanichelli, F., “A Technique for Adaptive Scheduling of Soft Real-Time Tasks,” Real-Time System Journal, vol.30, 2005.
Bernat, G., Burns, A., Llamosi, A., “Weakely hard real-time systems,” IEEE Transaction on Computers, 2001.
Buttazzo, G., Lipari, G., Abeni, L., “Elastic task model for adaptive rate control,” In Proc. IEEE Real-Time Systems Symposium, 1998.
Buttazzo, G., Abeni. L., “Adaptive workload management through elastic scheduling,” Real-Time System Journal, vol.23, 2002.
Buttazzo, G., Lipari, G., Abeni, L., “Elastic scheduling for flexible workload management,” IEEE Transactions on Computers, 2002.
Caccamo, M., Buttazzo, G., Sha, L., “Elastic Feedback Control,” In Proc. Euromicro Conference on Real-Time Systems, 2000.
Cervin, A., Eker, J., Bernhardsson, B., “Feedback-feedforward scheduling control tasks,” Real-Time System Journal, vol.23, 2002.
Chantem, T., Hu, X.S., Lemmon, M.D., “Generalized Elastic Scheduling,” In Proc. IEEE Real-Time Systems Symposium, 2006.
Chantem, T., Hu, X.S., Lemmon, M.D., “Generalized Elastic Scheduling for Real-Time Tasks,” 2007.
Kuo, T.W., Mok, A., “Load Adjustment in Adaptive real-time systems,” In Proc. IEEE Real-Time Systems Symposium, 1991.
Liu, C.L., Layland, J.W., “Scheduling Algorithm for multiprogramming in Hard Real-time environment,” Journal of ACM, vol.20, 1973.
Lu, C., Stankovic, J. A., Tao, G., Son, S. H., “Design and evaluation of a feedback control edf scheduling algorithm,” In Proc. IEEE Real-Time Systems Symposium, 1999.
Lu, C., Stankovic, J.A., Abdelzaher, T.F., Tao, G., Son, S.H., Marley, M., “Performance specifications and metrics for adaptive Real-time Systems,” In Proc. IEEE Real-Time Systems Symposium, 2000.
Mok, A., Wang, W., “Window constrained real time periodic scheduling,” In Proc. IEEE Real-Time Systems Symposium, 2001.
Ramanathan, P., “Overload management in real-time control application using (m,k) firm guarantee,” Transactions on Parallel and Distributed Systems, 1999.
Seto, D., Lehoczky, J., Sha, L., “Task period selection and schedulability in real-time systems,” In Proc. IEEE Real-Time Systems Symposium, 1998.
Stankovic, J. A., Lu, C., Son, S. H., Tao, G., “The case for feedback control real-time scheduling,” In Proc. Euromicro Conference on Real-Time Systems, 1999.
Stewart, D. B., Khosla, P. K., “Mechanisms for detecting and handling timing errors,” Communications of the ACM, vol.40, 1997.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Considerable effort has been devoted to deriving the Born rule (e.g. that $|\psi(x)|^2 dx$ is the probability of finding a system, described by $\psi$, between $x$ and $x + dx$) in quantum mechanics. Here we show that the Born rule is not solely quantum mechanical; rather, it arises naturally in the Hilbert space formulation of [*classical*]{} mechanics as well. These results provide new insights into the nature of the Born rule, and impact on its understanding in the framework of quantum mechanics.'
author:
- Paul Brumer and Jiangbin Gong
date: 'January 17, 2006'
title: The Born Rule in Quantum and Classical Mechanics
---
introduction
============
The Born rule [@born] postulates a connection between deterministic quantum mechanics in a Hilbert space formalism with probabilistic predictions of measurement outcomes. It is typically stated [@fine] as follows (without considering degeneracies): if an observable $\hat{O}$, with eigenstates $\{|O_{i}\rangle\}$ and spectrum $\{O_{i}\}$, is measured on a system described by the state vector $|\psi\rangle$, the probability for the measurement to yield the value $O_{i}$ is given by $|\langle
O_{i}|\psi\rangle|^{2}$. Alternatively, in the density matrix formulation used below, this rule states that the probability is $Tr\left[\rho_{\psi}
\rho_{O_i}\right]$, where $\rho_{\psi} = |\psi \rangle \langle\psi|$ and $\rho_{O_i} = |O_i \rangle \langle O_i|$. Most familiar is the textbook example that the probability of observing a system that is in a state $\psi$ in the coordinate range $x$ to $x + dx$ is given by $|\langle x |
\psi \rangle|^2 dx.$ Born’s rule appears as a fundamental postulate in quantum mechanics and is thus far in agreement with experiment. Hence, there is intense interest in providing an underlying motivation for, or derivation of, this rule.
Gleason’s theorem [@gleason], for example, provides a formal motivation of the Born rule, but it is a purely mathematical result about vectors in Hilbert spaces and does not provide insight into the physics of this postulate. For this reason there have been several attempts to provide a physical derivation of the Born rule. For example, Deutsch showed the possibility of deriving the Born rule from “the non-probabilistic axioms of quantum theory” and “the non-probabilistic part of classical decision theory” [@deutsch]. Deutsch’s approach was criticized by Barnum [*et al.*]{} [@barnum2] but was recently reinforced by Saunders [@saunders]. Hanson [@hanson] and Wallace [@wallace] analyzed the connection between possible derivations of the Born rule and Everett’s many worlds interpretation of quantum mechanics. Zurek recently proposed a significantly new approach, the so-called “envariance" approach [@zurek], for deriving the Born rule from within quantum mechanics. This approach, totally different from Deutsch’s method, was recently analyzed in detail by Schlosshauer and Fine [@fine] and by Zurek [@zurekpra]. Zurek’s “envariance" approach has also been analyzed and modified by Barnum [@barnum]. All these studies have attracted considerable interest in deriving Born’s rule by making some basic assumptions about quantum probabilities or expectation values of observables.
The Born rule is not expected to violate any future experiments. In this sense, even a strict derivation of the Born rule will not help predict any experimentally new physics. However, understanding the origin of the Born rule is important for isolating this postulate from other concepts in quantum mechanics and for understanding what is truly unique in quantum mechanics as compared with the classical physics. For example, based on the above-mentioned efforts to derive the Born rule [@deutsch; @saunders; @zurek; @fine; @zurekpra; @barnum], it seems now clear that the physical origin of Born’s rule is unrelated to the details (e.g., wavefunction collapse) of quantum measurement processes.
The main purpose of this work is to show that Born’s rule is not solely quantum mechanical and that it arises naturally in the Hilbert space formulation of [*classical*]{} mechanics. As such, the Born rule connecting probabilities with eigenvalues and eigenfunctions is not as “quantum" as it sounds. Indeed, quantum-classical correspondence, which played no role in Born’s original considerations [@born; @nobel], can then be arguably regarded as an interesting motivation for the Born rule in quantum mechanics. Similarly, exposing the Born rule in classical mechanics should stimulate new routes to understanding the physical origin of this rule in quantum mechanics. In particular, new and interesting questions can be asked in connection with the previous derivations of the Born rule.
To demonstrate that the Born rule exists in both quantum and classical mechanics we: (1) recall that both quantum and classical mechanics can be formulated in the Hilbert space of density operators [@koopman; @prigogine; @zwanzig; @fano; @wilkie; @jaffe], that the quantum and classical systems are represented by vectors $\rho$ and $\rho_c$, respectively, in that Hilbert space, and that $\rho$ and $\rho_c$ can be expanded in eigenstates of a set of commuting quantum and classical superoperators, respectively; and show (2) that the quantum mechanical Born rule can be expressed in terms of the expansion coefficient of a given density associated with eigendistributions of a set of superoperators in the Hilbert space of density matrix; and (3) that the classical interpretation of the phase space representation of $\rho_c$ as a probability density allows the extraction of Born’s rule in classical mechanics, and gives exactly the same structure as the quantum mechanical Born rule.
These results suggest that the quantum mechanical Born rule not only applies to cases of large quantum numbers, but also has a well-defined purely classical limit. Hence, independent of other subtle elements of the quantum theory, the inherent consistency with the classical Born rule for the macroscopic world imposes an important condition on any eigenvalue-eigenfunction based probability rule in quantum mechanics.
The Quantum Mechanical Born Rule in Density Matrix Formalism
============================================================
Consider first quantum mechanics in the Hilbert space of the density matrix [@fano; @wilkie]. Given an operator $\hat{O}\equiv \hat{K}_{N}$ for a system of $N$ degrees of freedom, we first consider the (classically) integrable case where there exist $N$ independent and commuting observables $\hat{K}_{i},\
i=1, \cdots, N$. Another extreme, the chaotic case, will be discussed in Sec. IV. For convenience we also assume that the $\hat{K}_{i}, \ i=1,2,\cdots, N$, have a discrete spectrum, but the central result below applies to cases with a continuous spectrum as well. The complete set of commuting superoperators in the quantum Hilbert space can be constructed as $$\begin{aligned}
\frac{1}{\hbar}[\hat{K}_{i},\ ],\ \ \frac{1}{2}[\hat{K}_{i},\ ]_{+}, \ \
(i=1,2, \cdots, N),\end{aligned}$$ where $[,]$ denotes the commutator and $[,]_{+}$ denotes the anticommutator, i.e., $[\hat{A},
\hat{B}]=\hat{A}\hat{B}-\hat{B}\hat{A}$, $[\hat{A},
\hat{B}]_{+}=\hat{A}\hat{B}+\hat{B}\hat{A}$. The simultaneous eigendensities of the complete set of superoperators are denoted $\rho_{{\bf \alpha},{\bf \beta}}$. That is, $$\begin{aligned}
\frac{1}{2}[\hat{K}_{i},\rho_{{\bf \alpha}, {\bf \beta}}]_{+} & = &
\alpha_{i}\rho_{{\bf \alpha}, {\bf \beta}}; \nonumber \\
\frac{1}{\hbar}\left[\hat{K}_{i},\rho_{{\bf \alpha}, {\bf \beta}}\right] & = & \beta_{i}\rho_{{\bf \alpha}, {\bf \beta}},\end{aligned}$$ where ${\alpha}\equiv (\alpha_{1}, \alpha_{2}, \cdots, \alpha_{N})$ is the collection of eigenvalues associated with $\frac{1}{2}[\hat{K}_{i},\ ]_{+}$, and ${\beta}\equiv (\beta_{1}, \beta_{2}, \cdots, \beta_{N})$ is the collection of eigenvalues associated with $\frac{1}{\hbar}[\hat{K}_{i},\ ]$.
The state of the quantum system is described by an arbitrary density matrix ${\rho}$ in the Hilbert space under consideration, and can be expanded in terms of the basis states $\rho_{{\bf \alpha}, {\bf \beta}}$ as $$\begin{aligned}
{\rho}=\sum_{{\bf \alpha},{\bf \beta}}D_{{\bf \alpha},{\bf \beta}}
\rho_{{\bf \alpha}, {\bf \beta}},
\label{qex}\end{aligned}$$ where the sum is over all eigendensities. The sum should be understood as an integral if the spectrum is continuous. Clearly, the expansion coefficients in Eq. (\[qex\]) are given by $$\begin{aligned}
D_{{\bf \alpha},{\bf \beta}}=Tr\left [{\rho}\rho_{{\bf \alpha}, {\bf
\beta}}^{\dagger}\right].
\label{qcoe}\end{aligned}$$ Equations (\[qex\]) and (\[qcoe\]), i.e. the expansion of $\rho$ in terms of the eigendensities $\rho_{{\bf \alpha},{\bf \beta}}$, is central to the analysis later below.
Consider now the quantum probability ${\cal P}_Q({\bf K}')$ of finding the quantum observables $\hat{K}_{i}$ with eigenvalues $K_{i}'$, $i=1,2, \cdots, N$, given that the system is in state $\rho$. We show here that the Born rule is then equivalent to the statement that ${\cal P}_{Q}({\bf K}')$ [*must be proportional to the expansion coefficient $D_{{\bf K}',{\bf 0}}$ of the given density $\rho$ associated with the common eigendistribution $\rho_{{\bf K}',{\bf 0}}$ of superoperators $\frac{1}{2}[\hat{K}_{i},\ ]_{+}$ with eigenvalues $K_{i}'$ and of superoperators $\frac{1}{\hbar}[\hat{K}_{i},\ ]$ with eigenvalue zero*]{}. To see this, consider first a quantum density for a pure quantum state, e.g., $\rho=|\psi\rangle\langle\psi|$ (the extension to mixed states is straightforward). Then the Born rule gives that $$\begin{aligned}
{\cal P}_Q({\bf K}')
& = & |\langle \psi|{\bf K}'\rangle|^{2}=
Tr \left[|{\bf K}'\rangle \langle {\bf K}'| |\psi\rangle\langle\psi| \right] \nonumber \\
& = & Tr \left[|{\bf K}'\rangle \langle {\bf K}'| \rho \right] ,
\label{qborn1}
\end{aligned}$$ where $|{\bf K}'\rangle$ is a common and normalized eigenfunction of operators $\hat{K}_{i},\ i=1,2,\cdots, N$. However, $$\begin{aligned}
\frac{1}{2}[\hat{K}_{i}, |{\bf K}'\rangle \langle {\bf K}'|]_{+} = K_{i}' |{\bf K}'\rangle \langle {\bf K}'|;
\ [\hat{K}_{i}, |{\bf K}'\rangle \langle {\bf K}'|] = 0,\end{aligned}$$ so that $|{\bf K}'\rangle \langle {\bf K}'|$ is seen to be the common eigendistribution of superoperators $\frac{1}{2}[\hat{K}_{i},\ ]_{+}$ with eigenvalues $K_{i}'$ and of superoperators $\frac{1}{\hbar}[\hat{K}_{i},\ ]$ with eigenvalue zero. That is, $$\begin{aligned}
|{\bf K}'\rangle \langle {\bf K}'|=\rho_{{\bf K}',{\bf 0}}.
\label{qborn2}\end{aligned}$$ Equations (\[qcoe\]), (\[qborn1\]), and (\[qborn2\]) then lead to $$\begin{aligned}
{\cal P}_{Q}({\bf K}')= D_{{\bf K}',{\bf 0}}.
\label{qeq}\end{aligned}$$ Equation (\[qeq\]) is a general restatement of the quantum mechanical Born rule based on the Hilbert space structure of density matrix. Note that the multidimensional result of Eq. (\[qborn1\]) has carefully accounted for possible degeneracies associated with $K_{N}'$. That is, the total probability of observing $K_{N}'$ would be obtained by summing ${\cal P}_{Q}({\bf K}')$ with all possible $K_{i}'$, $i=1,2, \cdots, (N-1)$, a necessary procedure not explicitly stated in Born’s rule.
The Born Rule in Classical Mechanics
====================================
Consider now classical mechanics. The mechanics has numerous equivalent formulations, such as Newton’s Laws, the Lagrangian or Hamiltonian formulations, Hamilton-Jacobi theory, etc. The less familiar Hilbert space formulation of classical mechanics used below was first established by Koopman [@koopman] and subsequently appreciated by, for example, Prigogine [@prigogine], Zwanzig [@zwanzig] and us [@wilkie; @jaffe] in some theoretical considerations. This being the case, the above eigenvalue-eigenfunction structure is not unique to quantum mechanics, a fact that may not be well appreciated and that is exploited below.
It is convenient to introduce the classical picture in the phase space representation, although abstract Hilbert space formulations may be used as well. Consider then the same case as above. Let the classical limit of the Wigner-Weyl representation of $\hat{K}_{i}$ be ${K}_{i}({\bf p},
{\bf q})$, where $({\bf p}, {\bf q})$ are momentum and coordinate space variables. The complete set of commuting superoperators on the classical Hilbert space is then [@wilkie] $$\begin{aligned}
i\{K_{j}({\bf p}, {\bf q}),\ \},\ K_{j}({\bf p}, {\bf q}), \ \ \ j=1,2, \cdots, N,\end{aligned}$$ where $K_{j}({\bf p}, {\bf q})$ are multiplicative operators and $\{\, ,
\}$ denotes the classical Poisson bracket. The simultaneous eigendensities of this complete set of classical operators, denoted $\rho^{c}_{{\bf
\alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q})$, satisfy: $$\begin{aligned}
K_{j}({\bf p}, {\bf q})\rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q})&=& \alpha_{j}^{c} \rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q}); \nonumber \\
i\{K_{j}({\bf p}, {\bf q}),\rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q})\}&= &\beta_{j}^{c} \rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q}),\end{aligned}$$ where the notation ${\bf \alpha}^{c}, {\bf \beta}^{c}$, $\alpha_{j}^{c}$, and $\beta_{j}^{c}$ is introduced in parallel with the quantum case. An arbitrary classical probability density $\rho_{c}({\bf p}, {\bf q})$ can be expanded as $$\begin{aligned}
\rho_{c}({\bf p}, {\bf q})=\sum_{{\bf \alpha}^{c}, {\bf \beta}^{c}} D^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}
\rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q}),
\label{classicalsum}\end{aligned}$$ where $$\begin{aligned}
D^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}} &=& \int d{\bf p} d{\bf q}\
\rho_{c}({\bf p}, {\bf q}) \left[ \rho^{c *}_{{\bf \alpha}^{c}, {\bf
\beta}^{c }}({\bf p}, {\bf q})\right] \nonumber \\
&\equiv& Tr\left[\rho_c \rho^{c \dagger}_{{\bf
\alpha}^{c}, {\bf \beta}^{c}}\right],\end{aligned}$$ and where the sum in Eq. (\[classicalsum\]) is over all eigendensities.
Consider now, within this formalism, the probability of finding ${\bf K}$ \[with ${\bf K} \equiv (K_{1},K_{2}, \cdots, K_{N})$\] between ${\bf K'}$ and ${\bf K'} + d{\bf K'}$. To proceed we make a canonical transformation between representations $({\bf p}, {\bf q})$ and $({\bf K}, {\bf Q})$, where ${\bf K}$ are the new momentum variables, and ${\bf Q}\equiv(Q_{1},Q_{2},
\cdots, Q_{N})$ denotes the new position variables conjugate to ${\bf K}$. The ${\bf Q}$ can be obtained by regarding ${\bf p}$ as a function of ${\bf q}$ and ${\bf K}$, defining the generating function $S({\bf q}, {\bf K})=\int_{{\bf q}_{0}}^{{\bf q}} {\bf p}({\bf q}', {\bf K})
\cdot d{\bf q}'$, and then noting that $Q_{i}=\partial S/\partial
K_{i}$ [@rasband]. In this representation the classical eigendensities $\rho^{c}_{\alpha^{c}, {\bf \beta}^{c}} $ take a rather simple form, $$\begin{aligned}
\rho^{c}_{{\bf \alpha}^{c}, {\bf \beta}^{c}}({\bf p}, {\bf q})& =>&
{\cal R}_{{\bf K}',{\bf \Lambda}}({\bf K}, {\bf Q}) \nonumber \\
&=&
\frac{1}{{(2\pi)}^{N/2}}\delta({\bf K}'-{\bf K})
\exp(i{\bf \Lambda}\cdot{\bf Q})
\label{eigendist}\end{aligned}$$ with eigenvalues [@footnote] $\alpha^{c}_{i} = K_{i}$ and $\beta^{c}_{i} = \Lambda_{i}$, for $i=1,2,\cdots, N$. The set of eigendistributions ${\cal R}_{{\bf K}',{\bf \Lambda}}$ are complete and orthogonal, i.e., $$\begin{aligned}
& & \int d{\bf K}d{\bf \Lambda}\ {\cal R}^{*}_{{\bf K},{\bf \Lambda}}({\bf K}', {\bf Q}')
{\cal R}_{{\bf K},{\bf \Lambda}}({\bf K}'', {\bf Q}'') \nonumber \\
&&=
\delta({\bf K}'-{\bf K}'') \delta({\bf Q}'-{\bf Q }''); \nonumber \\
&& \int d{\bf K} d{\bf Q}\ {\cal R}^{*}_{{\bf K}',{\bf \Lambda}'}({\bf K}, {\bf Q})
{\cal R}_{{\bf K}'',{\bf \Lambda}''}({\bf K}, {\bf Q})\nonumber \\
&&= \delta({\bf \Lambda}'-{\bf \Lambda}'')\delta({\bf K}'-{\bf K }'').\end{aligned}$$ The ${\cal R}_{{\bf K}',{\bf \Lambda}}$ are “improper states", insofar as they contain delta functions. However, this is consistent with the fact that they are eigendistributions of classical superoperators that have continuous spectra. If desired, a Rigged Hilbert space [@bohm] can be used to include these states more formally.
Given a classical probability density $\rho_{c} ({\bf p}, {\bf q})$ that describes the state of the system, we can convert to the ${\bf K}, {\bf Q}$ representation to obtain $\rho_{c}'
({\bf K}, {\bf Q})\equiv \rho_{c}\left[{\bf p}({\bf K}, {\bf Q}),
{\bf q}({\bf K}, {\bf Q})\right]$. The probability ${\cal P}_{c}({\bf K}')$ of finding the observables between ${\bf K}'$ and ${\bf K}'$+d${\bf K}$ is evidently given by $$\begin{aligned}
{\cal P}_{c}({\bf K}')=
\left[\int d{\bf K} d{\bf Q}\ \delta({\bf K}'- {\bf K})
\rho_{c}'({\bf K},{\bf Q})\right] d{\bf K}.
\label{obvious}\end{aligned}$$ This result has an enlightening interpretation in the Hilbert space formulation of classical mechanics, as can be seen by independently obtaining Eq. (\[obvious\]) using this approach. To do so, we first expand the given density in terms of the basis states ${\cal R}_{{\bf K}',{\bf \Lambda}}$. That is, $$\begin{aligned}
\rho_{c}'({\bf K}, {\bf Q})
= \int d{\bf K}' d{\bf \Lambda}
D^{c}_{{\bf K}',{\bf \Lambda}}{\cal R}_{{\bf K}',{\bf \Lambda}}({\bf K}, {\bf Q}),
\label{expansion}\end{aligned}$$ where the expansion coefficients are given by the overlap integrals $$\begin{aligned}
D^{c}_{{\bf K'},{\bf \Lambda}}&=&
\int d{\bf K}d{\bf Q}\ \rho_{c}'({\bf K}, {\bf Q})
{\cal R}^{*}_{{\bf K}',{\bf \Lambda}}({\bf K}, {\bf Q}).
\label{overlapeq}\end{aligned}$$ Because classical probabilities in ${\bf
K}$, that do not refer to ${\bf Q}$, are obtained by integrating over ${\bf Q}$, we obtain $$\begin{aligned}
{\cal P}_{c}({\bf K}')= \left[\int d{\bf Q}\ \rho_{c}'({\bf K}', {\bf Q})\right] d{\bf K}.
\label{peq1}\end{aligned}$$ Substituting Eq. (\[expansion\]) into Eq. (\[peq1\]) and using Eq. (\[eigendist\]) then yields $$\begin{aligned}
{\cal P}_{c}({\bf K}')
&=& (2\pi)^{N/2} D^{c}_{{\bf K}',{\bf 0}} d{\bf K}
\label{claborn}\end{aligned}$$ This result is equivalent to Eq. (\[obvious\]), but contains an important message from the perspective of mechanics in Hilbert space. That is, Eq. (\[claborn\]) indicates that, given a classical density $\rho'_{c}$ that describes the state, the probability of finding the observable $K_{j}$ ($j=1,2,\cdots, N$) in a regime $({\bf K}', {\bf K}' +d{\bf K})$, is [*proportional to the overlap between the given density $\rho'_{c}$ and the common eigendistribution ${\cal R}_{{\bf
K}',{\bf 0}}$ of multiplicative operators $K_{j}$ with eigenvalues $K_{j}'$ and of operators $i\{ K_{j}, \}$ with eigenvalue zero*]{}. This overlap is $D^{c}_{{\bf K}',{\bf 0}}$, the expansion coefficient of $\rho'_{c}$ in terms of ${\cal R}_{{\bf K}',{\bf 0}}$. Significantly, this connection between classical probabilities and the overlap between the given classical state density and a particular set of classical eigendistributions \[Eq. (\[claborn\])\] is the direct analog of the quantum mechanical Born rule in the density matrix formalism \[Eq. (\[qeq\])\]. That is, Eq. (\[claborn\]) is the Born rule in classical mechanics.
Quantum versus Classical Born rule in Chaotic Cases
===================================================
Classical-quantum correspondence of Hilbert space structures in chaotic cases is more subtle and complicated than in integrable cases [@wilkie]. Nevertheless, results above indicate that only a particular set of eigendensities are relevant in understanding the quantum versus classical Born rule. Indeed, it is straightforward to show that our previous considerations also apply to cases where there do not exist $(N-1)$ observables that commute with $K_{N}({\bf p}, {\bf q})$. For example, consider a chaotic spectrum case where $K_{N}({\bf p}, {\bf q})$ does not commute with any other smooth phase space functions $Z({\bf p}, {\bf q})$, i.e., $\{K_{N}({\bf p}, {\bf q}), Z({\bf p}, {\bf q})\}\ne 0 $ always holds. Following Ref. [@wilkie] we consider a function $\tau({\bf p}, {\bf q})$ that satisfies $\{\tau({\bf p}, {\bf q}), K_{N}({\bf p},
{\bf q})\}=1 $. Then the classical eigenfunction of the multiplicative operator $K_{N}({\bf p}, {\bf q})$ and of the operator $i\{ K_{N}({\bf p},
{\bf q}), \ \}$ with eigenvalue $\lambda$ is given by $\xi \delta
[K_{N}'-K_{N}({\bf p}, {\bf q})] \exp [i\lambda \tau({\bf p}, {\bf q})]$, where $\xi$ is a normalization constant. In particular, the eigenfunction with zero eigenvalue for the operator $i\{ K_{N}({\bf p}, {\bf q}), \ \}$ is $\xi\delta [K_{N}'-K_{N}({\bf p}, {\bf q})]$. This eigenfunction defines a $(2N-1)$-dimensional hypersurface in phase space (whereas the eigenfunction ${\cal R}_{{\bf K}',{\bf 0}}$ in the integral case defines an $N$-dimensional manifold). The overlap between this particular eigenfunction $\xi\delta [K_{N}-K_{N}({\bf p}, {\bf q})]$ and a given phase space probability density $\rho_{c}({\bf p}, {\bf q})$, i.e., $$\begin{aligned}
P_{c}(K_{N}')\equiv \xi \int d{\bf p} d{\bf q}\
\rho_{c}({\bf p}, {\bf q})\delta [K_{N}'-K_{N}({\bf p}, {\bf q})],
\label{chaotic}\end{aligned}$$ yields the probability $P_{c}(K_{N}')dK_{N}$ of finding $K_{N}({\bf p},
{\bf q})$ lying in the regime $[K_{N}', K_{N}'+dK_{N}]$. This is again in complete analogy to how the quantum probability ${\cal P}_Q(K_{N}')$ of finding the eigenvalue $K_{N}'$ is determined, i.e., ${\cal P}_Q(K_{N}')$ is given by the overlap between a given quantum density $|\psi\rangle\langle\psi|$ and the eigenfunction of superoperator $\frac{1}{2}[\hat{K}_{N},]_{+}$ with eigenvalue $K_{N}'$ and of superoperator $\frac{1}{\hbar}[\hat{K}_{N},\ ]$ with eigenvalue zero. Such a quantum eigenfunction is simply given by $|K_{N}'\rangle\langle K_{N}'|$, where $|K_{N}'\rangle$ is the eigenfunction of $\hat{K}_{N}$ with eigenvalue $K_{N}'$. These analyses make it clear that even for chaotic cases, the Born rule formulated in terms of eigen-densities in the associated (classical or quantum) Hilbert space exists in both quantum and classical mechanics.
Concluding Remarks
==================
In summary, with the quantum mechanical Born rule formulated in the Hilbert space of density matrix, we have demonstrated an analogous Born rule in classical mechanics. In so doing we never assumed that the quantum density goes smoothly, in the classical limit, to a classical density that already has a clear probabilistic interpretation. Rather, we have simply assumed that a system in either quantum or classical mechanics is described by a density operator in Hilbert space, and that these density operators serve the same descriptive purpose in both mechanics. This, plus the decomposition of the operator in terms of eigendistributions of a set of commuting superoperators, suffice to show that the Born rule applies in both quantum and classical mechanics. Hence, the quantum mechanical Born rule appears to be very natural in the light of quantum-classical correspondence in how Hilbert space structures embody measured probabilities. The recognition that Born’s rule is not really a unique quantum element should complement, as well as impact upon, previous attempts to derive the Born rule [@deutsch; @saunders; @zurek; @fine; @zurekpra; @barnum]. Further, it motivates numerous questions, such as, can the quantum mechanical Born rule be derived with fewer assumptions by taking advantage of the classical Born rule as a limit? How can one reconcile derivations involving purely quantum language with the existence of a classical Born rule?, etc. These, and related issues, are the subject of future work.
[**Acknowledgments:**]{} This work was supported by a grant from the National Science and Engineering Research Council of Canada. We thank Profs. R. Kapral and S. Whittington, University of Toronto, for comments on an earlier version of this manuscript.
[100]{} M. Born, Z. Phys. [**37**]{}, 863 (1926). M. Schlosshauer and A. Fine, Found. Phys. [**35**]{}, 197 (2005). A.M. Gleason, J. Math. Mech. [**6**]{}, 885, (1957). D. Deutsch, Proc. Roy. Soc. London [**A 455**]{}, 3129 (1999). H. Barnum, C.M. Caves, J. Finkelstein, C.A. Fuchs, and R. Schack, Proc. Roy. Soc. London [**A 456**]{}, 1175 (2000). S. Saunders, Proc. Roy. Soc. London [**A 460**]{}, 1771 (2004). R. Hanson, Found. Phys. [**33**]{}, 1129 (2003). D. Wallace, Stud. Hist. Philos. Mod. Phys. [**34B**]{}, 415 (2003). W.H. Zurek, , 120404 (2003). W.H. Zurek, , 052105 (2005). H. Barnum, e-print quant-ph/0312150; Phys. Rev. A (to be published). M. Born, “The Statistical Interpretation of Quantum Mechanics”, Nobel Lecture, Dec. 11, 1954. B.O. Koopman, Proc. Natl. Acad. Sci. U.S.A. [**17**]{}, 315 (1931). I. Progogine, [*Non-equilibrium Statistical Mechanics*]{}, Wiley, New York (1962). R.S. Zwanzig, [*Lectures in Theoretical Physics*]{}, Vol.3, Wiley, New York, (1961), pp 135. U. Fano, Rev. Mod. Phys. [**29**]{}, 74 (1959). J. Wilkie and P. Brumer, , 27 (1997); , 43 (1997). C. Jaffé and P. Brumer, J. Phys. Chem. [**88**]{}, 4829 (1984); J. Chem. Phys. [**82**]{}, 2330 (1985). (2000). S. N. Rasband, [*Dynamics*]{} (John Wiley & Sons, New York, 1983). The eigenvalues ${\bf \Lambda}$ can be either continuous or discrete, depending on the system. For example, the spectrum comprises integers when, for an integrable system, the ${\bf K}, {\bf Q}$ are action-angle variables. A. Bohm, [*Quantum Mechanics: Foundations and Applications*]{} (Springer, New York, 1993).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $I=(x^{v_1},\ldots,x^{v_q})$ be a square-free monomial ideal of a polynomial ring $K[x_1,\ldots,x_n]$ over an arbitrary field $K$ and let $A$ be the incidence matrix with column vectors ${v_1},\ldots,{v_q}$. We will establish some connections between algebraic properties of certain graded algebras associated to $I$ and combinatorial optimization properties of certain polyhedra and clutters associated to $A$ and $I$ respectively. Some applications to Rees algebras and combinatorial optimization are presented. We study a conjecture of Conforti and Cornuéjols using an algebraic approach.'
---
=-1cm =-1.5cm
[**Blowup algebras of square-free monomial ideals and some links\
to combinatorial optimization problems**]{}\
Isidoro Gitler, Enrique Reyes[^1],\
and\
Rafael H. Villarreal[^2]\
[Departamento de Matemáticas]{}\
[Centro de Investigación y de Estudios Avanzados del IPN]{}\
[Apartado Postal 14–740]{}\
[07000 México City, D.F.]{}\
[e-mail: [[email protected]]{}]{}
Introduction {#Int}
============
Let $R=K[x_1,\ldots,x_n]$ be a polynomial ring over a field $K$ and let $I$ be an ideal of $R$ of height $g\geq 2$ minimally generated by a finite set of square-free monomials $F=\{x^{v_1},\ldots,x^{v_q}\}$ of degree at least two. As usual we use $x^a$ as an abbreviation for $x_1^{a_1} \cdots x_n^{a_n}$, where $a=(a_1,\ldots,a_n)\in \mathbb{N}^n$. A [*clutter*]{} with vertex set $X$ is a family of subsets of $X$, called edges, none of which is included in another. We associate to the ideal $I$ a [*clutter*]{} $\cal C$ by taking the set of indeterminates $X=\{x_1,\ldots,x_n\}$ as vertex set and $E=\{S_1,\ldots,S_q\}$ as edge set, where $$S_k={\rm supp}(x^{v_k})=\{x_i\vert\, \langle e_i,v_k\rangle=1\}.$$ Here $\langle\ ,\, \rangle$ denotes the standard inner product and $e_i$ is the $i$[*th*]{} unit vector. For this reason $I$ is called the [*edge ideal*]{} of $\cal C$. To stress the relationship between $I$ and $\cal C$ we will use the notation $I=I({\cal C})$. A basic example of clutter is a graph. Algebraic and combinatorial properties of edge ideals and graded algebras associated to graphs have been studied in [@susan1; @herzog-survey; @ITG; @carlos-tesis; @Vi2]. The related notion of facet ideal has been studied by Faridi [@faridi; @faridi1] and Zheng [@zheng].
The [*blowup algebras*]{} studied here are the [*Rees algebra*]{} $$R[It]=R\oplus It\oplus\cdots\oplus I^{i}t^i\oplus\cdots
\subset R[t],$$ where $t$ is a new variable, and the [*associated graded ring*]{} $${\rm gr}_I(R)=R/I\oplus I/I^2\oplus\cdots\oplus
I^i/I^{i+1}\oplus\cdots\simeq R[It]\otimes_R(R/I),$$ with multiplication $(a+I^{i+1})(b+I^{j+1})=ab+I^{i+j+1}$, $a\in I^{i}$, $b\in I^{j}$.
In the sequel $A$ will denote the [*incidence matrix*]{} of order $n\times q$ whose column vectors are $v_1,\ldots,v_q$. In order to link the properties of these algebras with combinatorial optimization problems we consider the [*set covering polyhedron*]{} $$Q(A)=\{x\in\mathbb{R}^n\vert\, x\geq 0;\, xA\geq\mathbf{1}\},$$ and the related system of linear inequalities $x\geq 0;\ xA\geq\mathbf{1}$, where $\mathbf{1}=(1,\ldots,1)$. Recall that this system is called [*totally dual integral*]{} (TDI) if the maximum in the LP-duality equation $$\label{jun6-2-03}
{\rm min}\{\langle \alpha,x\rangle \vert\, x\geq 0; xA\geq \mathbf{1}\}=
{\rm max}\{\langle y,\mathbf{1}\rangle \vert\, y\geq 0; Ay\leq\alpha\}$$ has an integral optimum solution $y$ for each integral vector $\alpha$ with finite maximum. If the system is totally dual integral it is seen that $Q(A)$ has only integral vertices, this follows from [@Schr Theorem 22.1, Corollary 22.1.a, pp. 310-311].
We are able to express algebraic properties of blowup algebras in terms of TDI systems and combinatorial properties of clutters, such as the integrality of $Q(A)$ and the König property. An important goal here is to establish bridges between commutative algebra and combinatorial optimization, which could be beneficial to both areas. Necessary and/or sufficient conditions for the normality of $R[It]$ and the reducedness of ${\rm gr}_I(R)$ are shown. Some of our results give some support to a conjecture of Conforti and Cornuéjols (Conjecture \[conforti-cornuejols1\]). Applications to Rees algebras theory and combinatorial optimization are presented.
Along the paper we introduce some of the algebraic and combinatorial notions that are most relevant. For unexplained terminology and notation we refer to [@korte; @oxley; @Schr] and [@Mats; @Vas]. See [@cornu-book] for detailed information about clutters.
Vertex covers of clutters
=========================
The set of non-negative real numbers will be denoted by $\mathbb{R}_+$. To avoid repetitions throughout this article we shall use the notation and assumptions introduced in Section \[Int\]. For convenience we shall always assume that each variable $x_i$ occurs in at least one monomial of $F$.
A subset $C\subset X$ is a [*minimal vertex cover*]{} of the clutter $\cal C$ if: (i) every edge of $\cal C$ contains at least one vertex of $C$, and (ii) there is no proper subset of $C$ with the first property. If $C$ satisfies condition (i) only, then $C$ is called a [*vertex cover*]{} of $\cal C$.
The first aim is to characterize this notion in terms of the integral vertices of set covering polyhedrons and the minimal primes of edge ideals.
[*Notation*]{} The [*support*]{} of $x^a=x_1^{a_1}\cdots x_n^{a_n}$ is ${\rm supp}(x^a)= \{x_i\, |\,
a_i>0\}$.
\[1cover\] The following are equivalent[:]{}
[(a)]{} $\mathfrak{p}=(x_{1},\ldots,x_{r})$ is a minimal prime of $I=I({\cal C})$.
[(b)]{} $C=\{x_{1},\ldots,x_{r}\}$ is a minimal vertex cover of $\cal C$.
[(c)]{} $\alpha=e_{1}+\cdots+e_{r}$ is a vertex of $Q(A)$.
\(a) $\Leftrightarrow$ (b): It follows readily by noticing that the minimal primes of the square-free monomial ideal $I$ are face ideals, that is, they are generated by subsets of the set of variables, see [@monalg Proposition 5.1.3].
\(b) $\Rightarrow$ (c): Fix $1\leq i\leq r$. To make notation simpler fix $i=1$. We may assume that there is an $s_1$ such that $x^{v_j}=x_1m_j$ for $j=1,\ldots,s_1$ and $x_1\notin{\rm supp}(x^{v_j})$ for $j>s_1$. Notice that ${\rm supp}(m_{k_1})\cap(C\setminus\{x_1\})=\emptyset$ for some $1\leq k_1\leq s_1$, otherwise $C\setminus\{x_1\}$ is a vertex cover of $\cal C$ strictly contained in $C$, a contradiction. Thus ${\rm supp}(m_{k_1})\cap C=\emptyset$ because $I$ is square-free. Hence for each $1\leq
i\leq r$ there is $v_{k_i}$ in $\{v_1,\ldots,v_q\}$ such that $x^{v_{k_i}}=x_im_{k_i}$ and ${\rm supp}(m_{k_i})\subset\{x_{r+1},\ldots,x_n\}$. The vector $\alpha$ is clearly in $Q(A)$, and since $\{e_i\}_{i=r+1}^n\cup\{v_{k_1},\ldots,v_{k_r}\}$ is linearly independent, and $$\langle\alpha,e_i\rangle=0\ \ \ \ (i=r+1,\ldots,n);
\ \ \ \langle\alpha,v_{k_i}\rangle=1\ \ \ \ (i=1,\ldots,r),$$ we get that the vector $\alpha$ is a basic feasible solution. Therefore by [@bertsimas Theorem 2.3] $\alpha$ is a vertex of $Q(A)$.
\(c) $\Rightarrow$ (b): It is clear that $C$ intersects all the edges of the clutter $\cal C$ because $\alpha\in Q(A)$. If $C'\subsetneq C$ is a vertex cover of $\cal C$, then the vector $\alpha'=\sum_{x_i\in C'}e_i$ satisfies $\alpha'A\geq
\mathbf{1}$ and $\alpha'\geq 0$. Using that $\alpha$ is a basic feasible solution in the sense of [@bertsimas] it is not hard to verify that $\alpha'$ is also a vertex of $Q(A)$. By the finite basis theorem [@webster Theorem 4.1.3] we can write $$Q(A)=\mathbb{R}_+^n+{\rm conv}(V),$$ where $V$ is the vertex set of $Q(A)$. As $\alpha=\beta+\alpha'$, for some $0\neq \beta\in\mathbb{R}_+^n$, we get $$Q(A)=\mathbb{R}_+^n+{\rm conv}(V\setminus\{\alpha\}).$$ Hence the vertices of $Q(A)$ are contained in $V\setminus\{\alpha\}$ (see [@Bron Theorem 7.2]), a contradiction. Thus $C$ is a minimal vertex cover.
\[may12-06\] A vector $\alpha\in\mathbb{R}^n$ is an integral vertex of $Q(A)$ if and only if $\alpha$ is equal to $e_{i_1}+\cdots+e_{i_s}$ for some minimal vertex cover $\{x_{i_1},\ldots,x_{i_s}\}$ of $\cal C$.
By Proposition \[1cover\] it suffices to observe that any integral vertex of $Q(A)$ has entries in $\{0,1\}$ because $A$ has entries in $\{0,1\}$. See [@shiftcon Lemma 4.6].
A set of edges of the clutter $\cal C$ is [*independent*]{} if no two of them have a common vertex. We denote the smallest number of vertices in any minimal vertex cover of $\cal C$ by ${\alpha}_0({\cal C})$ and the maximum number of independent edges of ${\cal C}$ by $\beta_1({\cal
C})$. These numbers are related to min-max problems because they satisfy: $$\begin{aligned}
\lefteqn{\alpha_0({\cal C})\geq {\rm min}\{\langle
\mathbf{1},x\rangle \vert\, x\geq 0; xA\geq
\mathbf{1}\}}\\
&\ \ \ \ \ \ \ \ \ \ &
={\rm max}\{\langle y,\mathbf{1}\rangle \vert\, y\geq 0;
Ay\leq\mathbf{1}\}
\geq \beta_1({\cal C}). \end{aligned}$$ Notice that $\alpha_0({\cal C})=\beta_1({\cal C})$ if and only if both sides of the equality have integral optimum solutions.
These two numbers can be interpreted in terms of $I$. By Proposition \[1cover\] the height of the ideal $I$, denoted by ${\rm ht}(I)$, is equal to the [*covering number*]{} $\alpha_0({\cal C})$. On the other hand the [*independence number*]{} $\beta_1({\cal C})$ is equal to ${\rm mgrade}(I)$, the [*monomial grade*]{} of the ideal: $$\beta_1({\cal C})=\max\{r\vert\, \exists\, \mbox{ a regular sequence
of monomials }
x^{\alpha_1},\ldots,x^{\alpha_r}\in I\}.$$ The equality $\alpha_0({\cal C})=\beta_1({\cal C})$ is equivalent to require $x_1\cdots x_nt^g\in R[It]$, where $g$ is the covering number $\alpha_0({\cal C})$.
If $\alpha_0({\cal C})=\beta_1({\cal C})$ we say that the clutter $\cal C$ (or the ideal $I$) has the [*König property*]{}.
Rees algebras and polyhedral geometry {#valencia}
=====================================
Let ${\cal A}=\{v_1,\ldots,v_q\}$ be the set of exponent vectors of $x^{v_1},\ldots,x^{v_q}$ and let $${\cal
A}'=\{e_1,\ldots,e_n,(v_1,1),\ldots,(v_q,1)\}\subset\mathbb{R}^{n+1},$$ where $e_i$ is the $i$[*th*]{} unit vector. The [*Rees cone*]{} of ${\cal A}$ is the rational polyhedral cone, denoted by ${\mathbb R}_+{\cal A}'$, consisting of the linear combinations of ${\cal A}'$ with non-negative coefficients. Note $\dim(\mathbb{R}_+{\cal A}')=n+1$. Thus according to [@webster] there is a unique irreducible representation $$\label{okayama}
{\mathbb R}_+{\cal A}'=H_{e_1}^+\cap \cdots\cap H_{e_{n+1}}^+
\cap H_{a_1}^+\cap\cdots\cap H_{a_r}^+$$ such that $0\neq a_i\in\mathbb{Q}^{n+1}$ and $\langle a_i,e_{n+1}\rangle=-1$ for all $i$. As usual $H_{a}^+$ denotes the closed halfspace $$H_a^+=\{\alpha\in\mathbb{R}^{n+1}\vert\, \langle
\alpha,a\rangle\geq 0\}$$ and $H_a$ is the hyperplane through the origin with normal vector $a$.
\[gvvalencia-theo\] The function $\varphi\colon \mathbb{Q}^n\rightarrow\mathbb{Q}^{n+1}$ given by $\varphi(\alpha)=(\alpha,-1)$ induces a bijective map $$\varphi\colon V\longrightarrow\{a_1,\ldots,a_r\}$$ between the set of vertices $V$ of $Q(A)$ and the set $\{a_1,\ldots,a_r\}$ of normal vectors that occur in the irreducible representation of $\mathbb{R}_+{\cal A}'$.
First we show the containment $\varphi(V)\subset\{a_1,\ldots,a_r\}$. Take $\alpha$ in $V$. By [@bertsimas Theorem 2.3] $\alpha$ is a basic feasible solution. Hence $\langle \alpha,v_i\rangle\geq 1$ for $i=1,\ldots,q$, $\alpha\geq 0$, and there exist $n$ linearly independent vectors $v_{i_1},\ldots,v_{i_k},
e_{j_1},\ldots,e_{j_s}$ in ${\cal A}\cup\{e_1,\ldots,e_n\}$ such that $\langle \alpha,v_{i_{h}}\rangle=1$ and $\langle
\alpha,e_{j_m}\rangle=0$ for all ${h},m$. It follows that the set $$F=H_{(\alpha,-1)}\cap\mathbb{R}_+{\cal A}'$$ has dimension $n$ and $\mathbb{R}_+{\cal A}'\subset
H_{(\alpha,-1)}^+$. Therefore $F$ is a facet of $\mathbb{R}_+{\cal A}'$. Using [@webster Theorem 3.2.1] we obtain that $F=\mathbb{R}_+{\cal A}'\cap H_{a_p}$ for some $1\leq p\leq r$, and consequently $H_{(\alpha,-1)}=H_{a_p}$. Since the first $n$ entries of $a_p$ are non-negative and $\langle a_p,e_{n+1}\rangle=-1$ it follows that $\varphi(\alpha)=(\alpha,-1)=a_p$, as desired.
To show the reverse containment write $a_p=(\alpha,-1)$, with $1\leq
p\leq r$ and $\alpha\in\mathbb{R}^n$. We will prove that $\alpha$ is a vertex of $Q(A)$. Since Eq. (\[okayama\]) is an irreducible representation one has that the set $$F=H_{(\alpha,-1)}\cap\mathbb{R}_+{\cal A}'$$ is a facet of the Rees cone $\mathbb{R}_+{\cal A}'$, see [@webster Theorem 3.2.1]. Hence there is a linearly independent set $$\{(v_{i_1},1),\ldots,(v_{i_k},1),e_{j_1},\ldots,e_{j_s}\}\subset{\cal
A}'\ \ \ \ (k+s=n)$$ such that $$\begin{aligned}
\langle(v_{i_{h}},1),(\alpha,-1)\rangle=0&\Rightarrow&
\langle v_{i_{h}},\alpha\rangle=1\ \ \ \
({h}=1,\ldots,k),\label{oct15-03}\\
\langle e_{j_m},(\alpha,-1)\rangle=0&\Rightarrow&
\langle e_{j_m},\alpha\rangle=0 \ \ \ (m=1,\ldots,s).\label{oct15-1-03} \end{aligned}$$ It is not hard to see that $v_{i_1},\ldots,v_{i_k},e_{j_1},\ldots,e_{j_s}$ are linearly independent vectors in $\mathbb{R}^n$. Indeed if $$\lambda_1v_{i_1}+\cdots+\lambda_kv_{i_k}+\mu_1e_{j_1}+\cdots+\mu_se_{j_s}=0\
\ \ \
(\lambda_{h},\mu_m\in\mathbb{R}),$$ then taking inner product with $\alpha$ and using Eqs. (\[oct15-03\]) and (\[oct15-1-03\]) we get $$\lambda_1+\cdots+\lambda_k=0\ \Rightarrow\
\lambda_1(v_{i_1},1)+\cdots+\lambda_k(v_{i_k},1)+\mu_1e_{j_1}+
\cdots+\mu_se_{j_s}=0.$$ Therefore $\lambda_{h}=0$ and $\mu_m=0$ for all ${h},m$, as desired. From $\mathbb{R}_+{\cal A}'\subset H_{a_p}^+$ and $H_{a_p}^+=H_{(\alpha,-1)}^+$ we get $\alpha\geq 0$ and $\langle\alpha,v_i\rangle\geq 1$ for all $i$. Altogether we obtain that $\alpha$ is a basic feasible solution, that is, $\alpha$ is a vertex of $Q(A)$.
Let $\mathfrak{p}_1,\ldots,\mathfrak{p}_s$ be the minimal primes of the edge ideal $I=I({\cal C})$ and let $$C_k=\{x_i\vert\, x_i\in\mathfrak{p}_k\}\ \ \ \ (k=1,\ldots,s)$$ be the corresponding minimal vertex covers of the clutter $\cal C$. By Proposition \[1cover\] and Theorem \[gvvalencia-theo\] in the sequel we may assume that
$$a_k=(\textstyle\sum_{x_i\in C_k}e_i,-1) \ \ \ \ (k=1,\ldots,s).$$
[*Notation*]{} Let $d_k$ be the unique positive integer such that $d_ka_k$ has relatively prime integral entries. We set $\ell_k=d_ka_k$ for $k=1,\ldots,r$. If the first $n$ rational entries of $a_k$ are written in lowest terms, then $d_k$ is the least common multiple of the denominators. For $1\leq k\leq r$, we have $d_k=-\langle\ell_k,e_{n+1}\rangle$.
The set covering polyhedron $Q(A)$ is [*integral*]{} if all its vertices have integral entries.
\[1cover-1\] The irreducible representation of the Rees cone has the form $$\label{okayama-oct}
{\mathbb R}_+{\cal A}'=H_{e_1}^+\cap \cdots\cap H_{e_{n+1}}^+
\cap H_{\ell_1}^+\cap\cdots\cap H_{\ell_r}^+,$$ $d_k=1$ if and only if $1\leq k\leq s$, and $Q(A)$ is integral if and only if $r=s$.
It follows from Theorem \[gvvalencia-theo\] and Corollary \[may12-06\].
[*Notation*]{} In the sequel we shall always assume that $\ell_1,\ldots,\ell_r$ are the integral vectors of Eq. (\[okayama-oct\]).
Recall that the [*Simis cone*]{} of ${\cal A}$ is the rational polyhedral cone $${\rm Cn}({\cal A})=H_{e_1}^+\cap \cdots\cap H_{e_{n+1}}^+
\cap H_{\ell_1}^+\cap\cdots\cap H_{\ell_s}^+,$$ and the [*symbolic Rees algebra*]{} of $I$ is the $K$ algebra: $$R_s(I)=R+I^{(1)}t+I^{(2)}t^2+\cdots+I^{(i)}t^i+\cdots\subset R[t],$$ where $I^{(i)}=\mathfrak{p}_1^i\cap\cdots\cap\mathfrak{p}_s^i$ is the $i$[*th*]{} symbolic power of $I$. Symbolic Rees algebras have a combinatorial interpretation [@cover-algebras]. Notice the following description: $$I^{(b)}=(\{x^a\vert\, \langle (a,b),\ell_i\rangle\geq 0\mbox{ for
}i=1,\ldots,s\}).$$
A first use of the Simis cone is the following expression for the symbolic Rees algebra. In particular $R_s(I)$ is a finitely generated $K$-algebra [@Lyu3] by Gordan’s Lemma [@BHer].
\[oct24-03\] If $S=\mathbb{Z}^{n+1}\cap{\rm Cn}({\cal A})$ and $K[S]=K[\{x^{a}t^b\vert\, (a,b)\in S\}]$ is its semigroup ring, then $R_s(I)=K[S]$.
Let $\mathbb{N}{\cal A}'$ be the subsemigroup of $\mathbb{N}^{n+1}$ generated by ${\cal A}'$, consisting of the linear combinations of ${\cal A}'$ with non-negative integer coefficients. The Rees algebra of $I$ can be written as $$\begin{aligned}
R[It]&=&
K[\{x^at^b\vert\, (a,b)\in\mathbb{N}{\cal A}'\}]\label{may6-06-2}\\ &=&
R\oplus It\oplus\cdots\oplus I^{i}t^i\oplus\cdots
\subset R[t].\label{may6-06-3}\end{aligned}$$ According to [@monalg Theorem 7.2.28] and [@Vas1 p. 168] the integral closure of $R[It]$ in its field of fractions can be expressed as $$\begin{aligned}
\overline{R[It]}&=&K[\{x^at^b\vert\, (a,b)\in
\mathbb{Z}^{n+1}\cap \mathbb{R}_+{\cal A}'\}]\label{may6-06}\\
&=&R\oplus
\overline{I}t\oplus\cdots\oplus
\overline{I^i}t^i\oplus\cdots,\label{jun05-1-03} \end{aligned}$$ where $\overline{I^i}=(\{x^a\in R\vert\, \exists\, p\geq
1;(x^a)^{p}\in I^{pi}\})$ is the integral closure of $I^i$. Hence, by Eqs. (\[may6-06-2\]) to $(\ref{jun05-1-03})$, we get that $R[It]$ is a normal domain if and only if any of the following two equivalent conditions hold:
[(a)]{} $\mathbb{N}{\cal A}'=
\mathbb{Z}^{n+1}\cap\mathbb{R}_+{\cal A}'$.
[(b)]{} $I^{i}=\overline{I^i}$ for all $i\geq 1$.
If the second condition holds we say that $I$ is a [*normal*]{} ideal.
For $1\leq i\leq r$ we write $a_i=(a_i',-1)$. Let $B$ be the matrix with column vectors $a_1',\ldots,a_r'$ and let $Q=\mathbb{Q}_+^n+{\rm conv}(v_1,\ldots,v_q)$. Then
[(a)]{} $\overline{I^i}=(\{x^a\in R\vert\,
a\in{iQ}\cap\mathbb{Z}^n\})$.
[(b)]{} $Q=Q(B)=\{x\vert\, x\geq 0;\,
xB\geq\mathbf{1}\}$. In particular $Q(B)$ is integral.
Part (a) follows from Eq. (\[jun05-1-03\]) and part (b) follows from Eq. (\[okayama\]).
In the sequel $J_k^{(d_k)}$ will denote the ideal of $R[It]$ given by $$J_k^{(d_k)}=(\{x^at^b\in R[It]\vert\, \langle(a,b),
\ell_k\rangle\geq d_k\}) \ \ \ \ (k=1,\ldots,r)$$ and $J_k$ will denote the ideal of $R[It]$ given by $$J_k=(\{x^at^b\in R[It]\vert\, \langle(a,b),
\ell_k\rangle>0\}) \ \ \ \ (k=1,\ldots,r),$$ where $d_k=-\langle\ell_k,e_{n+1}\rangle$. If $d_k=1$, we have $J_k^{(1)}=J_k$. In general $J_k^{(d_k)}$ might not be equal to the $d_k$[*th*]{} symbolic power of $J_k$. The localization of $R[It]$ at $R\setminus \mathfrak{p}_k$ is denoted by $R[It]_{\mathfrak{p}_k}$.
\[oct28-03\] $J_1,\ldots,J_r$ are height one prime ideals containing $IR[It]$ and $J_k$ is equal to $\mathfrak{p}_kR[It]_{\mathfrak{p}_k}\cap R[It]$ for $k=1,\ldots,s$. If $Q(A)$ is integral, then $${\rm rad}(IR[It])=J_1\cap J_2\cap\cdots\cap J_s.$$
$IR[It]$ is clearly contained in $J_k$ for all $k$ by construction. To show that $J_k$ is a prime ideal of height one it suffices to notice that the right hand side of the isomorphism: $$R[It]/J_k\simeq K[\{x^at^b\in R[It]\vert\, \langle(a,b),\ell_k\rangle=0 \}]$$ is an $n$-dimensional integral domain, because $F_k=\mathbb{R}_+{\cal
A}'\cap H_{\ell_k}$ is a facet of the Rees cone for all $k$. Set $P_k=\mathfrak{p}_kR[It]_{\mathfrak{p}_k}\cap R[It]$ for $1\leq k\leq s$. This ideal is a minimal prime of $IR[It]$ (see [@HuSV]) and admits the following description $$\begin{aligned}
P_k&=&
\mathfrak{p}_kR_{\mathfrak{p}_k}[\mathfrak{p}_kR_{\mathfrak{p}_k}t]
\cap R[It]
\\
&=&\mathfrak{p}_k+(\mathfrak{p}_k^2\cap I)t+
(\mathfrak{p}_k^3\cap I^2)t^2+
\cdots+(\mathfrak{p}_k^{i+1}\cap I^i)t^i+\cdots\end{aligned}$$ Notice that $x^a\in\mathfrak{p}_k^{b+1}$ if and only if $\langle a,\sum_{x_i\in C_k}e_i\rangle\geq b+1$. Hence $J_k=P_k$.
Assume that $Q(A)$ is integral, i.e., $r=s$. Take $x^\alpha t^b\in J_k$ for all $k$. Using Eq. (\[okayama-oct\]) it is not hard to see that $(\alpha,b+1)\in\mathbb{R}_+{\cal A}'$, that is $x^\alpha t^{b+1}$ is in $\overline{R[It]}$ and $x^\alpha t^{b+1}\in
\overline{I^{b+1}}t^{b+1}$. It follows that $x^\alpha t^b$ is a monomial in the radical of $IR[It]$. This proves the asserted equality.
For use below recall that the [*analytic spread*]{} of $I$ is given by $$\ell(I)=\dim R[It]/\mathfrak{m}R[It];\ \ \ \mathfrak{m}=(x_1,\ldots,x_n).$$
\[march19-06\] If $Q(A)$ is integral, then $\ell(I)<n$.
Since $Q(A)$ is integral, we have $r=s$. If $\ell(I)=n$, then the height of $\mathfrak{m}R[It]$ is equal to $1$. Hence there is a height one prime ideal $P$ of $R[It]$ such that $IR[It]\subset\mathfrak{m}R[It]\subset P$. By Proposition \[oct28-03\] the ideal $P$ has the form $\mathfrak{p}_kR[It]_{\mathfrak{p}_k}\cap R[It]$, this readily yields a contradiction.
\[burch\] ${\rm inf}\{{\rm
depth}(R/I^i)\vert\, i\geq 1\}\leq\dim(R)-\ell(I)$. If ${\rm
gr}_I(R)$ is Cohen-Macaulay, then the equality holds.
By a result of Brodmann [@brodmann], the depth of $R/I^k$ is constant for $k$ sufficiently large. Broadmann improved this inequality by showing that the constant value is bounded by $\dim(R)-\ell(I)$. For a study of the initial and limit behavior of the numerical function $f(k)={\rm depth}\, R/I^k$ see [@depth].
\[nov16-03\] Let $x_1\in C_k$ for some $1\leq k\leq
s$. If $x^{v_i}=x_1x^{v_i'}$ for $1\leq i\leq p$ and $x_1\notin{\rm supp}(x^{v_i})$ for $i>p$, then there is $x^{v_j'}$ such that ${\rm supp}(x^{v_j'})\cap C_k=\emptyset$.
If ${\rm supp}(x^{v_j'})\cap C_k\neq\emptyset$ for all $j$, then $C_k\setminus\{x_1\}$ is a vertex cover of $\cal C$, a contradiction because $C_k$ is a minimal vertex cover.
If $P\in\{J_1,\ldots,J_s\}$, then $R[It]\cap IR[It]_P=P$.
Set $P=J_k$. We may assume that $x_1,\ldots,x_m$ (resp. $x^{v_1}t,\ldots,x^{v_p}t$) is the set of all $x_i$ (resp. $x^{v_i}t$) such that $x_i\in P$ (resp. $x^{v_i}t\in P$). Notice that $\mathfrak{p}_k$ is equal to $(x_1,\ldots,x_m)$ and set $C=\{x_1,\ldots,x_m\}$. In general the left hand side is contained in $P$. To show the reverse inclusion we first prove the equality $$\label{may2-06}
P=(x_1,\ldots,x_m,x^{v_1}t,\ldots,x^{v_p}t)R[It].$$ Let $x^at^b\in P$. Thus $x^at^b=x_1^{\mu_1}\cdots
x_n^{\mu_n}(x^{v_1}t)^{\lambda_1}\cdots(x^{v_q}t)^{\lambda_q}$ and $\langle (a,b),\ell_k\rangle>0$. Hence $\langle e_i,\ell_k\rangle>0$ for some $i$ or $\langle (v_j,1),\ell_k\rangle>0$ for some $j$. Therefore $x^at^b$ belongs to the right hand side of Eq. (\[may2-06\]), as required.
Case (I): Consider $x_\ell$ with $1\leq \ell\leq m$. By Lemma \[nov16-03\] there is $j$ such that $x^{v_j}=x_{\ell}x^{\alpha}$ and ${\rm supp}(x^{\alpha})\cap C=\emptyset$. Thus since $x^{\alpha}$ is not in $P$ (because of the second condition) we obtain $x_\ell\in R[It]\cap
IR[It]_P$.
Case (II): Consider $x^{v_\ell}t$ with $1\leq \ell\leq p$. Since $$\langle(v_\ell,1),e_1+\cdots+e_m-e_{n+1}\rangle\geq 1,$$ the monomial $x^{v_\ell}$ contains at least two variables in $C$. Thus we may assume that $x_1,x_2$ are in the support of $x^{v_\ell}$. Again by Lemma \[nov16-03\] there are $j,j_1$ such that $x^{v_j}=x_1x^{\alpha}$, $x^{v_{j_1}}=x_2x^{\gamma}$, and the support of $x^\alpha$ and $x^\gamma$ disjoint from $C$. Hence the monomial $x^{v_\ell}x^{\alpha+\gamma}t$ belongs to $I^2t$ and $x^{\alpha+\gamma}$ is not in $P$. Writing $$x^{v_\ell}t=({x^{v_\ell}x^{\alpha+\gamma}t})/{x^{\alpha+\gamma}},$$ we get $x^{v_\ell}t\in R[It]\cap IR[It]_P$.
${\rm rad}(J_k^{(d_k)})=J_k$ for $1\leq k\leq r$.
By construction one has ${\rm rad}(J_k^{(d_k)})\subset J_k$. The reverse inclusion follows by noticing that if $x^at^b\in J_k$, then $(x^at^b)^{d_k}\in J_k^{(d_k)}$.
\[primdec-irit\] If $R[It]$ is normal, then $IR[It]=J_1^{(d_1)}\cap\cdots\cap J_r^{(d_r)}$.
“$\subset$”: Let $x^a t^{b}\in IR[It]$. Since $x^a\in I^{b+1}$, we obtain $(a,b+1)\in\mathbb{N}{\cal A}'$. In particular we get $(a,b+1)\in\mathbb{R}_+{\cal A}'$. Therefore $$0\leq \langle(a,b+1),\ell_k\rangle=\langle(a,b),\ell_k\rangle-d_k$$ and consequently $x^a t^b\in J^{(d_k)}$ for $1\leq k\leq r$.
“$\supset$”: Let $x^a t^b\in J^{(d_k)}$ for all $k$. Since $(a,b+1)\in \mathbb{R}_+{\cal A}'\cap\mathbb{Z}^{n+1}$, using that $R[It]$ is normal yields $(a,b+1)\in\mathbb{N}{\cal A}'$. It follows that $x^a t^b\in I^{b+1}t^b\subset IR[It]$.
A similar formula is shown in [@bruns-tania]. The normality of $R[It]$ can be described in terms of primary decompositions of $IR[It]$, see [@HSV Proposition 2.1.3].
The following two nice formulas, pointed out to us by Vasconcelos, describe the difference between the symbolic Rees algebra of $I$ and the normalization of its Rees algebra. If $\mathfrak{q}_k=J_k\cap{R}$ for $k=1,\ldots,r$, then $$\begin{aligned}
R_s(I)=\bigcap_{k=1}^s R[It]_{\mathfrak{q}_k}\cap R[t];& &
\overline{R[It]}=\bigcap_{k=1}^r
\overline{R[It]}_{\mathfrak{q}_k}\cap R[t].\end{aligned}$$ These representations are linked to the so called Rees valuations of the ideal $I$, see [@bookthree Chapter 8] for further details.
\[oct30-03\] The following conditions are equivalent
[(a)]{} $Q(A)$ is integral.
[(b)]{} ${\mathbb R}_+{\cal A}'=H_{e_1}^+\cap \cdots\cap H_{e_{n+1}}^+
\cap H_{\ell_1}^+\cap\cdots\cap H_{\ell_s}^+$, i.e., $r=s$.
[(c)]{} $R_s(I)=\overline{R[It]}$.
[(d)]{} The minimal primes of $IR[It]$ are of the form $\mathfrak{p}_kR[It]_{\mathfrak{p}_k}\cap R[It]$.
\(a) $\Leftrightarrow$ (b) $\Leftrightarrow$ (c): These implications follow from Theorems \[gvvalencia-theo\] and \[oct24-03\]. The other implications follow readily using Proposition \[oct28-03\].
Let $x^{u_k}=\prod_{x_i\in C_k}x_i$ for $1\leq
k\leq s$. The ideal of [*vertex covers*]{} of ${\cal C}$ is the ideal $$I_c({\cal C})=(x^{u_1},\ldots,x^{u_s})\subset R.$$ The clutter of minimal vertex covers, denoted by $\cal D$ or $b({\cal C})$, is the [*blocker*]{} of $\cal C$.
In the literature $I_c({\cal C})$ is also called the [*Alexander dual*]{} of $I$ because if $\Delta$ is the Stanley-Reisner complex of $I$, then $I_c({\cal C})$ is the Stanley-Reisner ideal of the Alexander dual of $\Delta$. The survey article [@herzog-survey] explains the role of Alexander duality to prove combinatorial and algebraic theorems.
Let $I=(x_1x_2x_5,x_1x_3x_4,x_2x_3x_6,x_4x_5x_6)$. The clutter of $I$ is denoted by ${\cal Q}_6$. Using [*Normaliz*]{} [@B] and Proposition \[oct30-03\] we obtain: $$R[It]\subsetneq{R_s(I)}=\overline{R[It]}=R[It][x_1\cdots x_6t^2]\
\mbox{ and }\ R[I_c({\cal Q}_6)t]=\overline{R[I_c({\cal Q}_6)t]}.$$
\[nov3-03\] If $\overline{R[It]}=R_s(I)$ and $J=I_c({\cal C})$, then $\overline{R[Jt]}=R_s(J)$.
[[@cornu-book Theorem 1.17]]{} If $Q(A)$ is integral and $A'$ is the incidence matrix of the clutter of minimal vertex covers of $\cal C$, then $Q(A')$ is integral.
It follows at once from Propositions \[oct30-03\] and \[nov3-03\].
Let $X'=\{x_{i_1},\ldots,x_{i_r},x_{j_1},\ldots,x_{j_s}\}$ be a subset of $X$. A [*minor*]{} of $I$ is a proper ideal $I'$ of $R'=K[X\setminus
X']$ obtained from $I$ by making $x_{i_k}=0$ and $x_{j_\ell}=1$ for all $k,\ell$. The ideal $I$ is considered itself a minor. A [*minor*]{} of $\cal C$ is a clutter ${\cal C}'$ that corresponds to a minor $(0)\subsetneq
I'\subsetneq R'$.
Notice that ${\cal C}'$ is obtained from $I'$ by considering the unique set of square-free monomials of $R'$ that minimally generate $I'$.
If $\overline{I^i}=I^{(i)}$ for some $i\geq 2$ and $J=I'$ is a minor of $I$, then $\overline{J^i}=J^{(i)}$.
Assume that $J$ is the minor obtained from $I$ by making $x_1=0$. Take $x^a\in{J^{(i)}}$. Then $x^a\in I^{(i)}=\overline{I^i}$ because $J\subset I$. Thus $x^a\in\overline{I^i}$. Since $x_1\notin{\rm supp}(x^a)$ it follows that $x^a\in \overline{J^i}$. This proves $J^{(i)}\subset\overline{J^i}$. The other inclusion is clear because $J^{(i)}$ is integrally closed.
Assume that $J$ is the minor obtained from $I$ by making $x_1=1$. Take $x^a\in J^{(i)}$. Notice that $x_1^ix^a\in I^{(i)}=\overline{I^i}$. Indeed if $x_1\in\mathfrak{p}_k$, then $x_1^i\in\mathfrak{p}_k^i$, and if $x_1\notin\mathfrak{p}_k$, then $J\subset \mathfrak{p}_k$ and $x^a\in\mathfrak{p}_k^i$. Since $x_1\notin{\rm supp}(x^a)$ it follows that $x^a\in \overline{J^i}$.
\[nov19-1-03\] If $R_s(I)=\overline{R[It]}$, then $R_s(I')=\overline{R'[I't]}$ for any minor $I'$ of $I$.
\[nov2-03\] Let ${\cal D}$ be the clutter of minimal vertex covers of $\cal C$. If $\overline{R[It]}$ is equal to $R_s(I)$ and $|A\cap B|\leq 2$ for $A\in \cal C$ and $B\in \cal D$, then $R[It]$ is normal.
Let $x^at^b=x_1^{a_1}\cdots x_n^{a_n}t^b\in\overline{R[It]}$ be a minimal generator, that is $(a,b)$ cannot be written as a sum of two non-zero integral vectors in the Rees cone $\mathbb{R}_+{\cal A}'$. We may assume $a_i\geq 1$ for $1\leq i\leq m$, $a_i=0$ for $i>m$, and $b\geq 1$.
Case (I): $\langle(a,b),\ell_i\rangle>0$ for all $i$. The vector $\gamma=(a,b)-e_1$ satisfies $\langle\gamma,\ell_i\rangle\geq 0$ for all $i$, that is $\gamma\in\mathbb{R}_+{\cal A}'$. Thus since $(a,b)=e_1+\gamma$ we derive a contradiction.
Case (II): $\langle(a,b),\ell_i\rangle=0$ for some $i$. We may assume $$\{\ell_i\vert\, \langle(a,b),\ell_i\rangle=0\}=\{\ell_1,\ldots,\ell_p\}.$$
Subcase (II.a): $e_i\in H_{\ell_1}\cap\cdots\cap H_{\ell_p}$ for some $1\leq i\leq m$. It is not hard to verify that the vector $\gamma=(a,b)-e_i$ satisfies $\langle\gamma,\ell_k\rangle\geq 0$ for all $1\leq k\leq s$. Thus $\gamma\in\mathbb{R}_+{\cal A}'$, a contradiction because $(a,b)=e_i+\gamma$.
Subcase (II.b): $e_i\notin H_{\ell_1}\cap\cdots\cap H_{\ell_p}$ for all $1\leq i\leq m$. Since the vector $(a,b)$ belongs to $\mathbb{R}_+{\cal A}'$, it follows (see the proof of Theorem \[oct31-03\]) that we can write $$\label{may6-06-1}
(a,b)=\lambda_1(v_1,1)+\cdots+\lambda_q(v_q,1) \ \ \ \ (\lambda_i\geq 0).$$ By the choice of $x^at^b$ we may assume $0<\lambda_1<1$. Set $\gamma=(a,b)-(v_1,1)$ and notice that by Eq. (\[may6-06-1\]) this vector has non-negative entries. We claim that $\gamma$ is in the Rees cone. Since by hypothesis one has $0\leq \langle(v_1,1),\ell_j\rangle\leq 1$ for all $j$ we readily obtain $$\langle\gamma,\ell_k\rangle=\left\{\begin{array}{cccl}
\langle(a,b),\ell_k\rangle-\langle (v_1,1),\ell_k\rangle&=& 0
&\mbox{if }\ 1\leq k\leq p,\\
\langle(a,b),\ell_k\rangle-\langle(v_1,1),\ell_k\rangle&\geq& 0 &\mbox{otherwise.}
\end{array} \right.$$ Thus $\gamma\in\mathbb{R}_+{\cal A}'$ and $(a,b)=(v_1,1)+\gamma$. As a result $\gamma=0$ and $x^at^b\in R[It]$, as desired.
König property of clutters and normality
========================================
Let us introduce a little bit more notation and definitions. Recall that the [*Ehrhart ring*]{} of the lattice polytope $P={\rm conv}(v_1,\ldots,v_q)$ is the subring $$A(P)=K[\{x^at^i\vert\, a\in \mathbb{Z}^n \cap iP; i\in \mathbb{N}\}]\subset R[t],$$ and the homogeneous [*monomial subring*]{} generated by $Ft=\{x^{v_1}t,\ldots,x^{v_q}t\}$ over the field $K$ is the subring $K[Ft]\subset R[t]$.
\[oct31-03\] If $\overline{R[It]}=R_s(I)$ and $K[Ft]=A(P)$, then $R[It]$ is normal.
Let $x^at^b=x_1^{a_1}\cdots x_n^{a_n}t^b\in\overline{R[It]}$ be a minimal generator, that is $x^at^b$ cannot be written as a product of two non-constant monomials of $\overline{R[It]}$. We may assume $a_i\geq 1$ for $1\leq i\leq m$, $a_i=0$ for $i>m$, and $b\geq 1$.
Case (I): $\langle(a,b),\ell_i\rangle>0$ for all $i$. The vector $\gamma=(a,b)-e_1$ satisfies $\langle\gamma,\ell_i\rangle\geq 0$ for all $i$, that is $\gamma\in\mathbb{R}_+{\cal A}'$. Thus since $x_1$ and $x_1^{a_1-1}x_2^{a_2}\cdots x_{n}^{a_n}t^b$ are in $\overline{R[It]}$ we get a contradiction. In conclusion this case cannot occur.
Case (II): $\langle(a,b),\ell_i\rangle=0$ for some $i$. We may assume $$\{\ell_i\vert\, \langle(a,b),\ell_i\rangle=0\}=\{\ell_1,\ldots,\ell_p\}.$$
Subcase (II.a): $e_i\in H_{\ell_1}\cap\cdots\cap H_{\ell_p}$ for some $1\leq i\leq m$. For simplicity of notation assume $i=1$. The vector $\gamma=(a,b)-e_1$ satisfies $$\langle\gamma,\ell_k\rangle=\left\{\begin{array}{cccl}
\langle(a,b),\ell_k\rangle-\langle e_1,\ell_k\rangle&=& 0 &\mbox{if
}\ 1\leq k\leq p,\\
\langle(a,b),\ell_k\rangle-\langle e_1,\ell_k\rangle&\geq& 0 &\mbox{otherwise.}
\end{array} \right.$$ Thus $\gamma\in\mathbb{R}_+{\cal A}'$. Proceeding as in Case (I) we derive a contradiction.
Subcase (II.b): $e_i\notin H_{\ell_1}\cap\cdots\cap H_{\ell_p}$ for all $1\leq i\leq m$. The vector $(a,b)$ belongs to the polyhedral cone $$C=H_{\ell_1}\cap\cdots\cap H_{\ell_p}\cap \mathbb{R}_+{\cal A}'.$$ Hence we can write $$\begin{aligned}
(a,b)&=&\lambda_1(v_1,1)+\cdots+\lambda_q(v_q,1)+\mu_1e_1+\cdots+\mu_ne_n
\ \ \ \
(\lambda_i;\mu_j\geq 0),\\
\langle(a,b),\ell_k\rangle&=&\lambda_1\langle(v_1,1),\ell_k\rangle+\cdots
+\lambda_q\langle(v_q,1),\ell_k\rangle+\\
& &\mu_1\langle e_1,\ell_k\rangle+\cdots+
\mu_n\langle e_n,\ell_k\rangle=0\end{aligned}$$ for $k=1,\ldots,p$. From the first equality we get $\mu_i=0$ for $i>m$ because $a_i=0$ for $i>m$. If $\mu_i>0$ for some $1\leq i\leq m$, then $\langle e_i,\ell_k\rangle=0$ for $1\leq k\leq p$, a contradiction. Hence $\mu_i=0$ for all $i$. Therefore $a/b\in P$ and $a\in\mathbb{Z}^{n}\cap bP$. This proves $x^at^b\in A(P)=K[Ft]\subset
R[It]$, as desired.
If $\ell_k=d_ka_k$ has the form $$\ell_k=e_{i_1}+\cdots+e_{i_k}-d_ke_{n+1}\ \ \ \ \ (1\leq
i_1<\cdots<i_k\leq n)$$ for $k=1,\ldots,r$, then $A(P)[x_1,\ldots,x_n]=\overline{R[It]}$.
The Ehrhart ring is contained in $\overline{R[It]}$. Thus the equality follows using the proof of Theorem \[oct31-03\].
If $x^{v_1},\ldots,x^{v_q}$ have degree $d\geq 2$ and $\overline{I^b}=I^{(b)}$ for all $b$, then $\overline{I^b}$ is generated by monomials of degree $bd$ for $b\geq 1$.
The monomial ideal $\overline{I^b}$ has a unique minimal set of generators consisting of monomials. Take $x^a$ in this minimal set. Notice that $(a,b)\in\mathbb{R}_+{\cal A}'$. Thus we may proceed as in the proof of Theorem \[oct31-03\] to obtain that $(a,b)$ is in the cone generated by $\{(v_1,1),\ldots,(v_q,1)\}$. This yields $\deg(x^a)=bd$.
\[nov1-03\] If $x^{v_1},\ldots,x^{v_q}$ have degree $d\geq 2$, then $I^{i}=I^{(i)}$ for all $i\geq 1$ if and only if $Q(A)$ is integral and $K[Ft]=A(P)$.
$\Rightarrow$) By Proposition \[oct30-03\] the polyhedron $Q(A)$ is integral. Since $I^{(i)}$ is integrally closed [@monalg Corollary 7.3.15], we get that $R[It]$ is normal. Therefore applying [@ehrhart Theorem 3.15] we obtain $K[Ft]=A(P)$, here the hypothesis on the degrees of $x^{v_i}$ is essential.
$\Leftarrow$) By Proposition \[oct30-03\] $\overline{I^{i}}=I^{(i)}$ for all $i$, thus applying Theorem \[oct31-03\] gives $R[It]$ normal and we get the required equality. Here the hypothesis on the degrees of $x^{v_i}$ is not needed.
The clutter $\cal C$ satisfies the [*max-flow min-cut*]{} (MFMC) property if both sides of the LP-duality equation $$\label{jun6-2-03-1}
{\rm min}\{\langle \alpha,x\rangle \vert\, x\geq 0; xA\geq \mathbf{1}\}=
{\rm max}\{\langle y,\mathbf{1}\rangle \vert\, y\geq 0; Ay\leq\alpha\}$$ have integral optimum solutions $x$ and $y$ for each non-negative integral vector $\alpha$.
It follows from [@Schr pp. 311-312] that $\cal C$ has the MFMC property if and only if the maximum in Eq. (\[jun6-2-03-1\]) has an optimal integral solution $y$ for each non-negative integral vector $\alpha$. Thus the system $x\geq 0;\
xA\geq\mathbf{1}$ is TDI if and only if $\cal C$ has the max-flow min-cut property.
A ring is called [*reduced*]{} if $0$ is its only nilpotent element. For convenience let us state some known characterizations of the reducedness of the associated graded ring.
\[noclu\] The following conditions are equivalent
[(i) ]{} ${\rm gr}_I(R)$ is reduced.
[(ii) ]{} $R[It]$ is normal and $Q(A)$ is an integral polyhedron.
[(iii)]{} $I$ is normally torsion free, that is, $I^i=I^{(i)}$ for all $i\geq 1$.
[(iv)]{} $x\geq 0;\, xA\geq \mathbf{1}$ is a [TDI]{} system.
[(iv)]{} $\cal C$ has the max-flow min-cut property.
[[@CGM Theorem 1.3]]{} Let ${\cal D}$ be the clutter of minimal vertex covers of $\cal C$. If $Q(A)$ is integral and $|A\cap B|\leq 2$ for $A\in \cal C$ and $B\in \cal D$, then $x\geq 0;\, xA\geq \mathbf{1}$ is a [TDI]{} system.
By Proposition \[nov2-03\] the Rees algebra $R[It]$ is normal. To complete the proof apply Theorem \[noclu\].
\[dec25-02\] If $I$ is a monomial ideal of $R$, then the nilradical of the associated graded ring of $I$ is given by $${\rm nil}({\rm gr}_I(R))
=(\{\overline{x^\alpha}\in I^i/I^{i+1}\vert\,
x^{s\alpha}\in I^{si+1}; i\geq 0; s\geq 1\}).$$
The nilradical of ${\rm gr}_I(R)$ is graded with respect to the fine grading, and thus it is generated by homogeneous elements.
The matrix $A$ is [*balanced*]{} if $A$ has no square submatrix of odd order with exactly two $1$’s in each row and column. $A$ is [*totally unimodular*]{} if each $i\times i$ minor of $A$ is $0$ or $\pm 1$ for all $i\geq 1$.
\[dec30-03\] If $A$ is balanced, then ${\rm gr}_{I}(R)$ is reduced.
Let $\overline{x^{\alpha}}\in I^i/I^{i+1}$ be in ${\rm
nil}({\rm gr}_I(R))$, that is $x^{s\alpha}\in I^{is+1}$ for some $0\neq s\in\mathbb{N}$. By Lemma \[dec25-02\] we need only show $\overline{x^\alpha}=\overline{0}$. It follows rapidly that the maximum in Eq. (\[jun6-2-03\]) is greater or equal than $i+1/s$. By [@Schr Theorem 21.8, p. 305] the maximum in Eq. (\[jun6-2-03\]) has an integral optimum solution $y=(y_1,\ldots,y_q)$. Thus $y_1+\cdots+y_q\geq i+1$. Since $y$ satisfies $y\geq 0$ and $Ay\leq
\alpha$ we obtain $x^\alpha\in I^{i+1}$. This proves $\overline{x^\alpha}=\overline{0}$, as required.
\[balanced-mfmc-blocker\] If $A$ is balanced and $J=I_c({\cal C})$, then $R[Jt]=R_s(J)$.
Let $\cal D$ be the blocker of $\cal C$. By [@Schr2 Corollary 83.1a(v), p. 1441], we get that $\cal D$ satisfy the max-flow min-cut property. Thus the equality follows at once from Theorem \[noclu\].
\[oct20-03\] If ${\rm gr}_I(R)$ is reduced $($resp. $R[It]$ is normal$)$ and $I'$ is a minor of $I$, then ${\rm gr}_{I'}(R')$ is reduced $($resp. $R'[I't]$ is normal$)$.
Notice that we need only show the result when $I'$ is a minor obtained from $I$ by making $x_1=0$ or $x_1=1$. Using Lemma \[dec25-02\] both cases are quite easy to prove.
A clutter $\cal C$ satisfies the [*packing property*]{} (PP) if all its minors satisfy the König property, that is, $\alpha_0({\cal C}')=\beta_1({\cal C}')$ for any minor ${\cal C}'$ of $\cal C$.
\[oct21-03\] If the ring ${\rm gr}_I(R)$ is reduced, then $\alpha_0({\cal C}')=
\beta_1({\cal C}')$ for any minor ${\cal C}'$ of $\cal C$.
Let ${\cal C}'$ be any minor of $\cal C$ and let $I'$ be its clutter ideal. We denote the incidence matrix of ${\cal C}'$ by $A'$. By Proposition \[oct20-03\] the associated graded ring ${\rm gr}_{I'}(R')$ is reduced. Hence by Theorem \[noclu\] the clutter ${\cal C}'$ has the max-flow min-cut property. In particular the LP-duality equation $${\rm min}\{\langle\mathbf{1},x\rangle \vert\, x\geq 0; xA'\geq \mathbf{1}\}=
{\rm max}\{\langle y,\mathbf{1}\rangle \vert\, y\geq 0; A'y\leq\mathbf{1}\}$$ has optimum integral solutions $x$, $y$. To complete the proof notice that the left hand side of this equality is $\alpha_0({\cal C}')$ and the right hand side is $\beta_1({\cal C}')$.
If $I$ is the facet ideal of a simplicial tree, then ${\rm gr}_I(R)$ is reduced. This follows from [@faridi p. 174] using the proof of [@ITG Corollary 3.2, p. 399]. In particular $\cal C$ has the König property, this was shown in [@faridi1 Theorem 5.3].
\[jan20-05\] If $\cal C$ has the max-flow min-cut property, then $\cal C$ has the packing property.
It follows at once from Theorem \[noclu\] and Corollary \[oct21-03\].
Conforti and Cornuéjols conjecture that the converse is also true:
[[@cornu-book Conjecture 1.6]]{} \[conforti-cornuejols1\] If the clutter $\cal C$ has the packing property, then $\cal C$ has the max-flow min-cut property.
Next we state the converse of Corollary \[oct21-03\] as an algebraic version of this interesting conjecture which to our best knowledge is still open:
\[conforti-cornuejols\]If $\alpha_0({\cal C}')=\beta_1({\cal C}')$ for all minors $\cal C'$ of $\cal C$, then the ring ${\rm gr}_I(R)$ is reduced.
It is known [@cornu-book Theorem 1.8] that clutters with the packing property have integral set covering polyhedrons. As a consequence, using Theorem \[noclu\], this conjecture reduces to the following:
\[con-cor-vila\]If $\alpha_0({\cal
C}')=\beta_1({\cal C}')$ for all minors $\cal C'$ of $\cal C$, then $R[It]$ is normal.
In this paper we will give some support for this conjecture using an algebraic approach.
\[march18-06-1\] Let $J_i$ be the ideal obtained from $I$ by making $x_i=1$. If $Q(A)$ is an integral polyhedron, then the ideal $I$ is normal if and only if $J_i$ is normal for all $i$ and ${\rm depth}(R/I^k)\geq 1$ for all $k\geq 1$.
$\Rightarrow$) The normality of an ideal is closed under minors [@normali Proposition 4.3], hence $J_i$ is normal for all $i$. Using Theorem \[burch\] and Corollary \[march19-06\] we get that ${\rm depth}(R/I^i)\geq 1$ for all $i$.
$\Leftarrow$) It follows readily by adapting the arguments given in the proof of the normality criterion [@normali Theorem 4.4].
By Proposition \[march18-06-1\] we obtain that Conjecture \[conforti-cornuejols\] also reduces to:
If $\alpha_0({\cal C}')=\beta_1({\cal C}')$ for any minor ${\cal C}'$ of $\cal C$, then $${\rm depth}(R/I^i)\geq 1\ \mbox{ for all }i\geq 1.$$
[*Notation*]{} For an integral matrix $B\neq(0)$, the greatest common divisor of all the nonzero $r\times r$ subdeterminants of $B$ will be denoted by $\Delta_r(B)$.
If $x^{v_1},\ldots,x^{v_q}$ are monomials of degree $d\geq 2$ such that ${\rm gr}_I(R)$ is reduced and the matrix $$B=\left(\hspace{-1mm}
\begin{array}{ccc}
v_1&\cdots&v_q\\
1&\cdots &1
\end{array}\hspace{-1mm}
\right)$$ has rank $r$, then $\Delta_r(B)=1$ and $B$ diagonalizes over $\mathbb{Z}$ to an identity matrix.
By Proposition \[nov1-03\] we obtain $A(P)=K[Ft]$. Hence a direct application of [@ehrhart Theorem 3.9] gives $\Delta_r(B)=1$.
This result suggest the following weaker conjecture of Villarreal:
If $\alpha_0({\cal C}')=\beta_1({\cal C}')$ for all minors $\cal C'$ of $\cal C$ and $x^{v_1},\dots,x^{v_q}$ have degree $d\geq 2$, then $\Delta_r(B)=1$, where $r={\rm rank}(B)$.
Let $G$ be a matroid on $X$ of rank $d$ and let $\cal B$ be the collection of bases of $G$. The set of all square-free monomials $x_{i_1}\cdots x_{i_d}\in R$ such that $\{x_{i_1},\ldots,x_{i_d}\}\in {\cal B}$ will be denoted by $F_G$ and the subsemigroup (of the multiplicative semigroup of monomials of $R$) generated by $F_G$ will be denoted by $\mathbb{M}_G$. The [*basis monomial ring*]{} of $G$ is the monomial subring $K[F_G]=K[\mathbb{M}_G]$. The ideal $I({\cal
B})=(F_G)$ is called the [*basis monomial ideal*]{} of $G$. An open problem in the area is whether the toric ideal of $K[F_G]$ is generated by quadrics, see [@white1 Conjecture 12]. This has been shown for graphic matroids [@blasiak].
The next result implies the normality of the basis monomial ring of $G$.
\[nwhite\] If $x^a$ is a monomial of degree $\ell{d}$ for some $\ell\in\mathbb{N}$ such that $(x^{a})^p\in\mathbb{M}_G$ for some $0\neq p\in\mathbb{N}$, then $x^a\in\mathbb{M}_G$.
If $I=I({\cal B})$ and $\cal B$ satisfies the packing property, then ${\rm gr}_I(R)$ is reduced.
First we show the equality $A(P)=K[F_Gt]$. It suffices to prove the inclusion $A(P)\subset K[F_Gt]$. Take $x^at^b\in A(P)$, i.e., $x^a\in\mathbb{Z}^n\cap bP$. Hence $x^a$ has degree $bd$ and $(x^{a})^p\in\mathbb{M}_G$ for some positive integer $p$. By Proposition \[nwhite\] we get $x^a\in\mathbb{M}_G$. It is seen that $x^at^b$ is in $K[F_Gt]$. Since $Q(A)$ is integral [@cornu-book], using Theorem \[oct31-03\] we get that $R[It]$ is normal. Thus both conditions yield that ${\rm gr}_I(R)$ is reduced according to Theorem \[noclu\].
This proof can be simplified using that the basis monomial ideal of a matroid is normal [@matrof].
Let $X_1,\ldots,X_d$ be a family of disjoint sets of variables and let $M$ be the transversal matroid whose collection of basis is $${\cal C}=\{\{y_1,\ldots,y_d\}\vert\, y_i\in X_i\, \forall\, i\}.$$ If $I=I({\cal C})$, then ${\rm gr}_I(R)$ is reduced.
The combinatorial equivalencies in the next result are well known [@cornu-book; @CGM]. Our contribution here is to link the reducedness of the associated graded ring with the integrality of $Q(A)$.
If $\cal C$ is a simple graph, then the following are equivalent[:]{}
[(a)]{} ${\rm gr}_I(R)$ is reduced.
[(b)]{} $\cal C$ is bipartite.
[(c)]{} $Q(A)$ is integral.
[(d)]{} $\cal C$ has the packing property.
By [@ITG Theorem 5.9] (a) and (b) are equivalent. Applying Theorem \[noclu\] and Proposition \[nov2-03\] we obtain that (a) is equivalent to (c). By Corollary \[oct21-03\] condition (a) implies (d). Finally using [@cornu-book Theorem 1.8] we obtain that (d) implies (c).
\[nov9-2-03\] If $\cal C$ is a bipartite graph and $J=I_c({\cal C})$, then ${\rm gr}_J(R)$ is reduced.
The matrix $A$ is totally unimodular [@Schr p. 273], hence $Q(A)$ is integral. By Proposition \[nov3-03\] we get $\overline{R[Jt]}=R_s(J)$. On the other hand $R[Jt]$ is normal by Proposition \[nov2-03\]. Thus by Theorem \[noclu\] the ring ${\rm
gr}_J(R)$ is reduced.
If $\cal C$ has the packing property and $I=I({\cal C})$, then $I^2=\overline{I^2}$.
By induction on $n$. Assume $\overline{I^2}\neq I^2$ and consider $M=\overline{I^2}/I^2$. If $\mathfrak{p}\neq\mathfrak{m}=(x_1,\ldots,x_n)$ is a prime ideal of $R$, then by induction $M_\mathfrak{p}=(0)$. Thus $\mathfrak{m}$ is the only associated prime of $M$ and there is an embedding $R/{\mathfrak m}\hookrightarrow M$, $\overline{1}\mapsto
\overline{x^{a}}$, where $\overline{x^{a}}\in\overline{I^2}\setminus
I^2$ and $x_ix^{{a}}\in I^2$ for all $i$. Notice that by induction all the entries ${a}_i$ of ${a}$ are positive. We consider two cases. Assume $a_i\geq 2$ for some $i$, say $i=1$. Given a monomial $x^\alpha$, the monomial obtained from $x^\alpha$ by making $x_1=1$ is denoted by $x^{\alpha'}$. Then making $x_1=1$ and using that $x_1x^a\in I^2$ gives $x^{a'}=x^{v_1'}x^{v_2'}x^\delta$, hence $x^a=x_1^{a_1}x^{a'}=x^{v_1}x^{v_2}x^\gamma\in I^2$, a contradiction. On the other hand if $a_i=1$ for all $i$, then $x^a=x_1\cdots x_n\in I^g\subset I^2$, where $g={\rm ht}(I)$, a contradiction. Therefore $I^2=\overline{I^2}$.
Recall that $I$ is said to be [*unmixed*]{} if all the minimal vertex covers of $\cal C$ have the same cardinality.
If $I$ is an unmixed ideal and $\cal C$ satisfies the König property, then $x^{\mathbf 1}=x_1x_2\cdots x_n$ belongs to the subring $K[x^{v_1},\ldots,x^{v_q}]$.
We may assume $x^{\mathbf 1}=x^{v_1}\cdots x^{v_g}x^\delta$, where $g$ is the height of $I$. If $\delta\neq 0$, pick $x_n\in{\rm supp}(x^\delta)$. Since the variable $x_n$ occurs in some monomial of $I$, there is a minimal prime $\mathfrak{p}$ containing $x_n$. Thus using that $x^{v_1},\ldots,x^{v_g}$ have disjoint supports we conclude that $\mathfrak{p}$ contains at least $g+1$ variables, a contradiction.
\[nov20-1-03\] Let $I_i=I\cap
K[X\setminus\{x_i\}]$. If $I$ is an unmixed ideal such that the following conditions hold
[($\mathrm{a}_1$)]{} $Q(A)$ is integral,
[($\mathrm{a}_2$)]{} $I_i$ is normal for $i=1,\ldots,n$, and
[($\mathrm{a}_3$)]{} $\cal C$ has the König property,
then $R[It]$ is normal.
Take $x^at^b=x_1^{a_1}\cdots x_n^{a_n}t^b\in\overline{R[It]}$ a minimal generator. By the second condition we may assume $a_i\geq 1$ for all $i$. Set $g={\rm ht}(I)$. Notice that $x_1\cdots
x_nt^g$ is in $\overline{R[It]}$ because $Q(A)$ is integral, this follows from Corollary \[1cover-1\] and Eq. (\[may6-06\]). We claim that $b\leq g$. If $b>g$, consider the decomposition $$x^at^b=(x_1\cdots x_nt^g)(x_1^{a_1-1}\cdots x_n^{a_n-1}t^{b-g}).$$ To derive a contradiction consider the irreducible representation of the Rees cone $\mathbb{R}_+{\cal A}'$ given in Eq. (\[okayama-oct\]). Observe that $$\textstyle\sum_{x_i\in C_k}a_i\geq b\ \ \ \ \ (k=1,\ldots,s)$$ because $(a,b)\in\mathbb{R}_+{\cal A}'$. Now since $I$ is unmixed we get $$\textstyle\sum_{x_i\in C_k}(a_i-1)\geq b-g\ \ \ \ \ (k=1,\ldots,s),$$ and consequently $x_1^{a_1-1}\cdots x_n^{a_n-1}t^{b-g}\in\overline{R[It]}$, a contradiction to the choice of $x^at^b$. Thus $b\leq g$. Using the third condition we get $x_1\cdots x_n\in I^g\subset I^b$, which readily implies $x^at^b\in R[It]$.
According to Corollary \[dec11-03\] condition ($\mathrm{a}_3$) is redundant when $I$ is generated by monomials of the same degree.
Let $Y\subset X$ and let $I_{Y}=I\cap K[Y]$. If $I_Y$ has the König property for all $Y$ and $\overline{R[It]}$ is generated as a $K$-algebra by monomials of the form $x^at^b$, with $x^{a}$ square-free, then $R[It]$ is normal.
Take $x^at^b$ a generator of $\overline{R[It]}$, with $x^a$ square-free. By induction we may assume $x^at^b=x_1\cdots x_nt^b$. Hence, since $(1,\ldots,1,b)$ is in $\mathbb{R}_+{\cal A}'$, we get that $|C_k|\geq b$ for $k=1,\ldots,s$. In particular $g={\rm ht}(I)\geq b$. As $I$ has the König property, we get $x_1\cdots x_n\in I^g$ and consequently $x^at^b\in R[It]$.
Let $I_i=I\cap K[X\setminus\{x_i\}]$. If $I_i$ is normal for $i=1,\ldots,n$ and $$\label{jan6-04}
C=H_{\ell_1}\cap H_{\ell_2}\cap\cdots\cap H_{\ell_r}\cap
\mathbb{R}_{+}^{n+1}\neq(0),$$ then $R[It]$ is normal.
Let $x^at^b=x_1^{a_1}\cdots x_n^{a_n}t^b\in\overline{R[It]}$ be a minimal generator, that is $(a,b)$, cannot be written as a sum of two non-zero integral vectors in $\mathbb{R}_+{\cal A}'$. It suffices to prove that $0\leq b\leq 1$ because this readily implies that $x^a$ is either a variable or a monomial in $F$. Assume $b\geq 2$. Since $I_i$ is normal we may assume that $a_i\geq 1$ for all $i$. As each variable occurs in at least one monomial of $F$, using that $C$ is contained in $\mathbb{R}_+{\cal A}'$ together with Eq. (\[jan6-04\]), it follows that there is $(v_k,1)$ such that $\langle (v_k,1),\ell_i\rangle=0$ for $i=1,\ldots,r$. Therefore $$\langle(a-v_k,b-1),\ell_i\rangle\geq 0\ \ \ \ \ (i=1,\ldots,r).$$ Thus $(a,b)-(v_k,1)\in \mathbb{R}_+{\cal A}'$, a contradiction to the choice of $x^at^b$.
Some applications to Rees algebras and clutters
===============================================
Throughout this section we assume $\deg(x^{v_i})=d\geq 2$ for all $i$. By assigning $\deg(x_i)=1$ and $\deg(t)=-(d-1)$, the Rees algebra $R[It]$ becomes a standard graded $K$-algebra, i.e., it is generated by elements of degree $1$. The $a$-invariant of $R[It]$, with respect to this grading, is denoted by $a(R[It])$. If $R[It]$ is a normal domain, then according to a formula of Danilov-Stanley [@BHer Theorem 6.3.5] its [ *canonical module*]{} is the ideal of $R[It]$ given by $$\omega_{R[It]}=(\{x_1^{a_1}\cdots x_n^{a_n}
t^{a_{n+1}}\vert\, a=(a_i)
\in(\mathbb{R}_+{\cal A}')^{\rm o}\cap\mathbb{Z}^{n+1}\}),$$ where $(\mathbb{R}_+{\cal A}')^{\rm o}$ is the topological interior of the Rees cone.
\[nov8-03\] If ${\rm gr}_I(R)$ is reduced, then $$a(R[It])\geq -\left[n-(d-1)(\alpha_0({\cal C})-1)\right],$$ with equality if $I$ is unmixed.
It is well known (see [@BHer]) that the $a$-invariant can be expressed as $$a(R[It])=-{\rm min}\{\, i\, \vert\,
(\omega_{R[It]})_i\neq 0\}.$$ Set $\alpha_0=\alpha_0({\cal C})$. Using Eq. (\[okayama-oct\]) it is seen that the vector $(1,\ldots,1,\alpha_0-1)$ is in the interior of the Rees cone. Thus the inequality follows by computing the degree of $x_1\cdots x_nt^{\alpha_0-1}$.
Assume that $I$ is unmixed. Take an arbitrary monomial $x^at^b=x_1^{a_1}\cdots x_n^{a_n}t^b$ in the ideal $\omega_{R[It]}$, that is, $(a,b)\in(\mathbb{R}_+{\cal A}')^{\rm o}$. By Proposition \[oct30-03\] the vector $(a,b)$ has positive entries and satisfies $$\begin{aligned}
&-b+\sum_{x_i\in C_k}a_i\geq 1 \ \ \ \ (k=1,\ldots,s).&\label{nov9-03}\end{aligned}$$ If $\alpha_0\geq b+1$, we obtain the inequality $$\label{jan9-04}
\deg(x^at^b)=a_1+\cdots+a_n-b(d-1)\geq n-(d-1)(\alpha_0-1).$$ Now assume $\alpha_0\leq b$. Using the normality of $R[It]$ and Eqs. (\[okayama-oct\]) and (\[nov9-03\]) it follows that the monomial $$m=x_1^{a_1-1}\cdots x_n^{a_n-1}t^{b-\alpha_0+1}$$ belongs to $R[It]$. Since $x^at^b=mx_1\cdots x_nt^{\alpha_0-1}$, the inequality (\[jan9-04\]) also holds in this case. Altogether we conclude the desired equality.
\[nov9-1-03\] If $I$ is unmixed with $\alpha_0({\cal C})=2$ and ${\rm gr}_I(R)$ is reduced, then $R[It]$ is a Gorenstein ring and $$a(R[It])=-(n-d+1).$$
From the proof of Theorem \[nov8-03\] it follows that $x_1\cdots x_nt$ generates the canonical module.
Notice that if $\alpha_0({\cal C})\geq 3$, then $R[It]$ is not Gorenstein because the monomials $x_1\cdots x_nt^{\alpha_0-1}$ and $x_1\cdots x_nt$ are distinct minimal generators of $\omega_{R[It]}$. This holds in a more general setting (see Proposition \[nov9-3-03\] below).
Let $J=I_c({\cal C})$ be the ideal of vertex covers of $\cal C$. If $\cal C$ is a bipartite graph and $I=I({\cal C})$ is unmixed, then $R[Jt]$ is a Gorenstein ring and $$a(R[Jt])=-(n-\alpha_0({\cal C})+1).$$
Notice that $R[Jt]$ has the grading induced by assigning $\deg(x_i)=1$ and $\deg(t)=1-\alpha_0({\cal C})$. Thus the formula follows from Corollary \[nov9-1-03\] once we recall that ${\rm gr}_J(R)$ is a reduced ring according to Corollary \[nov9-2-03\].
[([@Bd], [@Vas p. 142])]{}\[jan10-03\] If $S$ is a regular local ring and $J$ is an ideal of $S$ generated by a regular sequence $h_1,...,h_g$, then $S[Jt]$ is determinantal: $$S[Jt]\simeq S[z_1,\ldots,z_g]/I_2
\left(\hspace{-1mm}
\begin{array}{ccc}
z_1&\cdots&z_g\\
h_1&\cdots&h_g
\end{array}\hspace{-1mm}
\right)$$ and its canonical module is $\omega_S(1,t)^{g-2}$.
\[nov9-3-03\] If $I$ has height $g\geq 2$ and $S=R[It]$ is Gorenstein, then $g=2$.
Since $I_{\mathfrak{p}}$ is a complete intersection for all associated prime ideals $\mathfrak{p}$ of $I$ and $S$ is Gorenstein one has $\omega_S\simeq\omega_R(1,t)^{g-2}$ [@Her]. Then $$\label{may5-06}
S\simeq \omega_S\simeq\omega_R(1,t)^{g-2}=R\oplus Rt
\oplus\cdots\oplus Rt^{g-2}\oplus It^{g-1}\oplus\cdots$$ Take a minimal prime $\mathfrak{p}$ of $I$ of height $g$. Then $S_\mathfrak{p}=R_\mathfrak{p}[I_\mathfrak{p}t]$ is the Rees algebra of the ideal $I_\mathfrak{p}$, which is generated by a regular sequence. Thus localizing the extremes of Eq. (\[may5-06\]) at $\mathfrak{p}$ and using Lemma \[jan10-03\] we obtain $$S_\mathfrak{p}=R_\mathfrak{p}[I_\mathfrak{p}t]\simeq
\omega_{R_\mathfrak{p}}(1,t)^{g-2} \simeq \omega_{S_\mathfrak{p}}.$$ Note that it is important to know a priori that the canonical module of $S_\mathfrak{p}$ is $\omega_{R_\mathfrak{p}}(1,t)^{g-2}$. Hence $S_\mathfrak{p}$ is Gorenstein. To finish the proof note that the only Gorenstein determinantal rings that occur in Lemma \[jan10-03\] are those with $g=2$. Here the hypothesis on the degrees of $x^{v_i}$ is not needed.
\[nov19-03\] If $\overline{R[It]}=R_s(I)$, then there is a minimal vertex cover $C_k$ of $\cal C$ such that $|{\rm supp}(x^{v_i})\cap C_k|=1$ for $i=1,\ldots,q$.
We claim that $J_k=\mathfrak{p}_kR[It]$ for some $1\leq k\leq s$. If not, using Eq. (\[may2-06\]), we can pick $x^{v_k}t\in J_k$ for $k=1,\ldots,s$. Then by Proposition \[oct28-03\] the product of these monomials is in the radical of $IR[It]$. Therefore $$\left[(x^{v_1}t)\cdots(x^{v_s}t)\right]^p\in IR[It]$$ for some $0\neq p\in\mathbb{N}$. Thus $(x^{v_1}\cdots x^{v_s})^p\in I^{sp+1}$. By degree considerations, using that $\deg(x^{v_i})=d$ for all $i$, one readily derives a contradiction. This proves the claim. Hence $\langle (v_i,1),\ell_k\rangle=0$ for all $i$ and ${v_1},\ldots,{v_q}$ lie on the hyperplane $$\textstyle\sum_{x_i\in C_k}x_i=1.$$ Therefore $|{\rm supp}(x^{v_i})\cap C_k|=1$ for all $i$, as desired.
If $\overline{R[It]}=R_s(I)$ and $I$ is unmixed, then $$H_{\ell_1}\cap H_{\ell_2}\cap\cdots\cap H_{\ell_r}\cap
\mathbb{R}_{+}^{n+1}\neq(0).$$
Let $J=I_c({\cal C})$ be the Alexander dual of $I$. Using Proposition \[nov3-03\] one has $\overline{R[Jt]}=R_s(J)$. Thus by Lemma \[nov19-03\] there is $v_k$ such that $|{\rm supp}(x^{v_k})\cap C_i|=1$ for $i=1,\ldots,r$. This means that $(v_{k},1)$ is in the intersection of $H_{\ell_1},\ldots,H_{\ell _r}$.
\[nov20-03\] If $\overline{R[It]}=R_s(I)$, then there are $C_1,\ldots,C_d$ mutually disjoint minimal vertex covers of $\cal C$ such that $\cup_{i=1}^q{\rm supp}(x^{v_i})=\cup_{i=1}^d C_i$ and $$|{\rm supp}(x^{v_i})\cap C_k|=1\ \ \ \ \forall\ i,k.$$
By induction on $d$. By Lemma \[nov19-03\] there is a minimal vertex cover $C_1$ of $\cal C$ such that $|{\rm supp}(x^{v_i})\cap C_1|=1$ for all $i$. Consider the ideal $I'$ obtained from $I$ by making $x_i=1$ for all $x_i\in C_1$. Then $I'$ is an ideal generated by monomials of degree $d-1$ and $\overline{R[I't]}=R_s(I')$ by Corollary \[nov19-1-03\]. Thus we can apply induction to get the required assertion.
\[dec11-03\] If $I$ is unmixed and $\overline{R[It]}=R_s(I)$, then both $\cal C$ and the clutter $\cal D$ of minimal vertex covers of $\cal C$ have the König property.
That $\cal D$ has the König property follows from Proposition \[nov20-03\], because $\alpha_0({\cal D})=d$ and $C_1,\ldots,C_d$ are independent edges of $\cal D$. Now $I_c({\cal
C})$ is unmixed, is generated by monomials of degree $\alpha_0({\cal C})$, and according to Proposition \[nov3-03\] one has $\overline{R[I_c({\cal C})]}=R_s(I_c({\cal C}))$. Thus, using again Proposition \[nov20-03\], we conclude that $\cal C$ has the König property.
Combining Corollary \[dec11-03\] with Proposition \[nov20-1-03\] we obtain:
\[may4-06\] Let $I_i=I\cap K[X\setminus\{x_i\}]$. If $I$ is unmixed and $Q(A)$ is integral, then ${\rm gr}_I(R)$ is reduced if and only if $I_i$ is normal for $i=1,\ldots,n$.
[10]{}
D. Bertsimas and J. N. Tsitsiklis, [*Introduction to Linear Optimization*]{}, Athena Scientific, Massachusetts, 1997.
J. Blasiak, The toric ideal of a graphic matroid is generated by quadrics, Combinatorica [**28**]{} (2008), no. 3, 283–297.
M. Brodmann, The asymptotic nature of the analytic spread, Math. Proc. Cambridge Philos. Soc. [**86**]{} (1979), 35–39.
W. Bruns and G. Restuccia, Canonical modules of Rees algebras, J. Pure Appl. Algebra [**201**]{} (2005), 189–203.
C. Escobar, R. H. Villarreal and Y. Yoshino, Torsion freeness and normality of blowup rings of monomial ideals, [*Commutative Algebra*]{}, Lect. Notes Pure Appl. Math. [**244**]{}, Chapman & Hall/CRC, Boca Raton, FL, 2006, pp. 69-84.
S. Faridi, The facet ideal of a simplicial complex, Manuscripta Math. [**109**]{} (2002), 159-174.
S. Faridi, Cohen-Macaulay properties of square-free monomial ideals, J. Combin. Theory Ser. A [**109(2)**]{} (2005), 299-329.
I. Gitler, E. Reyes and R. H. Villarreal, Blowup algebras of ideals of vertex covers of bipartite graphs, Contemp. Math. [**376**]{} (2005), 273–279.
I. Gitler, C. Valencia and R. H. Villarreal, A note on Rees algebras and the MFMC property, Beiträge Algebra Geom. [**48**]{} (2007), No. 1, 141-150.
J. Herzog, Alexander duality in commutative algebra and combinatorics, Algebra Colloq. [**11**]{} (2004), no. 1, 21–30.
J. Herzog and T. Hibi, The depth of powers of an ideal, J, Algebra [**291**]{} (2005), 534–550.
J. Herzog, T. Hibi and N. V. Trung, Symbolic powers of monomial ideals and vertex cover algebras, Adv. Math. [**210**]{} (2007), 304–322.
B. Korte and J. Vygen, [*Combinatorial Optimization Theory and Algorithms*]{}, Springer-Verlag, 2000.
C. Valencia and R. H. Villarreal, Canonical modules of certain edge subrings, European J. Combin. [**24(5)**]{} (2003), 471–487.
W. V. Vasconcelos, [*Integral Closure*]{}, Springer Monographs in Mathematics, Springer, New York, 2005.
R. H. Villarreal, Rees cones and monomial rings of matroids, Linear Algebra Appl. [**428**]{} (2008), 2933-2940.
N. L. White, The basis monomial ring of a matroid, Adv. in Math. [**24**]{} (1977), 292–297.
N. L. White, A unique exchange property for bases, Linear Algebra Appl. [**31**]{} (1980), 81–91.
X. Zheng, Resolutions of facet ideals, Comm. Algebra [**32(6)**]{} (2004), 2301–2324.
[^1]: Partially supported by COFAA-IPN.
[^2]: Partially supported by CONACyT grant 49251-F and SNI.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Dmitry Fuchs[^1] and Serge Tabachnikov[^2]\'
title: 'Self-dual polygons and self-dual curves'
---
Introduction {#intro}
============
The projective plane $P$ is the projectivization of 3-dimensional space $V$ (we consider the cases of two ground fields, ${{\mathbb R}}$ and ${{\mathbb C}}$); the points and the lines of $P$ are 1- and 2-dimensional subspaces of $V$. The dual projective plane $P^\ast$ is the projectivization of the dual space $V^\ast$. Assign to a subspace in $V$ its annihilator in $V^\ast$. This gives a correspondence between the points in $P$ and the lines in $P^\ast$, and between the lines in $P$ and the points in $P^\ast$, called the projective duality. Projective duality preserves the incidence relation: if a point $A$ belongs to a line $B$ in $P$ then the dual point $B^\ast$ belongs to the dual line $A^\ast$ in $P^\ast$. Projective duality is an involution: $(A^\ast)^\ast=A$.
Projective duality extends to polygons in $P$. An $n$-gon is a cyclically ordered collection of $n$ points and $n$ lines satisfying the incidences: two consecutive vertices lie on the respective side, and two consecutive sides pass through the respective vertex. We assume that our polygons are non-degenerate: no three consecutive vertices are collinear. Thus to every polygon $L\subset P$ there corresponds the dual polygon $L^\ast \subset P^\ast$. A polygon $L$ is called self-dual if there exists a projective map $P \to P^\ast$ that takes $L$ to $L^\ast$.
Projective duality also extends to locally convex smooth curves. A smooth curve $\gamma\subset P$ determines a one-parameter family of its tangent lines, and projective duality takes it to a one-parameter family of points in $P^\ast$, the dual curve $\gamma^\ast \subset P^\ast$. If $\gamma$ is locally convex then $\gamma^\ast$ is smooth as well. One has $(\gamma^\ast)^\ast=\gamma$. Projective duality further extends to a broader class of curves with inflections and cusps, called [*wave fronts*]{} (see end of Section \[polycurves\] for a precise definition). Projective duality interchanges inflections and cusps. One defines self-dual curves similarly to self-dual polygons.
A motivation for this work is the following problem (No 1994-17 in [@Arn]) of V. Arnold:
> Find all projective curves equivalent to their duals. [*The answer seems to be unknown even in ${{\mathbb{RP}}}^2$*]{}.
(A traditional interpretation of this question would be to consider algebraic curves, in which case the Plucker formulas play a critical role; see [@Hol]. In particular, the Plucker formulas imply that a [*non-singular*]{} self-dual algebraic curve is a conic; however, there exist other, singular, self-dual curves, for example, $y=x^3$, projectively equivalent to its dual $y=x^{3/2}$.)
The main result of this paper is a description of self-dual polygons in ${{\mathbb{CP}}}^2$. Let $A_1,A_3,\dots,A_{2n-1}\in P$ (where the indices are odd residues modulo $2n$) be the vertices of an $n$-gon, and let $B_2, B_4,
\dots,B_{2n}$ (where the indices are even residues modulo $2n$) be its respective sides: $B_{2i}=A_{2i-1}A_{2i+1}$ for all $i$. Let $m$ be an odd number, $1\le m\le n$. The $n$-gon $L$ is called [*$m$-self-dual*]{} if there exists a projective map $g\colon P \to P^\ast$ such that $g(A_i)=B_{i+m}^\ast$ for all $i$. An example of an $m$-self-dual $n$-gon, for arbitrary $m$, is a regular $n$-gon. Denote by ${\mathcal M}_{m,n}$ the moduli space of $m$-self-dual $n$-gons. Our result is as follows.
\[main\] If $(m,n)=1$ then ${\mathcal M}_{m,n}$ consists of one point, the class of a regular $n$-gon. If $m<n,\ (m,n)>1$ and $n\neq 2m$ then dim ${\mathcal M}_{m,n}=(m,n)-1$. Finally, dim ${\mathcal M}_{m,2m}=m-3$ and dim ${\mathcal M}_{n,n}=n-3$.
Note, for comparison, that the dimension of the moduli space of $n$-gons is $2n-8$. The proof of Theorem \[main\] occupies Sections \[m=n\] and \[m<n\]. These sections also contain explicit constructions of self-dual polygons.
The map $g\colon P \to P^\ast$, associated with an $m$-self-dual $n$-gon, determines a linear map $V\to V^\ast$, defined up to a factor, and therefore a bilinear form $F$ on $V$. We prove that if the polygon is not a multiple of another polygon then $F$ is symmetric if and only if $m=n$, see Proposition \[symmetric\]. We also show that if an $n$-self-dual $n$-gon is convex then the symmetric bilinear form $F$ is definite, Proposition \[convex\].
We make additional observations. First, every pentagon is 5-self-dual, see Proposition \[pentagons\] below (five is the first interesting number because all triangles are projectively equivalent, and so are all quadrilaterals). Secondly, if an $n$-gon with odd $n$ is inscribed in a conic and circumscribed about a conic then it is $n$-self-dual, see Proposition \[poncelet\]. However the moduli space of such “Poncelet" polygons has dimension two, which is less than $n-3$ for $n\geq 7$.
In the real case, one can also interpret a polygon as a closed polygonal curve. In Section \[polycurves\], we define a polygonal curve as a polygon in ${{\mathbb{RP}}}^2$ with two additional structures: every two consecutive vertices $A_{2i-1},A_{2i+1}$ partition the real projective line $B_{2i}$ into two segments, and one of these segments is chosen (as a side); every two consecutive sides $B_{2i}$ and $B_{2i+2}$ determine two pairs of vertical angles at the vertex $A_{2i+1}$, and one of these pairs is chosen (as an exterior angle). Polar duality naturally extends to these polygonal curves, and one can consider self-dual polygonal curves. A given $n$-gon $L$ gives rise to $2^{2n}$ polygonal curves. We prove (Proposition \[choices\]) that if $L$ is $m$-self-dual then, out of these $2^{2n}$ polygonal curves, $2^{(m,n)}$ are $m$-self-dual.
Section \[sdcurves\] concerns self-dual curves and wave fronts in the real projective plane $P$. We do not attempt to give a complete classification of such curves. A curve $\gamma(t) \subset P,\ t\in S^1={{\mathbb R}}/2\pi{{\mathbb Z}}$, is called self-dual if there exists a projective transformation $g\colon P\to P^\ast$ and a diffeomorphism $\varphi$ of $S^1$ such that $g(\gamma(\varphi (t))=\gamma^\ast(t)$. The diffeomorphism $\varphi$ is a continuous analog of the cyclic shift by $m$ in the definition of $m$-self-dual $n$-gons.
A number of results that we establish for polygons have analogs for curves. For example, the bilinear form $F$ is symmetric if and only if $\varphi^2=\id$, see Proposition \[curvesymm\]. If, in addition, a self-dual curve is convex then $F$ is definite, Proposition \[curvepos\].
We observe that curves of constant width $\displaystyle \frac\pi2$ on the unit sphere $S^2$ project to self-dual curves with $\varphi^2=\id$ in ${{\mathbb{RP}}}^2$. We construct such curves of constant width as Legendrian curves in the manifold of contact elements of $S^2$ satisfying certain monodromy conditions. We also give a similar description to self-dual curves with the diffeomorphism $\varphi$ having higher order than 2. This description leads to explicit formulas for self-dual curves.
Finally, we briefly discuss the Radon curves, the unit circles in two-dimensional normed spaces for which the orthogonality relation is symmetric. Radon curves have been extensively studied; they provide examples of projectively self-dual curves.
Let us finish this introduction with a question: can a smooth convex self-dual curve, other than a conic, be an oval of an algebraic curve?
#### Acknowledgments.
Many thanks to J. C. Alvarez, V. Ovsienko and R. Schwartz for stimulating discussions. It is a pleasure to acknowledge the hospitality of the Research in Pairs program at the Mathematical Institute at Oberwolfach (MFO) where this work was done. The second author was partially supported by an NSF grant DMS-0555803.
Polygons and duality {#poldual}
====================
We use the notation from Section \[intro\]. Let $P={{{\mathbb C}}}P^2$ and $V={{{\mathbb C}}}^3$. Let $L$ be an $n$-gon in $P$ whose vertices are $A_1,A_3,\dots,A_{2n-1}$ and whose sides are $B_2,B_4,
\dots,B_{2n}$. The dual $n$-gon $L^\ast$ in the dual projective plane $P^\ast$ has the vertices $B_2^\ast,B_4^\ast, \dots,B_{2n}^\ast$ and the sides $A_1^\ast,A_3^\ast,\dots,A_{2n-1}^\ast$.
Assume that $L$ is $m$-self-dual where $m$ be an odd number, $1\le m\le n$. Then there exists a linear isomorphism $f\colon V\to V^\ast$ that takes the line $A_i \subset V$ to the line $B_{i+m}^\ast \subset V^\ast$ for all $i$. Along with $f$, we shall consider the corresponding projective isomorphism $\hat f\colon P\to P^\ast$ and the bilinear form $F$ on $V$, $F(v,w)=\langle f(v),w\rangle$. Obviously, for a given $m$-self-dual polygon, $\hat f$ is unique, while $f$ and $F$ are unique up to a non-zero constant factor.
Along with an $n$-gon $L=A_1A_3,\dots,A_{2n-1}$, we can consider the $kn$-gon $kL=A_1A_3\dots
A_{2kn-1}$ with $A_i=A_{i+2n}$. Obviously, $(kL)^\ast=kL^\ast$, and if $L$ is $m$-self-dual, then $kL$ is $(m+2rn)$-self-dual for $r=0,1,\dots,k-1$. A polygon $L$ is called [*simple*]{}, if $L\ne kL'$ for any $k>1$ and $L'$.
\[symmetric\] Let $L$ be a simple $m$-self-dual $n$-gon and $f\colon V\to V^\ast$ be the corresponding isomorphism. Then $f$ is self-adjoint $($or, equivalently, $F$ is symmetric$)$ if and only if $m=n$.
#### *Proof.*
Notice that $$\begin{array} {rl} F(A_i,A_j)=0&\Leftrightarrow \langle
F(A_i),A_j\rangle=0\Leftrightarrow\langle B^\ast_{i+m},A_j\rangle=0\\ &\Leftrightarrow A_j\in
B_{i+m}=A_{i+m-1}A_{i+m+1}.\end{array}$$ In particular, $F(A_i,A_{i+m\pm1})=0$ for all $i$.
Let $F$ be symmetric. Then, for all $i$, $F(A_{i+m\pm 1},A_i)=0$, and hence $$\begin{array} {rl} A_i\in&A_{(i+m-1)+(m-1)}A_{(i+m-1)+(m+1)}=A_{i+2m-2}A_{i+2m},\\ A_i\in&A_{(i+m+1)+(m-1)}A_{(i+m+1)+(m+1)}=A_{i+2m}A_{i+2m+2}.\end{array}$$Since the points $A_{i+2m-2}, A_{i+2m}, A_{i+2m+2}$ are not collinear, this means that $A_i=A_{i+2m}$. Hence $m=n$ (the polygon $L$ is simple!).
Let $m=n$. For every $i$, $F(A_i,A_{i+m\pm1})=0$, and, in addition to that, $$\begin{array} {rl} F(A_{i+m-1},A_i)&=F(A_{i+m-1},A_{i+2m})=
F(A_{i+m-1},A_{(i+m-1)+(m+1)})=0,\\ F(A_{i+m+1},A_i)&=
F(A_{i+m+1},A_{i+2m})=F(A_{i+m+1},A_{(i+m+1)+(m-1)})=0.\end{array}$$ This implies that the linear forms $F(A_i,-)$ and $F(-,A_i)$ are proportional for every $i$, that is, there exist non-zero complex numbers $\lambda_i$ such that $ F(A_i,x)=\lambda_i F(x,A_i)$ for all $i$ and $x$. Hence $ F(A_i,A_j)=\lambda_i F(A_j,A_i)=\lambda_i\lambda_j F(A_i,A_j)$, so if $ F(A_i,A_j)\ne0$, then $\lambda_i=\lambda_j^{-1}$. But $ F(A_i,A_{i+m-3})\ne0$ (because $A_{i+m-3}$ does not belong to the line $A_{i+m-1}A_{i+m+1}$, which is the zero locus of the form $ F(A_i,x)$). Hence $$\lambda_i=\lambda_{i+m-3}^{-1}=\lambda_{i+2(m-3)}=\lambda_{i+3(m-3)}^{-1}=\dots=\lambda_{i+m(m-3)}^{-1}=
\lambda_i^{-1},$$so $\lambda_i=\pm1$. We state that all $\lambda_i$’s are the same. Indeed, if $\lambda_i=1,\lambda_j=-1$ for some $i,j$, then for every $k$, one of $ F(A_i,A_k),
F(A_j,A_k)$ must be 0, that is, $A_k$ belongs to one of two lines, $A_{i+m-1}A_{i+m+1}$ and $A_{j+m-1}A_{j+m+1}$, that is, all vertices of the polygon belong to two lines, which is impossible, since the three lines $A_1A_2,A_2A_3,A_3A_4$ are all different. We see that our form $ F$ is either symmetric or skew symmetric. However it cannot be skew-symmetric because it is non-degenerate, and all skew-symmetric forms in an odd-dimensional space are degenerate. [$\Box$]{}
\[coord\] [The class of projective equivalence of an $n$-gon $L$ is determined by a collection of $2n$ numbers $(p_1,q_1,p_3,q_3\dots,p_{2n-1},q_{2n-1})$ (the indices are odd residues modulo $2n$). The definition of these numbers refers to Figure \[cross\]: $$p_{2i+1}=[A_{2i-3},A_{2i-1},P,R], \ \ q_{2i+1}=[R,Q,A_{2i+3},A_{2i+5}]$$ where $P=B_{2i-2}\cap B_{2i+2}, R=B_{2i-2}\cap B_{2i+4}, Q=B_{2i}\cap B_{2i+4}$ and $[\ ,\ ,\ ,\ ]$ denotes the cross-ratio of four points on a projective line, see [@Sch1; @Sch2]. These $2n$ numbers are not independent: they satisfy 8 relations ensuring that the polygon is closed. Similarly, the dual polygon $L^\ast$ is characterized by the respective cross-ratios $(p^\ast_2,q^\ast_2,\dots,p^\ast_{2n},q^\ast_{2n})$ (the indices are even residues modulo $2n$). It is easy to see that $p^\ast_{2i}=q_{2i-1}$ and $q^\ast_{2i}=p_{2i+1}$. An $n$-gon $L$ is $m$-self-dual if and only if $p_i=p^\ast_{i+m}, q_i=q^\ast_{i+m}$, and hence iff $p_i=q_{i+m-1}, q_i=p_{i+m+1}$. This implies $2m$-periodicity of the sequence of cross-ratios: $p_i=p_{i+2m}, q_i=q_{i+2m}$, cf. Section \[m<n\]. ]{}
![Cross-ratios at the vertices of a polygon[]{data-label="cross"}](figeight.eps){width="2in"}
The case $m=n$ {#m=n}
==============
In this section, we consider $n$-gons with odd $n$ and with every vertex dual to the opposite side of a projectively equivalent $n$-gon. According to Proposition \[symmetric\], the bilinear form $ F$ is symmetric in this case, so it determines a (complex) Euclidean structure in space $V$. In an appropriate coordinate system, the projective duality becomes the [*polar duality*]{} $$(a,b,c)\mapsto \{ax+by+cz=0\}.$$ (Geometrically, this means that we apply to a point of the standard Euclidean plane the inversion in the unit circle centered at 0, then reflect the point in 0, and then take a line through the obtained point perpendicular to the position vector of this point.) For this duality, we will use the notations $A\mapsto A^\perp\mapsto(A^\perp)^\perp=A$. Two polygons, $n$-self-dual with respect to this duality, are projectively equivalent if and only if they are $O(3,{{\mathbb C}})$-equivalent.
For a polygon $A_1A_3\dots A_{2n-1}$, construct the star-like polygon $C_1C_2\dots C_n$ where $C_i=A_{1+(i-1)(n-1)}$.
\[star\] The polygon $A_1A_3\dots A_{2n-1}$ is $n$-self-dual $($with respect to the polar duality$)$ if and only if $C_{i+1}\in C_i^\perp$ for all $i$ $($with $i$ a residue modulo $n)$.
#### *Proof.*
Obvious. [$\Box$]{}
This leads to a simple explicit construction of all $n$-self-dual $n$-gons. Fix a point $C_1$, then a point $C_2\in C_1^\perp$ not equal to $C_1$ (the latter is relevant only if $C_1\in C_1^\perp$). Notice that, modulo the action of $O(3,{{\mathbb C}})$, there are four choices of the pair $C_1,C_2$ (depending on possible incidences $C_1\in C_1^\perp,C_2\in C_2^\perp$). Then choose $C_i\in C_{i-1}^\perp,C_i\ne
C_{i-2},C_{i-1}$ for $i=2,\dots,n-1$, with the additional requirement $C_{n-1}\ne C_1$. In conclusion, we put $C_n=C_{n-1}^\perp\cap C_1^\perp$. After this, we redenote the points, $C_i=A_{1+(i-1)(n-1)}$, and get an $n$-self-dual $n$-gon $A_1A_3\dots A_{2n-1}$. Moreover, up to a projective equivalence preserving the numeration of vertices, this construction gives all $n$-self-dual $n$-gons, one time each.
In particular, the moduli space of $n$-self-dual $n$-gons has dimension $n-3$ (each of the points $C_3,\dots C_{n-1}$ is arbitrarily chosen within a line with finitely many punctures).
#### Pentagons.
For an arbitrary $n\ge4$, the moduli space of all $n$-gons has dimension $2n-8$ ($2n$ for $n$ vertices, $-8$ for the action of the group $PSL(3,{{\mathbb C}})$). In general, this exceeds the dimension $n-3$ of $n$-self-dual $n$-gons, but for $n=5$ the two numbers coincide: both equal 2. Moreover, the following holds.
\[pentagons\] Every pentagon is 5-self-dual.
![Pentagons are self-dual[]{data-label="pentagon"}](figone.eps){width="1.8in"}
#### *Proof.*
We will prove this for a “generic" pentagon with no three vertices collinear; the general case can be resolved by a transition to limit. For a pentagon $ABCDE$ there are 5 cross-ratios $\rho_A,\dots,\rho_E$: $\rho_A$ is defined as the cross-ratio of the lines $AB,AC,AD,AE$, and the other four are defined in a similar way. These cross-ratios projectively determine a pentagon: the points $A,B,C,D$ can be moved to chosen locations, and after that the lines $BE$ and $CE$ are determined by $\rho_B$ and $\rho_C$. (Certainly, the five cross-ratios are not independent: generally, two of them determine the rest.) For the dual pentagon, these cross-ratos are $\rho_{AB},\dots,\rho_{AE}$ where $\rho_{AB}$ is the cross ratio of the points $DE\cap AB,A,B,DC\cap AB$ on the line $AB$, and the other four are defined in a similar way. Figure \[pentagon\] shows that $\rho_E=\rho_{BC}$, and four similar equalities hold as well. [$\Box$]{}
#### Poncelet polygons.
We begin with the following easy statement.
\[conic\] Let $C\subset P$ be a non-degenerate conic, and let $L=E_1\dots E_n$ be an $n$-gon inscribed in $C$. Then the $n$-gon whose sides are tangent to $C$ at points $E_1,\dots,E_n$ is $($projectively equivalent to the$)$ dual to $L$. More precisely, there exists a projective isomorphism $P\to P^\ast$ that takes $E_i$ to the tangent line to $C$ at $E_i$.
#### *Proof.*
Since all non-degenerate conics are projectively equivalent, we may assume that $C$ is a unit circle in the Euclidean plane. Let $E'_i$ be the point of $C$ opposite to $E_i$. Then the polygon $L$ is projectively equivalent to the polygon $L'=E'_1\dots E'_n$ and the tangent to $C$ at $E'_i$ is polar dual to $E_i$. [$\Box$]{}
![Poncelet pentagons[]{data-label="Poncelet"}](Poncelet.eps){width="3in"}
An $n$-gon with an odd $n$ is called a [*Poncelet polygon*]{} if it is both inscribed into a non-degenerate conic and circumscribed about a non-degenerate conic (see Figure \[Poncelet\]).
\[poncelet\] Every Poncelet $n$-gon is $n$-self-dual.
#### *Proof.*
This follows from Lemma \[conic\] and the following known result ([@Sch; @L-T]). Let $A_1A_3\dots A_{2n-1}$ be an $n$-gon (with an odd $n$) inscribed into a conic $C$ and circumscribed about a conic $C'$. Then there exists a projective involution $h\colon P\to P$ such that $h(A_i)$ is the tangency point of $A_{i+n-1}A_{i+n+1}$ and $C'$. [$\Box$]{}
Two more remarks. The first is that any non-degenerate pentagon is a Poncelet polygon, so Proposition \[pentagons\] follows from Proposition \[poncelet\]. The second is the following proposition.
\[ponceletmod\] For every odd $n\ge5$, the projective moduli space of Poncelet $n$-gons is two-dimensional.
#### *Proof.*
The conics $C,C'$ from the definition of Poncelet polygons determine a one-parameter family $\mathcal F$ of conics that have four common tangents. Generically, there exists a unique, up to a projective equivalence, such family $\mathcal F$ (to specify this family, it suffices to fix a generic quadruple of lines). For every $C\in\mathcal F$, there exists a finite number of $C'\in\mathcal F$ such that some $n$-gon inscribed in $C$ is circumsribed about $C'$ (see [@G-H] for an explicit condition due to Cayley). Moreover, for such a pair $C,C'$, every point of $C$ is a vertex of such an $n$-gon (Poncelet’s theorem, see [@B-K-O-R]). Thus, a projective class of a Poncelet $n$-gon is determined by two independent choices: the choice of a $C\in\mathcal F$ and the choice of a point in $C$. [$\Box$]{}
Thus, for an odd $n>5$, Poncelet $n$-gons form a small fraction of the space of $n$-self-dual $n$-gons.
#### The real Euclidean case.
Let $L$ be a real $n$-self-dual $n$-gon. Then $F$ is a real symmetric bilinear form (determined up to real non-zero factors), and there are two possibilities: the form $F$ may be definite or indefinite. In the definite (Euclidean) case, the construction of a self-dual polygon given in the beginning of this section looks especially simple. Consider the unit sphere $S\subset{{\mathbb R}}^3$. Choose an arbitrary point $C_1\in S$. Then choose a point $C_2$ at the distance $\displaystyle
\frac\pi2$ from $C_1$. Then choose a point $C_3$ at the distance $\displaystyle\frac\pi2$ from $C_2$, not equal to $\pm C_1$, then choose $C_4,C_5,\dots$. The last choice will be slightly different from the preceding ones: we choose the point $C_{n-1}$ at the distance $\displaystyle\frac\pi2$ from $C_{n-2}$, not equal to $\pm C_{n-3}$, and also not equal to $\pm C_1$. After this, we denote by $C_n$ a point at the distance $\displaystyle\frac\pi2$ from each of the points $C_{n-1}$ and $C_1$. (There are two such points, they form the intersection of two different great circles.) Then we put $C_i=A_{1+(i-1)(n-1)}$ and project the polygon $A_1A_3\dots A_{2n-1}$ onto $P$. This is our self-dual polygon (see Figure \[euclidean\]).
![Self-dual polygon on the unit sphere[]{data-label="euclidean"}](figthree.eps){width="5in"}
It is natural to ask, which $n$-self-dual $n$-gons correspond to definite forms. A partial answer to this question is provided by the following proposition.
\[convex\] If a real $n$-self-dual $n$-gon is projectively equivalent to an affine convex $n$-gon, then the corresponding symmetric bilinear form is definite.
#### *Proof.*
Let $A_1A_3\dots A_{2n-1}$ be our convex polygon. We will use only the convexity of the heptagon $A_1A_3A_{n-2}A_nA_{n+2}A_{n+4}A_{2n-1}$. We will assume that the points $A_{n-2},A_n,A_{n+2},A_{n+4}$ have projective coordinates $(0:1:1),(0:1:0),(1:0:0),(1:0:1)$. Then the lines $B_{n-1}=A_{n-2}A_n,\ B_{n+1}=A_nA_{n+2},\ B_{n+3}=A_{n+2}A_{n+4}$ are, respectively, the $y$-axis, the line at infinity, and the $x$-axis (see Figure \[conv7\]). Therefore $B^\ast_{n-1}=(1,0,0),B^\ast_{n+1}=(0,0,1), B^\ast_{n+3}=(0,1,0)$. The matrix of the isomorphism $G=F^{-1}\colon P^\ast\to P$ which takes $B^\ast_j$ into $A_{j+n}$ is symmetric by Proposition \[symmetric\]. Let it be $$G=\left[\begin{array} {ccc} a&b&c\\ b&d&e\\ c&e&f\end{array}\right].$$ Then $A_{2n-1}=
G(B^\ast_{n-1})=(a,b,c), A_1=G(B^\ast_{n+1})=(c,e,f), A_3=G(B^\ast_{n+3})=(b,d,e)$, and the affine coordinates of the points $A_{2n-1},A_1,A_3$ are$$x_{2n-1}=\frac ac , y_{2n-1}=\frac bc,\ x_1=\frac cf,\
y_1=\frac ef,\ x_3=\frac be,\ y_3=\frac de.$$
![Convex septagon: proof of Proposition \[convex\][]{data-label="conv7"}](figten.eps){width="2in"}
The conditions of convexity of our heptagon are $0<x_3<x_1<x_{2n-1}<1,$ $1>y_3>y_1>y_{2n-1}>0$ and $$\det\left[\begin{array} {cc}
x_1-x_{2n-1}&y_1-y_{2n-1}\\ x_1-x_3&y_1-y_3\end{array}\right]>0;$$the latter means$$x_1y_{2n-1}+x_3y_1+x_{2n-1}y_3-x_1y_3-x_3y_{2n-1}-x_{2n-1}y_1>0,$$ or$$2\cdot\frac bf+\frac ac\cdot\frac de-\left(\frac cf\cdot\frac de+\frac be\cdot bc+\frac ac\cdot\frac ef\right)>0.$$After multiplication by the positive number $c^2f^2x_{2n-1}y_1=acef$, this inequality becomes$$2acbe+a^2df-(ac^2d+ab^2f+
a^2e^2)>0,$$that is, $a\cdot\det G>0$. Also it follows from the convexity inequalities that $ad-b^2=f^2x_1y_1(x_{2n-1}y_3-x_3y_{2n-1})>0$. These two inequalities show that the form $F$ is definite. [$\Box$]{}
It should be noted that, as it is seen from the construction above (with a sequence of points on the sphere), a self-dual polygon with a definite form does not need to be convex. Still, it is true that a pentagon has a definite form if and only if it is projectively equivalent to a convex pentagon (we leave a proof to the reader).
The case $m<n$ {#m<n}
==============
Let $L=A_1A_3\dots A_{2n-1}$ be an $m$-self-dual polygon with $m<n$. By Proposition \[symmetric\], the form $F$ in this case is not symmetric. If $B$ is a point or a line in $P$, then $B^\perp$ is defined as $\{y\in P\mid F(y,x)=0\ \mbox{for all}\ x\in B\}$. There arises a projective transformation $G\colon P\to P,\ G(B)=(B^\perp)^\perp$, and obviously, in term of matrices, $G=F^{-1}F^t$.
\[shift\] If $L$ is $m$-self-dual then $G(A_i)=A_{i+2m}$.
#### *Proof.*
One has $A_i^\perp=A_{i+m-1}A_{i+m+1}$, and hence $G(A_i)=(A_i^\perp)^\perp=A_{i+m-1}^\perp\cap A_{i+m+1}^\perp=(A_{i+2m-2}A_{i+2m})\cap(A_{i+2m}A_{i+2m+2})=A_{i+2m}.$ [$\Box$]{}
Thus, $G$ makes a non-trivial cyclic permutation of vertices of $L$, and, in particular, $G^r=\id$ where $r=\displaystyle\frac n{(m,n)}$. The following result is an elementary fact from linear algebra.
\[canonical\] Let $F$ be a non-degenerate non-symmetic bilinear form in $V$. Then there exists a basis in $V$ with respect to which $F$ has one of the following matrices:$$H_\varphi=\left[\begin{array} {ccc} \phantom{-}\cos\varphi&\sin\varphi&0\\ -\sin\varphi&\cos\varphi&0\\ \phantom{-} 0&0&1\end{array}\right],\ J=\left[\begin{array} {ccc} \phantom{-} 1&1&0\\ -1&0&0\\ \phantom{-} 0&0&1\end{array}\right],\ K=\left[\begin{array} {ccc} \phantom{-} 1&1&0\\ -1&0&1\\ \phantom{-} 0&1&0\end{array}\right].$$
#### *Proof.*
There is a unique decomposition $F=F_++F_-$ where $F_+$ is symmetric and $F_-$ is skew-symmetric. Let $W=\Ker F_-$; since $F_-\ne0$, $\dim W=1$.
[*Case*]{} 1: $F_+|_W\ne 0,\rank F_+=3$. Let $e_3\in W,\ F_+(e_3,e_3)=1$. Let $Z$ be the orthogonal complement to $W$ with respect to $F_+$. Choose $e'_1,e'_2\in Z$ with $F_+(e'_i,e'_j)=\delta_{ij}$. Then choose a rescaling $e_1=\alpha e'_1, e_2=\alpha e'_2$ such that $F_+(e_1,e_1)^2+
F_-(e_1,e_2)^2=1$. (Notice that $F_+(e'_1,e'_1)^2+F_-(e'_1,e'_2)^2=\det F\ne0$.) Then the matrix of $F$ with respect to the basis $e_1,e_2,e_3$ is $H_\varphi$ with some (complex) $\varphi\ne\displaystyle\frac{k\pi}2$.
[*Case*]{} 2: $F_+|_W\ne 0, \rank F_+=2$. Let $e_3, Z$ denote the same as in Case 1, let $0\ne e'_2\in\Ker F_+$, and let $e_1\in Z-\Ker F_+,\ F_+(e_1,e_1)=1$. Choose $\alpha$ such that, for $e_2=\alpha e'_2,\ F_-(e_1,e_2)=1$. Then the matrix of $F$ with respect to the basis $e_1,e_2,e_3$ is $J$.
[*Case*]{} 3: $F_+|_W\ne 0, \rank F_+=1$. Let $e_3$ denote the same as in Cases 1 and 2, and let $e_1,e_2$ be a basis in $\Ker F_+$ such that $F_-(e_1,e_2)=1$. Then the matrix of $F$ with respect to the basis $e_1,e_2,e_3$ is $\displaystyle{\left[\begin{array} {ccc} \phantom{-} 0&1&0\\
-1&0&0\\ \phantom{-} 0&0&1\end{array}\right]}$ which is $H_{\frac\pi2}$.
[*Case*]{} 4: $F_+|_W=0, \rank F_+=3$. Choose a non-zero vector $e'_3\in W$. Let $Z$ be the orthogonal complement to $W$, and let $C$ be the “light cone" $\{x\in V\mid F_+(x,x)=0\}$. Since $C\not\subset Z$, we can choose an $e'_2\in V$ such that $F_+(e'_2,e'_2)=0,\ F_+(e'_2,e'_3)=1$. Let $U\in V$ be the subspace spanned by $e'_2$ and $Y$ be the orthogonal complement of $U$. The intersection $Y\cap Z$ is not contained in $W+U$: if a linear combination of $e'_2$ and $e'_3$ is orthogonal to both $e'_2$ and $e'_3$, then it must be 0. Take $e_1\in Y\cap Z$ with $F_+(e_1,e_1)=1$. Then $F_-(e_1,e'_2)\ne 0$ (otherwise $F_-$ would have been zero). Put $e_2=\alpha e'_2,\ e_3=\alpha^{-1} e'_3$ in such a way that $F_-(e_1,e_2)=1$. Then the matrix of $F$ with respect to the basis $e_1,e_2,e_3$ is $K$.
[*Case*]{} 5: $F_+|_W=0, \rank F_+<3$. Take a 1-dimensional space $U\subset\Ker F_+$. If $U=W$, then $\Ker F\supset W$ is non-zero, so $F$ is degenerate. If $U\ne W$, then both $F_+,F_-$ are zero on $U\oplus W$, which also means that $F$ is degenerate. [$\Box$]{}
Return now to the bilinear form $F$ related to our $m$-self-dual $n$-gon $L$.
\[rotation\] In an appropriate coordinate system, the matrix of $F$ is $H_\varphi$ with $r\varphi\in\pi\mathbb Z$. Moreover, if the $n$-gon $L$ is simple, then $s\varphi\notin\pi\mathbb Z$ for any positive $s<r$.
#### *Proof.*
According to Lemma \[canonical\], the matrix of $F$, in an appropriate basis, is $H_\varphi,\ J$, or $K$. But $$J^{-1}J^t=\left[\begin{array} {ccc} -1& \phantom{-}0&0\\
\phantom{-}2&-1&0\\ \phantom{-}0& \phantom{-}0&1\end{array}\right],\ K^{-1}K^t=\left[\begin{array} {ccc} 1&-2&0\\ 0&\phantom{-}1&0\\ 2&-2&1\end{array}\right]$$ and neither of these two matrices has finite order (for both, the Jordan form contains a non-trivial Jordan block). On the other hand, the matrix of $G$ is $(H_\varphi)^{-1}H_\varphi^t=H_{-\varphi}^2=H_{-2\varphi}$, that is (again, in an appropriate coordinate system), $H_{-2\varphi}A_i=A_{i+2m}$. First, this shows that $H_{-2\varphi}^r=I$, that is, $2r\varphi$ is a multiple of $2\pi$. Second, if $2s\varphi$ is a multiple of $2\pi$ for a positive $s<r$, then $A_i=A_{i+2sm}$ where $sm$ is not a multiple of $n$, so our polygon is not simple. [$\Box$]{}
Thus, $L$ contains $(m,n)$ regular $r$-gons, $A_iA_{i+2m}\dots
A_{i+(r-1)m},\ i=1,2,$ $\dots,(m,n)$. \[By a regular $n$-gon we understand a (maybe, self-intersecting) $n$-gon in the Euclidean plane with all lengths of the sides equal and all angles equal; thus, there are two projectively different types of regular pentagons, three projectively different types of regular heptagons, and so on.\] If $(m,n)=1$, then $L$ itself is regular (so, in this case, an $m$-self-dual $n$-gon is projectively unique). If $(m,n)>1$, then this uniqueness, in general, does not hold. Below, we give an explicit construction of all $m$-self-dual $n$-gons which will demonstrate this non-uniqueness.
First, notice that, in our case, the projective duality has a simple geometric description: we consider a Euclidean plane (with a fixed origin) and, for a point $A\ (\ne0)$, the dual line $A^\ast$ is obtained from the polar dual $A^\perp$ by a clockwise rotation about the origin by the angle of $\displaystyle
\frac{\pi(n-m)}n$.
Now, let us construct an arbitrary $m$-self-dual $n$-gon. In addition to $r=\displaystyle\frac n{(m,n)}$, put $k=\displaystyle\frac m{(m,n)}$ and also $d=(m,n)$; thus, $mr=kn$. In the Euclidean plane with a fixed origin $O$, choose an arbitrary point $A_1$. Then successive counter-clockwise rotations by the angle $\displaystyle\frac{2\pi m}n$ give the points $A_{2m+1}, A_{4m+1},\dots, A_{2(r-1)m+1}$, and also the lines $A^\ast_1=A_mA_{m+2},A^\ast_{2m+1}=A_{3m}A_{3m+2},A^\ast_{4m+1}=A_{5m}A_{5m+2},\dots, A^\ast_{2(r-1)m+1)}=A_{(2r-1)m}A_{(2r-1)m+2}$. Of the numbers $m,3m,5m,\dots,2(r-1)m$ one is $d$ modulo $2n$ ($d=um+v\cdot2n$ for a unique $u,\ 0\le u<2r$, and this $u$ must be odd). So, one of our lines should be $A_dA_{d+2}$; choose a point $A_d$ on this line. This choice gives also the points $A_{2m+d},A_{4m+d},\dots,A_{2(r-1)m+d}$ and the lines $A_{d+m-1}A_{d+m+1},
A_{d+3m-1}A_{d+3m+1},$ $A_{d+5m-1}A_{d+5m+1},\dots,A_{d+(2r-1)m-1}A_{d+(2r-1)m+1}$. By the way, one of these lines will be $A_{2n-1}A_1$. Our next choice will be a point $A_{d+2}$, again on the line $A_dA_{d+2}$. This will give us $r$ additional points (including $A_{d+2}$) and $r$ lines, dual to these points. One of these lines will be $A_1A_3$, and we choose a point $A_3$ on it. One of the lines coming with this point will be $A_{d+2}A_{d+4}$, and we choose a point $A_{d+4}$, and so on. Proceeding in these way, we choose the points in the following order: $A_1,A_d, A_{d+2},A_3,A_{d+4},
A_5,A_{d+6}A_7,\dots ,A_{2d-3},A_{d-2}$. Here we stop: the next choice should be $A_{2d-1}$, but this point will appear as the intersection of the line $A_{2d-1}A_{2d+1}$ coming with the point $A_d$ and the line $A_{2d-3}A_{2d-1}$ coming with the point $A_{d-2}$. After that, we have the points $A_1,A_3,
A_5,\dots,A_{2d-1}$, and hence we have all the vertices of our polygon.
The projective symmetry $G,\ G(A_i)=A_{i+2m}$, shows that every $m$-self-dual $n$-gon is also $m'$-self-dual for every $m'\equiv em\bmod n$ where $e$ is odd. In particular, if $(e,r)=1$, then $m$-self-dual $n$-gons and $m'$-self-dual $n$-gons are the same $n$-gons (although their self-dualities involve different projective isomorphisms $P\to P^\ast$). On the other hand, if $n$ is odd, then we see that every $m$-self-dual $n$-gon is also $n$-self-dual, that is, it belongs to the class of polygons considered in Section 3.
![Two 3-self-dual nonagons[]{data-label="9-gons"}](figfourtwo.eps "fig:"){width="1.4in"} ![Two 3-self-dual nonagons[]{data-label="9-gons"}](figfourthree.eps "fig:"){width="1.4in"}
Figure \[9-gons\] shows two 3-self-dual 9-gons. As was remarked above, they are also 9-self-dual (with respect to the polar duality). Note that no $n$-gon with $n$ even is $n$-self-dual (the definition of $m$-self-duality requires that $m$ is odd); but they must be centrally symmetric (with respect to the affine chart considered in this section), and the 12-gons of Figure \[12-gons\] are centrally symmetric indeed.
![Two 3-self-dual dodecagons[]{data-label="12-gons"}](figsix.eps "fig:"){width="2.2in"} ![Two 3-self-dual dodecagons[]{data-label="12-gons"}](figseven.eps "fig:"){width="2.2in"}
\[dimension\] Let $m<n$ and $(m,n)>1$. Then the moduli space of $m$-self-dual $n$-gons has the dimension $(m,n)-1$, for $n\ne2m$, and $(m,n)-3$, for $n=2m$.
#### *Proof.*
If the basis, in which the bilinear form $F$ has the canonical form $H_\varphi$, is chosen then, to specify our $m$-self dual $n$-gon (with $m<n$), we need to choose a point $A_1$ in the plane (minus one point), which depends on two parameters, and then $d-2,\ d=(m,n)$, points $A_d,A_{d+2},
A_3,A_{d+4},A_5,\dots,A_{d-2}$ which provide $(m,n)-2$ more parameters, with the total of $(m,n)$. From this number, we need to subtract the number of parameters on which the basis for a given form $F$ depends (in other words, the dimension of the Lie group of linear transformations of $V$ that preserve the form $H_\varphi$). If $2\varphi$ is not a multiple of $\pi$, which corresponds to the case $n\ne2m$, this dimension is 1. This is seen from Case 1 of the proof of Lemma \[canonical\]: the choice of $e_3$ provides no parameters (it is two-valued), then we choose $e'_1$ and $e'_2$ on the conic $F_+(x,x)=0$ with the condition $F_+(e'_1,e'_2)=0$, which provides one parameter, and then we multiply both $e'_1$ and $e'_2$ by the same complex number which we determine from a quadratic equation. So the total number of free parameters in this case is 1, and the dimension of the moduli space is $(m,n)-1$. If $n=2m$, then $\cos\varphi=0$, and this is Case 3 of the proof of Lemma \[canonical\]. In this case, the choice of $e_3$ does not provide any parameters while $e_1$ and $e_2$ are chosen up to the action of the group of transformations of $Z$ preserving the form $F_-$; it is $SL(2;{{\mathbb C}})$, the dimension is 3. Thus, if $n=2m$, then the dimension of the moduli space is $(m,n)-3=m-3$. [$\Box$]{}
Notice in conclusion that our results show that the space of moduli of $m$-self-dual hexagons has dimension 0, whatever $m$ is. Actually, the only self-dual hexagon (for any $m$) is the regular hexagon; again, we leave the details to the reader.
Polygonal curves {#polycurves}
================
A [*polygonal curve*]{} is a real polygon $A_1A_3\dots A_{2n-1}$ with the following two (independent) additional structures. (1) For every even $i$, one of the two segments into which the points $A_{i-1},A_{i+1}$ cut the real projective line $B_i$ is chosen; we will refer to this segment as an [*edge*]{} of the polygonal curve. (2) For every odd $i$, one of the two pairs of vertical angles formed by the real projective lines $B_{i-1},B_{i+1}$ is fixed; we will refer to these angles as [*exterior angles*]{} of the polygonal curve. Thus, every real $n$-gon gives rise to $2^{2n}$ polygonal curves.
For a polygonal curve $A_1A_3\dots A_{2n-1}$ (with its additional structures), there arises a [*dual polygonal curve*]{} in dual real projective plane $P^\ast$. This is the polygon $B^\ast_2B^\ast_4\dots B^\ast_{2n}$ with the following edges and exterior angles. The edge $B^\ast_{2i}B^\ast_{2i+2}$ is formed by points of $P^\ast$ dual to the lines through $A_{2i+1}$ in $P$ contained in the exterior angles at $A_{2i+1}$ of the given polygonal curve. The exterior angles at $B^\ast_{2i}$ is formed by lines in $P^\ast$ dual to the points of the edge $A_{2i-1}A_{2i+1}$ of the given polygonal curve.
Let $m\le n$ be a positive odd number. A polygonal curve $A_1A_3\dots A_{2n-1}$ is called $m$-self-dual if there exists a projective isomorphism $P\to P^\ast$ which takes $A_i$ into $B^\ast_{i+m}$ and also takes edges and exterior angles of the polygonal curve $A_1A_3\dots
A_{2n-1}$ into edges and exterior angles of the polygonal curve $B^\ast_2B^\ast_4\dots B^\ast_{2n}$.
\[choices\] Let $L=A_1A_3\dots A_{2n-1}$ be an $m$-self-dual $n$-gon. Then, of the $2^{2n}$ polygonal curves arising from the polygon $L$, $2^{(m,n)}$ are $m$-self-dual.
#### *Proof.*
Choose edges $A_1A_3,A_3A_5,\dots A_{2d-1}A_{2d+1}$ in an arbitrary way. If $m<n$, then apply to these edges $r=\displaystyle\frac nd$ consecutive rotations by the angle $\displaystyle\frac {2\pi m}n$; we will get a full set of edges. Then apply the duality to these edges, and this gives a choice of exterior angles for the dual polygon $L^\ast$. The projective isomorphism between $L$ and $L^\ast$ makes these angles exterior angles for $L$, and $L$ becomes an $m$-self-dual polygonal curve. Obviously, this construction gives all $m$-self-dual $n$-gonal curves. [$\Box$]{}
Polygonal curves are natural polygonal counterparts to smooth curves. Let us describe a unifying point of view.
#### The space of contact elements and projective duality.
A contact element of the real projective plane is a pair $(A,B)$ where $A\in P$ is a point, $B\subset P$ is a line, and $A\in B$. Denote the space of contact elements by $F$ (it is naturally identified with the space of full flags in ${{\mathbb R}}^3$). One has two projections $\pi_1\colon F\to P$ and $\pi_2\colon F\to P^\ast$ defined by the formulas: $\pi_1(A,B)=A,\ \pi_2(A,B)=B$. The space $F$ has a contact structure (a non-integrable two-dimensional distribution) defined by the condition that the velocity of point $A$ lies in the line $B$. The fibers of the projections $\pi_{1,2}$ are Legendrian curves (curves tangent to the contact distribution). The space of contact elements of the dual plane $P^\ast$ is canonically identified with $F$.
Projective duality is easily described in terms of the space of contact elements. Let $\gamma\subset P$ be a smooth curve. Assigning the tangent line to each point of $\gamma$ gives a lift $\Gamma\subset F$; this lifted curve is Legendrian. The curve $\pi_2(\Gamma)$ is the dual curve $\gamma^\ast\subset P^\ast$, and the lift of $\gamma^\ast$ to $F$ is again $\Gamma$. A wave front in $P$ is defined as the $\pi_1$-projection of a smooth Legendrian curve $\Gamma$ in $F$ to $P$; it has singularities (generically, semi-cubical cusps) at the points where $\Gamma$ is tangent to the fibers of $\pi_1$. The dual wave front is $\pi_2(\Gamma)$.
Likewise for polygonal curves. The set of edges of a polygonal curve provides a closed curve in $P$, but its lift to $F$ consists of disjoint arcs of the fibers of $\pi_2$; to connect these arcs by segments of the fibers of $\pi_1$, we need to choose exterior angles. Thus, for an $n$-gonal curve $L$ in $P$, its lift to $F$ is a $2n$-gonal curve, whose sides are segments of the fibers of the alternating projections $\pi_1$ and $\pi_2$, and whose projection to $P^\ast$ is the dual polygonal curve $L^\ast$.
Self-dual curves {#sdcurves}
================
Many a result from the preceding sections extends to self-dual wave fronts in ${{\mathbb{RP}}}^2$. Here we do not attempt to give a complete classification of such fronts; instead we describe several classes of examples, including explicit formulas for self-dual curves and fronts. Self-dual curves will be described as Legendrian curves in certain three-dimensional contact manifolds satisfying certain monodromy conditions.
In this section, $P={{\mathbb{RP}}}^2$ and $V={{\mathbb R}}^3$. Let $\gamma(t)\subset P$ be a self-dual parameterized closed curve (possibly, with cusps); we assume that the parameter $t$ takes values in $S^1={{\mathbb R}}/2\pi{{\mathbb Z}}$. The projectively dual curve also has a parameterization, $\gamma^\ast(t)$: the covector $\gamma^\ast(t)$, defined up to a non-zero multiplier, vanishes on the vectors $\gamma(t)$ and $\gamma'(t)$. As in Section \[poldual\], we have a linear isomorphism $f\colon V \to V^\ast$ that takes the line $\gamma(t)\subset V$ to the line $\gamma^\ast(\varphi^{-1}(t))\subset V^\ast$ where $\varphi$ is a diffeomorphism of $S^1$. This diffeomorphism plays the role of the cyclic shift through $m$ in the definition of $m$-self-dual polygons. As in Section \[poldual\], we consider the corresponding projective isomorphism $\hat f\colon P\to P^\ast$ and the bilinear form $F$ on $V$, $F(v,w)=\langle f(v),w\rangle$. As in Section \[m<n\], we consider the projective transformation $G=(\hat f)^{-1}\hat f^\ast \colon P\to P,\ G(B)=(B^\perp)^\perp$.
The next lemma is an analog of Lemma \[shift\].
\[curveshift\] One has: $G(\gamma(t))=\gamma(\varphi^2(t))$ for all $t$.
#### *Proof.*
The proof is essentially the same as that of Lemma \[shift\]. First note that the line $\{y\in P\ |\ \langle y,\gamma(t)\rangle=0\}$ is the tangent line $T_{\gamma^\ast(t)} \gamma^\ast$ to $\gamma^\ast$ at point $\gamma^\ast(t)$. It follows that $\gamma(t)^\perp = (\hat f)^{-1} (T_{\gamma^\ast(t)} \gamma^\ast)=T_{\gamma(\varphi(t))} \gamma$. Likewise, the point $\{y\in P\ |\ \langle y,T_{\gamma(\varphi(t))} \gamma\rangle=0\}$ is $\gamma^\ast(\varphi(t))$, and therefore $(\gamma(t)^\perp)^\perp = \gamma(\varphi^{2}(t))$. [$\Box$]{}
Analogs of the polygons considered in Section \[m=n\] are the curves for which the diffeomorphism $\varphi$ is an involution. Just like a polygon, a curve may be a multiple of another curve. The next proposition is an analog of Proposition \[symmetric\].
\[curvesymm\] Assume that a self-dual curve $\gamma$ is not a multiple of another curve. Then $\varphi^2=\id$ if and only if the bilinear form $F$ is symmetric.
#### *Proof.*
$F$ is symmetric if and only if $G$ is the identity. If $\varphi^2=\id$ then, by Lemma \[curveshift\], $G(\gamma(t))=\gamma(t)$ for all $t$. Since $\gamma$ contains four points in general position, $G=\id$. Conversely, if $G=\id$ then $\gamma(t)=\gamma(\varphi^2(t))$ for all $t$. Since $\gamma$ is not a multiple of another curve, $\varphi^2=\id$. [$\Box$]{}
An analog of Proposition \[convex\] holds as well.
\[curvepos\] If $\gamma$ is a convex self-dual curve such that $\varphi^2=\id$ then the symmetric bilinear form $F$ is definite.
#### *Proof.*
Assume not. Then, in an appropriate coordinate system, the respective quadratic form is $x^2+y^2-z^2$. The light cone $x^2+y^2=z^2$ projects to a circle $C\subset {{\mathbb{RP}}}^2$. If $p\in C$ then the line $p^\perp$ is the tangent line to $C$ at point $p$, and if a point $p$ is inside $C$ then the line $p^\perp$ lies in the exterior of $C$.
Due to the convexity, $\gamma(t)\notin \gamma(t)^\perp$ for all $t$, hence $\gamma$ does not intersect $C$. Therefore $\gamma$ lies either inside $C$ or outside of it. In the former case, the envelope $\gamma^\ast$ of the lines $\gamma(t)^\perp$ lies outside of $C$ and cannot coincide with $\gamma$. In the later case, one can find a tangent line $\ell$ to $\gamma$, disjoint from $C$; then the point $\ell^\perp\in\gamma^\ast$ lies inside $C$, and again $\gamma$ fails to coincide with $\gamma^\ast$. [$\Box$]{}
#### Spherical curves.
Let us consider the case when the symmetric bilinear form $F$ is (positive) definite. Then the correspondence between points and the dual lines is that between pairs of antipodal poles and the corresponding equators on the unit sphere (that doubly covers ${{\mathbb{RP}}}^2$). Thus the projective duality moves every point of a curve $\gamma\subset S^2$ distance $\displaystyle\frac\pi2$ in the normal direction to $\gamma$.
An example of a self-dual curve on $S^2$ is a circle of radius $\displaystyle\frac\pi4$. This circle is included into the family of curves of constant width $\displaystyle\frac\pi2$ that are all self-dual (of course, all distances are in the spherical metric). We interpret curves of constant width as Legendrian curves.
Let $M$ be the space of oriented geodesic segments of length $\displaystyle\frac\pi2$ on $S^2$. Then $M$ is diffeomorphic to $SO(3)\cong {{\mathbb{RP}}}^3$. Define a two-dimensional distribution $E$ on $M$ by the condition that the velocities of the end points of a segment are perpendicular to the segment. Assign an oriented contact element to a geodesic segment $AB$: the foot point is the midpoint of $AB$ and the direction is the oriented normal to $AB$. This provides an identification of $M$ with the space of oriented contact elements $F$ of $S^2$.
Under the diffeomorphism $F\cong M$, the standard contact structure in $F$ is identified with the distribution $E$.
#### *Proof.*
The space $E$ is generated by two vector fields corresponding to the following motions of a geodesic segment $AB$: the rotation of $AB$ about its midpoint, and the rotation of $AB$ about the axis $AB$ (so that the end points are fixed). The velocities of the corresponding motions of the midpoint of $AB$ are orthogonal to $AB$ which proves the lemma. [$\Box$]{}
Thus a curve of constant width can be constructed as a smooth Legendrian curve $A(t) B(t) \subset M,\ t\in {{\mathbb R}}$ satisfying the monodromy condition $A(\pi)=B(0), B(\pi)=A(0)$. Clearly, there is an abundance of such curves, in particular, analytic ones. This construction gives curves with cusps and inflections as well. (A similar approach is used in [@B-Zh] to construct billiard tables that possess one-parameter families of periodic trajectories, the case of two-periodic trajectories being that of curves of constant width.)
To construct such a curve, take a closed wave front on the sphere with an odd number of cusps (say, an odd-cusped hypocycloid), place a geodesic segment of length $\displaystyle\frac\pi2$ orthogonally to the front, so that its midpoint is on the front, and use the front as a guide to move the geodesic segment all the way around until its end points swap their positions.
![Self-dual cubic curve with one inflection and one cusp[]{data-label="cubic"}](fignine.eps){width="2in"}
Curves of constant width $\displaystyle\frac\pi2$ in $S^2$ project to ${{\mathbb{RP}}}^2$ as contractible curves. Similarly one can construct self-dual non-contractible curves in ${{\mathbb{RP}}}^2$. For this, one needs to modify the above monodromy condition: $A(\pi)=B(0), B(\pi)=-A(0)$. A non-contractible curve necessarily has an odd number of inflections, and therefore, by self-similarity, an odd number of cusps. An example is a cubic curve $y=x^3$; after a projective transformation, this curve looks like shown on Figure \[cubic\].
#### Rotationally symmetric curves.
Let us consider analogs of the polygons studied in Section \[m<n\]. Assume that $\gamma$ is a self-dual curve that lies in the affine plane (i.e., is disjoint from the line at infinity) and is star-shaped with respect to the origin $O$ (i.e., the tangent lines to $\gamma$ do not pass through $O$). We assume that $\gamma$ is not a multiple of another curve. We allow $\gamma$ to have inflections and cusps. Assume that the bilinear form $F$ is not symmetric, so, by Proposition \[curvesymm\], $\varphi^2\neq \id$.
Arguing as in Section \[m<n\], we choose a coordinate system in which the mapping $G$ is a rotation through some angle $\alpha$ (the cases of Jordan blocks and complex angles in Lemma \[canonical\] are excluded because in these cases orbits of $G$ would have accumulation points at infinity, in contradiction with Lemma \[curveshift\]). If $\alpha$ is $\pi$-irrational then the orbits of $G$ are dense in a circle and, by Lemma \[curveshift\], $\gamma$ is a circle. Otherwise, $\alpha=\displaystyle\frac{2\pi p}{q}$ where $p$ and $q$ are co-prime. Thus $q$ is the least period of $G$, and hence of the circle diffeomorphism $\varphi^2$ as well. Choose a parameterization $\gamma(t)$ so that $\displaystyle\varphi^2(t)=t+\frac{2\pi r}{q}$ where $r$ and $q$ are also co-prime. To summarize, we have the following analog of Lemma \[shift\].
\[cshift\] One has: $\displaystyle G(\gamma(t))=\gamma\left(t+\frac{2\pi r}{q}\right)$.
Note that the rotation number of $\gamma$ about the origin $O$ equals the least positive $k$ such that $kr=p$ mod $q$.
Let us describe an explicit construction of such self-dual curves. Let $H$ be the rotation about the origin through angle $\displaystyle\frac{\pi p}{q}$ (so that $H^2=G$) and set $c=\displaystyle\frac{\pi r}{q}$. As in Section \[m<n\], the projective duality has a simple geometric description in terms of the Euclidean metric: for point $A$, the dual line $A^\ast$ passes through the point $H(A)/|A|^2$ and is orthogonal to the vector $H(A)$.
Self-dual curves again can be described as Legendrian curves. Let $M$ consists of pairs of vectors $(u,v)$ such that $H(u)\cdot v=1$. Define a contact structure on $M$ by the condition $H(u)\cdot v'=0$ (or, equivalently, $H(u')\cdot v=0$); we leave it to the reader to check that this is indeed a contact structure. Let $(u(t),v(t))$ be a Legendrian curve in $M$ satisfying the monodromy condition $u(t+c)=v(t), v(t+c)=G(u(t))$. Then we can set: $\gamma(t)=u(t)$. The condition $H(u)\cdot v=1$ implies that $v$ belongs to the line $u^\ast$, and the Legendrian condition $H(u)\cdot v'=0$ that this line is tangent to the curve $\gamma$ at point $\gamma(t+c)$. Thus $\gamma$ is self-dual.
Now we give explicit formulas.
\[expform\] Let $\beta(t)$ be a smooth function such that $|\beta(t)|<\displaystyle\frac\pi4$ and $\beta(t+c)=-\beta(t)$. Let $\rho_1(t)$ and $\rho_2(t)$ satisfy the differential equations $$\label{form}
\rho'_1=(\beta'+1) \tan 2\beta,\ \ \rho'_2=(\beta'-1) \tan 2\beta.$$ Then the curve $\gamma(t)$ whose polar coordinates are $$\left(t-\beta(t)-\frac{\pi p}{q}, e^{\rho_1(t)}\right)$$ is self-dual.
#### *Proof.*
Using the above notation, the polar coordinates of the points $H(u(t))$ and $v(t)$ are $(t-\beta, e^{\rho_1})$ and $(t+\beta, e^{\rho_2})$. The differential equations (\[form\]) are the Legendrian conditions $H(u)\cdot v'=0$ and $H(u')\cdot v=0$ which together imply that $H(u)\cdot v$ is constant. One needs to satisfy the monodromy conditions $\rho_1(t+c)=\rho_2(t), \rho_2(t+c)=\rho_1(t)$. Due to (\[form\]), these equalities hold once one has $$0=\int_0^c \rho'_1(t)\ dt + \int_0^c \rho'_2(t)\ dt = 2\int_0^c \tan 2\beta\ d\beta=-\int_0^c d(\ln \cos 2\beta).$$ The latter is zero because $\beta(c)=-\beta(0)$, and we are done. [$\Box$]{}
#### Radon curves.
Let $U$ be a Minkowski plane (two dimensional normed space) and let $\gamma$ be its unit circle, a closed smooth strictly convex centrally symmetric curve centered at the origin. For a vector $u\in \gamma$, one defines its orthogonal complement as the tangent line to $\gamma$ at $u$. This orthogonality relation is not symmetric, in general. A Minkowski plane is called a Radon plane, and the curve $\gamma$ a [*Radon curve*]{}, if the orthogonality relation is symmetric. A Radon curve admits a one-parameter family of circumscribed parallelograms whose sides are orthogonal to each other. Introduced by J. Radon about 90 years ago, Radon curves abound (they have functional parameters), see [@M-S] for a survey.
The relevance of Radon curves to our subject is the following statement.
\[Radon\] Radon curves are projectively self-dual.
#### *Proof.*
Let $[\ ,\ ]$ be an area element (linear symplectic structure) in $U$. Identify $U^\ast$ with $U$ using this area form: $u^\ast = [\,\cdot ,u]$. With this identification, $u^\perp$ is a line $\ell$, parallel to $u$, and such that $[v,u]=1$ for every $v\in \ell$.
Similarly to the preceding discussion, a Radon curve can be realized as a curve $(u(t),v(t))$ in the space of pairs of non-zero vectors $(u,v)$, tangent to the distribution given by the conditions $[u,v']=0=[u',v]$ (these conditions mean that $u$ is orthogonal to $v$ and $v$ is orthogonal to $u$), and satisfying the monodromy conditions: $\displaystyle u\left(t+\frac\pi2\right)=v(t), v\left(t+\frac\pi2\right)=-u(t)$. The equalities $[u,v']=0=[u',v]$ imply that $[u,v]$ is constant or, after rescaling $\gamma$, that $[u,v]=1$. Therefore $u^\perp$ is the tangent line to $\gamma$ at point $v$, that is, $\gamma$ is self-dual. [$\Box$]{}
Let us conclude with two remarks. First, Radon planes can be also characterized as the Minkowski planes for which the unit circle is the solution to the isoperimetric problem (Busemann’s theorem). Secondly, the outer billiard around a Radon curve possesses a one-parameter family of 4-periodic trajectories, see [@G-T] for a study of such outer billiards in the context of sub-Riemannian geometry.
[99]{}
5 mm
V. Arnold. [*Arnold’s problems.*]{} Springer-Verlag, Berlin; PHASIS, Moscow, 2004.
Yu. Baryshnikov, V. Zharnitsky. [*Sub-Riemannian geometry and periodic orbits in classical billiards*]{}. Math. Res. Lett. [**13**]{} (2006), 587–598.
H. Bos, C. Kers, F. Oort, D. Raven. *Poncelet’s closure theorem*. Expos. Math. **5** (1987), 289–364.
D. Genin, S. Tabachnikov. [*On configuration spaces of plane polygons, sub-Riemannian geometry and periodic orbits of outer billiards.*]{} J. Mod. Dyn. [**1**]{} (2007), 155–173.
Ph. Griffiths, J. Harris. [*On Cayley’s explicit solution to Poncelet’s porism.*]{} Enseign. Math. [**24**]{} (1978), no. 1-2, 31–40.
T. Hollcroft. [*Conditions for self dual curves.*]{} Ann. of Math. [**27**]{} (1926), 258–270.
M. Levi, S. Tabachnikov. *The Poncelet grid and the billiard in an ellipse*. Amer. Math. Monthly, in print.
H. Martini, K. Swanepoel. [*Antinorms and Radon curves.* ]{} Aequationes Math. [**72**]{} (2006), 110–138.
R. Schwartz. *The pentagram map is recurrent.* Experiment. Math. 10 (2001), 519–528.
R. Schwartz. *Discrete monodromy, pentagrams, and the method of condensation*. Preprint.
R. Schwartz. *The Poncelet grid*. Adv. Geom. [**7**]{} (2007), 157–175.
[^1]: Department of Mathematics, University of California, Davis, CA 95616, USA; e-mail:
[^2]: Department of Mathematics, Pennsylvania State University, University Park, PA 16802, USA; e-mail:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The launch of Hinode satellite led to the discovery of rising plumes, dark in chromospheric lines, in quiescent prominences that propagate from large ($\sim 10$Mm) bubbles that form at the base of the prominences. These plumes present a very interesting opportunity to study Magnetohydrodynamic (MHD) phenomena in quiescent prominences, but obstacles still remain. One of the biggest issues is that of the magnetic field strength, which is not easily measurable in prominences. In this paper we present a method that may be used to determine a prominence’s plasma $\beta$ when rising plumes are observed. Using the classic fluid dynamic solution for flow around a circular cylinder with an MHD correction, the compression of the prominence material can be estimated. This has been successfully confirmed through simulations; application to a prominence gave an estimate of the plasma $\beta$ as $\beta=0.47 \pm 0.079$ to $1.13\pm 0.080$ for the range $\gamma=1.4$-$1.7$. Using this method it may be possible to estimate the plasma $\beta$ of observed prominences, therefore helping our understanding of a prominence’s dynamics in terms of MHD phenomena.'
author:
- 'Andrew Hillier, Richard Hillier and Durgesh Tripathi'
title: 'Determination of Prominence Plasma $\beta$ from the Dynamics of Rising Plumes'
---
Introduction {#intro}
============
Quiescent prominences are large clouds of relatively cool plasma supported against gravity by their magnetic field. It is known that the temperature of quiescent prominences is approximately 8000K [@TH1995] and density $\sim 10^{11}$cm$^{-3}$ [@HIR1986] which is a decrease and increase of approximately two orders of magnitude respectively from the surrounding corona. Using this value for the temperature the pressure scale height can be calculated as $\Lambda \sim 300$km, which is approximately $2$ orders of magnitude less than the characteristic height of a quiescent prominence [$\sim 25$Mm @TH1995]. There are many reviews that describe the current understanding of the nature of quiescent prominences [see, for example, @TH1995; @LAB2010; @MAC2010].
Observations by the Solar Optical Telescope [SOT; @TSU2008] on the Hinode satellite [@KOS2007] have shown that on very small scales quiescent prominences are highly dynamic and unstable phenomena. [@BERG2008] and [@BERG2010] reported dark upflows that propagate from large ($\sim 10$Mm in size) bubbles that formed at the base of some quiescent prominences, through a height of approximately $10$Mm before dispersing into the background prominence material. An example of these plumes is shown in Figure \[prom\]. Observations by [@BERG2011] show that the bubble and plumes have a minimum temperature of $250,000$K. The dark upflows were found to rise at constant velocity of $\sim 20$kms$^{-1}$. Often these plumes would separate from the cavity forming bubbles inside the prominence material.
![Panel (a) shows a quiescent prominence observed on the 3-Oct-2007 at 04:16UT. Panel (b) shows a zoom-in of the region in the box in panel (a). The plume and the bright prominence plasma above it are clearly visible.[]{data-label="prom"}](figure1_a.eps "fig:"){height="8cm"} ![Panel (a) shows a quiescent prominence observed on the 3-Oct-2007 at 04:16UT. Panel (b) shows a zoom-in of the region in the box in panel (a). The plume and the bright prominence plasma above it are clearly visible.[]{data-label="prom"}](figure1_b.eps "fig:"){height="8cm"}
[@BERG2010] hypothesized that the observed upflows were caused by the magnetic Rayleigh-Taylor instability, as a mixed mode perturbation with an interchange mode ($k$ perpendicular to $B$) and an undular mode ($k$ parallel to $B$), in the high Atwood number limit \[$A=(\rho_+ - \rho_-)/(\rho_+ + \rho_-) \rightarrow 1$\], where $k$ is the wave number, $\rho_+$ is the density of the region above the contact discontinuity and $\rho_-$ is the density of the below the contact discontinuity. [@RYU2010] described how the theoretically predicted growth rates and behaviour for the magnetic Rayleigh-Taylor instability well match the observations of quiescent prominence plumes. [@HILL2011] and [@HILL2012] showed how upflows could be created by the 3D mode of the magnetic Rayleigh-Taylor instability acting on the boundary between the Kippenhahn-Schlüter prominence model [@KS1957] and a hot tube inserted into the it. The upflows created were found to be qualitatively similar to the observed plumes.
Observations of the magnetic field of quiescent prominences show the field strength to be in the range $\sim 3$-$30$ G [@LER1989], where the average for polar-crown prominences is $\sim 5$G [@ANHEIN2007]. However, as it is very difficult to observe the prominence magnetic field directly, often other methods to determine the magnetic field of prominences have been employed. One common mechanism has been to extrapolate the photospheric magnetic field to obtain values for the coronal field where the prominence is observed [@DUD2008; @AUL2003]. Another common method is to analyse the oscillatory pattern of filaments [@ISO2006; @PINT2008; @TRIP2009] or filament threads [@LIN2007; @LIN2009] to calculate the magnetic field through the application of wave theory. This is known as prominence seismology [@AGG2012]. In this paper we present a new method to investigate the magnetic field strength relative to the gas pressure by determining the plasma $\beta$ (ratio of gas pressure to magnetic pressure) of a prominence through the study of compression of the prominence material by the Rayleigh-Taylor plumes.
The Observations Needed to Determine a Prominence’s Plasma $\beta$
==================================================================
Before the methodology through which the prominence plasma $\beta$ can be determined is presented, first we will quickly review the observational characteristics of the prominence plumes that are important for this analysis. The points we need to remember are [@BERG2010]:
- The plumes have a phase in their evolution where they rise at an approximately constant velocity
- The plumes have an approximately circular (elliptical) head
- The intensity of the prominence material at the top of the plume is greater than the average prominence intensity.
- Flows of material can be observed along the head of the plume
To make it possible to use these observables, a number of assumptions are necessary. The first is in relation to the density of the prominence. The prominence intensity is known to be a result of scattering. Therefore the intensity is directly related to the column density of the prominence. This allows us to assume that the intensity can be used as a proxy for the density. Hence the bright top of the plume would imply that there has been an increase in the density through compression of the prominence.
The next assumption is in relation to the modelling of the geometry of the problem. As the plume moves at an approximately constant velocity with an elliptical head, this can be used to simplify the system under study. The first step is to remove the stem of the plume, and just leave an ellipse in the prominence material. Then as this ellipse is moving at constant velocity through (what we assume to be) a constant medium, a shift in reference frame can make the situation of an ellipse, stationary in a constant flow. The effect of a non-stationary prominence will be investigated to some extent later. Now we must consider the 3D nature of the geometry under investigation. The observed plumes are known to be caused by the magnetic Rayleigh-Taylor instability [@BERG2010; @HILL2012] It is also known that the 3D magnetic Rayleigh-Taylor instability creates filamentary structures aligned with the magnetic field [@ISO2005; @ISO2006b; @Stone2007; @HILL2012] or the bi-sector of the upper an lower magnetic fields in the case where shear is present [@Stone2007]. Therefore we can view the problem as being similar to a elliptical cylinder in a constant flow. Then a coordinate transform to cylindrical coordinates would effectively give a circle in a constant flow. This has now reduced to a problem that has been significantly studied in fluid dynamics.
The transform described above is a conformal transform of the coordinate system. In an incompressible regime, such a conformal transform would automatically give the correct result, but in a compressible regime, where nonlinearity is important, the velocity will not be completely accurate. The smaller the eccentricity and the fast-mode Mach number, the more accurate the coordinate transform will be.
Derivation of equations to determine a prominence’s plasma $\beta$
==================================================================
Here we use a classic result from fluid dynamics to define a set of equations that can be used to interpret prominence observations and determine physical parameters relating to that prominence. First, we describe the thought process behind the method, followed by a derivation of the equations used to determine the plasma $\beta$ of the prominence.
The derivation presented here is analogous to that presented in @VD1975 for a compressible flow around a circular cylinder in a steady state. Here the derivation is altered to include a magnetic field that is perpendicular to the plane in which the plasma is flowing. In this derivation we assume that the magnetic field is in the $x$-direction and the flow has velocity components in the $y$- and $z$-direction. All derivatives in the $x$-direction are assumed to be $0$. The validity of these assumptions will be investigated through comparisons with simulations in section \[SIM\]. This analysis uses the equations of mass conservation, momentum and energy.
The equation of mass conservation, for a steady planar two-dimensional flow is: $$\frac{\partial \rho v}{\partial y}+\frac{\partial \rho w}{\partial z}=0$$ where $\rho$, $v$ and $w$ are the density, the $y$-component of the velocity and the $z$-component of the velocity. The equation presented above is for a steady state solution, so there is no temporal variation of the density and the velocity field. All these values have been normalised by the free stream reference values for density $\rho_{\infty}$ and velocity $U_{\infty}$, where the values with the subscript $\infty$ denote the values upstream of the plume head. This can then be rewritten to give: $$\label{mass_cons}
\frac{\partial v}{\partial y}+\frac{\partial w}{\partial z}=-\frac{v}{\rho}\frac{\partial \rho}{\partial y}-\frac{w}{\rho}\frac{\partial \rho}{\partial z}.$$
If the flow is irrotational, then the momentum equations can be simply replaced by the statement that the vorticity is zero, that is: $$\label{irrot}
\frac{\partial w}{\partial y}-\frac{\partial v}{\partial z}=0$$ which is satisfied by the velocity potential $\phi$ where: $$v=\frac{\partial \phi}{\partial y}; w=\frac{\partial \phi}{\partial z}$$
The next step is to relate $\rho$ to the velocity, so that Equation \[mass\_cons\] can be expressed purely in velocity derivatives, that is to say purely in derivatives of $\phi$. For a perfect gas under adiabatic contraction/expansion that has a magnetic field perpendicular to the plane of the flow, the MHD-Bernoulli’s equation can be written as: $$\frac{1}{2}\left(v^2+w^2\right)+\frac{1}{\gamma-1}\frac{p\gamma}{\rho}+\frac{B_x^2}{4\pi\rho}=C(S)$$ where $C(S)$ is used to show that the value is constant along streamlines. Therefore, the values of the physical variables at each point in the flow can related to the free stream values (the values of the flow at an infinite distance from the circular cylinder). $$\label{ENERGY}
\frac{\gamma}{\gamma-1}\frac{p p_{\infty}}{\rho \rho_{\infty}}+\frac{1}{2}\left( v^2 +w^2 \right)U^2_{\infty}+\frac{B_x^2}{\rho}\frac{B_{x \infty}^2}{4\pi \rho_{\infty}}=\frac{\gamma}{\gamma-1}\frac{p_{\infty}}{\rho_{\infty}}+\frac{1}{2}U^2_{\infty}+\frac{B_{x \infty}^2}{4\pi \rho_{\infty}}$$ where the subscript $\infty$ implies the free stream values that are used for normalisation.
If we define the free stream speed of sound as $C_s=\sqrt{\gamma p_{\infty}/\rho_{\infty}}$, the free stream Alfven velocity as $V_a=B_{x\infty}/\sqrt{4\pi \rho_{\infty}}$ and denote the free stream Mach number as $M_{\infty}=U_{\infty}/C_s$ then equation \[ENERGY\] becomes: $$\frac{1}{\gamma-1}\frac{p}{\rho}+\frac{1}{2}M_{\infty}^2(v^2+w^2)+\frac{B_x^2}{\rho}\frac{V_A^2}{C_s^2}=\frac{1}{\gamma-1}+\frac{1}{2}M_{\infty}^2+\frac{V_A^2}{C_s^2}.$$ We can then define a relation between the Alfven velocity and the sound speed as follows: $$\frac{V_A^2}{C_s^2}=\frac{2}{\gamma \beta_{\infty}}$$ where $\beta_{\infty}=B_{x\infty}^2/p_{\infty}$ is the free stream plasma $\beta$. This then gives: $$\label{Mach_eqn}
\frac{1}{\gamma-1}\frac{p}{\rho}+\frac{1}{2}M_{\infty}^2(v^2+w^2)+\frac{B_x^2}{\rho}\frac{2}{\gamma \beta_{\infty}}=\frac{1}{\gamma-1}+\frac{1}{2}M_{\infty}^2+\frac{2}{\gamma \beta_{\infty}}.$$
As we have assumed that the flow is irrotational, we can assume that the flow is constant entropy. For a perfect gas (in normalised form) this can be expressed as: $p/\rho^{\gamma}=1$. We can also use the identity that $B_x^2=p/\beta'$, where $\beta' \beta_{\infty}$ defines the plasma $\beta$ of system. Using these identities, equation \[Mach\_eqn\] becomes $$\rho^{\gamma-1}+\frac{\gamma-1}{2}M_{\infty}^2(v^2+w^2 -1)+\frac{\rho^{\gamma-1}}{\beta'}\frac{2(\gamma-1)}{\gamma \beta_{\infty}}=1+\frac{2(\gamma-1)}{\gamma \beta_{\infty}}$$ To tidy up the equations, we define: $$\begin{aligned}
D(y,z)=1+\frac{1}{\beta'}\frac{2(\gamma-1)}{\gamma \beta_{\infty}}\\
C=1+\frac{2(\gamma-1)}{\gamma \beta_{\infty}}\end{aligned}$$ giving: $$\rho^{\gamma-1}+\frac{\gamma-1}{2}\frac{1}{D(y,z)}M_{\infty}^2(v^2+w^2 -1)=\frac{C}{D(y,z)}$$
For simplicity, we will now make the assumption that the plasma $\beta$ is approximately uniform. In reality, we would have to solve for this using the induction equation, but for simplicity this is neglected in this analysis. This leads to the following equation: $$\label{reduced_en}
\rho^{\gamma-1}+\frac{\gamma-1}{2}\frac{1}{D}M_{\infty}^2(v^2+w^2 -1)=1.$$ Now it is possible to combine equation \[reduced\_en\] with equation \[mass\_cons\] to determine an equation for the velocity field.
If we follow the derivation by @VD1975, by transforming to cylindrical coordinates (where the length-scale is normalised by the radius of the cylinder) we can then calculate the velocity potential $\phi$ (with all $\partial / \partial x=0$): $$\label{firstMstar}
\phi=\left(r+\frac{1}{r} \right)\cos \theta - {M_*}^2\left[ \left(\frac{13}{12}\frac{1}{r}-\frac{1}{2}\frac{1}{r^3}+\frac{1}{12}\frac{1}{r^5} \right) \cos\theta -\left(\frac{1}{12}\frac{1}{r^3} - \frac{1}{4}\frac{1}{r} \right)\cos ^3 \theta \right]$$ where: $$\label{Meqn}
M_* = \sqrt{\frac{1}{D}}M_{\infty}$$ which can be viewed as the fast-mode Mach number. Using $\mathbf{v}= \nabla \phi$, we get: $$\begin{aligned}
v_r(r,\theta) = & \left(1-\frac{1}{r^2} \right)\cos \theta - {M_*}^2\left[ \left(-\frac{13}{12}\frac{1}{r^2}+\frac{3}{2}\frac{1}{r^4}-\frac{5}{12}\frac{1}{r^6} \right) \cos\theta +\left(\frac{1}{4}\frac{1}{r^4} - \frac{1}{4}\frac{1}{r^2} \right)\cos ^3 \theta \right] \notag \\
& + O({M_*}^4)\label{vr}\\
v_{\theta}(r,\theta) = & -\left(1+\frac{1}{r^2} \right)\sin \theta + {M_*}^2\left[ \left(\frac{13}{12}\frac{1}{r^2}-\frac{1}{2}\frac{1}{r^4}+\frac{1}{12}\frac{1}{r^6} \right) \sin\theta -\left(\frac{1}{12}\frac{1}{r^4} - \frac{1}{4}\frac{1}{r^2} \right)3 \cos ^2 \theta \sin \theta \right] \notag \\
& + O({M_*}^4)\label{vthe}\end{aligned}$$
Derivation of density distribution
----------------------------------
It is now necessary to derive an equation for the distribution of the density. We need to convert equation \[reduced\_en\] to cylindrical coordinates: $$\label{derivedeqn}
\rho^{\gamma-1}+\frac{\gamma-1}{2}{M_*}^2(v_r^2+v_{\theta}^2 -1)=1$$ Rewriting Equation \[derivedeqn\], we find: $$\label{dendist}
\rho=\left(1-\frac{\gamma-1}{2}{M_*}^2(v_r^2+v_{\theta}^2 -1)\right)^{1/(\gamma-1)}$$ This equation can now be used to determine the expected density distribution around a plume head.
Figure \[eqnden\] shows the density distribution around a circular cylinder. The high density region, created as a result of the compression of the plasma, is clearly visible at the top of the circular cylinder. This is for the case where the higher order terms were used, i.e. the distribution shown is from Equation \[dendist\].
![The density found for a compressible flow around a circular cylinder. Brighter colours imply higher densities.[]{data-label="eqnden"}](figure2.eps){height="8cm"}
Comparison with Simulation Results {#SIM}
==================================
As it is not possible to directly confirm the effectiveness of this method for determining the plasma $\beta$ of a prominence, it is important to apply it to simulations results to provide some validation. We apply it to the results of @HILL2011 and @HILL2012, where a full description of the simulations are presented. In this section we apply Equation \[dendist\] to solve the forward problem, i.e. we use the parameters of the simulation - including the plasma $\beta$ - to see how well the density distribution around the plume head can be reproduced.
The simulation used in this paper is of the 3D magnetic Rayleigh-Taylor instability in the Kippenhahn-Schülter prominence model. To destabilise the Kippenhahn-Schülter prominence model, a low density tube (density of 30% of the prominence) was placed inside the model. The boundary between this tube and the dense material was subjected to a random velocity perturbation allowing the magnetic Rayleigh-Taylor instability to develop. A 3D rendering of the initial conditions of this simulation are shown in Figure \[INIT\].
![3D rendering of the initial conditions of the simulation. The blue isosurface shows the density and the lines show the magnetic field.[]{data-label="INIT"}](figure3.eps){height="8cm"}
In this study, 3D conservative ideal MHD equations were used. Constant gravitational acceleration was assumed, but viscosity, diffusion, heat conduction and radiative cooling terms were neglected and an ideal gas was assumed. The equations were non-dimensionalised using the sound speed ($C_s=13.2$kms$^{-1}$), the pressure scale height ($\Lambda=C_s/(\gamma g)= R_gT/(\mu g)=6.1 \times 10^7$ cm), the density at the centre of the prominence ($\rho (x=0)=10^{-12}$ g cm$^{-3})$ and the temperature ($T_0=10^4$K), giving a characteristic timescale of $\tau=\Lambda/C_s=47$s. The ratio of specific heats was set as $\gamma=1.05$ and $\beta=0.5$. A full description of the simulations can be found in @HILL2012.
Description of simulated plumes {#SIMPlume}
-------------------------------
Panel (a) of Figure \[plumerise\] shows the density distribution at the centre of the the Kippenhahn-Schlüter model as the rising plumes, created by the magnetic Rayleigh-Taylor instability move upward through the dense prominence material. The plume in the box, more like a bubble in this image but really a low-density tube, is the plume we use for this comparison. Note that we didn’t have to choose a plume that has detached from its tail, as we are only interested in the density at the head of the plume.
![Panel (a) shows the rising plumes that are formed by the magnetic Rayleigh-Taylor instability in the Kippenhahn-Schlüter model. Panel (b) shows the time-distance diagram taken along the dashed line in panel (a). This allows the Mach number of the rise velocity of the plume to be calculated as $M=0.28$.[]{data-label="plumerise"}](figure4.eps){height="8cm"}
Panel (b) of Figure \[plumerise\] shows the time-distance plot taken along the slit in Panel (a). The constant rise velocity of the plumes, resulting from the force balance created at the head of the plume, is clearly shown [@HILL2012]. The rise velocity of the topmost plume was found to have a Mach number of $M=0.28$. We use the free-stream Mach number because if we transform the plume head to 0-velocity rest frame then the rise velocity transforms to a free-stream velocity of the dense material.
It can be instantly noticed in Figure \[plumerise\] that the plume is not perfectly circular. The first step of the transform is to change the coordinate size so that the size of the elliptical bubble is that of a unit ellipse (i.e. the area of the ellipse is $\pi$). This is achieved by dividing the $y$ and $z$ axis by $\sqrt{ab}$, where $a$ is the minor axis of the ellipse and $b$ is the major axis of the ellipse. A second transform needs to be performed to transform the unit ellipse to a unit circle. This is achieved by multiplying the $z$ axis by $\sqrt{b/a}$ and the $y$ axis by $\sqrt{a/b}$. This process is shown in Figure \[transform\]. The curved line marking the head of the plume is used in Figure \[plumeheadvel\] to compare the velocity distribution at the plume head to the expected distribution.
![Calculation of the coordinate shift for the plume head. panel (a) shows the original size with the plume dimensions marked ($a=1.05\Lambda$ and $b=0.9\Lambda$) and panel (b) shows the result of that transform. The arch at the top of the plume marks the size of the plume head.[]{data-label="transform"}](figure5.eps){height="8cm"}
Comparison of Simulated and Predicted Density Distribution and Velocity Field {#eqnapp}
-----------------------------------------------------------------------------
To apply Equation \[dendist\], various parameters have to be determined. The Mach number of the rising flow is shown in Figure \[plumerise\] to be $0.28$. The density distribution needs to take in mass conservation, so the column density (integrating in the x-direction) needs to be used. We calculate the average density over $x=0$-$2.5\Lambda$, where x is the direction of the horizontal magnetic field in the simulation domain. In a similar method to determining the density distribution, the average plasma $\beta$ is calculated in the same way, giving $\beta \sim 0.55$. The ratio of specific heats ($\gamma$) was taken as $\gamma=1.05$ for the simulations presented here.
The next important step is to determine the background values. We determine the background density profile as the density profile at the same height at the dashed line along the left-hand edge of Figure \[avden\]. This is then removed from the distribution above the head of the plume to show the density increase due to the compression by the rising plume. Finally the background distribution, required to normalise the density distribution, is calculated from the average distribution (over $x\in [0,2.5\Lambda]$) as $\rho_{av}\sim0.67$. One final point to note is that the background distribution has a weak upflow (Mach number $M=0.07$) that needs to be included in the calculation to move the rising plume into a stationary reference frame with a free stream Mach number of $M_{\infty}=0.21$. It is important to note that because the frame of reference is changed, this free stream Mach number is considered as the Mach number of the flow toward a stationary plume head.
Figure \[avden\] shows the density profile of the simulation at the head of the plume, with contours from the predicted distribution to highlight the expected distribution. The high density region at the head of the plume and the rarefied regions at the sides of match well between the predicted and simulated densities.
![This shows the average density (averaging in the x-direction). The density increase at the head of the plume can be clearly seen. The white contours show the predicted distribution. The dashed black line shows the position of the slit used in Figure \[denprofile\]. The background distribution has been removed.[]{data-label="avden"}](figure6.eps){height="8cm"}
Next we compare the velocity distribution at the plume head with the distribution predicted by Equations \[vr\] and \[vthe\]. Figure \[plumeheadvel\] shows the predicted distributions (solid line) and the simulated distributions (dashed line) for the vertical and horizontal velocity. The horizontal scale uses the same scale as that of Figure \[transform\], but centers the zero position at the plume head. The velocity distributions are in general good agreement. It should be noted that the velocities are multiplied by a factor of $\rho(x=0)/\rho_{av}$ to give the correct magnitude of the velocity (i.e. the maximum value of $W/Cs \sim 0.21$ ). Note the good agreement in the velocity field even though the circular solution has been modified to apply to an elliptical body.
![Velocity distribution along the head of the plume (as marked in Figure \[transform\] panel (a)).The solid line shows the simulated distribution and the dashed line shows the predicted distribution.[]{data-label="plumeheadvel"}](figure7.eps){height="8cm"}
Figure \[denprofile\] shows the density distribution (minus the background distribution) at the top of a plume (solid line). The dashed line shows the expected density distribution when using the parameters discussed above to give a value of $M_* \approx 0.19$. The plots demonstrate that the theoretically predicted density distribution and velocity field well describes those from the simulation.
![Density profile along dotted line shown in Figure \[avden\] minus the background density distribution (shown with the solid line) and the expected distribution using the parameters of the simulation explained in Section \[eqnapp\] (shown with dashed line).[]{data-label="denprofile"}](figure8.eps){height="8cm"}
To investigate this further, we can fit the density distribution using Equation \[dendist\] and solve for ${M_*}^2$. This, in turn, will make it possible to determine the value of the plasma $\beta$. The method is explained in greater detail in the next section. The fit for the density distribution gives ${M_*}^2=0.038$, which gives a plasma $\beta$ of $\beta = 0.59$ when solving for the unknown (i.e. $\beta$) in Equation \[Meqn\]. This value is slightly higher than the input value, but sufficiently close to show that the method can estimate the plasma $\beta$ using the compression of the plume. Figure \[chisqu\] shows the chi-squared for the density distribution for models using different plasma $\beta$ values and the simulated distribution. The minimum of the $\chi^2$ distribution at a $\beta$ value that is close the value of plasma $\beta$ in the simulation.
![$\chi^2$ distribution calculate for different values of plasma $\beta$ giving different density distributions and the simulated distribution.[]{data-label="chisqu"}](figure9.eps){height="8cm"}
Application to Observed Prominence Data {#PROM_SEIS}
=======================================
In this section we apply the method presented above to determine the plasma $\beta$ of the plume displayed in Figure \[prom\]. In contrast to Section \[SIM\], this is treated as an inverse problem, where using the intensity and rise velocity allows the plasma $\beta$ of the prominence to be deduced. Due to the timescales of these compressions ($\sim100$s) being shorter than the timescales for the cooling ($\sim 1000$s), we can assume to first-order approximation that the increased intensity comes through purely adiabatic effects. These adiabatic effects would be an increased column density and a temperature rise from the compression. Both of these can be calculated from the model described above.
Figure \[prom\] panel (a) shows a quiescent prominence seen on the NW solar limb on 2007 October 03 observed by the SOT with the Ca [II]{} H filter at a cadence of 30 s. The time series of this observation was between 01:16UT and 04:59UT. This prominence presents many interesting dynamic features, for example the start of this observation (01:16UT) a large bubble has formed inside the prominence similar to those described in [@BERG2010]. There are also a number of bright threads and downwardly propagating knots that occur during the duration of the observations as well as the upwardly ejected plasma blobs presented in [@HILL2011]. In this paper we are interested in the rising plume and, in particular, the bright rim that forms at the top of the plume. This is shown in panel (b) of Figure \[prom\].
Calculation of the Plume Mach Number {#MACH}
------------------------------------
Before the plasma $\beta$ is calculated, it is necessary to calculate the Mach number of the rising plume. Following the method applied in Section \[SIMPlume\], we create a time-distance plot to show how the height of the plume changes with time. The dashed vertical line in Panel (a) of Figure \[prom\] shows the position of the slit used.
Panel (a) of Figure \[timedist\] shows the time-distance plot, where the bright, rising region is the rising lip of the plume. The cross marks show the top of the plume. This is calculated as the pixel with the greatest intensity change from the pixel below change multiplied by the pixel intensity. Using these pixels, the velocity of the plume and the velocity error were calculated to be $v_{plume}=12.3 \pm 0.6$kms$^{-1}$. As this rise velocity is calculated from a time-distance diagram, it should be viewed as a lower limit on the velocity. The dashed line in the figure shows a velocity of $\sim 12$kms$^{-1}$. Throughout the rise of the plume, the rise velocity can be seen to be approximately constant.
![Panel (a) shows time-distance diagram along the slit shown in Figure \[prom\] panel (b). The cross marks show the position of the top of the plume. The rise velocity of the plume is found to be $\sim 12$kms$^{-1}$. Panel (b) shows the dimensions of the ellipse used to model the rising plume.[]{data-label="timedist"}](figure10_a.eps "fig:"){height="8cm"} ![Panel (a) shows time-distance diagram along the slit shown in Figure \[prom\] panel (b). The cross marks show the position of the top of the plume. The rise velocity of the plume is found to be $\sim 12$kms$^{-1}$. Panel (b) shows the dimensions of the ellipse used to model the rising plume.[]{data-label="timedist"}](figure10_b.eps "fig:"){height="8cm"}
To calculate the Mach number of a flow it is necessary to know the sound speed of the ambient material. The sound speed of the dense prominence material is $C_s=[\gamma (R/\mu)T_{prom}]^{1/2}=11$kms$^{-1}$ for $\gamma=5/3$, $R=8.3\times 10^7$ergK$^{-1}$mol$^{-1}$, $\mu=0.9$ and $T_{prom}=8000$K. Using the velocity calculated above we find the Mach number of the flow to be $M_{\infty}=1.12 \pm 0.05$ in the case where $\gamma=5/3$.
Panel (b) of Figure \[timedist\] shows the dimensions of the ellipse used. In this case $a=1.3$arcsec and $b=2.35$arcsec. These lengthscales can now be used to determine the normalising length scales for the system, so that the plume head becomes circular. Now the prominence plasma $\beta$ can be determined.
Determining Plasma $\beta$ of Observed Prominences
--------------------------------------------------
Figure \[intfit\] shows the normalised intensity. Using the emissions for heights greater than $4$ in Figure \[intfit\] a linear fit was used to de-trend the data (as was applied to the simulations). This intensity is used as a proxy for the column density assuming the emission is scattering dominate ($\propto \rho$). In this paper we do not take into account the change in emission from the change in temperature. The fitting for the density is shown by the dotted line. This fit is performed using the IDL routine curvefit.pro. The fit solves for the value of ${M_*}^2$ giving the values shown in Table \[gam\_table\].
It is important to determine the error associated with this fit, so that the error in the plasma $\beta$ can later be calculated. There are two important errors that are associated with this fit: the intensity error and the error between the observed curve and the model curve. These errors then determine the error for ${M_*}^2$.
The intensity error is determined in a simple fashion. The fluctuations of intensity in the corona above the prominence are used. Here we assume that the fluctuations in the Ca II H line (which should not have a coronal signal) are due to stray light, and so we assume this is representative of the error in the intensity. For a region in the corona above the prominence of size \[200,100\] pixel taken over 100 consecutive frames, the standard deviation of the intensity fluctuations is given as $\sigma_{cor}=1.33$. As the fitting is normalised, the error should also be normalised. Performing this normalisation results in a normalised standard deviation of $\hat{\sigma}_{cor}=0.028$. This values is set as the intensity error of the prominence, i.e. $\hat{\sigma}_{prom}=0.028$. It should be noted that this value is approximately the same as the standard deviation of the fluctuations shown in the prominence in Figure \[intfit\] for heights greater than $4$.
The fitting error is determined as the standard deviation of the difference between the normalised intensity and the fit at each pixel. These errors are shown in Table \[gam\_table\]. The error bars equivalent to $2\hat{\sigma}_{prom}$ and $2\hat{\sigma}_{fit}$ are shown in Figure \[intfit\]. The errors for ${M_*}^2$ from the fitting routine are also shown in Table \[gam\_table\].
![The graph shows a plot of intensity along the slit (solid line) and the fitting assuming the intensity proportional to density (dashed line) for $\gamma=1.65$. Error bars equivalent to twice the standard deviation are presented for the observed intensity and the fit.[]{data-label="intfit"}](figure11.eps){height="8cm"}
Using the value of $M_*$ from the fitting of the intensity and the definition of $M_*$ as: $$M_*=\sqrt{\frac{1}{D}}M_{\infty}=\sqrt{\frac{\gamma \beta_{\infty}}{\gamma \beta_{\infty}+2(\gamma-1)}}M_{\infty}$$ It is possible to solve for the plasma $\beta$ of the prominence, where $\beta$ is given by: $$\beta=\frac{2(\gamma-1)}{\gamma}\frac{M_*^2}{M_{\infty}^2-M_*^2}$$ The results for the plasma $\beta$ for different values of $\gamma$ are given in Table \[gam\_table\]. The range of values covers two orders of magnitude from $\beta=0.044 \pm 0.078$ for $\gamma=1.05$ to $\beta=2.666 \pm 0.080$ for $\gamma=2.0$. However, the value for $\gamma$ in a prominence is likely to be in the range $\gamma=1.4$-$1.7$, giving a range of plasma $\beta$ of $\beta=0.47 \pm 0.079$ to $1.13\pm 0.080$. This is a range of only a factor of $\sim2.5$.
$\gamma$ $\chi^2$ of fit $M_*^2$ $\sigma(M_*^2)$ $\beta$ $\sigma(\beta)$
---------- ----------------- --------- ----------------- --------- -----------------
$1.05$ $1.410$ $0.61$ $0.027$ $0.04$ $0.078$
$1.1$ $1.407$ $0.62$ $0.027$ $0.09$ $0.079$
$1.15$ $1.404$ $0.63$ $0.027$ $0.14$ $0.079$
$1.2$ $1.401$ $0.63$ $0.028$ $0.20$ $0.079$
$1.25$ $1.400$ $0.64$ $0.028$ $0.26$ $0.079$
$1.3$ $1.394$ $0.65$ $0.029$ $0.32$ $0.079$
$1.35$ $1.390$ $0.65$ $0.029$ $0.39$ $0.079$
$1.4$ $1.386$ $0.66$ $0.030$ $0.47$ $0.079$
$1.45$ $1.382$ $0.67$ $0.030$ $0.56$ $0.080$
$1.5$ $1.377$ $0.67$ $0.030$ $0.64$ $0.080$
$1.55$ $1.372$ $0.68$ $0.036$ $0.74$ $0.080$
$1.6$ $1.367$ $0.68$ $0.031$ $0.86$ $0.090$
$1.65$ $1.362$ $0.69$ $0.031$ $0.99$ $0.080$
$1.7$ $1.357$ $0.70$ $0.031$ $1.13$ $0.080$
$1.75$ $1.351$ $0.71$ $0.032$ $1.30$ $0.080$
$1.8$ $1.345$ $0.71$ $0.032$ $1.49$ $0.080$
$1.85$ $1.339$ $0.72$ $0.032$ $1.72$ $0.080$
$1.9$ $1.333$ $0.73$ $0.033$ $1.98$ $0.080$
$1.95$ $1.327$ $0.74$ $0.033$ $2.29$ $0.080$
$2.0$ $1.321$ $0.75$ $0.033$ $2.67$ $0.080$
: Results for $M_*^2$ and plasma $\beta$ (with errors) for various values of $\gamma$
\[gam\_table\]
Discussion
==========
In this paper, using Equation \[dendist\] we found that for the range $\gamma=1.4$-$1.7$ the observed prominence has a plasma $\beta$ of $\beta=0.47 \pm 0.079$ to $1.13\pm 0.080$. Using Equation \[dendist\] and the analysis presented in this paper, it is possible to make some predictions on the nature of the density enhancement at the head of a plume:
- The larger the value of $M_*$ (i.e. the larger $\beta$ and $M_{\infty}$ are), the larger the increase in density at the head of the plume. This is a simple consequence of the greater perturbation from the incompressible state the system receives.
- The larger the rising plume, the greater the thickness of the region at the head of the plume over which the density increases. This has important implications relating to which plumes this mechanism can be accurately applied.
It is clear from Table \[gam\_table\] that $\gamma$ presents the biggest uncertainty relating to the value of the plasma $\beta$ of the prominence. It may be expected that by looking at Equation \[dendist\] the fit may provide some constraints on the value of $\gamma$, but the $\chi^2$ of the fit is almost the same for each value of $\gamma$ therefore statistically it is not possible to do this. Separate methods that can constrain the value of $\gamma$ would work well in conjunction with this method to determine the plasma $\beta$ of a prominence.
The compressions that are under investigation here happen on timescales of approximately $1$Mm/$10$kms$^{-1}\sim 100$s, which is shorter than the radiative cooling timescales of the prominence ($\sim 1000$s), so the compression should be well modelled as adiabatic. Currently there is no model to describe accurately how the intensity of prominences changes under adiabatic compression. With this in mind, we will only use the column density, and ignore the effect of heating. We can at least be confident that the heating will not move the prominence plasma out of the observable temperature range, as the temperature increase is given as $T'=\rho'^{\gamma -1}\sim \rho'^{2/3}$ when $\gamma=5/3$. Therefore, an increase of the density of $\sim1.4$ (which is approximately what we find in this paper), would increase the temperature from $8000$K to $10^4$K which is still in the emission range for Ca II H $\sim 5000$-$10^4$K so would not move the emission out of the passband. A model of the intensity change due to the increase in temperature under these conditions would increase the accuracy of the calculation from plasma $\beta$.
The value for plasma $\beta$ found here has a strong dependence on the hydrodynamic Mach number as the values found were not small enough to make it unimportant. Therefore, if downflows of prominence material interact with the head of the plume, they would increase the $M_{\infty}$, which would reduce the value of the plasma $\beta$. In section \[MACH\] we assume a static prominence, but if we use a prominence background where the material is falling at sound-speed [based on the downflows observed by @Chae2010], we can calculate a new value for the plasma $\beta$ of $\beta\sim0.14$ for $\gamma=1.65$. Figure \[timedist\] does not show any obvious downflows interacting with the head of the plume, so the assumption of $\sim11$kms$^{-1}$ downflows is probably a huge over estimate, but this calculation is shown to highlight the potential importance of downflows when using this method to calculate the plasma $\beta$. Therefore, when downflows are not included in a calculation, as is the case in this paper, then the plasma $\beta$ value should always be viewed as an upper limit.
Another point that should be noted is the assumption of an irrotational flow as used in Equation \[irrot\]. The most important point is that the plume is rising through the prominence material, pushing it out of the way as it rises. In a compressible medium with finite Reynolds number, the downstream flow (flow behind the plume head) would develop turbulence and a wake, breaking the irrotational assumption. However, the compression that is being modelled is at the head of the plume, where the flows are much less complex. There are also observations that suggest prominences are weakly turbulent [@LEO2012], which would imply that to a some extent the irrotational assumption is also broken upstream of the plume. But it can be expected that to first order the properties of the mean flow are not affected by the turbulence, but the turbulence would be modified by the mean flow potentially changing the characteristics of the turbulence downstream of the plume head, i.e. turbulence distortion. To summarise, the assumption used is broken downstream, but holds to first order in the area we are interested in. Therefore, even though we couldn’t use this to describe all the flows associated with the prominence plumes, it does apply for the area that is of interest allowing the plasma $\beta$ to be determined.
One interesting point to note is that we have equations that define the velocity around the plume head, as analysed in Figure \[plumeheadvel\]. Therefore, analysis of the velocities at the observed plume head (either through Doppler shift analysis or the cork-tracking used in @BERG2010 and @Chae2010 will allow the projection of the plume on the plane-of-sky to be determined. This would be important as it would remove some uncertainty in the width of the plume head, making the parameters of the fitted ellipse more accurate. With this information, it should be possible to determine the prominence’s plasma $\beta$ with greater confidence and also, through application of knowledge of the 3D magnetic Rayleigh-Taylor instability, estimate the direction of the magnetic field in the prominence. We plan to present this work in a future paper.
It should be noted that [@CAR1996] studied a similar geometry through numerical simulations to investigate the movement of a magnetic cloud through a magnetised atmosphere in relation to CME propagation. Therefore, the method presented in this paper may be applicable to more MHD phenomena than just prominences, an investigation into what phenomenon this can be applied to may open up some very rewarding research areas.
Hinode is a Japanese mission developed and launched by ISAS/JAXA, with NAOJ as domestic partner and NASA and STFC (UK) as international partners. It is operated by these agencies in co-operation with ESA and NSC (Norway). The authors would like to thank the staff and students of Kwasan and Hida observatories for their support and comments. Special mention goes to Drs. H. Watanabe and H. Isobe, as well as Mr. K. Hiroi of Dept. Astronomy, Kyoto University, Dr. M. Cheung and Dr. S. Gunar. We would also like to thank the anonymous referee, whose comments helped to greatly improve the clarity of the manuscript. This work was supported in part by the Grant-in-Aid for the Global COE program “The Next Generation of Physics, Spun from Universality and Emergence” from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan. Part of this work was performed by AH during a visit to the Inter-University Centre for Astronomy and Astrophysics.
Anzer, U., & Heinzel, P. 2007, , 467, 1285
Arregui, I., Oliver, R., & Ballester, J. L. 2012, Living Reviews in Solar Physics, 9, 2
Aulanier, G., & D[é]{}moulin, P. 2003, , 402, 769
Berger, T. E., et al. 2008, , 676, L89
Berger, T. E., et al. 2010, , 716, 1288
Berger, T., et al. 2011, , 472, 197
Cargill, P. J., Chen, J., Spicer, D. S., & Zalesak, S. T. 1996, , 101, 4855
Chae, J. 2010, , 714, 618
Dud[í]{}k, J., Aulanier, G., Schmieder, B., Bommier, V., & Roudier, T. 2008, , 248, 29
Engvold, O. 1981, , 70, 315
Hillier, A., Isobe, H., Shibata, K., & Berger, T. 2011b, , 736, L1
Hillier, A., Berger, T., Isobe, H., & Shibata, K. 2012, , 746, 120
Hirayama, T. 1986, NASA Conference Publication, 2442, 149
Isobe, H., Miyagoshi, T., Shibata, K., & Yokoyama, T. 2005, , 434, 478
Isobe, H., & Tripathi, D. 2006a, , 449, L17
Isobe, H., Miyagoshi, T., Shibata, K., & Yokoyama, T. 2006b, , 58, 423
Kippenhahn, R., & Schl[ü]{}ter, A. 1957, , 43, 36
Kosugi, T., et al. 2007, , 243, 3
Labrosse, N., Heinzel, P., Vial, J.-C., Kucera, T., Parenti, S., Gun[á]{}r, S., Schmieder, B., & Kilper, G. 2010, , 151, 243
Leonardis, E., Chapman, S. C., & Foullon, C. 2012, , 745, 185
Lerche, I., & Low, B. C. 1980, , 67, 229
Leroy, J. L. 1989, Dynamics and Structure of Quiescent Solar Prominences, 150, 77
Lin, Y., Engvold, O., Rouppe van der Voort, L. H. M., & van Noort, M. 2007, , 246, 65
Lin, Y., Soler, R., Engvold, O., et al. 2009, , 704, 870
Mackay, D. H., Karpen, J. T., Ballester, J. L., Schmieder, B., & Aulanier, G. 2010, , 151, 333
Pint[é]{}r, B., Jain, R., Tripathi, D., & Isobe, H. 2008, , 680, 1560
Priest, E. R. 1982, Dordrecht, Holland ; Boston : D. Reidel Pub. Co. ; Hingham,, 74P
Ryutova, M., Berger, T., Frank, Z., Tarbell, T., & Title, A. 2010, , 170
Stone, J. M., & Gardiner, T. 2007, , 671, 1726
Tandberg-Hanssen, E. 1995, Astrophysics and Space Science Library, 199,
Tripathi, D., Isobe, H., & Jain, R. 2009, , 149, 283
Tsuneta, S., et al. 2008, , 249, 167
van Dyke, M. 1975, NASA STI/Recon Technical Report A, 75, 46926
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Ming Li
- and Alex Kovner
title: ' JIMWLK Evolution, Lindblad Equation and Quantum-Classical Correspondence'
---
Introduction
============
This paper examines the status of the JIMWLK evolution equation [@Balitsky:1995ub; @JalilianMarian:1997jx; @JalilianMarian:1997gr; @JalilianMarian:1997dw; @Kovner:2000pt; @Iancu:2000hn; @Ferreiro:2001qy; @Mueller:2001uk] in relation to the effective density matrix of a high energy hadronic system. We are motivated to address this question by the discussion in a recent paper [@Armesto:2019mna] which suggested an extension of JIMWLK framework to include a wider set of observables other than just color charge density $j^a(x)$ in the hadronic wave function. The starting point of [@Armesto:2019mna] is the interpretation of JIMWLK evolution equation as the equation for diagonal matrix elements of the density matrix in the color charge density basis. Although this interpretation is natural when the color charge density is large, it is not quite clear how to formalize it for low density, since in this regime the commutator of the color charge density operators is not negligible and a basis in which all components of $j^a(x)$ are diagonal obviously does not exist. On the other hand, as shown a while ago [@Kovner:2005uw] the calculations of averages in this regime as well can be formulated in terms of the functional integral over classical fields $j^a(x)$, which suggests that perhaps such interpretation albeit possibly modified, can be put forward after all.
An interesting suggestion of [@Armesto:2019mna] is that the rapidity evolution of the generalized CGC density matrix is of the Lindblad type [@Gorini:1975nb],[@Lindblad:1975ef]. This type of evolution is very general in quantum mechanical systems where one follows only part of the degrees of freedom by reducing the density matrix over the “environment" (the unobserved degrees of freedom in the Hilbert space). If the “environment" degrees of freedom have only short range correlation in time, the dynamics of the observed part of the system is Markovian and is therefore governed by a differential equation. The Lindblad form of such evolution is the most general one that preserves the properties of the density matrix stemming from its probabilistic nature (normalization and positivity of all eigenvalues). Although Lindblad equation naturally arises in [*time evolution*]{} of quantum systems, JIMWLK evolution is of a somewhat different nature. It describes the change of the system with rapidity (or energy) but not in time. It is thus not obvious whether one should expect Lindblad form to be generic in this context and if yes, under what conditions.
In this paper we try to address these questions. We arrive at two basic results. First, we show that JIMWLK evolution can indeed be understood as evolution of a density matrix. Within the JIMWLK framework however, the density matrix is not generic, but is rather assumed to depend only on the operators $\hat j^a(x)$ which satisfy the standard $SU(N)$ commutation relations. The fact that $\hat\rho$ depends only on the generators of the $SU(N)$ group means that it has a quasi diagonal form - i.e. it does not connect states belonging to different representations of $SU(N)$. It is in this sense that the reduced density matrix is (almost) diagonal even if the commutators of $j^a$ cannot be neglected. The consequence of this strong assumption on the nature of the density matrix is that the states in the Hilbert space of the reduced system are completely specified by their $SU(N)$ transformation properties at every transverse position $x$, and therefore in the reduced space one loses track of the gluon longitudinal momentum as well as polarization.
Second, we show that on this Hilbert space the JIMWLK evolution is indeed equivalent to Lindblad type equation for this restricted set of density matrices. The same applies to the so called KLWMIJ evolution which describes the dilute regime. In analogy with time evolution of quantum mechanical systems, the Lindblad equation arises in the situation when the correlations in the “unobserved" part of the system are short range in rapidity. However we also argue that in general (i.e. away from the dense - JIMWLK and dilute - KLWMIJ limits) the high energy evolution is unlikely to be of Lindblad type. This follows from certain general properties of our derivation of the evolution of the density matrix based on the calculation of the CGC wave function presented in [@Altinoluk:2009je] . Although the calculations of [@Altinoluk:2009je] are strictly valid only in the aforementioned limits, the general features of the derivation are expected to be more universal. The reason that the Lindblad form is not expected to arise, is that in the high energy evolution framework, the rapidity does not just play the role of the evolution parameter, but also that of the label on the quantum states of the gluons which are integrated out. In this situation in general one does not expect the Lindblad form for the differential equation. Thus to ensure Lindblad form nontrivial conditions on dependence of the matrix elements on gluon rapidities have to be satisfied. We discuss this point in detail in Section IV.
Another result of this paper is the precise mathematical relation between the effective density matrix, and the “probability density function" $W[{\mathbf j}]$ that usually appears in the literature as the subject of JIMWLK evolution. We confirm that the quantum mechanical averaging with the density matrix $\hat\rho$ is mapped into the calculation of observables in terms of functional integral over [*classical*]{} fields $j^a(x)$ with the weight functional $W[{\mathbf{j}}]$, as indeed always done in the CGC literature. This functional integral must be regarded as an integral over the [*phase space*]{} variables of the classical system, and not its [*configuration space*]{} variables. This quantum-classical correspondence between the quantum density matrix and the classical functional of phase space variables $W[{\mathbf{j}}]$ is deeply related to the correspondence between the density matrix and Wigner function in ordinary quantum mechanics. In the context of high energy evolution we require a generalization of the original Wigner-Weyl correspondence[@Hillery:1983ms] since the phase space of the theory is spanned not by operators $q$ and $p$ which constitute the Heisenberg algebra, but rather by operators $j^a$ with the $SU(N)$ algebra. Nevertheless the basic correspondence involves mappings between quantum operators and Hilbert space on one side and classical quantities and phase space on the other side in the sense of Weyl’s correspondence rule. The weight functional $W[{\mathbf{j}}]$ is consequently identified as the Wigner functional [@Hillery:1983ms] and can indeed be considered as a quasi-probability distribution on the phase space.
The outline of this paper is the following. In Sec. II, we give a brief review of the Lindblad equation for density matrix of an open quantum system and a recap of the Hamiltonian formalism of CGC effective theory. In Sec. III we explain how to define the reduced CGC density matrix, and show that its rapidity evolution has the Kraus form, which is a general evolution that preserves probabilistic interpretation of a density matrix, but is not necessarily differential. In Sec. IV we derive the differential evolution of the density matrix using the analog of Markovian porperty, i.e. the fact that the correlation of the “environment" is short range in rapidity. We show that in the dilute (KLWMIJ) and dense (JIMWLK) limits the differential evolution equation is indeed of the Lindblad type. We also discuss the properties of the derivation which suggest that the standard Lindblad form is bound to be modified away from these limits. To be clear, in this paper we do not go beyond the original JIMWLK setup in the sense that we consider density matrices that only depend on color charge density operators, and thus presently our derivation does not extend to ideas put forward in [@Armesto:2019mna]. In Sec. V, we derive the explicit relation between the standard JIMWLK approach and the density matrix approach described in this paper. We show that the two are related by a variant of the Wigner-Weyl quantum-classical correspondence and spell out explicitly the correspondence rules which transform one setup into the other. The JIMWLK and KLWMIJ equations are then reproduced by mapping the Lindblad equation for the density matrix in the appropriate (dense and dilute) limits into the classical phase space. Finally Sec. VI contains a short discussion.
Review of Basics
================
Lindblad equation for open quantum systems
------------------------------------------
In this section we present a short review of Lindblad equation for open systems.
Lindblad equation is the most general Markovian and non-unitary evolution equation for density matrix of an open quantum system. This equation preserves the properties of density matrix: hermiticity, unit trace and positivity. Here we follow the heuristic discussions by Preskill [@Preskill:2019]. More physical derivations and applications can be found in the books [@Carmichael:1993; @Breuer:2007].
Consider a bipartite system involving two subsystems: the “observed system" and the “environment" with the Hamiltonians $\hat{H}_s$ and $\hat{H}_e$, respectively. The two subsystems interact via the Hamiltonian $\hat{H}_{se}$. The total density matrix of the complete system evolves according to the quantum Liouville equation $$\frac{d \hat{\rho}}{dt} = -i [\hat{H}, \hat{\rho}]$$ with $\hat{H} = \hat{H}_s + \hat{H}_e +\hat{H}_{se}$. Formally, the solution can be expressed as $$\hat{\rho}(t) = \hat{U}(t) \hat{\rho}(0) \hat{U}^{\dagger}(t)$$ with $\hat{U}(t) = e^{-i\hat{H} t}$. To obtain the density matrix of the observed subsystem after a finite time, one traces over the Hilbert space of the environment. Let us assume that the initial total density matrix is a direct product of the density matrices of the observed system and the environment $\hat{\rho}(0) = \hat{\rho}_s(0) \otimes \hat{\rho}_e(0) = \hat{\rho}_s(0) \otimes |0_e\rangle\langle 0_e|$. For simplicity let us take the environment to be initially in a pure state denoted by $|0_e\rangle$, which can be thought of as the ground state without loss of generality. The density matrix of the observed system is then expressed as $$\label{eq:kraus_rep}
\hat{\rho}_s(t) = \mathrm{Tr}_e \hat{\rho} (t) = \sum_{n}\langle n |\hat{U}(t) | 0_e\rangle \hat{\rho}_s (0) \langle 0_e| \hat{U}^{\dagger}(t) |n\rangle =\sum_n \hat{M}_n(t) \hat{\rho}_s(0) \hat{M}^{\dagger}_n(t).$$ Here $\left\{ |n\rangle \right\}$ represents a complete basis in the Hilbert space of the environment. The objects $\hat{M}_n(t) = \langle n |\hat{U}(t) | 0_e\rangle$, sometimes called superoperators, are operators on the Hilbert space of the observed system and govern the evolution of its density matrix. As far as the dynamics of the environment is considered, $\hat{M}_n(t)$ represents the transition amplitude for the environment, which is initially in the state $|0_e\rangle$, to be in the state $|n\rangle$ after a finite time $t$. They satisfy the property $\sum_n \hat{M}_n^{\dagger}(t)\hat{M}_n(t) =1$ following from the unitarity of $\hat{U}(t)$.
The time evolution of density matrix in Eq. has been expressed in an operator summation form which is also called a Kraus representation. It is easy to check that the Kraus representation preserves the hermiticity, unit trace and positivity of the density matrix. It is believed that any reasonable time evolution of density matrices should have a Kraus representation.
The general Kraus representation Eq. does not have the form of a differential equation for the evolution of the density matrix. It is only under the Markovian approximation that an equivalent expression in terms of a differential equation becomes possible. The Markovian approximation holds if the typical correlation time between the environment degrees of freedom $t_{corr}$ is shorter than the typical inverse frequency of the observed system $\Delta t_s$, which is of the order of the relevant “discretization" time step for approximate differential time evolution. If this is the case the state of environment is only affected by the state of the observed system at the particular time of observation (measured with accuracy $\Delta t_s$), and thus the back reaction - the effect of the environment on the observed system is local in time. We note that this is the typical Born-Oppenheimer situation, when the environment is associated with fast degrees of freedom, while the observed system is relatively slow. In the opposite regime it is clear that local (differential) time evolution is impossible, since the backreaction of the environment on the system will depend on the state of the system at some past time.
In Markovian regime one then proceeds as follows. For an infinitesimal period of time, only terms linear in $dt$ should be kept on the right hand side of Eq. . The superoperators for $n>0$, have the structure $$\label{amp}
\hat{M}_n(dt) = \sqrt{dt}\hat{L}_n, \quad n>0$$ The argument here is that $\hat{M}^\dagger_n(t)\hat{M}_n(t)$ is the probability for the environment to “jump” to the state $n$ during the time $t$. For small enough $t$ (but such that $t>t_{corr}$) this probability should grow linearly with $t$. The operators $\hat L_n$ are called Lindblad operators or jump operators as they involve transitions of the environment to different states after an infinitesimal time.
The remaining superoperator has the form $$\hat{M}_0(dt) = 1 + (-i \hat{H}_s + \hat{K}) dt$$ with $H_s$ and $K$ being Hermitian. This is the transition amplitude for the environment to be in its original state after an infinitesimal time and should be linear in time for small enough times. The operator $\hat{K}$ is related to the wave function renormalization effect and $\hat{H}_s$ governs the unitary evolution of the system without causing any changes to the environment.
The Kraus normalization condition $\sum_n\hat{M}^{\dagger}_n(dt) \hat{M}_n (dt)=1$ relates the wavefunction renormalization operator $\hat{K}$ to the jump operators by $$\hat{K} = -\frac{1}{2}\sum_{n>0} \hat{L}^{\dagger}_n \hat{L}_n\, .$$ Taking the limit $dt\rightarrow 0$, the Kraus representation then becomes an differential equation $$\frac{d\hat{\rho}_s}{dt} = -i[\hat{H}_s, \hat{\rho}_s] + \sum_{n>0} \left( \hat{L}_n\hat{\rho}_s \hat{L}^{\dagger}_n -\frac{1}{2}\hat{L}_n^{\dagger}\hat{L}_n \hat{\rho}_s-\frac{1}{2} \hat{\rho}_s \hat{L}_n^{\dagger}\hat{L}_n \right)\, .$$ This is the Lindblad equation, or sometimes known as Gorini-Kossakowski-Lindblad-Sudarshan master equation [@Gorini:1975nb; @Lindblad:1975ef].
The soft gluon vacuum and the CGC
---------------------------------
We now review the derivation of the high energy evolution[@Kovchegov:2012mbw] . There exist two equivalent approaches to the derivation of the CGC effective theory. One is based on the Lagrangian formalism [@McLerran:1993ni; @McLerran:1993ka; @JalilianMarian:1997jx; @JalilianMarian:1997gr; @JalilianMarian:1997dw; @Iancu:2000hn; @Ferreiro:2001qy] and the other on the Hamiltonian formalism [@Kovner:2005pe; @Kovner:2007zu]. We briefly review the Hamiltonian formalism as it will be the starting point for deriving the Lindblad equation for the CGC density matrix.
The derivation of the JIMWLK evolution equation starts with establishing the ground state wave function of soft gluon modes in the background of more energetic gluons which are described by a color charge density field.
In the light cone gauge $A^+=0$, the Hamiltonian of the pure gluonic sector of QCD is $$H = \int dx^-d\mathbf{x}_{\perp} \left(\frac{1}{2} \Pi^-_a(x^-, \mathbf{x}_{\perp}) \Pi^-_a(x^-,\mathbf{x}_{\perp}) + \frac{1}{4}F_{ij}^a(x^-,\mathbf{x}_{\perp}) F^a_{ij}(x^-,\mathbf{x}_{\perp})\right)$$ with the chromoelectric and chromomagnetic parts being $$\begin{split}
&\Pi^-_a(x^-, \mathbf{x}_{\perp}) = \partial_-A^-_a(x^-,\mathbf{x}_{\perp})=\frac{1}{\partial_-}\left( D_i^{ab} \partial_-A_i^b(x^-,\mathbf{x}_{\perp})\right)\, ,\\
&F_{ij}^a(x^-,\mathbf{x}_{\perp}) = \partial_iA_j^a(x^-,\mathbf{x}_{\perp}) -\partial_j A_i(x^-,\mathbf{x}_{\perp}) -gf^{abc}A_i^b(x^-,\mathbf{x}_{\perp})A_j^b(x^-,\mathbf{x}_{\perp})
\end{split}$$ The covariant derivative is defined as $D^{ab}_i = \partial_i\delta^{ab} -gf^{acb} A_i^c$ and $\partial_- = \partial/\partial x^-$ is the longitudinal spatial derivative. The $1/\partial_-$ operator in the expression of the chromoelectric field has to be regularized as it contains a singularity at vanishing longitudinal momentum, $k^+=0$. This singularity is ultimately related to the zero mode in the $A_i^a(x^-,\mathbf{x}_{\perp})$ fields and is regulated by imposing a residual gauge fixing condition. We choose the residual gauge fixing $$\label{eq:gauge_fixing}
\partial_i A_i^a(x^-\rightarrow-\infty) =0\,$$ One separates the gluonic degrees of freedom imposing a longitudinal momentum separation scale $\Lambda^+$. In the high energy limit, the dominant interaction between soft gluons ($k^+<\Lambda^+$) and hard gluons ($k^+>\Lambda^+$) has the form of eikonal coupling $A^-_a J_a^+$ with $J^+_a$ representing the color charge density of the hard gluons and $A^-_a$ representing soft gluons. This interaction term emerges from the chromoelectric part of the Hamiltonian and involves the specific expressions $J^+_a = -gf^{abc}A_i^b\partial_-A_i^c$ and $A^-_a = \frac{1}{\partial_-}\Pi^-_a$. Furthermore, as far as soft gluons are concerned, the hard gluon dynamics can be viewed as frozen in time so that the color current $J^+_a \equiv J^+_a(x^-,\mathbf{x}_{\perp})$ is time independent at the lowest order. All in all, the Hamiltonian for the soft gluonic modes becomes $$\label{eq:cgc_hamiltonian}
H_{CGC} = \int dx^-d\mathbf{x}_{\perp} \left(\frac{1}{2} (\Pi^-_a(x^-, \mathbf{x}_{\perp}) + \frac{1}{\partial_-} J^+_a)^2 + \frac{1}{4}F_{ij}^a(x^-,\mathbf{x}_{\perp}) F^a_{ij}(x^-,\mathbf{x}_{\perp})\right)$$ Canonical quantization is implemented by promoting the *normal modes* of the full $A_i^a$ fields to operators and imposing the equal (light cone) time commutation relation $$\label{eq:commutation_fields}
[\hat{A}_i^a(x^-, \mathbf{x}_{\perp}), \hat{A}_j^b(y^-,\mathbf{y}_{\perp})] = -\frac{i}{2} \epsilon(x^- - y^-) \delta^{ab}\delta_{ij}\delta(\mathbf{x}_{\perp} - \mathbf{y}_{\perp})$$ with the sign function defined as $\epsilon(x) = \frac{1}{2}(\Theta(x) - \Theta(-x))$. In terms of the canonical creation and annihilation operators, the normal modes $\hat{A}_i^a$ have the expansion $$\hat{A}_i^a(x^-, \mathbf{x}_{\perp}) = \int_0^{+\infty} \frac{dk^+}{2\pi} \frac{1}{\sqrt{2k^+}} \left( \hat{a}_i^a(k^+, \mathbf{x}_{\perp}) e^{-ik^+x^-} + \hat{a}^{\dagger a }_i(k^+, \mathbf{x}_{\perp}) e^{ik^+x^-}\right)$$ with $$\left[ \hat{a}_i^a(k^+,\mathbf{x}_{\perp}), \hat{a}_j^{\dagger b}(p^+, \mathbf{y}_{\perp})\right] = (2\pi) \delta^{ab}\delta_{ij}\delta(k^+-p^+)\delta(\mathbf{x}_{\perp} -\mathbf{y}_{\perp})\, .$$ The color charges in the leading order are taken to have the extreme Lorentz contracted form $J^+_a(x^-,\mathbf{x}_{\perp}) = \delta(x^-)j^a(\mathbf{x}_{\perp})$ with the transverse color charge density $$\label{eq:color_current}
\hat{ j}^a(\mathbf{x}_{\perp}) = igf^{abc}\int_{k^+>\Lambda^+} \frac{dk^+}{2\pi} \hat{a}_i^{\dagger b}(k^+,\mathbf{x}_{\perp}) \hat{a}_i^c(k^+,\mathbf{x}_{\perp})\,$$ The components of color charge satisfy the commutation relations of the $SU(N)$ algebra $$\label{eq:commutation_current}
\left[\hat{j}^a(\mathbf{x}_{\perp}),\hat{j}^b(\mathbf{y}_{\perp}) \right] = igf^{abc} \hat{j}^c(\mathbf{x}_{\perp})\delta(\mathbf{x}_{\perp} -\mathbf{y}_{\perp})\, .$$ The Hamiltonian system Eqs. , , together with the commutation relations Eqs. , constitute the starting point for the derivation of the CGC effective theory.
The first goal is to find the ground state wave function of the soft glue. This is in general a very complicated problem, but is simplifies in two parametric regimes. One interesting regime is when the color charge density is small $j^a\sim g$ (dilute limit). Here one can treat the interaction with color charges perturbatively. The other regime is the dense limit where the color charge is parametrically large $j^a \sim \frac{1}{g}$. Here the simplification is that the commutator of the color charges is (almost) negligible and they can be treated as (almost) classical fields.
In the dilute limit, the ground state wave function can be found by a direct perturbative calculation. The resulting vacuum wave function can be written as $$|\psi_0\rangle = \hat{\mathcal{C}} |0\rangle$$ with the coherent operator $$\label{eq:coherent_operator}
\hat{\mathcal{C}} = \mathrm{Exp}\left\{i\int d\mathbf{x}_{\perp} b_i^a(\mathbf{x}_{\perp}) \int^{\Lambda^+}_{\Lambda^+e^{-\Delta y}} \frac{dk^+}{\pi |k^+|^{1/2}}\left(\hat{a}_i^{\dagger a}(k^+, \mathbf{x}_{\perp}) + \hat{a}_i^{a}(k^+, \mathbf{x}_{\perp})\right)\right\}$$ where $E$ is the energy of the process.
Here $b_i^a(\mathbf{x}_{\perp})$ satisfy the equations $$\begin{split}
& \partial_i b_i^a(\mathbf{x}_{\perp}) = j^a(\mathbf{x}_{\perp})\, ,\\
&\partial_i b_j^a(\mathbf{x}_{\perp}) - \partial_j b_i^a(\mathbf{x}_{\perp}) -gf^{abc} b_i^b(\mathbf{x}_{\perp}) b_j^c(\mathbf{x}_{\perp}) =0 \, .\\
\end{split}$$ or, in the dilute regime $$b_i^a(\mathbf{x}_{\perp})=\int d^2y \frac{\partial_i}{\partial^2}(x,y)j^a(y)$$ In the dense limit, similar analysis applies except now $b_i^a(\mathbf{x}_{\perp})\sim 1/g$ and additional order $\mathcal{O}(1)$ quantum fluctuations on top of the $b_i^a$ fields need to be considered. One can still use perturbative expansion in $g$, but resumming terms of order $gb$. In the leading order the Hamiltonian is diagonalized by a nontrivial Bogoliubov transformation. The detailed analysis appears in [@Kovner:2007zu]. The resulting ground state wavefunction is $$|\psi_0\rangle = \hat{\mathcal{C}} \hat{\mathcal{B}} |0\rangle \, .$$ The additional Bogoliubov operator $\hat{\mathcal{B}}$ can be formally expressed as $$\label{eq:bogoliubov_operator}
\hat{\mathcal{B}} = \mathrm{Exp}\left\{ \hat{a}_{\alpha}^{\dagger} \Lambda_{\alpha\beta} \hat{a}_{\beta}^{\dagger} +\hat{a}_{\alpha}\Lambda_{\alpha\beta}^{\ast} \hat{a}_{\beta} \right\}\, .$$ Here $\alpha, \beta$ represent all the possible indices (color, spatial coordinates, polarization, and longitudinal momentum which varies between $\Lambda^+e^{-\Delta y}$ and $\Lambda^+$). The explicit expression of the symmetric matrix $M_{\alpha\beta}$ is not available, however, the action of the Bogoliubov operator on the fundamental degrees of freedom $\hat{A}_i^a$ and $\hat{j}^a$ have been derived. The nontrivial structure of the soft gluon ground state leads to appearance of induced color charge density due to the soft gluons modes. This additional color charge density serves as an additional source for even softer gluons which arise in the evolution to even higher rapidities. This is the basic physics of the high energy evolution.
The Reduced CGC Density Matrix and Its Evolution
================================================
Having found the vacuum of the soft gluons, we can now address the evolution at high energy. We take here a different perspective on this derivation than given in the literature, and discuss the evolution from the point of view of quantum density matrix.
Given that we have separated our degrees of freedom into soft and hard gluons, we can view our system naturally as bipartite. At some initial rapidity, the soft gluons are in the perturbative vacuum state, and thus the total density matrix is separable $$\hat{\rho} = \hat\rho_v \otimes |0\rangle \langle 0|$$ where the density matrix $\hat\rho_v$ is an operator on the hard gluon Hilbert space.
The assumption inherent in the derivation of the JIMWLK equation is that the only relevant degrees of freedom on this Hilbert space are components of the color charge density $\hat j^a({\mathbf x}_\perp)$. This is a crucial assumption. If the valence Hilbert space could be factorized into a direct product of the space spanned by $\hat j^a({\mathbf x}_\perp)$ and its complement, reducing over the complement would rigorously define $\hat\rho_v[{\mathbf j}]$. However the full Hilbert space of the valence modes does not have such a direct product structure. It is thus not clear whether a well defined mathematical procedure of “integrating out" exists which may reduce the density matrix so that in general it depends only on $\hat j^a({\mathbf x}_\perp)$. Nevertheless one can simply assume that at initial rapidity the density matrix indeed has such a form. It is then true (as we will see below) that this form persists throughout the evolution to higher rapidities. We will thus abide by this assumption and will treat $\hat\rho_v$ as an operator that depends only on $\hat j^a({\mathbf x}_\perp)$.
After boosting the system by a finite rapidity $\Delta y$, the total density matrix changes due to the emission of soft gluons into the newly opened rapidity interval. $$\hat{\rho}(\Delta y) = \hat{\Omega} |0\rangle \hat{\rho}_v \langle 0|\hat{\Omega}^{\dagger}\, .$$ The gluon emission operator as discussed above can be written as $$\hat{\Omega} \equiv \Omega [\hat{j}^a, \hat{a}_i^{a \dagger}, \hat{a}_i^a; \Delta y] = \hat{\mathcal{C}}\hat{\mathcal{B}}$$ with $\hat{\mathcal{C}}$ and $\hat{\mathcal{B}}$ defined in eqs. and , respectively. This form applies both for the dilute and dense regime of the evolution. Note that $\hat{\Omega} $ depends on the soft gluon creation and annihilation operators as well as the color charge density operator. While the $\hat{a}_i^{a \dagger}, \hat{a}_i^a$ act on the soft vacuum state $|0\rangle$, the $\hat{j}^a$ acts on the valence (hard) density matrix $\hat\rho_v$. Dependence on $\Delta y$ of $\hat{\Omega}$ is crucial in obtaining the evolution equation. This point will be elaborated in the following.
Our next goal is to derive the reduced density matrix by tracing over the “environment" degrees of freedom. The purpose of this reduction of the Hilbert space is to integrate out all the additional degrees of freedom associated with soft gluons that emerged after boosting the wave function, [*except*]{} the additional color charge density that they contribute. The reason for this exception is, that in the next step in the evolution the even softer gluons will couple to the total color charge density, including that due to gluons in the rapidity interval between $y$ and $y+\Delta y$. Our current soft gluons give a nontrivial contribution to this charge density, and we have to keep this extra contribution explicitly, rather than integrate it out.
Defining the charge shift operator
----------------------------------
Put in different words, we are interested in a general set of observables that depend on rapidity integrated color charge density. Before evolution those are averages of the form $$\langle O(\hat j^a)\rangle={\rm Tr}[O(\hat j^a)\hat\rho_v]$$ while after a step $\Delta y$ of the evolution $$\label{oav}
\langle O(\hat j^a+\hat j^a_{\mathrm{soft}})\rangle={\rm Tr}[O(\hat j^a+\hat j^a_{\mathrm{soft}})\hat\rho(\Delta y)]$$ Here $\hat{j}^a_{\mathrm{soft}}(\mathbf{x}_{\perp})$ has the explicit expression eq. with the longitudinal momentum integration restricted in the rapidity range $\Delta y$.
It is thus clear that we should not simply reduce the density matrix over the Hilbert space of soft gluons, but “partially" trace over the soft gluons integrating out all degrees of freedom except the color charge density. To facilitate this partial tracing over soft gluons, we introduce the operator $\hat{R}$, which is defined by its action on $\hat{j}^a$, $$\label{rj}
\hat{R}^{\dagger} \hat{j}^a(\mathbf{x}_{\perp})\hat{R} = \hat{j}^a(\mathbf{x}_{\perp}) + \hat{j}^a_{\mathrm{soft}}(\mathbf{x}_{\perp})\, .$$ so that $$\label{oj}
\hat{R}^{\dagger}O( \hat{j}^a(\mathbf{x}_{\perp}))\hat{R} = O(\hat{j}^a(\mathbf{x}_{\perp}) + \hat{j}^a_{\mathrm{soft}}(\mathbf{x}_{\perp}))\, .$$ for any operator $O$. It may not be obvious that $\hat R$ can be properly defined as an operator on the Hilbert space, given that different components of $\hat{j}^a(x)$ are noncommuting operators. As we show now, this nevertheless is the case.
Let us introduce the operator $\hat{\Phi}^a(\mathbf{x}_\perp)$ via $$\label{R}\hat{R} = \mathrm{Exp}\{ - i\int d^2\mathrm{x}_{\perp} \hat{j}^a_{\mathrm{soft}}(\mathbf{x}_{\perp})\hat{\Phi}^a(\mathbf{x}_{\perp})\}$$
We will look for $\hat{\Phi}^a$ (we omit the transverse coordinate dependence for simplicity) as a set of operators acting on the same Hilbert space as $\hat j^a$ satisfying the following commutation relations $$\label{com}
\begin{split}
&[\hat{\Phi}^a, \hat{\Phi}^b] =0\, ,\\
&[\hat{\Phi}^a, \hat{j}^b] = M^{ab}(\hat\Phi)\, \\
\end{split}$$ with $M$ chosen to satisfy the requirement $$\label{eq:shift_requirement}
\mathrm{exp}\left\{i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a\right\} \hat{j}^e\, \mathrm{exp}\left\{-i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a\right\} = \hat{j}^e + \hat{j}^e_{\mathrm{soft}}\, .$$ In calculating the action of $\hat R$ we assume that the operators $\hat j^a$ satisfy $SU(N)$ algebra, and so do the operators $\hat j^a_{\mathrm{soft}}$, while the two set of operators commute with each other.
We use the Baker-Hausdorff formula $$e^X Y e^{-X} = Y + [X, Y] + \frac{1}{2!}[X, [X, Y]] + \frac{1}{3!}[X,[X,[X,Y]]]+\ldots + \frac{1}{n!}[X, [X, [\ldots [X, Y]\ldots]]] + \ldots$$ With the commutation relations eq.(\[com\]) we have (for adjoint representation $-if^{abc} = T^a_{bc}$) $$[i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] = i\hat{j}^a_{\mathrm{soft}} M^{ae}(\hat\Phi)$$ $$%\begin{split}
\frac{1}{2!}[i\hat{j}^b_{\mathrm{soft}} \hat{\Phi}^b, [i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] ]= \frac{1}{2!} \, i\hat{j}^a_{\mathrm{soft}} \left(igT^b \hat{\Phi}^b M(\hat\Phi)\right)_{ae}
%
%&= \frac{1}{2!} [\hat{j}^b_{\mathrm{soft}}, \hat{j}^a_{\mathrm{soft}}] i^2 \hat{\Phi}^b M^{ae}(\Phi) = \frac{1}{2!} igf^{bac} \hat{j}^{c}_{\mathrm{soft}} i^2 \hat{\Phi}^b M^{ae}(\Phi) \\
%&= \frac{1}{2!}\, i\hat{j}^c_{\mathrm{soft}} \left(igT^b_{ca} \hat{\Phi}^b \right)M^{ae}(\Phi)\\
%&\\
%\end{split}$$ $$\frac{1}{3!}[i\hat{j}^c_{\mathrm{soft}} \hat{\Phi}^c,[i\hat{j}^b_{\mathrm{soft}} \hat{\Phi}^b, [i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] ] ] = \frac{1}{3!} i\hat{j}^a_{\mathrm{soft}} \left((igT^b \hat{\Phi}^b)^2 M(\hat\Phi)\right)_{ae}$$ Let us take the ansatz $$M^{ab}(\hat\Phi) = -i\sum_{n=0}^{\infty} c_n \left[\chi^n\right]_{ab}\, , \quad\mathrm{with}\,\,\, \chi = igT^b\hat{\Phi}^b\, .$$ Clearly, $c_0 =1$ follows from the requirement eq. . This requirement further imposes the constraint $$i + \sum_{k=0}^{\infty} \frac{1}{(k+1)!} \left[\chi^k M(\chi)\right]^{ab}=0$$ which after substituting the ansatz for $M(\chi)$ becomes $$1= \sum_{k=0}^{\infty}\sum_{m=0}^{\infty}\frac{c_m}{(k+1)!} \chi^{k+m}_{ab}$$ which is equivalent to the following recursive relations $$\label{eq:recursive_relations}
c_N = -\sum_{m=0}^{N-1} \frac{c_m}{(N-m+1)!} \, , \quad \mathrm{with}\,\, c_0=1.$$ These relations are satisfied by $$\label{eq:M_expression}
M^{ab}(\chi) = -i \left[\frac{\chi}{2} \coth{\frac{\chi}{2}} - \frac{\chi}{2}\right]^{ab}\, .$$ One can check explicitly that Taylor expansion of eq. in $\chi$ reproduces all the coefficients calculated using the recursive relations in eq. .
Once the function $M$ in Eq. has been determined, the algebra of $\hat j^a$ and $\hat\Phi^a$ is completely defined.
We note that in order for this algebra to be consistent, the commutators have to satisfy Jacoby identity $$[[\hat{\Phi}^a, \hat{j}^b], \hat{j}^c] + [[\hat{j}^b, \hat{j}^c],\hat{\Phi}^a] + [[\hat{j}^c, \hat{\Phi}^a], \hat{j}^b] =0\, ,$$ which is equivalent to an additional constraint on $M$ $$\label{eq:consistence_check}
[M^{ab}, \hat{j}^c] - [M^{ac}, \hat{j}^b] = igf^{bcd} M^{ad}\, .$$ In Appendix A we verify that the Jacoby identity is in fact satisfied at least to fourth order in expansion of eq. in powers of $\hat\Phi$. Although we do not have a complete all order proof, one could in principle continue the order-by-order proof. We believe that the algebra eqs., is in fact consistent and we will continue our analysis under this assumption.
We have thus found the algebra of operators $\hat\Phi^a$ and $\hat j^a$ that implements eq.. Note that expansion of $M$ in powers of $\hat\Phi$ can be recast as formal expansion of $\hat\Phi$ in powers of $\frac{\delta}{\delta\hat j^a}$. Thus to leading order we have $\hat\Phi^a=-i\frac{\delta}{\delta \hat j^a}+...$. In the dense regime where the commutators of charge densities can be neglected, our operator $\hat R$ therefore reduces precisely to the shift operator $\exp \{-\int_{\mathrm{x}_\perp}j_{\mathrm{soft}}^a(\mathrm{x}_\perp)\frac{\delta}{\delta j^a(\mathrm{x}_\perp)}\}$ used extensively in the existing literature. The previous discussion puts its use also in the dilute regime on firm mathematical basis, provided the commutation relations of $\hat\Phi^a$ are modified according to eq..
The evolution
-------------
Having defined the charge density shift operator $\hat R$ we can write Eq.(\[oav\]) in the form $${\rm Tr}[O(\hat j^a+\hat j^a_{\mathrm{soft}})\hat\rho(\Delta y)]={\rm Tr}[\hat{R}^{\dagger}O( \hat{j}^a)\hat{R} \hat\rho(\Delta y)]={\rm Tr}[O( \hat{j}^a)\hat{R} \hat\rho(\Delta y)\hat{R}^{\dagger}]$$ The operator $\hat R$ in this expression can be understood as acting on the density matrix $\hat\rho_v$ rather than on the observable $O$. Using this form we can define the reduced density matrix which when traced with the operator $O(\hat j)$ gives the same result as $\hat\rho(\Delta y)$ traced with $O(\hat j+\hat j_{\mathrm{soft}})$. We thus define the evolved CGC reduced density matrix by tracing over soft gluons $$\label{eq:cgc_kraus_rep}
\hat\rho_v(\Delta y) = \mathrm{Tr}_s [\hat R\hat{\rho}(\Delta y) \hat R^\dagger]= \sum_n \langle n|\hat{R} \hat{\Omega} |0\rangle \hat\rho_v \langle 0|\hat{\Omega}^{\dagger}\hat{R}^{\dagger} |n\rangle = \sum_n \hat{M}_n \hat\rho_v \hat{M}^{\dagger}_n$$ with $\hat{M}_n = \langle n|\hat{R} \hat{\Omega} |0\rangle$. The complete basis $\{|n\rangle \}$ represents the Fock states in the soft gluon Hilbert space. This procedure technically is very similar to the standard reduction of the Hilbert space discussed in the previous section with $\hat\rho_v(\Delta y)$ playing the role of the reduced density matrix in a bipartite system.
When formulated in this way, the rapidity evolution of the CGC density matrix is formally very similar to the time evolution of the reduced density matrix of a bipartite system with the operator $\hat R\hat \Omega$ playing the role of the time evolution operator $\hat U$.
Eq. gives the change of density matrix in the form of a Kraus representation. As a consequence, $\hat\rho_v(\Delta y)$ has all the properties of a density matrix as long as $\hat\rho_v$ is a density matrix initially. Note that, $\sum_n \hat{M}_n^{\dagger} \hat{M}_n =1$ as both $\hat{\Omega}$ and $\hat{R}$ are unitary operators.
The Differential Form of the Evolution - the Lindblad Equation
==============================================================
To extract a differential equation from the Kraus representation, we need to evaluate the superoperators $\hat{M}_n$ and analyze their $\Delta y$ dependence. The calculations of $\hat{M}_n$ can be simplified by working in the Leading Logarithmic Approximation (LLA) so that only terms that are proportional to $\alpha_s \Delta y$ on the right hand side of eq. are kept.
The dilute limit
----------------
We start by considering the dilute limit, i.e. assume that parametrically $b_i^a \sim \mathcal{O}(g)$. In this regime the gluon emission operator is just the coherent operator and $\Lambda =0$. This is the so called KLWMIJ limit introduced in [@Kovner:2005nq] . $$\Omega= \mathrm{Exp}\left\{i\int d\mathbf{x}_{\perp} b_i^a(\mathbf{x}_{\perp}) \int \frac{d\eta}{2\pi}\sqrt{2}\left(\hat{a}_i^{\dagger a}(\eta, \mathbf{x}_{\perp}) + \hat{a}_i^{a}(\eta, \mathbf{x}_{\perp})\right)\right\}$$ Note that we have changed the integration variable from longitudinal momentum $k^+$ to rapidity $\eta$ and an explicit numerical factor $\sqrt{2}$ follows [@Altinoluk:2009je]. In this limit the dependence on $\Delta y$ becomes very transparent $$\hat{M}_n = \langle n | \hat{R}\hat{\Omega} |0\rangle=\mathrm{Exp}\left\{-\frac{\Delta y}{2\pi} b_{\alpha}b_{\alpha}\right\} \langle n| \hat{R}\, \mathrm{Exp}\left\{i\sqrt{2}b_{\alpha}\int\frac{d\eta}{2\pi} \hat{a}^{\dagger}_{\alpha}(\eta) \right\}|0\rangle \, .$$ Note that the operator $\hat R$ has a nontrivial action on the $n$-gluon state. It does not change the number of soft gluons in a Fock state but rotates their color indices according to its definition in eq.(\[R\]) $$\label{eq:R_on_aa_dagger}
\begin{split}
&\hat{R} \hat{a}_i^b(k^+, \mathbf{y}_{\perp}) \hat{R}^{\dagger} = [\hat{\mathcal{R}}(\mathbf{y}_{\perp})]_{bd} \, \hat{a}^d_i(k^+, \mathbf{y}_{\perp})\, ,\\
&\hat{R} \hat{a}_i^{b\dagger}(k^+, \mathbf{y}_{\perp}) \hat{R}^{\dagger} = [\hat{\mathcal{R}}(\mathbf{y}_{\perp})]_{bd} \, \hat{a}^{d\dagger}_i(k^+, \mathbf{y}_{\perp})\,
\end{split}$$ with $$\hat{\mathcal{R}}(\mathbf{y}_{\perp}) = e^{igT^a\hat{\Phi}^a(\mathbf{y}_{\perp})}$$
In the LLA we need to collect terms which contribute at order $O(\alpha_s)$ to the evolution. For the virtual term we have $$\hat{M}_0 = 1 -\frac{\Delta y}{2\pi} b_{\alpha}b_{\alpha} + \mathcal{O}(g^4)$$ and $$\hat{M}_0 \hat{\rho}_v \hat{M}^{\dagger}_0 = \hat{\rho}_v -\frac{\Delta y}{2\pi} \left( b_{\alpha}b_{\alpha} \hat{\rho}_v + \hat{\rho}_v b_{\alpha}b_{\alpha}\right) + \mathcal{O}(g^4)\, .$$ It is obvious that, for Fock states with even numbers of gluons, $\hat{M}_{2m}$ is at least of order $ \mathcal{O}(g^2)$ and thus will not contribute to the evolution at LLA. The same holds for $\hat{M}_{2m+1}$ associated with Fock states of odd numbers of gluons. The only exception is $\hat{M}_1$ related to the single gluon Fock state. For a one-gluon Fock state $|1_{\{\alpha_1, \mathbf{w}_1,\eta_1\}}\rangle = a_{\alpha_1}^{\dagger}(\mathbf{w}_1,\eta_1)|0\rangle$ with transverse position $\mathbf{w}_1$, rapidity $\eta_1$ and color index $\alpha_1$ we have, $$\hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}} = i\sqrt{2}b_{\alpha}(\mathbf{w}_1) \mathcal{R}_{\alpha \alpha_1}(\mathbf{w}_1)$$ Summing over all possible one-gluon Fock states, $$\int \frac{d\eta_1}{2\pi} \int d\mathbf{w}_{1} \,\hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}} \hat{\rho}_v \hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}}^{\dagger} = \frac{\Delta y}{\pi} \bar{b}_{\alpha} \hat{\rho}_v \bar{b}_{\alpha}$$ with $\bar{b}_{\alpha} = \mathcal{R}^{\dagger}_{\alpha\beta}b_{\beta}$ and again we used the compact notation with index $\alpha, \beta$ representing colors, transverse coordinates and polarizations. The evolution equation for the density matrix follows $$\label{lind}
\frac{d\hat{\rho}_v}{dy} = -\frac{1}{2\pi} (\bar{b}_{\alpha}\bar{b}_{\alpha} \hat{\rho}_v + \hat{\rho}_v \bar{b}_{\alpha}\bar{b}_{\alpha} - 2\bar{b}_{\alpha}\hat{\rho}_v \bar{b}_{\alpha} )=-\frac{1}{2\pi}\int_{x}[\bar b_a^i(x),[\bar b_a^i(x),\hat \rho_v]].$$ In this equation we have written the virtual terms in terms of $\bar b$ rather than $b$, since the unitary operator $R$ drops out of this expression anyway. This is the Lindblad equation for the CGC density matrix in the dilute limit.
Eq.(\[lind\]) is written in a somewhat convoluted form in terms of the operators $\bar b$, which contain the operator $\mathcal{R}$. It is perhaps worthwhile to make explicit the operational meaning of various factors of $\mathcal{R}$ in the right hand side of eq.(\[lind\]). As already mentioned, the virtual terms do not actually involve $\mathcal R$ since for unitary $\mathcal R$ $$\bar b_\alpha\bar b_\alpha=b_\alpha b_\alpha$$ As for the real term, we have (suppressing the transverse coordinate) $$\bar b_{\alpha}\hat{\rho}_v \bar{b}_{\alpha} =b_\gamma \mathcal{R}_{\gamma\alpha}\hat\rho_v[\hat j]\mathcal{R}^\dagger_{\alpha\beta}b_\beta=b_\gamma \left[\hat\rho_v[\hat j^a-gT^a]\right]_{\gamma\beta}b_\beta$$ where the last term is defined by Taylor expanding of $\hat\rho_v$, shifting the argument $\hat j^a$ in every term by the matrix $T^a$ and finally taking the $\gamma\beta$ matrix element of the the whole expression in all products of $T$’s that arise.
In the above explicit calculation, the LLA automatically picks up terms that are linear in $\Delta y$ thus making the extraction of a differential equation from the Kraus representation straightforward. Physically indeed we can understand this from the point of view of Markovian nature of the process. The variable analogous to time in the present discussion is rapidity. Thus the requirement of short range correlations in time of the “environment" in the CGC case translates into the requirement of short range in rapidity correlations for the soft gluons, which are integrated over. This is indeed the case. In the LLA the relevant “time" scale for the change of the density matrix is $O(1/\alpha_s)$, as obvious from the differential equation eq.(\[lind\]). The soft gluons in our approximation do not interact with each other, and thus their correlation function is free. The free propagator is proportional to $1/ k^+ \sim e^{-y}$, and thus the typical correlation length in rapidity space is $O(1)$. The evolution is therefore clearly in the Markovian regime which allows, at least naively speaking for the existence of differential evolution in the Lindblad form. We will come back to the discussion of Lindblad form later.
Equation eq.(\[lind\]) may look slightly unfamiliar as it does not quite have the form of the KLWMIJ equation discussed in [@Kovner:2005nq] . This is because it is written for density matrix and not the weight functional $W[j]$. To get to the latter form one needs to perform an extra step, i.e. Weyl transformation. This will be the subject of the next section. But before we do that, we consider the evolution of the density matrix in the dense regime.
The dense limit.
----------------
As we have seen, in the limit where the hadronic wave function contains a small number of partons (the dilute limit), the Lindblad form of the evolution equation follows directly using the straightforward perturbation theory at low x. We now turn our attention to the dense limit, where we assume that the color charge density in the wave function is large, parametrically of order $1/g$. The wave function in this limit has been calculated several years ago in [@Altinoluk:2009je]. In this section we use the results of that paper and reinterpret them from our current point of view.
To prepare for the calculation, note that the soft gluon emission operator $\hat{\Omega}$, when acting on the vacuum state $|0\rangle$ can be written as $$\label{ome}
\hat{\Omega}|0\rangle = \mathrm{Exp}\left\{i\sqrt{2} b_{\alpha} \int\frac{d\eta}{2\pi}[\hat{a}^{\dagger}_{\alpha}(\eta) + \hat{a}_{\alpha}(\eta)]\right\} \mathrm{Exp}\left\{-\frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi} \Lambda_{\beta\gamma}(\eta,\xi)\hat{a}^{\dagger}_{\beta}(\eta) \hat{a}^{\dagger}_{\gamma}(\xi)\right\}\mathcal{N}(\Lambda) |0\rangle$$ Here we write out the dependence on rapidity explicitly. Other indices (color, polarization, transverse position) are collectively represented by the Greek letters $\alpha, \beta, \gamma$. The matrix $\Lambda^{ab}_{ij}({\mathbf x}_\perp,{\mathbf y}_\perp,\eta_1,\eta_2)$ determines the amount of “squeezing” of the soft gluon vacuum. As we mentioned above, it has not been calculated explicitly in [@Altinoluk:2009je], however its properties relevant to the JIMWLK limit are known (see later). The $\mathcal{N}(\Lambda)$ is a normalization constant that only depends on $\Lambda$. Note that both $b_{\alpha}$ and $\Lambda_{\beta\gamma}$ are operators in the Hilbert space of hard gluons as they depend on the color charge density ${\mathbf j}$, and so in principle they do not commute. In eq.(\[ome\]) all the factors of $\Lambda$ should be understood as placed to the right of $b_{\alpha}$. In the JIMWLK limit however, where parametrically, $b=O(1/g)$ while $\Lambda=O(1)$, as was shown in [@Altinoluk:2009je] the order of the factors does not matter. In fact in showing that the operator $\hat \Omega$ Eq. diagonalizes the QCD Hamiltonian to leading order, ref.[@Kovner:2007zu; @Altinoluk:2009je] explicitly used this argument and assumed commutativity of the various factors of $b$ and $\Lambda$. We will not deviate from this assumption here and will treat these factors as commuting.
We further separate the annihilation operator $\hat{a}_{\alpha}(\eta)$ from the coherent state operator and move it to the far right acting on the vacuum state: $$\label{eq:omega_expression}
\begin{split}
\hat{\Omega}|0\rangle =& \mathrm{Exp}\left\{-\frac{\Delta y}{2\pi} b_{\alpha}(1-\Lambda_{0})_{\alpha\beta} b_{\beta}\right\}\mathrm{Exp}\left\{i \sqrt{2}b_{\alpha}(1-\Lambda_{0})_{\alpha\beta} \int \frac{d\eta}{2\pi} a^{\dagger}_{\beta}(\eta)\right\}\\
&\times \mathrm{Exp}\left\{-\frac{1}{2} \int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi} a^{\dagger}_{\alpha}(\eta)\Lambda_{\alpha\beta}(\eta,\xi) a^{\dagger}_{\beta}(\xi)\right\} \mathcal{N}(\Lambda)|0\rangle
\end{split}$$ where we have defined $$\Lambda_{0,\alpha\beta} = \int_{-\Delta y}^{\Delta y} \frac{d\zeta}{2\pi} \Lambda_{\alpha\beta}(\zeta,\eta)$$ Since $\Lambda_{\alpha\beta}(\zeta,\eta)$ depends only on the rapidity difference $\zeta-\eta$ [@Altinoluk:2009je], $\Lambda_{0,\alpha\beta}$ is rapidity independent. It does however have a nontrivial dependence on the width of the evolution step $\Delta y$. The nature of this dependence is very important. As we discussed above, we expect to have a bona fide differential evolution equation only if the correlations of the soft gluons in rapidity are short range. The function $\Lambda(\eta,\xi)$ is in fact the inverse of the correlator of the soft gluon modes. It should therefore decrease exponentially for rapidity difference greater than $\sim 1$. For such a function $\Lambda$ the dependence of $\Lambda_0$ on $\Delta y$ should be smooth with $\Lambda_0$ approximately constant for $1< \Delta y< 1/\alpha_s$. We will assume here that this is indeed the case and will treat $\Lambda_0$ as a constant independent of $\Delta y$. The results of [@Altinoluk:2009je] suggest that this is valid in the JIMWLK limit, i.e. when the dense hadron scatters on a dilute target, which is the regime that concerns us in this paper. We note that going beyond the JIMWLK limit posed some problems in [@Altinoluk:2009je], precisely for the reason that some of the soft modes in general seemed to possess long range correlations in rapidity. Our current understanding is that such long range correlations indeed are incompatible with the differential form of the evolution. It is thus possible that in order to go beyond the JIMWLK limit one would have to rethink the way in which the bipartitioning into the “observable" system and “environment" is done. This is however beyond the scope of the present paper.
In eq. , the first exponential represents wavefunction renormalization effects that have an overall $\Delta y $ factor. The second exponential contains the single gluon emission vertex $i b_{\alpha}(1-\Lambda_0)_{\alpha\beta}$ which is “renormalized" relative to the dilute case by the presence of the Bogoliubov operator $B$, while the third exponential contains the double gluon emission vertex $\Lambda_{\alpha\beta}(\xi,\eta)$.
Two gluons emitted from the same double gluon emission vertex are in general correlated in rapidity, while two gluons emitted from two single gluon emission vertexes are uncorrelated. We are now ready to calculate the superoperators. The fundamental difference with the dilute case, is that now not only one gluon state, but states with arbitrary number of soft gluons yield nontrivial jump operators that contribute to the evolution of the density matrix. For an $n$ soft gluon state we have $$\label{eq:Mn_general_expression}
\begin{split}
\hat{M}_n = & \langle n | \hat{R}\hat{\Omega} |0\rangle=\mathrm{Exp}\left\{-\frac{\Delta y}{2\pi} b_{\alpha}b_{\beta}(1-\Lambda_0)_{\alpha\beta}\right\} \mathcal{N}(\Lambda) \\
&\langle n| \hat{R}\, \mathrm{Exp}\left\{i\sqrt{2}b_{\alpha}(1-\Lambda_0)_{\alpha\beta}\int\frac{d\eta}{2\pi} \hat{a}^{\dagger}_{\beta}(\eta) \right\}\mathrm{Exp}\left\{- \frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi}\Lambda_{\alpha\beta}(\eta,\xi)\hat{a}^{\dagger}_{\alpha}(\eta) \hat{a}^{\dagger}_{\beta}(\xi)\right\} |0\rangle\\
\end{split}$$ Depending on the Fock state $|n\rangle$ being considered, we separately discuss the situations when the Fock state contains zero gluons, odd number of gluons and even number of gluons.
### Wavefunction renormalization operator
The superoperator $\hat{M}_0$ represents the wavefunction renormalization effects $$\hat{M}_0 = \langle 0| \hat{R} \hat{\Omega} |0\rangle = \mathrm{Exp}\left\{-\frac{\Delta y}{2\pi} b_{\alpha}b_{\beta}(1-\Lambda_0)_{\alpha\beta}\right\} \mathcal{N}(\Lambda)\, .$$ Up to terms linear in $\Delta y$, $$\hat{M}_0 \approx 1 -\frac{\Delta y}{2\pi}\left[b_{\alpha}b_{\beta}(1-\Lambda_0)_{\alpha\beta} \right] + \mathcal{O}(\Delta y^2)\,$$ Note that the wavefunction renormalization operator $\hat{M}_0$ is independent of $\hat{R}$ and we have ignored the normalization $\mathcal{N}(\Lambda)$ factor, since it is irrelevant in the JIMWLK limit [@Altinoluk:2009je]. The superoperator $\hat{M}_0$ contributes to the change of density matrix through the term $$\hat{M}_0 \hat{\rho}_v \hat{M}_0^{\dagger} = \hat{\rho}_v - \frac{\Delta y}{2\pi} \left[b_{\alpha}b_{\beta}(1-\Lambda_0)_{\alpha\beta} \, \hat{\rho}_v + \hat{\rho}_v\, (1-\Lambda_0^{\dagger})_{\alpha\beta}b_{\alpha}b_{\beta} \right] + \mathcal{O}(\Delta y^2) \, .$$
### Jump operators with odd number of gluons
For Fock states with odd numbers of gluons, one needs odd number of single-gluon-emission vertices in calculating the jump operators. However, every single gluon emission brings an extra power of $\Delta y$, since gluons produced from different single-gluon-emission vertices are uncorrelated in rapidity. The integral over rapidity of every such gluons in the amplitude and conjugate amplitude brings therefore an extra power of $\Delta y$. Thus one needs to keep only one single-gluon-emission vertex in $\hat M_{2i+1}$ in order to calculate the relevant jump operators that contribute to the differential form of the evolution equation.
The explicit expression for a jump operator follows from eq. $$\label{mn}
\begin{split}
\hat{M}_n&= \langle n| \hat{R} \left(i\sqrt{2}b_{\alpha}(1-\Lambda_0)_{\alpha\beta}\int\frac{d\zeta}{2\pi} \hat{a}^{\dagger}_{\beta}(\zeta) \right) \mathrm{Exp}\left\{- \frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi}\Lambda_{\alpha\beta}(\eta,\xi)\hat{a}^{\dagger}_{\alpha}(\eta) \hat{a}^{\dagger}_{\beta}(\xi)\right\}|0\rangle\\
& =\left( i \mathcal{R}^{\dagger}_{\gamma \delta} [\sqrt{2}b(1-\Lambda_0)]_{\delta} \right) \langle n |\int\frac{d\zeta}{2\pi} \hat{a}_{\gamma}^{\dagger}(\zeta) \mathrm{Exp}\left\{- \frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi}\bar{\Lambda}_{\alpha\beta}(\eta,\xi)\hat{a}^{\dagger}_{\alpha}(\eta) \hat{a}^{\dagger}_{\beta}(\xi)\right\}|0\rangle.
\end{split}$$
Here $$\label{lambar}
\bar{\Lambda}_{\alpha\beta} (\eta,\xi)= \mathcal{R}^{\dagger}_{\alpha\gamma} \Lambda_{\gamma\delta}(\eta,\xi) \mathcal{R}_{\delta \beta}.$$
To arrive at this expression we have inserted the factor $\hat{R}^{\dagger}\hat{R} =1$ next to the soft gluon vacuum state $|0\rangle$, used the fact that $\hat R|0\rangle=|0\rangle$ and evaluated the action of $\hat{R}$ on the soft gluon creation and annihilation operators using Eq..
Importantly, the operator ordering in Eq. is such that all the operators $\mathcal{R} $ are understood to be placed to the left of all the factors of the $b_{\alpha}$ and $\Lambda_{\alpha\beta}$. This follows from the fact that the operator $\hat R$ in the original expression is acting directly on the $n$-gluon state, and thus all the factors of $\hat \Phi$ indeed are ordered to the left of all $j$-dependent factors in the original expression. Thus for example in the definition Eq. the action of $\mathcal{R}_{\delta \beta}$ on $\Lambda$ is understood only as a color matrix rotation. This comment also applies to the rest of the formulae in this section.
In the following, we explicitly calculate a few expressions of the jump operators and their action on the density matrix. This will make the dependence on $\Delta y$ more transparent.
For a one-gluon Fock state $|1_{\{\alpha_1, \mathbf{w}_1,\eta_1\}}\rangle = a_{\alpha_1}^{\dagger}(\mathbf{w}_1,\eta_1)|0\rangle$ with transverse position $\mathbf{w}_1$, rapidity $\eta_1$ and color index $\alpha_1$, the jump operator is $$\hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}} = \int d^2\mathbf{z}_{1} i\sqrt{2}b_{\alpha}(\mathbf{z}_1) [1 - \Lambda_0]_{\alpha\beta}(\mathbf{z}_1,\mathbf{w}_1) \mathcal{R}_{\beta \alpha_1}(\mathbf{w}_1)$$ Note that the jump operator associated with one-gluon Fock state is independent of the rapidity index $\eta_1$. Integration over all the one-gluon Fock states produces an overall factor $\Delta y$ in the evolution of the density matrix. The one-gluon jump operators contribute to this evolution through $$\label{oneg}
\begin{split}
&\hat{M}_{1}\hat{\rho}_v \hat{M}_1^{\dagger} = \sum_{\alpha_1} \int d\mathbf{w}_1\int\frac{d\eta_1}{2\pi}\hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}} \hat{\rho}_v\hat{M}_{1\{\alpha_1, \mathbf{w}_1,\eta_1\}}^{\dagger} \\
=&\frac{\Delta y}{\pi} \int d\mathbf{z}_1 d\mathbf{z}_2b_{\alpha}(\mathbf{z}_1) \left[\int d\mathbf{w}_1(1-\Lambda_0)_{\alpha\beta}(\mathbf{z}_1,\mathbf{w}_1) [\mathcal{R}\hat{\rho}_v\mathcal{R}^{\dagger}]_{\beta\gamma} (1-\Lambda_0^{\dagger})_{\gamma\delta}(\mathbf{w}_1, \mathbf{z}_2)\right]b_{\delta}(\mathbf{z}_2)\\
=&\frac{\Delta y}{\pi} [\bar{b}(1-\bar{\Lambda}_0]_{\alpha} \hat{\rho}_v [(1-\bar{\Lambda}_0^{\dagger})\bar{b}]_{\alpha}
\end{split}$$ In the last line we have reverted to the convoluted notation where single index $\alpha$ represents the transverse position, color and polarization. Barred quantities here and below indicate the quantities that are rotated by the $\mathcal{R}$ matrix.
For a three-gluon Fock state $
|3_{\{\alpha_i, \mathbf{w}_i,\eta_i;i=1,2,3\}}\rangle = a^{\dagger}_{\alpha_1}(\mathbf{w}_1,\eta_1) a^{\dagger}_{\alpha_2}(\mathbf{w}_2,\eta_2) a^{\dagger}_{\alpha_3}(\mathbf{w}_3,\eta_3) |0\rangle $, the jump operator is $$\begin{split}
&\hat{M}_{3_{\{\alpha_i, \mathbf{w}_i, \eta_i; i=1,2,3\}}} = -\int d^2\mathbf{z}_1 i\sqrt{2}b_{\alpha}(\mathbf{z}_1) \Bigg([1-\Lambda_0]_{\alpha\beta}(\mathbf{z}_1, \mathbf{w}_1) \Lambda_{\kappa\lambda}(\mathbf{w}_2,\eta_2, \mathbf{w}_3,\eta_3) \\
&+ [1-\Lambda_0]_{\alpha\kappa}(\mathbf{z}_1, \mathbf{w}_2) \Lambda_{\beta\lambda}(\mathbf{w}_1,\eta_1, \mathbf{w}_3,\eta_3) +[1-\Lambda_0]_{\alpha\lambda}(\mathbf{z}_1, \mathbf{w}_3) \Lambda_{\beta\kappa}(\mathbf{w}_1,\eta_1; \mathbf{w}_2,\eta_2) \Bigg)\\
&\times \mathcal{R}_{\beta\alpha_1}(\mathbf{w}_1)\mathcal{R}_{\kappa\alpha_2}(\mathbf{w}_2)\mathcal{R}_{\lambda\alpha_3}(\mathbf{w}_3)
\end{split}$$ It contains sum of all possible terms where two out of the three gluons are emitted from the same two-gluon-emission vertex. Note that $\Lambda$ is a symmetric matrix. The contribution of the three gluon jump operator to the evolution of the density matrix is $$\hat{M}_3 \hat{\rho}_v \hat{M}_3^{\dagger} =\sum_{\alpha_1,\alpha_2,\alpha_3}\int d\mathbf{w}_1d\mathbf{w}_2d\mathbf{w}_3 \int\frac{d\eta_1}{2\pi}\frac{d\eta_2}{2\pi}\frac{d\eta_3}{2\pi}\hat{M}_{3_{\{\alpha_i, \mathbf{w}_i, \eta_i; i=1,2,3\}}}\hat{\rho}_v\hat{M}_{3_{\{\alpha_i, \mathbf{w}_i, \eta_i; i=1,2,3\}}}^{\dagger}$$ This expression has in principle nine terms. However, for terms that involve the same two gluons connected to a two gluon emission vertex both in $\hat M_3$ and $\hat M_3^\dagger$, integration over rapidity produces higher than linear powers in $\Delta y$. For example $$\int\frac{d\eta_1}{2\pi}\frac{d\eta_2}{2\pi}\frac{d\eta_3}{2\pi}\Lambda_{\kappa\lambda}(\mathbf{w}_{1}, \eta_{1}; \mathbf{w}_{2}, \eta_{2})\Lambda^{\dagger}_{\rho\delta}(\mathbf{w}_{1},\eta_{1};\mathbf{w}_{2},\eta_{2}) = (\Delta y)^2\int\frac{d\zeta}{2\pi}\Lambda_{\kappa\lambda}(\mathbf{w}_{1},\mathbf{w}_{2};\zeta)\Lambda^{\dagger}_{\rho\delta}(\mathbf{w}_{1},\mathbf{w}_{2};\zeta).$$ This term therefore does not contribute to the differential form of the evolution.
On the other hand, for the two-gluon-emission vertexes connected to different pairs of gluons, only one explicit $\Delta y$ factor arises $$\int\frac{d\eta_1}{2\pi}\frac{d\eta_2}{2\pi}\frac{d\eta_3}{2\pi}\Lambda_{\kappa\lambda}(\mathbf{w}_{1}, \eta_{1}; \mathbf{w}_{2}, \eta_{2})\Lambda^{\dagger}_{\rho\delta}(\mathbf{w}_{1},\eta_{1};\mathbf{w}_{3},\eta_{3}) = \Delta y\, \Lambda_{0,\kappa\lambda}(\mathbf{w}_{1}, \mathbf{w}_{2})\Lambda^{\dagger}_{0,\rho\delta}(\mathbf{w}_{1},\mathbf{w}_{3})\, .$$ These terms do contribute.
The contribution of the three gluon jump operators is thus $$\label{m3}
\begin{split}
\hat{M}_3\hat{\rho}_v \hat{M}_3^{\dagger}
=& \frac{\Delta y}{\pi} \sum_{\alpha_1, \alpha_2,\alpha_3}\int d\mathbf{w}_1d\mathbf{w}_2d\mathbf{w}_3 [\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_{1}} (\mathbf{w}_{1}) \bar{\Lambda}_{0, \alpha_{2}\alpha_{3}}(\mathbf{w}_{2}, \mathbf{w}_{3})\hat{\rho}_v\\
& \quad \times \bar{\Lambda}^{\dagger}_{0,\alpha_{1}\alpha_{2}}(\mathbf{w}_{1},\mathbf{w}_{2}) [(1-\bar{\Lambda}_0^{\dagger})\bar{b}]_{\alpha_{3}}(\mathbf{w}_{3})\\
=&\frac{\Delta y}{\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})] \bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L} [(1-\bar{\Lambda}_{0,R}^{\dagger})\bar{b}_R]\hat{\rho}_v
\end{split}$$ In the last line we have used superscripts $L$ and $R$ to indicate the position of various factors relative to the density matrix $\hat\rho_v$, thus $\bar\Lambda_{0,L}$ indicates that this factor $\bar\Lambda_0$ is placed to the left of $\hat\rho_v$ etc. The ordering is important as the various operators do not commute with $\hat\rho_v$. The reason to write the expression in this particular way is that we can use convenient matrix notations, so that products in eq. (\[m3\]) are matrix products over all indexes carried by $\bar \Lambda_0$ and $\bar b$, i.e. color, polarization and transverse coordinate.
This pattern clearly generalizes to any odd $n$. Linear in $\Delta y$ contributions arise only from terms where no two gluons are emitted from the same two gluon emission vertex both in $M_n$ and $M^\dagger_n$. Diagrammatically the terms that yield linear in $\Delta y$ contributions are depicted in Fig.1.
![The diagrams involving the odd number of gluons that contribute terms linear in $\Delta y$ to the evolution of $\hat\rho$. []{data-label="f1"}](odd_gluons.pdf){width="12cm"}
Generalizing the above analysis to jump operators with $2m+1$ numbers of gluons we obtain $$\label{eq:jump_operator_odd_gluons}
\begin{split}
\hat{M}_{2m+1_{\{\alpha_i,\mathbf{w}_i,\eta_i; i=1,\ldots, 2m+1\}}}
=&\sum_{P=\{i_1, \ldots, i_{2m+1}\}}[i\sqrt{2}b(1-\Lambda_0)]_{\beta_{i_1}}(\mathbf{w}_{i_1}) \prod^m_{k\neq l\neq 1} \Lambda_{\beta_{i_k}\beta_{i_l}}(\mathbf{w}_{i_k},\eta_{i_k};\mathbf{w}_{i_l},\eta_{i_l})\\
&\times \prod_{q=1}^{2m+1} \mathcal{R}_{\beta_{i_q} \alpha_{q}}(\mathbf{w}_{q})\\
\end{split}$$ Here $i_1, i_2, \ldots, i_{2m+1}$ is a permutation of $1, 2,\ldots, 2m+1$. The sum over P goes over all the possible permutations. The action of $\hat M_{2m+1}$ on the density matrix after summing over all the possible Fock states with $2m+1$ gluons and performing the rapidity integrations, becomes $$\begin{split}
&\hat{M}_{2m+1}\hat{\rho}_v \hat{M}^{\dagger}_{2m+1} = \frac{\Delta y}{\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})](\bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L})^{m}[(1-\bar{\Lambda}_{0,R}^{\dagger})\bar{b}_R]\hat{\rho}_v\, .\\
\end{split}$$ Here, just like for $\hat M_3$ the contribution comes only from those terms that do not contain a single pair of gluons emitted from the same two gluon emission vertex in $M_{2n+1}$ and $M^\dagger_{2n+1}$ as illustrated on Fig.1.
Now adding all the jump operators associated with odd numbers of gluons, their action on the density matrix is $$\sum_{m=0}\hat{M}_{2m+1}\hat{\rho}_v \hat{M}^{\dagger}_{2m+1} = \frac{\Delta y}{\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})](1-\bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L})^{-1}[(1-\bar{\Lambda}_{0,R}^{\dagger})\bar{b}_R]\hat{\rho}_v\, .$$
### Jump operators with even number of gluons
If the number of gluons in the Fock state is even, the gluons can either be emitted from a two-gluon-emission vertex or from even number of single-gluon-emission vertexes. Since gluons emitted from single-gluon-emission vertexes are uncorrelated in rapidity, the more single gluon emission vertexes are involved, the higher power of $\Delta y$ is generated. Recall that we only need to keep terms linear in $\Delta y$. To extract these terms, we allow either no gluons or two gluons to be emitted from the single-gluon-emission vertexes. The rest of the contributions, as we wil see are subleading in powers of $\Delta y$.
The expression for the jump operators with even numbers of gluons by $\hat{M}_{2n}$ follows from eq. $$\begin{split}
\hat{M}_{2n}=
&\hat{M}_{2n}^{0} + \hat{M}_{2n}^{2} =\langle 2n| \hat{R}\, \mathrm{Exp}\left\{- \frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi}\Lambda_{\alpha\beta}(\eta,\xi)\hat{a}^{\dagger}_{\alpha}(\eta) \hat{a}^{\dagger}_{\beta}(\xi)\right\} |0\rangle\\
+&\langle 2n| \hat{R}\,\frac{1}{2!}\left(i\sqrt{2}b_{\gamma}(1-\Lambda_0)_{\gamma\delta}\int\frac{d\zeta}{2\pi} \hat{a}^{\dagger}_{\delta}(\zeta) \right)^2\mathrm{Exp}\left\{- \frac{1}{2}\int\frac{d\eta}{2\pi}\frac{d\xi}{2\pi}\Lambda_{\alpha\beta}(\eta,\xi)\hat{a}^{\dagger}_{\alpha}(\eta) \hat{a}^{\dagger}_{\beta}(\xi)\right\} |0\rangle\, .\\
\end{split}$$ We have denoted the parts with no single-gluon-emission vertex and with two single-gluon-emission vertexes by $\hat{M}_{2n}^{0}$ and $\hat{M}^{2}_{2n}$, respectively. The action on the density matrix becomes $$\label{eq:explicit_EJO_dm}
\hat{M}_{2n}\hat{\rho}_v \hat{M}_{2n}^\dagger = \hat{M}_{2n}^0\hat{\rho}_v \hat{M}_{2n}^{0 \dagger}+ \hat{M}_{2n}^{0}\hat{\rho}_v \hat{M}_{2n}^{2\dagger}+ \hat{M}_{2n}^{2}\hat{\rho}_v \hat{M}_{2n}^{0\dagger}+ \hat{M}_{2n}^{2}\hat{\rho}_v \hat{M}_{2n}^{2\dagger}$$ Just like in the case of odd number of gluons, not all the terms in eq. contribute to differential evolution. The last term in eq. contains an overall factor of $(\Delta y)^2$ and therefore can be discarded. The first term in eq. does contain terms that are only linear in $\Delta y$, however in the dense limit it is suppressed by a power of $\alpha_s$ relative to the second and third terms, as in the dense limit $b\sim 1/g$, while $\Lambda \sim 1$. Similar terms arise in the expansion of the normalization factor $\mathcal{N}(\Lambda)$, which we have neglected above. We therefore discard these terms in the week coupling limit. Only the second and third terms in eq. are to be evaluated and contribute to the differential evolution of the density matrix.
We first evaluate $\hat{M}_{2n}^{0}$. For example, for a two-gluon Fock state $|2_{\{\alpha_1, \mathbf{w}_1,\eta_1; \alpha_2, \mathbf{w}_2,\eta_2\}}\rangle = a^{\dagger}_{\alpha_1}(\mathbf{w}_1,\eta_1)a^{\dagger}_{\alpha_2}(\mathbf{w}_2,\eta_2) |0\rangle $ $$\begin{split}
\hat{M}_{2_{\{\alpha_1,\mathbf{w}_1,\eta_1; \alpha_2,\mathbf{w}_2,\eta_2\}}}^{0}
& = - \Lambda_{\kappa\lambda}(\mathbf{w}_1,\eta_1;\mathbf{w}_2,\eta_2) \mathcal{R}_{\kappa\alpha}(\mathbf{w}_1) \mathcal{R}_{\lambda\alpha_2}(\mathbf{w}_2)= -\bar{\Lambda}_{\alpha_1\alpha_2}(\mathbf{w}_1,\eta_1;\mathbf{w}_2,\eta_2)\\
\end{split}$$ Generalization to Fock states with $2m$ gluons is straightforward $$\begin{split}
\hat{M}_{2m_{\{\alpha_i,\mathbf{w}_i,\eta_i; i=1,\ldots, 2m\}}}^{0}
=&\left(-\frac{1}{2}\right)^m \frac{1}{m!} \sum_{P=\{i_1, \ldots, i_{2m}\}} \prod_{k=1}^m \bar{\Lambda}_{\alpha_{i_{2k-1}}\alpha_{i_{2k}}}(\mathbf{w}_{i_{2k-1}},\eta_{i_{2k-1}};\mathbf{w}_{i_{2k}},\eta_{i_{2k}}) \\
\end{split}$$ The summation is over all the permutations. The prefactor $(1/2)^m$ is needed to account for the fact that $\Lambda$ is a symmetric matrix. The factor $1/m!$ takes care of the fact that two permutations that differ only by ordering of some pairs of indices and nothing else, give identical contributions which only need to be counted once.
To calculate $\hat{M}_{2n}^{2}$ we start with the simple situation of two-gluon Fock state $$\begin{split}
\hat{M}_{2_{\{\alpha_1,\mathbf{w}_1,\eta_1; \alpha_2,\mathbf{w}_2,\eta_2\}}}^{2}
&=- \int d^2\mathbf{z}_1 d^2\mathbf{z}_3 b_{\alpha}(\mathbf{z}_1) b_{\rho}(\mathbf{z}_3)\bigg( [1-\Lambda_0]_{\alpha\beta}(\mathbf{z}_1, \mathbf{w}_1)[1-\Lambda_0]_{\rho\delta}(\mathbf{z}_3, \mathbf{w}_2)\bigg)\\
&\quad \times \mathcal{R}_{\beta\alpha_1}(\mathbf{w}_1) \mathcal{R}_{\delta\alpha_2}(\mathbf{w}_2)\\
&=-[\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_1}(\mathbf{w}_1) [\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_2}(\mathbf{w}_2)
\end{split}$$ Generalization to Fock states with $2m$ gluons leads to $$\begin{split}
\hat{M}^{2}_{2m_{\{\alpha_i,\mathbf{w}_i,\eta_i;i=1,\ldots, 2m\}}}&=-\frac{1}{(m-1)!}\left(-\frac{1}{2}\right)^{m-1} \sum_{P=\{i_1,\ldots, i_{2m}\}} [\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_{i_1}}(\mathbf{w}_{i_1}) [\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_{i_2}}( \mathbf{w}_{i_2})\\
&\quad\times\prod_{k=2}^{m}\bar{\Lambda}_{\alpha_{i_{2k-1}}\alpha_{i_{2k}}}(\mathbf{w}_{i_{2k-1}},\eta_{i_{2k-1}}, \mathbf{w}_{i_{2k}},\eta_{i_{2k}})\,.\\
\end{split}$$ The action of the two gluon jump operator on the density matrix is $$\begin{split}
\hat{M}_2^{2}\hat{\rho}_v \hat{M}_2^{0\dagger}& = \sum_{\alpha_1\alpha_2}\int d\mathbf{w}_1d\mathbf{w}_2\frac{d\eta_1}{2\pi}\frac{d\eta_2}{2\pi}[\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_1}(\mathbf{w}_1) [\bar{b}(1-\bar{\Lambda}_0)]_{\alpha_2}(\mathbf{w}_2) \hat{\rho}_v \bar{\Lambda}^{\dagger}_{\alpha_1\alpha_2}(\mathbf{w}_1,\eta_1;\mathbf{w}_2,\eta_2)\\
&=\frac{\Delta y}{2\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})]\bar{\Lambda}_{0,R}^{\dagger}[\bar{b}_L(1-\bar{\Lambda}_{0,L})] \hat{\rho}_v\, .
\end{split}$$ For the four gluon operator we similarly find $$\hat{M}_4^{2}\hat{\rho}_v \hat{M}_4^{ 0\dagger} = \frac{\Delta y}{2\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})]\bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L} \bar{\Lambda}_{0,R}^{\dagger}[\bar{b}_L(1-\bar{\Lambda}_{0,L})] \hat{\rho}_v\, .$$ Summing up all the jump operators with even numbers of gluons, we get $$\sum_{m=1}^{\infty}\hat{M}_{2m}^{2}\hat{\rho}_v \hat{M}_{2m}^{0\dagger} =\frac{\Delta y}{2\pi} [\bar{b}_L(1-\bar{\Lambda}_{0,L})](1-\bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L})^{-1} \bar{\Lambda}_{0,R}^{\dagger}[\bar{b}_L(1-\bar{\Lambda}_{0,L})] \hat{\rho}_v\, .$$ and its complex conjugate $$\sum_{m=1}^{\infty}\hat{M}_{2m}^{0}\hat{\rho}_v \hat{M}_{2m}^{2\dagger} =\frac{\Delta y}{2\pi} [\bar{b}_R(1-\bar{\Lambda}_{0,R}^{\dagger})]\bar{\Lambda}_{0,L}(1-\bar{\Lambda}_{0,R}^{\dagger}\bar{\Lambda}_{0,L})^{-1}[\bar{b}_R(1-\bar{\Lambda}_{0,R}^{\dagger})] \hat{\rho}_v\, .$$
### All together now.
We now put together the above results. The wave function renormalization operator and the jump operators with even numbers of gluons contribute to what one might call “the virtual part" of the evolution: $$\begin{split}
&\hat{M}_0\hat{\rho}_v \hat{M}^{\dagger}_0 +\sum_{n=1}^\infty\hat{M}_{2n} \hat{\rho}_v \hat{M}^{ \dagger}_{2n}\\
= &\hat{\rho}_v- \left(\frac{\Delta y}{2\pi}\right)\left[ \bar{b}_L(1-\bar{\Lambda}_{0,L})(1-\bar{\Lambda}^{\dagger}_{0,R}\bar{\Lambda}_{0,L})^{-1} (1-\bar{\Lambda}^{\dagger}_{0,R})\bar{b}_L\right] \, \hat{\rho}_v\\
&- \left(\frac{\Delta y}{2\pi} \right)\left[\bar{b}_R(1-\bar{\Lambda}_{0,L})(1-\bar{\Lambda}^{\dagger}_{0,R}\bar{\Lambda}_{0,L})^{-1} (1-\bar{\Lambda}^{\dagger}_{0,R})\bar{b}_R\right]\,\hat{\rho}_v\, .
\end{split}$$ Adding contributions from jump operators with odd numbers of gluons we get $$\label{eq:deltaY_changes}
\begin{split}
&\hat{M}_0\hat{\rho}_v \hat{M}^{\dagger}_0 +\sum_{n=1}^\infty\hat{M}_{2n} \hat{\rho}_v \hat{M}^{ \dagger}_{2n} + \sum_{n=0}^\infty\hat{M}_{2n+1} \hat{\rho}_v \hat{M}^{ \dagger}_{2n+1}\\
=&\hat{\rho}_v -\left(\frac{\Delta y}{2\pi}\right)\left[(\bar{b}_L -\bar{b}_R)(1-\bar{\Lambda}_{0,R}^{\dagger})(1- \bar{\Lambda}_{0,L}\bar{\Lambda}_{0,R}^{\dagger})^{-1}(1-\bar{\Lambda}_{0,L}) (\bar{b}_L -\bar{b}_R)\right]\hat{\rho}_v\, .
\end{split}$$
### Operator ordering.
As we have mentioned earlier, the relative ordering of operators of $b$ and $\Lambda_0$ in Eq. is not important. It is however important to keep track of the ordering of the classical fields $b_L$ and $b_R$ relative to the density matrix. More precisely, in the JIMWLK limit one can change the order of various factors in Eg. as long as each factor $(\bar b_L-\bar b_R)$ is kept as a unit and is commuted with any other operator in question..
The argument for that was given in [@Altinoluk:2009je] , and we reproduce it here for completeness.
Recall that the JIMWLK limit is obtained when parametrically $b\sim O(1/g)$ and $\Lambda_0\sim O(1)$, and additionally the evolution equation should be expanded to order $\alpha_s$. The latter expansion gives the leading contribution when the dense hadron scatters on a dilute target.
Since both $b$ and $\Lambda_0$ are functions of $j$, the commutator between $b$ and $\Lambda_0$ can be estimated as $$(b\Lambda_0 - \Lambda_0 b) = \frac{\delta b}{\delta j^{a}} [j^a, j^b] \frac{\delta \Lambda_0}{\delta j^b} = igf^{abc} \frac{\delta b}{\delta j^{a}} j^c \frac{\delta \Lambda_0}{\delta j^b} \sim \mathcal{O}(g)\, .$$ The difference $\bar{b}_L -\bar{b}_R$ can also be estimated as $$(b_L -b_R)\hat\rho_v = \left(b\hat{\rho}_v - \hat{\rho}_v b\right) = \frac{\delta b}{\delta j^a} [j^a, j^b] \frac{\delta \hat{\rho}_v}{\delta j^b} = igf^{abc} \frac{\delta b}{\delta j^a} j^c \frac{\delta \hat{\rho}_v}{\delta j^b} \sim \mathcal{O}(g)\hat\rho_v$$ and barred quantities are color rotated by $\hat{\mathcal{R}}\sim \mathcal{O}(1)$.The difference between $\Lambda_{0,L}$ and $\Lambda_{0,R}$ is estimated similarly $$( \Lambda_{0,L} -\Lambda_{0,R})\hat\rho_v = \Lambda_0 \hat{\rho}_v -\hat{\rho}_v \Lambda_0 = \frac{\delta \Lambda_0}{\delta j^a} [j^a , j^b] \frac{\delta \hat{\rho}_v}{\delta j^b} = igf^{abc} \frac{\delta \Lambda_0}{\delta j^a} j^c \frac{\delta \hat{\rho}_v}{\delta j^b}\sim \mathcal{O}(g^2)\hat\rho_v\, .$$ Since the factor $(\bar{b}_L -\bar{b}_R)^2$ on the right hand side of eq. is already of order of $\alpha_s$, the ordering between $b$ and $\Lambda_0$ and the difference between $\Lambda_{0,L}$ and $\Lambda_{0,R}$ contribute to higher orders in $\alpha_s$ and therefore can be ignored in the JIMWLK limit as long as one does not order the operators differently in the terms containing $b_L$ and $b_R$. Thus for example, one can substitute in Eq. $$(1-\bar{\Lambda}_{0,R}^{\dagger})(1- \bar{\Lambda}_{0,L}\bar{\Lambda}_{0,R}^{\dagger})^{-1}(1-\bar{\Lambda}_{0,L})\rightarrow )(1-\bar{\Lambda}_{0,R}^{\dagger})(1- \bar{\Lambda}_{0,R}\bar{\Lambda}_{0,R}^{\dagger})^{-1}(1-\bar{\Lambda}_{0,R})$$or $$(1-\bar{\Lambda}_{0,R}^{\dagger})(1- \bar{\Lambda}_{0,L}\bar{\Lambda}_{0,R}^{\dagger})^{-1}(1-\bar{\Lambda}_{0,L})\rightarrow)(1-\bar{\Lambda}_{0,L}^{\dagger})(1- \bar{\Lambda}_{0,L}\bar{\Lambda}_{0,L}^{\dagger})^{-1}(1-\bar{\Lambda}_{0,L})$$ as a whole, without breaking the factor $(\bar{b}_L -\bar{b}_R)^2$ into separate pieces.
### The Lindblad form, finally.
We now simplify eq. using the results of [@Altinoluk:2009je]. First we note that the function of $\bar\Lambda$ appearing in eq. can be represented as a square if we indeed forget about the difference between $\Lambda_L$ and $\Lambda_R$. Define formally $$\Theta=\sqrt{(1-\Lambda\Lambda^\dagger)^{-1}}$$ We can then write $$(1-\bar{\Lambda}^{\dagger})(1- \bar{\Lambda}\bar{\Lambda}^{\dagger})^{-1}(1-\bar{\Lambda})=\bar N^\dagger\bar N$$with $$\label{n}
N=\Theta(1-\Lambda)$$ In fact the matrix $\Theta$ appears naturally in the calculation of [@Altinoluk:2009je]. Since the soft gluon vacuum is a squeezed state due to the presence of the Bogoliubov operator ${\cal\hat B}$, it is the Fock space vacuum of the Bogoliubov transformed set of creation and annihilation operators, related to the original gluon operators $a^\dagger$ and $a$ by $$\hat{\beta}_{\rho} = \Theta_{\rho\sigma } \hat{a}_{\sigma} + \Phi_{\rho\sigma} \hat{a}^{\dagger}_{\sigma}; \ \ \ \ \ \ \ \ \ \ \
\hat{\beta}_{\rho}^{\dagger} = \Theta^{\ast}_{\rho\sigma} \hat{a}^{\dagger}_{\sigma} + \Phi^{\ast}_{\rho\sigma} \hat{a}_{\sigma}$$ where $\Theta$ and $\Phi$ are constrained by the unitarity condition $$\Theta\Theta^{\dagger} -\Phi\Phi^{\dagger}=1$$ The following relation was also derived in [@Altinoluk:2009je] $$\Lambda = \Theta^{-1} \Phi$$ Using these relations we find $$\label{eq:multiparticle_correction}
(1-\bar{\Lambda}^{\dagger})(1- \bar{\Lambda}\bar{\Lambda}^{\dagger})^{-1}(1-\bar{\Lambda})
=(\bar{\Theta}^{\dagger} - \bar{\Phi}^{\dagger}) (\bar{\Theta} -\bar{\Phi}).$$ with the usual definitions $\bar{\Theta} = \mathcal{R}^{\dagger}\Theta \mathcal{R}$ and $\bar{\Phi} = \mathcal{R}^{\dagger} \Phi\mathcal{R}$. This is indeed the same as eq. with $$N=\Theta-\Phi$$
After integration over rapidities, eq. becomes $$\label{eq:N_perp_to_Lambda}
\frac{\Delta y}{2\pi}(1-\bar{\Lambda}_{0}^{\dagger})(1- \bar{\Lambda}_{0}\bar{\Lambda}_{0}^{\dagger})^{-1}(1-\bar{\Lambda}_{0}) = \frac{\Delta y}{2\pi} \bar{N}_{\perp}^{\dagger} \bar{N}_{\perp}\, ,$$ where $N_{\perp} = \int d\eta (\Theta -\Phi)$ has been calculated in [@Altinoluk:2009je] $$\label{eq:Nperp_exp}
N_{\perp} = [1-l-L] = \left[\delta_{ij} - \partial_i \frac{1}{\partial^2}\partial_j -D_i\frac{1}{D^2} D_j\right]^{ab} (\mathbf{x}_{\perp}, \mathbf{y}_{\perp})\, .$$ where the covariant derivative is defined as $D_i^{ab} = \delta^{ab}\partial_i -igT^{e}_{ba}b_i^e$.
Substituting eq. into eq. , we obtain $$\label{eq:classical_approximated_equation}
\hat{\rho}_v(\Delta y) = \hat{\rho}_v - \frac{\Delta y}{2\pi} (\bar{b}_L -\bar{b}_R) \bar{N}_{\perp}^{\dagger} \bar{N}_{\perp} (\bar{b}_L - \bar{b}_R)\hat{\rho}_v\, .$$ To write this explicitly as an operator equation we need to make the choice of whether to place the factor $\bar N^\dagger \bar N$ to the right or to the left of the density matrix, as both choices, as well as some others are equivalent in the JIMWLK limit. Additionally we need to specify the ordering between the operators $b$ and $\Lambda$ on one hand and factors of $R$ on another, which has been scrambled in the calculation above. It is not our goal here to carefully restore the correct ordering, as only the JIMWLK limit of this expression is strictly speaking under control. We therefore simply choose the specific ordering which reproduces JIMWLK as well as gives an evolution equation which preserves the Hermiticity of the density matrix also away from the JIMWLK limit.
We take $$\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp} = \mathcal{R}^{\dagger} (N^{\dagger}_{\perp}N_{\perp})\mathcal{R}, \qquad \bar{b} = b\mathcal{R}\, , \, \ \ \ \bar b^\dagger =\mathcal R^\dagger b$$ where the operator ordering is now specified. With these definitions we write the evolution as $$\label{eq:nonlindblad_eq_cgc}
\begin{split}
\frac{d\hat{\rho}_v}{dy} =& -\frac{1}{4\pi} \left(\hat{\rho}_v \bar{b}_{\alpha} (\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp})_{\alpha\beta} \bar{b}^{\dagger}_{\beta} + \bar{b}^{\dagger}_{\beta}\bar{b}_{\alpha} \hat{\rho}_v (\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp})_{\alpha\beta} - \bar{b}_{\alpha} \hat{\rho}_v (\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp})_{\alpha\beta} \bar{b}_{\beta}^{\dagger} - \bar{b}^{\dagger}_{\beta} \hat{\rho}_v \bar{b}_{\alpha} (\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp})_{\alpha\beta} \right)+h.c.\\
=&-\frac{1}{4\pi} \left[\bar{b}^{\dagger}_{\beta}, \left[\bar{b}_{\alpha}, \hat{\rho}_v\right](\bar{N}^{\dagger}_{\perp} \bar{N}_{\perp})_{\alpha\beta} \right]+h.c.\\
\end{split}$$
Written out explicitly $$\label{eq:nonlindblad_eq_cgc_explicit}
\begin{split}
\frac{d\hat{\rho}_v}{dy} =& -\frac{1}{4\pi} \Big[\hat{\rho}_v b_{\alpha} (N^{\dagger}_{\perp} N_{\perp})_{\alpha\beta} b_{\beta} + \mathcal{R}^{\dagger}_{\alpha\omega}b_{\omega}b_{\rho}\mathcal{R}_{\rho\beta} \hat{\rho}_v \mathcal{R}^{\dagger}_{\beta\lambda}(N^{\dagger}_{\perp} N_{\perp})_{\lambda\kappa} \mathcal{R}_{\kappa\alpha}\\
& - b_{\rho}\mathcal{R}_{\rho\alpha} \hat{\rho}_v \mathcal{R}^{\dagger}_{\alpha\omega}(N^{\dagger}_{\perp} N_{\perp})_{\omega\beta}b_{\beta} - \mathcal{R}^{\dagger}_{\alpha\omega}b_{\omega} \hat{\rho}_v b_{\beta} (N^{\dagger}_{\perp} N_{\perp})_{\beta\kappa}\mathcal{R}_{\kappa\alpha} \Big]+h.c.\, .\\
\end{split}$$
Several words on the nature of the evolution of $\hat\rho_v$ as given by . As we discussed earlier, the only relevant characteristic of a state in the valence Hilbert space is its representation of the color $SU(N)$ (at each spatial point). Therefore the valence Hilbert space on which $\hat\rho_v$ is defined is a direct sum of all the possible subspaces labelled by different representations of $SU(N)$, i.e. the values of all the Casimir operators at each point $\mathcal{J}= \{\hat{j}^a(\mathbf{x}_{\perp} )\hat{j}^a(\mathbf{x}_{\perp}), \ d^{abc}\hat{j}^a(\mathbf{x}_{\perp} )\hat{j}^b(\mathbf{x}_{\perp})\hat{j}^c(\mathbf{x}_{\perp})\ ...\}$, so that $\mathcal{H} = \oplus \mathcal{H}_{\mathcal{J}}$. Since the density matrix itself depends only on $j^a$, it is a block diagonal operator on this Hilbert space and has nonvanishing matrix elements only between states that belong to the same representation ${\mathcal J}$. The same is true for the operators $b_{\alpha}$ and $N_{\alpha\beta}$ since they also are functions of $j^a$ only. Thus if not for the operator ${\mathcal R}$ in Eq. the evolution would mix the matrix elements of $\hat\rho_v$ in a given representation ${\mathcal J}$ only between themselves. The presence of the operator $\hat{\mathcal{R}}$ changes the nature of the evolution. It shifts $j^a$ to $j^a + gT^a$ and therefore mixes matrix elements of $\hat\rho_v$ in one representation with those in another representation with an additional adjoint added in. The operator $\hat{\mathcal{R}}$ is thus the only source of communication between subspaces of different $\mathcal{J}$ in the evolution. Note that even though such cross talk between different representations exists throughout the evolution, the matrix $\hat\rho_v$ remains block diagonal if it was chosen to be block diagonal at the initial rapidity, since for such an initial condition the right hand side of eq. is a function of $j^a$.
Note that Eq. is somewhat more general than the JIMWLK equation. To obtain the original JIMWLK equation (apart from invoking quantum-classical correspondence which is the subject of the next section) one has to expand to second order in in $\Phi^a$. In fact eq. as written here contains both, the JIMWLK limit when expanded in $\Phi^a$ as well as the KLWMIJ limit when expanded in $j^a$. It can therefore be viewed as an interpolating form of the evolution equation for density matrix between the dense and dilute regimes, just like the corresponding equation for the (quasi)probability density functional $W[{\mathbf j}]$ in [@Altinoluk:2009je]. The expansion to leading order in $\hat\Phi^a$ can be performed directly in Eq.. One has to be careful however, since apart from expanding the explicit dependence on $\hat\Phi^a$ in operators ${\mathcal R}$ one also needs to expand the commutators of $\hat\rho$ and $b$. This is easier done in a somewhat roundabout way, namely transforming the operator equation into the equation for the quasi probability function, performing the expansion there, and then returning to the operator equation using the Wigner - Weyl transformation. We will do precisely this in the next section after introducing the quasiclassical correspondence between the quantum dynamics and dynamics on classical phase space. Here we only present the result of this exercise. The evolution equation in the JIMWLK limit turns out to be $$\label{qlin}
\frac{d\hat{\rho}_v}{dy}=-\frac{1}{2\pi} \int d^2\mathbf{z}_{\perp} [\hat Q_i^{a}(\mathbf{z}_{\perp}),[ \hat Q_i^a(\mathbf{z}_{\perp}),\hat\rho_v]]$$ where $$\label{eq:BL_BR1}
\hat Q_i^a({\mathbf x})=\int_{{\mathbf z}}\left[U({\mathbf x})\left(D_i\frac{1}{D^2} - \partial_i\frac{1}{\partial^2}\right)D\partial\right]^{ab}(\mathbf{x},\mathbf{z}) \hat\Phi^b({\mathbf z})\, .$$ and the matrix $U$ is defined as $$\label{u}
U(\mathbf{x}_{\perp}) = \mathcal{P} \mathrm{exp}\left[ ig\int_{\mathcal{C}} d^2\mathbf{y}_{\perp} \cdot b^c(\mathbf{y}_{\perp})T^c \right]$$ with the path $\mathcal{C}$ starting from infinity on the transverse plane and ending at some point $\mathbf{x}_{\perp}$[^1].
We note that this is precisely the equation for density matrix proposed in [@Armesto:2019mna].
A comment is in order on the form of the evolution equation. First we note that Eq. is of the Lindblad type. Thus we find that both in the dilute and dense limits the CGC density matrix satisfies Lindblad type equations. Interestingly however, our interpolating equation Eq., does not have the Lindblad form. One might wonder if this absence of Lindblad form is simply an artifact of our approximation. After all, the calculations that lead to Eq. are under full control only in the two limits. However the reason for deviation from Lindblad in rapidity evolution seems to be very general, and it rather looks like very special conditions have to be satisfied in order for Lindblad form to hold.
We drew an analogy from open quantum systems by treating the valence gluons as “the system" and the soft guons as “the environment". However, time evolution and rapidity evolution are very different concepts. In the former situation, the system degrees of freedom and the environment degrees of freedom are well specified from the beginning and do not change over time. The interaction between the system and the environment is assumed to be Markovian which holds if the system degrees of freedom are slow while the environment degrees of freedom are fast. Time correlations of the environment degrees of freedom are assumed to be local in time compared to the long time scale on which the changes of the system occur, and this leads to Lindblad form of the evolution equation via Eq..
In the case of rapidity evolution, however, the separation between the “environment" - the fast soft gluonic degrees of freedom and the “system“ - the slow valence partons is not fixed, but instead the separation boundary moves together with the evolution parameter. As the rapidity increases one therefore integrates over additional degrees of freedom, namely those whose rapidity label is between the old and new values of the evolution parameter - $\Delta y$. The rapidity thus appears not only as a parameter of the evolution analogous to time, but also as the label of the quantum states which are being integrated out in the process. This integration over additional degrees of freedom is part and parcel of rapidity evolution, and the increment in the density matrix $\Delta\hat\rho_v$ proportional to $\Delta y$ arises due to this integration. Thus the ”Markovian" regime (i.e. short correlation length of the environment modes in rapidity) albeit sufficient to guarantee existence of differential evolution, does not guarantee that this evolution is of Lindblad form. In fact it is easy to trace that the additional integration over the rapidity label is in fact the reason why the general argument that leads to Lindblad form for time evolution in quantum mechanics, is violated in our calculation.
The crucial missing piece is Eq.. As discussed in Section II, in an evolving quantum system for small $\Delta t$ the probability $\hat M^\dagger_n(dt)\hat M_n(dt)\propto dt$ is proportional to $dt$ [*for every environment state $n$*]{} save the vacuum. One can then define the jump operator $\hat L_n$ via $\hat M_n(dt)=\sqrt{dt}\hat L_n$ and the Lindblad form follows. On the other hand, in the case of rapidity evolution the factor $\Delta y$ arises only as a result of the integration over rapidity label of the gluon states in the rapidity window $[y\ ; \ y+\Delta y]$. As a result we have $$\int_{\eta=y}^{y+\Delta y} \hat M^\dagger_n(\eta,\Delta y)\hat M_n(\eta,\Delta y)\propto \Delta y$$ but the probabilities for individual states with fixed $\eta$ do not scale with $\Delta y$. It thus does not appear to be possible in general to define a jump operator unless $\hat M_n(\eta)$ has very special properties. Indeed, examining our calculation, for example in Eq., we realize that the fact that only the terms $\hat M^0_{2n}\hat\rho_v\hat M^{2\dagger}_{2n}+\hat M^2_{2n}\hat\rho_v\hat M^{0\dagger}_{2n}$ contribute at linear order in $\Delta y$ is precisely due to the integration over the rapidities of the $2n$ gluons. This is also the reason this contribution cannot be written in the standard Lindblad form $\Delta y \hat L_{2n}\hat\rho_v\hat L^\dagger_{2n}$. The same is evidently true also for the odd gluon contributions.
Nevertheless in some special cases the Lindblad form may be attainable. For example, if $\hat M(\eta,\Delta y)=\hat M(\Delta y)$, i.e. if the probability of a particular state does not depend on the rapidity of the gluon, the jump operator can indeed be defined. This is precisely the situation we encounter in the derivation in the dilute (KLWMIJ) limit. In this case the coherent operator ${\mathcal C}$ involves only the gluon creation operator integrated over rapidity and as a result the probability $\hat M^\dagger_nM_n$ does not depend on the rapidity label of the gluons in the state $n$. This then allows to take the “square root” of the probability and define the corresponding jump operator $\hat L_n$ which ensures that evolution is in Lindblad form. It is more difficult to trace the origin of the Lindblad form in the dense limit. However, given that JIMWLK and KLWMIJ limits are dual to each other, it is not surprising that such a form indeed exists.
In this section we have discussed the energy evolution in terms of the CGC density matrix. This is not the way it has been formulated in the literature so far. In the next section we show how to relate the two formulations.
From the Lindblad Equation to the JIMWLK Equation via Quantum-Classical Correspondence
======================================================================================
In the previous section we have derived the rapidity evolution equation for the CGC density matrix. To turn this into the conventional JIMWLK evolution equation we will invoke a variant of the Quantum-Classical correspondence, which for simple quantum mechanical systems has been studied 50 years ago, for review see [@Hillery:1983ms]. To start with, we present this analysis as it appears in [@Hillery:1983ms; @Cahill:1969iq; @Agarwal:1971wc; @Agarwal:1971wb].
Quantum mechanics in phase space
--------------------------------
For simplicity, consider a system with one degree of freedom equipped with the canonical variables $\hat{q}$ and $\hat{p}$ that satisfy the cannonical commutation realtion $[\hat{q}, \hat{p}] = i$. The state of the system is described by the density matrix operator $\hat{\rho}$. Observables are expressed as functions of $\hat{q}, \hat{p}$, e.g. $\hat{A}\equiv A[\hat{p}, \hat{q}]$.
Suppose we want to represent a calculation of quantum expectation values in a form similar to averaging over classical distribution in phase space, i.e. $$\mathrm{Tr}(\hat{\rho}\hat{A}) = \int dp dq {\mathcal A}[q, p] W[q,p]$$ This can be achieved if one can find a one to one correspondence between an arbitrary quantum operator $ A(\hat q,\hat p)$ and a corresponding classical function ${\mathcal A}(q,p)$, and additionally similar correspondence for the density matrix $\hat\rho\rightarrow W(q,p)$.
In principle one can devise different mappings that achieve this goal. One widely used mapping is the Wigner-Weyl transformation that maps fully symmetrized quantum operators in the Hilbert space to the corresponding classical functions in the phase space (and back). The familiar form of the Wigner transformation uses eigenstates of the $\hat{q}$ operator and defines the classical phase space functions via $$W[q,p] = \int dz e^{ipz} \langle q-\frac{z}{2} |\hat{\rho}| q + \frac{z}{2}\rangle$$ and $$\label{eq:wigner_transformation}
A_w[q,p] =\int dz e^{ipz} \langle q-\frac{z}{2} |A[\hat{q},\hat{p}]| q + \frac{z}{2}\rangle \, .$$ The Wigner function $W[q,p]$ that corresponds classically to the density matrix is often called the quasi probability distribution function on the phase space. To formulate the mapping in a basis-independent way, one follows Weyl’s correspondence rule which associates fully symmetrized operators in Hilbert space to classical functions in phase space.
Consider the following representation of an operator $G(\hat p,\hat q)$ $$\label{eq:Gpq_fourier_transform}
G[\hat{p}, \hat{q}] = \int f(u, v) e^{i (u\hat{p} + v\hat{q})} du dv\, .$$ This can be regarded as an operator Fourier transformation. First of all, note that the operator $G$ written in this form is necessarily symmetric under permutations of $\hat q$ and $\hat p$. This is straightforward to see by expanding the exponential in Taylor series. This is however not a restriction on the set of operators one can consider, as [*any*]{} operator function can be written in a symmetric form utilizing the commutation relation between $\hat p$ and $\hat q$. The simplest example of such symmetrization is $\hat p\hat q=\frac{1}{2}(\hat p\hat q+\hat q\hat p)-\frac{i}{2}$. One can easily convince oneself that any polynomial of $\hat p$ and $\hat q$ can be written in a symmetric form of this type.
Given the representation eq. we define a classical function on the phase space via $$\label{eq: Fpq_fourier_transform}
F(p,q) = \int f(u, v) e^{i(up + vq)} du dv\, .$$ This is Weyl’s rule for correspondence between quantum operators and classical functions on phase space. Note that under Weyl’s rule, the same Fourier kernel $f(u, v)$ is used in eqs and .
From Eqs. and , the mapping between $F(p,q)$ and $G[\hat{p}, \hat{q}]$ can be represented as $$\label{eq:basis_independent_weyl_mapping}
\begin{split}
& G[\hat{p},\hat{q}] = \int \frac{dp}{2\pi}\frac{dq}{2\pi} F(p,q) \Delta_w(p-\hat{p}, q-\hat{q})\, ,\\
& F(p,q) = \mathrm{Tr} \big( G[\hat{p},\hat{q}]\Delta_w(p-\hat{p}, q-\hat{q})\big)\\
\end{split}$$ with $$\Delta_w(p-\hat{p}, q-\hat{q}) = \int du dv e^{ -i [u(p-\hat{p}) + v(q-\hat{q})]}\, .$$ The Weyl mapping kernel $\Delta_w(p-\hat{p}, q-\hat{q})$ is a functional of both canonical operators $\hat{p},\hat{q}$ and phase space classical variables $q, p$. It has the following properties $$\Delta_w(p-\hat{p}, q-\hat{q}) =\Delta^\dagger_w(p-\hat{p}, q-\hat{q}) ;
\ \ \ \ \ \ \mathrm{Tr} [\Delta_w(p-\hat{p}, q-\hat{q}) ] =1$$ $$\mathrm{Tr} [\Delta_w(p-\hat{p}, q-\hat{q}) \Delta_w^{\dagger}(p^{\prime}-\hat{p}, q^{\prime}-\hat{q}) ] = (2\pi)^2\delta(p-p^{\prime})\delta(q-q^{\prime}).$$
The basis independent mapping in eqs. reproduces the familiar Wigner transformation eq. when the eigenbasis of $\hat{q}$ is used. $$\begin{split}
F(p,q)
&=\int\frac{dq^{\prime}}{2\pi} \langle q^{\prime}| G[\hat{p},\hat{q}]\Delta_w(p-\hat{p}, q-\hat{q})|q^{\prime}\rangle\\
& = \int\frac{dq^{\prime}}{2\pi}\int du dv e^{ -i (up+vq)}e^{iv(q^{\prime}-\frac{1}{2}u)} \langle q^{\prime}|G[\hat{p},\hat{q}]|q^{\prime}-u\rangle \\
& =\int\frac{dq^{\prime}}{2\pi}\int du e^{-iup} 2\pi\delta(q-q^{\prime} + \frac{1}{2}u) \langle q^{\prime}|G[\hat{p},\hat{q}]|q^{\prime}-u\rangle \\
& = \int du e^{-iup} \langle q+\frac{1}{2}u|G[\hat{p},\hat{q}] |q-\frac{1}{2} u\rangle
\end{split}$$ where we have used the Baker-Campbell-Hausdorff formula, and $$e^{iu\hat{p}}|q^{\prime}\rangle = |q^{\prime} -u\rangle; \ \ \ \ \ \
e^{i(u\hat{p} + v\hat{q})}|q^{\prime}\rangle = e^{iu\hat{p}} e^{iv\hat{q}} e^{-\frac{i}{2} uv} |q^{\prime}\rangle = e^{ivq^{\prime}} e^{-\frac{i}{2} uv} |q^{\prime} -u\rangle\, .$$ Thus the Weyl’s correspondence rule provides a basis independent mapping between quantum operators in Hilbert space and classical functions in phase space.\
Using eq. one finds that expectation values of quantum observables can be calculated as weighted integrals in the phase space $$\mathrm{Tr}(\hat{\rho}\hat{A}) = \int dp dq A_w[q, p] W[q,p]$$ with $$W(p,q)=\mathrm{Tr} \big( \hat \rho\Delta_w(p-\hat{p}, q-\hat{q})\big)$$
Note that once the operator $A_s[\hat{q}, \hat{p}]$ is written in a fully symmetrized form with respect to $\hat{q}, \hat{p}$, its associated Wigner-Weyl mapped classical function can be obtained by simply replacing the quantum operators $\hat{q},\hat{p}$ with their classical counterparts $q, p$, $A_s[\hat{q}, \hat{p}] \rightarrow A_s[q, p] = A_w[q,p]$.
Consider now a product of two operators $\hat{F}[\hat{q},\hat{p}] = \hat{A}[\hat{q}, \hat{p}] \hat{B}[\hat{q}, \hat{p}]$. Even if both $\hat{A}[\hat{q},\hat{p}]$ and $\hat{B}[\hat{q},\hat{p}]$ are fully symmetrized, their product as a function of $\hat p$ and $\hat q$ does not necessarily have a fully symmetrized form, and therefore the Wigner-Weyl transform of a product is not equal to product of two Wigner-Weyl transforms. Instead, the correct procedure to obtain the Wigner-Weyl transformation for a product of two observables is $$F_w[q,p] = A_w[q, p] e^{\frac{\Lambda}{2i}} B_w[q,p] = B_w[q,p] e^{-\frac{\Lambda}{2i}} A_w[q,p]$$ with $\Lambda = \frac{\overleftarrow{\partial}}{\partial p} \frac{\overrightarrow{\partial}}{\partial q}-\frac{\overleftarrow{\partial}}{\partial q} \frac{\overrightarrow{\partial}}{\partial p}$, where the derivatives act on functions on the left or on the right as indicated by the arrows.
An alternative representation for the transformation of products of operators can be achieved by introducing the left and right Bopp operators $$\label{boppl}
Q_L = q + \frac{i}{2}\frac{\partial}{\partial p}\, ,\quad P_L = p - \frac{i}{2}\frac{\partial}{\partial q}$$ and $$\label{boppr}
Q_R = q - \frac{i}{2}\frac{\partial}{\partial p}\, , \quad P_R = p + \frac{i}{2}\frac{\partial}{\partial q}\, .$$ The two sets of Bopp operators labelled by “L” and “R” are operators in phase space rather than Hilbert space, as they act on classical functions of $p$ and $q$. Note that the pairs $(Q_L,P_L)$ and $(Q_R,-P_R)$ as operators in phase space form the same Heisenberg algebra as do $(\hat q,\hat p)$ in the Hilbert space. Using Bopp operators, the Wigner-Weyl transformation of a product of two observables is expressed as $$\label{fqp}
F_w[q,p] = A_s[Q_L, P_L] B_w[q,p] = B_s[Q_R,P_R]A_w[q,p]\, .$$ Here $A_s[Q_L, P_L]$ is obtained by replacing $\hat{q}$ and $\hat{p}$ in $ A_s[\hat{q}, \hat{p}]$ by $Q_L$ and $P_L$, respectively, and similarly for $B[Q_R, P_R]$. One can use either set of Bopp operators, depending on whether one uses them in the left factor or the right factor of the product. The two expressions in eq.(\[fqp\]) are equivalent. The action of Bopp operators represents the additional symmetrization rearrangement necessary in order to represent a product of two symmetrized operators in a completely symmetrized form.
As an application of the Wigner-Weyl transformation formalizm, consider the equation of motion for classical quasi distribution $W$ that follows for the quantum Liouville equation for the density matrix $$\label{liouville}
\frac{d\hat{\rho}}{dt} = -i [\hat{H},\hat{\rho}]$$ with the Hamiltonian $\hat{H}\equiv H[\hat{q},\hat{p}]$. Performing the Wigner-Weyl transformation of eq.(\[liouville\]) we obtain $$\begin{split}
\frac{d W[q,p]}{dt} &= -i (H_w[q,p] e^{\frac{\Lambda}{2i}} W[q,p] - W[q,p] e^{\frac{\Lambda}{2i}} H_w[q,p])\\
& = 2 H_w[q,p] \sin{\left(\frac{\Lambda}{2i}\right)} W[q,p]
\end{split}$$ or equivalently using the Bopp operators $$\frac{d W[q,p]}{dt} = -i\left(H[Q_L, P_L] - H[Q_R, P_R]\right) W[q,p]\, .$$ On the other hand, for a fully symmetrized observable $A_s[\hat{q}, \hat{p}]$, the Heisenberg equation is $$\frac{dA_s[\hat{q}, \hat{p}]}{dt} = i [\hat{H}, A_s[\hat{q}, \hat{p}]]\, .$$ Its phase space formulation becomes $$\begin{split}
\frac{dA_s[q,p]}{dt} &= i (H(Q_L, P_L) - H(Q_R, P_R)) A_s[q, p]\\
&= \left\{ i [\hat{H}, A_s[\hat{q}, \hat{p}]]\right\}_{s}\bigg|_{\hat{q}, \hat{p}\rightarrow q, p}
\end{split}$$ In the second line, symmetrization for $H[\hat{q},\hat{p}] A_s[\hat{q}, \hat{p}]$ and $A_s[\hat{q}, \hat{p}]H[\hat{q},\hat{p}] $ with respect to $\hat{q},\hat{p}$ is performed before replacing $\hat{q}, \hat{p}$ with $q, p$, respectively.
Quantum-classical correspondence for $SU(N)$ charges
----------------------------------------------------
In the context of the JIMWLK evolution, the relation between the “probability density functional" $W[\mathbf{j}]$ and the density matrix $\hat\rho$ is similar to that between $W[p,q]$ and $\hat\rho$ in quantum mechanics as described in the previous subsection. In this subsection we make this relation explicit. Much of this discussion already appears in the literature, e.g. in [@Kovner:2005uw], but the relation to the classical-quantum correspondence and Wigner-Weyl transformation has not been elucidated in the past.
Unlike the canonical case discussed above, we are now dealing with the system whose phase space is spanned by the generators of the $SU(N)$ group $j^a(\mathbf{x}_{\perp})$. It is important to stress that the components of color charge density are coordinates on the phase space, and not on configurations space. The Hilbert space of the corresponding quantum system is spanned by the quantum operators $\hat{j}^a(\mathbf{x}_{\perp})$ that satisfy the $SU(N)$ commutation relations $[\hat{j}^a(\mathbf{x}_{\perp}, \hat{j}^b(\mathbf{y}_{\perp})] = igf^{abc}\hat{j}^c(\mathbf{x}_{\perp})\delta(\mathbf{x}_{\perp} - \mathbf{y}_{\perp})$. Note that although the full Hilbert space of CGC requires introduction of the operators $\hat\Phi^a({\mathbf x}_\perp)$, the observables that are currently considered in all calculations are only functions of the color charge density $\hat j^a({\mathbf x}_\perp)$. It is thus sufficient for our purposes to discuss the quantum-classical correspondence for operators that depend only on ${\mathbf j}({\mathbf x}_\perp)$. One must keep in mind however that if one wishes to generalize the framework along the lines of [@Armesto:2019mna], this correspondence has to be extended to include also functions of $\hat\Phi({\mathbf x}_\perp)$.
For quantum systems of spins, most notably the $SU(2)$ group, the mapping between the Hilbert space and the phase space has long been studied [@Stratonovich:1956; @Varilly:1989sv; @Brif:1997km; @Brif:1998pw]. These studies mostly rely on introducing a particular (over)complete basis (ususally generalized coherent states) and working in a fixed representation of the underlying Lie group. Our situation is slightly different, since as discussed above the valence Hilbert space is a direct sum of different representations of $SU(N)$. We therefore cannot fix the representation and instead will rely on the operator properties of the quantum-classical correspondence. The purpose of this section is, drawing analogy to the canonical case to provide the mapping between quantum operators in valence Hilbert space and classical variables in non-Abelian phase space, as well as relation between the quantum density matrix and “classical" quasi probability distribution.
We concentrate on operators in the Hilbert space which can be written as functions of $\hat j^a(\mathbf{x})$ and that are fully symmetric with respect to interchange of the different color components of $\hat{j}^a$. If an operator is not written in a fully symmetrized form, it can always be brought into such form by repeated use of the basic commutation relations of $\hat j^a$. This has been explicitly demonstrated in [@Kovner:2005uw].
We will construct the analog of Weyl’s quantum-classical correspondence where the classical counterpart of a quantum operator is obtained by replacing an operator $\hat{j}^a$ with its classical counterpart $ j^a$ once a quantum operator $ \mathcal{\hat O}$ is expressed in terms of fully symmetrized products of $\hat j^{a_i}$, so that $ \mathcal{\hat O}=\mathcal{O}_s(\hat{j}^a)$ , where the subscript “s” indicates a fully symmetric function. This is the most straightforward generalization of Weyl’s quantum-classical correspondence. In the following the spatial coordinates of $\hat{j}^a(\mathbf{x}_{\perp})$ are suppressed since $\hat{j}^a(\mathbf{x}_{\perp})$ with different transverse coordinates commute. All the nontrivial action therefore happens at the same transverse coordinate.
In analogy with the discussion of the Weyl’s correspondence rules for canonical operators in the previous subsection, we adopt the following rules for correspondence between an operator and a classical function on the phase space $$\label{eq:weyl_rule_color}
F[\mathbf{j}] = \int f(\mathbf{\alpha}) e^{\,i \mathbf{\alpha}\cdot \mathbf{j}} d\mathbf{\alpha}\, ,\quad G_{s}[\hat{\mathbf{j}}] = \int f(\mathbf{\alpha}) e^{\,i \mathbf{\alpha}\cdot\hat{ \mathbf{j}}} d\mathbf{\alpha}\, .$$ Note that for $SU(N_c)$, the color index runs from $1$ to $N_c^2-1$. The above Fourier transformations are understood as $N_c^2-1$-variate transformations. It is easy to see by Taylor expanding the second of eq. that $G_s$ is a fully symmetric function of $\hat j^a$.
Note that eq. is an operator relation, and is not limited to any particular representation of the $SU(N)$ group, but is rather valid on all the valence Hilbert space.
Eq. leads to the following relation between $G_s[\hat{\mathbf{j}}]$ and $F[\mathbf{j}]$ $$\label{eq:operators_to_functions}
G_s[\hat{\mathbf{j}}] = \int\, d\mathbf{j}\, F[\mathbf{j}] \Delta_{W} (\mathbf{j} ,\hat{\mathbf{j}})$$ with the mapping kernel $$\Delta_{W} (\mathbf{j} , \hat{\mathbf{j}}) = \int d\mathbf{\alpha} \, e^{-i \mathbf{\alpha} \cdot \mathbf{j}}e^{ i\mathbf{\alpha}\cdot \hat{\mathbf{j}}}\, .$$ Our definition is such that the integration $\int d\mathbf{j}$ is over all real valued $j^a$ with a simple integration measure on $R^{N^2-1}$.
The $\Delta_W(\mathbf{j}, \hat{\mathbf{j}})$ maps classical functions $F[\mathbf{j}]$ to quantum operators $G_s[\hat{j}^a]$. By requiring that for a fully symmetrized operator $G_s[\hat{\mathbf{j}}]$ the corresponding classical function in phase space is just $G_s[\mathbf{j}]$, so that $F[\mathbf{j}] = G_s[\mathbf{j}]$ we can also find the inverse mapping.
Let us write this mapping in the suggestive form: $$\label{eq:functions_to_operators}
G_s[\mathbf{j}] = \mathrm{Tr} \left(G_s[\hat{\mathbf{j}}] \tilde{\Delta}_W(\mathbf{j}, \hat{\mathbf{j}})\right)\, .$$ Substituting eq. into eq. , one obtains the condition that $\tilde{\Delta}_W[\mathbf{j}, \hat{\mathbf{j}}]$ must satisfy $$\label{eq:kernel_conditions}
\mathrm{Tr}\left(\Delta_W[\mathbf{j}_1, \hat{\mathbf{j}}]\tilde{\Delta}_W[\mathbf{j}_2, \hat{\mathbf{j}}]\right) = \delta(\mathbf{j}_1-\mathbf{j}_2)\,.$$ The following expression of $\tilde{\Delta}_W[\mathbf{j}, \hat{\mathbf{j}}]$ solves this constraint: $$\label{eq:deltatilde}
\tilde{\Delta}_W[\mathbf{j}, \hat{\mathbf{j}}] = \int dg_{\alpha}\, e^{i\alpha\cdot \mathbf{j}} \, d_r(\hat{\mathbf{j}}) \, e^{i\alpha \cdot \hat{\mathbf{j}}}$$ Here $dg_{\alpha}$ is the Haar measure over the $SU(N)$ group. Each group element $g_{\alpha}$ is labelled by the parameters $\alpha$. The factor $d_r(\hat{\mathbf{j}})$ denotes the dimension of the particular representation $r$ and is viewed here as a function of $\hat{\mathbf{j}}$ which depends only on the Casimir operators of the Lie algebra. The function is such that for a given representation its numerical value is equal to the dimension of this representation.
To prove that $\tilde{\Delta}_W[\mathbf{j}, \hat{\mathbf{j}}] $ satisfies eq. , we use the Peter-Weyl theorem [@Barut:1986] for representations of Lie group $$\sum_{r=1}^{\infty} \sum_{j,k=1}^{d_r} d_r D^{(r)T}_{jk}(g_{\alpha_1})D^{(r)}_{jk}(g_{\alpha_2}) = \delta(g_{\alpha_1} - g_{\alpha_2}).$$ Here $D^{(r)}(g_{\alpha})$ is the representation of group element $g_{\alpha}$ and $r$ indicates all the irreducible, inequivalent representations. As above, $d_r$ is the dimension of the representation $r$.
Using eq. in the left hand side of eq., and remembering that summation over all representations $r$ is a part of tracing over the valence Hilbert space, we recover the right hand side of eq..
The above expression is the formal definition of the kernel $\tilde\Delta$, however for all practical purposes one does not need to know its explicit form. This is because the classical counterpart of the operator $ G[\hat{\mathbf j}]$ is simply obtained by substitution $\hat{\mathbf j}\rightarrow {\mathbf j}$ once the operator $\hat G$ is written in the fully symmetrized form.
Now we can establish the relation between quantum average of operators in Hilbert space and phase space weighted integrations $$\mathrm{Tr}(\hat{\rho}G_s[\hat{\mathbf{j}}] ) = \int\, d\mathbf{j} \, G_s[\mathbf{j}] \, \mathrm{Tr}(\hat{\rho} \Delta_{W} (\mathbf{j}, \hat{\mathbf{j}}) ) = \int\, d\mathbf{j} \, G_s[\mathbf{j}] \, W[\mathbf{j}]\, .$$ with the classical weight functional $$\label{eq:density_matrix_to_wigner_function}
W[\mathbf{j}] = \mathrm{Tr}\left(\hat{\rho} \Delta_{W} (\mathbf{j}, \hat{\mathbf{j}})\right)$$ One can check that $\int d\mathbf{j} \, W[\mathbf{j}] =1$ using $\mathrm{Tr} \hat{\rho} =1$ and so the classical weight function $W[\mathbf{j}]$ has the interpretation of quasi probability distribution.
The mapping back from the classical weight functional to the density matrix is through the kernel $\tilde{\Delta}_W[\mathbf{j}, \hat{\mathbf{j}}]$. Again, one does not need an explicit form of $\tilde\Delta$ to perform this mapping. The practical way to do it, is to expand $W[\mathbf j]$ in Taylor series, and then substitute in every term $$j^{a_1}...j^{a_n}\rightarrow \frac{1}{n!}\sum_{i_1,...,i_n}\hat j^{a_{i_1}}...\hat j^{a_{i_n}}$$ where the summation goes over all possible permutations of $(1,...,n)$.
The other issue we need to understand in order to formulate evolution in the classical phase space approach is how to extend the mapping for products of quantum operators. In principle, this involves generalizing Moyal’s star-product to a general Lie algebra. However, rather than taking this general mathematical approach, the particular realization for $SU(N)$ algebra has been worked out in [@Kovner:2005uw; @Altinoluk:2013rua]. Consider a product of two operators with each one written in the symmetrized form $$G[\hat{\mathbf{j}}]=A_s[\hat{\mathbf{j}}]B_s[\hat{\mathbf{j}}]$$ The analog of the transformation eq. for the present case is $$G_s[{\mathbf{j}}]=A_s[{\mathbf{j}}_L]B_s[{\mathbf{j}}]=B_s[{\mathbf{j}}_R]A_s[{\mathbf{j}}]$$ where the appropriate Bopp operators are defined as $$\begin{split}
j_L^a =&j^b\left[\frac{\tau}{2}\coth{\frac{\tau}{2}} + \frac{\tau}{2}\right]^{ba}\\
= &j^a + \frac{1}{2} j^b \left(gT^e\frac{\delta}{\delta j^e}\right)_{ba}+\frac{1}{12} j^b \left(gT^e\frac{\delta}{\delta j^e}\right)^2_{ba} -\frac{1}{720}j^b\left(gT^e\frac{\delta}{\delta j^e}\right)^4_{ba} +\ldots\\
j_R^a =&j^b\left[\frac{\tau}{2}\coth{\frac{\tau}{2}} - \frac{\tau}{2}\right]^{ba}\\
=& j^a - \frac{1}{2} j^b \left(gT^e\frac{\delta}{\delta j^e} \right)_{ba}+\frac{1}{12} j^b \left(gT^e\frac{\delta}{\delta j^e}\right)^2_{ba} -\frac{1}{720}j^b\left(gT^e\frac{\delta}{\delta j^e}\right)^4_{ba}+ \ldots\\
\end{split}$$ with $$\tau = gT^e\frac{\delta}{\delta j^e}$$
The Bopp operators $j^a_L$ and $j^b_R$ act on functions in the phase space rather than the Hilbert space. It is straightforward if somewhat tedious to explicitly check that, similarly to Bopp operators defined in Eqs.,, the $SU(N)$ phase space Bopp operators $j^a_L$ and $-j^a_R$ form the same $SU(N)$ algebra as the operators $\hat{j}^a$ on the Hilbert space. In addition, $j^a_L$ and $j^b_R$ commute $[j^a_L, j^b_R] =0$.
We find it interesting to note that the functional form of the Bopp operators $j^a_R$ involves exactly the same function as in eq., which ensured correct operator properties of the charge shift operator $\hat R$.
As a corollary to this discussion consider Hermitian conjugation in Hilbert space $$(\hat A\hat B)^\dagger=\hat B^\dagger \hat A^\dagger$$ As discussed above the classical correspondence is $$\hat A\hat B\rightarrow A_s(J_L)B_s(j); \ \ \ \ \ \ \hat B^\dagger\hat A^\dagger\rightarrow B^*(j)A^*(J_R)$$ Thus the Hermitian conjugation operation is represented by complex conjugation in conjunction with changing left (right) Bopp operators int right (left) Bopp operators $$\label{herm}
(... )^\dagger\rightarrow (L\leftrightarrow R)^*$$
The evolution equation for the quasi probability distribution.
--------------------------------------------------------------
Using the correspondence rules described above we can now rewrite the evolution equations eq. , for the density matrix as the evolution equation for the quasi probability distribution $W[{\mathbf{j}}]$.
The right hand side of eq. contains product of operators $(N_{\perp}^{\dagger}N_{\perp})$, $b$, $\mathcal{R}$ and $\hat\rho$. Performing Wigner-Weyl transformation the density matrix $\hat\rho$ becomes $W[j^a(\mathbf{x})]$. The operator $b$ becomes $b_L^{\alpha}=b^{\alpha}[j_L^a(\mathbf{x}_{\perp}); \mathbf{x}_{\perp}]$ or $b_R^{\alpha}=b^{\alpha}[j_R^a(\mathbf{x}_{\perp}); \mathbf{x}_{\perp}]$, depending on its position relative to the factor $\hat\rho$ in eq.. The operator $N_{\perp}^{\dagger}N_{\perp}$ becomes $N_{\perp,R}^{\dagger}N_{\perp,R}$ with $N_{\perp, R} \equiv N_{\perp}[j^a_R(\mathbf{x}_{\perp})]$.
Additionally we need to understand how the operator $\hat{\mathcal{R}}$ is mapped to the phase space. To do this, we note that for a fully symmetrized operator $\hat{O}_s[\hat{j^a}]$, the action of $\hat{\mathcal{R}}$ operator is $$\hat{\mathcal{R}}\hat{O}_s[j^a] \hat{\mathcal{R}} ^{\dagger}= \hat{O}_s[\hat{j}^a +gT^a] \rightarrow \mathcal{R}_p O_s[j^a]$$ with the phase space shift operator $$\mathcal{R}_p = e^{gT^a\frac{\delta}{\delta j^a}}$$ Therefore, we should simply replace the action of the operator $\hat{\mathcal{R}}$ with the phase space shift operator $\mathcal{R}_p$.
It is now straightforward to write the equation for $W$. We obtain
$$\label{eq:phase_space_lindblad}
\frac{d W[j^a]}{dy} = -\frac{1}{4\pi} \Bigg[(\tilde{b}_L^{\alpha} - b_R^{\alpha})^{\dagger}(N^{\dagger}_{\perp,R}N_{\perp,R} )_{\alpha\beta}(\tilde{b}_L^{\beta} -b_R^{\beta}) +h.c.\Bigg]W[j^a]$$
with $\tilde{b}_L^{\alpha} = b_L^{\beta}\mathcal{R}_{p}^{\beta\alpha}$ and hermitian conjugation defined in Eq.. Eq. is the final form of the evolution equation in the classical phase space formulation. When $j^a$ are considered as coordinates on a classical phase space, this equation is interpretable as a Focker-Planck equation for the quasi probability phase space distribution $W$.
One can now take various limits to reproduce the results known in the literature. In particular, assuming that ${\mathbf{j}}$ is small and expanding the right hand side of eq. to second order in ${\mathbf{j}}$ one straightforwardly recovers the so called KLWMIJ equation [@Kovner:2005nq],[@Kovner:2005en].
Alternatively, keeping all orders in ${\mathbf{j}}$, but expanding to second order in $\delta/\delta{\mathbf{j}}$ one reproduces the JIMWLK equation. This last expansion is a little more involved, but it is performed explicitly in [@Kovner:2007zu]. We reproduce the derivation here for completeness.
To reproduce the JIMWLK kernel, we truncate $\mathcal{R}_p$ to first order in $\delta/\delta j^a$ and expand $b_L$ and $b_R$ around $b(j^a)$. We only need to keep first order terms, since Eq. contains a factor $(\tilde b_L-b_R)^2$.
At this order there is no need to expand $N_{\perp,R}$ or $N_{\perp,L}$ so that both are substituted by $N_\perp (j)$.We have $$\begin{split}
&b_L\mathcal{R}_p - b_R = b_i^a[j_L] \mathcal{R}^{ab}_p - b_i^b[j_R] \\
\simeq & \left(b_i^a[j] + \frac{\delta b_i^a}{\delta j^e}(j_L^e -j^e)\right)\left(\delta_{ab}+ gT^d_{ab}\frac{\delta}{\delta j^d}\right) - \left(b_i^b[j] +\frac{\delta b_i^b}{\delta j^e} (j_R^e -j^e)\right)\\
\simeq&gb_i^aT^d_{ab}\frac{\delta}{\delta j^d} + \frac{\delta b_i^b}{\delta j^e}g j^cT^d_{ce}\frac{\delta}{\delta j^d}\\
=&gb_i^a(\mathbf{x}_{\perp})T^d_{ab}\frac{\delta}{\delta j^d(\mathbf{x}_{\perp})} + \int d^2\mathbf{z}_{\perp}\frac{\delta b_i^b(\mathbf{x}_{\perp})}{\delta j^e(\mathbf{z}_{\perp})} gj^c(\mathbf{z}_{\perp})T^d_{ce}\frac{\delta}{\delta j^d(\mathbf{x}_{\perp})}\\
=&i\left[\partial_i - D_i\frac{1}{\partial D} D\partial \right]^{bd}(\mathbf{x},\mathbf{z})\, \frac{\delta}{\delta j^d(\mathbf{z}_{\perp})}
\end{split}$$
In the last line we have used $igT^{e}_{ba}b_i^e = \delta^{ab}\partial_i - D_i^{ab} $ and $
igT^e_{ba}j^e = -(\partial D - D\partial)^{ab} = igj^eT^a_{eb} $ as well as $
\frac{\delta b_i^b(\mathbf{x}_{\perp})}{\delta j^e(\mathbf{z}_{\perp})} = \left[D_i\frac{1}{\partial D}\right]^{be}(\mathbf{x}_{\perp} ,\mathbf{z}_{\perp})\, $. Additionally using eq. , we calculate $$\label{eq:BL_BR}
Q_i^a({\mathbf x})\equiv [UN_{\perp} (b_L\mathcal{R}_p -b_R)]^a_i({\mathbf x})=-i\left[U({\mathbf x})\left(D_i\frac{1}{D^2} - \partial_i\frac{1}{\partial^2}\right)D\partial\right]^{ab}(\mathbf{x},\mathbf{z}) \frac{\delta }{\delta j^b(\mathbf{z})}\, .$$ where the standard eikonal scattering matrix $U$ is defined in Eq.. Eq. now becomes $$\label{halfj}
\frac{d W[{\mathbf j}]}{dy} = -\frac{1}{2\pi} \int_{\mathbf x}Q_i^{a\dagger}({\mathbf x)}Q_i^a{(\mathbf x})W[{\mathbf j}]$$
One can express the JIMWLK equation in terms of single gluon scattering matrix $U(\mathbf{x}_{\perp})^{ab}$ rather than the color charge density $j^a(\mathbf{x}_{\perp})$. The relation between the two was derived in [@Kovner:2007zu]. Now we can relate functional derivatives with respect to the $U$ matrix to those with respect to the color current. $$\frac{\delta}{\delta j^a(\mathbf{x}_{\perp})} = \int d\mathbf{z}_{\perp} \frac{\delta U^{cd}(\mathbf{z}_{\perp})}{\delta j^a(\mathbf{x}_{\perp})} \frac{\delta}{\delta U^{cd}(\mathbf{z}_{\perp})}$$ with $$\begin{split}
\frac{\delta U^{cd}(\mathbf{z}_{\perp})}{\delta j^a(\mathbf{x}_{\perp})} &=\int d^2\mathbf{y}_{\perp} \frac{\delta U^{cd}(\mathbf{z}_{\perp})}{\delta b_l^b(\mathbf{y}_{\perp})}\frac{\delta b_l^b(\mathbf{y}_{\perp})}{\delta j^a(\mathbf{x}_{\perp})} \\
& =ig \int_{\mathcal{C}} d^2 y_l \left(U^{\dagger bm}(\mathbf{y}_{\perp})\, T^{m}_{cn}\right) U^{nd}(\mathbf{z}_{\perp}) \left[D_l\frac{1}{\partial D}\right]^{ba}(\mathbf{y}_{\perp} , \mathbf{x}_{\perp})\\
& = ig \int_{\mathcal{C}} d^2 y_l \left(U^{\dagger bm}(\mathbf{y}_{\perp})\, T^{m}_{cn}\right) U^{nd}(\mathbf{z}_{\perp}) \left[U^{\dagger} \partial_l U\frac{1}{\partial D}\right]^{ba}(\mathbf{y}_{\perp} , \mathbf{x}_{\perp})\\
&= igT^{e}_{cn}U^{nd}(\mathbf{z}_{\perp}) \int_{\mathcal{C}} d^2 y_l \left[ \partial_l U\frac{1}{\partial D}\right]^{ea}(\mathbf{y}_{\perp} , \mathbf{x}_{\perp})\\
&= igT^{e}_{cn}U^{nd}(\mathbf{z}_{\perp}) U^{eb}(\mathbf{z}_{\perp})\left[ \frac{1}{\partial D}\right]^{ba}(\mathbf{z}_{\perp} ,\mathbf{x}_{\perp}) \\
& = ig U^{cm}(\mathbf{z}_{\perp}) T^{m}_{db} \left[ \frac{1}{\partial D}\right]^{ba}(\mathbf{z}_{\perp}, \mathbf{x}_{\perp}) \\
& =- ig\left[ U(\mathbf{z}_{\perp}) T^d \frac{1}{\partial D} \right]^{ca}(\mathbf{z}_{\perp} , \mathbf{x}_{\perp})
\end{split}$$ In the above we have used $U^{\dagger} T^a U = U^{ab}T^b$ and $D_l = U^{\dagger}\partial_l U$. We have also used $$\begin{split}
\frac{\delta U^{cd}(\mathbf{z}_{\perp})}{\delta b_l^b(\mathbf{y}_{\perp})} =& U^{ce}(\mathbf{y}_{\perp})( ig T^b_{ef} )\int_{\mathcal{C}} d\mathbf{w}_l \delta(\mathbf{y}_{\perp} - \mathbf{w}_{\perp})[U^{\dagger}(\mathbf{y}_{\perp}) U(\mathbf{z}_{\perp})]^{fd}\\
=&ig \left(U^{\dagger bm}(\mathbf{y}_{\perp})\, T^{m}_{cn}\right) U^{nd}(\mathbf{z}_{\perp})\int_{\mathcal{C}} d\mathbf{w}_l \delta(\mathbf{y}_{\perp} - \mathbf{w}_{\perp})
\end{split}$$ Finally, we obtain $$\begin{split}
\frac{\delta}{\delta j^a(\mathbf{x}_{\perp})} &= \int d\mathbf{z}_{\perp} \frac{\delta U^{cd}(\mathbf{z}_{\perp})}{\delta j^a(\mathbf{x}_{\perp})} \frac{\delta}{\delta U^{cd}(\mathbf{z}_{\perp})}\\
&=ig\int d\mathbf{z}_{\perp}U^{cm}(\mathbf{z}_{\perp}) T^{m}_{db} \left[ \frac{1}{\partial D}\right]^{ba}(\mathbf{z}_{\perp}, \mathbf{x}_{\perp})\frac{\delta}{\delta U^{cd}(\mathbf{z}_{\perp})}\\
&=ig\int d\mathbf{z}_{\perp} \mathrm{Tr}\left(U(\mathbf{z}_{\perp}) T^b \frac{\delta}{\delta U^{\dagger}(\mathbf{z}_{\perp})}\right) \left[\frac{1}{\partial D} \right]^{ba}(\mathbf{z}_{\perp} , \mathbf{x}_{\perp})\\
&= -ig \int d\mathbf{z}_{\perp} \left[\frac{1}{ D\partial} \right]^{ae}(\mathbf{x}_{\perp} , \mathbf{z}_{\perp})\mathcal{J}^e_R(\mathbf{z}_{\perp})\, .
\end{split}$$ We have denoted $$\mathcal{J}^b_R(\mathbf{z}_{\perp}) = -\mathrm{Tr}\left(U(\mathbf{z}_{\perp}) T^b \frac{\delta}{\delta U^{\dagger}(\mathbf{z}_{\perp})}\right)\, .$$ Also recall that $$\partial_i \frac{1}{\partial^2}(\mathbf{x}_{\perp}, \mathbf{y}_{\perp}) = \frac{1}{2\pi} \frac{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})_i}{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})^2}\,, \quad D_i \frac{1}{D^2}(\mathbf{x}_{\perp}, \mathbf{y}_{\perp}) = \frac{1}{2\pi}U^{\dagger}(\mathbf{x}_{\perp}) \frac{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})_i}{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})^2}U(\mathbf{y}_{\perp})\,.$$ Finally, eq. becomes $$\label{qu}
Q_i^a[\mathbf{x}_{\perp};U] = -\frac{g}{2\pi}\int d\mathbf{y}_{\perp} \frac{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})_i}{(\mathbf{x}_{\perp}-\mathbf{y}_{\perp})^2} \left[ U(\mathbf{y}_{\perp}) -U(\mathbf{x}_{\perp})\right]^{ab} \mathcal{J}_R^b(\mathbf{y}_{\perp})$$ The standard JIMWLK kernel is reproduced as $$\begin{split}
& -\frac{1}{2\pi} \int d^2\mathbf{z}_{\perp} Q_i^{a\dagger}[\mathbf{z}_{\perp}; U] Q_i^a[\mathbf{z}_{\perp}; U] \\
=& - \frac{\alpha_s}{2\pi^2} \int_{\mathbf{z}_{\perp}, \mathbf{y}_{\perp},\mathbf{x}_{\perp}}\frac{(\mathbf{z}_{\perp} - \mathbf{x}_{\perp})_i}{(\mathbf{z}_{\perp} - \mathbf{x}_{\perp})^2}\frac{(\mathbf{z}_{\perp} - \mathbf{y}_{\perp})_i}{(\mathbf{z}_{\perp} - \mathbf{y}_{\perp})^2}\mathcal{J}^e_{R}(\mathbf{y}_{\perp})\\
&\times [ 1+ U^{\dagger}(\mathbf{y}_{\perp}) U(\mathbf{x}_{\perp}) - U^{\dagger}(\mathbf{y}_{\perp}) U(\mathbf{z}_{\perp}) - U^{\dagger}(\mathbf{z}_{\perp}) U(\mathbf{x}_{\perp})]^{ed} \mathcal{J}^d_R(\mathbf{x}_{\perp})\, .
\end{split}$$
From JIMWLK back to Lindblad.
-----------------------------
Finally using Eq. and Eq. we can transform the evolution equation back to Hilbert space. First we note, that the amplitude $Q_i^a$ is hermitian as an operator on the phase space. This is obvious from Eq., since $\mathcal{J}_R^b(\mathbf{y}_{\perp})$ can be commuted to the left through the factors of $U$, as the right color index on $U$ is contracted with the index of $ \mathcal{J}_R^b(\mathbf{y}_{\perp})$. One can then write the JIMWLK equation as $$\frac{d W[{\mathbf j}]}{dy} = -\frac{1}{2\pi} \int_{\mathbf x}\big[Q_i^{a}({\mathbf x)},\big[Q_i^a{(\mathbf x}),W[{\mathbf j}]\big] \big]$$ where the commutator is understood as the commutator of the operators on phase space. Transforming this back to Hilbert space we see that to this order all we need to do is substitute $-i\frac{\delta }{\delta j^a(\mathbf{z})}\rightarrow \hat\Phi^a(\mathbf{z})$ and keep the structure of the double commutator. This procedure gives Eq. as claimed.
Discussion
==========
This paper is devoted to analysis of the high energy limit of hadronic scattering, and its energy evolution formulated as effective quantum theory.
The dynamics of this effective theory is governed by a density matrix. Here we were able to define this “reduced” density matrix in a way reminiscent to an open quantum system, i.e. bi partitioning the degrees of freedom into the “system” and “environment” and integrating over the environment. In the present case the bi partitioning is into the “valence" gluons as the system and “soft“ gluons as the environment. Despite some similarities, there is a significant difference between the high energy limit considered here and bi partitioning in a quantum open system. In a quantum system one normally integrates completely the environment and then considers only observables that depend on the degrees of freedom of the ”system". This is not the case for the high energy limit, since the soft gluons contribute nontrivially to the color charge density, which is the basic observable in the effective theory. Defining the reduced density matrix is therefore rather nontrivial. Nevertheless we were able to do it.
We have then followed the usual assumption made in the derivation of the high energy evolution, i.e. that only the distribution of the color charge density in the transverse plane is relevant for determining hadronic properties at high energy. Assuming that the reduced density matrix of a hadronic system depends only on the components of the color charge density results in a quasi diagonal density matrix in the sense that its matrix elements between states belonging to different color representation vanish. This is true both in dense and dilute limits, and generalizes the notion of diagonal density matrix discussed in [@Armesto:2019mna] to arbitrary parametric values of color charge density.
Under this assumption we have shown that the rapidity evolution of the reduced density matrix is of the Lindblad type in the two limiting cases - the dense (JIMWLK) and the dilute (KLWMIJ) limits. This is true even though the nature of the energy evolution is in principle quite different from the nature of time evolution of a dynamical quantum system.
Interestingly the evolution equation that interpolates between the two limits, Eq. does not have a Lindblad form. Although the derivation of this interpolating equation is not under parametric control, the basic features of the derivation are generic, and our analysis shows that the absence of Lindblad form should be a rule rather than exception. The basic reason is that the rapidity plays a dual role in high energy evolution: it is the analog of the evolution time on one hand, and is a quantum number that labels the quantum states of the environment that are integrated out on the other hand. This invalidates in principle the usual argument for the Lindblad form of the differential evolution equation.
Finally, we have shown how to rigorously relate the reduced density matrix description of the evolution with the approach used in most pertinent literature based on the probability density functional $W[\mathbf{j}]$. To this end we have explored the Wigner-Weyl transformation, which maps the Hilbert space description of a quantum system in terms of density matrix $\hat\rho$ into the classical phase space description in terms of quasi probability distribution $W$. By adapting this transformation to the present case we have shown that the Lindblad evolution equation for $\hat\rho$ is indeed equivalent to a Fokker-Planck type equation for $W$, where the components of the color charge density $\mathbf{j}$ are considered as coordinates on a classical non Abelian phase space. This Fokker-Planck equation reduces to KLWMIJ and JIMWLK equations in the appropriate dilute and dense limits.
In quantum optics, it has been known for decades that the Lindblad master equation maps to the Fokker-Planck equation through quantum-classical correspondence. Here we have established the same in the context of high energy evolution with the JIMWLK (or KLWMIJ) playing the role of the Focker-Planck equation.
We stress again that this paper deals only with the conventional JIMWLK/KLWMIJ setup, where the density matrix is assumed to depend only on the color charge density degrees of freedom. Recently it was suggested that this framework may be too restrictive and may not be adequate for studying some interesting observables at high energy [@Armesto:2019mna]. Such observables, like correlations between the transverse momentum and density in the transverse plane may be formally subleading at high energy, but could be of great interest in the study of correlations in particle production. In order to include these into consideration one has to extend the conventional framework and allow for density matrices that depend not just on color charge density $\hat{\mathbf{j}}$, but also on their conjugate variables, which in the present paper we have identified as the operators $\hat\Phi^a$. It was suggested in [@Armesto:2019mna] that the evolution of this more general density matrix is also given by the same Lindblad equation. Although this seems very likely to be the case, our current derivation does not cover this interesting more general situation. It should be possible to extend our current method to deal with this intriguing problem. This investigation is currently under way.
We now comment on several questions/issues that arise from our results.
First, the fact that the evolution of the reduced density matrix beyond dense-dilute limit is most likely not of Lindblad type begs an interesting general question. It is known that Lindblad equation preserves the properties of the density matrix, namely normalization and positivity. Is this also the case for Eq. even though it is not in Lindblad form? It is quite obvious that the normalization of the density matrix is preserved under Eq., since its right hand side is a commutator, and therefore has a vanishing trace. As for the positivity, it is more difficult to establish. We note however, that the differential evolution follows from the Krauss representation Eq. which does preserve positivity [@Preskill:2019]. We therefore believe that the differential evolution Eq. does indeed preserve positivity and thus is a consistent evolution of a density matrix. If this is the case, one is lead to a general conclusion that the set of possible differential evolutions of a density matrix is not limited to equations of Lindblad type.
Second, we note that one of the useful perspectives on the JIMWLK evolution is that of a Langevin equation for Brownian motion in the space of Wilson line $U^{ab}(\mathbf{x}_{\perp})$ [@Weigert:2000gi; @Blaizot:2002np]. The bi partitioning into the “system" of the hard gluons and “environment" due to the soft gluons harmonizes nicely with the random walk picture. After a boost by $\Delta y$, the soft gluons can be emitted into any of the multigluon Fock states $\{ |n\rangle \}$. This emission contributes to a random addition to the color charge density $j^a_{y+\Delta y} \sim j^a_{y} + \delta j^a$ with $\delta j^a$ being a random variable, which therefore random walks in the color space. The Langevin equation is a reformulation of the Focker-Planck equation, which is equivalent to JIMWLK. It is then interesting to ask whether such a Langevin description can be extended beyond the leading order. The NLO JIMWLK equation has been derived some years ago [@nlo:balitsky; @nlo:kovner; @nlo:lublinsky; @nlo:caron-huot]. Naturally the derivation involves integration over gluon and quark degrees of freedom in the rapidity interval $\Delta y$. As opposed to the leading order, where as we discussed the probability to create all single gluon states is equal, independent of their rapidities, at NLO there is a genuine integration over rapidity of two soft parton states. This suggests that the evolution equation for the density matrix is not of Lindblad type for the same reason Eq. is not. If that is the case one does not expect it to be equivalent to a Focker-Planck equation for the quasi probability function and thus the Langevin description may well not be possible.
We hope that the new perspective on high energy evolution discussed in this paper will be useful not only for a more fundamental understanding of JIMWLK equation but will also prove useful for future developments.
The Operator $\hat{\Phi}^a$
===========================
In this Appendix we present the derivation of the quantum “phase" operator $\hat\Phi^a$, which via eq. defines the quantum shift operator of the color charge density. Part of this derivation appears in the text, but we keep here all the details for completeness.
We are looking for $M^{ab}(\Phi)$, as a functinal of operator $\hat{\Phi}^a$ that satisfies the commutation relations $$\begin{split}
&[\hat{\Phi}^a, \hat{\Phi}^b] =0\, ,\\
&[\hat{\Phi}^a, \hat{j}^b] = M^{ab}(\Phi)\, \\
\end{split}$$ so that the color charge density shift operation is $$\label{eq:shift_requirement1}
\mathrm{exp}\left\{i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a\right\} \hat{j}^e\, \mathrm{exp}\left\{-i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a\right\} = \hat{j}^e + \hat{j}^e_{\mathrm{soft}}\, .$$ Using the Baker-Hausdaurff formula $$e^X Y e^{-X} = Y + [X, Y] + \frac{1}{2!}[X, [X, Y]] + \frac{1}{3!}[X,[X,[X,Y]]]+\ldots + \frac{1}{n!}[X, [X, [\ldots [X, Y]\ldots]]] + \ldots$$ and expanding eq. in commutators, the first three terms are $$[i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] = i\hat{j}^a_{\mathrm{soft}} M^{ae}(\Phi)\, ,$$ $$\begin{split}
\frac{1}{2!}[i\hat{j}^b_{\mathrm{soft}} \hat{\Phi}^b, [i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] ] &= \frac{1}{2!} [\hat{j}^b_{\mathrm{soft}}, \hat{j}^a_{\mathrm{soft}}] i^2 \hat{\Phi}^b M^{ae}(\Phi) = \frac{1}{2!} igf^{bac} \hat{j}^{c}_{\mathrm{soft}} i^2 \hat{\Phi}^b M^{ae}(\Phi) \\
&= \frac{1}{2!}\, i\hat{j}^c_{\mathrm{soft}} \left(igT^b_{ca} \hat{\Phi}^b \right)M^{ae}(\Phi)\\
&= \frac{1}{2!} \, i\hat{j}^a_{\mathrm{soft}} \left(igT^b \hat{\Phi}^b M(\Phi)\right)_{ae}\, ,\\
\end{split}$$ $$\frac{1}{3!}[i\hat{j}^c_{\mathrm{soft}} \hat{\Phi}^c,[i\hat{j}^b_{\mathrm{soft}} \hat{\Phi}^b, [i\hat{j}^a_{\mathrm{soft}} \hat{\Phi}^a, \hat{j}^e] ] ] = \frac{1}{3!} i\hat{j}^a_{\mathrm{soft}} \left((igT^b \hat{\Phi}^b)^2 M(\Phi)\right)_{ae}\, .$$ Here we have used $-if^{abc} = T^a_{bc}$.
From the above explicit calculations, it is natural to make the following ansatz $$M^{ab}(\Phi) = -i\sum_{n=0}^{\infty} c_n \left[\chi^n\right]_{ab}\, , \quad\mathrm{with}\,\,\, \chi = igT^b\hat{\Phi}^b\, .$$ Clearly, $c_0 =1$ follows from the requirement eq.. This requirement further imposes the constraint $$i + \sum_{k=0}^{\infty} \frac{1}{(k+1)!} \left[\chi^k M(\chi)\right]^{ab}=0$$ which after substituting the ansatz for $M(\chi)$ becomes $$1= \sum_{k=0}^{\infty}\sum_{m=0}^{\infty}\frac{c_m}{(k+1)!} \chi^{k+m}_{ab}\, .$$ Note that for $k=0, m=0$, the $c_0=1$ automatically satisfies the above condition. For $N=k+m\geq 1$, the coefficients of $\chi^N$ have to be vanishing, one then obtains $$\sum_{m=0}^N \frac{c_m}{(N-m+1)!} =0$$ which is equivalent to the following recursive relations $$\label{eq:recursive_relations1}
c_N = -\sum_{m=0}^{N-1} \frac{c_m}{(N-m+1)!} \, , \quad \mathrm{with}\,\, c_0=1.$$ A few examples can be explicitly calculated $$\begin{split}
& c_1 = -\frac{c_0}{2!} = -\frac{1}{2} \,, \\
&c_2 = - \frac{c_0}{3!} - \frac{c_1}{2!} = \frac{1}{12} \, , \\
&c_3 = -\frac{c_0}{4!} - \frac{c_1}{3!} - \frac{c_2}{2!} = 0\, ,\\
&c_4 = -\frac{c_0}{5!} - \frac{c_1}{4!} -\frac{c_2}{3!} -\frac{c_3}{2!} = - \frac{1}{6!} = -\frac{1}{720}\,, \\
&c_5 = -\frac{c_0}{6!} -\frac{c_1}{5!} -\frac{c_2}{4!}-\frac{c_4}{2!} =0\, , \\
&c_6 = -\frac{c_0}{7!} -\frac{c_1}{6!} -\frac{c_2}{5!} -\frac{c_4}{3!} = \frac{1}{6}\frac{1}{7!} =\frac{1}{30240}\, ,\\
&c_7 = -\frac{c_0}{8!} -\frac{c_1}{7!} -\frac{c_2}{6!} -\frac{c_4}{4!}-\frac{c_6}{2!} = 0 \, , \\
&c_8 = -\frac{c_0}{9!} -\frac{c_1}{8!} -\frac{c_2}{7!} -\frac{c_4}{5!}-\frac{c_6}{3!} = - \frac{1}{30}\frac{1}{8!} = -\frac{1}{1209600}\, .
\end{split}$$ It turns out that these numbers correspond to the coefficients in expanding the function $$\label{eq:M_expression1}
M^{ab}(\chi) = -i \left[\frac{\chi}{2} \coth{\frac{\chi}{2}} - \frac{\chi}{2}\right]^{ab}\, .$$ One can now explicitly check that Taylor expansion of eq. in $\chi$ reproduces all the coefficients calculated using the recursive relations in eq. .
In addition one needs to check the consistence of the Lie algebra constructed from $\hat{\Phi}^a$ and $\hat{j}^a$. This consistency requires that the following Jacobi identity holds $$[[\hat{\Phi}^a, \hat{j}^b], \hat{j}^c] + [[\hat{j}^b, \hat{j}^c],\hat{\Phi}^a] + [[\hat{j}^c, \hat{\Phi}^a], \hat{j}^b] =0\, ,$$ which is equivalent to $$\label{eq:consistence_check1}
[M^{ab}, \hat{j}^c] - [M^{ac}, \hat{j}^b] = igf^{bcd} M^{ad}\, .$$ On the left hand side $$\begin{split}
[M^{ab}, \hat{j}^c] &= -i\sum_{n=0}^{\infty}c_n [\chi^n_{ab}, \hat{j}^c]\\
&=-i\sum_{n=1}^{\infty} c_n \sum_{k=0}^{n-1} \chi^k_{ae_1}\, [\chi_{e_1e_2}, \hat{j}^c] \, \chi^{n-1-k}_{e_2b}\\
&=-i\sum_{n=1}^{\infty} c_n \sum_{k=0}^{n-1} \chi^k_{ae_1}\, ig T^d_{e_1e_2}[\hat{\Phi}^d, \hat{j}^c] \, \chi^{n-1-k}_{e_2b}\\
&=-i\sum_{n=1}^{\infty} c_n \sum_{k=0}^{n-1} \chi^k_{ae_1}\, ig T^d_{e_1e_2} \, \chi^{n-1-k}_{e_2b} M^{dc}\\
&=(-i)^2\sum_{n=1}^{\infty} c_n \sum_{k=0}^{n-1} \chi^k_{ae_1}\, ig T^d_{e_1e_2} \, \chi^{n-1-k}_{e_2b} \sum_{m=0}^{\infty} c_m \chi^m_{dc}\, .\\
\end{split}$$ On the right hand side $$igf^{bcd}M^{ad} = \sum_{N=0}^{\infty} gf^{bcd}c_N \chi^N_{ad}\, .$$ We here check the consistency condition eq. order by order to verify that it indeed is satisfied. We do not have a proof for the case of general N, but we believe the same procedure can be carried out to high order terms.
The left hand side of eq. has $N= m+n-1$ in terms of power of $\hat{\Phi}^a$. Also note that $m\geq 0$ and $n\geq 1$. In the following, we calculate the cases $N=0, 1, 2, 3$ in details.
$N=0$ {#n0 .unnumbered}
-----
For $N=0$, the right hand side is $gf^{bcd} c_0\delta_{ad} = gf^{bca}$, the left hand side can have $n=1, m=0$ $$\begin{split}
&(-i)^2 c_0c_1 \delta_{ae_1}\left( igT^{d}_{e_1e_2}\right)\delta_{e_2b} \delta_{dc} -(b\leftrightarrow c)\\
=&\frac{1}{2} igT^c_{ab} -(b\leftrightarrow c) = \frac{1}{2} ig (-if^{cab}) - (b\leftrightarrow c)\\
=& gf^{abc}\, .
\end{split}$$ For $N=0$ terms eq. holds.
$N=1$ {#n1 .unnumbered}
-----
For $N=1$, the right hand side becomes $$gf^{bcd} c_1 \chi_{ad} = -\frac{1}{2} gf^{bcd} (igT^e_{ad} \Phi^e) = \frac{1}{2}g^2 T^d_{bc} T^e_{ad} \Phi^e\, .$$ For the left hand side $N=m+n-1 =1$, we have two possibilities $n=1, m=1$ and $n=2, m=0$. $$\begin{split}
&- c_0c_2 \sum_{k=0}^1 \chi_{ae_1}^k (igT^d_{e_1e_2}) \chi^{1-k}_{e_2b} \delta_{dc} - c_1c_1 \delta_{ae_1} (igT^d_{e_1e_2} )\delta_{e_2b} \chi_{dc} - (b\leftrightarrow c)\\
=&-c_0c_2[\delta_{ae_1} (igT^d_{e_1e_2}) (igT^e_{e_2b}\Phi^e) \delta_{dc} + (igT^e_{ae_1}\Phi^e) (igT^d_{e_1e_2}) \delta_{e_2b}\delta_{dc}] -c_1^2(igT^d_{ab})(igT^e_{dc}\Phi^e) - (b\leftrightarrow c)\\
=&-c_0c_2 (ig)^2 [-T^c_{ae_2} T^b_{e_2e}+T^e_{ae_1}T^c_{e_1b}]\Phi^e -c_1^2(ig)^2 T^b_{ad}T^c_{de}\Phi^e - (b\leftrightarrow c)\\
=&-c_0c_2(ig)^2[-(T^cT^b)_{ae} + T^e_{ae_1}T^{e_1}_{bc}]\Phi^e - c_1^2 (ig)^2 (T^bT^c)_{ae} \Phi^e -(b\leftrightarrow c)\\
=&-c_0c_2(ig)^2[-if^{cbd}T^d_{ae} + 2 T^e_{ae_1}T^{e_1}_{bc}]\Phi^e - c_1^2 (ig)^2 if^{bcd}T^d_{ae}\Phi^e\\
=&-g^2\left(-3c_0c_2-c_1^2\right)T^d_{bc}T^e_{ad}\Phi^e = \frac{1}{2} g^2 T^d_{bc}T^e_{ad}\Phi^e\, .\\
\end{split}$$ Therefore for $N=1$ terms eq. holds.
$N=2$ {#n2 .unnumbered}
-----
For $N=2$, the right hand side becomes $$gf^{bcd} c_2 \chi^2_{ad} = (ig)^3 c_2 T^d_{bc}(T^{e_1}T^{e_2})_{ad}\Phi^{e_1}\Phi^{e_2}\, .$$ For the left hand side $N=m+n-1=2$, there are three possibilities $n=1, m=2$; $n=2, m=1$; $n=3, m=0$. From $c_3=0$, we can only consider the first two possibilities. $$\begin{split}
&(-i)^2 c_1c_2 \sum_{k=0}^1 \chi^k_{ae_1} (igT^d_{e_1e_2})\chi^{1-k}_{e_2b} (igT^{e_3}_{dc}\Phi^{e_3}) + (-i)^2 c_2c_1 \delta_{ae_1}(igT^d_{e_1e_2})(igT^{e_3}\Phi^{e_3} igT^{e_4}\Phi^{e_4})_{dc}\delta_{e_2b}\\
=&-(ig)^3 c_1c_2 (T^d_{ae_2}T^{e_3}_{dc}T^{e_4}_{e_2b} + T^{e_4}_{ae_1} T^d_{e_1b}T^{e_3}_{dc})\Phi^{e_3}\Phi^{e_4} - (ig)^3 c_2c_1 T^d_{ab}(T^{e_3}T^{e_4})_{dc} \Phi^{e_3}\Phi^{e_4}\\
=&-(ig)^3 c_1c_2[ -(T^bT^aT^c)_{e_4e_3} + (T^cT^bT^a)_{e_3e_4} + (T^bT^{e_3}T^c)_{ae_4}] \Phi^{e_3}\Phi^{e_4}\\
=&-(ig)^3 c_1c_2[ 2 (T^cT^bT^a)_{e_3e_4} -(T^bT^{e_4}T^c)_{ae_3}+ (T^bT^{e_3}T^c)_{ae_4}] \Phi^{e_3}\Phi^{e_4}
\end{split}$$ Note that $e_3$ and $e_4$ are symmetric. The last two terms in last equality cancel. After subtracting the $(b\leftrightarrow c)$ part, one obtains $$-(ig)^3 c_1c_2 2 if^{cbe}(T^eT^a)_{e_3e_4}\Phi^{e_3}\Phi^{e_4} =- (ig)^3 2c_1c_2 T^e_{bc}(T^{e_3}T^{e_4})_{ae} \Phi^{e_3}\Phi^{e_4}$$ Since $c_1 = -1/2$, clearly for $N=2$ terms eq. holds.
$N=3$ {#n3 .unnumbered}
-----
For $N=3$ the right hand side vanishes because $c_3 =0$. We have to show that the left hand side also vanishes. First note that for $N=m+n-1=3$, there are four possibilities: $(n=1, m=3)$; $(n=2, m=2)$; $(n=3, m=1)$ ; $(n=4, m=0)$. Only the two cases $n=2,m=2$ and $n=4, m=0$ contribute. We then need to show that the following terms cancel. $$\begin{split}
&-c_2^2 \Big[(T^dT^{d_3})_{ab}(T^{d_1}T^{d_2})_{dc} + (T^{d_3}T^d)_{ab}(T^{d_1}T^{d_2})_{dc}\Big] -c_0c_4 \Big[(T^cT^{d_1}T^{d_2}T^{d_3})_{ab} \\
&+ (T^{d_1}T^cT^{d_2}T^{d_3})_{ab}+(T^{d_1}T^{d_2}T^cT^{d_3})_{ab} +(T^{d_1}T^{d_2}T^{d_3}T^c)_{ab}\Big] -(b\leftrightarrow c)\\
\end{split}$$ The guiding principles of organizing these terms are the following:
- $d_1, d_2, d_3$ are symmetric indices, we are free to interchange among them.
- We rearrange terms according to their $b, c$ indices. If $b,c$ are indices of the same matrix like $T^m_{bc}$, there is no need for further simplication. If we have $T^b, T^c$ in adjacent position like $T^bT^c$, the $b\leftrightarrow c$ subtraction gives a commutator, which results in placing $b,c$ indices on the same matrix $T$. Then no further simplification is needed.
- If $T^b, T^c$ are not directly in the adjacent position, we rearrange the $b,c$ as the indices of matrix element not the label of $T$ matrix.
- We use the cancellations between terms like $(T^aT^{d_1}T^{d_2}T^{d_3})_{bc}$ and $(T^{d_1}T^{d_2}T^{d_3}T^a)_{cb}$.
First note the last two terms in the second bracket $$(T^{d_1}T^{d_2}T^{d_3}T^c)_{ab} -(b\leftrightarrow c) = 2 (T^{d_1}T^{d_2}T^{d_3})_{ae}T^{e}_{bc}\,$$ and $$\begin{split}
&(T^{d_1}T^{d_2}T^cT^{d_3})_{ab} -(b\leftrightarrow c)\\
=&-(T^{d_1}T^{d_2}T^cT^b)_{ad_3} - (b\leftrightarrow c)\\
=&(T^{d_1}T^{d_2}T^{d_3})_{ae}T^e_{bc}\, .\\
\end{split}$$ the first two terms in the second bracket $$\begin{split}
&(T^cT^{d_1}T^{d_2}T^{d_3})_{ab} + (T^{d_1}T^cT^{d_2}T^{d_3})_{ab} - (b\leftrightarrow c)\\
=&2(T^cT^{d_1}T^{d_2}T^{d_3})_{ab} -T^e_{d_1c}(T^eT^{d_2}T^{d_3})_{ab} - (b\leftrightarrow c)\\
=&-2(T^aT^{d_1}T^{d_2}T^{d_3})_{cb} + (T^{d_1}T^aT^{d_2}T^{d_3})_{cb} +2(T^aT^{d_1}T^{d_2}T^{d_3})_{bc} - (T^{d_1}T^aT^{d_2}T^{d_3})_{bc} \\
=&-2if^{ad_1e}(T^eT^{d_2}T^{d_3})_{cb} - 2if^{ad_2e}(T^{d_1}T^eT^{d_3})_{cb} - 2if^{ad_3e}(T^{d_1}T^{d_2}T^e)_{cb} + if^{ad_2e}(T^{d_1}T^eT^{d_3})_{cb}\\
=&- 5if^{ad_2e}(T^{d_1}T^eT^{d_3})_{cb} - 2if^{ad_1e}if^{ed_2 h}(T^hT^{d_3})_{cb} -2if^{ad_3e}if^{d_2eh}(T^{d_1}T^h)_{cb}\\
=&- 5if^{ad_2e}(T^{d_1}T^eT^{d_3})_{cb} - 2if^{ad_1e}if^{ed_2h} if^{hd_3 d}T^d_{cb}\\
=&- 5if^{ad_2e}(T^{d_1}T^eT^{d_3})_{cb}+2(T^{d_1}T^{d_2}T^{d_3})_{ad}T^{d}_{bc}\\
\end{split}$$ In obtaining the third equality, we have moved the $T^a$ matrix in the first term gradually to the far right so that the resulting term will cancel the third term. We also moved the $T^a$ matrix in the second term passing $T^{d_2}$, which then will cancel the fourth term. Now consider the terms in the first bracket. $$\begin{split}
&(T^dT^{d_3})_{ab}(T^{d_1}T^{d_2})_{dc} + (T^{d_3}T^d)_{ab}(T^{d_1}T^{d_2})_{dc} - (b\leftrightarrow c)\\
=&-(T^{d_3}T^aT^{d_1}T^{d_2})_{bc} - (T^{d_3}T^bT^{d_1}T^{d_2})_{ac} - (b\leftrightarrow c)\\
=&-(T^{d_3}T^aT^{d_1}T^{d_2})_{bc} - (T^bT^{d_3}T^{d_1}T^{d_2})_{ac} +T^e_{d_3 b}(T^eT^{d_1}T^{d_2})_{ac} - (b\leftrightarrow c)\\
=&-2(T^{d_3}T^aT^{d_1}T^{d_2})_{bc} + (T^aT^{d_3}T^{d_1}T^{d_2})_{bc} +2(T^{d_3}T^aT^{d_1}T^{d_2})_{cb} - (T^aT^{d_3}T^{d_1}T^{d_2})_{cb} \\
=&-2if^{ad_1e}(T^{d_3}T^eT^{d_2})_{bc} + if^{ad_3e}(T^eT^{d_1}T^{d_2})_{bc} + if^{ad_1e}(T^{d_3}T^eT^{d_2})_{bc} + if^{ad_2e}(T^{d_3}T^{d_1}T^{e})_{bc}\\
=&if^{ad_1e}(T^{d_3}T^eT^{d_2}) _{bc}+ if^{ad_3e}if^{ed_1 h}(T^hT^{d_2})_{bc} + if^{ad_2e} if^{d_1eh}(T^{d_3}T^h)_{bc}\\
=&if^{ad_1e}(T^{d_3}T^eT^{d_2})_{bc} + if^{ad_3e}if^{ed_1h} if^{hd_2 d}T^d_{bc}\\
=&if^{ad_1e}(T^{d_3}T^eT^{d_2})_{bc} + (T^{d_3}T^{d_1}T^{d_2})_{ad}T^d_{bc}\\
\end{split}$$ Note that $c_2^2 = (1/12)^2 = \frac{5}{6!}$ and $c_0c_4 = -\frac{1}{6!}$. Therefore for $N=3$ terms eq. holds.
We have checked four leading terms in the expansion of the Jacobi identity. The same explicit procedure can be followed for higher order terms as well, but it becomes increasingly cumbersome. We therefore stop at this point.
We thank Nestor Armesto and Michael Lublinsky for interesting and useful discussion on the subject of this paper. The work is supported by the NSF Nuclear Theory grant 1913890.
[99]{}
I. Balitsky, *“Operator expansion for high-energy scattering”*, *Nucl. Phys. B* [**463**]{}, 99 (1996)
J. Jalilian-Marian, A. Kovner, A. Leonidov and H. Weigert, *“The BFKL equation from the Wilson renormalization group”*, *Nucl. Phys. B* [**504**]{}, 415 (1997)
J. Jalilian-Marian, A. Kovner, A. Leonidov and H. Weigert, *“The Wilson renormalization group for low x physics: Towards the high density regime”*, *Phys. Rev. D* [**59**]{}, 014014 (1998)
J. Jalilian-Marian, A. Kovner and H. Weigert, *“The Wilson renormalization group for low x physics: Gluon evolution at finite parton density”*, *Phys. Rev. D* [**59**]{}, 014015 (1998)
A. Kovner, J. G. Milhano and H. Weigert, *“Relating different approaches to nonlinear QCD evolution at finite gluon density”*, *Phys. Rev. D* [**62**]{}, 114005 (2000)
E. Iancu, A. Leonidov and L. D. McLerran, *“Nonlinear gluon evolution in the color glass condensate. 1”*, *Nucl. Phys. A* [**692**]{}, 583 (2001)
E. Ferreiro, E. Iancu, A. Leonidov and L. McLerran, *“Nonlinear gluon evolution in the color glass condensate. 2”*, *Nucl. Phys. A* [**703**]{}, 489 (2002)
A. H. Mueller, *“A Simple derivation of the JIMWLK equation”*, *Phys. Lett. B* [**523**]{}, 243 (2001)
N. Armesto, F. Dominguez, A. Kovner, M Lublinsky and V. Skokov, *“The Color Glass Condensate density matrix: Lindblad evolution, entanglement entropy and Wigner functional”*, *JHEP* 05 (2019) 025;
A. Kovner and M. Lublinsky, *“Dense-dilute duality at work: Dipoles of the target”*, *Phys. Rev. D* [**72**]{}, 074023 (2005)
V. Gorini, A. Kossakowski and E. C. G. Sudarshan, *“Completely Positive Dynamical Semigroups of N Level Systems”*, *J. Math. Phys.* [**17**]{}, 821 (1976).
G. Lindblad, *“On the Generators of Quantum Dynamical Semigroups”*, *Commun. Math. Phys. * [**48**]{}, 119 (1976).
T. Altinoluk, A. Kovner, M. Lublinsky and J. Peressutti, *“QCD Reggeon Field Theory for every day: Pomeron loops included”*, *JHEP* [**0903**]{}, 109 (2009)
M. Hillery, R. F. O’Connell, M. O. Scully and E. P. Wigner, *“Distribution functions in physics: Fundamentals”*, *Phys. Rept.* [**106**]{}, 121 (1984).
J. Preskill, *“Lecture notes for physics 219: Quantum computation”*, available at $http://www.theory.caltech.edu/~preskill/ph219/ph219\_2018-19$, 2019,
H. J. Carmichael, *“An Open Systems Approach to Quantum Optics ”*, Springer (1993), Berlin.
H. P. Breuer, *“The Theory of Open Quantum Systems ”* , Oxford University Press, USA(March 29, 2007).
Y. V. Kovchegov and E. Levin, *“Quantum chromodynamics at high energy”*, Camb. Monogr. Part. Phys. Nucl. Phys. Cosmol. [**33**]{}, 1 (2012).
L. D. McLerran and R. Venugopalan, *“Computing quark and gluon distribution functions for very large nuclei”*, *Phys. Rev. D* [**49**]{}, 2233 (1994);
L. D. McLerran and R. Venugopalan, *“Gluon distribution functions for very large nuclei at small transverse momentum”*, *Phys. Rev. D* [**49**]{}, 3352 (1994)
A. Kovner, *“High energy evolution: The Wave function point of view”*, *Acta Phys. Polon. B* [**36**]{}, 3551 (2005)
A. Kovner, M. Lublinsky and U. Wiedemann, *“From bubbles to foam: Dilute to dense evolution of hadronic wave function at high energy”*, *JHEP* [**0706**]{}, 075 (2007)
A. Kovner and M. Lublinsky, *“In pursuit of Pomeron loops: The JIMWLK equation and the Wess-Zumino term”* *Phys. Rev. D* [**71**]{}, 085004 (2005)
K. E. Cahill and R. J. Glauber, *“Density operators and quasi probability distributions”*, *Phys. Rev. * [**177**]{}, 1882 (1969).
G. S. Agarwal and E. Wolf, *“Calculus for functions of noncommuting operators and general phase-space methods in quantum mechanics. i. mapping theorems and ordering of functions of noncommuting operators”*, *Phys. Rev. D* [**2**]{}, 2161 (1970).
G. S. Agarwal and E. Wolf, *“Calculus for functions of noncommuting operators and general phase-space methods in quantum mechanics. ii. quantum mechanics in phase space”*, *Phys. Rev. D* [**2**]{}, 2187 (1970).
R. L. Stratonovich, *Zh. Eksp. Teor, Fiz*. [**31**]{}, 1012 (1956), \[*Sov. Phys. JETP* [**4**]{}, 891 (1957)\].
J. C. Varilly and J. M. Gracia-Bondia, *“The Moyal representation for spin”*, *Annals Phys. * [**190**]{}, 107 (1989).
C. Brif and A. Mann, *“A general theory of phase space quasi probability distributions”*, *J. Phys. A* [**31**]{}, L9 (1998)
C. Brif and A. Mann, *“Phase space formulation of quantum mechanics and quantum state reconstruction for physical systems with Lie group symmetries”*, *Phys. Rev. A* [**59**]{}, 971 (1999)
A. Barut and R. Raczka, *“Theory of Group Representations and Applications”*, World Scientific (1987).
T. Altinoluk, C. Contreras, A. Kovner, E. Levin, M. Lublinsky and A. Shulkin, *“QCD Reggeon Calculus From KLWMIJ/JIMWLK Evolution: Vertices, Reggeization and All”*, *JHEP* [**1309**]{}, 115 (2013)
A. Kovner and M. Lublinsky, *“From target to projectile and back again: Selfduality of high energy evolution”*, *Phys. Rev. Lett. * [**94**]{}, 181603 (2005)
H. Weigert, *“Unitarity at small Bjorken x”*, *Nucl. Phys. A* [**703**]{}, 823 (2002)
J. P. Blaizot, E. Iancu and H. Weigert, *“Nonlinear gluon evolution in path integral form”*, *Nucl. Phys. A* [**713**]{}, 441 (2003)
I. Balitsky and G. A. Chirilli, *“Next-to-leading order evolution of color dipoles”*, *Phys. Rev. D* [**77**]{}, 014019 (2008)
A. Kovner, M. Lublinsky and Y. Mulian, *“Jalilian-Marian, Iancu, McLerran, Weigert, Leonidov, Kovner evolution at next to leading order”*, *Phys. Rev. D* [**89**]{}, no. 6, 061704 (2014)
M. Lublinsky and Y. Mulian, *“High Energy QCD at NLO: from light-cone wave function to JIMWLK evolution”*, *JHEP* [**1705**]{}, 097 (2017)
S. Caron-Huot, *“Resummation of non-global logarithms and the BFKL equation”* *JHEP* [**1803**]{}, 036 (2018)
[^1]: The most common form of the eikonal scatttering amplitude one finds in the literature is a lightlike Wilson line. This is the right definition in a gauge which has a nonvanishing light cone component of the vector potential $A^-$. Our discussion here is set in the lightcone gauge in which $A^-=0$. In this gauge the scattering amplitude is given by the transverse Wilson line at $x^+\rightarrow\infty$, which is defined in Eq..
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The (Berry-Aharonov-Anandan) geometric phase acquired during a cyclic quantum evolution of finite-dimensional quantum systems is studied. It is shown that a pure quantum state in a $(2J+1)$-dimensional Hilbert space (or, equivalently, of a spin-$J$ system) can be mapped onto the partition function of a gas of independent Dirac strings moving on a sphere and subject to the Coulomb repulsion of $2J$ fixed test charges (the Majorana stars) characterizing the quantum state. The geometric phase may be viewed as the Aharonov-Bohm phase acquired by the Majorana stars as they move through the gas of Dirac strings. Expressions for the geometric connection and curvature, for the metric tensor, as well as for the multipole moments (dipole, quadrupole, etc.), are given in terms of the Majorana stars. Finally, the geometric formulation of the quantum dynamics is presented and its application to systems with exotic ordering such as spin nematics is outlined.\
\
Phys. Rev. Lett. **108**, 240402 (2012) \[http://link.aps.org/doi/10.1103/PhysRevLett.108.240402\]\
selected for a Viewpoint in *Physics* \[http://link.aps.org/doi/10.1103/Physics.5.65\]
author:
- Patrick Bruno
date: 'received 10 April 2012; accepted 26 April, 2012'
title: |
Quantum geometric phase in Majorana’s stellar representation:\
Mapping onto a many-body Aharonov-Bohm phase
---
The concept of geometric phase associated with a cyclic quantum evolution [@berry1984; @aharonov_anandan1987] has by now become a central unifying concept of quantum mechanics [@shapere_wilczek1989; @bohm2003; @chrucinski2004]. Its importance stems from the fact that its local expression, the geometric curvature, controls the quantum dynamics.
The prototype of geometric phase is that of a spin $1/2$: in this case, it can be thought of as the Aharonov-Bohm (AB) phase of a unit electrical charge moving on the Bloch sphere in the field of a Dirac monopole of unit magnetic charge located at the center of the Bloch sphere [@berry1984]. For a spin-$J$ system evolving within the manifold of coherent states (CSs) (i.e., quasi-classical states), this interpretation goes over straightforwardly, the electric charge being now of magnitude $2J$ [@berry1984]. For a spin-$J$ system in an arbitrary quantum state, except for for the specific case of a quantum evolution consisting only of global rotations (i.e., due to a, possibly time-dependent, magnetic field) [@bruno2004], the geometric phase associated with an arbitrary cyclic evolution of a spin $J$ system [@hannay1998] does not seem to be amenable to any physically appealing AB-like interpretation, which is somehow unsatisfactory.
In the present Letter, I propose a novel theory of the geometric phase of spin systems, based upon a mapping onto a (fictitious) many-body system. From the latter emerges quite naturally a novel AB-like, physically transparent, understanding of the geometric phase of spin systems. Since any system with a finite dimensional Hilbert space of dimension $2J+1$ can be thought of as a spin-$J$ system, the present study actually holds, at least formally, for any finite quantum system.
Since any two state vectors $|\Psi_1^{(\! J)}\rangle$ and $|\Psi_2^{(\! J)}\rangle$ of the spin-$J$ Hilbert space $\mathcal{H}^{(\! J)}=\mathbb{C}^{2J+1}\! -\!\{0\}$ satisfying $|\Psi_2^{(\! J)}\rangle = c |\Psi_1^{(\! J)}\rangle$ (with $c\in \mathbb{C}, c\neq0$) yield the same expectation value for any observable, they represent the same physical state and belong to the same equivalence class ($|\Psi_2^{(\! J)}\rangle \!\! \sim \!\! |\Psi_1^{(\! J)}\rangle$); thus the manifold of physical states (projective Hilbert space) is the quotient space of equivalence classes $\mathcal{P}^{(\! J)}\equiv \mathcal{H}^{(\! J)}/\!\! \sim \ = \mathbb{C}P^{2J}$ [@aharonov_anandan1987]. As in Ref. , the approach to be used here is purely geometric and relies upon Majorana’s stellar representation [@majorana1932] for $\mathcal{P}^{(\! J)}$. Majorana’s representation is most easily understood by noticing that spin-$J$ states can be obtained as fully symmetrized states of system of $2J$ spins $1/2$ [@bloch_rabi1945]. This idea is at the heart of the Schwinger boson (SB) representation [@schwinger1952], in which the spin-$1/2$ CS pointing along the direction $\mathbf{\hat{n}}$ of spherical angles $\theta$ and $\varphi$ is $|\mathbf{\hat{n}}^{(1/2)}\rangle \equiv \hat{a}^\dag_\mathbf{\hat{n}} |\emptyset\rangle$, with $\hat{a}^\dag_\mathbf{\hat{n}} \equiv \cos\left(\frac{\theta}{2}\right) \hat{a}^\dag_\uparrow + \sin\left(\frac{\theta}{2}\right) \mathrm{e}^{\mathrm{i}\varphi}\hat{a}^\dag_\downarrow$. Let us pick $2J$ (non necessarily distinct) unit vectors $\{\mathbf{\hat{u}}_1 ,\ldots ,\mathbf{\hat{u}}_{2J} \}\equiv \mathbf{U}$, and form the state $| \Psi_{\mathbf{U}}^{(\! J)} \rangle \equiv \frac{1}{\sqrt{(2J)!}} \left( \prod_{i=1}^{2J} \hat{a}^\dag_{-\mathbf{\hat{u}}_i} \right) |\emptyset\rangle$. Obviously, being a superposition of states with $2J$ SBs, such a state is a spin-$J$ state. In particular, the states $|\mathbf{\hat{n}}^{(\! J)}\rangle \equiv \frac{1}{\sqrt{(2J)!}} (\hat{a}^\dag_\mathbf{\hat{n}})^{2J} |\emptyset\rangle$ are the spin-$J$ CSs [@perelomov1986]; their scalar product is given by $
\langle \mathbf{\hat{n}}_1^{(\! J)} | \mathbf{\hat{n}}_2^{(\! J)} \rangle = \left( \frac{1+ \mathbf{\hat{n}}_1 \cdot \mathbf{\hat{n}}_2}{2} \right)^{J} \mathrm{e}^{\mathrm{i} J \Sigma \left(\mathbf{\hat{z}}, \mathbf{\hat{n}}_1, \mathbf{\hat{n}}_2\right) } ,
$ where $\Sigma \left(\mathbf{\hat{z}}, \mathbf{\hat{n}}_1, \mathbf{\hat{n}}_2\right)$ is the oriented area of the spherical triangle $\left(\mathbf{\hat{z}}, \mathbf{\hat{n}}_1, \mathbf{\hat{n}}_2\right)$, and they satisfy the following resolution of unity: $
\mathbf{1}_{J} \equiv \frac{2J+1}{4\pi}\int_{S^2} \mathrm{d}^{2}\mathbf{\hat{n}} \ |\mathbf{\hat{n}}^{(\! J)}\rangle \langle \mathbf{\hat{n}}^{(\! J)}| .
$ The rotated SB creation and annihilation operators satisfy the commutation relations: $[\hat{a}_\mathbf{\hat{n}},\hat{a}_\mathbf{\hat{n}^\prime}] = [\hat{a}^\dag_\mathbf{\hat{n}}, \hat{a}^\dag_\mathbf{\hat{n}^\prime}]=0$ and $[\hat{a}_\mathbf{\hat{n}}, \hat{a}^\dag_\mathbf{\hat{n}^\prime}] = \langle \mathbf{\hat{n}}^{(\! 1/2)} | \mathbf{\hat{n}}^{\prime (\! 1/2)} \rangle$. Let us introduce the CS representation $\Psi_\mathbf{U}^{(\! J)}(\mathbf{\hat{n}})\equiv \langle \mathbf{\hat{n}}^{(\! J)}| \Psi_\mathbf{U}^{(\! J)}\rangle$, which is a wavefunction over the sphere $S^2$, with probability distribution $
Q^{(\! J)}_\mathbf{U} (\mathbf{\hat{n}}) \equiv |\Psi_\mathbf{U}^{(\! J)}(\mathbf{\hat{n}})|^{2}$ (Husimi function). Simple algebraic manipulations yield $\Psi_\mathbf{U}^{(\! J)}(\mathbf{\hat{n}}) = \prod_{i=1}^{2J} \left( \frac{1-\mathbf{\hat{n}}\cdot\mathbf{\hat{u}}_i}{2} \ \mathrm{e}^{\mathrm{i}\Sigma (\mathbf{\hat{z}},\mathbf{\hat{n}},-\mathbf{\hat{u}}_i)} \right)^{1/2}$. Conversely, using the decomposition in the familiar $|JM\rangle$ basis, one sees that a generic spin-$J$ state vector $|\Psi^{(\! J)}\rangle$ can be expressed as $|\Psi^{(\! J)}\rangle = P_\Psi (\hat{a}^\dag_\uparrow, \hat{a}^\dag_\downarrow) |\emptyset\rangle$, where $P_\Psi (\hat{a}^\dag_\uparrow, \hat{a}^\dag_\downarrow)$ is a homogenous polynomial of degree $2J$ of $\hat{a}^\dag_\uparrow$ and $\hat{a}^\dag_\downarrow$, which can be factorized (up to an unimportant prefactor) in the above form $|\Psi_\mathbf{U}^{(\! J)}\rangle$ with a unique multiset $\mathbf{U}$ (Fundamental Theorem of Algebra). Thus, $\mathcal{P}^{(\! J)}$ can be univocally parameterized by the constellation $\mathbf{U}$ of $2J$ Majorana stars (MSs) (zeros $\mathbf{\hat{u}}_i$ of the Husimi function), and $|\Psi_\mathbf{U}^{(\! J)}\rangle$ can be taken as a fiducial state in $\mathcal{H}^{(\! J)}$ to describe $\mathcal{P}^{(\! J)}$. Addition or removal of $n$ stars to/from a given spin-$J$ constellation generates a state of spin $(J+n/2)$ or $(J-n/2)$, respectively. While the SB formalism was inspired by Majorana’s representation [@schwinger2000], the underlying geometric aspects have not been fully explored so far; to carry out this task is one of the aims of the present paper.
![Typical diagrams used to compute $D_\mathbf{U}^{(\! J,n)}$ (a), $D^{(\! J,n)}_{\mathbf{U}\mu}$ (b), and $D^{(\! J,n)}_{\mathbf{U}\mu\nu}$ (c). The solid dots represent the MSs (labeled from 1 to $2J$), the open dots (labeled $\mu$, $\nu$, ...) represent the auxiliary stars used to compute the multipole moments. The rules are: (i) draw all possible distinct diagrams with $n$ pairing links (each dot, solid or open, may be linked only once in a given diagram); (ii) calculate the contribution of each diagram as indicated below, and then sum over all diagrams; (iii) an unlinked solid dot yields a factor 1; (iv) an unlinked open dot yields a factor 0; (v) a link between 2 solid dots $i$ and $j$ yields a factor $d_{ij}$; (vi) a link between a solid dot $i$ and an open dot $\mu$ yields a factor $\hat{u}_{i\mu}$; (vii) a link between 2 open dots $\mu$ and $\nu$ yields a factor $-2\delta_{\mu\nu}$. The diagrams (a), (b), (c) shown here yield the contributions $d_{12}d_{45}$, $\hat{u}_{5\mu} d_{23} d_{46}$, and $\hat{u}_{3\mu} \hat{u}_{6\nu}$ to $D_\mathbf{U}^{(3,2)}$, $D^{(3,3)}_{\mathbf{U}\mu}$ and $D^{(3,2)}_{\mathbf{U}\mu\nu}$, respectively. \[fig\_diagrams\]](Fig1.eps){width="1.0\columnwidth"}
Noticing that the two-dimensional (2D) Coulomb interaction on the sphere is $\tilde{V}(\mathbf{\hat{u}}_1,\mathbf{\hat{u}}_2)= -\ln(d_{12})$, where $2 d_{12} \equiv 2 \sin^{2}(\theta_{12}/2) = 1-\mathbf{\hat{u}}_1 \cdot\mathbf{\hat{u}}_2$ is the chordal distance between $\mathbf{\hat{u}}_1$ and $\mathbf{\hat{u}}_2$ [@caillol1981], and writing $
Q^{(\! J)}_\mathbf{U} (\mathbf{\hat{n}}) = \exp (-\tilde{\beta}\tilde{U}_\mathbf{U} (\mathbf{\hat{n}}) ) ,
$ with $\tilde{\beta}\equiv1$ and $
\tilde{U}_\mathbf{U} (\mathbf{\hat{n}}) \equiv \sum_{i=1}^{2J} \tilde{V}(\mathbf{\hat{n}},\mathbf{\hat{u}}_i) ,
$ we can interpret the (rescaled) norm of $|\Psi_\mathbf{U}^{(\! J)}\rangle$, $$\begin{aligned}
\label{}
\tilde{Z}(\mathbf{U}) &\equiv & \frac{\langle\Psi_\mathbf{U}^{(\! J)} |\Psi_\mathbf{U}^{(\! J)}\rangle}{2J+1} = \frac{1}{4\pi}\int_{S^2}\!\! \mathrm{d}^{2} \! \mathbf{\hat{n}} \ Q^{(\! J)}_\mathbf{U} (\mathbf{\hat{n}}) \nonumber \\
&=& \frac{1}{4\pi}\int_{S^2}\!\! \mathrm{d}^{2} \! \mathbf{\hat{n}} \ \ \exp \left(-\tilde{\beta}\tilde{U}_\mathbf{U} (\mathbf{\hat{n}}) \right) ,\end{aligned}$$ as the (fictitious) partition function, at inverse temperature $\tilde{\beta}\equiv1$, of a classical gas of independent particles (of density $Q_\Psi^{(\! J)} (\mathbf{\hat{n}})$) living on the sphere and interacting via the 2D spherical Coulomb repulsion with $2J$ fixed test charges located at the $\mathbf{\hat{u}}_i$’s. The corresponding fictitious free energy, $\tilde{F}(\mathbf{U}) \equiv - \tilde{\beta}^{-1} \ln \tilde{Z}(\mathbf{U})$, expresses a fictitious indirect interaction among the MSs, mediated by the gas particles in thermal equilibrium at temperature $\tilde{\beta}^{-1}$; we thus obtain a mapping of a spin-$J$ quantum state onto a $2J$-body classical system [@coulomb]. The partition function can be expressed in terms of the pairwise chordal distances between the MSs, $2 d_{ij}\equiv 1-\mathbf{\hat{u}}_i \cdot \mathbf{\hat{u}}_j$, as [@lee1988] $$\label{}
\tilde{Z}(\mathbf{U} ) = \frac{1}{2J+1}\sum_{n=0}^{[J]} \left[(-1)^{n}\frac{(2J-n)!}{(2J)!}\ D_\mathbf{U}^{(\! J,n)}\right] ,$$ with $[J]\equiv J$ (resp. $[J]\equiv J-\frac{1}{2}$) for $2J$ even (resp. $2J$ odd), and where the expression of $D_\mathbf{U}^{(\! J,n)}$ in terms of diagrams is explained in Fig. \[fig\_diagrams\].
Let us now show how the expectation value of the various multipole moments can be expressed in terms of the MS. They are obtained from the expectation values of the irreducible spherical tensor operators $\mathcal{\hat{Y}}_l^m (\mathbf{J})$, which in turn have the following P-representation [@gilmore1976] $$\label{}
\mathcal{\hat{Y}}_l^m (\mathbf{J})\! = \! \frac{(2J+1+l)!}{(2J+1)! 2^l} \frac{2J+1}{4\pi}\! \int_{S^2}\!\!\! \mathrm{d}^{2}\mathbf{\hat{n}} |\mathbf{\hat{n}}^{(\! J)}\rangle Y_l^m(\mathbf{\hat{n}}) \langle \mathbf{\hat{n}}^{(\! J)}| .$$ For the expectation values of the dipole moment $\hat{J}_\mu$ and the quadrupole moment $\hat{Q}_{\mu\nu} \equiv \frac{\hat{J}_\mu \hat{J}_\nu +\hat{J}_\nu \hat{J}_\mu}{2} - \frac{J(J+1)}{3} \delta_{\mu\nu}$ (where $\mu,\nu,\ldots$ label the cartesian axes), this yields
$$\begin{aligned}
\langle \hat{J}_\mu \rangle &=& (J+1) \langle \hat{n}_\mu \rangle , \\
\left\langle \hat{Q}_{\mu\nu} \right\rangle &=& (J+1)\left( J+\frac{3}{2} \right) \left( \left\langle \hat{n}_\mu \hat{n}_\nu \right\rangle -\frac{\delta_{\mu\nu}}{3} \right) ,\end{aligned}$$
where $$\langle f(\mathbf{\hat{n}}) \rangle \equiv \frac{\int_{S^2}\! \mathrm{d}^{2}\mathbf{\hat{n}}\ f(\mathbf{\hat{n}})\ Q^{(\! J)}_\mathbf{U} (\mathbf{\hat{n}}) }{\int_{S^2}\! \mathrm{d}^{2}\mathbf{\hat{n}}\ Q^{(\! J)}_\mathbf{U} (\mathbf{\hat{n}}) } .$$ To calculate the averages $\langle \hat{n}_\mu \rangle$ and $\langle \hat{n}_\mu \hat{n}_\nu \rangle$, I remark that if we form the spin-$(J+1/2)$ state $\mathbf{U}^{\prime}$ obtained from the spin-$J$ state $\mathbf{U}$ by adding" the star $\mathbf{\hat{u}}^\prime$, and the spin-$(J+1)$ state $\mathbf{U}^{\prime\prime}$ obtained by adding one further star $\mathbf{\hat{u}}^{\prime\prime}$, the corresponding fictitious free energies are given by
$$\begin{aligned}
\tilde{F}(\mathbf{U}^{\prime}) &=& \tilde{F}(\mathbf{U}) + \tilde{F}(1/2) - \ln \left[ 1 - \hat{u}^\prime_\mu \langle \hat{n}_\mu \rangle \right] \\
\tilde{F}(\mathbf{U}^{\prime\prime}) &=& \tilde{F}(\mathbf{U}) + 2\tilde{F}(1/2) \nonumber \\
&-& \ln \left[ 1 - (\hat{u}^\prime_\mu + \hat{u}^{\prime\prime}_\mu )\langle \hat{n}_\mu \rangle
+ \hat{u}^\prime_\mu \hat{u}^{\prime\prime}_\nu \langle \hat{n}_\mu \hat{n}_\nu \rangle \right] ,\end{aligned}$$
where $\tilde{F}(1/2) \equiv \ln 2$ (here and further below, Einstein’s convention of summation over repeated is used, unless explicitly specified). Thus we see that we can obtain the dipole and quadrupole moments by adding 1 or 2 auxiliary stars, respectively, from the variation of the free energy as these auxiliary stars are moved around the sphere. A careful but straightforward calculation yields:
$$\begin{aligned}
\langle \hat{n}_\mu \rangle \! &=& \! \frac{1}{2(J \! + \! 1)} \frac{\sum_{n=1}^{[J+1 \! / 2]} (-1)^n (2J\! + \! 1 \! - \! n)! D_{\mathbf{U}\mu}^{(\! J,n)} }{\sum_{n=0}^{[J]} (-1)^n (2J-n)! D_\mathbf{U}^{(\! J,n)}} , \\
\langle \hat{n}_\mu \hat{n}_\nu \rangle &=& \! \frac{1}{2(J \!\! + \! 1) (2J \! \!+ \! 3)} \frac{\sum_{n=1}^{[J+1]} (-1)^n (2J\! \! + \! 2 \! - \! n)! D_{\mathbf{U}\mu\nu}^{(\! J,n)} }{\sum_{n=0}^{[J]} (-1)^n (2J-n)! D_\mathbf{U}^{(\! J,n)}} , \nonumber \\
&&\end{aligned}$$
where the expressions of $D_{\mathbf{U}\mu}^{(\! J,n)}$ and $D_{\mathbf{U}\mu\nu}^{(\! J,n)}$ in terms of diagrams are given in Fig. \[fig\_diagrams\]. For example, the spin-1 dipole and quadrupole moments are, respectively,
$$\begin{aligned}
\langle \hat{J}_\mu \rangle &=& -\frac{\hat{u}_{1\mu}+\hat{u}_{2\mu}}{2-d_{12}} , \\
\!\!\! \langle \hat{Q}_{\!\mu\nu}\! \rangle\! &=& \! \frac{1}{2\! -\! d_{12}} \! \left( \!\!\frac{\hat{u}_{1\mu}\hat{u}_{2\nu}\! +\! \hat{u}_{2\mu}\hat{u}_{1\nu}}{2} - \mathbf{\hat{u}}_1 \!\! \cdot \! \mathbf{\hat{u}}_2 \frac{\delta_{\mu\nu}}{3} \!\! \right) .\end{aligned}$$
The extension of this procedure to higher-order multipole moments is straightforward. In turn, the method presented here allows to express, in terms of the MSs, the expectation value $H(\mathbf{U})$ of the Hamiltonian and its derivatives $\frac{\partial H(\mathbf{U})}{\partial{\mathbf{\hat{u}}_i}}$, which will be used further below to describe the quantum dynamics.
Let us now come to the geometric phase and the quantum metric. The geometric phase acquired as the systems is parallel-transported along a closed circuit $\mathcal{C}$ in $\mathcal{P}^{(\! J)}$ is given by [@berry1984; @aharonov_anandan1987]
$$\begin{aligned}
\phi_{B} &=& \oint_{\mathcal{C}} \mathbf{A} \cdot \mathrm{d}\mathbf{U} \equiv \sum_{i=1}^{2J} \oint_{\mathcal{C}} \mathbf{a}_i \cdot \mathrm{d}\mathbf{\hat{u}}_i \label{eq_Berry_a}\\
&=& \frac{1}{2}\sum_{i,j=1}^{2J} \sum_{\alpha ,\beta =1}^2 \int_{\mathcal{S}\ (\partial\mathcal{S}=\mathcal{C})} f_{ij}^{\alpha\beta} \ \mathrm{d}\hat{u}_i^\alpha \wedge \mathrm{d}\hat{u}_j^\beta , \label{eq_Berry_b}\end{aligned}$$
with
$$\begin{aligned}
a_i^\alpha &\equiv& \frac{\mathrm{i}}{2} \left( \overrightarrow{\partial}_{\!\!\hat{u}_i^\alpha} - \overleftarrow{\partial}_{\!\!\hat{u}_i^\alpha} \right) \mathrm{ln} \langle \Psi_\mathbf{U}^{(\! J)} | \Psi_\mathbf{U}^{(\! J)} \rangle
, \\
f_{ij}^{\alpha\beta} &\equiv & \partial_{\hat{u}_i^\alpha} a_j^\beta - \partial_{\hat{u}_j^\beta} a_i^\alpha = -2\ \mathrm{Im}( h_{ij}^{\alpha\beta} ) , \\
h_{ij}^{\alpha\beta} & \equiv & \overleftarrow{\partial}_{\!\!\hat{u}_i^\alpha} \overrightarrow{\partial}_{\!\!\hat{u}_j^\beta} \mathrm{ln} \langle \Psi_\mathbf{U}^{(\! J)} | \Psi_\mathbf{U}^{(\! J)} \rangle ,\end{aligned}$$
where $\overleftarrow{\partial}$ (resp. $\overrightarrow{\partial}$) indicates derivative of the bra $\langle \Psi_\mathbf{U}^{(\! J)} |$ (resp. ket $| \Psi_\mathbf{U}^{(\! J)} \rangle$) only. In the above equations, $\alpha,\beta =1,2$ label some spherical coordinates for the MS, with the tangent unit vectors $\mathbf{\hat{e}}_i^1$ and $\mathbf{\hat{e}}_i^2= \mathbf{\hat{u}}_i \times \mathbf{\hat{e}}_i^1$. In going from Eq. (\[eq\_Berry\_a\]) to Eq. (\[eq\_Berry\_b\]), Stokes’ theorem has been used, and $\mathcal{S}$ is an oriented surface bounded by the oriented path $\mathcal{C}$. Here, $a_i^\alpha$ and $f_{ij}^{\alpha\beta}$ are, respectively, the (gauge-dependent) Berry connection and the (gauge-independent) Berry curvature tensor; they have the physical meaning of a vector potential” and of a flux density”, respectively, in $\mathcal{P}^{(\! J)}$. The other important geometric structure is the quantum metric (Fubini-Study metric), corresponding to a distance between $|\Psi \rangle$ and $| \Phi\rangle$ defined as $D_{\mathrm{FS}}(\Psi,\Phi )\equiv 2 \arccos \left(\frac{|\langle \Psi | \Phi\rangle|}{\langle \Psi | \Psi \rangle^{1/2} \langle\Phi | \Phi\rangle^{1/2}} \right)$, whose infinitesimal expression is $\mathrm{d}s^2= g_{ij}^{\alpha\beta} \mathrm{d}\hat{u}_i^\alpha \mathrm{d}\hat{u}_j^\beta$, with metric tensor $g_{ij}^{\alpha\beta}= 4\ \mathrm{Re}(h_{ij}^{\alpha\beta})$ [@anandan_aharonov1990].
The direct calculation of the geometric phase [@hannay1998] is complicated because of the need of taking care of the commutation relations among SB operators, yielding physically obscure results. This difficulty can be overcome by inserting the CS resolution of unity between bra and kets, and, after some algebraic manipulations, one obtains $\mathbf{a}_i = \left\langle \mathbf{\tilde{a}}_i (\mathbf{\hat{n}}) \right\rangle$, where $\mathbf{\tilde{a}}_i (\mathbf{\hat{n}}) \equiv \frac{-1}{2} \left( \frac{\mathbf{\hat{z}}\times \mathbf{\hat{u}}_i }{1- \mathbf{\hat{z}}\cdot \mathbf{\hat{u}}_i} - \frac{\mathbf{\hat{n}}\times \mathbf{\hat{u}}_i }{1- \mathbf{\hat{n}}\cdot \mathbf{\hat{u}}_i}\right)$ is readily seen to be the vector potential, at $\mathbf{\hat{u}}_i$, due to a (unit flux) Dirac string entering the sphere along the $z$-axis and exiting at $\mathbf{\hat{n}}$. This means that the particles of our fictitious classical gas actually carry a Dirac string; thus the MSs are surrounded by a flux density (of total flux equal to that of a Dirac monopole of unit magnetic charge) proportional to the gas density $Q_\mathbf{U}^{(\! J)}(\mathbf{\hat{n}})$. The geometrical phase is then naturally interpreted as the AB phase acquired by the MSs as they perform a cyclic motion on the sphere. The most salient feature of this novel interpretation is the fluid” character of the flux density, which results from the Coulomb repulsion between the flux carrying gas particles and the MSs. For a CS circuit, Berry’s result [@berry1984] is recovered, albeit with a different AB-like interpretation. Skipping technical algebraic details, the final expression for the Berry connection is (no Einstein convention here)
$$\begin{aligned}
\label{}
\mathbf{a}_i &=& \frac{-1}{2} \left( \frac{\mathbf{\hat{z}}\times \mathbf{\hat{u}}_i }{1- \mathbf{\hat{z}}\cdot \mathbf{\hat{u}}_i} - \left\langle \frac{\mathbf{\hat{n}} }{1- \mathbf{\hat{n}} \cdot \mathbf{\hat{u}}_i} \right\rangle \times \mathbf{\hat{u}}_i \right) \\
&=& \frac{-1}{2} \left( \frac{\mathbf{\hat{z}}\times \mathbf{\hat{u}}_i }{1- \mathbf{\hat{z}}\cdot \mathbf{\hat{u}}_i} - \frac{ \langle \mathbf{\hat{n}}\rangle^\prime_{i} \times \mathbf{\hat{u}}_i }{1- \langle\mathbf{\hat{n}}\rangle^\prime_{i} \cdot \mathbf{\hat{u}}_i } \right) \label{berry_connection_b} \\
&=& \frac{-1}{2} \left( \frac{\mathbf{\hat{z}}\times \mathbf{\hat{u}}_i }{1- \mathbf{\hat{z}}\cdot \mathbf{\hat{u}}_i} - \partial_{\mathbf{\hat{u}}_i}\tilde{F} \times \mathbf{\hat{u}}_i \right) .\end{aligned}$$
In Eq. (\[berry\_connection\_b\]), the notation $\langle f(\mathbf{\hat{n}}) \rangle^\prime_i$ indicates that the average is taken for the spin-$(J \! - \!1/2)$ state obtained by removing the star $\mathbf{\hat{u}}_i$ from the Majorana constellation of the spin-$J$ state $\mathbf{U}$; similarly $\langle f(\mathbf{\hat{n}}) \rangle^\prime_{ij}$, to be used further below, indicates the average taken over the spin-$(J\! -\! 1)$ obtained by removing the two stars $\mathbf{\hat{u}}_i$ and $\mathbf{\hat{u}}_j$. The first term in the above equations corresponds to the uniform flux density of a Dirac monopole, while the second one corresponds to the non-uniform part (with zero average) of the flux density.
The metric and Berry curvature and tensors are obtained in a similar manner. For the former, one gets (no Einstein convention in Eqs. (\[eq\_gij\]–\[eq\_gii\]))
$$\begin{aligned}
\label{eq_gij}
g_{ij}^{\alpha\beta} &=& \delta_{ij}\delta_{\alpha\beta} + \mathbf{\hat{e}}_i^\alpha \cdot \frac{\partial^2 \tilde{F}}{\partial \mathbf{\hat{u}}_i \otimes \partial \mathbf{\hat{u}}_j} \cdot \mathbf{\hat{e}}_j^\beta \nonumber \\
&+& \mathcal{J}\mathbf{\hat{e}}_i^\alpha \cdot \frac{\partial^2 \tilde{F}}{\partial \mathbf{\hat{u}}_i \otimes \partial \mathbf{\hat{u}}_j} \cdot \mathcal{J}\mathbf{\hat{e}}_j^\beta , \\
&=& \mathbf{\hat{e}}_i^\alpha \cdot \overline{\overline{\mathbf{d}}}_{ij} \cdot \mathbf{\hat{e}}_j^\beta + \mathcal{J}\mathbf{\hat{e}}_i^\alpha \cdot \overline{\overline{\mathbf{d}}}_{ij} \cdot \mathcal{J}\mathbf{\hat{e}}_j^\beta ,\end{aligned}$$
with $\mathcal{J}\mathbf{\hat{e}}_i^1 \equiv \mathbf{\hat{e}}_i^2$, $\mathcal{J}\mathbf{\hat{e}}_i^2 \equiv - \mathbf{\hat{e}}_i^1$, and $$\begin{aligned}
\overline{\overline{\mathbf{d}}}_{ij} &\equiv& \left\langle \frac{\mathbf{\hat{n}}}{1- \mathbf{\hat{u}}_i \cdot \mathbf{\hat{n}}} \otimes \frac{\mathbf{\hat{n}}}{1- \mathbf{\hat{u}}_j \cdot \mathbf{\hat{n}}} \right\rangle \nonumber \\
&-& \left\langle \frac{\mathbf{\hat{n}}}{1- \mathbf{\hat{u}}_i \cdot \mathbf{\hat{n}}} \right\rangle \otimes \left\langle \frac{\mathbf{\hat{n}}}{1- \mathbf{\hat{u}}_j \cdot \mathbf{\hat{n}}} \right\rangle .\end{aligned}$$
For $i\neq j$, this yields $$\begin{aligned}
\label{}
\overline{\overline{\mathbf{d}}}_{ij} &=& \frac{\langle \mathbf{\hat{n}}\otimes \mathbf{\hat{n}} \rangle^\prime_{ij} }{1- ( \mathbf{\hat{u}}_i + \mathbf{\hat{u}}_j ) \cdot \langle \mathbf{\hat{n}}\rangle^\prime_{ij} + \mathbf{\hat{u}}_i \cdot \langle \mathbf{\hat{n}}\otimes \mathbf{\hat{n}} \rangle^\prime_{ij} \cdot \mathbf{\hat{u}}_j } \nonumber \\
&-& \frac{\langle \mathbf{\hat{n}} \rangle^\prime_{i} }{1 - \langle \mathbf{\hat{n}} \rangle^\prime_{i} \cdot \mathbf{\hat{u}}_i } \otimes \frac{\langle \mathbf{\hat{n}} \rangle^\prime_{j} }{1 - \langle \mathbf{\hat{n}} \rangle^\prime_{j} \cdot \mathbf{\hat{u}}_j } ,\end{aligned}$$ whereas for $i=j$, one gets $$\label{eq_gii}
g_{ii}^{\alpha\beta} = \delta_{\alpha\beta} \frac{ 1 - \left( \langle \mathbf{\hat{n}} \rangle^\prime_{i} \right)^2 }{\left(1 - \langle \mathbf{\hat{n}} \rangle^\prime_{i} \cdot \mathbf{\hat{u}}_i \right)^2 } .$$ The Berry curvature tensor is obtained from the metric tensor from the following identities
$$\begin{aligned}
-2 f_{ij}^{12} &=& 2 f_{ij}^{21} = g_{ij}^{11} = g_{ij}^{22} , \\
2 f_{ij}^{11} &=& 2 f_{ij}^{22} = g_{ij}^{12} = -g_{ij}^{21} .\end{aligned}$$
Mathematically, this follows from the Kählerian nature of the projective Hilbert space [@chrucinski2004].
To describe the quantum dynamics, I write down Schrödinger’s equation for the state vector $|\psi (t) \rangle \equiv \mathrm{e}^{\mathrm{i}\varphi (t)} \frac{|\Psi_{\mathbf{U}(t)} \rangle}{\langle \Psi_{\mathbf{U}(t)} |\Psi_{\mathbf{U}(t)} \rangle^{1/2}}$, which yields $\dot{\varphi} = - H(\mathbf{U})+ \mathbf{A}\cdot \dot{\mathbf{U}}$ (setting $\hbar \equiv 1$); the latter result is nothing but as the infinitesimal version of the Aharonov-Anandan decomposition of the total phase variation into the dynamical and geometric terms [@aharonov_anandan1987]. It is not gauge invariant with respect to a change of the phase choice for the fiducial states $| \Psi_\mathbf{U}\rangle$. In order to obtain a gauge invariant equation of motion, we use the fact that, due to the unitarity of Hamiltonian evolution, the relative phase of any two (non-orthogonal) states, defined as $\varphi_{12}(t) \equiv \arg \langle \psi_1 (t) | \psi_2(t) \rangle$, is time-independent. Doing this for states at $\mathbf{U}$ and $\mathbf{U}+\delta\mathbf{U}$, one finally obtains $$\label{eq_dynamics}
f_{ij}^{\alpha\beta} \partial_t \hat{u}_j^\beta = \partial_{\hat{u}_i^\alpha}H(\mathbf{U}) .$$ The above equation has the form of the classical equation of motion of a system of $2J$ coupled particles evolving on a spherical phase space, with symplectic form given by $\frac{1}{2} f_{ij}^{\alpha\beta} \mathrm{d}\hat{u}_i^\alpha \wedge \mathrm{d}\hat{u}_j^\beta$, and a Hamilton function given by $H(\mathbf{U})$ [@arnold1978]. An alternative (equivalent) formulation of the spin dynamics in terms of the Majorana stars has been given earlier by Leb[œ]{}uf [@leboeuf1991]; however, the symplectic-Hamiltonian nature of the dynamics is displayed more transparently in the present formulation. I note that the $2J$ particles are coupled to each other, not only dynamically through the Hamilton function, but also kinematically via the symplectic form. Eq. (\[eq\_dynamics\]) is the quantum mechanical counterpart of the Landau-Lifshitz equation for spin dynamics [@landau_lifshitz1935], which we would recover if we would restrict our description to (quasi-classical) CSs. For quantum spin systems, such as molecular magnets [@gatteschi2006], the latter is clearly inadequate, and a fully quantum description as given in Eq. (\[eq\_dynamics\]) is necessary.
Finally, I briefly address the question of systems of interacting spins in magnetically ordered systems. The usual treatment, spin-wave theory, amounts to use a variational wave-function given as a tensorial product of CS with site-dependent unit vectors $\mathbf{\hat{n}}(\mathbf{R}_i,t)$, and solve the linearized coupled Landau-Lifshitz equations. This approach is clearly not adequate for systems with exotic (e.g., quadrupole or higher-multipole) ordering such as spin nematics [@chen_levy1971], and the effective field theory proposed for spin nematics [@ivanov_kolezhuk2003] applies only to spin-1 systems and cannot be extended to systems with spin $J>1$. This clearly calls for a more general theory of spin nematics. I argue here that the most natural description of such systems is the geometric description offered by the Majorana representation. It consists in setting up a quantum field theory based upon Majorana’s constellations instead of the quasi-classical CSs; one thus obtains a path-integral theory for the $2J$ coupled $O(3)$ fields $\mathbf{\hat{u}}_i$. This theory will be developed in detail in a forthcoming paper.
I gratefully acknowledge helpful discussions and/or correspondence with Robert Whitney, Efim Kats, and John Hannay.
[99]{} M.V. Berry, Proc. Roy. Soc. A **392**, 45 (1984).
Y. Aharonov and J. Anandan, Phys. Rev. Lett. **58**, 1593 (1987).
A. Shapere and F. Wilczek (Eds.), *Geometric phases in physics*, World Scientific, Singapore (1989).
A. Bohm *et al.*, *The Geometric Phase in Quantum Systems*, Springer-Verlag, Berlin (2003).
D. Chr[ú]{}ci[ń]{}ski and A. Jamio[ł]{}kowski, *Geometric Phases in Classical and Quantum Mechanics*, Birkhäuser, Boston (2004).
P. Bruno, Phys. Rev. Lett. **93**, 247202 (2004).
J.H. Hannay, J. Phys. A **31**, L53 (1998).
E. Majorana, Nuovo Cimento **9**, 43 (1932).
F. Bloch and I.I. Rabi, Rev. Mod. Phys. **17**, 237 (1945).
J. Schwinger, US Atomic Energy Commission, Report NYO-3071 (1952); later published in *Quantum Theory of Angular Momentum*, ed. by L.C. Biendenharn and H. Van Dam, Academic Press, New York (1965).
A. Perelomov, *Generalized Coherent States and Their Applications*, Springer-Verlag, Berlin (1986).
J. Schwinger, unpublished (1945, 1959); reprinted in *A quantum legacy: seminal papers of Julian Schwinger*, ed. by K.A. Milton, World Scientific, Singapore (2000).
J.M. Caillol, J. Physique Lett. **42**, L245 (1981). Note that the expression of the 2D spherical Coulomb interaction was given there without proof; a proof may be found in *Log-Gases and Random Matrices*, P.J. Forrester, Princeton University Press (2010).
Such an identification of a quantum state with the thermodynamic partition function of a fictitious classical gas is familiar from the theories of the fractional quantum Hall effect (R.B. Laughlin, Phys. Rev. Lett. **50**, 1395 (1983)) and of the valence bond solids in quantum antiferromagnets (D.P. Arovas, A. Auerbach and F.D.M. Haldane, Phys. Rev. Lett. **60**, 531 (1988)).
C.T. Lee, J. Phys. A **21**, 3749 (1988).
R. Gilmore, J. Phys. A **9**, L65 (1976).
J. Anandan and Y. Aharonov, Phys. Rev. Lett. **65**, 1697 (1990).
V.I. Arnold, *Mathematical Methods of Classical Mechanics*, Springer-Verlag, Berlin (1978).
P. Leb[œ]{}uf, J. Phys. A **24**, 4575 (1991).
L.D. Landau and E.M. Lifshitz, Phys. Z. Sowjetunion **8**, 153 (1935); reprinted in *Collected Papers of Landau*, edited by D. ter Haar (Pergamon Press, Oxford 1965), p. 101.
D. Gatteschi, R. Sessoli, and J. Villain, *Molecular Nanomagnets*, Oxford University Press (2006).
H.H. Chen and P.M. Levy, Phys. Rev. Lett. **27**, 1383 (1971); A.F. Andreev and I.A. Grishchuk, Sov. Phys. JETP **60**, 267 (1984); P. Chandra and P. Coleman, Phys. Rev. Lett. **66**, 100 (1991).
B.A. Ivanov and A.K. Kolezhuk, Phys. Rev. B **68**, 052401 (2003).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the Löwner differential equation generating univalent maps of the unit disk (or of the upper half-plane) onto itself minus a single slit. We prove that the circular slits, tangent to the real axis are generated by Hölder continuous driving terms with exponent 1/3 in the Löwner equation. Singular solutions are described, and the critical value of the norm of driving terms generating quasisymmetric slits in the disk is obtained.'
address:
- |
Department of Mathematics and Mechanics,\
Saratov State University,\
Astrakhanskay Str. 83,\
410012, Saratov\
Russia
- |
Department of Mathematics,\
University of Bergen,\
Johannes Brunsgate 12, Bergen 5008,\
Norway
author:
- Dmitri Prokhorov
- 'Alexander Vasil’ev'
date: 'May 2, 2008'
title: Singular and tangent slit solutions to the Löwner equation
---
[^1]
Introduction
============
Let $\mathbb{D}=\{z\in\mathbb{C}:\,\,|z|<1\}$ be the unit disk and $\mathbb T:=\partial
\mathbb D$. The famous Löwner equation was introduced in 1923 [@Loewner] in order to represent a dense subclass of the whole class of univalent conformal maps $f(z)=z(1+c_1z+\dots)$ in $\mathbb D$ by the limit $$f(z)=\lim\limits_{t\to\infty}e^tw(z,t),\quad z\in \mathbb D,$$ where $w(z,t)=e^{-t}z(1+c_1(t)z+\dots)$ is a solution to the equation $$\frac{dw}{dt}=-w\frac{e^{iu(t)}+w}{e^{iu(t)}-w}, \quad w(z,0)\equiv z, \label{LE}$$ with a continuous driving term $u(t)$ on $t\in [0,\infty)$, see [@Loewner page 117]. All functions $w(z,t)$ map $\mathbb D$ onto $\Omega(t)\subset \mathbb D$. If $\Omega(t)= \mathbb D\setminus \gamma(t)$, where $\gamma(t)$ is a Jordan curve in $\mathbb D$ except one of its endpoints, then the driving term $u(t)$ is uniquely defined and we call the corresponding map $w$ a [*slit map*]{}. However, from 1947 [@Kufarev] it is known that solutions to (\[LE\]) with continuous $u(t)$ may give non-slit maps, in particular, $\Omega(t)$ can be a family of hyperbolically convex digons in $\mathbb D$.
Marshall and Rohde [@Marshall] addressed the following question: [*Under which condition on the driving term $u(t)$ the solution to (\[LE\]) is a slit map?*]{} Their result states that if $u(t)$ is Lip(1/2) (Hölder continuous with exponent 1/2), and if for a certain constant $C_{\mathbb D}>0$, the norm $\|u\|_{1/2}$ is bounded $\|u\|_{1/2}<C_{\mathbb D}$, then the solution $w$ is a slit map, and moreover, the Jordan arc $\gamma(t)$ is s quasislit (a quasiconformal image of an interval within a Stolz angle). As they also proved, a converse statement without the norm restriction holds. The absence of the norm restriction in the latter result is essential. On one hand, Kufarev’s example [@Kufarev] contains $\|u\|_{1/2}=3\sqrt{2}$, which means that $C_{\mathbb D}\leq 3\sqrt{2}$. On the other hand, Kager, Nienhuis, and Kadanoff [@Kadanoff] constructed exact slit solutions to the half-plane version of the Löwner equation with arbitrary norms of the driving term.
Let us give here the half-plane version of the Löwner equation. Let $\mathbb H=\{z: {\text{\rm Im }}z>0\}$, $\mathbb R=\partial \mathbb H$. The functions $h(z,t)$, normalized near infinity by $h(z,t)=z-2t/z+b_{-2}(t)/z^2+\dots$, solving the equation $$\frac{dh}{dt}=\frac{-2}{h-\lambda(t)}, \quad h(z,0)\equiv z, \label{LE2}$$ where $\lambda(t)$ is a real-valued continuous driving term, map $\mathbb H$ onto a subdomain of $\mathbb H$. The question about the slit mappings and the behaviour of the driving term $\lambda(t)$ in the case of the half-plane $\mathbb H$ was addressed by Lind [@Lind]. The techniques used by Marshall and Rohde carry over to prove a similar result in the case of the equation (\[LE2\]), see [@Marshall page 765]. Let us denote by $C_{\mathbb H}$ the corresponding bound for the norm $\|\lambda\|_{1/2}$. The main result by Lind is the sharp bound, namely $C_{\mathbb H}=4$.
In some papers, e.g., [@Kadanoff; @Lind], the authors work with equations (\[LE\], \[LE2\]) changing (–) to (+) in their right-hand sides, and with the mappings of slit domains onto $\mathbb D$ or $\mathbb H$. However, the results remain the same for both versions.
Marshall and Rohde [@Marshall] remarked that there exist many examples of driving terms $u(t)$ which are not Lip(1/2), but which generate slit solutions with simple arcs $\gamma(t)$. In particular, if $\gamma(t)$ is tangent to $\mathbb T$, then $u(t)$ is never Lip(1/2).
Our result states that if $\gamma(t)$ is a circular arc tangent to $\mathbb R$, then the driving term $\lambda(t)\in$Lip(1/3). Besides, we prove that $C_{\mathbb D}=C_{\mathbb H}=4$, and consider properties of singular solutions to the one-slit Löwner equation.
The authors are greateful for the referee’s remarks which improved the presentation.
Circular tangent slits
======================
We shall work with the half-plane version of the Löwner equation and with the sign (+) in the right-hand side, consequently with the maps of slit domains onto $\mathbb H$.
We construct a mapping of the half-plane $\mathbb H$ slit along a circular arc $\gamma(t)$ of radius 1 centered on $i$ onto $\mathbb H$ starting at the origin directed, for example, positively. The inverse mapping we denote by $z=f(w,t)=w-2t/w+\dots$. Then $\zeta=1/f(w,t)$ maps $\mathbb H$ onto the lower half-plane slit along a ray co-directed with $\mathbb R^+$ and having the distance 1/2 between them. Let $\zeta_0$ be the tip of this ray. Applying the Christoffel-Schwarz formula we find $f$ in the form $$\frac{1}{f(w,t)}=\int\limits_{0}^{1/w}\frac{(1-\gamma w)\, dw}{(1-\alpha w)^2(1-\beta w)}=
\frac{\beta-\gamma}{(\alpha-\beta)^2}\log\frac{w-\alpha}{w-\beta}+\frac{\alpha-\gamma}
{\alpha-\beta}\frac{1}{w-\alpha},\label{Eq3}$$ where the branch of logarithm vanishes at infinity, and $f(w,t)$ is expanded near infinity as $$f(w,t)=w-\frac{2t}{w}+\dots$$ The latter expansion gives us two conditions: there is no constant term and the coefficient is $-2t$ at $w$, which implies $\gamma=2\alpha+\beta$ and $\alpha(\alpha+2\beta)=-6t$. The condition ${\text{\rm Im }}\zeta_0=-1/2$ yields $$\frac{-2\alpha}{(\alpha-\beta)^2}=\frac{1}{2\pi}.$$ Then, $\beta=\alpha+2\sqrt{-\alpha\pi}$, and $\alpha(3\alpha+4\sqrt{-\alpha\pi})=-6t$. Considering the latter equation with respect to $\alpha$ we expand the solution $\alpha(t)$ in powers of $t^{1/3}$. Hence, $$\alpha(t)=-\left(\frac{9}{4\pi}\right)^{1/3}t^{2/3}+A_2t+A_3t^{4/3}+\dots$$ and $$\beta(t)=(12\pi)^{1/3}t^{1/3}+B_2t^{2/3}+\dots$$ Formula (\[Eq3\]) in the expansion form regarding to $1/w$ gives $$\frac{\beta-\alpha}{2\pi}\;\frac{1}{w}+\frac{\beta^2-\alpha^2}{4\pi}\;\frac{1}{w^2}+\dots+
\left(1+2\frac{\alpha}{\beta}+2\frac{\alpha^2}{\beta^2}+\dots\right)\left(\frac{1}{w}+
\frac{\alpha}{w^2}+\dots\right)=\zeta.\label{Eq4}$$ Remember that this formula is obtained under the conditions $\gamma=2\alpha+\beta$ and $(\alpha-\beta)^2=4\alpha\pi$. We substitute the expansions of $\alpha(t)$ and $\beta(t)$ in this formula and consider it as an equation for the implicit function $w=h(z,t)$. Calculating coefficients $B_2\dots B_4$ in terms of $A_2,\dots, A_4$, and verifying $A_2=-3/4\pi$ we come to the following expansion for $h(z,t)$: $$w=h(z,t)=h(\frac{1}{\zeta},t)=\frac{1}{\zeta}
+2\zeta t+\frac{3}{2}(12\pi)^{1/3}t^{4/3}+\dots.$$
This version of the Löwner equation admits the form $$\frac{dh}{dt}=\frac{2}{h-\lambda(t)}, \quad h(z,0)\equiv z.\label{Eq5}$$ Being extended onto $\mathbb R\setminus \lambda(0)$ the function $h(z,t)$ satisfies the same equation. Let us consider $h(z,t)$, $z\in \widehat{\mathbb H}\setminus \lambda(0)$ with a singular point at $\lambda(0)$, where $\widehat{\mathbb H}$ is the closure of $\mathbb H$. Then $$\lambda(t)=h(z,t)-\frac{2}{dh(z,t)/dt}=\lambda(0)+(12\pi)^{1/3}t^{1/3}+\dots$$ about the point $t=0$. Thus, the driving term $\lambda(t)$ is Lip(1/3) about the point $t=0$ and analytic for the rest of the points $t$.
The radius of the circumference is not essential for the properties of $\lambda(t)$. Passing from $h(z,t)$ to the function $\frac{1}{r}h(rz,t)$ we recalculate the coefficients of the function $h(z,t)$ and the corresponding coefficients in the expansion of $\lambda(t)$ that depend continuously on $r$. Therefore, they stay within bounded intervals whenever $r$ ranges within the bounded interval.
In particular, the expansion for $h(z,t)$ reflects the Marshall and Rohde’s remark [@Marshall page 765] that the tangent slits can not be generated by driving terms from Lip(1/2).
Singular solutions for slit images
==================================
Suppose that the Löwner equation (\[Eq5\]) with driving term $\lambda(t)$ generates a map $h(z,t)$ from $\Omega(t)=\mathbb H\setminus\gamma(t)$ onto $\mathbb H$, where $\gamma(t)$ is a quasislit. Extending $h$ to the boundary $\partial\Omega(t)$ we obtain a correspondence between $\gamma(t)\subset\partial\Omega(t)$ and a segment $I(t)\subset\mathbb
R$, while the remaining boundary part $\mathbb R=\partial\Omega(t)\setminus\gamma(t)$ corresponds to $\mathbb R\setminus I(t)$. The latter mapping is described by solutions to the Cauchy problem for the differential equation (\[Eq5\]) with the initial data $h(x,0)=x\in\mathbb R\setminus\lambda(0)$. The set $\{h(x,t): x\in\mathbb
R\setminus\lambda(0)\}$ gives $\mathbb R\setminus I(t)$, and $\lambda(t)$ does not catch $h(x,t)$ for all $t\geq0$, see [@Lind] for details.
The image $I(t)$ of $\gamma(t)$ can be also described by solutions $h(\lambda(0),t)$ to (\[Eq5\]), but the initial data $h(\lambda(0),0)=\lambda(0)$ forces $h$ to be singular at $t=0$ and to possess the following properties.
\(i) There are two singular solutions $h^-(\lambda(0),t)$ and $h^+(\lambda(0),t)$ such that $I(t)=[h^-(\lambda(0),t),h^+(\lambda(0),t)]$.
\(ii) $h^{\pm}(\lambda(0),t)$ are continuous for $t\geq0$ and have continuous derivatives for all $t>0$.
\(iii) $h^-(\lambda(0),t)$ is strictly decreasing and $h^+(\lambda(0),t)$ is strictly increasing, so that $h^-(\lambda(0),t)<\lambda(t)<h^+(\lambda(0),y)$.
We will focus on studying the singularity character of $h^{\pm}$ at $t=0$.
Let the Löwner differential equation (\[Eq5\]) with the driving term $\lambda\in\text{Lip(1/2)}$, $\|\lambda\|_{1/2}=c$, generate slit maps $h(z,t): \mathbb
H\setminus\gamma(t)\to\mathbb H$ where $\gamma(t)$ is a quasislit. Then $h^+(\lambda(0),t)$ satisfies the condition $$\lim_{t\to 0+}\sup\frac{h^+(\lambda(0),t)-
h^+(\lambda(0),0)}{\sqrt t}\leq\frac{c+\sqrt{c^2+16}}{2},$$ and this estimate is the best possible.
Assume without loss of generality that $h^+(\lambda(0),0)=\lambda(0)=0$. Denote $\varphi(t):=h^+(\lambda(0),t)/\sqrt t$, $t>0$. This function has a continuous derivative and satisfies the differential equation $$t\varphi'(t)=\frac{2}{\varphi(t)-\lambda(t)/\sqrt{t}}-\frac{\varphi(t)}{2}.$$ This implies together with property (iii) that $\varphi'(t)>0$ iff $$\frac{\lambda(t)}{\sqrt
t}<\varphi(t)<\varphi_1(t):=\frac{\lambda(t)}{2\sqrt t}+\sqrt{\frac{\lambda^2(t)}{4t}+4}.$$ Observe that $\varphi_1(t)\leq A:=(c+\sqrt{c^2+16})/2$.
Suppose that $\lim_{t\to0+}\sup\varphi(t)=B>A$, including the case $B=\infty$. Then there exists $t^*>0$, such that $\varphi(t^*)>B-\epsilon>A$, for a certain $\epsilon>0$. If $B=\infty$, then replace $B-\epsilon$ by $B'>A$. Therefore, $\varphi'(t^*)<0$ and $\varphi(t)$ increases as $t$ runs from $t^*$ to 0. Thus, $\varphi(t)>B-\epsilon$ for all $t\in(0,t^*)$ and we obtain from (\[Eq5\]) that $$\frac{dh^+(\lambda(0),t)}{dt}\leq\frac{2}{\sqrt
t(B-\epsilon-c)},$$ for such $t$. Integrating this inequality we get $$h^+(\lambda(0),t)\leq\frac{4\sqrt t
}{B-\epsilon-c}<\frac{4\sqrt t}{A-c},$$ that contradicts our supposition. This proves the estimate of Theorem 3.1.
In order to attain the equality sign in Theorem 3.1, one chooses $\lambda(t)=c\sqrt t$. Then $h^+(\lambda(0),t)=A\sqrt t$ solves equation (\[Eq5\]) with singularity at $t=0$. This completes the proof.
Estimates similar to Theorem 3.1 hold for the other singular solution $h^-(\lambda(0).t)$.
Let us compare Theorem 3.1 with the results from Section 2. The image of a circular arc $\gamma(t)\subset\mathbb H$ tangent to $\mathbb R$ is $I(t)=[h^-(\lambda(0),t),h^+(\lambda(0),t)]$, where $h^-(\lambda(0),t)=\alpha(t)=-(9/4\pi)^{1/3}t^{2/3}+\dots$, and $h^+(\lambda(0),t)=\beta(t)=(12\pi)^{1/3}t^{1/3}+\dots$, so that $h^-(\lambda(0),t)\in\text{Lip}(2/3)$ and $h^+(\lambda(0),t)\in\text{Lip}(1/3)$.
Singular solutions to the differential equation (\[Eq5\]) appear not only at $t=0$ but at any other moment $\tau>0$. More precisely, there exist two families $h^-(\gamma(\tau),t)$ and $h^+(\gamma(\tau),t)$, $\tau\geq0$, $t\geq\tau$, of singular solutions to (\[Eq5\]) that describe the image of arcs $\gamma(t)$, $t\geq\tau$ under map $h(z,t)$. They correspond to the initial data $h(\gamma(\tau),\tau)=\lambda(\tau)$ in (\[Eq5\]) and satisfy the inequalities $h^-(\gamma(\tau),t)<\lambda(t)<h^+(\gamma(\tau),t)$, $t>\tau$. These two families of singular solutions have no common inner points and fill in the set $$\{(x,t): h^-(\lambda(0),t)\leq
x\leq h^+(\lambda(0),t), 0\leq t\leq t_0\},$$ for some $t_0$.
Critical norm values for driving terms
======================================
In this section we discuss the results and techniques of Marshall and Rohde [@Marshall] and Lind [@Lind]. The authors of [@Marshall] proved the existence of $C_{\mathbb D}$ such that driving terms $u(t)\in\text{Lip}(1/2)$ with $\|u\|_{1/2}<C_{\mathbb D}$ in (\[LE\]) generate quasisymmetric slit maps. This result remains true for an absolute number $C_{\mathbb
H}$ in the half-plane version of the Löwner differential equation (\[LE2\]), see e.g. [@Lind].
Lind [@Lind] claimed that the disk version (\[LE\]) of the Löwner differential equation is ‘more challenging’, than the half-plane version (\[LE2\]). Working with the half-plane version she showed that $C_{\mathbb H}=4$. The key result is based on the fact that if $\lambda(t)\in\text{Lip}(1/2)$ in (\[LE2\]), and $h(x,t)=\lambda(t)$, say at $t=1$, then $\Omega(t)=h(\mathbb H,t)$ is not a slit domain and $\|\lambda\|_{1/2}\geq4$. Moreover, there is an example of $\lambda(t)=4-4\sqrt{1-t}$ that yields $h(2,1)=\lambda(1)$. Although there may be more obstacles for generating slit half-planes than that of the driving term $\lambda$ catching up some solution $h$ to (\[LE2\]), Lind showed that this is basically the only obstacle. The latter statement was proved by using techniques of [@Marshall].
We will modify here the main Lind’s reasonings so that they could be applied to the disk version of the Löwner equation. After that it remains to refer to [@Marshall] and [@Lind] to state that $C_{\mathbb D}$ also equals 4.
Suppose that slit disks $\Omega(t)$ correspond to $u\in\text{Lip}(1/2)$ in (\[LE\]) with the sign ‘+’ in its right-hand side instead of ‘-’. Then the maps $w(z,t)$ are extended continuously to $\mathbb T\setminus\{e^{iu(0)}\}$. Let $z_0\in\mathbb T\setminus\{e^{iu(0)}\}$, and let $\alpha(t,\alpha_0):=\arg w(z_0,t)$ be a solution to the following real-valued initial value problem $$\frac{d\alpha(t)}{dt}=\cot\frac{\alpha-u}{2},\;\;\;\alpha(0)=\alpha_0.\label{Eq6}$$
Similarly, suppose that slit half-planes $\Omega(t)$ correspond to $\lambda\in\text{Lip}(1/2)$ in (\[LE2\]) with the sign ‘+’ in its right-hand side instead of ‘-’. Then the maps $h(z,t)$ are extended continuously to $\mathbb R\setminus\lambda(0)$. Let $x_0\in\mathbb
R\setminus\lambda(0)$ and let $x(t,x_0):=h(x_0,t)$ be a solution to the following real-valued initial value problem $$\frac{dx(t)}{dt}=\frac{2}{x(t)-\lambda(t)},\;\;\;x(t_0)=x_0.\label{Eq7}$$
For all $t\geq0$, $\tan((\alpha(t)-u(t))/2)\neq0$ in (\[Eq6\]), and $x(t)-\lambda(t)\neq0$ in (\[Eq7\]) (see [@Lind] for the half-plane version). Let us show a connection between the solutions $\alpha(t)$ to (\[Eq6\]), and $x(t)$ to (\[Eq7\]), where the driving terms $u(t)$ and $\lambda(t)$ correspond to each other.
Given $\lambda(t)\in\text{\rm Lip}(1/2)$, there exists $u(t)\in\text{\rm Lip}(1/2)$, such that equations (\[Eq6\]) and (\[Eq7\]) have the same solutions. Conversely, given $u(t)\in\text{\rm Lip}(1/2)$ there exists $\lambda(t)\in\text{\rm Lip}(1/2)$, such that equations (\[Eq6\]) and (\[Eq7\]) have the same solutions.
Given $\lambda(t)\in\text{Lip}(1/2)$, denote by $x(t,x_0)$ a solution to the initial value problem (\[Eq7\]). Then the solution $\alpha(t,\alpha_0)$ to the initial value problem (\[Eq6\]) is equal to $x(t,\alpha_0)$ when $$\tan\frac{\alpha-u}{2}=\frac{x-\lambda}{2},$$ and $$x_0=\lambda(0)+2\tan\frac{\alpha_0-u(0)}{2}.$$ The function $u(t)$ is normalized by choosing $$u(0)=x_0-\arctan\frac{x_0-\lambda(0)}{2}.$$ This condition makes $\alpha_0$ and $x_0$ equal. Hence, the first part of Lemma 1 is true if we put $$u(t)=x(t,x_0)-2\arctan\frac{x(t,x_0)-\lambda(t)}{2}.\label{Eq8}$$ Obviously, (\[Eq8\]) preserves the $\text{Lip}(1/2)$ property.
Conversely, given $u(t)\in\text{Lip}(1/2)$, a solution $x(t,x_0)$ is equal to $\alpha(t,\alpha_0)$ when $$\lambda(t)=\alpha(t,\alpha_0)-2\tan\frac{\alpha(t,\alpha_0)-u(t)}{2}.
\label{Eq9}$$ Again (\[Eq9\]) preserves the $\text{Lip}(1/2)$ property. This ends the proof.
Observe that in some extreme cases relations (\[Eq8\]) or (\[Eq9\]) preserve not only the Lipschitz class but also its norm. Lind [@Lind] gave an example of the driving term $\lambda(t)=4-4\sqrt{1-t}$ in (\[Eq7\]). It is easily verified that $x(t,2)=4-2\sqrt{1-t}$. If $t=1$, then $x(1,2)=\lambda(1)=4$, and $\lambda$ cannot generate slit half-plane at $t=1$. This implies that $C_{\mathbb H}\leq4$. Going from (\[Eq7\]) to (\[Eq6\]) we use (\[Eq8\]) to put $$u(t)=x(t,2)-2\arctan\frac{x(t,2)-\lambda(t)}{2}=4-2\sqrt{1-t}-2\arctan\sqrt{1-t}.$$ From Lemma 4.1 we deduce that $\alpha(1,2)=u(1)$. Hence $u$ cannot generate slit disk at $t=1$, and $C_{\mathbb D}\leq\|u\|_{1/2}$. Since $$\sup_{0\leq
t<1}\frac{u(1)-u(t)}{\sqrt{1-t}}=\sup_{0\leq
t<1}\left(2+2\frac{\arctan\sqrt{1-t}}{\sqrt{1-t}}\right)=4,$$ we have that $\|u\|_{1/2}\leq4$. It is now an easy exercise to show that $\|u\|_{1/2}=4$. This implies that $C_{\mathbb
D}\leq4$.
Let $u\in\text{\rm Lip}(1/2)$ in (\[Eq6\]) with $u(0)=0$ and $\alpha_0\in(0,\pi)$. Suppose that $\alpha(t)$ is a solution to (\[Eq6\]) and $\alpha(1)=u(1)$. Then $\|u\|_{1/2}\geq4$.
Observe that $\alpha(t)$ is increasing on $[0,1]$, and $\alpha(t)-u(t)>0$ on $(0,1)$. Let $u\in\text{Lip}(1/2)$ in (3), and $\|u\|_{1/2}=c$. Then, $$\alpha(t)-u(t)\leq\alpha(1)-u(1)+c\sqrt{1-t}=c\sqrt{1-t}.\label{Eq10}$$ Given $\epsilon>0$, there exists $\delta>0$, such that $$\tan\frac{c\sqrt{1-t}}{2}<\frac{c\sqrt{1-t}}{2}(1+\epsilon),$$ for $1-\delta<t<1$ and all $0<c\leq4$. We apply this inequality to (\[Eq6\]) and obtain that $$\frac{d\alpha}{dt}\geq\cot\frac{c\sqrt{1-t}}{2}>\frac{2}{c\sqrt{1-t}(1+\epsilon)}.$$ Integrating gives that $$\alpha(1)-\alpha(t)\geq\frac{4\sqrt{1-t}}{c(1+\epsilon)}.$$
This allows us to improve (\[Eq10\]) to $$\alpha(t)-u(t)\leq\alpha(1)-
\frac{4\sqrt{1-t}}{c(1+\epsilon)}-u(1)+c\sqrt{1-t}=\left(c-\frac{4}{c(1+\epsilon)}
\right)\sqrt{1-t}.\label{Eq11}$$
Repeating these iterations we get $$\alpha(t)-u(t)\leq c_n\sqrt{1-t},$$ where $c_0=c$, $c_{n+1}=c-4/[(1+\varepsilon)c_n]$, and $c_n>0$. Let $g_n$ be recursively defined by (see Lind [@Lind]) $$g_1(y)=y-\frac{4}{y},\;\;\;g_n(y)=y-\frac{4}{g_{n-1}(y)},
\;\;\;n\geq2.$$ It is easy to check that $c_n<g_n((1+\varepsilon)c)<(1+\varepsilon)c_n$
Lind [@Lind] showed that $g_n(y_n)=0$ for an increasing sequence $\{y_n\}$, and $g_{n+1}(y)$ is an increasing function from $(y_n,\infty)$ to $\mathbb R$. So $c(1+\epsilon)>y_n$ for all $n$, and it remains to apply Lind’s result [@Lind] that $\lim_{n\to\infty}y_n=4$. Hence, $c\geq4/(1+\epsilon)$. The extremal estimate is obtained if $\epsilon\to0$ which leads to $c\geq4$. This completes the proof.
Now Lind’s reasonings in [@Lind] based on the techniques from [@Marshall] give a proof of the following statement.
If $u\in\text{\rm Lip}(1/2)$ with $\|u\|_{1/2}<4$, then the domains $\Omega(t)$ generated by the Löwner differential equation (\[LE\]) are disks with quasislits.
In other words, Proposition 4.1 states that $C_{\mathbb D}=C_{\mathbb H}=4$.
[9]{}
W. Kager, B. Nienhuis, L. P. Kadanoff, [*Exact solutions for Löwner evolutions*]{}, J. Statist. Phys. [**115**]{} (2004), no. 3-4, 805–822.
J. Lind, [*A sharp condition for the Löwner equation to generate slits*]{}, Ann. Acad. Sci. Fenn. Math. [**30**]{} (2005), no. 1, 143–158.
K. Löwner, [*Untersuchungen über schlichte konforme Abbildungen des Einheitskreises. I*]{}, Math. Ann. [**89**]{} (1923), no. 1-2, 103–121.
D. E. Marshall, S. Rohde, [*The Löwner differential equation and slit mappings*]{}, J. Amer. Math. Soc. [**18**]{} (2005), no. 4, 763–778.
P. P. Kufarev, [*A remark on integrals of Löwner’s equation*]{}, Doklady Akad. Nauk SSSR (N.S.) [**57**]{}, (1947). 655–656 (in Russian).
[^1]: The first author was partially supported by the Russian Foundation for Basic Research (grant 07-01-00120) and the second by the grants of the Norwegian Research Council \#177355/V30, and of the European Science Foundation RNP HCAA.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present deep single-dish observations of a sample of six nearby E+A galaxies ($0.05<z<0.1$). A non-negligible fraction of a local sample of E+As are detected in HI. In four galaxies, we have detected up to a few times $10^9$ [M$_\sun$]{} of neutral gas. These E+A galaxies are almost as gas-rich as spiral galaxies with comparable luminosities. There appears to exist no direct correlation between the amount of present in an E+A galaxy and its star-formation rate as traced by radio continuum emission. Moreover, the end of the starburst does not necessarily require the complete exhaustion of the neutral gas reservoir. Most likely, an intense burst of star formation consumed the dense molecular clouds, which are the sites of massive star formation. This effectively stops star formation, even though copious amounts of diffuse neutral gas remain. The remaining reservoir may eventually lead to further episodes of star formation. This may indicate that some E+As are observed in the inactive phase of the star-formation duty cycle.'
author:
- 'P. Buyle, D. Michielsen, S. De Rijcke, D.J. Pisano, H. Dejonghe, K. Freeman'
title: The content of E+A galaxies
---
Introduction {#intro}
============
At $z \approx 0.5$, galaxy clusters contain a population of blue, distorted galaxies that is missing in local clusters: the so-called Butcher-Oemler effect [@bo78]. In their spectroscopic study of blue galaxies in three clusters at z$\sim$0.31 @cs87 found that sixty percent of these blue galaxies (classified as Type 3 galaxies) have optical spectra characterised by strong Balmer absorption lines, typical for a very young stellar population, but with weak, if any, emission lines, such as [\[\]]{} $\lambda$3727[Å]{}. Since these spectra are a superposition of an old stellar population, resembling that of an elliptical ’or E’ galaxy, and a young stellar population, dominated by A stars, these galaxies are called E+A galaxies. In the notation of @dr99, these galaxies would be classified as k+a/a+k, satisfying the criteria EW([\[\]]{})$< 5$[Å]{} and EW(H$\delta)>$3[Å]{}. While at $z \approx 0.4$, about 20% of all cluster galaxies are classified as E+As [@belloni95], they constitute less than 1% of the present-day clusterpopulation [@fabricant91].
Apparently, these galaxies are observed during a quiescent phase, which explains the lack of emission lines, soon after a vigorous starburst, which explains the strong Balmer absorption lines. Due to the short lifetimes of the stars causing the Balmer absorption, the starburst must have ended no more than $\sim 1$ Gyr ago [@dg83; @p99]. E+As often have disturbed morphologies, e.g. tidal tails, suggestive of recent merger or interaction events. They span the whole morphological range, from bulge-dominated with underlying disks to disk-dominated [@tran03; @yang04]. Their high surface brightness sets them apart from the elliptical and lenticular galaxies in the Fundamental Plane. Over time, fading of the stellar population will drive them towards the locus of the E/S0s [@yang04; @p99]. Internal velocity dispersions of galaxies classified as E+As appear to increase as a function of redshift, going from $\sigma \simeq 150$ [kms$^{-1}$]{} at $z
= 0.3$ to $\sigma \simeq 250$ [kms$^{-1}$]{} at $z=0.83$ [@tran03]. This trend suggests that massive galaxies undergo an E+A phase, i.e. are observed in a post-starburst phase, at earlier cosmic times than less massive ones. This is reminiscent of the “down-sizing” phenomenon in star-forming galaxies [@c96], according to which the masses of galaxies hosting star formation decrease as the Universe ages.
------------------------------ ------------ -------------------------------- ------- ------------------ ------------------ ------------------ -----------------------
Galaxy RA (J2000) $\delta$ (J2000) $M_B$ $v_{\rm hel}$ $\Delta v$ $\int S(v)\,dv$ H[i]{} mass
(h,m,s) ($^\circ$,$\arcmin$,$\arcsec$) (mag) ([kms$^{-1}$]{}) ([kms$^{-1}$]{}) (Jy km s$^{-1}$) ($10^9$ [M$_\sun$]{})
[SDSS J210258.87+103300.6]{} 21 02 58.9 +10 33 01 -21.7 27821 440 $0.18 \pm 0.02$ $6.5 \pm 0.8$
[SDSS J230743.41+152558.4]{} 23 07 43.4 +15 25 58 -20.4 20894 240 $0.04 \pm 0.01$ $0.9 \pm 0.3$
[SDSS J233453.20+145048.7]{} 23 34 53.2 +14 50 49 -20.2 19388 380 $0.15 \pm 0.02$ $2.7 \pm 0.3$c
LCRS B101120.1-024053 (EA17) 10 13 52.4 -02 55 48 -18.2 18258 $<0.18$ $< 2.9 $
LCRS B002018.8-415015 (EA18) 00 22 47.1 -41 33 37 -18.9 17941 660 $0.15 \pm 0.02$ $2.3 \pm 0.3$
LCRS B020551.6-453502 (EA19) 02 07 49.7 -45 20 50 -18.9 19186 $<0.07$ $< 1.2 $
------------------------------ ------------ -------------------------------- ------- ------------------ ------------------ ------------------ -----------------------
\
------------------------------ ------------- ----------- ------- -------------- ---------------------------
Galaxy \[O[ii]{}\] H$\delta$ $r_e$ Sérsic index SFR
([Å]{}) ([Å]{}) (kpc) ([M$_\sun$]{}/yr)
[SDSS J210258.87+103300.6]{} -0.68 5.13 4.6 3.2 $10.2^{^{+15.7}_{-6.5}}$
[SDSS J230743.41+152558.4]{} -0.92 6.78 2.0 3.1 $1.2^{^{+1.8}_{-0.8}}$
[SDSS J233453.20+145048.7]{} -1.27 5.18 2.1 3.3 $5.6^{^{+10.1}_{-3.7}}$
LCRS B101120.1-024053 (EA17) 1.68 6.92 1.5 2.7/1.3 $< 8.2^{^{+18.1}_{-5.9}}$
LCRS B002018.8-415015 (EA18) 1.75 5.96 1.7 2.2 $5.3^{^{+10.4}_{-3.7}}$
LCRS B020551.6-453502 (EA19) 0.98 6.08 2.1 1.1 $<1.8^{^{+2.7}_{-1.1}}$
------------------------------ ------------- ----------- ------- -------------- ---------------------------
Although the majority of low-$z$ E+A galaxies have a smooth light distribution, many of them also show slightly disturbed morphologies (e.g. warps and dust lanes). Based on this, @zab96 argued that the E+A phase is the aftermath of a vigorous starburst, triggered by a major merger or interaction. This is corroborated by a study of E+A galaxies drawn from the 2dFGRS catalogue [@blake04]. About three quarters of all E+A galaxies are found in the field (i.e. outside clusters), simply because most of the galaxies in the universe do not reside in clusters. However, the *fraction* of E+As is four times higher in clusters than in the field [@zab96; @tran04]. The spatial distribution of E+A galaxies in clusters is more extended than that of quiescent galaxies, but less extended than that of emission-line galaxies [@dr99], suggesting that processes such as galaxy harassment or ram-pressure stripping, which are specific to clusters, can also cause the E+A phenomenon. Based on the 2dFGRS E+A sample, @blake04 find that the distribution of [*local*]{} environments of E+A galaxies closely traces that of the ensemble of 2dFGRS galaxies and conclude that whatever causes the E+A phenomenon, must be a very local mechanism, such as encounters of galaxy pairs. This is corroborated by a recent analysis of the environments of E+As selected from the SDSS [@goto05b].
Most E+As have E/S0-like morphologies, with a small fraction of ongoing interactions. Their luminosity distribution is more similar to the distribution of spectroscopically defined elliptical galaxies than to the luminosity distribution of the ensemble of 2dFGRS galaxies [@blake04]. However, not all E+As can be associated with mergers and, obviously, more than one evolutionary pathway can lead to a post-starburst galaxy [@tran03; @dr99]. Numerical simulations show that E+As can indeed be formed via a major merger of two gas-rich spiral galaxies [@bekki05]. A disk-disk merger event then triggers a starburst, which consumes, or, by feedback, expels most of the available gas and then subsides. The young stars then dominate the optical spectrum for the following few hundred Myr while emission lines are absent. During this time-span, a galaxy would be classified as an E+A. In this case, one expects star-formation to be centrally concentrated, leading to radially decreasing Balmer line strengths [@pracy05]. Alternatively, star formation can be truncated more or less instantaneously over the whole disk of the galaxy without a starburst, e.g. by the gas being swept away by ram pressure stripping. In this case, as the young star population fades, the older bulge population causes the strengths of the Balmer lines to be radially increasing [@pracy05].
The red colours of some H$\delta$-strong E+As cannot be explained by any plausible starburst model [@cs87; @blake04], leaving heavy dust extinction as the only viable explanation. This hypothesis can be tested by using dust insensitive tracers of star formation. Since radio continuum emission is synchrotron radiation from electrons accelerated in supernova remnants, it is an indirect tracer of star formation. @mo01 observed part of the @zab96 sample and detected radio-continuum emission in only two out of fifteen galaxies. @smg99 detect five out of eight post-starburst galaxies at radio wavelengths. Radio-continuum observations of a sample of 36 E+As drawn from the SDSS yielded no detections [@goto05]. If this apparent lack of ongoing star-formation in E+As is true, then no dust obscuration, hiding the star-formation sites, needs to be invoked. Near-infrared studies [@gal00; @bal05] have shown that the u$-$g and r$-$k colours, and the H$\delta$ line-strengths of E+As can be well explained by dust-free models in which more than 5% of the stellar mass has recently been produced in a starburst. Hence, the presence of dust is still uncertain because of these contradictory observations.
Up to now only one search for in E+As was conducted [@chang01]. VLA observations of five E+As from the sample of @zab96 resulted in the detection of only one field E+A galaxy, EA1, with a total H[i]{} mass of $7.1 \pm 0.4 \times 10^9\,M_\odot$ (assuming $H_0 = 70$ [kms$^{-1}$]{}Mpc$^{-1}$). For the four other galaxies, upper limits of order $10^9\,M_\odot$ could be derived. EA1 consists of two components, and is most likely a merger remnant. However, other galaxies in this sample are also optical mergers but they do not contain detectable amounts of gas. We started an study of a sample of E+A galaxies in order to constrain the amount of neutral gas present in these systems. In section \[obs\] we describe the sample and the observations. The results are presented in section \[results\]. We discuss these results in section \[discussion\] and summarise our conclusions in section \[concl\].
 \
 \
 
\[HIspectra\]
Observations {#obs}
============
The sample
----------
Since [*(i)*]{} the occurrence of the E+A phenomenon seems to be determined predominantly by the local environment, [*(ii)*]{} the mechanisms triggering the E+A phenomenon, i.e. mergers, interactions, ram-pressure stripping, act both at high and low redshift, and [*(iii)*]{} for a given flux, the mass scales with distance squared, we opted to use distance as our main selection criterion to maximise our chances of a detection. We selected the three closest E+A galaxies from the large catalogue compiled by @goto03 from the SDSS, and the three closest E+A galaxies from the compilation of @zab96. The E+A sample selected from the SDSS by @goto03 satisfies very strict criteria and contains only galaxies with EW(H$\delta)>4{\rm\AA}+\Delta{\rm EW}({\rm H}\delta)$, with $\Delta{\rm EW}({\rm H}\delta)$ the 1$\sigma$ error on the H$\delta$ equivalent width, and no detectable [\[\]]{} and H$\alpha$ emission, quantified by the constraints EW([\[\]]{}$)<\Delta$EW([\[\]]{}) and EW(H$\alpha)<\Delta{\rm EW}({\rm H}\alpha)$, respectively. The Zabludoff sample satisfies the following criteria: EW(H$\beta\gamma\delta)>5.5{\rm\AA}$ and EW([\[\]]{}$)<2.5$[Å]{}, with EW(H$\beta\gamma\delta$) the mean equivalent width of the H$\beta$, H$\gamma$, and H$\delta$ absorption lines. Table \[EAproperties\] summarises the properties of the galaxies in our sample.
Arecibo observations {#arecibo}
--------------------
We observed the galaxies SDSS J210258.87+103300.6, [[SDSS J230743.41+152558.4]{}]{}, and [[SDSS J233453.20+145048.7]{}]{} for 7.5 h each, including overhead, with the 305m Arecibo Radio Telescope[^1] in Puerto Rico. The observations were scheduled on the nights of 23$-$25 June, 13$-$15 July and 28$-$30 July 2005. Each galaxy was observed for 2.5 hours per day during night-time to minimise solar interference. We used the L-wide receiver which has an average system temperature of $\approx 27$ K (depending on the elevation of the source). We selected the interim correlator in both linear polarisations to process the data. This resulted in final spectra with a total bandwidth of 25 MHz and 12.5 MHz divided over 1024 channels, resulting in a velocity resolution of 6.3 [[kms$^{-1}$]{}]{} and 3.15 [[kms$^{-1}$]{}]{} respectively. For [[SDSS J230743.41+152558.4]{}]{} and [[SDSS J233453.20+145048.7]{}]{}, we used the radar blanker to minimise the effect of the FAA Airport radar at 1330 MHz and 1350 MHz. The beam size of the L-wide receiver is $3.1'\times3.5'$. We applied the standard position-switching algorithm. Each galaxy was observed for 5 minutes, followed by a 5 minute offset by 5$'$ in right ascension to blank sky, such that we tracked the same azimuth and zenith angle as the on-source scan. This mode was used for all galaxies. We reduced the data by means of the standard Arecibo [IDL]{} routines, written by P. Perrilat. We calibrated each bandwidth individually and the polarisations were averaged together before a second-order baseline was fit across the interference-free part of the spectrum. We checked, with NED, the recession velocities of all galaxies inside the Arecibo beam to avoid confusion with other objects. The results are presented in section \[results\].
Parkes observations {#parkes}
-------------------
We observed the galaxies LCRS B101120.1-024053 (EA17), LCRS B002018.8-415015 (EA18), and LCRS B020551.6-453502 (EA19) with the 64m ATNF Parkes Radio Telescope[^2] in Australia from dusk till dawn from 11 till 15 October 2005. EA17 was observed during sunrise. We used the Multibeam Correlator in the MB13 configuration (beam-switching mode) which enabled us to observe with 7 beams simultaneously, with one beam on the source while the other six were pointed adjacently on the sky. The beams were switched in position each 5 minutes. This way, we derived spectra with a bandwidth of 64 MHz divided over 1024 channels which yields a spectral resolution of 13.19 [[kms$^{-1}$]{}]{} and a beam width of 14.1$'$. The integration times, including overhead, were 10 h for EA17, 21 h for EA18, and 10 h for EA19. This resulted in one clear 3$\sigma$-detection of EA18. All observational quantities are listed in table \[EAproperties\]. The data were reduced by means of the [Livedata]{} data reduction pipeline, which is especially developed for the Multibeam Correlator. No obvious radio interference could be observed and a second order polynomial was used to fit the spectral baseline, after masking out a region around the optical velocity. Afterwards, all data were combined with the help of the [Gridzilla]{} software package using the median of the weighted values as an estimator. Finally, residual baseline ripples were removed by using the [mbspect]{} fitting algorithm in the [MIRIAD]{} [@sault95] software package. Again we checked the recession velocities of all other objects within the Parkes beam to avoid confusion. The results are presented in section \[results\].
{width="7cm"} {width="7cm"}\
{width="7cm"} {width="7cm"}\
{width="7cm"} {width="7cm"}
Results
=======
masses
-------
In Figure \[HIspectra\], we show the spectra of our six E+As. Three of them have a clear 3$\sigma$ detection: SDSS J210258.87+103300.6, [[SDSS J233453.20+145048.7]{}]{}, and EA18. In the case of [SDSS J230743.41+152558.4]{}, there is a peak at the correct velocity, over a velocity width of 240 [kms$^{-1}$]{}, which could be a tentative 2.5$\sigma$ detection of this galaxy. We did not detect EA17 and EA19. The observations of EA17 were badly affected by solar interference though.
We calculate masses for the detected objects and 3$\sigma$ upper limits for the undetected galaxies using the formula $$M_{\rm H\,I} = 2.36 \times 10^5\,M_\odot\, D^2 \int S(v)\, dv$$ with the distance $D$ in Mpc and $\int S(v) \,dv$ the total flux density in Jy [kms$^{-1}$]{}. The distance $D$ was calculated as the Hubble distance $D = v/H_0$, using $H_0 = 70$ [kms$^{-1}$]{}Mpc$^{-1}$. We find masses of $6.5 \pm 0.8 \times 10^9\,M_\odot$ for [[SDSS J210258.87+103300.6]{}]{}, $0.9 \pm
0.3 \times 10^9\,M_\odot$ for [[SDSS J230743.41+152558.4]{}]{}, $2.7 \pm 0.3 \times
10^9\,M_\odot$ for [[SDSS J233453.20+145048.7]{}]{}, and $2.3 \pm 0.3 \times 10^9\,M_\odot$ for EA18, as listed in Table \[EAproperties\]. To estimate the error on the mass of a galaxy, we generated 50000 statistically equivalent renditions of its radio spectrum by adding Gaussian noise to the original spectrum, using the measured 1$\sigma$ noise on the original datapoints. The mass error is taken to be the rms of the 50000 masses measured from these spectra. The $3\sigma$ upperlimits for the gas content of EA17 and EA19 are $2.9 \times 10^9\,M_\odot$ and $1.2 \times 10^9\,M_\odot$ respectively, assuming a total velocity width of 450 [[kms$^{-1}$]{}]{} (the average velocity width of the detected galaxies).
Dust content
------------
Given the small angular sizes of these galaxies, their dust content can only be studied accurately with the supreme spatial resolution of the Hubble Space Telescope (HST). The HST archive contains images of EA17, EA18, and EA19, obtained with the Advanced Camera for Surveys (ACS) through the F435W and F625W filters. In Fig. \[ima\_EA\], we show the F625W images and F425W$-$F625W color maps of these galaxies. From the images it is obvious that the E+A class of galaxies comprises very different morphological types.
EA17 has a smooth appearance, but it also contains a dust lane, which is clearly visible in both the colour-map and plain image. The only galaxy of this sample of three, that was detected in , is EA18, which is an irregular-looking galaxy with lots of red and blue patches along the major axis, pointing to dust and/or young stars. EA18 appears also to be warped. Finally, EA19 is a spiral (Sab) galaxy, with two spiral arms nicely visible in both images. The spiral arms are slightly bluer than the rest of the disk and the centre contains blue and red patches, pointing again to dust and/or young stars.
From just this small sample, it is already clear that dust is present near the centres of some E+A galaxies. Along with the presence of neutral gas this might indicate that on-going star formation could in fact be hidden by dust.
Environment
-----------
Although E+A galaxies are predominately located in the field (i.e. outside clusters), their fraction is four times higher in clusters [@zab96; @tran04]. By investigating their spatial distribution inside clusters, @dr99 showed that an E+A phenomenon can be caused by environmental processes such as galaxy harassment or ram-pressure stripping. Such processes act mostly on the neutral hydrogen content of galaxies and can therefore be investigated by means of our observations.
We adopted the determination of cluster membership of @zab96 and also checked the number of neighbours within a radius of 0.5Mpc of each galaxy. Those E+As that were listed in @zab96 as being cluster members (EA4 and EA11) have more than 10 known neighbours. Others, such as EA1, EA3, EA18 have 1 or 2 known neighbours and are clearly field galaxies, while the remaining ones have 3-7 known neighbours. For the SDSS galaxies from @goto03, cluster membership is not given. Based on the number of known near neighbours, only [SDSS J230743.41+152558.4]{} is a possible cluster galaxy.
Hence, of the five E+As detected at 21 cm up to now, one is a cluster member ( SDSS J230743.41+152558.4) and four are not. Of those not detected, two are cluster members and four are not. From this result, it is clear that it is premature to draw conclusions on the environment by means of neutral hydrogen observations. In order to do so, a larger sample is required.
Optical emission line strengths
-------------------------------
As E+A galaxies are defined by means of their optical spectra, and more precisely by the equivalent widths of primarily the emission lines [\[\]]{}and H$\delta$ [@zab96; @goto03], one can investigate any trend between the neutral hydrogen content and the equivalents widths. For our sample we list these values in Table \[EAproperties\]. We restate that our E+A galaxies are compiled from two different samples with each different constraints concerning the equivalent widths of [\[\]]{} and H$\delta$. In Table \[othersample\] we list the EW([\[\]]{}) and EW(H$\delta$) for the undetected galaxies from @chang01.
Both groups of detected and undetected E+A galaxies at 21 cm contain a mixture of [\[\]]{} absorption and emission lines. Similarly, E+A galaxies with low and high EW(H$\delta$) are detected. We conclude that again a larger sample is needed in order to investigate any trend between the optical emission lines and their neutral hydrogen content.
-------- -------------- ----------- ----------------------- -- -- -- -- -- --
Galaxy [[\[\]]{}]{} H$\delta$ H[i]{} mass
([Å]{}) ([Å]{}) ($10^9$ [M$_\sun$]{})
EA01 1.80 8.98 7.1
EA02 1.25 7.98 $<3.1$
EA03 -0.29 8.13 $<3.9$
EA04 1.37 9.82 $<2.0$
EA11 2.16 6.96 $<4.7$
-------- -------------- ----------- ----------------------- -- -- -- -- -- --
: Properties of the E+A galaxies of @chang01.[]{data-label="othersample"}
Star-formation rate
-------------------
We estimate the star-formation rate (SFR) associated with the observed H[i]{} masses using the relation $$\Sigma_{\rm SFR} \approx 2.5^{\pm 0.7} \times 10^{-10} \left( \frac{\Sigma_{\rm
gas}}{M_\odot \, {\rm pc}^{-2}} \right)^{1.40^{\pm 0.15}} M_\odot\,{\rm
pc}^{-2}\,{\rm yr}^{-1}, \label{sfr}$$ [@k98]. We substituted the total H[i]{} mass divided by $\pi
R_{\rm e}^2$, with $R_{\rm e}$ the half-light radius, for the gas surface density $\Sigma_{\rm gas}$. From the archival F625W HST/ACS images of EA17, EA18 and EA19, we derived surface brightness profiles as a function of radius by integrating the light in circular apertures centered on the galaxy. A similar method was applied to derive the half-light radii for the SDSS galaxies from Sloan r-band images. We fitted seeing or PSF convolved Sérsic profiles to the surface brightness profiles of all galaxies. The seeing and PSF profiles are determined from about 10 stars in each image. We used the surface brightness profile of a galaxy, extrapolated beyond the last data point by the best fitting Sérsic profile, to measure the half-light radii of these E+As. The SFR surface density, $\Sigma_{\rm SFR}$, was then converted into a global SFR by multiplying with $\pi R_{\rm
e}^2$. In all cases but SDSS J230743.41+152558.4 and EA19, we found the E+As detected at 21 cm to have SFRs in the range $5-10\,M_\odot\,{\rm yr}^{-1}$ (see Table \[EAproperties\]), which is higher than expected for post-starburst galaxies. Hence, these gas-rich E+As could be actively forming stars at quite high rates but the star-formation sites are obscured by dust. Alternatively, although much gas is present, almost no stars are being formed. The radio continuum observations of [@mo01] rule out star formation at a rate higher than $1.0-1.5\,M_\odot\,{\rm yr}^{-1}$ in the case of EA17 and EA18. These authors measure a star formation rate of $1.5\,M_\odot\,{\rm yr}^{-1}$ in EA19. While the star formation rate of EA19 agrees roughly with our upper limit, the high star formation rates of EA17 and EA18 derived from their H[i]{} content are in clear contradiction with the very low rates derived from radio continuum observations. Unfortunately, no radio continuum observations have been performed of [[SDSS J210258.87+103300.6]{}]{}, [[SDSS J230743.41+152558.4]{}]{}, and [[SDSS J233453.20+145048.7]{}]{} so far.
Discussion
==========
Three out of the six galaxies observed by us contain detectable amounts of neutral gas and one is a border case. Both the SDSS and the Zabludoff samples satisfy very strict selection criteria and surely do not show any optical evidence for star formation: these appear to be true post-starburst galaxies. They do, however, contain significant amounts of neutral gas. EA19 is not detected by us in 21 cm line emission, with a 3$\sigma$ upperlimit on the mass of $1.2 \times 10^9\,M_\odot$. EA18 was detected by us, with a 21 cm flux consistent with an mass of $2.3 \pm 0.3 \times 10^9\,M_\odot$ of .
![$\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$. Spiral galaxies [@h04] are depicted by spiral symbols, elliptical galaxies [@g01] by blue circles (the shade of blue is an indication of spectroscopic age and the vertical black line shows the luminosity limit of log($L_B$)=9.5 that was used by @g01), and E+A galaxies by red circles (this work) and orange circles [@chang01]. The predicted trend for late-type galaxies by the semi-analytical models of @ny04 is plotted in green. \[logLM\]](f3.eps){width="8.5cm"}
In Fig. \[logLM\], we plotted $\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$, $L_B$ the B-band luminosity expressed in solar B-band luminosities, for a sample of spiral galaxies [@h04], elliptical galaxies [@g01], and E+A galaxies [this work and @chang01] along with the predicted $\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$ relation for late-type galaxies [@ny04]. For the E+As from the SDSS sample, we converted the SDSS g and r magnitudes to the B-band magnitude using the conversion formulae of [@j05]. The b$_J$ magnitudes of the LCRS E+As were converted into B-band magnitudes using the relation $m_B = m_J + 0.28 ({\rm B}-{\rm V})$ from [@mes90]. Using the mean B$-$V colour of the SDSS E+As, $\langle ({\rm B}-{\rm V}) \rangle = 0.78$ mag, in this equation, we find $m_B \approx m_J + 0.22$ mag. The absolute B-band magnitudes found by applying these conversion formulae are also listed in Table 1. Ellipticals with young spectroscopic ages ($\lesssim 1-2$ Gyr) are generally more gas-rich than older elliptical galaxies, which @g01 interpret as evidence for a merger origin for elliptical galaxies: the starburst following the merger rapidly consumes the gas reservoir and subsequent quiescent star formation consumes the gas at a much slower rate. It should be noted that @g01 include only ellipticals brighter than $M_B = -18.5$ mag, or $\log(L_B) = 9.5$, in their sample because fainter galaxies are likely to have experienced different evolutionary histories. The resulting sample of ellipticals spans about the same luminosity range as the E+A samples.
As is clear from Fig. \[logLM\], some E+As are more gas-rich than most young elliptical galaxies. This is most likely not a selection effect, since the E+A data set was assembled based on optical, spectral properties and the elliptical galaxies were selected according to optical morphological considerations, not on H[i]{} mass. One would, moreover, expect the @g01 data set to be complete for galaxies with large H[i]{} masses. We performed a Kolmogorov-Smirnov (KS) and Wilcoxon (W) test on the $\log(M_{\rm
HI}/L_B)$ distribution (within the same magnitude limits, both in $L_B$ and $\log(M_{\rm HI}/L_B)$, i.e. in the region defined by the detected E+As) and found that the distribution of E+A galaxies (i) corresponds with that of spiral and young elliptical galaxies with a significance of respectively 99% (KS) or 94% (W) and 65% (KS) or 79% (W) and (ii) differs from that of old elliptical galaxies with a confidence level of 79% (KS) or 78% (W). This statistical test and Fig. \[logLM\] might suggest a gas depletion time sequence, with E+As being observed less than $\sim 500$ Myr after the termination of the starburst that followed the putative merger [@yang04], young ellipticals after $\lesssim 1-2$ Gyr, and old ellipticals at later times. However, one should note that the difference between the distributions of E+As and elliptical galaxies has just a significance of 1$\sigma$. In order to derive conclusive evidence for this suggestion, a larger statistical sample of H[i]{} masses and spectroscopic ages of E+As is essential. Nevertheless, if correct, this ties together with the morphological study of 5 E+As from the Zabludoff (1996) sample with HST by @yang04, who suggest that E+As are likely to evolve to elliptical galaxies with power-law density profiles.
One should also note that the end of a starburst does not necessarily require the complete exhaustion of the neutral gas reservoir. This could be explained by the fact that neutral gas itself is not the raw material for star formation: stars form in the dense cores of molecular clouds. @ko02 have obtained CO($1\!\rightarrow\! 0$) and HCN($1\!\rightarrow\! 0$) observations of the nearby post-starburst galaxy NGC5195, which, interestingly, forms an interacting pair with NGC5194. These authors note a central decrease of the mass fraction of dense molecular cores, leaving only diffuse molecular gas, evidenced by a very low central $L_{\rm HCN}/L_{\rm
CO}$ value. Most likely, an intense central burst of star formation $\sim 1$ Gyr ago, responsible for the observed population of A stars, evaporated the dense molecular clouds, which are the sites where massive stars form. This effectively stopped further star formation, although large amounts of diffuse neutral and molecular gas remain. A similar mechanism may be responsible for switching off the starburst in E+A galaxies without the necessity of consuming the complete gas reservoir. The remaining reservoir in some E+As may eventually lead, after gas has been allowed to condense into molecular cores, to further episodes of star formation. Eq. (\[sfr\]), which is based on observations of normal spiral galaxies that host a balanced mix of neutral and molecular gas, may therefore not be applicable to post-starburst galaxies.
Conclusions {#concl}
===========
We present deep single-dish observations of a sample of six nearby E+A galaxies ($0.05<z<0.1$). We find masses of $6.5 \pm
0.8 \times 10^9\,M_\odot$ for [[SDSS J210258.87+103300.6]{}]{}, $0.9 \pm 0.3 \times
10^9\,M_\odot$ for [[SDSS J230743.41+152558.4]{}]{}, $2.7 \pm 0.3 \times 10^9\,M_\odot$ for [[SDSS J233453.20+145048.7]{}]{}, and $2.3 \pm 0.3 \times 10^9\,M_\odot$ for EA18. The $3\sigma$ upperlimits for the gas content of EA17 and EA19 are $2.9
\times 10^9\,M_\odot$ and $1.2 \times 10^9\,M_\odot$ respectively, assuming a total velocity width of 450 [[kms$^{-1}$]{}]{} (the average velocity width of the detected galaxies). The three galaxies from the SDSS sample satisfy very strict selection criteria and, for all practical purposes, can be considered to be truly post-starburst galaxies.
The E+A galaxies detected in 21 cm line emission are almost as gas-rich as spiral galaxies with comparable luminosities. By plotting these E+As, spiral galaxies, and elliptical galaxies in a $\log(M_{\rm
HI}/L_B)$ versus $\log(L_B)$ diagram, and performing a Kolmogorov-Smirnov and Wilcoxon test we suggest that existence of a gas depletion time sequence, with E+As being observed very shortly after the termination of the starburst that ensued from the putative merger, young ellipticals after $\lesssim 1-2$ Gyr, and old ellipticals at even later times. This would tie together with the morphological study of 5 E+As from the Zabludoff (1996) sample with HST by @yang04, who suggest that E+As are likely to evolve to elliptical galaxies with power-law density profiles. However, the conclusions drawn here are just based on 1$\sigma$ trends. A larger sample of detections and spectroscopic ages of E+As is required to investigate this hypothesis.
The presence of in an E+A galaxy can be explained in two ways.\
(i) We can interpret the lack of on-going star-formation in E+A galaxies, suggested by previous radio continuum observations [@mo01] and indirectly by their selection criteria, and the fact that the end of the starburst does not necessarily require the complete exhaustion of the neutral gas reservoir, as being due to the effect the starburst has on the dense molecular cores which are responsible for the massive star formation. An intense burst of star formation can evaporate the dense molecular clouds, effectively stopping further star formation, even though copious amounts of diffuse neutral and molecular gas remain. The remaining reservoir in some E+As may eventually lead, after the gas has again condensed into molecular cores, to further episodes of star formation. This may indicate that E+As are observed in the inactive phase of the star-formation duty cycle.\
(ii) A second possibility is what previously has been proposed by Couch & Sharples (1987) and Blake et al. (2004). There might still be on-going star-formation associated with the presence of gas, which is hinted from the high SFR (see Table \[EAproperties\]) of our E+A galaxies; however it cannot be observed since it is obscured by dust [@smg99], which is suggested by the ACS images in Fig. \[ima\_EA\], and/or optical emission lines might be a poor way of isolating true post-starbust systems. In this case E+A galaxies are in the active star-formation phase and will presumably exhaust all their gas content.
PB acknowledges financial support from the Bijzonder OnderzoeksFonds (BOF). DM is supported by the MAGPOP EU Marie Curie Training and Research Network. SD acknowledges financial support from the Fonds voor Wetenschappelijk Onderzoek – Vlaanderen (FWO). This research was performed while D. J. P. held a National Research Council Research Associateship Award at the Naval Research Laboratory. Basic research in astronomy at the Naval Research Laboratory is funded by the Office of Naval Research. This work has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Based on observations made with the NASA/ESA Hubble Space Telescope, obtained from the data archive at the Space Telescope Institute. STScI is operated by the association of Universities for Research in Astronomy, Inc. under the NASA contract NAS 5-26555.
Balogh, M. L., Miller, C., Nichol, R., Zabludoff, A., & Goto, T. 2005, , 360, 587 Bekki, K., Couch, W. J., Shioya, Y., & Vazdekis, A. 2005, , 359, 949 Belloni, P., Bruzual, A. G., Thimm, G. J., Roser, H.-J., 1995, A&A, 297, 61 Blake, C., Pracy, M. B., Couch, W. J., Kenji, B., Lewis, I., Glazebrook, K., Baldry, I. K., Baugh, b. M., Bland-Hawthorn, J., et al., 2004, , 355, 713 Butcher, H., Oemler, A., Jr., 1978, ApJ, 219, 18 Chang, T. C., van Gorkum, J. H., Zabludoff, A. I., Zaritsky, D., & Mihos, J. C. 2001, , 121, 1965 Couch, W. J., Sharples, R. M., 1987, MNRAS, 229, 423 Cowie, L. L., Songaila, A., Hu, Ester M., Cohen, J. G., 1996, AJ, 112, 839 Dressler, A., Gunn, J. E., 1983, ApJ, 270, 7 Dressler, A., Smail, I., Poggianti, B. M., Butcher, H., Couch, W. J., Ellis, R. S., & Oemler, A., Jr., 1999, ApJS, 122, 51 Fabricant, D.G., McClintock, J. E., Bautz, M.W., 1991, , 381, 33 Galaz, G. 2000, , 119, 2118 Georgakakis, A., Hopkins, A. M., Caulton, A., Wiklind, T., Terlevich, A. I., Forbes, D., A., 2001, , 326, 1431 Goto, T., Nichol, R. C., Okamura, S., Sekiguchi, M., Miller, C. J., Bernardi, M., Hopkins, A., Tremonti, C., Connolly, A., Castander, F. J., et al. 2003, , 55, 771 Goto, T. 2004, A&A, 427, 125 Goto, T. 2005, MNRAS, 357, 937 Helmboldt, J. F., Walterbos, R. A. M., bothun, G. D., O’Neil, K., de Blok, W. J. G., 2004, , 613, 914 Jester, S., Schneider, D. P., Richards, G. T., Green, R. F., Schmidt, M., Hall, P. B., Strauss, M. A., Vanden Berk, D. E., et al., 2005,, 130, 873 Kennicutt, R. C., Jr., 1998, ARA&A, 36, 189 Kohno, K., Tosaki, T., Matsushita, S., Vila-Vilaó, B., Shibatsuka, T., Kawabe, R., 2002, PASJ, 54, 541 Maddox, S. J., Efstathiou, G., Sutherland, W. J., 1990, , 246, 433 Miller, N. A., & Owen, F. N. 2001, , 554, 125 Nagashima, M. & Yoshii, Y., 2004, , 610, 23 Poggianti, B. M., Smail, I., Dressler, A., Couch, W. J., Barger, J., Butcher, H., Ellis, E. S., & Oemler, A., Jr., 1999, ApJ, 518, 576 Pracy, M. B., Couch, W. J., Blake, C., Bekki, K., Harrison, C., Colless, M., Kuntschner, H., & de Propris, R. 2005, , 359, 1421 Sault, R.J., Teuben, P.J. & Wright, M.C.H., 1995, in Shaw R., Payne H.E., Hayes, J.J.E., eds, ASP Conf Ser. Vol. 77, Astronomical Data Analysis Software and Systems IV. Astron. Soc. Pac., San Fransisco, p. 433 2005, , 359, 1421 Smail, I., Morrison, G., Gray, M. E., Owen, F. N., Ivison, R. J., Kneib, J.-P., Ellis, R. S. Gray, M. E., et al. 1999, ApJ, 525, 609 Tran, K.-V. H., Franx, M., Illingworth, G., Kelson, D. D., & van Dokkum, P. 2003, , 599, 865 Tran, K.-V. H., Franx, M., Illingworth, G. D., van Dokkum, P., Kelson, D. D., & Magee, D. 2004, , 609, 683 Yang, Y., Zabludoff, A. I., Zaritsky, D., Lauer, T. R., & Mihos, J. C. 2004, , 607, 258 Zabludoff, A. I., Zaritsky, D., Lin, H., Tucker, D., Hashimoto, Y., Shectman, S. A., Oemler, A., & Kirshner, R. P. 1996, , 466, 104
[^1]: The Arecibo Observatory is part of the National Astronomy and Ionosphere Center, which is operated by Cornell University under a cooperative agreement with the National Science Foundation.
[^2]: The Parkes telescope is part of the Australia Telescope which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
K. Hagino,$^{1,2}$ and N. Takigawa,$^{2}$\
\
\
[*Seattle, Washington 98915* ]{}\
[*$^2$ Department of Physics, Tohoku University, Sendai 980–8578, Japan*]{}\
title: ' Effects of finite width of excited states on heavy-ion sub-barrier fusion reactions'
---
=22.5cm
[**Abstract**]{}
We discuss the effects of coupling of the relative motion to nuclear collective excitations which have a finite lifetime on heavy-ion fusion reactions at energies near and below the Coulomb barrier. Both spreading and escape widths are explicitly taken into account in the exit doorway model. The coupled-channels equations are numerically solved to show that the finite resonance width always hinders fusion cross sections at subbarrier energies irrespective of the relative importance between the spreading and the escape widths. We also show that the structure of fusion barrier distribution is smeared due to the spreading of the strength of the doorway state.
PACS number(s): 25.70.Jj, 24.10.Eq,24.30.Cz,25.70.Mn
[**I. INTRODUCTION**]{}
Extensive experimental as well as theoretical studies during the past two decades have led to a well-established idea that cross sections for heavy-ion fusion reactions are considerably enhanced at sub-barrier energies compared with predictions of the one dimensional potential model [@B88]. It has subsequently been concluded that these enhancements of fusion cross sections can be attributed to couplings between the relative motion of the colliding nuclei and several nuclear collective motions as well as transfer reactions [@BT98]. A standard way to address the effects of channel coupling on fusion cross sections is to numerically solve the coupled-channels equations. In coupled-channels calculations involving low-lying collective excitations of medium mass nuclei, which show the large enhancements of fusion cross sections, the excited states are usually assumed to have an infinite lifetime. However, when the excitation energies of the collective modes exceed the threshold energy for particle emission or the typical excitation energy of incoherent modes of excitations, like giant resonances in stable nuclei, they have finite lifetimes [@BB94]. Although the excitation energy of giant resonances in stable nuclei is in general so large that the effects of their excitations and thus their width can be well described by a static potential renormalisation [@THAB94; @HTDDL97], the effects of the finite width of excited states are expected to become important in discussing fusion reactions of weakly bound nuclei like $^{6,7}$Li, $^{9}$Be [@SYK83; @HS92; @RGMT97; @TMS97] or nuclei far from the stability lines[@YSF96; @HPCD92; @CDLH95; @TKS93; @DV94]
Some attempts to include the effects of finite width in coupled-channels calculations have been made recently by Hussein [*et al.*]{} [@HT94; @HPT95]. They used the exit doorway model [@BM69; @BT85] to discuss the effects of spreading width $\Gamma^{\downarrow}$ on fusion reactions. Instead of numerically solving the resultant coupled-channels equations as they are, they introduced the constant coupling approximation to diagonalise the coupling matrix[@DLW83] and claimed that the spreading width further enhances fusion cross sections compared with the case where the excited state has an infinite life time. As for the effects of escape width $\Gamma^{\uparrow}$, they took another model, i.e. a model which uses a dynamical polarization potential to account for the effects of the break-up reaction. They thus showed that the escape width strongly reduces fusion cross sections.
Although the results of Refs. [@HT94; @HPT95] are interesting, there are still some unsatisfactory aspects in their approach. First the constant coupling approximation used in Refs. [@HT94; @HPT95] does not provide satisfactory results in heavy-ion fusion reactions where the coupling extends outside the Coulomb barrier[@DL87; @DNA92; @HTB97]. Furthermore, the constant coupling approximation leads to complex values of eigen-energies as well as weight factors if one eliminates the internal degrees of freedom which couple to the doorway state. Refs. [@HT94; @HPT95], however, did not fully take this fact into consideration. The exact coupled-channels calculations are, therefore, urged in order to draw a definite conclusion on the effects of spreading width $\Gamma^{\downarrow}$ on heavy-ion fusion reactions. Secondly, the effects of escape width $\Gamma^{\uparrow}$ are not transparent in the polarization potential formalism used in Refs. [@HT94; @HPT95]. Also it will be hard to evaluate the polarization potential for each reaction. For example, in fusion reactions of $^9$Be, most of the states of $^9$Be which are excited during fuion will eventually decay into the $n + 2\alpha$ channel, because the separation energy of $\alpha$ particle is small in this nucleus. It is not so straight forward to derive a poralization potential in a reliable way for such four-body problems.
In this paper, we extend the model in Refs. [@HT94; @HPT95] to treat both spreading and escape widths on an equal footing. This enables us to see explicitly the effects of both widths on subbarrier fusion reactions and to easily discuss the case where both of them are present simultaneously. We then carry out exact coupled-channels calculations to investigate the effects of the finite width on heavy-ion fusion reactions. It will be shown that both spreading and escape widths reduce fusion cross sections, contrary to the conclusions in Refs. [@HT94; @HPT95]. We also discuss the effects of the finite width on the fusion barrier distribution defined as the second derivative of the product of the fusion cross section and the center of mass energy, $E \sigma$ [@RSS91; @LDH95].
The paper is organised as follows. In Sec. II, we briefly review the exit doorway model and derive coupled-channels equations which account for the effects of the finite width of excited states. We consider three cases, i) where only the spreading width exists, ii) only the escape width exists, and iii) both the spreading and escape widths are present simultaneously. In Sec. III, we present numerical results of the coupled-channels calculations for these three cases and discuss the effects of finite width on fusion cross sections as well as on fusion barrier distributions. The summary is given in Sec. IV. Finally in Appendix A, a time dependent approach to discuss the effects of escape width is presented.
[**II. COUPLING TO RESONANCE CHANNEL**]{}
[**A. Effect of spreading width**]{}
We first discuss the effects of the spreading width on sub-barrier fusion reactions. To this end, we use the exit doorway model [@HT94; @HPT95; @BM69; @BT85]. In this model, the relative motion between the colliding nuclei couples to many excited states only through the doorway state, i.e. the collective state. We assume the following Hamiltonian for the fusing system: $$H=
-\frac{\hbar^2}{2\mu}\frac{d^2}{dr^2}
+\frac{J(J+1)\hbar^2}{2\mu r^2}
+V_N(r)+\frac{Z_PZ_Te^2}{r}+H_{int}(\xi)+V_{coup}(r,\xi),$$ where $r$ is the coordinate of the relative motion between the projectile and the target, and $\mu$ is the reduced mass. $V_N$ is the bare nuclear potential, $Z_P$ and $Z_T$ are the atomic numbers of the projectile and the target, respectively. $H_{int}$ describes the intrinsic excitations in one of the colliding nuclei, and $V_{coup}$ the coupling between these excitations (generically denoted by $\xi$) and the relative motion. In writing Eq. (1), we used the no-Coriolis approximation and replaced the angular momentum operator for the relative motion by the total angular momentum $J$ [@TI86; @HTBB95]. Following Refs. [@HT94; @HPT95; @BT85], we assume that the intrinsic Hamiltonian $H_{int}$ and the coupling Hamiltonian $V_{coup}$ are given respectively by $$H_{int}= |d>E_d<d| + \sum_j|j>e_j<j|
+\sum_j\left[ |j> \Delta_j <d| + |d> \Delta^*_j <j| \right],$$ $$V_{coup}=f(r)\left(|0><d| + |d><0| \right),$$ where $|0>, |d>$, and $|j>$ denote the ground state, the doorway state, and the other intrinsic states, respectively. $E_d$ and $e_j$ are the energy of the doorway state and that of the state $|j>$, respectively. In our example in this subsection where we discuss the effects of spreading width, $|j>$ are uncorrelated $1p1h$ states or more complicated many particle many hole states within the same nucleus. The former yields the Landau damping, while the latter the spreading width [@BB94]. In Eqs. (2) and (3), $\Delta$ and $f(r)$ are the coupling strength between the doorway state and $|j>$, and between the doorway state and the ground state, respectively. Inherently, the former is independent of $r$. For simplicitly we have assumed that the doorway state $|d>$ linearly couples to the ground state $|0>$. The extension to the case where there exist higher order couplings [@HTD97] is straight forward.
The intrinsic Hamiltonian $H_{int}$ can be diagonalised by introducing the normal modes $|\varphi_i>$ by $$\begin{aligned}
H_{int}&=&\sum_i |\varphi_i> E_i <\varphi_i|, \\
|\varphi_i>&=&\alpha_i |d> + \sum_j \beta_{ij} |j>.\end{aligned}$$ When the states $|j>$ are distributed with equal energy spacing from $-\infty$ to $\infty$, and the coupling strengths $\Delta_j$ are independent of $j$, i.e., $$\begin{aligned}
e_j&=&jD~~~~~(j=0,\pm 1, \pm 2, \cdots), \\
\Delta_j &=& \kappa,\end{aligned}$$ $E_i$ and $\alpha_i$ in Eqs. (4) and (5) are given by $$\begin{aligned}
E_i&=& E_d + \frac{\pi\kappa^2}{D}\cot\frac{\pi E_i}{D}, \\
\left|\alpha_i\right|^2&=& \frac{1}{\frac{2\pi}{D}}
\frac{\Gamma^{\downarrow}}{(E_i-E_d)^2
+\frac{\Gamma^{\downarrow 2}}{4}},\end{aligned}$$ respectively (the Breit-Wigner distribution) [@BM69]. Here $\Gamma^{\downarrow}$ is the spreading width, including the Landau width, and is defined by $\Gamma^{\downarrow}=2\pi\kappa^2/D$. In obtaining Eq. (9), we assumed that the coupling strength $\kappa$ is much larger than the energy spacing $D$ and neglected a term of the order of $D^2/\kappa^2$ in the denominator.
Expanding the total wave function with the eigen-states $|\varphi_i>$ as $$\Psi(r,\xi)=\frac{u_0(r)}{r}|0> + \sum_i\frac{u_i(r)}{r}|\varphi_i>,$$ the coupled-channels equations read $$\begin{aligned}
&&\left[
-\frac{\hbar^2}{2\mu}\frac{d^2}{dr^2}
+\frac{J(J+1)\hbar^2}{2\mu r^2}
+V_N(r)+\frac{Z_PZ_Te^2}{r}-E\right]u_0(r)
+\sum_j\alpha_j f(r) u_j(r)=0, \\
&&\left[
-\frac{\hbar^2}{2\mu}\frac{d^2}{dr^2}
+\frac{J(J+1)\hbar^2}{2\mu r^2}
+V_N(r)+\frac{Z_PZ_Te^2}{r}-E+E_j\right]u_j(r)
+\alpha^*_j f(r) u_0(r)=0. \nonumber \\\end{aligned}$$ These equations may be solved by imposing the incoming wave boundary condition inside the Coulomb barrier, i.e. $$\begin{aligned}
u_i(r)&=& T_i \exp\left(-i \int^r_{r_{abs}} k_i(r')dr'\right)
~~~~~~~~~~~r\leq r_{abs}, \\
&=&H_J^{(-)}(k_ir)\delta_{i,0} + R_i H_J^{(+)}(k_ir)
~~~~~~~~r\to\infty,\end{aligned}$$ where $k_i(r)$ is the local wave length for the $i$-th channel, and $r_{abs}$ is the absorption radius where the incoming wave boundary condition is imposed. $H_J^{(-)}$ and $H_J^{(+)}$ are the incoming and the outgoing Coulomb waves, respectively. The fusion cross section is then obtained as $$\sigma(E)=\frac{\pi}{k^2}\sum_J(2J+1)
\left(
\frac{k_0(r_{abs})}{k}
|T_0|^2 + \sum_i \frac{k_i(r_{abs})}{k}|T_i|^2\right),$$ where $k=k_0(\infty)$ is the wave length of the entrance channel at $r\to\infty$. Note that in contrast to the works by Hussein [*et al.*]{} which eliminate the intrinsic degree of freedom $|j>$ and introduce the complex $Q$-value [@HT94; @HPT95], we treat it explicitly in the coupled-channels calculations.
[**B. Effect of escape width**]{}
We next consider the effects of escape width. In this case, the states $|j>$ represent particle continuum states. The coupled-channels equations to be solved are exactly the same as Eqs. (11) and (12). One may imagine the box normalisation of these states in order to match with the discretisation of the continuum states $|j>$. The difference appears at the final stage. We define the complete fusion as such a process, where the whole projectile is absorbed by the whole target without emitting any particle prior to the fusion. Therefore, the particle continuum states $|j>$ have to be excluded from the final states in obtaining the cross section for the complete fusion. Since the probability to find a state $|j>$ in the state for the $i$-th normal mode $|\varphi_i>$ is given by $|\beta_{ij}|^2$, the complete fusion cross section is given by $$\sigma(E)=\frac{\pi}{k^2}\sum_J(2J+1)
\left(\frac{k_0(r_{abs})}{k}|T_0|^2 + \sum_i |\alpha_i|^2
\frac{k_i(r_{abs})}{k}|T_i|^2\right).$$ In deriving Eq. (16), we used the normalisation condition of the normal states, i.e. $|\alpha_i|^2=1-\sum_j|\beta_{ij}|^2$.
Another approach to fusion reactions in the presence of a break-up channel is to evaluate the loss of flux during fusion [@HPCD92; @CDLH95; @TKS93]. The relation between such approach and the present formalism is given in Appendix by using a time dependent theory.
[**C. Interplay between spreading and escape widths**]{}
Lastly we consider the case where both spreading and escape widths are present simultaneously and interplay with each other. In this case, the intrinsic Hamiltonian is given by $$\begin{aligned}
H_{int} &=& |d>E_d<d|
+ \sum_j|j^{\uparrow}>e^{\uparrow}_j<j^{\uparrow}|
+ \sum_j|j^{\downarrow}>e^{\downarrow}_j<j^{\downarrow}| \nonumber \\
&&+\sum_j\left[ |j^{\uparrow}> \Delta^{\uparrow}_j <d|
+ |d> \Delta^{\uparrow *}_j <j^{\uparrow}| \right]
+\sum_j\left[ |j^{\downarrow}> \Delta^{\downarrow}_j <d|
+ |d> \Delta^{\downarrow *}_j <j^{\downarrow}| \right], \nonumber \\\end{aligned}$$ where $|j^{\uparrow}>$ and $|j^{\downarrow}>$ represent particle continuum states and complicated particle hole bound states, respectively. We diagonalise this Hamiltonian by introducing the normal states defined by $$|\varphi_i>=\alpha_i |d> + \sum_j \beta^{\uparrow}_{ij} |j^{\uparrow}>
+\sum_j \beta^{\downarrow}_{ij} |j^{\downarrow}>.$$ As in Sec. II.A, if we assume a uniformely spaced sequence of energies $e_j^{\uparrow}$ and $e_j^{\downarrow}$ and state independent coupling strengths $\Delta_j^{\uparrow}$ and $\Delta_j^{\downarrow}$, the eigen-values $E_i$ and the doorway amplitude $\alpha_i$ are given by $$\begin{aligned}
E_i&=& E_d + \frac{\pi(\kappa^{\uparrow 2}+\kappa^{\downarrow 2})}{D}
\cot\frac{\pi E_i}{D}, \\
\left|\alpha_i\right|^2&=& \frac{1}{\frac{2\pi}{D}}
\frac{\Gamma^{\uparrow}+\Gamma^{\downarrow}}
{(E_i-E_d)^2+\frac{(\Gamma^{\uparrow}+\Gamma^{\downarrow})^2}{4}},\end{aligned}$$ respectively. We have assumed that the energy spacing of the particle continuum states $|j^{\uparrow}>$ is the same as that of the particle-hole bound states $|j^{\downarrow}>$. Here $\Gamma^{\uparrow}$ and $\Gamma^{\downarrow}$ are escape and spreading widths given by $2\pi\kappa^{\uparrow 2}/D$ and $2\pi\kappa^{\downarrow 2}/D$, respectively. Excluding the particle continuum states $|j^{\uparrow}>$ from the final states, the complete fusion cross section is given by $$\sigma(E)=\frac{\pi}{k^2}\sum_J(2J+1)
\left[\frac{k_0(r_{abs})}{k}|T_0|^2 + \sum_i \left(|\alpha_i|^2+
\frac{\Gamma^{\downarrow}}{\Gamma^{\uparrow}+\Gamma^{\downarrow}}
\right)
\frac{k_i(r_{abs})}{k}|T_i|^2\right].$$ In deriving Eq. (21), we used the identity $\sum_j|\beta_{ij}^{\downarrow}|^2=
\Gamma^{\downarrow}/(\Gamma^{\uparrow}+\Gamma^{\downarrow})$. The complete fusion cross section is thus initimately related to the ratio $\Gamma^{\downarrow}/(\Gamma^{\uparrow}+\Gamma^{\downarrow})$.
[**III. EFFECTS OF FINITE WIDTH ON FUSION CROSS SECTIONS AND BARRIER DISTRIBUTIONS**]{}
We now present the results of our calculations of fusion cross sections and fusion barrier distributions. We consider fusion reactions between $^{16}$O and $^{144}$Sm in the presence of low-lying octupole phonon excitations in the latter nucleus. We artificially set its excitation energy $E_d$ to be 2 MeV and assume that it has a hypothetical total width of 1 MeV. Our conclusions do not depend so much on the particular choice for these parameters. We use the collective model for the coupling form factor $f(r)$, i.e. $$f(r)=\frac{\beta_3}{\sqrt{4\pi}}\left(-R_T\frac{dV_N}{dr}
+\frac{3}{2\lambda+1}Z_PZ_T\frac{R_T^{\lambda}}{r^{\lambda+1}}
\right),$$ where $\lambda=3$ is the multipolarity of the excitation and $R_T$ is the radius of $^{144}$Sm. $\beta_3$ is the deformation parameter of the phonon excitation, which was chosen to be 0.205 with the target radius of $R_T=1.06 A^{1/3}$ fm. We used the same nuclear potential $V_N$ as that in Refs. [@HTDDL97; @HTD97], i.e. a Woods-Saxon potential whose depth, range parameter and surface diffuseness are $V$=105.1 MeV, $r_0$=1.1 fm, and $a$=0.75 fm, respectively. In the actual calculations, we introduced a cut-off energy and a finite energy spacing for the excited states by considering the normal states between $E_d - 1$ and $E_d + 1$ MeV with the energy spacing $D$ of 0.2 MeV. We have checked that our conclusions do not qualitatively alter when the cut-off energy is taken to be larger and/or the energy spacing smaller.
We first discuss the effects of spreading width. The upper panel of Fig. 1 shows the fusion cross section of this system. The solid line was obtained by numerically solving the coupled-channels equations Eqs. (11) and (12) with $\Gamma^{\downarrow}$=1 MeV, while the dashed line is the result when the doorway state has an infinite lifetime. The figure also contains the result for the no coupling case (the dotted line) for comparison. One can see that the spreading width slightly reduces the fusion cross section, though it is still enhanced compared with the case of no coupling. Our result contradicts the conclusion in Refs. [@HT94; @HPT95] where it was claimed that the spreading width enhances the fusion cross section. This discrepancy could be associated with the constant coupling model and/or the incorrect treatments of complex potentials and weight factors in Refs. [@HT94; @HPT95].
The fact that the spreading width reduces the enhancement of fusion cross section can be understood in the following way. After eliminating the intrinsic states $|j>$, the coupled-channels problem given by Eqs. (11) and (12) reduces to the two channel problem with the coupling matrix [@HT94; @HPT95] (see also Appendix) $$\left(
\begin{array}{cc}
0&f(r)\\ f(r)&E_d-i\frac{\Gamma}{2}
\end{array}
\right).$$ If we diagonalise this coupling matrix with a bi-orthogonal basis, the lower potential barrier is given by [@HT94; @HPT95] $$V_-(r)=V_N(r)+\frac{Z_PZ_Te^2}{r}
+\frac{1}{2}\left(E_d-i\frac{\Gamma^{\downarrow}}{2}
-\sqrt{E_d^2-\frac{\Gamma^{\downarrow 2}}{4}
+4f(r)^2-iE_d\Gamma}\right).$$ The real part of the lower barrier thus always increases when the width $\Gamma$ is non-zero, leading to smaller penetrabilities at energies below the barrier. When the doorway energy $E_d$ is much larger than the coupling form factor $f(r)$, Eq. (24) is transformed to $$V_-(r)=V_N(r)+\frac{Z_PZ_Te^2}{r}
-f(r)^2\frac{E_d+i\frac{\Gamma^{\downarrow}}{2}}
{E_d^2+\frac{\Gamma^{\downarrow 2}}{4}},$$ which is identical to the adiabatic barrier derived from Eq. (B.21) in Ref. [@BT85].
One might expect that fusion cross section is enhanced if the spreading width is finite, since the doorway state then couples to the environmental background whose energy is lower than that of the doorway state itself and thus the effective excitation energy of the doorway state becomes lower. However, this intuitive picture does not hold if there is non-negligible coupling $f(r)$ around the barrier position, which modifies the potential barrier according to Eq. (24). Since the fusion cross section is much more sensitive to the barrier height at energies well below the Coulomb barrier than the energy transfer which takes place before the projectile hits the Coulomb barrier, the net effects of the spreading width cause the hindrance of the fusion cross section. A similar importance of the potential renormalisation concerning the effects of transfer reaction with positive $Q$-value on subbarrier fusion reactions has been pointed out in Ref. [@L84].
This contrasts to the situation where the doorway state couples to a small number of intrinsic states $|j>$. In such cases, the coupling could further enhance the fusion cross section, as is usually the case in double phonon couplings [@HKT98]. When the doorway state couples to a large number of surrounding states, as is discussed in this paper, the relaxation time becomes very short and consequently the couplings begin to reveal a dissipative character [@BT85]. The hindrance of fusion cross sections due to the spreading width thus resembles the dissipative quantum tunneling, which has been an extremely popular subject during the past two decades in many fields of physics and chemistry[@CL81; @JJAP93].
The above conclusions have been reached by assuming the Breit-Wigner distribution given by Eq. (9) for the non-collective states $|j>$. In order to test the sensitivity to the property of the distribution, we repeated the calculations by assuming the Lorentzian distribution (not shown) and obtained the similar conclusions concerning the role of the finite resonance width in heavy-ion fusion reactions. The difference is negligible especially when the doorway energy $E_d$ is larger than the width $\Gamma$.
The lower panel of Fig. 1 shows fusion barrier distributions for this system. This quantity is defined as the second derivative of $E\sigma$ with respect to the energy $E$ [@RSS91], and has been experimentally shown to be very sensitive to the nuclear structure of the colliding nuclei[@LDH95]. We used the 3-point differece formula with an energy spacing of 1.8 MeV to obtain the second derivative from the fusion cross sections[@LDH95]. The meaning of each line is the same as that in the upper panel. When there exists no coupling between the ground and the doorway states, the barrier distribution has only a single peak, corresponding to a single potential barrier, i.e. the bare potential barrier. If the coupling is turned on, the single barrier splits to two and the barrier distribution has two peaks (the dashed line). This double peaked structure is somewhat smeared when the width of the doorway state is finite, since it re-distributes the strength of the doorway state (the solid line). As a consequence, the higher peak of the fusion barrier distribution becomes less apparent.
We next discuss the effects of escape width. The solid line in Fig. 2 was obtained by setting $\Gamma^{\uparrow}$=1 MeV and $\Gamma^{\downarrow}$=0 MeV, and using Eq. (16). The dotted and the dashed lines are the same as in Fig. 1. As will be discussed in Appendix, the escape width is intimately related to a loss of flux due to the break-up reaction, and strongly hinders the fusion cross section over a wide range of bombarding energies. The fusion cross section is smaller even than that in the absense of the channel couplings at energies above the Coulomb barrier. The escape width also lowers the height of the main peak of the fusion barrier distribution and at the same time broadens the fusion barrier distribution.
When there exist both spreading and escape widths simultaneously, one expects to have a situation intermediate between depicted in Figs. 1 and 2. In order to demonstrate this, Fig. 3 shows the result when the both widths are set to 0.5 MeV. Since both spreading and escape widths always reduce the fusion cross section, they are smaller than those in the infinite lifetime case over the entire energy range shown in Fig. 3. One finds that the degree of hindrance is intermediate between Figs. 1 and 2, as expected.
[**IV. SUMMARY**]{}
We have derived coupled-channels equations which take into account the effects of finite width of an excited state. This formalism treats the spreading and the escape widths on an equal footing, and thus enables one to discuss an interplay between them. To this end, we used the exit doorway model. Numerical solutions of the coupled-channels equations showed that both widths hinder the fusion cross section. The degree of hindrance is moderate when the spreading width dominates the total width, while the fusion cross section is considerably reduced in the opposite case, i.e. when the escape width dominates the total width. We also investigated the effects of finite width on the fusion barrier distribution. We demonstrated that the spreading width smears the structure of the fusion barrier distribution and also that the escape width lowers the height of the main peak of the fusion barrier distribution. These considerations are important when one analyses high precision measurements of the fusion reactions of fragile nuclei like $^{6,7}$Li or $^{9}$Be, which have been undertaken recently, or when one discusses fusion reactions of unstable nuclei. We will report analyses of these experimental data in a separate paper.
[**ACKNOWLEDGMENTS**]{}
The authors thank M. Dasgupta, D.J. Hinde, and C.R. Morton for useful discussions, and the Australian National University for its hospitality and for partial support for this project. The work of K.H. was supported by the Japan Society for the Promotion of Science for Young Scientists. This work was also supported by the Grant-in-Aid for General Scientific Research, Contract No.08640380, Monbusho International Scientific Research Program: Joint Research, Contract No. 09044051, from the Japanese Ministry of Education, Science and Culture.
[**APPENDIX A: TIME DEPENDENT APPROACH OF THE EFFECT OF ESCAPE WIDTH**]{}
In this appendix, we discuss the relation between the coupled-channels formalism discussed in Sec. II.B and an approach using the flux loss during fusion due to a break-up reaction. To this end, we use a time dependent approach. Assuming that the total wave function at time $t$ is given by $$|\Psi (t)>=a_0(t)|0> + \sum_i a_i(t)e^{-iE_it/\hbar}|\varphi_i>,$$ the time dependent coupled equations which corresponds to Eqs. (11) and (12) read [@CRHT94; @BCHT96] $$\begin{aligned}
\dot{a}_0(t)&=&\sum_i\frac{1}{i\hbar}\alpha_i f(r(t))e^{-iE_it/\hbar}
a_i(t), \\
\dot{a}_i(t)&=&\frac{1}{i\hbar}\alpha^*_i f(r(t))e^{iE_it/\hbar}
a_0(t). \end{aligned}$$ If we assume the Breit-Wigner distribution Eq. (9) for the incoherent states $|j>$, these two equations can be combined to give [@CRHT94; @BCHT96] $$\dot{a}_0(t)=-\frac{f(r(t))}{\hbar^2}\int^t_{-\infty}
dt'e^{-i(E_d-i\frac{\Gamma}{2})(t-t')/\hbar}a_0(t')
f(r(t')).$$ In deriving this equation, we have assumed that the energy spacing of the states $|j>$ is small enough and replaced the summation over the normal states $|\varphi_i>$ with an integration over the energy of these states $E_i$.
Here we introduce the doorway amplitude by $$a_d(t)=\frac{1}{i\hbar}\int^t_{-\infty}
dt'e^{-i(E_d-i\frac{\Gamma}{2})(t-t')/\hbar}a_0(t')
f(r(t')).$$ Eq. (29) can then be written in a form of two coupled equations as $$\begin{aligned}
\dot{a}_d(t)&=&\frac{1}{i\hbar}\left(E_d-i\frac{\Gamma}{2}\right)a_d(t)
+\frac{1}{i\hbar}f(r(t))a_0(t), \\
\dot{a}_0(t)&=&\frac{1}{i\hbar}a_d(t)f(r(t)).\end{aligned}$$ These equations provide a two level problem with the coupling matrix given by Eq. (23) [@HT94; @HPT95].
When the escape width dominates the total width, i.e. $\Gamma\sim\Gamma^{\uparrow}$, the survival probability of the system at time $t$ is given by $$P_s(t)=\left|<0|\Psi(t)>\right|^2 + \left|<d|\Psi(t)>\right|^2.$$ Noticing that $<d|\Psi(t)>$ is nothing but the doorway amplitude $a_d(t)$, one can easily show that the time dependence of the survival probability is given by $$\frac{d}{dt}P_s(t)=-\frac{\Gamma^{\uparrow}}{\hbar}
\left|a_d(t)\right|^2.$$ The survival probability $P_s(t)$ is thus a decreasing function of time $t$, and $1-P_s(t)$ represents the probability of the flux loss caused by the particle emission.
[99]{}
M. Beckerman, Rep. Prog. Phys. [**51**]{}, 1047(1988).
A.B. Balantekin and N. Takigawa, Rev. of Mod. Phys., [**70**]{}, 77(1998).
G.F. Bertsch and R.A. Broglia, [*Oscillations in finite quantum systems*]{}, (Cambridge university press, Cambridge, 1994); G.F. Bertsch, P.F. Bortignon, and R.A. Broglia, Rev. Mod. Phys. [**55**]{}, 287 (1983).
N. Takigawa, K. Hagino, M. Abe, and A.B. Balantekin, Phys. Rev. C[**49**]{}, 2630(1994).
K. Hagino, N. Takigawa, M. Dasgupta, D.J. Hinde, and J.R. Leigh, Phys. Rev. Lett. [**79**]{}, 2014(1997).
Y. Sakuragi, M. Yahiro, and M. Kamimura, Prog. Theo. Phys. [**70**]{}, 1047(1983).
Y. Hirabayashi and Y. Sakuragi, Phys. Rev. Lett. [**69**]{}, 1892(1992).
K. Rusek, J. Gómez-Camacho, I. Martel-Bravo, and G. Tungate, Nucl. Phys. [**A614**]{}, 112(1997).
J. Takahashi, M. Munhoz, E.M. Szanto, N. Carlin, N. Added, A.A.P. Suaide, M.M. de Moura, R. Liguori Neto, A. Szanto de Toledo, and L.F. Canto, Phys. Rev. Lett. [**78**]{}, 30(1997).
A. Yoshida, C. Signorini, T. Fukuda, Y. Watanabe, N. Aoi, M. Hirai, M. Ishihara, H. Kobinata, Y. Mizoi, L. Mueller, Y. Nagashima, J. Kaneko, T. Nomura, Y.H. Pu, and F. Scarlassara, Phys. Lett. [**B389**]{}, 457(1996).
M.S. Hussein, M.P. Pato, L.F. Canto, and R. Donangelo, Phys. Rev. C[**46**]{}, 377(1992).
L.F. Canto, R. Donangelo, P. Lotti, and M.S. Hussein, Phys. Rev. C[**52**]{}, R2848(1995).
N. Takigawa, M. Kuratani, and H. Sagawa, Phys. Rev. C[**47**]{}, R2470(1993).
C.H. Dasso and A. Vitturi, Phys. Rev. C[**50**]{}, R12(1994).
M.S. Hussein and A.F.R. de Toledo Piza, Phys. Rev. Lett. [**72**]{}, 2693(1994).
M.S. Hussein, M.P. Pato, and A.F.R. de Toledo Piza, Phys. Rev. C[**51**]{}, 846(1995).
A. Bohr and B.R. Mottelson, [*Nuclear Structure I*]{}, (Benjamin, New York, 1969), p. 302.
A.B. Balantekin and N. Takigawa, Ann. Phys. (N.Y.) [**160**]{}, 441(1985).
C.H. Dasso, S. Landowne, and A. Winther, Nucl. Phys. [**A407**]{}, 221(1983).
C.H. Dasso and S. Landowne, Phys. Lett. [**B183**]{}, 141(1987).
M. Dasgupta, A. Navin, Y.K. Agarwal, C.V.K. Baba, H.C. Jain, M.L. Jhingan, and A. Roy, Nucl. Phys. [**A539**]{}, 351(1992).
K. Hagino, N. Takigawa, and A.B. Balantekin, Phys. Rev. C[**56**]{}, 2104(1997).
N. Rowley, G.R. Satchler, and P.H. Stelson, Phys. Lett. [**B254**]{}, 25(1991).
J.R. Leigh, M. Dasgupta, D.J. Hinde, J.C. Mein, C.R. Morton, R.C. Lemmom, J.P. Lestone, J.O. Newton, H. Timmers, J.X. Wei, and N. Rowley, Phys. Rev. C[**52**]{}, 3151(1995).
N. Takigawa and K. Ikeda, in [*Proceedings of the Symposium on Many Facets of Heavy Ion Fusion Reactions*]{}, edited by W. Henning [*et al.*]{}(Argonne National Laboratory Report No. ANL-PHY-87-1), 1986, p.613.
K. Hagino, N. Takigawa, A.B. Balantekin, and J.R. Bennett, Phys. Rev. C[**52**]{}, 286 (1995).
K. Hagino, N. Takigawa, M. Dasgupta, D.J. Hinde, and J.R. Leigh, Phys. Rev. C[**55**]{}, 276 (1997).
S.Y. Lee, Phys. Rev. C[**29**]{}, 1932(1984).
K. Hagino, S. Kuyucak, and N. Takigawa, Phys. Rev. C[**57**]{}, 1349 (1998).
A.O. Caldeira and A.J. Leggett, Phys. Rev. Lett., [**46**]{}, 211(1981).
edited by M. Tsukada [*et al.,*]{} Japanese Journal of Applied Physics Series Vol. 9 ( Publication Office of Japanese Journal of Applied Physics, Tokyo, 1993).
L.F. Canto, A. Romanelli, M.S. Hussein, and A.F.R. de Toledo Piza, Phys. Rev. Lett. [**72**]{}, 2147(1994).
C.A. Bertulani, L.F. Canto, M.S. Hussein, and A.F.R. de Toledo Piza, Phys. Rev. C[**53**]{}, 334(1996).
[**Figure Captions**]{}
[**Fig. 1:**]{} Effects of the spreading width on the fusion cross section (the upper panel) and fusion barrier distribution (the lower panel). The dotted line is for the case of no coupling. The solid line takes into account the effects of coupling of the relative motion to a doorway state at 2 MeV with the width 1 MeV, while the dashed line assumes that the doorway state has an infinite lifetime.
[**Fig. 2:**]{} Same as fig.1, but for the escape width.
[**Fig. 3:**]{} Same as fig.1, but in the simultaneous presence of spreading and escape widths. Both widths are assumed to be 0.5 MeV.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Sensor-based human activity recognition (HAR) requires to predict the action of a person based on sensor-generated time series data. HAR has attracted major interest in the past few years, thanks to the large number of applications enabled by modern ubiquitous computing devices. While several techniques based on hand-crafted feature engineering have been proposed, the current state-of-the-art is represented by deep learning architectures that automatically obtain high level representations and that use recurrent neural networks (RNNs) to extract temporal dependencies in the input. RNNs have several limitations, in particular in dealing with long-term dependencies. We propose a novel deep learning framework, [<span style="font-variant:small-caps;">TrASenD</span>]{}, based on a purely attention-based mechanism, that overcomes the limitations of the state-of-the-art. We show that our proposed attention-based architecture is considerably more powerful than previous approaches, with an average increment, of more than $7\%$ on the F1 score over the previous best performing model. Furthermore, we consider the problem of personalizing HAR deep learning models, which is of great importance in several applications. We propose a simple and effective transfer-learning based strategy to adapt a model to a specific user, providing an average increment of $6\%$ on the F1 score on the predictions for that user. Our extensive experimental evaluation proves the significantly superior capabilities of our proposed framework over the current state-of-the-art and the effectiveness of our user adaptation technique.
Code is available at: <https://github.com/DavideBuffelli/TrASenD>
author:
- |
Davide Buffelli\
Department of Information Engineering\
University of Padova\
Padova, Italy\
`[email protected]`\
Fabio Vandin\
Department of Information Engineering\
University of Padova\
Padova, Italy\
`[email protected]`\
bibliography:
- 'scibib.bib'
title: 'Attention-Based Deep Learning Framework for Human Activity Recognition with User Adaptation'
---
Introduction
============
Sensor-based human activity recognition (HAR) is a time series classification task that involves predicting the movement or action of a person (e.g. walking, running, etc.) based on sensor data. HAR has many practical applications, such as fitness tracking, video surveillance, and gesture recognition. Despite being a well studied and mature problem, HAR has been a very active research area in recent years, due to the rise of ubiquitous computing enabled by smartphones, wearables, and Internet-of-Things devices [@NWEKE2018233; @94ca4272fbb244418572d2d29c6ef6ae; @Bianchi_2019].
Several previously proposed approaches tackled the problem by hand-crafting features [@Figo2010; @Stisen_2015]. These kind of approaches, based on trial-and-error, require a lot of human effort, and therefore time, and are not guaranteed to generalize well to unseen subjects. Deep learning enables automatic feature extraction and can hierarchically compose features to obtain high level representations, which have more discriminative power than handcrafted features based on human expertise. These properties allow deep learning models to be more robust and with higher generalization properties, and make deep learning the state-of-the-art technique for HAR [@NWEKE2018233; @Wang_2019]. In particular, the state-of-the-art is given by the DeepSense framework [@Yao_2017], with an architecture based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs).
While RNNs have been used in several domains to capture sequential relationships, they still have some shortcomings, in particular in terms of learning from long input sequences [@Hochreiter:01book; @Cho_2014]. An attractive strategy to enhance, or replace, RNNs is provided by *attention models*, first introduced in encoder-decoder neural networks in the context of natural language processing (NLP) [@Bahdanau2015NeuralMT]. The main idea behind attention mechanisms is to act as a memory-access mechanism that allows the decoder to selectively access the most important parts of the input sequence based on the current context. Attention models alleviate RNNs difficulties in learning from long input sequences, and successive developments have led to NLP models based solely on attention mechanisms [@Vaswani2017AttentionIA]. To the best of our knowledge, the use of *pure* attention models in deep learning architectures to extract temporal dependencies in multimodal data, such as multi-sensor HAR data, has not been explored.
The human activity recognition task is highly “personal”, in the sense that a single smartphone or smartwatch is usually used by just one person, and the style of walking, running or climbing stairs is peculiar to each individual. It is then natural to aim at the development of deep learning techniques that can be adapted to a specific user. However, the exploration of personalized deep learning models for HAR has been hitherto ignored.
Our Contribution
----------------
We expand the deep learning approaches for HAR with a new purely attention-based framework, [<span style="font-variant:small-caps;">TrASenD</span>]{}, that builds upon the state-of-the-art while significantly outperforming it on three different HAR datasets. [<span style="font-variant:small-caps;">TrASenD</span>]{} builds on the observation that RNNs do not provide the best way to capture the temporal relationships in the data, and significantly outperforms the state-of-the-art deep learning method for HAR. We also consider other variants of DeepSense, designed by replacing RNNs with more powerful attention enhanced RNNs mechanisms to capture temporal dependencies, and we show that while they do perform better than DeepSense, they are still less performing than our purely attention based [<span style="font-variant:small-caps;">TrASenD</span>]{}. In addition, we propose a personalization framework to adapt the model to a specific user over time, increasing the accuracy of the predictions for the user. To achieve this result we use a lightweight *transfer learning* approach that continues the training of only a small portion of the model with data acquired from the user. We empirically show that this approach significantly improves the performance of the model on data from a specific user.
Our contributions can be summarized as follows:
- We make use of a purely attention-based mechanism to develop a novel deep learning framework, [<span style="font-variant:small-caps;">TrASenD</span>]{}, for multimodal temporal data. [<span style="font-variant:small-caps;">TrASenD</span>]{} builds on the state-of-the-art frameworks [@Yao_2017; @Yao_2019] and significantly enhances them by replacing RNNs with a pure attention architecture inspired by the Transformer [@Vaswani2017AttentionIA].
- We extensively evaluate [<span style="font-variant:small-caps;">TrASenD</span>]{} against the current state-of-the-art and some of its variants that we design. We show that [<span style="font-variant:small-caps;">TrASenD</span>]{} significantly outperforms other methods on 3 different HAR datasets, with an average increment of more than $7\%$ on the F1 score over the previous best performing model. We also test the impact of data augmentation on the performances of the models, showing that it does play an important role on the generalization capabilities of the models.
- We propose a new transfer learning technique to adapt a model to specific user, in order to exploit the “personal” nature of the HAR task. We obtain an average increment of $6\%$ on the F1 score on the predictions for that user.
- We empirically prove the effectiveness of our personalization technique. In fact, we show that it is capable of significantly improving the performance of every model we analyze, on each dataset. Furthermore, we empirically demonstrate the learning capabilities of the proposed transfer learning approach.
Related Work
============
We now present the previous work related to our contributions. We first present the most relevant work on deep learning approaches for HAR (Section \[dlhar\]), then on attention mechanisms (Section \[attention\]), and finally on transfer learning and personalization in HAR (Section \[tl\_and\_pers\]).
Deep Learning for HAR {#dlhar}
---------------------
Following the taxonomy defined in recent surveys [@NWEKE2018233; @Wang_2019], deep learning techniques for sensor-based HAR fall into three main categories. The first category includes architectures composed of RNNs only [e.g. @Guan_2017; @Inoue_2017]. The second category includes architectures based on CNNs only, and can be further divided in two subcategories of models: *Data Driven* and *Model Driven* [@Wang_2019]. Data Driven models [e.g. @Hammerla2016DeepCA; @Sathyanarayana2016ImpactOP; @Pourbabaee_2018] use CNNs directly on the raw data coming from the sensors (each dimension of the data is seen as a channel). On the other hand, Model Driven approaches [e.g. @Li_2016; @Ravi_2016; @Singh_2017] first preprocess the data to get a grid-like structure, and then use CNNs. The third category is represented by those models that use both CNNs and RNNs [@Ord_ez_2016; @Singh_2017; @Yao_2017; @Yao_2019; @Ma_2019]. Finally, other deep learning techniques used for HAR are autoencoders [@Wang_2016; @almaslukh2017effective], and Restricted Boltzmann Machines [@Hammerla2015PDDS; @Li_2016_2; @Radu_2016]. DeepSense [@Yao_2017] is a deep learning framework for HAR that belongs to the third category, and constitutes the state-of-the-art for HAR. DeepSense starts with a CNN to extract features from intervals of data obtained from different sensors. The features extracted from different sensors in a time interval are then concatenated and other convolutional layers are used to extract higher level multi-sensor features. Successively a RNN layer (in particular, a stacked GRU layer) is used to learn temporal dependencies between the features extracted at different time intervals. A final layer is then easily customizable to adapt the framework for classification, regression or segmentation tasks.
The authors of DeepSense recently proposed a new version of the framework, SADeepSense [@Yao_2019], where they introduce a *self-attention* mechanism that automatically balances the contributions of multiple sensor inputs. SADeepSense maintains the same architecture of the original DeepSense framework, and adds an attention module to balance the contribution of different sensors based on their sensing quality. Additionally, in the RNN layer, another attention module is used to selectively attend to the most meaningful timesteps. This approach differs significantly from ours as the self-attention module of SADeepSense is used to address the issue of heterogeneity in the sensing quality from multiple sensors, and to select the most relevant timesteps for the final prediction, while [<span style="font-variant:small-caps;">TrASenD</span>]{} employs a purely attention-based mechanism directly as a mean to extract temporal dependencies in the data. Furthermore, SADeepSense retains the stacked GRU layer of the original DeepSense framework, while our approach replaces the GRU layer entirely. Another recently proposed architecture based on the DeepSense framework, which adopts a similar attention strategy to SADeepSense is AttnSense [@Ma_2019].
Attention Models {#attention}
----------------
Attention models were first introduced in encoder-decoder neural networks in the context of natural language processing (NLP) [@Bahdanau2015NeuralMT]. The main idea behind attention mechanisms is to allow the decoder to selectively access the most important parts of the input sequence based on the current context. This technique serves as a memory-access mechanism, and overcomes RNNs difficulties in learning from long input sequences. Attention has then been used for image captioning in an architecture that made use of both CNNs and RNNs [@pmlr-v37-xuc15]. Since then, attention models have become very popular in the deep learning community as an effective and powerful tool to enhance the capabilities of RNNs [e.g. @luong-etal-2015-effective; @Chaudhari2019AnAS; @Toshevska_2019]. Furthermore, Vaswani et al. ([-@Vaswani2017AttentionIA]) introduced the Transformer architecture, which is the current state-of-the-art for NLP, and completely removes RNNs with an attention-only mechanism to model temporal relationships.
In HAR, attention models have only been used in addition to a RNN (as described in Section \[dlhar\]), and not as a mean to directly capture temporal dependencies, which is the approach we propose in [<span style="font-variant:small-caps;">TrASenD</span>]{}.
Transfer Learning and Personalization in HAR {#tl_and_pers}
--------------------------------------------
Transfer learning is not new to HAR. In particular transfer learning has been leveraged to compensate for the amount of labeled data when training a model for activity recognition in different environments/circumstances [@Lopes_2011; @Cook_2013].
A previous (non-deep learning) transfer learning approach for personalized HAR, was proposed by Saeedi et al. ([-@Saeedi_2018]), and used the Locally Linear Embedding (LLE) algorithm to construct activity manifolds, which are used to assign labels to unlabeled data that can be used to develop a personalized model for the target user. Other different approaches to personalized HAR have been made with *incremental learning* [@inproceedings] on some classifiers that however were not based on deep learning, and with *Hidden Unit Contributions* [@Matsui_2017], a small layer inserted in between CNNs and learned from user data. In our approach we use transfer learning to train a small portion of the neural network architecture on data provided by a specific user. We show empirically that this simple and easy to implement technique is in fact capable of adapting the framework to the user. Some preliminary work in this direction can be found in Rokni et al. ([-@Rokni2018PersonalizedHA]). We greatly expand on it by: providing quantitative results on the improvements given by this personalization process; comparing with state-of-the-art techniques; and applying the personalization procedure to multiple, different, deep learning architectures. We also present an empirical evaluation of the learning capabilities of the proposed transfer learning technique.
Data Preprocessing {#sec:preproc}
==================
In this section we present the preprocessing of the sensor measurements that is performed for [<span style="font-variant:small-caps;">TrASenD</span>]{}[^1]. For each sensor $\mathcal{S}^{(i)}$, $i \in \{ 1,...,k \}$, let matrix $\bm{V}^{(i)}$ describe its measurements, and vector $\bm{u}^{(i)}$ define the timestamp of each measurement. $\bm{V}^{(i)}$ has size $d^{(i)} \times n^{(i)}$, where $d^{(i)}$ is the number of dimensions for each measurement from sensor $S^{(i)}$ (e.g., $3$ for both accelerometer and gyroscope as they measure data along the $x$, $y$, and $z$ axes) and $n^{(i)}$ is the number of measurements. $\bm{u}^{(i)}$ has size $n^{(i)}$. For each sensor $\mathcal{S}^{(i)}$, $i \in \{ 1,...,k \}$, the preprocessing procedure is defined as follows:
- Split the input measurements $\bm{V}^{(i)}$ and $\bm{u}^{(i)}$ along time to generate a series of *non-overlapping* intervals with width $\tau$. These intervals define the set $\mathbb{W}^{(i)} = \{ (\bm{V}^{(i)}_{t}, \bm{u}^{(i)}_{t}) \}$, where $|\mathbb{W}^{(i)}| = T$ and $t \in {1,..., T}$.
- For each pair belonging to $\mathbb{W}^{(i)}$ apply the Fourier transform and stack the inputs into a $d^{(i)} \times 2f \times T$ tensor $\mathbf{X}^{(i)}$, where $f$ is the dimension of the frequency domain containing $f$ magnitude and phase pairs.
Finally, we group all the tensors in the set $\mathbb{X} = \{ \mathbf{X}^{(i)} \}, i \in 1,...,k$, which is then the input to our [<span style="font-variant:small-caps;">TrASenD</span>]{} framework.
In practice, we first divide the measurements into samples with a length of 5 seconds *(with no overlap)*, and then apply the procedure with $\tau = 0.25$ seconds and $f = 10$. From now on, with the term *timestep* we refer to a given $\tau$-length interval. In order to deal with uneven sampling intervals that might appear in the data we first interpolate the measurements in each $\tau$-length interval, sample $f$ evenly separated points, and then apply the Fourier transform to those points. The interpolation is done with a linear interpolation along each measurement axis. The measurements in a 5 seconds sample *of each sensor* are passed to the architecture as a matrix of size *$T \times$ features dimension*, where $T = 20$ and *features dimension* $ = d^{(i)} \times 2f$ (each training and evaluation example is fed to the network with one matrix per sensor). Notice that applying a convolution operation with filters having a receptive field that spans a single row is like extracting features from each $\tau$-length interval separately.
#### Data Augmentation.
Similarly to Yao et al. ([-@Yao_2017]), for each training example we added other 9 artificial examples obtained by adding noise (with a normal distribution with zero mean and variance of $0.5$ for the accelerometer and of $0.2$ for the gyroscope). The idea behind this procedure is that the data generated by the sensors are already noisy, so having more samples with slightly different noise should make the network more robust to it. We analyze the impact of data augmentation in our experimental section.
Architecture {#sec:architecture}
============
In this section we present our framework [<span style="font-variant:small-caps;">TrASenD</span>]{}. We start with a description of the architectural template defined by the DeepSense framework [@Yao_2017][^2]. We then present the unique characteristics of [<span style="font-variant:small-caps;">TrASenD</span>]{} and its redesigned temporal extraction strategy that is based purely on attention. Finally, we present two additional variants of [<span style="font-variant:small-caps;">TrASenD</span>]{} with the goal of studying different temporal extraction strategies not based purely on attention, but still more advanced than the stacked GRU layer of DeepSense.
DeepSense {#ds_arc}
---------
DeepSense’s architecture (Figure \[fig:ds\]) can be divided in three parts: *convolutional layers*, *recurrent layers*, and *output layer*. The *convolutional layers* can be further divided into two subnetworks: an individual convolutional subnetwork for each sensor and a unique merge convolutional subnetwork. Each individual convolutional subnetwork (one per sensor) takes as input a matrix with dimension *$T \times $ features dimension* (see Section \[sec:preproc\]) and is composed of three convolutional layers with 64 filters each. The first layer has filters with dimension $1 \times 6d^{(i)}$ with a stride of $(1, d^{(i)} \times 2)$ [^3]. The second and the third individual convolutional layers have filters with dimension $1 \times 3$. The convolutions in all three layers are applied without padding and are followed by batch normalization [@Ioffe:2015:BNA:3045118.3045167], and a ReLu activation. Furthermore dropout [@Srivastava2014DropoutAS] is applied in between the layers, with probability $0.2$. The output of the individual layers are then concatenated, obtaining a tensor with dimension $T \times \text{\textit{number of sensors}} \times features \times channels$ (where *features* depends of the dimension of filters at the previous layers and *channels* is equal to the number of filters of the last individual convolutional layers), and passed to the merge convolutional subnetwork. This subnetwork is composed of three convolutional layers with 64 filters each. For each layer the dimensions of the filters are respectively $1 \times \text{\textit{number of sensors}} \times 8$, $1 \times \text{\textit{number of sensors}} \times 6$, $1 \times \text{\textit{number of sensors}} \times 4$, this time with padding. Again, after each layer, batch normalization and a ReLu activation are performed, with dropout in between layers (with probability $0.2$).
![Scheme of the DeepSense framework [@Yao_2017]. Individual convolutional subnetworks and the merge convolutional subnetwork share weights across timesteps.[]{data-label="fig:ds"}](Fig1.jpeg){height="3.0in"}
The *recurrent layers* are composed of two stacked GRU [@Chung2014EmpiricalEO] layers with 120 cells each. Dropout (with probability $0.5$) and recurrent batch normalization [@Cooijmans2017RecurrentBN] are performed between the two layers. Then the mean of the outputs at each time step is taken, and passed to the output layer.
Finally, the *output layer* is a simple dense layer with a number of units equal to the number of activities to predict. The *softmax* activation is used to get a probability distribution between the activities, and the cross-entropy is used as loss function: $$L = \sum_{i}^{N} \sum_{c}^{C} -\bm{y}^{(true)}_{i, c} \log(\bm{y}^{(pred)}_{i, c})$$ where $N$ is the number of training examples, $C$ is the number of different classes, $\bm{y}^{(true)}_{i, c}$ is the $c$-th element of the one-hot encoded ground truth for the $i$-th training example, and $\bm{y}^{(pred)}_{i, c}$ is the $c$-th element of the output of the architecture (after *softmax*) for the $i$-th training example.
[<span style="font-variant:small-caps;">TrASenD</span>]{}
---------------------------------------------------------
Recurrent Neural Networks (RNNs) present several problems, from the difficulty to learn long-term dependencies [@Hochreiter:01book; @Cho_2014], to their low computational efficiency. We propose a new framework, building on the architectural template defined in Section \[ds\_arc\], that replaces the *stacked-GRU recurrent layer* with an attention-based technique that better exploits temporal dependencies in the data.
We first introduce the attention operator, which is at the core of our attention-based technique for the extraction of temporal dependencies, and then present in more detail the architecture of our proposed framework. Figure \[fig:ds\_variants\] (a) shows a scheme of the architecture of our temporal dependencies extractor.
### Attention Operator
An attention operator takes as input three matrices: a *Query* matrix $\bm{Q}$, a *Key* matrix $\bm{K}$, and a *Value* matrix $\bm{V}$, where each row of the matrices indicates the query, key, or value vector of a specific item (where item usually refers to a feature vector). The attention operator attends every query to every key and obtains a similarity score (also called attention score) which is used to obtain weights for all the value vectors (rows of the *Value* matrix). Following [@Vaswani2017AttentionIA], we obtain the similarity score using the *scaled dot-product*, and then the attention weights by applying *softmax*. Finally, the values are scaled with their respective attention weight. The whole process can be written as: $$attention(\bm{Q}, \bm{K}, \bm{V}) = softmax\left( \frac{\bm{Q}\bm{K}^{T}}{\sqrt{d^{k}}} \right) \bm{V}$$ where $d^{k}$ is the dimension of query and key vectors. The weights are such that, for every query, the values related to the keys with the highest similarity score are given a higher weight (i.e., more importance). In other words, the weights are used to give more *attention* to the values that are more pertinent to the given query. We talk about *self-attention* when *Query*, *Key*, and *Value* matrices are all referring to items of the same sequence. A *multi-headed* mechanism is such that, for each item, different multiple *Query*, *Key*, and *Value* matrices are created and the attention operator is applied to all of them. The outputs of all the heads are then combined together.
### Architecture {#architecture}
[<span style="font-variant:small-caps;">TrASenD</span>]{} follows the feature extraction procedure and the feed-forward output layer of DeepSense, but completely replaces the *recurrent layers*. In fact, we only use *attention* to extract temporal dependencies in the data, with a temporal information extractor layer inspired by the Transformer [@Vaswani2017AttentionIA]. In more detail, we create a temporal information extractor using a *8-headed self-attention* mechanism. To pass the data to the temporal layer, we reshape the output of the merge convolutional subnetwork to have dimension $T \times \text{\textit{features}}$ (where *features* depends from the size and the number of filters in the merge convolutional subnetwork). The features at different timesteps will be the input of the *self-attention* mechanism. Every sublayer of the temporal block has output with size $T \times \text{\textit{features}}$ to allow residual connections.
We start by applying the positional embedding described by Vaswani et al. ([-@Vaswani2017AttentionIA]) to introduce a notion of relative order between the features extracted at different timesteps. Then, for each *head*, we first multiply the input with 3 different learnable matrices to obtain the *query, key, value* matrices $\bm{Q}, \bm{K}, \bm{V}$ (each row of these matrices represents query, key, and value vectors for each timestep). We then obtain the attention score using the *scaled dot-product*, where we used $d^{k} = 64$ and set the dimension of the values to be the same. The attention weights obtained from each *head* are then concatenated and multiplied by a learnable matrix to return to a matrix with dimension $T \times \text{\textit{features}}$. This matrix is then summed with the original inputs (creating a residual connection), and Layer Normalization [@Ba2016LayerN] is applied. The data in each timestep is passed through a *position-wise* dense layer[^4] with ReLu activation. Finally another residual connection with Layer Normalization is applied to obtain the output of the temporal information extraction block which is then passed to the feedforward output layer. A scheme of the temporal information extraction block can be found in Figure \[fig:ds\_variants\] (a).
Other Architectural Variants
----------------------------
We now present two variants of [<span style="font-variant:small-caps;">TrASenD</span>]{} where we replace the purely attention based temporal information extraction block, with other (simpler, but more advanced than regular RNNs) techniques to capture temporal dependencies in the input.
#### [<span style="font-variant:small-caps;">TrASenD</span>]{}-BD.
The first variant substitutes the pure attention temporal block with a bidirectional-RNN (BRNN) [@Schuster_1997]. A BRNN generalizes the concept of RNNs by connecting two hidden layers of opposite directions to the same output (we continue using GRUs as forward and backward hidden layers). This allows the network to get information from past and future inputs simultaneously. At each timestep we now get the state of both forward and backward cells, so we concatenate them, and finally take the average of the concatenated outputs at each timestep and pass them to the output layer.
#### [<span style="font-variant:small-caps;">TrASenD</span>]{}-CA.
Inspired by the work by Xu et al. ([-@pmlr-v37-xuc15]), we use a GRU layer (we keep it with 120 cells) with an attention mechanism over the output features of the merge convolutional subnetwork. We first average the features extracted from the first $\tau$-length interval (first *timestep*) and pass it through a dense layer to obtain the initial state for the GRU layer. We then use the following attention mechanism: at each timestep, we pass the features extracted by the CNN layers and the current state of the GRU through two different dense layers *without applying any activation function*. We then sum the two outputs and apply $tanh$ before passing it to *softmax* to obtain the *attention weights*. Finally, the features are scaled with their attention weights. The sum of the scaled feature vectors forms the *context vector* which is then concatenated to the original features for the current timestep and passed as input to the GRU. A scheme of this attention mechanism can be found in Figure \[fig:ds\_variants\] (b). The rest of the architecture remains unchanged.
Transfer Learning Personalization {#sec:tl_pers}
---------------------------------
To make the system capable of adapting to a specific user over time, we propose a simple *transfer learning* strategy. Transfer learning is a method where a model developed for a task is reused as the starting point to learn a model on a second task. The typical scenario in a transfer learning setting is to have a trained base network, which is repurposed by training on a target dataset. The idea is that the pre-trained weights in the base network can ease the training on the target dataset. We slightly depart from this scenario by extracting the output layer from a trained [<span style="font-variant:small-caps;">TrASenD</span>]{} model (and other proposed variants); that is, we are using transfer learning only on the output layer. More in detail, the data coming from the sensors will be passed to the [<span style="font-variant:small-caps;">TrASenD</span>]{} architecture, up to the end of the temporal layer. The output layer becomes a separate network that receives the output of the temporal layer as input, and will be trained with the data generated by the user. This can be implemented in a practical scenario by first using a model trained on one of the datasets, and after each prediction, asking the user to manually insert the activity he was performing. We then use these new data samples to retrain *only* the output layer, which is a single layer dense network that can easily be trained on-device. This procedure allows the architecture to take advantage of the complex general *feature extracting* mechanism that reduces multimodal time series to a fixed size vector, and to successively learn user-specific feature characteristics.
Experimental Evaluation
=======================
We present here the datasets and the procedure used to evaluate the performance of [<span style="font-variant:small-caps;">TrASenD</span>]{}, performing comparisons with multiple other methods for HAR, and to evaluate the proposed personalization process. Furthermore, we present an empirical study of the benefits of the data augmentation procedure.
Datasets
--------
We present below the three human activity recognition datasets used in our tests. Our choices were based on the statistics shown in Table 3 of the survey by Wang et al. ([-@Wang_2019]): we considered the datasets that had data from at least 9 subjects (to better test generalization properties), with at least 2 different sensing modalities (to test the various methods on multimodal data), and then took the datasets with the largest number of samples. A summary of the chosen datasets can be found in Table \[tab:dataStat\].
#### HHAR
[@Stisen_2015]. The Heterogeneity Activity Recognition Data Set contains data from accelerometer and gyroscope of 12 different devices (8 smartphones and 4 smartwatches) used by 9 different subjects while performing 6 activities. We only considered data coming from smartphones.
#### PAMAP2
[@Reiss_2012_2; @Reiss_2012]. The Physical Activity Monitoring dataset contains data of 12 different physical activities, performed by 9 subjects wearing 3 inertial measurement units and a heart rate monitor. We only considered data coming from the inertial measurement units (IMU), which were positioned in three different body areas (hand, chest, ankle) during the measurements. From each IMU we considered data measured by the first accelerometer, the gyroscope and the magnetometer. This provides a scenario with data coming from 9 input sensors.
#### USC-HAD
[@Zhang_2012]. The University of Southern California Human Activity Dataset uses high precision specialised hardware, and has a focus on the diversity of subjects, balancing the participants based on gender, age, height and weight. The dataset contains measurements from accelerometer and gyroscope obtained from 14 different subjects while performing 12 activities.
\[tab:dataStat\]
**Dataset** **Subjects** **Activities** **Input Sensors**
------------- -------------- ---------------- -------------------
**HHAR** 9 6 2
**PAMAP2** 9 12 9
**USC-HAD** 14 12 2
: Summary of the multi-modal HAR datasets used for our tests.
Baselines
---------
We choose an extensive collection of deep learning, and non-deep learning methods to compare to [<span style="font-variant:small-caps;">TrASenD</span>]{} and its variants. For all considered models, we use the implementation provided by the authors when it is available, and we implement it from scratch following the description from the papers when it is not. Unless otherwise specified we use the model hyperparameters defined by the authors.
#### Deep Learning Baselines.
We test our algorithm against all the DeepSense-based architectures, and additional deep learning techniques. In particular for the DeepSense-based architectures we test against the original DeepSense [@Yao_2017], and the two latest attention enhanced versions: SADeepSense [@Yao_2019], and AttnSense [@Ma_2019]. We then consider DeepConvLSTM [@Ord_ez_2016] which is a CNN+LSTM approach, and its new attentive version proposed in [@Murahari_2018] that we call DeepConvLSTM-Att. All the attention models considered thus far add an attention module to a RNN layer, while we remember that our algorithm [<span style="font-variant:small-caps;">TrASenD</span>]{} completely removes RNNs in favour of a purely attention-based temporal information extraction technique. We also provide some results for a basic LSTM based architecture (we implement it with 2 LSTM layers, each with 256 cells, followed by a fully connected layer that outputs the predicted class). Finally, to take into consideration also other deep learning techniques we consider MultiRBM [@Radu_2016], where a Restricted Boltzman Machine (RBM) is used for each sensor, and a single final RBM is used to then merge all the outputs for the sensors and obtain the predicted class.
#### Non-Deep Learning Baselines.
As non-deep learning baselines we considered a Random Forest (RF) classifier (one of the most used and most effective shallow classifiers for HAR [@Stisen_2015]) on the same raw frequency domain features fed to the deep learning approaches (denoted with **RF-FF**), and then on the most used handcrafted frequency domain features (DC Component, Spectral Energy, and Information Entropy; denoted with **RF-HC**).
Experimental Setup
------------------
For all tests we performed *leave-one-user-out* cross validation: we train on data from all subjects except one, and we use the data from the excluded subject as test set. We perform this procedure for each subject and then average the results. To evaluate the personalization process we divide the data of each activity of the excluded user into two equal *time-contiguous* parts. One part is used to personalize the output layer after the model has been learned on all other users, and the other is used as test set. We also make sure to feed the data, both for training and validation, in time-contiguous samples (simulating the real-world personalization procedure described in Section \[sec:tl\_pers\]).
Due to the imbalance in the number of samples per-class we use the F1 score as the measure to quantify the performance of the models. All [<span style="font-variant:small-caps;">TrASenD</span>]{} models were implemented[^5] using TensorFlow [@abadi2016tensorflow].
To ensure a fair comparison and to avoid “hyperparameter hacking” we kept all the values for the *architecture hyperparameters* (filters size, dropout probability, number of filters, number of GRU units, etc.: see Section \[sec:architecture\].) equal for each DeepSense-based model. Furthermore, for all models, the only optimized hyperparameter was the *learning rate*. To do so we took out 1 user and tried the training and evaluation procedure on the HHAR dataset, with *learning rate* $\in {\{10^{-2}, 10^{-3}, 10^{-4}\}}$. We then considered the setting that gave the highest F1 score on the user’s data and used it for *all* datasets (no optimization for each different dataset). In the training procedure we trained for 30 epochs for each user and took the model of the epoch with the highest performance. All [<span style="font-variant:small-caps;">TrASenD</span>]{} based models were trained using the Adam Optimizer [@article]. The other methods were trained with the optimization technique suggested by the authors. For the personalization process, we retrain the output layer for 1 epoch (per each new data point separately) with TensorFlow’s default Adam optimizer parameters: $\alpha = 0.001$, $\beta1 = 0.5$, $\beta2 = 0.9$, and $\epsilon = 10^{-8}$.
Results
-------
-------------------------------------------------------------------- ----------- ----------- ----------- -- -- --
**Model**
HHAR PAMAP2 USC-HAD
RF-FF 0.569 0.512 0.417
RF-HC [@Stisen_2015] 0.575 0.501 0.474
MultiRBM [@Radu_2016] 0.647 0.589 0.598
LSTM 0.663 0.583 0.612
DeepConvLSTM [@Ord_ez_2016] 0.701 0.633 0.658
DeepConvLSTM-Att [@Murahari_2018] 0.735 0.647 0.682
DeepSense [@Yao_2017] 0.720 0.647 0.670
SADeepSense [@Yao_2019] 0.753 0.661 0.688
AttnSense [@Ma_2019] 0.762 0.657 0.685
**[<span style="font-variant:small-caps;">TrASenD</span>]{}*-BD*** 0.798 0.650 0.681
**[<span style="font-variant:small-caps;">TrASenD</span>]{}*-CA*** 0.797 0.659 0.687
**[<span style="font-variant:small-caps;">TrASenD</span>]{}** **0.848** **0.723** **0.702**
-------------------------------------------------------------------- ----------- ----------- ----------- -- -- --
Table \[tab:res2\] summarizes the F1 score results for [<span style="font-variant:small-caps;">TrASenD</span>]{} and the other methods we considered, on the three datasets We can observe that [<span style="font-variant:small-caps;">TrASenD</span>]{} and its variants present higher F1 score than DeepSense on all the three datasets. Furthermore we notice that [<span style="font-variant:small-caps;">TrASenD</span>]{} always achieves the highest performance with a big margin. In fact, [<span style="font-variant:small-caps;">TrASenD</span>]{} shows an F1 score that is, on average, $7\%$ higher then the previous best performing model. These results confirm that our attention-based technique (without RNNs) is highly capable of extracting temporal dependencies. Most notably we can see that [<span style="font-variant:small-caps;">TrASenD</span>]{} significantly outperforms the newer SADeepSense and AttnSense, whose performance are comparable to the ones of [<span style="font-variant:small-caps;">TrASenD</span>]{}*-BD* and [<span style="font-variant:small-caps;">TrASenD</span>]{}*-CA*, which are far from [<span style="font-variant:small-caps;">TrASenD</span>]{}’s.
-------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- -----------
**Model**
NP P NP P NP P
DeepSense [@Yao_2017] 0.720 0.775 0.647 0.693 0.670 0.712
SADeepSense [@Yao_2019] 0.753 0.790 0.661 0.699 0.688 0.749
AttnSense [@Ma_2019] 0.762 0.801 0.657 0.689 0.685 0.746
**[<span style="font-variant:small-caps;">TrASenD</span>]{}*-BD*** 0.798 0.821 0.650 0.699 0.681 0.748
**[<span style="font-variant:small-caps;">TrASenD</span>]{}*-CA*** 0.797 0.819 0.659 0.701 0.687 0.726
**[<span style="font-variant:small-caps;">TrASenD</span>]{}** **0.848** **0.889** **0.723** **0.749** **0.702** **0.759**
-------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- -----------
Table \[tab:res3\] presents the F1 score of the DeepSense-based models when evaluated on the datasets with and without applying personalization. The results confirm the effectiveness of our transfer learning personalization process giving an average $\approx6.2\%$ increase on the F1 score independently of dataset and base architecture.
We can therefore conclude that pure attention based models are effective also outside of the Natural Language Processing scenario they were proposed for, and they are a very powerful technique that shows great results on sensor-based data.
### Validating the Personalization Process
To prove that the training of the output layer alone can significantly impact on the performance of the network we first train the full model of Section \[ds\_arc\] on the HHAR dataset with randomly permuted labels, and then we perform the personalization process on correctly labeled data. The resulting F1 scores (on the test set) are 0.166 and 0.523, respectively. We can notice that the model trained on data with randomly permuted labels has the performance of a uniform random classifier, as one would expect, and the personalization process is capable of significantly boosting the performance of the model. This result shows that in fact the re-training of the output layer alone can largely affect the outcome of the model.
**Model** **F1 score on Test Set**
------------------------ --------------------------
DeepSense 0.166
Personalized DeepSense 0.523
: Performance of the network on the test set after training it on data with permuted labels with and without personalization on correctly labeled data.[]{data-label="tab:resPers"}
![Performance of the Deep Learning models on HHAR when trained with different number of augmented samples.[]{data-label="fig:aug_plot"}](Fig4.pdf){width="0.7\linewidth"}
### Impact of Data Augmentation
To asses the benefits of the data augmentation procedure, we evaluate all the deep learning models based on the DeepSense framework on HHAR with and without augmented data. The results, shown in Table \[tab:resAug\], confirm that data augmentation is important to train a model that is more robust to noise, and in fact we can see a significant increase in the F1 score. Figure \[fig:aug\_plot\] shows how the performance of the analyzed DeepSense variants change when trained with different number of augmented samples. It’s interesting to see that using 4 augmented samples for each real sample, already provides an important performance gain. We also notice that [<span style="font-variant:small-caps;">TrASenD</span>]{} is always superior to the other architectures, and performs significantly better than the others even when trained without augmented samples. Furthermore, we see that SADeepSense and [<span style="font-variant:small-caps;">TrASenD</span>]{} are the two architectures showing the smallest gap between highest and lowest F1 score result, confirming their superior generalization properties, with [<span style="font-variant:small-caps;">TrASenD</span>]{} achieving a higher overall F1 score.
\[tab:resAug\]
---------------------------------------------------------------- ----------- -----------
**Model**
A NA
DeepSense [@Yao_2017] 0.720 0.621
SADeepSense [@Yao_2019] 0.753 0.682
AttnSense [@Ma_2019] 0.762 0.687
[<span style="font-variant:small-caps;">TrASenD</span>]{}*-BD* 0.798 0.646
[<span style="font-variant:small-caps;">TrASenD</span>]{}*-CA* 0.797 0.638
**[<span style="font-variant:small-caps;">TrASenD</span>]{}** **0.848** **0.761**
---------------------------------------------------------------- ----------- -----------
: Performance on HHAR with (A) and without (NA) data augmentation.
Conclusions
===========
In this paper we presented [<span style="font-variant:small-caps;">TrASenD</span>]{}, a new deep learning frameworks for multimodal time series, and also proposed a transfer learning procedure to personalize the model to a specific user for the human activity recognition tasks. [<span style="font-variant:small-caps;">TrASenD</span>]{} is designed to improve the extraction of temporal dependencies in the data by replacing RNNs with a purely attention based temporal information extraction block. Our extensive experimental evaluation shows that [<span style="font-variant:small-caps;">TrASenD</span>]{} significantly outperform the state-of-the-art and that, in general, replacing RNNs with attention-based strategies leads to significant improvements. In particular, we obtain an average increment of more than $7\%$ on the F1 score over the previous best performing model. We also show the effectiveness of our simple personalization process, which is capable of an average $6\%$ increment on the F1 score on data from a specific user, and the impact of data augmentation.
The personalization procedure we propose may impact the user experience while using an application that implements our technique. In fact, asking too many times for feedback about the model’s predictions may not be feasible. Future research directions include the optimization of the personalization process to minimize the feedback required from the user, for example by using data augmentation or curriculum training techniques [@Bengio_2009].
Work partially supported by MIUR, the Italian Ministry of Education, University and Research, under PRIN Project n. 20174LF3T8 AHeAD (Efficient Algorithms for HArnessing Networked Data) and under the initiative “Departments of Excellence" (Law 232/2016), and by the grant STARS2017 from the University of Padova.
[^1]: DeepSense [@Yao_2017] applies a similar procedure, however, we also report some details, like the interpolation of the measurements, and the exact values of the parameters, that were not specified in [@Yao_2017].
[^2]: In [@Yao_2017] the authors do not specify several architectural parameters (filter dimensions, strides, presence of padding, dropout probability, training optimizer, learning rate, etc.). We refer to the parameters that can be found on the author’s implementation available at <https://github.com/yscacaca/DeepSense>.
[^3]: Intuitively, the filters have a receptive field that covers three measurement points, and have a stride of one measurement point (after the Fourier transform each point is represented by two numbers: magnitude and phase).
[^4]: The same feedforward network is used for each timestep. It is equivalent to a one-dimensional convolutional layer over timesteps with kernel size 1.
[^5]: Code is available at: <https://github.com/DavideBuffelli/TrASenD>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study high–Higgs boson production at hadron colliders to order 3 in hadron collisions. In particular, we investigate the process $g+q/\bar{q}\ra q/\bar{q}+H$, where $q=u,d,c,s,$ or $b$, for the LHC (a $\sqrt{s}=$14 TeV, proton–proton collider). Our results are compared to the calculation. The associated production of a high–$p_T$ Higgs boson with a $b$–quark or anti–quark is comparable to the calculation because of the large top quark mass and the additional contribution of electroweak gauge and Goldstone bosons. The associated production of light quarks, however, is not significant. We also comment on new physics effects in the framework of the electroweak chiral Lagrangian.'
---
3[[W]{}\^[3]{}]{} å[[A]{}]{} | 3 \#1[0= 0=0 1= 1=1 0>1 \#1 / ]{} \#1\#2[[0=1=1>0.51-.500-.50-.5110>1 .50-.51]{}]{} 0[DØ]{}
MSUHEP–050627\
CITHE–68–2006
[**S. Mrenna**]{}[^1]
[Lauritsen Laboratory\
California Institute of Technology\
Pasadena, CA 91125]{}
and
[**C.–P. Yuan**]{}[^2]
[Department of Physics and Astronomy\
Michigan State University\
East Lansing, MI 48824]{}
PAC codes: 12.15.Lk, 14.80.Bn
Introduction
============
With the discovery of the top quark [@CDF], the only remaining element of the Standard Model (SM) particle spectrum is the Higgs boson. Experimentally, there are only lower bounds on $M_H$. LEP-I has placed the limit $M_H > 64.5$ GeV [@LEPH0]. Theoretically, there are upper bounds in the SM from unitarity and triviality arguments [@upperH]. One goal of the future High Energy Physics experimental program is to discover the Higgs boson and verify its properties or determine the alternative mechanism of electroweak symmetry breaking. The search for the Higgs boson at LEP-II is strictly limited by the available center of mass energy and luminosity, so that only $M_H <$ 90–95 GeV can be probed for an energy of 190 GeV and 500 pb$^{-1}$ of data [@LEPII]. The reach of a high–luminosity Tevatron collider is better, but becomes challenging above $M_H =$ 110 GeV [@marciano]. The LHC, on the other hand, is hoped to have enough energy, luminosity, and instrumentation to decisively probe the energy scale associated with electroweak symmetry breaking. This task is not as straight–forward as it may seem. One possible alternative to the SM is the Minimal Supersymmetric Standard Model (MSSM) with a constrained multi–dimensional parameter space [@arnowitt; @lopez; @kane]. The constrained MSSM models predict that the couplings of the lightest Higgs boson to SM particles is very SM–like, i.e. $\sin^2(\beta-\alpha)\simeq 1$, and its mass should be less than about 140GeV [@susyH]. In this case, to deduce supersymmetry, one must observe a superpartner directly or discern its presence in quantum corrections. Another alternative, a model with a strongly interacting scalar sector [@strongH], predicts a greatly enhanced Higgs boson width even for a Higgs mass of a few hundred GeV, so that the Higgs boson signal can be hidden by backgrounds. Regardless of the scenario, a full verification of the properties of the Higgs boson requires a deep theoretical understanding of its properties. In this investigation, we concentrate on the high–production of Higgs bosons, which is sensitive to loop corrections. The $\cal{O}$$(\alpha_s^3\gf)$ contribution to high–Higgs boson production was calculated previously [@ellis], where $\gf = (\sqrt{2}v^2)^{-1}$ and the vacuum expectation value $v =$ 246 GeV. Here, we extend that calculation to include the $\cal{O}$$(\alpha_s\gfc)$ contributions from electroweak gauge bosons, Goldstone bosons, and quarks. In particular, since the top mass is large, we expect to see an enhancement in the associated production of a Higgs boson with a $b$–quark or anti–quark in some kinematic region. We also expect this channel to be sensitive to the coupling of the electroweak gauge bosons and Goldstone bosons to the Higgs boson, since it does not vanish in the limit that the $U(1)_Y$ and $SU(2)_L$ gauge couplings vanish.[^3] We study this sensitivity in the framework of the electroweak chiral Lagrangian, which allows us to construct the most general effective Lagrangian that is consistent with $SU(2)_L\times U(1)_Y\ra U(1)_{em}$ symmetry breaking. We show that with new physics, the 3 contribution can be comparable to the contribution for high–Higgs boson production. Another source of high–Higgs bosons is the ${\cal O}$$(\alpha_s\gf)$ tree level process. We show that this is small by examining the processes $q+\bar{q}\ra b+\bar{b}+H$ and $g+g\ra b+\bar{b}+H$. We also argue that any interference between this order amplitude and one of higher order is suppressed because the bottom quark mass $m_b$ is much less than the electroweak symmetry breaking scale $v$.
High–production of the Higgs boson to 3
=======================================
To 3, the Higgs boson is produced at high–from quark–gluon, antiquark–gluon, and quark–antiquark initial states. Since they are the most interesting, we will concentrate on the first two processes for the purpose of this discussion. The quark–antiquark annihilation process is typically an order of magnitude smaller at the LHC for $M_H \le 400$ GeV. We chose to perform the calculation in the helicity formalism, since the amount of algebra is reduced significantly. Furthermore, we used the Feynman rules in the ’t Hooft–Feynman gauge, since the electroweak gauge bosons and their associated Goldstone bosons have the same mass and, hence, loop integrals involving gauge bosons and Goldstone bosons have the same denominators. This choice is also advantageous for investigating the electroweak chiral Lagrangian.
We consider the process $g(p_g)+q(p_j)\ra q(p_i)+H(p_H)$ to 3, where the $p_g$ and $p_j$ are the four-momenta of the incoming particles and $p_i$ and $p_H$ are the four-momenta of the outgoing particles. The quark–antiquark initial state can be generated by the substitution $p_i\ra-p_i, p_g\ra -p_g$ and a reevaluation of the color factor. Contributions to the loop integral come from internal lines involving the weak isospin quark partner of $q$ (we use the simplification $V_{ud}=V_{cs}=V_{tb}=1$), gauge bosons, and Goldstone bosons. Some representative Feynman diagrams are illustrated in Figure 1.
The amplitude for this process can be written as: $$\begin{aligned}
\label{eqc0}
{\cal M}_{\lambda_i \lambda_j \lambda_g} & = & i g_s
\bar{u}({\lambda_i},p_i)[{\cal F}
\gamma_{\mu}+{\cal G}_{\mu}{\slashchar{p}_{H}}]
u({\lambda_j},p_j)\epsilon^{\mu}_{\lambda_g} ,\end{aligned}$$ where $\lambda_i$ and $\lambda_j$ are the fermion helicity indices, $\lambda_g$ is the gluon polarization, $\gamma_\mu$ are $4\times 4$ gamma matrices, $\slashchar{a} = a^\mu \gamma_\mu$ for 4–vector $a^\mu$, and equations of motion have been applied to the on–shell, 4–component spinors.
In addition, we have taken the limit $m_q\ra 0$ for all $q$ except $t$. The complex scalar ${\cal F}$ and the complex vector ${\cal G}_{\mu}$ are form factors resulting from the integration of the loop momentum, and their explicit expression are given in the Appendices I and II. The helicity amplitude can be re-written in terms of 2–component Weyl spinors using the bra–ket notation:[^4]
$$\begin{aligned}
u_{-}(\lambda=-1/2,p_a)& = &\omega^a_{+} |p_a-\rangle, \\
v_{-}(\lambda=+1/2,p_a)& = &-\omega^a_{+} |p_a-\rangle, \\
u_{+}(\lambda=+1/2,p_a)& = &\omega^a_{+} |p_a+\rangle, \\
v_{+}(\lambda=-1/2,p_a)& = &-\omega^a_{+} |p_a+\rangle,\end{aligned}$$
where $\omega^a_{+}=\sqrt{2 E_a}$ for massless fermions with energy $E_a$, and $\langle p\pm| = (|p\pm\rangle)^\dagger$. These are the only components for massless fermions. For $p^\mu=(E,p s_\theta c_\phi,p s_\theta s_\phi,p c_\theta)$, where $s_\psi$ and $c_\psi$ are shorthand for $\sin\psi$ and $\cos\psi$, $|p+\rangle = (\cos\theta/2,e^{i\phi}\sin\theta/2)^{\rm T}$, and $|p-\rangle = (-e^{-i\phi}\sin\theta/2, \cos\theta/2)^{\rm T}$[^5]. Also, the gluon polarization 4–vectors for left-handed $(L)$ and right-handed $(R)$ helicities can be written as: $$\begin{aligned}
& \epsilon^{\mu}_{\scriptscriptstyle (L)}
= \displaystyle{\frac{e^{-i\phi}}{\sqrt{2}}}
[0,i s_\phi+c_\phi c_\theta,-i c_\phi + s_\phi c_\theta,-s_\theta], & \\
& \epsilon^{\mu}_{\scriptscriptstyle (R)}
= \displaystyle{\frac{e^{i\phi}}{\sqrt{2}}}
[0,i s_\phi-c_\phi c_\theta,-i c_\phi - s_\phi c_\theta,s_\theta], &\end{aligned}$$ where $\phi$ and $\theta$ are spherical coordinates of the gluon momentum. In the helicity basis, there are four non–vanishing amplitudes (as $m_q\ra0$). The parton–level cross section is: $$\begin{aligned}
& d\hat{\sigma}
={\displaystyle\frac{1}{F}\frac{1}{S}}C_{gb}
{\displaystyle\sum_{\lambda_g=L,R}^{ }}
(|{\cal M}_{--\lambda_g}|^2 + |{\cal M}_{++\lambda_g}|^2 )dR_2,& \\
& {\cal M}_{--\lambda_g} = i g_s \omega^i_+\omega^j_+
\langle p_i -|[{\cal F}{\gamma_+}_{\mu}+{\cal G}_{\mu}
{\slashchar{p}_{H}}_+]|p_j -\rangle \epsilon^{\mu}_{\lambda_g}, & \\
& {\cal M}_{++\lambda_g} = i g_s \omega^i_+\omega^j_+
\langle p_i +|[{\cal F}{\gamma_-}_{\mu}+{\cal G}_{\mu}
{\slashchar{p}_{H}}_-]|p_j +\rangle \epsilon^{\mu}_{\lambda_g}, &\end{aligned}$$ where the flux factor $F = 2\hat{s}$ for ${\hat{s}}=(p_g+p_j)^2$; the spin average factor $S = 2\times2$; the color factor is $C_{gq}=4/(3\times 8)$ for $g+q\ra q+H$, and $C_{qq}=4/(3\times 3)$ for $q+\bar{q}\ra g+H$; the gluon polarization is specified by $\lambda_g$; $dR_s$ is the two–body phase space; $\gamma_\pm^\mu$ are the $2\times 2$ matrices $({\bf 1},\mp\sigma_i)$,[^6] and $\slashchar{a}_\pm = a_\mu{\gamma_\pm^\mu}$. In the above result, ${\cal M}_{++\lambda_g}$, which only contains contributions from $Z^0$–bosons, are small for two reasons. First, since we are only interested in initial and final states without $t$–quarks, the internal quark is always light (because of the neutral current) and has a tiny coupling to the Goldstone boson. Secondly, the left– and right–handed couplings of the $Z^0$–boson, which are smaller than the purely left–handed coupling of the $W^\pm$–bosons, appear in the squared matrix element to the fourth–power. For all practical purposes, then, the $Z^0$ contributions can be ignored, leaving only two independent helicity amplitudes, ${\cal M}_{--\lambda_g}$, differing only in the gluon polarization[^7].
Because of gauge invariance, each amplitude satisfies the Ward Identity resulting from replacing the gluon polarization vector with the gluon four momentum. This simplifies to ${\cal F}+p_g\cdot{\cal G} = 0$. The form factors are calculated numerically using the FF Fortran library [@FF], so the Ward Identity can be verified numerically. Rotational and Lorentz invariance are also checked in the same manner.
The high–production of the Higgs boson at a hadron collider is calculated by folding the parton–level cross section with the parton distribution functions (PDF). We use CTEQ2L parton distribution functions and evaluate coupling constants at the momentum scale $Q^2=\hat{s}$. Unless otherwise stated, we use $m_t$=175 GeV in all calculations.
Numerical Results
=================
High–Higgs boson production at a 2 TeV $\bar{{\rm p}} {\rm p}$ collider is too small to be observed for all practical purposes.[^8] We present results only for the LHC (a 14 TeV ${\rm p}{\rm p}$ collider). In Table I, we list the production cross section to 3 for several Higgs boson masses as well as the contribution for the transverse momentum of the Higgs boson $p_T^H>30$ GeV. There are separate columns for the associated production of the Higgs boson with $b$–quarks and $u,d,s,c$–quarks. For all of these results, we include both the quark and antiquark contribution. For the associated production of $H$ with a $b$–quark or anti–quark, the total 3 cross section for $p_T^H > 30$ GeV is as large as 10–20% of the for $M_H = 100-200$GeV. As the transverse momentum of the Higgs boson increases, the 3 contribution becomes relatively more important. For this process, the $m_t$ dependence is minimal. For instance, for $M_H = 110$GeV, the total 3 cross section of $b/\bar{b}+H$ is 16.9fb and 17.5fb for $m_t = 160$GeV and 190GeV, respectively. For the associated production of $H$ with light quark, the 3 cross section for $p_T^H > 30$ GeV is never more than about 3% of the for $M_H = 100-200$GeV. The cross section for $q+\bar{q}\ra H+g$ is much smaller than for the corresponding process $g+q/\bar{q}\ra q/\bar{q}+H$, having values (1.5,.15)fb for $M_H$ = (110,400)GeV, and will not be discussed further.
We also studied the $p_T^H$ dependence of the cross section as a function of $M_H$. In Figure 2 we show the cross section integrated above $p_T^H$ for $b/\bar{b}+H$ production to 3 and in the same $M_H$ range as in Table I. The mean $p_T^H$ of the 3 process, for $p_T^H$ in the range 50–350 GeV, is a slowly varying function of $M_H$ below $M_H$ = 180 GeV, with a value of approximately 120 GeV. Above $M_H$ = 180 GeV, the mean $p_T^H$ ranges takes on the values (139,152) GeV for $M_H$ = (200,400) GeV. The mean $p_T^H$ of the process is strongly dependent on the lower transverse momentum cutoff needed to regulate the $p_T^H \ra 0$ divergence associated with the gluon progagator, and is somewhat smaller than that for the 3 process.
[|c|c|c|c|c|]{}\
& &\
(GeV) & & & &\
110 & 17.1 & 125.4 & 46.5 & 2.2$\times 10^3$\
140 & 15.1 & 88.8 & 37.0 & 1.6$\times 10^3$\
180 & 7.4 & 61.6 & 25.2 & 1.2$\times 10^3$\
400 & 1.9 & 29.6 & 0.5 & 0.7$\times 10^3$\
Finally, we address the issue of the ${\cal O}$$(\alpha_s\gf)$ process $g+b/\bar{b}\ra b/\bar{b}+H$, which is a tree level process. To estimate the size of this cross section, we used the processes $q+\bar{q} \ra b+\bar{b}+H$ and $g+g\ra b+\bar{b}+H$ in Pythia 5.7 [@pythia] with $p_T^H > 50$ GeV. We obtained the values (4.5,3.0,2.0,.3) fb for $M_H$ = (110,140,180,400) GeV. In the limit that $m_b\ra0$, there is no interference between the tree level process and the higher order amplitudes[^9], so any observed cross section is primarily the and 3 processes.
In summary, we find that to accurately predict the cross section and test the properties of an intermediate mass Higgs boson produced at high–$p_T$ in association with $b$ quarks and anti–quarks, the 3 contributions should be included with the contributions. Because of the large top quark masss, the 3 contributions are larger than the tree-level contributions of ${\cal O}(\alpha_s\gf)$ for large $p_T^H$. Although the production rate of 3 for the associated production of the Higgs boson with a light quark or anti–quark is not negligible, it is only a few percent of the rate, and therefore is probably not distinguishable from the uncertainty in the PDF.
The Electroweak Chiral Lagrangian and Non-Standard Model Couplings
==================================================================
The process under consideration is sensitive to electroweak symmetry breaking in three different sets of couplings. First, there is the $t$-$b$-$W$ vertex. The additional non-standard couplings can be deduced from the chiral Lagrangian [@tbWnew]: $$\begin{aligned}
{\cal L}& = &-\sqrt{2}\kappa_L^{CC}\bar{t_L}\gamma^\mu{b_L}\Sigma^{+}_{\mu}
- \sqrt{2}{\kappa_L^{CC}}^\dagger\bar{b_L}\gamma^\mu{t_L}\Sigma^{-}_{\mu}
\nonumber \\
& & - \sqrt{2}\kappa_R^{CC}\bar{t_R}\gamma^\mu{b_R}\Sigma^+_\mu
- \sqrt{2}{\kappa_R^{CC}}^\dagger\bar{b_R}\gamma^\mu{t_R}\Sigma^-_\mu,\end{aligned}$$ where $\Sigma^\pm_\mu = \frac{1}{\sqrt{2}} (\Sigma^1_\mu\mp i\Sigma^2_\mu)$ for $\Sigma^a_\mu = -\frac{i}{2}{\rm Tr}
(\sigma^a\Sigma^\dagger D_\mu\Sigma)$, and the action of the covariant derivative is $D_\mu\Sigma=\partial_\mu\Sigma-i gW^a_\mu\frac{\sigma^a}{2}\Sigma +
i g^{'}\Sigma B_\mu\frac{\sigma^3}{2}$. The matrix field $\Sigma ={\rm{exp}}\left ( i\frac{\phi^{a}\sigma^{a}}{v}\right )$, where $\sigma^{a},\, a=1,2,3, $ are the Pauli matrices, and $\phi^a$’s are the Goldstone bosons. Second, there is the Yukawa coupling between the Higgs boson and top quark. The most general Yukawa coupling of the fermion doublet $F$ in the chiral Lagrangian is: $$\begin{aligned}
{\cal L} = -\frac{c_0}{v}H \bar{F} M F,\end{aligned}$$ where M is a $2\times 2$ mass matrix. In the Standard Model, $c_0 = 1$. Third, the coupling of the Higgs boson and the electroweak Goldstone bosons comes from the Lagrangian: $$\begin{aligned}
\label{eqc1}
{\cal L} & = &\frac{1}{2}\partial_{\mu}H\partial^{\mu} H -
{\frac{1}{2}} M_H^2 H^2 - V(H) \nonumber \\
& & + (\frac{c_1}{2} v H + \frac{c_2}{4} H^2)
{\rm Tr}\left(D_\mu\Sigma^\dagger D^\mu\Sigma\right).\end{aligned}$$ In the Standard Model, $c_1 =c_2 = 1$.
To illustrate that new physics effects may enhance the 3 rate but not the rate, we study the effects of new physics arising from the scalar sector of the Lagrangian, as shown in Eq.(\[eqc1\]). In this case, the contribution is not modified. As shown in Ref. [@strongH], some models of the symmetry breaking sector allow the coefficient $c_1$ in Eq.(\[eqc1\]) to be larger than 1. ($c_2$ is irrelevant to the processes of interest.) For instance, $c_1=\sqrt{8/3}$ was discussed in Ref. [@strongH]. Because new physics can simutaneously modify the interactions of $t$-$b$-$W$ and $W$-$W$-$H$, we do not intend to give predictions for any specific model. For simplicity, we only study the effects of new physics due to $c_1$ in the limit that the $SU(2)_L\times U(1)_Y$ gauge couplings $g$ and $g^\prime$ vanish. Table II contains some of our results.
[|c|c|c|c|c|c|]{}\
\
&\
(GeV) & $-.5$ & 0.0 & 0.5 & 1.0 & 1.5\
110 & 1.09 & 0.87 & 0.78 & 0.83 & 1.01\
140 & 1.26 & 0.79 & 0.87 & 1.49 & 2.64\
180 & 0.74 & 0.69 & 2.00 & 4.67 & 8.71\
200 & 0.42 & 0.63 & 1.76 & 3.82 & 6.81\
250 & 0.19 & 0.49 & 1.32 & 2.70 & 4.60\
300 & 0.12 & 0.36 & 1.00 & 2.01 & 3.41\
400 & 0.06 & 0.21 & 0.59 & 1.20 & 2.03\
Although these rates do not represent the true rates of the process $g+b/\bar{b} \rightarrow b/\bar{b}+H$, they illustrate that the rates can vary by about a factor of 2 for a heavier Higgs boson. If the electroweak corrections to high–$p_T$ Higgs production are substantially modified by new physics at the order ${\cal O}$$(\alpha_s\gfc)$, then this can be observed in $b/\bar{b}+H$ production at future hadron colliders.
Discussion and Conclusions
==========================
Because the top quark mass is large, of the order of $v$, the interaction of the top quark and the Goldstone bosons is strong and, therefore, can be sensitive to the electroweak symmetry breaking sector. For the associated production of $b/\bar{b}$ with the Higgs boson at high $p_T$, the SM electroweak corrections of 3, involving the large top quark mass, is comparable to the the QCD corrections of . On the other hand, the associated production of light quarks and anti–quarks with the Higgs boson is not significant because no large fermion mass is involved.
Once the Higgs boson is discovered, it is important to test whether it is a Standard Model Higgs boson or some other non-standard scalar particle. The cross section of the Higgs boson production at large transverse momentum can be sensitive to new physics which modify either $t$-$b$-$W$, $t$-$t$-$H$, or $W$-$W$-$H$ vertices. Among them, only the $t$-$t$-$H$ vertex can modify the contributions. In contrast, all of them can modify the 3 contributions. As illustrated in Table II, it is possible that the 3 rate is enhanced by more than a factor of 2 due to new physics effects. Therefore, 3 contributions should also be included when testing SM predictions.
Acknowledgments {#acknowledgments .unnumbered}
=================
C.-P.Y. thanks A. Abbasabadi, D. Bower-Chao, and W. Repko for useful discussions. S.M. was supported in part by DOE grant DE–FG03–92–ER40701. The work of C.P.Y. was supported in part by NSF grant No. PHY-9309902.
Appendix I: Loop Integration {#appendix-i-loop-integration .unnumbered}
==============================
In calculating the helicity amplitudes, one must evaluate loop integrals of the form $$\begin{aligned}
X \equiv \frac{1}{i\pi^2} \int_{}^{}
\!\!\frac{d^n Q\{1,Q^{\mu},Q^{\mu}Q^{\nu}\}}
{(Q^2 - m_1^2)((Q+P)^2 - m_2^2)\cdots}.\end{aligned}$$ For triangle diagrams, $X=C$, and for box diagrams, $X=D$.
Triangle Diagrams {#triangle-diagrams .unnumbered}
-----------------
The scalar function for triangle diagrams, showing explicitly its dependent variables, is: $$\begin{aligned}
C_0(m_1^2,m_2^2,m_3^2,p_1^2, p_2^2,p_3^2) =
\frac{1}{i\pi^2} \int
\!\!\frac{d^n Q}{(Q^2 - m_1^2)[(Q+p_1)^2 - m_2^2]
[(Q+p_1+p_2)^2 - m_3^2]},\end{aligned}$$ where the internal line masses $m_i$ are labelled by the external lines, $p_1$ is the momentum flowing between the lines with masses $m_1$ and $m_2$, $p_2$ between $m_2$ and $m_3$, and $p_3 = -p_1-p_2$ between $m_3$ and $m_1$. The vector integral over $Q^\mu$ is $$\begin{aligned}
C_{11} p_1^{\mu}+C_{12} p_2^{\mu}.\end{aligned}$$ Similary, the tensor integral over $Q^{\mu}Q^{\nu}$ is $$\begin{aligned}
C_{21}p_1^{\mu}p_1^{\nu}+C_{22}p_2^{\mu}p_2^{\nu}+
C_{23}\{p_1^{\mu}p_2^{\nu}+p_1^{\nu}p_2^{\mu}\}+C_{24}g^{\mu\nu}.\end{aligned}$$
Box Diagrams {#box-diagrams .unnumbered}
------------
The scalar function for box diagrams, similar to the triangle diagrams, can be written as $D_0(m_1^2,m_2^2,m_3^2,m_4^2,
p_1^2,p_2^2,p_3^2,p_4^2,(p_1+p_2)^2,(p_2+p_3)^2)$. The notation is an obvious generalization of that for the triangle diagrams. The vector integral over $Q^\mu$ is $$\begin{aligned}
D_{11} p_1^{\mu}+D_{12} p_2^{\mu}+D_{13} p_3^{\mu}.\end{aligned}$$ The tensor integral over $Q^{\mu}Q^{\nu}$ is $$\begin{aligned}
& D_{21}p_1^{\mu}p_1^{\nu}+D_{22}p_2^{\mu}p_2^{\nu}
+D_{23}p_3^{\mu}p_3^{\nu}+& \\
& D_{24}\{p_1^{\mu}p_2^{\nu}+p_1^{\nu}p_2^{\mu}\}+
D_{25}\{p_1^{\mu}p_3^{\nu}+p_1^{\nu}p_3^{\mu}\}+
D_{26}\{p_2^{\mu}p_3^{\nu}+p_2^{\nu}p_3^{\mu}\}
+D_{27}g^{\mu\nu}. &\end{aligned}$$
Appendix II: Form Factors {#appendix-ii-form-factors .unnumbered}
==========================
In ’t Hooft-Feynman gauge, there are 20 Feynman diagrams containing $W^\pm$–bosons and $\phi^\pm$ Goldstone bosons involved in the process $g+b/\bar{b}\ra b/\bar{b}+H$ at 3 for $m_b = 0$. As discussed in the text, the 3 diagrams involving $Z^0$–bosons and $\phi^0$ Goldstone bosons are negligible. Some typical diagrams are shown in Figure 1. In this appendix, we list the individual contributions to the form factors \[cf. Eq.(\[eqc0\])\] from each Feynman diagram for the process $g(q_g)+b(q_3)\ra b(q_1)+H(q_2)$. All momenta are defined pointing [*in*]{} to the Feynman diagram, i.e. the outgoing quark $(q_1)$ and Higgs boson $(q_2)$ four momenta have a negative energy component. There are 12 triangle diagrams, with terms labelled ${\cal F}_9-{\cal F}_{20}$, and 8 box diagrams, with terms ${\cal F}_1-{\cal F}_8$ and ${\cal G}^\mu_1-{\cal G}^\mu_8$. The full form factors are ${\cal F} = \sum_{i=1}^{20}{\cal F}_i$ and ${\cal G}^\mu = \sum_{i=1}^{8}{\cal G}^\mu_i$. The following expressions contain the invariant masses $s_{ij}=(q_i+q_j)^2$ and $s=(q_1+q_2+q_3)^2=q_g^2$. We have used the relation $M_W = \frac{1}{2}g v$ to re–express the electroweak coupling constants in terms of masses and the vacuum expectation value $v$. The process involving light quarks in the initial and final state can be deduced by setting $m_t = 0$. The limit $g,g^\prime\ra 0$, which we take to study the electroweak chiral Lagrangian, is obtained by eliminating all terms with an explicit $M_W$ dependence. All triangle diagrams contain progagators for one fermion and two gauge or Goldstone bosons. The box diagrams fall into two categories, those containing two fermion and two gauge or Goldstone boson propagators (denoted $t$-$t$-$W$-$W$) and those containing three fermion and one gauge or Goldstone boson propagators (denoted $t$-$t$-$t$-$W$).
Triangle Diagrams {#triangle-diagrams-1 .unnumbered}
-----------------
$$\begin{aligned}
{\cal F}_9 & = & -8 C_{12} M_W^4/v^{3} \\
{\cal F}_{10} & = & -2 C_{12} m_t^2 M_H^2/v^{3} \\
{\cal F}_{11} & = & 2 m_t^2 M_W^2 (-2 C_{00} - C_{12})/v^{3} \\
{\cal F}_{12} & = & -2 m_t^2 M_W^2 (-C_{00} + C_{12})/v^{3}\\
{\cal F}_{13} & = & -8 C_{12} M_W^4/v^{3} \\
{\cal F}_{14} & = & -2 C_{12} m_t^2 M_H^2/v^{3}\\
{\cal F}_{15} & = & 2 m_t^2 M_W^2 (C_{00} - C_{12})/v^{3}\\
{\cal F}_{16} & = & -2 m_t^2 M_W^2 (2 C_{00} + C_{12})/v^{3}\\
{\cal F}_{17} & = & 2 m_t^2 M_W^2 (-2 C_{00} - 4 C_{11})/v^{3}\\
{\cal F}_{18} & = & 2 m_t^4 (-C_{00} - 2 C_{11})/v^{3}\\
{\cal F}_{19} & = & 2 m_t^2 M_W^2 (-2 C_{00} - 4 C_{11})/v^{3}\\
{\cal F}_{20} & = & 2 m_t^4 (-C_{00} - 2 C_{11})/v^{3}\end{aligned}$$
Note that there are no tensor contributions from the triangle diagrams, i.e. no $C_{2i}$.
Box Diagrams {#box-diagrams-1 .unnumbered}
------------
$$\begin{aligned}
\mbox{$t$-$t$-$W$-$W$}~ diagrams & \\
{\cal F}_1 & = & -8 M_W^4 (2 D_{27} - D_{00} m_t^2 + (D_{22}-D_{24}
+ D_{25}- D_{26})M_H^2 \\
& & + (D_{11}+D_{25})s + (D_{24} - D_{25})s_{12}
- (D_{11}+ D_{12}- D_{25}+ D_{26})s_{23})/v^{3} \\
{\cal F}_2 & = &
-2 m_t^2 M_H^2 (2 D_{27} - D_{00} m_t^2 + (D_{22} - D_{24}
+ D_{25} - D_{26})M_H^2 + \\
& & (D_{13} + D_{25})s + (D_{12} - D_{13}
+ D_{24} - D_{25})s_{12} - (D_{25} - D_{26}) s_{23})/v^{3} \\
{\cal F}_3 & = & -2 m_t^2 M_W^2 (-2 D_{27} + D_{13} s +
(2 D_{00} + D_{12}- D_{13})s_{12})/v^{3} \\
{\cal F}_4 & = &
2 m_t^2 M_W^2 (2 D_{27} + (D_{00}+D_{11})s +(- 2 D_{00}- D_{11}
+ D_{12})s_{23})/v^{3} \\
{\cal G}^{\mu}_1 & = & -16 M_W^4 ((-D_{11}-
D_{24})q_1^{\mu} +(- D_{12}- D_{22})q_2^{\mu}
+(- D_{12} - D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_2 & = &
-4 m_t^2 M_H^2 ((-D_{12} - D_{24})q_1^{\mu}
+(- D_{12} - D_{22})q_2^{\mu} +(- D_{13} -
D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_3 & = &
-4 m_t^2 M_W^2 ((-2 D_{00} - 2 D_{11} -
D_{12} - D_{24})q_1^{\mu} \\
& & +(- 2 D_{00} -
3 D_{12} - D_{22})q_2^{\mu} +(- 3 D_{13}-
D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_4 & = &
4 m_t^2 M_W^2 ((-2 D_{00}- 2 D_{11} +
D_{12} + D_{24})q_1^{\mu} \\
& & +(- D_{12}+ D_{22})q_2^{\mu} +(- D_{13} + D_{26})q_3^{\mu})/v^{3} \\
\mbox{$t$-$t$-$t$-$W$}~ diagrams \\
{\cal F}_5 & = & 4 m_t^2 M_W^2 (D_{00} m_t^2 +(- D_{22} + D_{24}
- D_{25} + D_{26})M_H^2 - (2 D_{11}+ D_{25}) s \\
& &+(- D_{24}+ D_{25})s_{12}
+(- D_{00}+ 2 D_{11} - 2 D_{12} + D_{25} - D_{26})s_{23})/v^{3} \\
{\cal F}_6 & = & 2 m_t^4 (D_{00} m_t^2 +
(- D_{22} + D_{24} - D_{25} + D_{26})M_H^2 - (2 D_{13} s + D_{25}) s \\
& & +(- D_{00} - 2 D_{12} + 2 D_{13} - D_{24} + D_{25})s_{12}
+ (D_{25} - D_{26})s_{23})/v^{3} \\
{\cal F}_7 & = & 4 m_t^2 M_W^2 (D_{00} m_t^2 +
(- D_{22}+ D_{24}- D_{25}+ D_{26})M_H^2 \\
& & + (D_{00}- D_{11}+ D_{13}- D_{25})s +(- D_{00}
+ D_{12}- D_{13}- D_{24}+ D_{25})s_{12} \\
& & + (D_{11}- D_{12}+ D_{25} - D_{26})s_{23})/v^{3} \\
{\cal F}_8 & = & 2 m_t^4 (D_{00} m_t^2 +(-D_{22}+ D_{24}- D_{25}
+ D_{26})M_H^2 \\ & & + (D_{00}+ D_{11}- D_{13}- D_{25})s
+(- D_{12}+ D_{13}- D_{24} + D_{25})s_{12} \\
& &+(- D_{00}- D_{11}+ D_{12}
+ D_{25}- D_{26})s_{23})/v^{3} \\
{\cal G}^{\mu}_5 & = &
8 m_t^2 M_W^2 ((3 D_{11} +
2 D_{24})q_1^{\mu} + (D_{00} + 3 D_{12} +
2 D_{22})q_2^{\mu} \\
& & + (D_{00} + 2 D_{12} +
D_{13} + 2 D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_6 & = & 4 m_t^4 ((D_{00}+ D_{11}+
2 D_{12}+ 2 D_{24})q_1^{\mu} + (D_{00} +
3 D_{12} + 2 D_{22})q_2^{\mu} \\
& & + (3 D_{13} + 2 D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_7 & = & 8 m_t^2 M_W^2 ( (D_{11} +
2 D_{24})q_1^{\mu} + (D_{12} + 2 D_{22})q_2^{\mu} \\
& & +(- D_{00} + 2 D_{12}- D_{13}+
2 D_{26})q_3^{\mu})/v^{3} \\
{\cal G}^{\mu}_8 & = & 4 m_t^4 ((-D_{00} - D_{11}+
2 D_{12} + 2 D_{24})q_1^{\mu} + (D_{12} +
2 D_{22})q_2^{\mu} \\
& & + (D_{13} + 2 D_{26})q_3^{\mu})/v^{3}\end{aligned}$$
[99]{} F. Abe [*et al.*]{}, Phys. Rev. Lett. [**73**]{}, 225 (1994);\
S. Abachi [*et al.*]{}, Phys. Rev. Lett. [**72**]{}, 2138 (1994). P. Darriulat, conference summary of the 27th International Conference on High Energy Physics, Glasgow, 1994 (unpublished). J.F. Gunion, H.E. Haber, G.L. Kane, S. Dawson, [**The Higgs Hunter’s Guide**]{}, Addison–Wesley, 1990. P. Giacomelli, CERN-PPE-93-18, 2nd Trieste Conference on Recent Development in the Phenomenology of Particle Physics (1993). A. Stange, W. Marciano, and S. Willenbrock, Phys. Rev. [**D49**]{} (1994) 1354;\
J.F. Gunion and T. Han, Phys. Rev. [**D51**]{} (1995) 1051;\
S. Mrenna and G.L. Kane, hep-ph/9406337, 1994. R. Arnowitt and P. Nath, Phys. Rev. Lett. [**70**]{} (1993) 3696. J.L. Lopez, D.V. Nanopoulos, H. Pois, X. Wang, and A. Zichichi, Phys. Lett. [**B306**]{} (1993) 73. G.L. Kane, C. Kolda, L. Roszkowski, and J.D. Wells, Phys. Rev. [**D49**]{} (1994) 6173. G.L. Kane, C. Kolda and J.D. Wells, Phys. Lett. [**70**]{} (1993) 2686. R. Chivukula, M. Dugan, and M. Golden, Phys. Lett. [**B336**]{} (1994) 62. R.K. Ellis, I. Hinchliffe, M. Soldate, and J.J. van der Bij, Nucl. Phys. [**B297**]{} (1988) 221. G.J. van Oldenborgh, “FF, A Package to Evaluate One–Loop Feynman Diagrams”, NIKHEF–H/90–15, (1990). H.U. Bengtsson and T. Sjöstrand, Computer Physics Commun. [**43**]{} (1987) 43. E. Malkawi and C.–P. Yuan, Phys. Rev. [**D50**]{} (1994) 4462;\
C.-P. Yuan, “Top Quark Physics,” to appear in the proceedings of the VIth Mexican School of Particles and Fields, Villahermosa, Mexico, 3–7 October, 1994.
[^1]: [email protected]
[^2]: [email protected]
[^3]: The contribution does not depend on the electroweak gauge couplings, but is only sensitive to the coupling of the top quark to the Higgs boson.
[^4]: The 2–component Weyl spinors are defined by the relation $u_\pm={\frac{1}{2}}(1\pm\gamma_5)u$, etc.
[^5]: The superscript T denotes taking the transpose
[^6]: $\sigma_i$ are Pauli matrices satisfying Tr($\sigma_i\sigma_j$) = $2\delta_{ij}$.
[^7]: In our numerical results, we include all the contributions.
[^8]: The rate at 2 TeV ($\bar{{\rm p}} {\rm p}$) is about two orders of magnitude smaller than that at 14 TeV (${\rm p}{\rm p}$).
[^9]: We note that the tree level amplitude and the higher order amplitudes have different helicity structure in the $m_b\ra 0$ limit.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider quantum fidelity between two states $\rho$ and $\sigma$, where we fix $\rho$ and allow $\sigma$ to be sent through a quantum channel. We determine the minimal fidelity where one minimizes over (a) all unital channels, (b) all mixed unitary channels, and (c) arbitrary channels. We derive results involving the minimal eigenvalue of $\rho$, which we can interpret as a convex combination coefficient. As a consequence, we give a new geometric interpretation of the minimal fidelity with respect to the closed, convex set of density matrices and with respect to the closed, convex set of quantum channels. We further investigate the geometric nature of fidelity by considering density matrices arising as normalized projections onto subspaces; in this way, fidelity can be viewed as a geometric measure of distance between two spaces. We give a connection between fidelity and the canonical (principal) angles between the subspaces.'
address:
- 'Department of Mathematics & Computer Science, Brandon University, Brandon, MB, Canada R7A 6A9'
- 'Department of Mathematics & Statistics, University of Guelph, Guelph, ON, Canada N1G 2W1'
- 'Department of Mathematics & Computer Science, Brandon University, Brandon, MB, Canada R7A 6A9'
author:
- Jin Li
- Rajesh Pereira
- Sarah Plosker
title: Some geometric interpretations of quantum fidelity
---
Introduction
============
The quantum fidelity $F(\rho, \sigma)$ is a measure of the distance between two quantum states $\rho$ and $\sigma$ that quantifies the accuracy of state transfer through a channel; the ideal case being a fidelity value of 1, which represents perfect state transfer. Physically, one begins with initial state $\rho$ at time 0, and allows the quantum system to evolve over time. At time $t$, one measures the overlap of the two states $\rho$ and $\sigma$; this overlap decreases over time due to the evolution and perturbation of the system.
Fidelity has been considered in the context of quantum communication via unmeasured and unmodulated spin chains which are used to transmit quantum states [@Bose], and plays a role in quantum decision tree algorithms [@LB14]. Geometric interpretations of fidelity have been given in [@MMPZ08; @MZC08] and elsewhere, however our approach and results are distinctly different from the literature at present.
Formally, we have the following definition:
\[defn:fidelity\] Let $\rho$ and $\sigma$ be two $n\times n$ positive semidefinite matrices. The *(quantum) fidelity* between $\rho$ and $\sigma$ is $$\begin{aligned}
F(\rho, \sigma)&=&{\operatorname{Tr} }(\sqrt{\sqrt{\rho}\sigma\sqrt{\rho}})\\
&=&{\operatorname{Tr} }(\sqrt{\sqrt{\sigma}\rho\sqrt{\sigma}}).\end{aligned}$$
Any positive semidefinite matrix has a unique positive square root, and so the quantum fidelity between two quantum states (density matrices) is well-defined and yields a non-negative real number.
The transition probability between two states, which is the square of the fidelity, was defined in [@U76], although the idea stems from two earlier papers [@K48; @B69] in a more general context. Jozsa [@J94] proposed four axioms that the transition probability (which he called fidelity) must satisfy (we have re-written these axioms in terms of fidelity)[^1]
1. $0\leq F(\rho, \sigma)\leq 1$ with $F(\rho, \sigma)=1$ iff $\rho=\sigma$;
2. The fidelity is symmetric: $F(\rho, \sigma)=F(\sigma, \rho)$;
3. If $\rho=|\psi\rangle\langle\psi|$ is a pure state, then $F(\rho, \sigma)=\sqrt{{\langle \psi |}\sigma{| \psi \rangle}}$;
4. \[ax4\] The fidelity is invariant under unitary transformations on the state space: $$F(U\rho U^\dagger, U\sigma U^\dagger)=F(\rho, \sigma) \quad \textnormal{ for any unitary } U,$$ where $^\dagger$ represents complex conjugate transposition.
Note that, if both $\rho=|\psi\rangle\langle\psi|$ and $\sigma=|\phi\rangle\langle\phi|$ are pure states, then the quantum fidelity $F(\rho, \sigma)$ reduces to $|\langle \psi|\phi\rangle|$. Supposing ${| \psi \rangle}$ and ${| \phi \rangle}$ are unit vectors in $\R^n$ (rather than in $\C^n$), we can write ${| \psi \rangle}=(\sqrt{p_1}, \dots, \sqrt{p_n})$ and ${| \phi \rangle}=(\sqrt{q_1}, \dots, \sqrt{q_n})$, where $p=\{p_j\}$ and $q=\{q_j\}$ are two probability distributions. This then yields the classical fidelity between the two probability distributions $p=\{p_j\}$ and $q=\{q_j\}$, which is defined as $F(p, q)=\sum_j\sqrt{p_jq_j}$. The term classical fidelity is used in quantum information theory; outside of QIT, classical fidelity is referred to as the *Bhattacharyya coefficient*.
The quantum fidelity between unitary orbits of two density matrices $\rho$ and $\sigma$ is $F(V\rho V^\dagger, W\rho W^\dagger)$ for unitary $V, W$, which, in light of axiom (\[ax4\]), reduces to $F(\rho, U\sigma U^\dagger)$ for unitary $U$. The maximum and minimum quantum fidelity between the unitary orbits of $\rho$ and $\sigma$ have been characterized as follows:
[@MMPZ08; @ZF14]\[thm:ZF\] The quantum fidelity between unitary orbits ${\mathcal }U_\rho$ and ${\mathcal }U_\sigma$ satisfies the following relations: $$\begin{aligned}
\max_{U\in U(H_d)}F(\rho, U\sigma U^\dagger)&=&F(\lambda^\downarrow(\rho), \lambda^{\downarrow}(\sigma))\\
\min_{U\in U(H_d)}F(\rho, U\sigma U^\dagger)&=&F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma)),\end{aligned}$$ where $U(H_d)$ is the set of all $d\times d$ unitary matrices on a $d$-dimensional Hilbert space $H_d$ and $\lambda^{\downarrow}(\rho)$ (respectively, $\lambda^{\uparrow}(\rho)$), is the vector of eigenvalues of $\rho$, listed in non-increasing (respectively, non-decreasing) order, including multiplicities.
The quantum fidelity $F(\rho, \sigma)$ was originally found to satisfy the bounds of theorem \[thm:ZF\] in [@MMPZ08], although the result found in [@MMPZ08] was formulated in terms of the closely related Bures distance.
For practical purposes, one wishes to maximize $F(\rho, \sigma)$; however, it is also useful to consider minimal fidelity, which represents the worst-case scenario of quantum information state transfer.
We generalize theorem \[thm:ZF\] by characterizing the following minimum quantum fidelities:
1. the minimum quantum fidelity $F(\rho, \Phi(\sigma))$, where $\Phi$ is any quantum channel (completely positive, trace-preserving, linear map),
2. the minimum quantum fidelity $F(\rho, \Phi(\sigma))$, where $\Phi$ is any unital channel (a quantum channel satisfying $\Phi(I)=I$), and
3. the minimum quantum fidelity $F(\rho, \Phi(\sigma))$, where $\Phi$ is any mixed unitary channel (a quantum channel of the form $\Phi(\rho)=\sum_ip_iU_i\rho U_i^\dagger$, where $U_i$ are unitaries and $p_i$ form a probability distribution).
In our derivations, we take the point of view that $\rho$ is fixed (given) and $\sigma$ is sent through the channel $\Phi$. Our motivation is the case where one has access to the output $\Phi(\sigma)$ of the state $\sigma$ after it has been sent through a channel $\Phi$, but one does not have direct access to $\sigma$. Thus, it is of interest to see how far away $\rho$ and $\sigma$ can become through the use of the channel $\Phi$.
We will show (in corollary \[cor:min\]) that if $\Phi$ is a unital channel, then the quantity $\min_{\Phi}F(\rho, \Phi(\sigma))$, where the minimum is taken over all quantum channels $\Phi$, reduces to $F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma))=\min_{U\in U(H_d)}F(\rho, U\sigma U^\dagger)$ from theorem \[thm:ZF\].
Our methods make extensive use of majorization of vectors of eigenvalues. It is interesting to note that majorization is also used to characterize the more specialized situation when one entangled state can be transformed into another through the use of quantum operations described by local operations and classical communication (LOCC) [@Nie99].
The paper is organized as follows. In section \[sec:qf\] we review majorization, the main tool used in proving the results of this section, and we derive the minimum $F(\rho, \Phi(\sigma))$ where $\Phi$ is a quantum channel, under the restrictions listed above. In section \[sec:geom\], we give a geometric interpretation of our results. In section \[sec:proj\] we continue our geometric approach, this time focusing on density matrices arising as normalized projections onto subspaces. It appears that studying fidelity in terms of projections and subspace geometry has not been done previously. We obtain a number of new interpretations of fidelity, including theorem \[thm:Fcanangles\], which links fidelity with the canonical (principal) angles between the subspaces. Section \[sec:dis\] is devoted to a discussion on various related topics, linking our work with related results on fidelity as well as results in other areas of mathematics.
Quantum fidelity when one state is sent through a quantum channel {#sec:qf}
=================================================================
Majorization
------------
\[defn:maj\] Let $x=(x_{1},x_{2},...,x_{d})$ and $y=(y_{1},y_{2},...,y_{d})$ be two $d$-tuples of real numbers. We say that $(x_{1},x_{2},...,x_{d})$ is majorized by $(y_{1},y_{2},...,y_{d})$, written $x\prec y$, if $$\begin{aligned}
\sum_{j=1}^{k}x^{\downarrow}_{j}\leq \sum_{j=1}^{k}y^{\downarrow}_{j}\quad 1\leq k \leq d,\end{aligned}$$ with equality for $k=d$.
If equality does not necessarily hold when $k=d$, we say that $x$ is *sub-majorized* by $y$ and we write $x\prec_w y$, where the $w$ stands for “weak”.
If we order the components of the vectors in *non-decreasing* order, indicated by $^\uparrow$, then $x$ is majorized by $y$ if $$\begin{aligned}
\sum_{j=1}^k x^{\uparrow}_j \geq \sum_{j=1}^k y^{\uparrow}_j\quad 1\leq k \leq d,\end{aligned}$$ with equality when $k=d$. This definition is equivalent to the definition of majorization given above. If equality does not necessarily hold when $k=d$, we say that $x$ is *super-majorized* by $y$ and we write $x\prec^w y$.
Minimum Fidelity
----------------
A function $f: \R^n\rightarrow \R$ is *Schur-concave* if $x\prec y\Rightarrow f(x)\geq f(y)$.
Although there are several ways of proving proposition \[prop:Schur-concave\], we shall prove it using Ostrowski’s theorem:
[@Ost52], [@MOA11 Theorem 3.A.7] Let $D=\{(x_1, \dots, x_n)\,|\, x_1\geq \cdots \geq x_n\}$. Let $\phi$ be a real-valued function defined and continuous on $D$. Then $$\phi(x)\leq \phi(y)\textnormal{ whenever } x\prec^w y\textnormal{ on } D$$ if and only if $$0\geq \phi_1(z)\geq \cdots \geq \phi_n(z)\,\,\forall z\textnormal{ in the interior of } D,$$ where $\phi_i(z)=\frac{\partial \phi(z)}{\partial z_i}$.
\[prop:Schur-concave\] Let $\{p_j \}_{j=1}^{n}$ be fixed non-negative numbers that sum to one. The function $f(q_1,...,q_n)=\sum_j\sqrt{p_j^{\uparrow}q_j^\downarrow}$ is Schur-concave.
We know that the square root function is concave, so the sum of the square root functions acting on each of the components $q_1, \dots, q_n$ is Schur-concave.
Now, consider $\phi(z)=\sum_i\sqrt{p_i}\sqrt{z_i}$ where $z\in D$ so $z_1\geq \cdots \geq z_n$. In the minimum case, want $p_1\leq \cdots \leq p_n$. We find that $\phi_i(z)=\frac12\frac{\sqrt{p_i}}{\sqrt{z_i}}$, which increases as $i$ increases. Thus $-\phi$ satisfies Ostrowski’s theorem. So if $x\prec^w y$ then $\phi(x)\geq \phi(y)$; that is, the function $f(q_1,...,q_n)=\sum_j\sqrt{p_j^{\uparrow}q_j^\downarrow}$ is Schur-concave. It follows that the absolute minimum of $\phi$ over any subset $S$ of $D$ if it exists must be at a point of $D$ which is maximal with respect to the supermajorization order.
We can use the theory of majorization to find the minimum fidelity between a fixed state $\rho$ and $\Phi(\sigma)$, the image of a second fixed state under any unital quantum channel. We note that this result, while related to theorem \[thm:ZF\], is not a direct consequence of it since there exist unital quantum channels which are not the convex combination of unitary transforms [@LS90].
\[cor:min\] If we consider unital $\Phi$, then we have $$\min_{\Phi}F(\rho, \Phi(\sigma))=F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma))$$ where the minimum is taken over all possible unital quantum channels $\Phi$.
Suppose $\rho$ and $\sigma$ are density matrices and $\Phi$ is a quantum channel. Then by Uhlmann’s theorem, $\Phi(\sigma)\prec \sigma$, provided $\Phi$ is unital. Thus $f(\Phi(\sigma))\geq f(\sigma)$ for all Schur-concave functions $f$. In particular, we take the $f$ from proposition \[prop:Schur-concave\] with $p_j=\lambda_j(\rho)$ to obtain $$\begin{aligned}
\sum_j\sqrt{\lambda_j^\uparrow(\rho)\lambda_j^\downarrow(\Phi(\sigma))}&\geq &\sum_j\sqrt{\lambda_j^\uparrow(\rho)\lambda_j^\downarrow(\sigma)}\\
\textnormal{i.e.\ } F(\lambda_j^\uparrow(\rho), \lambda_j^\downarrow(\Phi(\sigma)))&\geq &F(\lambda_j^\uparrow(\rho),\lambda_j^\downarrow(\sigma))\\
\textnormal{equivalently } F(\lambda_j^\downarrow(\rho), \lambda_j^\uparrow(\Phi(\sigma)))&\geq &F(\lambda_j^\downarrow(\rho),\lambda_j^\uparrow(\sigma)).\end{aligned}$$ Thus $\min_{\Phi}F(\rho, \Phi(\sigma))$ (where the minimum is restricted to unital $\Phi$) is achieved precisely when $\Phi$ is the unitary transformation making the eigenvalues of $\sigma$ the same as those of $\rho$, with the eigenvalues lining up in the opposite direction, giving: $$\begin{aligned}
\min_{\Phi}F(\rho, \Phi(\sigma))=F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma)).\end{aligned}$$
\[prop:minF\] Let $H$ be a Hilbert space and $S(H)$ be the state space of $H$. Let $\rho\in S(H)$ and $K$ be a subset of $S(H)$ containing all of the pure states in $S(H)$. Then $\min_{\sigma \in K}F(\rho, \sigma)=(\lambda_{\min}(\rho))^{1/2}$, where $\lambda_{\min}(\rho)$ represents the minimal eigenvalue of $\rho$.
Any mixed state $\sigma$ can be represented as a convex combination of pure states: $\sigma=\sum_ip_i{| \psi_i \rangle}{\langle \psi_i |}$. The quantum fidelity is concave in each of its variables [@U76; @J94], and so $F(\rho, \sigma )\geq \sum_i p_iF(\rho , {| \psi_i \rangle}{\langle \psi_i |})$. Hence $\sqrt{{\langle \psi_0 |}\rho{| \psi_0 \rangle}}=F(\rho, {| \psi_0 \rangle}{\langle \psi_0 |})\leq F(\rho, \sigma)$ for at least one of the pure states ${| \psi_0 \rangle}$. By the Courant-Fisher theorem, we minimize $\sqrt{{\langle \psi_0 |}\rho{| \psi_0 \rangle}}$ as a function of ${| \psi_0 \rangle}$ by choosing ${| \psi_0 \rangle}$ to be the eigenvector corresponding to the minimal eigenvalue of $\rho$ which gives us $F(\rho, \sigma)\geq \sqrt{{\langle \psi_0 |}\rho{| \psi_0 \rangle}}=(\lambda_{\min}(\rho))^{1/2}$. Since ${| \psi_0 \rangle}{\langle \psi_0 |}$ is a pure state, it is in $K$ and our result follows.
As a corollary we have the following result:
\[prop:minev\] We have $$\min_{\Phi}F(\rho, \Phi(\sigma))=(\lambda_{\min}(\rho))^{1/2}$$ where the minimum on the left hand side of the equation is taken over all possible quantum channels $\Phi$.
Let $K=\{\Phi(\sigma): \Phi$ is a quantum channel$\}$. We note that $K$ contains all pure states: To see this consider $F(\rho, \Phi(\sigma))$ and take $\Phi(\cdot)={\operatorname{Tr} }(\cdot){| \psi \rangle}{\langle \psi |}$. The map $\Phi$ is clearly completely positive, trace-preserving, and linear, and so $\Phi$ is a quantum channel and ${| \psi \rangle}{\langle \psi |}\in K$. The result now follows from the previous proposition.
If we choose $\Phi$ to be the quantum channel $\Phi(\cdot)={\operatorname{Tr} }(\cdot)\rho$, then this choice gives us $F(\rho,\Phi(\sigma))=F(\rho, \rho)=1$, so the maximum value of $F(\rho, \Phi(\sigma))$ is one. It is therefore trivial to find the maximum of $F(\rho, \Phi(\sigma))$ when $\Phi$ is any quantum channel. The problem becomes interesting when we restrict to unital or to mixed unitary channels, since in these special cases we no longer have complete freedom. However, we do not have results such as proposition \[prop:Schur-concave\] and the Courant-Fisher theorem at our disposal, so finding the maximum is not a straightforward task.
If we consider mixed unitary channels $\Phi$ (all channels of the form $\Phi(\cdot)=\sum_jp_jU_j(\cdot)U_j^\dagger$ where $\{p_j\}$ is a probability distribution and $U_j$ are unitaries), then $$\min_{\Phi}F(\rho, \Phi(\sigma))=F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma))$$ where the minimum is taken over all mixed unitary channels $\Phi$.
The concavity of the quantum fidelity gives us that the minimum of $F(\rho, \Phi(\sigma))$ will occur at an extreme point $\Phi$ of the set of quantum channels. Thus, if we are considering the set of mixed unitary channels, then the minimum must occur at a unitary channel: a channel of the form $\Phi(\cdot)=U(\cdot)U^\dagger$. Thus, by theorem \[thm:ZF\], it follows that $$\min_{\Phi}F(\rho, \Phi(\sigma))=F(\lambda^\downarrow(\rho), \lambda^{\uparrow}(\sigma))$$ where the minimum is taken over all mixed unitary channels $\Phi$.
Again we stress that the maximum value of the fidelity could potentially occur at any point, so finding the maximum is a much more difficult matter.
Geometric Interpretation of Minimum Quantum Fidelity {#sec:geom}
====================================================
The set of all states is a compact convex set. At its center is the maximally mixed state $\frac1nI$; its boundary is made up of all singular (non-invertible) density matrices $\omega$. Any state $\rho$ can be written as a convex combination $$\rho=p\omega+(1-p)\left(\frac1nI\right)$$ for some $\omega$ on the boundary, where $0\leq p\leq 1$.
Similar convex combinations have been studied in quantum information theory, and in many other fields of mathematics and computer science. For instance, the set of all channels is a compact convex set. At its center is the completely depolarizing channel $\Omega:\rho\mapsto \frac1nI$; its boundary is made up of all channels $\Psi$ whose Choi matrix $C_{\Psi}$ is singular. Recall that the Choi matrix $C_{\Phi}$ corresponding to a channel $\Phi$ is defined by $$C_\Phi= \left (I_n\otimes\Phi \right ) \left (\sum_{ij}E_{ij}\otimes E_{ij} \right ) = \sum_{ij}E_{ij}\otimes\Phi(E_{ij}),$$ where $E_{ij}$ are the matrix units. The Choi matrix for a channel is singular precisely when the number of Kraus operators $V_i$ in the decomposition $\Phi(\rho)=\sum_{i=1}^k V_i\rho V_i^\dagger$ minimizing $k$ is strictly less than $n^2$. (Thus most channels that arise naturally are on the boundary of the set of all quantum channels). Any channel $\Phi$ can be written as a convex combination $$\begin{aligned}
\label{eq:Phi}
\Phi=p\Psi+(1-p)\Omega\end{aligned}$$ for some $\Psi$ on the boundary, where $0\leq p\leq 1$.
A more specific example along these lines is that of [@Wat09], where if one can write $$\Phi=p\Psi+(1-p)\Omega$$ for some unital quantum channel $\Psi$, where $0\leq p\leq \frac1{d^2-1}$, then $\Phi$ is a mixed unitary channel. Note here that the set of all mixed unitary channels forms a subset of the set of all unital channels, both sets are compact convex sets, and $\Omega$ is their common centroid.
A result along the same vein [@Mar05] gives $p$ for which $$\Phi=p\omega+(1-p)\left(\frac1nI\right)$$ is a real, rank-one correlation matrix, where $\omega$ is a real correlation matrix (a positive semi definite matrix with 1’s along the diagonal).
In [@OP13], the authors consider a similar convex combination problem involving $H$-unistochastic and bistochastic matrices.
Relating this to the results herein, the value of $\lambda_{\min}(\rho)$ tells us how close $\rho$ is to the maximally mixed state, or, equivalently, how close it is to the “extreme” states (singular density matrices). The value of $\lambda_{\min}(\rho)$ gets larger as $\rho$ gets closer to the maximally mixed state, and smaller as $\rho$ gets closer to the boundary of singular density matrices. In this way, $\min_{\sigma \in K}F(\rho, \sigma)$ of proposition \[prop:minF\] measures how far away your state $\rho$ is from the boundary.
Similarly, in the case of quantum channels, $\Phi$ maps $\sigma$ to a density matrix with larger and larger $\lambda_{\min}$ as $p\rightarrow 0$ in equation (\[eq:Phi\]).
Fidelity, projections and subspace geometry {#sec:proj}
===========================================
Let $S$ be an $m$-dimensional subspace of a $d$-dimensional Hilbert space $H_d$ and $P_{S}$ be the orthogonal projection onto the subspace $S$, then $\rho_{S}=\frac1m P_{S}$ is a density matrix. Let $T$ be an $n$-dimensional subspace of $H_d$. The main goal of this section is to examine the relationship between the geometry of two subspaces $S$ and $T$ and the quantity $F(\rho_{S},\rho_{T})$.
In the special case where $\rho_{S}$ and $\rho_{T}$ commute, we have $$\begin{aligned}
F(\rho_{S},\rho_{T})&=&{\operatorname{Tr} }\left(\sqrt{\left(\frac1m P_{S}\right)^{1/2}\frac1nP_{T}\left(\frac1m P_{S}\right)^{1/2}}\right)\\
&=&\frac1{(mn)^{1/2}}{\operatorname{Tr} }(\sqrt{P_SP_TP_S})\\
&=&\frac{{\operatorname{Tr} }\sqrt{P_{S\cap T}}}{\sqrt{m n}}\\
&=&\frac{\dim(S\cap T)}{\sqrt{mn}}\end{aligned}$$ so it appears that in this case the fidelity measures the proportion of overlap between the two subspaces, giving 0 when $S$ and $T$ are disjoint, and 1 when $S=T$.
We note that this result can be used to find the maximum and minimum of $F(\rho_{S},\rho_{T})$ where $S$ and $T$ range over all $m$-dimensional and $n$-dimensional subspaces of $H_d$, respectively. By theorem \[thm:ZF\], both the maximum and the minimum will occur at a choice of $S$ and $T$ for which $\rho_S$ and $\rho_T$ commute. Since $\max(m+n-d,0)\le \dim(S\cap T)\le \min(m,n)$, we get the following result.
Let $S$ and $T$ be subspaces of $H_d$ with dimension $m$ and $n$ respectively and let $\rho_S=\frac{1}{m}P_S$ and $\rho_T=\frac{1}{n}P_T$. Then $\frac{\max(m+n-d,0)}{\sqrt{mn}}\le F(\rho_{S},\rho_{T})\le \min(\sqrt{\frac{m}{n}},\sqrt{\frac{n}{m}})$.
We are interested in properties of the fidelity of two density matrices when one or both are normalized orthogonal projections. We have the following inequality.
\[prop:proj\] Let $S$ be an $m$-dimensional subspace of $H_d$. Let $\rho$ be a density matrix with eigenvalues $\lambda_1\geq \lambda_2\geq ...\geq \lambda_{d-1}\geq \lambda_d$. Then $\sum_{j=d-m+1}^{d}\sqrt{ \frac{\lambda_j}{m}}\le F(\rho, \rho_S)\le \sum_{j=1}^{m}\sqrt{\frac {\lambda_j}{m}}$. For any fixed $\rho$, there are choices of $S$ for which $F(\rho, \rho_S)$ achieves the upper and lower bounds of the inequality respectively.
Let $\mu_{1}\ge \mu_{2}\ge ...\ge \mu_{m}$ be the eigenvalues of $P_S\rho P_S$ when considered as an operator on $S$. By Cauchy’s interlacing theorem, we have $\lambda_{j+d-m} \le \mu_j \le \lambda_{j}$. Since $F(\rho, \rho_S)=\frac{1}{\sqrt{m}} tr((P_S\rho P_S)^{\frac{1}{2}})$, the result follows. We may attain the upper and lower bound of the inequalities by choosing $S$ to be the span of the eigenvectors corresponding respectively to the $m$ largest and $m$ smallest eigenvalues of $\rho$.
Proposition \[prop:minF\] follows as a corollary of proposition \[prop:proj\]. Indeed, pure states are rank-one projections, so the dimension of the set of all pure states $K$ is $m=1$. By concavity, the minimum will occur at the boundary of the state space $S(H)$, which is precisely the pure states. We thus obtain the lower bound of proposition \[prop:proj\]: $(\lambda_{\min}(\rho))^{1/2}$, which is precisely the result of proposition \[prop:minF\].
If the dimensions of $S$ and $T$ are equal, we can obtain an interesting interpretation of $F(\rho_{S},\rho_{T})$ as the average of the cosine of the canonical angles between $S$ and $T$. Before introducing our result, we remind the reader of the definition of the canonical angles between $S$ and $T$; this concept first appears in the work of Camille Jordan in 1875 [@J75].
Let $S$ and $T$ be finite dimensional subspaces of a Hilbert space $\mathcal{H}$ and let $\ell=\min\{\dim (S), \dim(T)\}$. Then the first canonical angle is the unique number $\theta_1\in [0,\frac{\pi}{2}]$ such that $\cos(\theta_1)=\max \{ \vert \langle x,y\rangle\vert: x\in S, y\in T, \Vert x\Vert=\Vert y\Vert=1\}$. Let $x_1$ and $y_1$ be unit vectors in $S$ and $T$ respectively where the previous maximum is attained. Then we define the second canonical angle as the unique number $\theta_2\in [0,\frac{\pi}{2}]$ such that $\cos(\theta_2)=\max \{ \vert \langle x,y\rangle\vert: x\in S, y\in T, \Vert x\Vert=\Vert y\Vert=1, x\perp x_1, y\perp y_1\}$. Let $x_2$ and $y_2$ be the unit vectors in $S$ and $T$ respectively where the previous maximum is attained. Now for any $k\le \ell$, $\theta_k$ is the unique number such that $\cos(\theta_k)=\max \{ \vert\langle x,y\rangle\vert: x\in S, y\in T, \Vert x\Vert=\Vert y\Vert=1, x\perp x_1, x_2, \dots x_{k-1} , y\perp y_1, y_2, \dots y_{k-1}\}$.
Canonical angles are also called principal angles. We will use a characterization of the canonical angles first given in [@BG73].
\[BG\] Let $S$ and $T$ be subspaces of a Hilbert space $H$ with dimensions $m$ and $n$ respectively, and let $Q_S$ and $Q_T$ be matrices whose column vectors are the elements of orthonormal bases of $S$ and $T$ respectively. Then the cosine of the canonical angles are the singular values of the matrix $Q_S^\dagger Q_T$: $$\cos(\theta_k)=\sigma^\downarrow_k(Q_S^\dagger Q_T),$$ for all $k=1, \dots, \ell=\min\{m, n\}$.
We are now ready to state our main result of this section.
\[thm:Fcanangles\] Let $S$ and $T$ be subspaces of a Hilbert space $H$ with dimensions $m$ and $n$ respectively, with $\ell=\min\{m, n\}$. Let $P_S$ and $P_T$ be the orthogonal projections onto $S$ and $T$ respectively and let $\rho_{S}=\frac1m P_{S}$ and $\rho_{T}=\frac1n P_{T}$. The fidelity $F(\rho_{S},\rho_{T})=\frac{1}{\sqrt{mn}}\sum_{k=1}^{\ell}\cos(\theta_k)$ where $\{ \theta_{k}\}_{k=1}^{\ell}$ are the canonical angles between $S$ and $T$.
Note that if $\dim(S)=\dim(T)$, then $F(\rho_{S},\rho_{T})$ is the arithmetic mean of the cosines of the canonical angles.
Note that $P_S=Q_SQ_S^\dagger $ and $P_T=Q_TQ_T^\dagger $ where $Q_S$ and $Q_T$ are any matrices whose column vectors are the elements of orthonormal bases of $S$ and $T$ respectively. Then $\rho_S^{\frac{1}{2}}\rho_T\rho_{S}^\frac{1}{2}$ is similar to $\frac{1}{mn}P_SP_T$ which is equal to $\frac{1}{mn}Q_SQ_S^\dagger Q_TQ_T^\dagger $ which has the same non-zero eigenvalues as $\frac{1}{mn}(Q_S^\dagger Q_T)(Q_S^\dagger Q_T)^\dagger $ with the same multiplicities. Therefore the non-zero eigenvalues of $(\rho_S^{\frac{1}{2}}\rho_T\rho_{S}^\frac{1}{2})^\frac{1}{2}$ are exactly the same as the non-zero singular values of $\frac{1}{\sqrt{mn}}Q_S^\dagger Q_T$. The result now follows from theorem \[BG\].
The Bures angle between two states $\rho$ and $\sigma$ is $\arccos (F(\rho, \sigma))$. We note here that if the density matrices are normalized orthogonal projections onto subspaces $S$ and $T$ of the same dimension, then the cosine of the Bures angle between the two states is the arithmetic mean of the cosines of the canonical angles.
Discussion {#sec:dis}
==========
In this section we discuss connections between the work herein and resutls found elsewhere in the mathematics and quantum information theory literature.
Rearrangement Inequality {#sec:rearr}
------------------------
Theorem \[thm:ZF\] is in fact a stronger version (in the sense that it deals with non-commutative operators) of the rearrangement inequality for non-negative numbers: $$\begin{aligned}
\label{eq:rearr}
x_ny_1+\cdots x_1y_n\leq x_{\sigma(1)}y_1+\cdots +x_{\sigma(n)}y_n\leq x_1y_1+\cdots x_ny_n\end{aligned}$$ for any choice of real numbers $$\begin{aligned}
\label{eq:xy}
x_1\leq \cdots\leq x_n\quad\textnormal{and}\quad y_1\leq \cdots\leq y_n\end{aligned}$$ and for any permutation $\sigma$ of $\{1, \dots, n\}$. If we have all strict inequalities in (\[eq:xy\]), then the lower bound of the inequality (\[eq:rearr\]) is attained only for the permutation that reverses the order, i.e. $\sigma(i)=n-i+1$ for all $i\in \{1, \dots, n\}$ and the upper bound is attained only for the identity $\sigma(i)=i$ for all $i\in \{1, \dots, n\}$.
Indeed, consider theorem \[thm:ZF\] under the special case where both density matrices $\rho$ and $\sigma$ are diagonal and the unitaries are permutations. With this setup, we restate theorem \[thm:ZF\] as $$\begin{aligned}
\max_{U\in U(H_d)}(F(\rho, U\sigma U^\dagger))^2&=&\sum_{i=1}^n\lambda_i(\rho)\lambda_i(\sigma),\end{aligned}$$ which is the upper bound of inequality (\[eq:rearr\]) and $$\begin{aligned}
\min_{U\in U(H_d)}(F(\rho, U\sigma U^\dagger))^2&=&\sum_{i=1}^n\lambda_i(\rho)\lambda_{n-i+1}(\sigma),\end{aligned}$$ which is the lower bound of inequality (\[eq:rearr\]), with $x_i=\lambda_i(\rho)$ and $y_i=\lambda_i(\sigma)$. All other permutation matrices $U$ just yield something in between these two bounds, thus giving inequality (\[eq:rearr\]).
The Spectral Geometric Mean
---------------------------
Let $A\# B=A^{1/2}(A^{-1/2}BA^{-1/2})^{1/2}A^{1/2}$ be the geometric mean between positive semidefinite matrices $A$ and $B$. In [@UC09], the authors show that for bipartite states $\rho$ and $\sigma$, the fidelity of Alice’s reduced states $\rho^A$ and $\sigma^A$ is related to the geometric mean of Bob’s reduced states $\rho^B$ and $\sigma^B$: $$F(\rho^A, \sigma^A)={\operatorname{Tr} }(\rho^B\# \sigma^B).
$$ Here we show that for general states $\rho$ and $\sigma$ (not necessarily bipartite), their fidelity is intimately related to their *spectral geometric mean*.
The spectral geometric mean between positive semidefinite matrices $A$ and $B$ is given by $A\diamond B =(A^{-1}\# B)^{1/2}A(A^{-1}\# B)^{1/2}$ [@FP97], which has the useful feature that $(A\diamond B)^2 $ is similar to $AB$.
We have, for positive semidefinite matrices $A$ and $B$, $$\begin{aligned}
{\operatorname{Tr} }(A\diamond B)&=&{\operatorname{Tr} }((A^{-1/2}(A^{1/2}BA^{1/2})^{1/2}A^{-1/2})^{1/2}A(A^{-1/2}(A^{1/2}BA^{1/2})^{1/2}A^{-1/2})^{1/2})\\
&=&{\operatorname{Tr} }(A^{-1/2}(A^{1/2}BA^{1/2})^{1/2}A^{-1/2}A)\\
&=&{\operatorname{Tr} }((A^{1/2}BA^{1/2})^{1/2}).
\end{aligned}$$ In particular, for density matrices $\rho$ and $\sigma$, we have $$\begin{aligned}
F(\rho, \sigma)={\operatorname{Tr} }(\rho\diamond \sigma)
\end{aligned}$$ with $\rho\diamond \sigma$ similar to $\sqrt{\rho\sigma}$ (since $(\rho\diamond \sigma)^2$ is similar to ${\rho\sigma}$). Re-writing the trace as a sum of eigenvalues, we have $$\begin{aligned}
F(\rho, \sigma)=\sum_i\lambda_i(\rho\diamond\sigma)=\sum_i\lambda_i(\sqrt{\rho\sigma}).
\end{aligned}$$
Maximum Fidelity
----------------
The *maximum output fidelity* of two channels $\Phi, \Psi$ is define as [@Ros09; @KW00] $$F_{\max}(\Phi, \Psi)=\max_{\rho, \sigma}F(\Psi(\rho), \Phi(\sigma)),$$ where the maximum is taken over all density matrices $\rho$ and $\sigma$. This maximum fidelity is connected to the diamond norm $\|\cdot\|_\diamond$, the dual of the completely bounded norm, via the following lemma
[@KW00] Let $\Phi,\Psi:\mathfrak{B}(H)\rightarrow\mathfrak{B}(K) $ be quantum channels with Stinespring dilations $$\begin{aligned}
\Phi(X)&=&{\operatorname{Tr} }_BUXU^\dagger\\
\Psi(X)&=&{\operatorname{Tr} }_BVXV^\dagger,\end{aligned}$$ where $U,V:H\rightarrow B\otimes K$ are unitaries. Let $\Gamma$ be the linear map given by $\Gamma(X)={\operatorname{Tr} }_KUXV^\dagger$. Then $F_{\max}(\Phi, \Psi)=\|\Gamma\|_\diamond$.
Taking the maximum allows one to interpret fidelity in terms of the diamond norm. This interpretation has been used in [@Ros09; @KW00] with respect to quantum interactive proof systems.
However, $F_{\max}(\Phi, \Psi)=1$ whenever the ranges of the two channels overlap. Thus, we propose $F_{\min}(\Phi, \Psi)$ as a more informative measure of distance between two channels, in the sense that it will only give 1 when the channels are equal, allowing for more useful comparisons between channels.
Acknowledgements {#acknowledgements .unnumbered}
================
R.P. was supported by NSERC Discovery Grant number 400550. S.P. was supported by NSERC Discovery Grant number 1174582. R.P. and S.P. wish to acknowledge the Sanya International Mathematics Forum, which hosted the twelveth Workshop on Numerical Ranges and Numerical Radii, where this work was initiated. The authors would like to thank Dr. Lin Zhang for many helpful suggestions about references.
[50]{}
Å. Björck and G. H. Golub, *Numerical methods for computing angles between linear subspaces*, Mathematics of Computation, **27** (1973), 579-594.
S. Bose, *Quantum communication through an unmodulated spin chain*, Phys. Rev. Lett., **91** (2003), 207901.
D.J.C. Bures, *An extension of Kakutani’s theorem on infinite product measures to the tensor product of semifinite w$^*$-algebras*, Trans. Amer. Math. Soc., **135** (1969), 199-212.
M. Fiedler and V. Pták, *A new positive definite geometric mean of two positive definite matrices*, Linear Algebra Appl. **251** (1997), 1-20. C. Jordan, *Essai sur la géométrie à $n$ dimensions* Bull. Soc. Math. France **3** (1875), 103-174.
R. Jozsa, *Fidelity for mixed quantum states*, J. Mod. Optics, **41** (1994), 2315-2323. S. Kakutani, *On equivalence of infinite product measures*, Ann. of Math., **49** (1948), 214-226. A. Kitaev and J. Watrous, *Parallelization, amplification, and exponential time simulation of quantum interactive proof systems*, Proc. 32nd ACM Symp. Theor. Comput. (2000), 608–617.
L. J. Landau and R. F. Streater, *On Birkhoff’s theorem for doubly stochastic completely positive maps of matrix algebras*, Linear Algebra. Appl., **193** (1993), 107-127.
S. Lu and S. L. Braunstein, *Quantum decision tree classifier*, Quantum Inf. Process, **13** (2014), 757-770. Z. Ma, F.L. Zhang, and J.L. Chen, *Geometric interpretation for the $A$ fidelity and its relation with the Bures fidelity*, Phys. Rev. A, **78** (2008), 064305. D. Markham, J.A. Miszczak, Z. Pucha[ł]{}a, and K. Życzkowski, *Quantum state discrimination: A geometric approach*, Phys. Rev. A **77** (2008), 042111. M. Marshall, *Error estimates in the optimization of degree two polynomials on a discrete hypercube*. SIAM J. Optim. **16** (2005), 297-309. A. W. Marshall, I. Olkin, and B. C. Arnold, *Inequalities: Theory of Majorization and its Applications*, 2nd ed. Springer: New York, 2011. M. Nielsen, *Conditions for a Class of Entanglement Transformations*, Phys. Rev. Lett. **83**, no. 2 (1999), 436–439. C. O’Meara and R. Pereira, *Self-dual and symmetric bistochastic matrices*. Linear Multilinear Algebra, **61** (2013), 23-34. A. Ostrowski, *Sur quelques applications des functions convexes et concaves au sens de I. Schur*. J. Math. Pures. Appl. (9) **31**(1952), 253-292 W. Rosgen, *Computational Distinguishability of quantum channels*. PhD thesis, University of Waterloo. Available at http://arxiv.org/pdf/0909.3930.pdf A. Uhlmann, *The “transition probability” in the state space of a $^*$-algebra*, Rep. Math. Phys., **9** (1976), 273-279.
A. Uhlmann and B. Crell, *Geometry of state spaces*. Lect. Notes Phys., Springer-Verlag, **768** (2009), 1-60.
J. Watrous, *Mixing doubly stochastic quantum channels with the completely depolarizing channel*. Quantum Info. Comp. **9** (2009), 406-413. L. Zhang and S.-M. Fei, *Quantum fidelity and relative entropy between unitary orbits*, J. Phys. A: Math. Theor. **47**, 055301 (2014).
[^1]: The language in the literature is not consistent. Some authors take the point of view of Jozsa: transition probability = fidelity =$F^2(\rho, \sigma)$, using the notation of definition \[defn:fidelity\]. They then call $F(\rho, \sigma)$ the square root fidelity.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the interstellar medium, as well as in the Universe, large density fluctuations are observed, that obey power-law density distributions and correlation functions. These structures are hierarchical, chaotic, turbulent, but are also self-organizing. The apparent disorder is not random noise, but can be described by a fractal, with a deterministic fractal dimension. We discuss the theories advanced to describe these fractal structures, and in particular a new theory of the self-gravity thermodynamics, that could explain their existence, and predict their fractal dimension. The media obeying scaling laws can be considered critical, as in second order phase transitions for instance.'
author:
- Francoise
title: |
FRACTAL STRUCTURES DRIVEN by SELF-GRAVITY:\
Molecular clouds and the Universe
---
Introduction: Fractal Concepts
==============================
Fractals have been introduced by Mandelbrot (1975) to define geometrical ensembles, or mathematical sets, that have a fractional dimension. He pioneered the study of very irregular mathematical sets, where the methods of classical calculus cannot be applied. Fractals are not smooth nor differentiable; they are characterized by self-similarity. Their geometrical structure has details at all scales, and the details are representative of the whole.
Fractals are very rich, since they give the best approximation for many natural phenomena, that cannot be represented by regular and smooth geometries (natural systems are only fractal between two boundaries, with upper and lower cut-offs, and not true mathematical fractals, of course). There can exist a large part of randomness in fractals, and the fractal dimension and exponents of scaling laws are the tools to quantify the hidden order in them.
There are many definitions of the dimension, trying to quantify how much space a system fills. We will use mainly the Hausdorff dimension $D$, based on the Hausdorff measure, that generalizes the notion of length, area and volume. When the length scale is magnified by $\lambda$, a regular and smooth set’s area and volume are respectively scaled by $\lambda^2$, $\lambda^3$, while the fractal set measure is scaled by $\lambda^D$. For the astrophysical systems that we describe below, the mass contained within a scale $r$ is M$\propto r^D$, with $D$ a fractional number between 1 and 3.
Interstellar Clouds
===================
The gaseous medium pervading the galaxy in between stars is highly structured. This medium consists of clouds of hydrogen, either atomic or molecular, according to its density or column density. A quick look at the Milky Way reveals at once the very irregular and clumpy structure of the interstellar medium (ISM), while the stellar distribution is much smoother. Sizes of the clouds range from 100pc (the so-called Giant Molecular Clouds or GMC) down to 20 AU the smallest structures observed through VLBI in the vicinity of the Sun, in absorption in front of quasars (e.g. Davis et al 1996). This corresponds to 6 orders of magnitude in size, and about 10 in masses.
The geometrical appearance is very irregular, but can be caracterized by sheets and filaments of great contrast. The aspect is self-similar, which can be glanced from clouds in the Galaxy at very different distances from the Sun. The dynamics of the ISM has always been mysterious, since it was expected that clouds collapse in a few million years to form stars. But molecular clouds, at all scales, are found to be relatively stable over times long compared to the free-fall time of the cloud at the given scale.
Scaling Laws
-------------
The various structures of interstellar clouds are not distributed at random, but obey power-law relations between size, linewidth and mass (cf Larson 1981). These power-laws demonstrate the self-similar nature, and the scaling properties of the ISM: no peculiar scale exists (except at the two boundaries of course, lower and upper cut-offs). The mass of the clouds is always a very uncertain quantity to obtain, since there is no good universal tracer. The H$_2$ molecule does not radiate in the cold conditions of the bulk of the ISM (10-15 K), since it is symmetric, with no dipole moment. The first tracer is the most abundant molecule CO (10$^{-4}$ with respect to H$_2$), but it is most of the time optically thick, or photo-dissociated (cf below). More direct quantities to measure are the sizes $R$ and the line-widths or velocity dispersion $\sigma$, and the two are linked through a power-law relation:
$$\sigma \propto R^q$$ with $q$ between 0.3 and 0.5 (e.g. Larson 1981, Scalo 1985, Solomon et al 1987, cf \[fig1\]). Besides, molecular clouds appear to be virialised (at least within the uncertainties of mass determination) over all scales, so that
$$\sigma^2 \propto M/R$$ and the size-mass relation follows:
$$M \propto R^D$$ with $D$ the Hausdorff fractal dimension between 1.6 and 2. It can be deduced also that the mean density over a given scale R decreases as 1/R$^\alpha$, where $\alpha$ is between 1 and 1.4.
Hierarchical and Tree Interpretations
--------------------------------------
The observations of molecular clouds reveal that the structure is highly hierarchical, smaller clumps being embedded within the larger ones. Is this a completely exclusive relation, or are there isolated clumps? This is difficult to disentangle, since we have no real 3D picture of the ISM, the third dimension being traced by the radial velocity, and the latter being turbulent and not systematic. It has been possible, however, to build a tree structure where each clump has a parent for instance for clouds in Taurus, a very nearby region (Houlahan & Scalo 1992).
This recursive structure might tell us about the formation mechanism, as will be detailed later on. Basically, if self-gravity is the dominant force, density contrast will be build through Jeans instability, and this is a recursive process, in a quasi isothermal medium. The cooling is very efficient in the dense ISM, and it can be considered almost isothermal.
Fractal Dimensions
-------------------
There are various ways to quantify the observed self-similarity, and to define fractal dimensions. One of them is just to measure in 2D the surface versus the perimeter of a given structure. This method has been used in 2D maps, like the IRAS continuum flux, or the extinctions maps of the sky. In all cases, this method converge towards the same fractal dimension $D_2$. For a curve of fractal dimension $D_2$ in a plane, the perimeter P and area A are related by $$P \propto A^{D_2/2}$$ Falgarone et al (1991) find a dimension $D_2$ = 1.36 for CO contours both at very large (degrees) and very small scales (arcmin), and the same is found for IRAS 100$\mu$ contours in many circumstances (e.g. Bazell & Desert 1988). Comparable dimensions ($D_2$ between 1.3 and 1.5) are found with any tracer, for instance HI clouds (Vogelaar & Wakker 1994).
Note that the projection of a fractal of dimension $D$ is not mandatorily a fractal, but if it is one with dimension $D_p$ it is impossible a priori to deduce its fractal dimensions, except that
$D_p = D $ if $D \leq 2$
$D_p = 2 $ if $D \geq 2$
(Falconer 1990).
Technical Biases
------------------
It is a very difficult task to trace quantitatively the fractal structure of the ISM. Dense gas is molecular, and cold H$_2$ molecules do not radiate (Combes & Pfenniger 1997). The trace molecules such as CO are either optically thick, or not thermally excited (in low-density regions), or photo-dissociated near ionizing stars. The large range of scales is also a source of bias: the small scales are not resolved, and observed maps are smoothed out. This process, which confuses the fractal with a diffuse medium of fractal dimension 3, can lead to underestimates of the mass by factors more than 10 (e.g. simulations in Pfenniger & Combes 1994).
The mass spectrum of density fluctuations has often been studied as another way to characterise the ISM structure, and may be also to predict the mass spectrum of stars that form within these clouds. The differential mass spectrum dN/dm has been found to obey a power-law $$dN/dm \propto m^\gamma$$ with $\gamma = -1.5\pm 0.2$, between masses of 1 to 10$^6$ M$_\odot$ (Casoli et al 1984, Solomon et al 1987, Brand & Wouterloot 1995). But this result has been obtained through molecular lines surveys. When extinction surveys are used, to determine a size spectrum (that can be directly related to the mass spectrum), a much flatter power law is found, and this has been attributed to occlusion (Scalo & Lazarian 1996), i.e. due to blocking of clouds by larger foreground clouds. Although this affects much less spectral studies, there could also be occlusion at a given velocity.
More recently, Heithausen et al (1998) have extended this relation over 5 orders of magnitudes in masses, down to Jupiter masses, and they found a steeper slope $\gamma = -1.84$. Their mass-size relation is M$\propto r^{2.31}$, also much steeper than previous studies, but the estimation of masses at small scales is quite uncertain (in particular the conversion factor between CO and H$_2$ mass could be much higher).
Turbulence
-----------
Everybody agrees that the word characterising the best the ISM is “turbulence”. In laboratory it is well known that a laminar flow can turn turbulent when the Reynolds number is larger than a critical value, i.e. $$R_e = v l /\nu > R_c$$ where $v$ is the velocity, $l$ a typical dimension, and $\nu$ the kinematic viscosity. This means that the advection term v . $\nabla$ v dominates the viscous term in the fluid equation. The turbulent state is characterised by unpredictable fluctuations in density and pressure, and a cascade of whirls. In the ISM, the viscosity can be estimated from the product of the macroturbulent velocity (or dispersion) and the mean-free-path of cloud-cloud collisions (since the molecular viscosity is negligible). But then the Reynolds number is huge ($\approx 10^9$), and the presence of turbulence is not a surprise.
This fact has encouraged many interpretations of the ISM structure in terms of what we know from incompressible turbulence. In particular, the Larson relations have been found as a sign of the Kolmogorov cascade (Kolmogorov 1941). In this picture, energy is dissipated into heat only at the lower scales, while it is injected only at large scale, and transferred all along the hierarchy of scales. Writing that the energy transfer rate $v^2/(r/v)$ is constant gives the relation $$v \propto r^{1/3}$$ which is close to the observed scaling law, at least for the smallest cores (Myers 1983). The source of energy at large scale would then be the differential galactic rotation and shear (Fleck 1981). This idealized view has been debated (e.g. Scalo 1987): it is not obvious that the energy cascades down without any dissipation in route (or injection), given the large-scale shocks, flows, winds, etc... observed in the ISM. Also, the interstellar medium is highly compressible, and its behaviour could be quite different from ordinary liquids in laboratory.
Besides, many features of ordinary turbulence are present in the ISM. For instance, Falgarone et al (1991) have pointed out that the existence of non-gaussian wings in molecular line profiles might be the signature of the intermittency of the velocity field in turbulent flows. More precisely, the $^{13}$CO average velocity profiles have often nearly exponential tails, as shown by the velocity derivatives in experiments of incompressible turbulence (Miesch & Scalo 1995). Comparisons with simulations of compressible gas give similar results (Falgarone et al 1994). Also the curves obtained through 2D slicing of turbulent flows have the same fractal properties as the 2D projected images of the ISM; their fractal dimension D$_2$ obtained from the perimeter-area relation is also 1.36 (Sreenivasan & Méneveau 1986).
More essential, the ISM is governed by strong fluctuations in density and velocity. It appears chaotic, since it obeys highly non-linear hydrodynamic equations, and there is coupling of phenomena at all scales. This is also related to the sensibility to initial conditions that defines a chaotic system. The chaos is not synonymous of random disorder, there is a remarquale ordering, which is reflected in the scaling laws. The self-similarity over several orders of magnitude in scale and mass means also that the correlation functions behave as power-laws, and that there is no finite correlation length. This characterizes critical media, experiencing a second order phase transition for example. This analogy will be developped further in this chapter.
Self-Gravity
--------------
Although the ISM is a self-organizing, multi-scale medium, comparable to what is found in laboratory turbulence, there are very special particularities that are not seen but in astrophysics. Self-gravity is a dominant, while it has not to be considered in atmospheric clouds for instance. It has been recognized by Larson (1981) and by many others that at each scale the kinetic energy associated with the linewidths balances the gravitational energy: clouds are virialized approximately, given their very irregular geometry.
This property of course has to stop at the largest scale, when the influence of the galactic gravity, and associated shear, intervenes. This scale is that of the Giant Molecular Clouds, of the order of 100pc in size, and 10$^6$ M$_\odot$ in mass. They are the largest self-gravitating structures in the Galaxy. At the other extremity, the smallest scales are not well known. The observations through molecular lines, in nearby clouds with millimetric interferometry, detect structures down to 0.01pc currently, and even in high-latitude clouds distant by 100 pc, structures down to 400 AU (2 10$^{-3}$ pc). Through VLBI, it is possible to go much further. HI is detected and mapped in absorption in front of remote quasars (Diamond et al 1989, Davis et al 1996), and contrasted structures of 20 AU are ubiquitous. This small-scale structure is also traced by the interstellar scattering, and in particular extreme scattering events or ESE (Fiedler et al 1987, 1994). The statistics on ESE is so large that we know approximately the number of small clumps in the Galaxy: they are about 1000 times more numerous than stars. If these objects are self-gravitating, at the end of the ISM hierarchy, their mass is of the order of 10$^{-3}$ M$_\odot$, and they represent a significant mass component of the Galaxy (Pfenniger & Combes 1994).
Self-gravity is widely accepted as dominant process in the ISM. Gravitational collapse is accompanied by fragmentation in a system with very efficient cooling, and this process can provide the turbulent motions observed. The theory was first proposed by Hoyle (1953) who showed that the isothermal collapse of a cloud led to recursive fragmentation, since the Jeans length decreases faster than the cloud radius. Rees (1976) has determined the size of the smallest fragments, when they become opaque to their own radiation. They correspond roughly to the smallest scales observed in the ISM (sizes of 10 AU, and masses of 10$^{-3}$ M$_\odot$, see the physical parameters of the “clumpuscules” in Pfenniger & Combes 1994). Many other physical processes play a role in the turbulent ISM, as for instance rotation and magnetic fields. But they cannot be identified as the motor and the origin of the structure. Galactic rotation certainly injects energy at the largest scales, but angular momentum cannot cascade down the hierachy of clouds; indeed if the rotational velocity is too high, the structure is unstable to clump formation (cf Toomre criterium, 1964), and the non-axisymmetry evacuates angular momentum outside the structure. Magnetic fields are certainly enhanced by the turbulent motions, and could reach a certain degree of global equipartition with gravitational and kinetic energies in tbe virialised clouds. But they cannot be alone at the origin of the hierarchical structure, gravity has to trigger the collapse first. Besides, there is no observational evidence of the gas collapse along the field lines, polarisation measurements give contradictory results for the field orientation with respect to the gas filaments. Therefore, although rotation, turbulence, magnetic fields play an important role in the ISM, they are more likely to be consequences of the formation of the structure.
Simulations
-------------
A large number of hydrodynamical simulations have been run, in order to reproduce the hierachical density structure of the interstellar medium. However, these are not yet conclusive, since the dynamical range available is still restricted, due to huge computational requirements.
It has been argued that self-similar statistics alone can generate the observed structure of the ISM, in pressureless turbulent flows without self-gravity (Vazquez-Semadeni 1994); however, only three levels of hierarchical nesting can be traced.
From the size-linewidth relation $ \sigma \propto R^{1/2}$, and the second observed scaling law $ \rho \propto R^{-1}$, it can be deduced that
$$\sigma \propto \rho^{-1/2}$$ and therefore, if the turbulent pressure $P$ is defined as usual by
$$dP /d\rho = \sigma^2$$ it follows that $$P \propto log\rho$$ which is the logatropic equation of state, or “logatrope”. This behaviour has been tested in simulations (e.g. Vazquez-Semadeni et al 1998), but the logatrope has not been found adequate to represent dynamical processes occuring in the ISM (either hydro, or magnetic 2D simulations). The equation of state of the gas would be more similar to a polytrope of index $\gamma \approx 2$. But the results could depend whether the clouds are in approximate equilibrium or not (cf McLaughlin & Pudritz 1996).
Vazquez-Semadeni et al (1997) have searched for Larson relations in the results of 2D self-gravitating hydro (and MHD) simulations of turbulent ISM: they do not find clear relations, but instead a large range of sizes at a given density, and a large range of column densities; they suggest that the observational results could be artefacts or selection effects (existence of a threshold in column density for UV-shielding for example).
Galaxy Distributions
=====================
It has long been recognized that galaxies are not distributed homogeneously in the sky, but they follow a hierarchical structure: galaxies gather in groups, that are embedded in clusters, then in superclusters, and so on (Shapley 1934, Abell 1958). Moreover, galaxies and clusters appear to obey scaling properties, such as the power-law of the two point-correlation function: $$\xi(r) \propto r^{-\gamma}$$ with the slope $\gamma$, the same for galaxies and clusters, of $\approx$ 1.7 (e.g. Peebles, 1980, 1993).
Correlation Functions
----------------------
The correlation function is defined as $$\xi(r) = \frac{<n(r_i).n(r_i+r)>}{<n>^2} -1$$ where $n(r)$ is the number density of galaxies, and $<...>$ is the volume average (over $d^3r_i$). One can always define a correlation length $r_0$ by $\xi(r_0) = 1$.
This definition involves the average density $<n>$, which depends on the scale for the galaxy distribution, since it is a fractal, at least over a certain range of scales. There has been considerable debate about this (see Davis 1997, Pietronero et al 1997). If everybody agrees that the universe is a fractal below 200 Mpc scales, the question is not settled as of the scale beyond which the universe is homogeneous. Pietronero et al (1997) claim that this limiting scale has not yet been reached in the present catalogs, since large-scale structures are still found at any scale. On the contrary, Davis & Peebles (1983) or Hamilton (1993) argue that the galaxy-galaxy correlation length $r_0$ is rather small. The most frequently reported value is $r_0 \approx 5 h^{-1}$ Mpc (where $h = H_0$/100km s$^{-1}$Mpc$^{-1}$).
The problem is that the definition of $\xi(r)$ includes a normalisation by the average density of the universe, which, if the homogeneity scale is not reached, depends on the size of the galaxy sample. [^1].
This implies a correlation length that should increase with the distance limits of galaxy catalogs, as it indeed does (Davis et al 1988). The same problem occurs for the two-point correlation function of galaxy clusters; the corresponding $\xi(r)$ has the same power law as galaxies, their length $r_0$ has been reported to be about $r_0 \approx 25 h^{-1}$ Mpc, and their correlation amplitude is therefore about 15 times higher than that of galaxies (Postman, Geller & Huchra 1986, Postman, Huchra & Geller 1992). The latter is difficult to understand, unless there is a considerable difference between galaxies belonging to clusters and field galaxies (or morphological segregation). The other obvious explanation is that the normalizing average density of the universe was then chosen lower.
Assuming that the average density is a constant, while homogeneity is not yet reached, could perturb significantly the correlation function, and its slope, as shown by Coleman, Pietronero & Sanders (1988) and Coleman & Pietronero (1992). The function $\xi(r)$ has a power-law behaviour of slope $-\gamma$ for $r< r_0$, then it turns down to zero rather quickly at the statitistical limit of the sample. This rapid fall leads to an over-estimate of the small-scale $\gamma$. Pietronero (1987) introduces the conditional density $$\Gamma(r) = \frac{<n(r_i).n(r_i+r)>}{<n>}$$ which is the average density around an occupied point. For a fractal medium, where the mass depends on the size as $$M(r) \propto r^D$$ $D$ being the fractal (Haussdorf) dimension, the conditional density behaves as $$\Gamma(r) \propto r^{D-3}$$
It is possible to retrieve the correlation function as $$\xi(r) = \frac{\Gamma(r)}{<n>} -1$$ In the general use of $\xi(r)$, $<n>$ is taken for a constant, and we can see that $$D = 3 - \gamma \quad .$$ If for very small scales, both $\xi(r)$ and $\Gamma(r)$ have the same power-law behaviour, with the same slope $-\gamma$, then the slope appears to steepen for $\xi(r)$ when approaching the length $r_0$. This explains why with a correct statistical analysis (Di Nella et al 1996, Sylos Labini & Amendola 1996, Sylos Labini et al 1996), the actual $\gamma \approx 1-1.5$ is smaller than that obtained using $\xi(r)$ (cf \[fig2\]). This also explains why the amplitude of $\xi(r)$ and $r_0$ increases with the sample size, and for clusters as well.
Homogeneity Hypothesis and Cosmological Principle
--------------------------------------------------
Isotropy and homogeneity are expected at very large scales from the Cosmological Principle (e.g. Peebles 1993). However, this does not imply local or mid-scale homogeneity (e.g. Mandelbrot 1982, Sylos Labini 1994): a fractal structure can be locally isotropic, but inhomogeneous. The main observational evidence in favor of the Cosmological Principle is the remarkable isotropy of the cosmic background radiation (e.g. Smoot et al 1992), that provides information about the Universe at the matter/radiation decoupling. There must therefore exist a transition between the small-scale fractality to large-scale homogeneity. This transition is certainly smooth, and might correspond to the transition from linear perturbations to the non-linear gravitational collapse of structures. The present catalogs do not yet see the transition since they do not look up sufficiently back in time. It can be noticed that some recent surveys begin to see a different power-law behavior at large scales ($\lambda \approx 200-400 h^{-1}$ Mpc, e.g. Lin et al 1996).
The Problem and Methods
-----------------------
It is generally recognized that the galaxy structures in the Universe have developped by gravitational collapse from primordial fluctuations. Once unstable, density fluctuations do not grow as fast as we are used to for Jeans instability (exponential), since they are slowed down by expansion. The rate of growth is instead a power-law. Let us call the density contrast $\delta$: $$\delta(\vec x) = (\rho(\vec x) -<\rho>)/<\rho>$$ where $<\rho>$ is the mean density of the Universe, assumed homogeneous at very large scale. If $\vec r$ is the physical coordinate, the comoving coordinate $\vec x$ is defined by: $$\vec r = a(t) \vec x$$ where a(t) is the scale factor, accounting for the Hubble expansion (normalised to a(t$_0$) = 1 at the present time). Since the Hubble constant verifies $H(t) = \dot{a}/a$, the peculiar velocity is defined by $$\vec v = \dot{\vec r} - H \vec r = a \dot{\vec x}$$ In comoving coordinates, the Poisson equation becomes: $$\nabla_x^2\Phi = 4\pi G a^2 (\rho -<\rho>)$$ It can be shown easily that in a flat universe, the density contrast in the linear regime grows as the scale factor $a(t) = (t/t_0)^{2/3}$ (this is also approximately true for any universe in the early times). An exact solution for the non-linear collapses exist only in very special conditions, such as the spherical collapse. For the well-known top-hat perturbation, an overdensity reaches the singularity in a finite collapse time $t_c$, when its corresponding linear density contrast would have reached the value $$\delta_c \approx 1.69$$ The evolution of its radius follows the Friedman solution for a density above critical; it reaches a maximum radius, before virializing to a radius equal to half that one. Its final density contrast is 178 (both figures, although indicative, are widely used in the domain).
The level of fluctuations was very small (about 10$^{-5}$ at the scale of COBE resolution, i.e. 7$^\circ$) at the last scattering surface, just before matter recombination, about 10$^5$ yr after the Big-Bang. At first, the development of the structures is easy to compute, since they are in the linear regime, and interactions between scales can be neglected (cf Peebles 1980). Then, in the non-linear regime, no analytical solution exists, and one should resort to approximations, or full N-body simulations.
### Numerical simulations
N-body computations have been widely used, to gravitationnally follow the non-linear evolution of the fluctuations. From a comparison of the results with today galaxy distribution, one can hope to trace back the initial mass spectrum of fluctuations, and to test postulated cosmologies such as CDM and related variants (cf Ostriker 1993). However, the evolution depends on many free parameters, the gas physics, the star formation feedback, the amount and physics of dark matter. This approach has not yet yielded definite results, also because numerical limitations (restricted dynamical range due to the softening and limited volume) have often masked the expected self-similar behavior (Colombi et al 1996).
### Zel’dovich approximation
The first approximation has been pioneered by Zel’dovich (1970): it consists to assume that the pressure forces are negligible in the first collapsing structures, since they are much bigger than the Jeans length. This is particularly adapted to the top-down scenario of adiabatic fluctuations, where photons and baryons both fluctuate, such that entropy is conserved. In that case, the matter-photon coupling will prevent the development of large amplitudes fluctuations at small-scales. Large-scales are then the first to collapse, and they fragment in smaller structures. In this approximation, the particles maintain their peculiar velocities in comoving space, and collapse in 1D filaments, or 2D sheets or pancakes (cf formation of caustics). After collapse, the approximation fails since particles diffuse away; an artificial viscosity has then been introduced, so that particles cannot cross each other, and the pancakes remain coherent, it is the “adhesion model” (Shandarin & Zel’dovich 1989). The velocity field is governed by a Burgers equation, for which analytical solutions are known (Vergassola et al 1994).
Related to this approach is the Lagrangian approximation, which in fact pushes the Zel’dovich approximation to further orders. Since this approach fails as soon as orbits are crossing (i.e. multi-streaming) it is assumed that this multi-streaming at small scales has negligible consequences at large-scales (Lachièze-Rey 1993).
The success of this approximation has been boosted by the fact that the self-gravitating Universe appears filamentary. The reason for this is probably that many collapsing structures are larger than the Jeans mass; also in realistic scenarios, the collapse is not spherical, but does form sheets and filaments (Larson 1985). The latter have the advantage that in 1D, pressure forces have always the same dependence with radius than the gravity forces, and therefore cannot halt the collapse, whatever the cooling (Uehara et al. 1996; Inutsuka & Miyama 1997).
### BBGKY hierarchy
A second approach, which should work essentially in the linear (or weakly non-linear) regime, is to solve the BBGKY hierarchy of equations, a method which has been successfully used in plasma physics. However, the hierachy is here infinite, and approximations should be made for closure (Davis & Peebles 1977; Balian & Schaeffer 1989). The main assumption is that the $N$-points correlation functions are scale-invariant and behave as power-laws like is observed for the few-body correlation functions. Crucial to this approach is the determination of the void probability, which is a series expansion of the $N$-points correlation functions (White 1979). The hierachical solutions found in this frame agree well with the simulations, and with the fractal structure of the universe at small-scales (Balian & Schaeffer 1988).
### Thermodynamical approach
A third approach is the thermodynamics of gravitating systems, developped by Saslaw & Hamilton (1984), which assumes quasi thermodynamic equilibrium. The latter is justified at the small-scales of non-linear clustering, since the expansion time-scale is slow with respect to local relaxation times. Indeed the main effect of expansion is to subtract the mean gravitational field, which is negligible for structures of mean densities several orders of magnitude above average. The predictions of the thermodynamical theory have been successfully compared with N-body simulations (Itoh et al 1993), but a special physical parameter (the ratio of gravitational correlation energy to thermal energy) had to be adjusted for a better fit (Bouchet et al 1991, Sheth & Saslaw 1996, Saslaw & Fang 1996).
Mass Function, Press-Schechter Formalism
------------------------------------------
In 1974, Press and Schechter developped a formalism to predict the mass function of galaxies and structures in the Universe, assuming that gravitational collapse in an expanding universe non-linearly evolves through a self-similar spectrum (PS74). The resulting mass function is independent of the initial spectrum of condensations assumed. The non-linear N-body interactions randomize the initial positions and generate perturbations to all other scales: the growth of large scale condensations from non-linear clumping of smaller ones occuring much faster than their linear growth, a self-similar universal spectrum is established, only from seeds corresponding to the Jeans mass at recombination ($\approx$ 10$^7$ M$_\odot$).
This led to an excellent fit of the galaxy luminosity function proposed by Schechter (1976): $$\Phi(L) \propto L^{-\alpha} exp(-L/L_*)$$ where $L_*$ is a characteristic galaxy luminosity, and $\alpha\sim 1$.
In order to derive the mass function, i.e. the number density of structures of mass between M and M+dM, $n(M)$, PS74 first assume that a structure collapses as soon as the extrapolated linear density contrast $\delta_c$ is of the order of 1. Since in the spherical top-hat case (see section 3.3) the collapse occurs in a time-scale corresponding to a linear $\delta_c$ = 1.69, this indicative value of $\delta_c$ is taken. The fraction of collapsed mass, at any resolution smaller than $R$ is:
$$F(<R) = \int_{\delta_c} ^\infty P(\delta, R) d\delta$$
For gaussian random phase fluctuations, the probability $ P(\delta, R)$ is $$P(\delta, R) = (2\pi)^{-1/2}/\delta_* \, exp(-\delta^2/(2 \delta_*^2))$$
where $\delta_*$ is the standard deviation, obtained at scale $R$ $$\delta_* = (<M^2> - <M>^2)^{1/2}/M$$
Then the fraction $F(<R)$ becomes: $$F(<R) = 1/2 \, {\rm erfc}\, (2^{-1/2} \delta_c/ \delta_*)$$ where $erfc$ is the complementary error function.
To relate this to the mass associate to the scale R, we use $ M = 4\pi <\rho> R^3/3$, justified for the top-hat smoothing function (to obtain the average $\rho$). The function $F(<R)$ gives the probability for a scale $R$ to be bound, but it could also be part of another larger bound scale, so the fraction of independent masses, with mass between $M$ and $M+dM$, that have collapsed is the derivative of it with respect to mass $M$: $$dF/dM = 1/2 (2\pi)^{-1/2} \delta_c /\delta_* M^{-1}
\, exp(-\frac{1}{2} \frac{\delta_c^2}{\delta_*^2} )$$
The mass density due to condensations of mass $M$ is then obtained by multiplying by the total number density at the same epoch $\rho/M$. However, PS74 realised that the integral of the fraction F was 1/2 instead of 1, i.e. that the formula accounted for only one half of the mass in the universe. They multiplied by a factor 2 to include the underdense regions that would accrete on the collapsed objects. This is a typical problem of the linear theory, where only half of the mass is in overdense regions. Also called the cloud-in-cloud problem, this factor 2 puzzle has been better explained later (Cole 1989, Bond et al 1991, White & Kauffman 1994).
This formalism has been recently widely used, since it gives excellent agreement with N-body simulations, and provides a way to compute easily the merging histories of galaxies (Lacey & Cole 1993, 1994).
If we make the particular assumption that the power spectrum of the fluctuations $P(k)$ (i.e. the square of the module of $\delta_k$ the Fourier transform of $\delta$) is $P(k) \propto k^n$, then $\delta_* \propto M^{-(n+3)/6}$, and the PS formula becomes $$n(M) dM = <\rho> /M_*^2 /sqrt(2\pi) (n+3)/6 (M/M_*)^{(n+3)/6-2}$$ $$exp (-1/2 (M/M_*)^{(n+3)/3}) dM$$ where $M_*$ is the mass corresponding to the scale at which the amplitude $\delta_*$ reaches $\delta_c$.
Multifractals
--------------
The idea of fractals as a good representation of the self-similar clustering hierachy of galaxies has been proposed very early (de Vaucouleurs 1960, 1970; Mandelbrot 1975). Since then, many authors have shown that a fractal distribution indeed reproduces quite well the aspect of galaxy catalogs, for example by simulating a fractal and observing it, as with a telescope (Scott, Shane & Swanson, 1954; Soneira & Peebles 1978). When going into details, however, the distribution of galaxies might appear more complex: first of all, catalogs are weighted by the luminosity distribution, and not the mass distribution, and these are not equivalent. For instance, there is a luminosity segregation in clusters of galaxies, and slightly different fractal dimensions can be obtained with different galaxy-types (elliptical/spirals/dwarfs). Also determination of masses of structures through different ways (X-rays, gravitational lensing, etc..) suggest that the dark matter is not exactly distributed as light, and there could be a significant bias. All this has led to the introduction of multifractality to represent the Universe (e.g. Sylos-Labini & Pietronero 1996). In a multifractal system, local scaling properties slightly evolve, and can be defined by a continuous distribution of exponents. This is a mere generalisation of a simple fractal, that links the space and mass distributions. Mutlifractality may also better account for the transition to homogeneity, with a fractal dimension varying with scale (Balian & Schaeffer 1989, Castagnoli & Provenzale 1991, Martinez et al 1993, Dubrulle & Lachièze-Rey 1994).
Bases for a Statistical Field theory
====================================
Both for the ISM and for galaxy distributions in the Universe, self-similar structures are observed over large ranges in scales. Scaling laws are observed, which translate by an average density decreasing with scale as a power-law, of slope $-\gamma$ between -1.5 and -1, corresponding to a fractal dimension $D = 3-\gamma$ between 1.5 and 2.
We propose in the following that self-gravity is a dominant factor in the two media, and try to establish a statistitical theory in the hope to explain the fractal structure. The theory should not only account for the existence of the structure, but also be able to predict its fractal dimension and others critical exponents (de Vega, Sanchez & Combes, 1996a,b, 1998).
Since gravity is scale-independent, there are opportunities for a mechanism to propagate over scales in a self-similar fashion. For the ISM, in a quasi isothermal regime, a fractal structure could be build through recursive Jeans instability and fragmentation. This recursive fragmentation proceeds until the density is high enough to reach the adiabatic regime. Self-gravity could be the principal origin of the fractal, with generated turbulent motions in virial equilibrium at each scale. For galaxy formation, the smallest structures collapse first, and these influence the largest scale in a non-linear manner. It is obvious that in both cases, the system does not tend to a stationary point, but develops fluctuations at all scales, and these must be studied statistically.
In order to study its thermodynamics properties, we develop the grand partition function of the ensemble of self-gravitating particles (section \[stat\]). In transforming the partition function through a functional integral (section \[functional\]), it can be shown that the system is exactly equivalent to a scalar field theory (section \[scalar\]). The theory does not diverge, since the system is considered only between two scale limits: the short-scale and large-scale cut-offs. Through a perturbative approach it can be demonstrated that the system has a critical behaviour, for any parameter (effective temperature and density). That is, we can consider the self-gravitating gaseous medium as correlated at any scale, as for the critical points phenomena in phase transitions (as was first suggested by Totsuji & Kihara 1969).
Since scaling behaviours are the best studied through renormalization group theory, we use these methods to derive by analogy the critical exponents of the system (section \[renorm\]). The aim, that will be developped more fully in the following, is to relate the critical exponents of well known universal critical phenomena, to the Haussdorf dimension of the astrophysics fractals.
Hamiltonian of the Self-Gravitating Ensemble of N-bodies {#stat}
---------------------------------------------------------
Let us consider a gas of particles submitted only to their self-gravity, in thermal equilibrium at temperature T $(kT = \beta^{-1}) $. In the interstellar medium, quasi isothermality is justified, due to the very efficient cooling. For unperturbed gas in the outer parts of galaxies, gas is in equilibrium with the cosmic background radiation at $T \approx 3K$ (Pfenniger et al 1994, Pfenniger & Combes 1994). For a system of collapsing structures in the universe, this can be a valid approximation, as soon as the gradient of temperature is small over a given scale.
This isothermal character is essential for the description of the gravitational systems as critical systems, as will be shown later, so that the canonical ensemble appears the best adapted system. We do not consider isolated gravitational systems, for which the microcanonical system is generally used (e.g. Horwitz & Katz 1978a,b; Padmanabhan 1990). However, for easy mathematical development, we develop the partition function in the grand canonical ensemble, allowing for a variable number of particles $N$ (for application to real systems, constant masses will later be considered). The grand partition function ${\cal Z}$ and the Hamiltonian $H_N$ are
$${\cal Z} = \sum_{N=0}^{\infty}\; {{z^N}\over{N!}}\; \int\ldots \int
\prod_{l=1}^N\;{{d^3p_l\, d^3q_l}\over{h^3}}\; e^{- \beta H_N}$$
$$H_N = \sum_{l=1}^N\;{{p_l^2}\over{2m}} - G \, m^2 \sum_{1\leq l < j\leq N}
{1 \over { |{\vec q}_l - {\vec q}_j|}}$$ where $z$ is the fugacity = $exp(-\beta \mu_c)$ in terms of the gravito-chemical potential $\mu_c$.
This can be transformed, with continuous density $\rho({\vec r})= \sum_{j=1}^N\; \delta({\vec r}- {\vec q}_j)\; $
$$\frac12 \, \beta G \, m^2 \sum_{1\leq l \neq j\leq N}
{1 \over { |{\vec q}_l - {\vec q}_j|}} = \frac12\, \beta \, G \, m^2
\int_{ | {\vec x} - {\vec y}|> a}\;
{{d^3x\, d^3y}\over { | {\vec x} - {\vec y}|}}\; \rho({\vec x})
\rho({\vec y}) \;$$
The cutoff $ a $ is here introduced naturally, it corresponds to the size of the smallest fragments, or clumpuscules (of the order of $\sim 10$ AU). In fact, we consider that the particles of the system interact with the Newton law of gravity ($1/r$) only within the size range of the fractal, where self-gravity is predominant. At small scale, other forces enter into account, and we can adopt a model of hard spheres to schematize them. Also at large scales, beyond the upper cutoff, different forces must be introduced. The phenomenological potential thus considered does not possess any singularity.
Grand Partition Function as a Functional {#functional}
-----------------------------------------
Using the potential in $1/r$, and its inverse operator $-\frac{1}{4\pi}\nabla^2$ (but see also a similar derivation, with $[1 - \theta (a-r) ]/r$ and its corresponding inverse operator, for the phenomenological potential with cutoff, in de Vega et al 1996b), the exponent of the potential energy can be represented as a functional integral (Stratonovich 1958, Hubbard 1959)
$$e^{ \frac12\, \beta G \, m^2
\int \;
{{d^3x\, d^3y}\over { | {\vec x} - {\vec y}|}}\; \rho({\vec x})
\rho({\vec y})} = \int\int\; {\cal D}\xi \; e^{ -\frac12\int d^3x \; (\nabla
\xi)^2 \; + \; 2 m \sqrt{\pi G\beta}\; \int d^3x \; \xi({\vec x})\;
\rho({\vec x}) }$$
$$\begin{aligned}
{\cal Z} &=& \sum_{N=0}^{\infty} {1 \over{N!}}
\left [z\left({2\pi m \over{h^2 \beta}}\right)^{3/2}\right]^N\
\int\int {\cal D}\xi e^{ -\frac12\int d^3x (\nabla \xi)^2}
\int\ldots \int
\prod_{l=1}^N d^3q_l e^{ 2 m \sqrt{\pi G\beta} \sum_{l=1}^N
\xi({\vec q}_l)} \cr \cr
&=& \int\int {\cal D}\xi e^{ -\frac12\int d^3x (\nabla \xi)^2}
\sum_{N=0}^{\infty}{1 \over{N!}}
\left [ z\left({2\pi m \over{h^2 \beta}}\right)^{3/2}\right]^N
\left[ \int d^3q e^{ 2 m \sqrt{\pi G\beta} \xi({\vec q})}
\right]^N \cr \cr
&=& \int\int\; {\cal D}\xi \; e^{ -\int d^3x \left[ \frac12(\nabla \xi)^2\;
- z \left({m \over{2\pi \beta}}\right)^{3/2}\; e^{ 2 m \sqrt{\pi
G\beta}\;\xi({\vec x})}\right]}\nonumber\end{aligned}$$
With the following change of variables: $$\phi({\vec x}) \equiv 2 m \sqrt{\pi G\beta}\;\xi({\vec x})$$
$${\cal Z} = \int\int\; {\cal D}\phi\; e^{ -{1\over{T_{eff}}}\;
\int d^3x \left[ \frac12(\nabla\phi)^2 \; - \mu^2 \; e^{\phi({\vec
x})}\right]}$$ where $$\mu^2 = {{\pi^{5/2}}\over {h^3}}\; z\; G \, (2m)^{7/2} \, \sqrt{kT} \;
\quad , \quad T_{eff} = 4\pi \; {{G\; m^2}\over {kT}} \quad$$ (note that the “equivalent” temperature in the field theory is in fact inversely proportional to the physical temperature). It can be shown that the parameter $\mu$ is equal to the inverse of the Jeans length, itself of the order of the cutoff $a$.
Introduction of a Local Scalar Field, with Exponential Self-interaction {#scalar}
------------------------------------------------------------------------
The main point of the previous derivation is that the partition function for the gas of particles in gravitational interaction has been transformed into the partition function for a single scalar field $\phi({\vec x})$ with local action
$$S[\phi(.)] \equiv {1\over{T_{eff}}}\;
\int d^3x \left[ \frac12(\nabla\phi)^2 \; - \mu^2 \; e^{\phi({\vec
x})}\right]$$
Apparently, the term $ - \mu^2 \; e^{\phi({\vec x})} $ makes the $\phi$-field energy density diverge, which is related to the attractive character of the gravitational force. But the physical short-distance cutoff $a$ eliminates the zero distance singularity. It is then possible to compute the statistical average value of the density $\rho({\vec r})$
$$<\rho({\vec r})> = {\cal Z}^{-1} \sum_{N=0}^{\infty} {1 \over{N!}}
\left [ z\left({m \over{2\pi \beta}}\right)^{3/2}\right]^N
\int\int \prod_{l=1}^N d^3q_l\rho({\vec r})
e^{ \frac12 \beta G m^2 \sum_{1\leq l \neq j\leq N}
{1 \over { |{\vec q}_l - {\vec q}_j|}} }$$
Or in the $\phi$-field language, the particle density expresses as $$<\rho({\vec r})> = -{1 \over {T_{eff}}}\;<\nabla^2 \phi({\vec r})>=
{{\mu^2}\over{T_{eff}}} \; <e^{\phi({\vec r})}>$$ where $ <\ldots > $ means functional average over $ \phi(.) $ with statistical weight $ e^{S[\phi(.)]} $. Density correlators are
$$C({\vec r_1},{\vec r_2}) \equiv
<\rho({\vec r_1})\rho({\vec r_2}) > -<\rho({\vec r_1})><\rho({\vec r_2}) >$$ $$C({\vec r_1},{\vec r_2}) =
{{\mu^4}\over{T_{eff}}^2} \; \left[
<e^{\phi({\vec r_1})} \; e^{\phi({\vec r_2})}> -
<e^{\phi({\vec r_1})}> \; <e^{\phi({\vec r_2})}> \right]$$
Stationary Points and Hydrostatic Solutions
-------------------------------------------
The equation for stationary points: $$\nabla^2\phi = -\mu^2\, e^{\phi({\vec x})} \;$$ can be expressed in terms of the gravitational potential $U({\vec x})$ $$\nabla^2U({\vec r}) = 4 \pi G \, z \, m
\left({{2\pi mkT}\over{h^2}}\right)^{3/2} \, e^{ - \frac{m}{kT}\,U({\vec r})}$$ This corresponds to the Poisson equation for an ideal gas in hydrostatic equilibrium: $$\nabla P({\vec r}) = - m \, \rho({\vec r}) \; \nabla U({\vec r})$$ where $ P({\vec r}) $ stands for the pressure. Combined with the equation of state for the ideal gas $$P = kT \rho$$ this yields for the particle density $$\rho({\vec r}) = \rho_0 \; e^{ - \frac{m}{kT}\,U({\vec r})}$$ where $ \rho_0 $ is a constant.
The $\phi$-field has special properties under scale transformations $${\vec x} \to {\vec x}_{\lambda} \equiv \lambda{\vec x}$$ where $\lambda$ is an arbitrary real number. For any solution $
\phi({\vec x}) $ of the stationary point equations, there is a family of dilated solutions of the same equation $$\phi_{\lambda}({\vec x}) \equiv \phi(\lambda{\vec x}) +\log\lambda^2$$ In addition, $ S[\phi_{\lambda}(.)] = \lambda^{-1} \; S[\phi(.)] $. A rotationally invariant stationary point is given by $$\phi^c(r) = \log{{2}\over { \mu^2 r^2}}$$ This singular solution, where can be recognized the isothermal sphere, is invariant under the scale transformations, i.e. $$\phi^c_{\lambda}(r) =\phi^c(r)$$ The only constant stationary solution is the singular $ \phi_0 = -\infty $.
With a perturbative method, starting from the stationary solution $ \phi_0 = -\infty $, it can be obtained for large distances (de Vega et al 1996b) $$C({\vec r_1},{\vec r_2}) \buildrel{ | {\vec r_1} - {\vec r_2}|\to
\infty}\over = {{ \mu^4 }\over {2\, C_D^2 \;
| {\vec r_1} - {\vec r_2}|^{2}}} + O\left( \; | {\vec
r_1} - {\vec r_2}|^{-3}\right)$$ showing that the $\phi$-field theory scales, and behaves critically for a continuum set of values of $\mu$ and $ T_{eff} $.
Renormalization Group Methods {#renorm}
==============================
The renormalization methods are very powerful to deal with self-similar systems obeying scaling laws, like critical phenomena. In the latter case, examplified by second order phase transitions, there exist critical divergences, where physical quantities become singular as power-laws of parameters called critical exponents. These critical systems reveal a collective behaviour, organized from microscopic degrees of freedom, through giant fluctuations and statistical correlations. Hierarchical structures are built up, coupling all scales together, replacing an homogeneous system in a scale-invariant system. It can be shown that local forces are not important to describe the collective behaviour, which is only due to the statistical coupling of local interactions. Therefore, critical exponents depend only on the statistical distribution of microscopic configurations, i.e. on the dimensionalities or symmetries of the system. There exist wide universality classes, that allow to draw quantitative predictions on the system from only a qualitative knowledge of its properties (e.g. Parisi, 1988; Zinn-Justin 1989; Binney et al 1992).
Critical Phenomena
-------------------
Critical phenomena occur at second order phase transitions, i.e. continuous transitions without latent heat. The paradigm of these systems is the transition at the Curie point (T=T$_c$= 1043K) from paramagnetic iron, where the magnetic moment is proportional to the applied field m=$\mu$ B, to ferromagnetic state, where there exists a permanent magnetic moment m$_0$ even in zero field. Another well known example is the critical point of water, at which the transition from the liquid to gas becomes continuous (at T$_c$ = 647 K, $\rho_c$ = 0.323 g cm$^{-3}$).
Although the permanent magnet tends to zero continuously at T$_c$, there are divergences: for instance the heat capacity C behaves as C$ \propto | T - T_c |^{-\alpha}$, with $\alpha > 0$. Also the magnetic susceptibility $$\chi_T = \partial{m}/\partial{B}_T \propto | T - T_c |^{-\gamma}$$ (or for the case of H$_2$O, the compressibility $\kappa_T \propto
| T - T_c |^{-\gamma}$).
At the critical point, it is easy to understand that the compressibility which tends to infinity generates large density fluctuations, and therefore light is strongly diffused by the varying optical index: this is the critical opalescence. The extraordinary fact is that microscopic forces can give rise to large-scale fluctuations, as if the medium was organized at all scales.
### Order parameter and correlation function
The order parameter is defined by the characteristic physical quantity which experiences large fluctuations at criticality. For the H$_2$O case, it is a scalar field $$\Phi(x) = \rho(x) -\rho_{gas}(x)$$ for a spin system, it is a vector field ($\Sigma s_i$), according to the dimensionality adopted. The critical point is particularly well characterized through the correlation functions. The two-point correlation function is $$G^{(2)} (r) = < \Phi(0). \Phi(r) >$$ where brackets mean statistical (thermal) average, over all configurations, and the connected correlation function is $$G^{(2)}_c (r) = < \Phi(0). \Phi(r) > - |<\Phi>|^2$$ which is independent of the mean value. At critical point, for large distances $r$ $$G^{(2)}_c (r) \propto r^{2-d-\eta}$$ where $d$ id the space dimension, and $\eta$ a critical exponent. Far from the critical point, $$G^{(2)}_c (r) \propto \, exp(-r/\xi)$$ where $\xi \propto | T - T_c |^{-\nu}$ is the correlation length, and $\nu$ another critical exponent.
### Universality, Dimensionality, Symmetry
Experiments have shown that the critical exponents for a wide variety of systems are the same, and more precisely they belong to universality classes, depending only on the dimensionality $d$ of space and $D$ of the order parameter (for instance if the field is scalar or a vector with dimension $D$). This universality means that the details of the local forces are unimportant; therefore the local interactions can be simply modelled, through a schematic hamiltonian supposed to hold the relevant symmetries of the system.
Generally, the thermodynamic functions, such as the free energy $F$ (from which the heat capacity $C = -T \partial^2F/\partial T^2$ can be derived) can be decomposed in a regular part, and a singular part. The latter contain non-integer exponents (such as some derivatives diverge) as a function of $|T - T_c|$. Only the singular part with its critical exponents are universal, but the regular part could be dominant for some functions.
The 6 critical exponents, $\alpha$, $\beta$, $\gamma$, $\delta$, $\nu$ and $\eta$ are in fact not independent. From the Widom & Kadanoff scaling hypotheses, four relations can be derived between them. These relations can be demonstrated rigorously through renormalization theory. They are: $$2 \beta + \gamma = 2 - \alpha = 2 \beta \delta - \gamma$$ $$\gamma= \nu (2 - \eta)$$ $$\nu d = 2 - \alpha$$ where d is the space dimension of the problem. There exists therefore only 2 independent critical exponents.
The Ising Model
----------------
The essential physical phenomena occuring when a system undergoes a continuous phase transition are reproduced in simplified models, that have played the role of prototypes. They consider a field (order parameter) defined on a lattice of N sites, of dimension $d$. For the Ising model of ferromagnets, the field is the spin value at each site, i.e. $s_i$. We will consider only two possible values for $s_i$, +1 or -1, so that the order parameter is a scalar ($D=1$).
The Ising model was resolved for $d=1$ by Ising in 1925, for $d=2$ by Onsager in 1944, and with B non zero only recently (Zamalodchikov 1989).There is not yet any analytical solution for $d=3$.
The Hamiltonian of the system can be written: $$H= 1/2 \sum_{ij} J_{ij} s_i s_j -B \sum_i s_i$$ where $B$ is the external magnetic field, and the interaction constant $$J_{ij} = J$$ if $i$ and $j$ are neighbouring sites, and 0 otherwise. For ferromagnets, $J < 0$ and spins tend to align parallel to one another, while $J > 0$ will correspond to anti-ferromagnets.
The partition function of the system in zero $B$ can be written as: $$Z = \sum_{si} exp(-\beta H) = \sum_{si} exp(-0.5\beta \sum_{ij} J_{ij} s_i s_j)$$ Any physical quantity can then be derived, by summing over the possible configurations $\alpha$, such as the average energy: $$U = <E> = 1/Z \sum_\alpha E\alpha \, \, exp(-\beta E\alpha) = - \partial
logZ/\partial \beta_V$$ and the heat capacity in particular: $$<E^2> - <E>^2 = \partial^2logZ/\partial\beta^2_V = C_V /k\beta^2$$ In the same way, correlation functions are expressed as second derivatives of the partition function. Since these summations are in general intractable analytically, one has recourse to numerical techniques.
Monte-Carlo Numerical Calculations
----------------------------------
The principle is to calculate directly the statistical averages $$<X> = 1/Z \sum_{\alpha} X\alpha \, \, exp(-\beta E\alpha)$$ But the number of terms to compute grows exponentially with the size of the system; let us consider for example the Ising model in $d=2$, on a small lattice 3x3, the number of configurations is $2^9 = 512$, which induces reasonable computations. But for a simple lattice 10x10, there are already $2^{100} \approx 1.3 \, 10^{30}$ configurations; even with a sustained computational speed of 1 GFlop, this will require more than 3 $10^{13}$ yr, i.e. more than a thousand Hubble times. Clearly, the method is to consider judicious samples of the ensemble of configurations. There are configurations that are much less probable, and they will contribute negligibly.
The most simple method is the Metropolis one; it computes the statistical averages in a certain number of steps (typically 10$^6$ steps or more), each step corresponding to a configuration of the system. Initially, the system is placed in a simple state (for instance all spins aligned), but far from equilibrium. A certain number of steps are run without computing averages, to let the system evolve far from the inital state, and the averages be independent of it. The rule to change the configuration $\alpha$ to $\alpha'$ from one step to the next is the simplest possible, i.e. reverse only one spin (so that the energy change between the two states involves only the nearest neighbours), and the probability that this state is effectively selected is: $$P(\alpha \longrightarrow \alpha') =1
\; {\rm if}\; \Delta E = E\alpha'-E\alpha < 0$$ $${\rm and} \;
P(\alpha \longrightarrow \alpha') = exp(-\beta \Delta E)
\; {\rm otherwise} \;$$ In this manner, the system is always in a very probable state, with the probability corresponding exactly to the actual one. The averages can therefore be taken with equal weight for each of these states.
One big problem of this method, is that the successive states are not independent, but correlated, and the more so as we approach the critical point. If the states were quasi-independent, the noise on the computed averages will go down as $n_{step}^{-1/2}$, but in fact it goes down slower. The correlation time between configurations varies according to the correlation length $\xi$ as $$\tau = \xi^z$$ where $z$ is a power close to 2. Near the critical point, where $\xi$ diverges, there is a considerable slowing down of the method.
Other methods can cure this problem, such as the Swendsen-Wang algorithm or the Wolff method. The idea is to reverse blocks of spins simultaneously, from one step to the next, and to reduce existing correlations, in such a manner that existing clusters of aligned spins are broken in smaller clusters, or different clusters. The algorithm consists of linking all aligned spins by bonds, and so defining connected clusters. These bonds are then destroyed randomly, but with probability $exp(4 \beta J)$. The smaller clusters resulting from this are re-oriented (up or down) randomly with equal probability. In some cases, it can be shown that the power $z$ is then reduced to zero. In general computing the critical exponents through direct Monte-Carlo methods is not precise enough, and renormalization techniques are required.
Renormalization
-----------------
Renormalization techniques were developped much earlier in quantum field theory (e.g. Gell-Mann & Low 1954), but their application in statistical physics awaited the 1970s (Wilson & Kogut 1974; Wilson 1975, 1983), although Kadanoff (1966) presented already ideas anunciating how critical exponents could be extracted simply, with great intuition about the physical processes giving rise to critical phenomena.
The principle of a renormalization in real space (as opposed to conjugate space), is schematized in fig \[fig3\]. In the transformation of renormalization, the scales are divided by a certain factor $k$, blocks of a certain number of sites ($k^2$) are replaced by one site, and since the system is scale-independent, we should be able to find an hamiltonian for the blocks which is of the same structure as the original one. The new system is less critical than the previous one, since the correlation length $\xi$ has also been divided by $k$. It is a way to reduce the number of degrees of freedom of the system.
If the transformation of renormalization is called [**R**]{} successive hamiltonians are related by $$H_{n+1} = {\bf R} H_n$$ At each step, a similar hamiltonian must be found, with however different values of its parameters: coupling constant, temperature, etc..
The fixed points of the transformation, which obey $$H = {\bf R} H$$ can be of various nature: attractive or repulsive (or mixed). They are attractive when after several iterations, neighbouring points are trapped there. In many problems, there exist trivial points, in the asymptotic regimes of low ($T \longrightarrow 0$) or high ($T \longrightarrow \infty$) temperatures. But in between, there can exist fixed points, corresponding to the critical points ($T = T_c$).
To give a very simple example, let us renormalize the Ising $d=1$ problem, which can be carried out easily (cf Lesne 1996). The renormalization transformation will consist in considering only the even-numbered sites, and drop the odd-numbered ones. The scale is then divided by 2. The partition function can be written as: $$Z = \sum_j exp( K_0 \, s_{2j+1} (s_{2j} + s_{2j+2}))$$ It is easy to demonstrate, since the spins can take only the values +1 and -1, that: $$exp( K_0 (s_3s_2 + s_3s_4)) = 2 exp (K_1) exp( K_1 s_2s_4)$$ Then the partition function is exactly of the same form as the original one, after the division by 2, with $$th(K_1) = th^2 (K_0)$$ (or $exp(2 K_1) = ch^2 (K_0)$). Let us consider now the fixed points, where $K_1$ = $K_0$, i.e. $th (K)$ = 0, or $th (K)$ = 1. This arrives in the two extreme cases, $K \longrightarrow 0$ and $K \longrightarrow \infty$. $K$ is in fact the product of the coupling constant $J$, and $\beta$. The first case corresponds, for a finite temperature, to zero coupling, that is a purely thermal system, paramagnetic only; at zero $B$ there is no correlation, and the system is not critical. Since the transform of $K$ in this region is $\sim K^2$, and $\xi$ is divided by 2 at each transform, we can deduce $\xi(K) = 2 \xi(K^2)$, and the behaviour $\xi(K) \propto 1/|log(K)|$, i.e. $\xi \longrightarrow 0$. The second fixed point corresponds to a zero temperature ($K \longrightarrow \infty$). Then we can show that $\xi(K) \propto exp(2 K)$, so $\xi$ diverges, with large correlations. The system is truly critical.
In the general case, it is quite complex to find the easiest way to build blocks, and the new hamiltonian. A possible way is to Monte-Carlo renormalize, i.e. represent the transformation matrix of [**R**]{} as a statistical average on the several possible choices. The critical exponent is then the largest eigen value of this matrix.
Mean Field Approximation
-------------------------
This simple method to resolve the problem consists, for the Ising model of ferromagnet as an example, to replace the action of the ensemble of spins on a particular site, by the mean magnetic field at this position (it is also independent of site by translational invariance). This gives approximately when the system becomes critical (in general $T_c$ is overestimated), and approximate values of the critical exponents that do not depend on the spatial dimensionality (more exact when the latter is large, or for infinite-range interactions). For the 3D Ising model ($d=3$), the critical exponents can be determined numerically to: $\alpha$ = 0.107, $\gamma$ = 1.239, $\nu$ = 0.631 and $\eta$ = 0.037; while in the mean field theory $\alpha$ = 0, $\gamma$ = 1, $\nu$ = 0.5 and $\eta$ = 0; the approximation appears therefore relatively bad.
Functional Integrals
---------------------
To efficiently use the renormalization methods, it is fruitful to go back to the field theory, domain in which the renormalization was first introduced, and numerous techniques are available. For example, in the case of the system of spins in a discrete lattice $s_i$, the discrete configurations of the lattice are replaced by a continuous field $s(\vec x)$. The summation on the sites $\Sigma_i$ are replaced by spatial integrals $\int d\vec x$; the hamiltonian becomes a functional of the field $s(\vec x)$, $$H(s,T) = \beta \int A(\vec s,\vec x) d\vec x$$ The summation over various configurations, to obtain for instance the partition function $Z$, is replaced by a functional integral, of variable the field $s(\vec x)$: $$Z = \int_{\vec s} exp(-H(\vec s, T)$$ Often the calculations are more easy in the Fourier conjugate space, the renormalised hamiltonian $H_1$ is defined of the same form as the original one $H_0$, as $$exp(H_1) = \int_{\phi} exp(H_0(\Phi_0 + \phi))$$ The renormalised hamiltonian $H_1$ can be expressed as a perturbative development, with the help of a diagrammatic analysis.
Statistical Self-Gravity
==========================
As was shown is section 4, it appears that the self-gravitating system is critical for a large range of the parameters, and it is difficult to isolate a critical point, to identify diverging behaviours. However, it is well known (Wilson 1975, Domb & Green 1976), that physical quantities diverge only for infinite volume systems, at the critical point. Since our systems are also finite and bounded, they only approach asymptotically the divergences.
Identification of the Fractal Dimension
---------------------------------------
If $ \Lambda $ measures the distance to the critical point, (in spin systems for instance, $ \Lambda $ is proportional to $| T - T_c |$), the correlation length $ \xi $ diverges as, $$\xi( \Lambda ) \sim \Lambda^{-\nu} \;$$ and the specific heat (per unit volume) $ {\cal C} $ as, $${\cal C} \sim \Lambda^{-\alpha} \;$$ But in fact, for a finite volume system, all physical quantities are finite at the critical point. When the typical size $R$ of the system is large, the physical magnitudes take large values at the critical point, and the infinite volume theory is used to treat finite size systems at criticality. In particular, for our system, the correlation length provides the relevant physical length $ \xi \sim R $, and we can write $$\Lambda \sim R^{-1/\nu} \;$$
Our system has the symmetries $d=3$ and $D=1$ (scalar field), which should indicate the universality class to which it corresponds. It remains to identify the correponding operators. Already in the previous sections, it was suggested that the field $\phi$ corresponds to the potential, and the mass density $$m\, \rho({\vec x}) = m\, \, e^{\phi({\vec x})}$$ can be identified with the energy density in the renormalization group (also called the ‘thermal perturbation operator’).
We note that the state of zero density (or zero fugacity), corresponds to a singular point, around which we develop the physical functions (and we choose $\Lambda$ accordingly). At this point $ \mu^2/T_{eff} = 0 $, the partition function $ {\cal Z}
$ is singular $$\Lambda \equiv {{\mu^2}\over{T_{eff}}} = z\,
\left({{2\pi mkT}\over{h^2}}\right)^{3/2} \;$$ i.e., the critical point $ \Lambda = 0 $ corresponds to zero fugacity $z$. We can write $ {\cal Z} $ as a function of the action $S^*$ at the critical point $${\cal Z}(\Lambda) =
\int\int\; {\cal D}\phi\; e^{ -S^* + \Lambda
\int d^3x \; e^{\phi({\vec x})}\;}$$
Let us now decompose log$ {\cal Z} $ in its singular and regular parts (we know that the second derivative will give the heat capacity that diverges in $\Lambda^{-\alpha}$)
$${1 \over V} \; \log{\cal Z}(\Lambda) = {K \over{(2-\alpha)(1-\alpha)}}\;
\Lambda^{2-\alpha} + F(\Lambda) \;$$
where $ F(\Lambda) $ is an analytic function of $ \Lambda $ around the origin $$F(\Lambda) = a \; \Lambda + \frac12 \, b \; \Lambda^2 + \ldots$$ $ V = R^3 $ stands for the volume and $ K, \; a $ and $ b $ are constants.
By derivation with respect to $ \Lambda $ $${1 \over V} \;{{\partial}\over{\partial\Lambda}}\log{\cal Z}(\Lambda)=
a + {K \over{1-\alpha}}\,
\Lambda^{1-\alpha}
+ \ldots = {1 \over V} \int d^3x \; <e^{\phi({\vec x})}>\;$$ Now, using the standard relation $ \alpha = 2 - \nu d = 2 - 3 \nu$, this gives $${{\partial}\over{\partial\Lambda}}\log{\cal Z}(\Lambda)= V \, a +
{K \over{1-\alpha}}\, R^{1/\nu} + \ldots$$ The mass contained in a region of size $ R $ is $$M(R) = m \int^R e^{\phi({\vec x})} \; d^3x$$ $$<M(R)> = m \, V \, a +m \, {K \over{1-\alpha}}\; R^{ \frac1{\nu}} +
\ldots$$ We already see that, apart a possible constant density, the average mass obeys a singular power-law.
The 2-points density correlator varies as $$C({\vec r_1},{\vec r_2})\sim |{\vec r_1} -{\vec r_2}|^{\frac2{\nu} -6}$$ The perturbative calculation matches with this result for $ \nu = \frac12 $, the mean field value for the exponent $ \nu $. Pursuing further to the second derivative of $ \log{\cal Z}(\Lambda) $ with respect to $\Lambda$ $${{\partial^2}\over{\partial\Lambda^2}}\log{\cal Z}(\Lambda)= V\left[
\Lambda^{-\alpha} \, K + b + \ldots \right]$$ $${{\partial^2}\over{\partial\Lambda^2}}\log{\cal Z}(\Lambda)=
\int d^3x\; d^3y\; C({\vec x},{\vec y}) \sim R^D \int^R
{{ d^3x}\over{x^{6 - 2d_H}}} \sim \Lambda^{-2}\sim R^D \;
\Lambda^{-\alpha}$$ we can find the mass fluctuations and corresponding dispersion: $$(\Delta M(R))^2 \equiv \; <M^2> -<M>^2 \sim
\int d^3x\; d^3y\; C({\vec x},{\vec y}) \sim R^{2/\nu}$$ $$\Delta M(R) \sim R^{1/\nu}$$ This is the definition relation of the fractal, with dimension $d_H$, and the scaling exponent $\nu$ can be identified with the inverse Haussdorf dimension of the system $$d_H = \frac1{\nu}$$ The velocity dispersion follows $$\Delta v \sim R^{q}$$ with $$q =\frac12\left(\frac1{\nu} -1\right) =\frac12(d_H -1) \;$$
Numerical Values
----------------
The scaling exponents $ \nu , \; \alpha $ can be computed through the renormalization group approach. The case of a single component (scalar) field has been extensively studied in the literature (Hasenfratz & Hasenfratz 1986, Morris 1994a,b). Very probably, there is a unique, infrared stable fixed point in 3D: the Ising model fixed point. Such non-perturbative fixed point is reached in the long scale regime independently of the initial shape of the interaction ($\phi$). For the Ising model $d=3$, the exponents are: $\nu$ = 0.631, from which we deduce $d_H$ = 1.585, $\eta$ = 0.037, and $\alpha$= 0.107.
The value of the dimensionless coupling constant $g^2 = \mu T_{eff}$ should decide whether the fixed point chosen by the system is the mean field (weak coupling) or the Ising one (strong coupling). At the tree level, we estimate $g \approx \frac{5}{\sqrt{N}}$, where $N$ is the number of points in a Jeans volume $d_J^3$. The coupling constant appears then of the order of 1, and we cannot settle this question without effective computations of the renormalization group equations. In any case, both the Ising and mean field values are in agreement with the astronomical observations (the mean field exponents are $\nu$ = 0.5 $d_H$ = 2, $\eta$ =0 and $\alpha$ = 0).
Important Differences with the Spin Models
-------------------------------------------
For the gravitational gas we find scaling behaviour for a full range of temperatures and couplings, while for spin models scaling only appears at the critical value of the temperature. At $T = T_c$, the correlation length $\xi$ is infinite, and the theory is massless. In fact, since the spin systems are not infinite, even in the critical domain, $ \xi $ is finite and the correlation functions decrease as $ \sim e^{ - r/\xi} $ for large distances $ r $ (only $\xi$ is large, of the order of the system size). Fluctuations of the relevant operators support perturbations which can be interpreted as massive excitations.
Such (massive) behaviour does not appear for the gravitational gas. The density correlators scale, exhibiting power-law behaviour. This feature is connected with the scale invariant character of the Newtonian force and its infinite range.
Observational Tests
-------------------
It was shown that the predicted fractal dimension of a self-gravitating critical medium is compatible with that of the astrophysical applications (interstellar medium, galaxies), within the observational uncertainties. However, there are other predictions of the theory, that could be checked. It is well known in critical phenomena that there exists two independent critical exponents: the second one concerns the correlations of the potential, corresponding to the field $\phi$ $$<\phi({\vec r})\phi(0)> \sim r^{-1-\eta} \;$$
The potential is then predicted to vary as a power law with size, with a slope $-1/2 -\eta/2$ = -0.518. It is not easy to observe directly this quantity, but tracers of the potential could be obtained through light rays that are deviated by astrophysical masses. As for the ISM, for example, the potential fluctuations could be traced by the micro-lensing produced on the light of remote quasars, while they are crossing the ISM of intervening galaxies, at high redshift. Time fluctuations of the quasars flux is regularly observed, and interpretation in terms of micro-lensing has been proposed (Lewis et al 1993).
Conclusion
===========
We have emphasized the existence of two astrophysical fractals, the interstellar medium, with structures ranging from 10 AU to 100 pc, and the large-scale structures of galaxies, from 50 kpc to 150 Mpc at least. The first one is in statistical equilibrium, while the second one is still growing to larger scales. In both cases, we can describe these media as developping large-scale fluctuations with large correlations as is familiar in critical phenomena. We propose that in both cases, self-gravity is the main force governing these fractal structures.
The statistical thermodynamic approach is developped, and it is shown that the phenomenological potential, which is in $1/r$ between two cutoffs (at small and large-scale), can be described by a scalar field theory. We use the renormalization group methods for this scale-independent system, to find the universality class of the problem. The fractal dimension, and the potential correlations exponent can be derived from the critical exponents of the Ising $d=3$ model. The stability of the results with respect to perturbations has been studied; the results are quite robust with respect to perturbations of external forces. Of course, if the external energy is large with respect to the self-gravitational energy of the gas (in the vicinity of violent star-formation in the ISM for instance), then the hierachical structure will be destroyed. The observed fractal dimensions are compatible with the predictions, and other observational tests are proposed.
Abell G.O.: 1958, ApJS 3, 211 Balian R., Schaeffer R.: 1988, ApJ 335, L43 Balian R., Schaeffer R.: 1989, A&A 226, 373 Bazell D., Desert F.X.: 1988, ApJ 333, 353 Binney J.J., Dowrick N.J., Fisher A.J., Newman M.E.J.: 1992 ‘The Theory of Critical Phenomena’, Oxford Science Publication. Bond J.R., Cole S., Efstathiou G., Kaiser N.: 1991, ApJ 379, 440 Bouchet F.R., Schaeffer R., Davis M.: 1991, ApJ 383, 19 Brand P., Wouterloot J.G.: 1995, A&A ??? Casoli F., Combes F., Gerin M.:1984, A&A 133, 99 Castagnoli C., Provenzale A.: 1991, A&A 246, 634 Cole S., 1989, PhD Thesis, Univ. of Cambridge Coleman P.H., Pietronero L., Sanders R.H.: 1988, A&A 200, L32 Coleman P.H., Pietronero L.: 1992, Phys. Rep. 231, 311 Colombi S., Bouchet F.R., Hernquist L.: 1996, ApJ 465, 14 Combes F., Pfenniger D.: 1997, A&A 327 453 Dame T.M., Elmegreen B.G., Cohen R.S., Thaddeus P.: 1986, ApJ 305, 982 Davis M.A., Meiksin M.A., Strauss L.N., da Costa and Yahil A.: 1988, ApJ 333, L9 Davis M.A.: 1997 in ‘Critical Dialogs in Cosmology’, ed. N. Turok, astro-ph/9610149 Davis M.A., Peebles P.J.E.: 1977, ApJS 34, 425 Davis M.A., Peebles P.J.E.: 1983, ApJ 267, 465 Davis R.J., Diamond P.J., Goss W.M.: 1996, MNRAS 283, 1105 de Vaucouleurs G.: 1960, ApJ 131, 585 de Vaucouleurs G.: 1970, Science 167, 1203 de Vega H., Sánchez N., Combes F.: 1996a, Nature 383, 53 de Vega H., Sánchez N., Combes F.: 1996b, Phys. Rev. D54, 6008 de Vega H., Sánchez N., Combes F.: 1998, ApJ in press Diamond P.J., Goss W.M., Romney J.D. et al: 1989, ApJ 347, 302 Di Nella H., Montuori M., Paturel G., Pietronero L., Sylos Labini F.: 1996, A&A 308, L33 Domb C., Green M.S.: 1976, ‘Phase transitions and Critical Phenomena’, vol. 6, Academic Press Dubrulle B., Lachièze-Rey M.: 1994, A&A 289, 667 Falconer K.J.:1990, Fractal geometry, Wiley, Chichester Falgarone, E., Phillips, T.G., Walker, C.K.: 1991, ApJ 378, 186 Falgarone, E., Lis D.C., Phillips, T.G. et al: 1994, ApJ 436, 728 Fleck R.C.: 1981, ApJ 246, L151 Fiedler R.L., Dennison B., Johnston K., Hewish A.: 1987, Nature 326, 675 Fiedler R.L., Pauls T., Johnston K., Dennison B.: 1994, ApJ 430, 595 Gell-Mann M., Low F.E.: 1954, Phys. Rev. 95, 1300 Hamilton A.J.S.: 1993, ApJ 417, 19 Hasenfratz A., Hasenfratz P., 1986, Nucl.Phys. B270, 687 Heithausen A.: 1996, A&A 314, 251 Heithausen A., Bensch F., Stutzki J., Fakgarone E., Panis J.F.: 1998, A&A 331, L65 Horwitz G., Katz J., 1978a, ApJ 222, 941 Horwitz G., Katz J., 1978b, ApJ 223, 311 Houlahan P., Scalo J.: 1992, ApJ 393, 172 Hoyle F.: 1953, ApJ 118 513 Hubbard J., 1959, Phys. Rev. Lett, 3, 77 Inutsuka S-I., Miyama S.M.: 1997, ApJ 480 681 Itoh M., Inagaki S., Saslaw W.C.: 1993, ApJ 403, 459 Kadanoff L.P.: 1966 Physics 2, 263 Kolmogorov A.: 1941, in “Compt. Rend. Acad. Sci. URSS” 30, 301 Lacey C.G., Cole S.: 1993, MNRAS 262, 627 Lacey C.G., Cole S.: 1994, MNRAS 271, 676 Lachièze-Rey M.: 1993, ApJ 407, 1 & ApJ 408, 403 Larson R.B., 1981, MNRAS 194, 809 Larson R.B., 1985, MNRAS 214, 379 Lewis G.F., Miralda-Escude J., Richardson D.C., Wambsganss J.: 1993, MNRAS 261, 647 Lesne A.: 1996, in “Méthodes de renormalisation”, Eyrolles Sciences Lin H. et al: 1996, ApJ 471, 617 Magnani L., Blitz L., Mundy L. 1985, ApJ 295, 402 Mandelbrot B.B.: 1975, ‘Les objets fractals’, Paris, Flammarion Mandelbrot B.B.: 1982, ‘The fractal geometry of nature’, New York: Freeman Martinez V.J., Paredes S., Saar E.: 1993, MNRAS 260, 365 McLaughlin D.E. & Pudritz R.E.: 1996, ApJ 469, 194 Miesch M., & Scalo J.M.: 1995, ApJ 450, L27 Morris T.R., 1994a: Phys. Lett. B329, 241 Morris T.R., 1994b: Phys. Lett. B334, 355 Myers P.C.: 1983, ApJ 270, 105 Ostriker J.P.: 1993, ARAA 31, 689 Padmanabhan 1990, Phys. Rep. 188, 285 Parisi G.: 1988, ‘Statistical field theory’, Addison Wesley, Redwood City Peebles P.J.E.: 1980, ‘The Large-scale structure of the Universe’, Princeton Univ. Press Peebles P.J.E.: 1993, ‘Principles of physical cosmology’ Princeton Univ. Press Pfenniger D., Combes F., Martinet L.: 1994, A&A 285 79 Pfenniger D., Combes F.: 1994, A&A 285, 94 Pietronero L., Montuori M., Sylos Labini F.: 1997, in ‘Critical Dialogs in Cosmology’, ed. N. Turok, astro-ph/9611197 Pietronero L.: 1987, Physica A, 144, 257 Postman M., Geller M.J., Huchra J.P.: 1986, AJ 91, 1267 Postman M., Huchra J.P., Geller M.J.: 1992, ApJ 384, 404 Press W.H., Schecter P.: 1974, ApJ 187 425 Rees M.J.: 1976, MNRAS 176, 483 Saslaw W.C., Fang F.: 1996, ApJ 460, 16 Saslaw W.C., Hamilton A.J.S.: 1984, ApJ 276, 13 Scalo J.M.: 1985, in [*Protostars and Planets II*]{}, ed. D.C. Black & M.S. Matthews, Univ. of Arizona Press, Tucson, p. 201 Scalo J.M., 1987 in ‘Interstellar Processes’, D.J. Hollenbach and H.A. Thronson Eds., D. Reidel Pub. Co, p. 349 Scalo J.M., Lazarian A.: 1996, ApJ 469, 189 Schechter P.L.: 1976, ApJ 203, 297 Scott E.L., Shane S.D., Swanson M.D.: 1954, ApJ 119, 91 Shandarin S.F., Zel’dovich Ya. B.: 1989, Rev. Mod. Phys. 61, 185 Shapiro P.R., Kang H.: 1987, ApJ 318 32 Shapley H.: 1934, MNRAS 94, 791 Sheth R.K., Saslaw W.C.: 1996, Apj 470, 78 Smoot G., et al: 1992, ApJ 396, L1 Solomon P.M., Rivolo A.R., Barrett J.W., Yahil A.: 1987, ApJ 319, 730 Soneira R.M., Peebles P.J.E.: 1978, AJ 83, 845 Sreenivasan K.R., & Méneveau C.: 1986, J. Fluid Mech. 173, 357 Stratonovich R.L., 1958, Doklady, 2, 146 Sylos Labini F., Amendola L.: 1996, ApJ 438, L1 Sylos Labini F., Gabrielli A., Montuori M., Pietronero L.: 1996, Physica A 226, 195 Sylos Labini F., Pietronero L.: 1996, ApJ 469, 26 Sylos Labini F.: 1994, ApJ 433, 464 Toomre A.: 1964, ApJ 139, 1217 Totsuji H., Kihara T.: 1969; PASJ 21, 221 Vazquez-Semadeni E.: 1994, ApJ 423, 681 Vazquez-Semadeni E., Ballesteros-Paredes J., Rodriguez L.F.: 1997, ApJ 474, 292 Vazquez-Semadeni E., Canto J., Lizano S.: 1998, ApJ 492, 596 Uehara H., Susa H., Nishi R., Yamada M., Nakamura T.: 1996, ApJ 473 L95 Vergassola M., Dubrulle B., Frisch U., Noullez A.: 1994, A&A 289, 325 Vogelaar M.G.R., Wakker B.P.: 1994, A&A 291, 557 White S.D.M., Rees M.J.: 1978, MNRAS 183 341 White S.D.M.: 1979, MNRAS 186, 145 White S.D.M., Kauffmann G.: 1994, in “The formation and evolution of galaxies”, V Canary Islands Winter School, ed. C. Munoz-Tunon & F. Sanchez, p. 455 Wilson K.G., Kogut, J.: 1974, Phys. Rep. 12, 75 Wilson K.G.: 1975, Rev. Mod. Phys. 47, 773 Wilson K.G.: 1983, Rev. Mod. Phys. 55, 583 Zamalodchikov A.B. 1989, Int. J. Mod. Phys. A4, 4235 Zel’dovich Ya. B.: 1970, Astrofizika 6, 319 Zinn-Justin J.: 1989, ‘QFT and Critical Phenomena’, Clarendon Press, Oxford
[^1]: The notion of correlation length $\xi_0$ is usually different in physics, where $\xi_0$ characterizes the exponential decay of correlations $ (\sim
e^{- r/ \xi_0} ) $. For power decaying correlations, it is said that the correlation length is infinite
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we establish an attainability result for the minimum time function of a control problem in the space of probability measures endowed with Wasserstein distance. The dynamics is provided by a suitable controlled continuity equation, where we impose a nonlocal nonholonomic constraint on the driving vector field, which is assumed to be a Borel selection of a given set-valued map. This model can be used to describe at a macroscopic level a so-called *multiagent system* made of several possible interacting agents.'
author:
- Giulia Cavagnari
- Antonio Marigonda
title: Attainability property for a probabilistic target in Wasserstein spaces
---
Introduction {#sec:intro}
============
We consider a finite-dimensional *multiagent system*, i.e., a system in $\mathbb R^d$ where the number of agents is so large that only a *macroscopic* description is available. As usual in this framework, in order to describe the behaviour of the system at a certain time $t$, we introduce a Borel positive measure $\mu_t$ on $\mathbb R^d$ whose meaning is the following: given a Borel set $A\subseteq \mathbb R^d$ the quantity $\dfrac{\mu_t(A)}{\mu_t(\mathbb R^d)}$ represents the fraction of the total number of agents that are present in $A$ at the time $t$. We will assume that the system is *isolated*, thus the total number of agents remains constant in time. Hence, by normalizing the measure $\mu_t$, we can always assume $\mu_t(\mathbb R^d)=1$, i.e., $\mu_t$ is a probability measure for all $t$.
The macroscopic evolution of the system is thus given by a curve $t\mapsto\mu_t$ in the space of probability measures. Due to the mass-preserving character of the evolution, we can assume that such an evolution is governed by the *continuity equation* $$\partial_t\mu_t+\mathrm{div}(v_t\mu_t)=0,$$ to be satisfied in a distributional sense, where $v_t$ is a suitable time-depending Borel vector field describing the macroscopic mass flux during the evolution.
It can be easily proved, see e.g. [@CM-LSSC17], that for a.e. $t$ and $\mu_t$-a.e. $x\in\mathbb R^d$ the vector field $v_t(x)$ can be constructed as a weigthed average of the velocities of the agents passing through the point $x$ at time $t$, where the weights are given by the fraction of the mass carried by each agent w.r.t. the total amount of mass flowing through $x$ at time $t$. In particular, possibly nonlocal nonholonomic constraints on the agents’ motion will reflect into constraints for the possible choices of $v_t$.
In this paper we consider a situation where each agent is constrained to follow the trajectories of a differential inclusion with a *nonlocal* dependence on the overall configuration of the agents. This fact models the possible *nonlocal interaction* among the agents. Examples of such interactions are quite commmon in the models of pedestrian dynamics, flocks of animals and social dynamics in general.
Due to the potential applications, the literature on control of multi-agent systems is growing quite fast in the recent years. Among the most recent contributions, we mention [@CFPT], where the authors investigate a controllability problem for a leader-follower model in a finite-dimensional setting and their aim is to achieve an alignment consensus for a mass of indistinguishable agents when the action of an external policy maker is *sparse*, i.e. concentrated on few individuals. In [@FPR] it is provided a mean-field formulation of the same model through Gamma-convergence techniques.
The relevance of such kind of results is enhanced when dealing with problems involving a considerable number of individuals, in order to circumvent the bounds coming from the curse of dimensionality: indeed, the mean-field limit can be used as a realistic approximation when the number of agents is huge. Results in this direction are provided for example by [@FLOS] or the preprint paper [@CLOS], where the authors study a Gamma-convergence result for an optimal control problem of a $N$-particles system subject to a nonlocal dynamics when $N\to +\infty$.
Controllability conditions in the space of probability measures are also analyzed in the preprints [@DMR1], [@DMR2]. In particular, the aim of the authors is to provide sufficient conditions in order to steer an initial configuration of agents into a desired final one, by acting through a control term on the vector field, under the constraint that the action can be implemented only in a certain fixed space region.
Also the extension of classical *viability theory* to multi-agent systems is attracting an increasing interest in the community. Similarly to the finite-dimensional framework, a subset $\mathscr K$ of probability measures is said to be viable for a controlled dynamics if it is possible to keep the evolution confined inside $\mathscr K$ by acting with an admissible control when starting with a initial state in $\mathscr K$. We refer to [@Aver] for first results in this direction.
It is worth pointing out that a key feature of all these studies, and many others available in the literature, is the combined use of tools, concepts, and techniques from optimal transport theory, measure theory, and from optimal control theory.
In our framework we will consider a time-optimal control problem, i.e., we deal with a target set of desired final configurations, and the minimum time needed by the agents to obtain it from an initial datum and obeying to the nonholonomic constraints.
Such construction defines the so-called *minimum time function*, which is a central object in optimal control theory. In the case without interactions, in [@CMNP] the authors proved that the minimum time function solves in a suitable viscosity sense an Hamilton-Jacobi-Bellman equation in the spaces of measures provided that it is continuous, and further development on this theory have been recently done in [@MQ; @JMQ]. We refer the reader to [@AGa], [@GT2019], [@GS] for an introduction to Hamilton-Jacobi equations in Wasserstein spaces.
Our aim is to provide a sufficient condition for the continuity of the minimum time function in this framework, i.e., sufficient conditions granting Small Time Local Attainability (STLA) in the sense of [@KQ]. The paper is structured as follows: in Section \[sec:notation\] we fix the notation and review some basic results about measure theory, optimal transport and set-valued analysis, in Section \[sec:admtraj\] we prove some basic properties of the admissible trajectories in the space of measures, in Section \[sec:gentarget\] we discuss some geometric properties of the target sets, and finally in Section \[sec:mintime\] we state our main result concerning the continuity of the minimum time function.
Preliminaries and notation {#sec:notation}
==========================
In this section we review some concepts from measure theory, optimal transport and set-valued analysis. Our main references for this part are [@AGS], [@AuF], and [@Vil].
We will use the following notation.
--------------------------------------- -------------------------------------------------------------------------------------------------------------
$B(x,r)$ the open ball of center $x\in X$ and radius $r$ of a normed space $X$,
i.e., $B(x,r):=\{y\in X:\,\|y-x\|_X<r\}$;
$\overline K$ the closure of a subset $K$ of a topological space $X$;
$d_K(\cdot)$ the distance function from a subset $K$ of a metric space $(X,d)$,
i.e. $d_K(x):=\inf\{d(x,y):\,y\in K\}$;
$C^0_b(X;Y)$ the set of continuous bounded function from a Banach space $X$ to $Y$,
endowed with $\|f\|_{\infty}=\displaystyle\sup_{x\in X}\|f(x)\|_Y$ (if $Y=\mathbb R$, $Y$ will be omitted);
$C^0_c(X;Y)$ the set of compactly supported functions of $C^0_b(X;Y)$,
with the topology induced by $C^0_b(X;Y)$;
$\Gamma_I$ the set of continuous curves from a real interval $I$ to $\mathbb R^d$;
$\Gamma_T$ the set of continuous curves from $[0,T]$ to $\mathbb R^d$;
$AC([0,T])$ the set of absolutely continuous curves from $[0,T]$ to $\mathbb R^d$;
$e_t$ the evaluation operator $e_t:\mathbb R^d\times\Gamma_I\to\mathbb R^d$
defined by $e_t(x,\gamma)=\gamma(t)$ for all $t\in I$;
$\mathscr P(X)$ the set of Borel probability measures on a Banach space $X$,
endowed with the weak$^*$ topology induced by $C^0_b(X)$;
$\mathscr M(\mathbb R^d;\mathbb R^d)$ the set of vector-valued Borel measures on $\mathbb R^d$ with values in $\mathbb R^d$,
endowed with the weak$^*$ topology induced by $C^0_c(\mathbb R^d;\mathbb R^d)$;
$|\nu|$ the total variation of a measure $\nu\in \mathscr M(\mathbb R^d;\mathbb R^d)$;
$\ll$ the absolutely continuity relation between measures defined on the same
$\sigma$-algebra;
$\mathrm{m}_p(\mu)$ the $p$-moment of a probability measure $\mu\in \mathscr P(X)$;
$r\sharp\mu$ the push-forward of the measure $\mu$ by the Borel map $r$;
$\mu\otimes\eta_x$ the product measure of $\mu\in\mathscr P(X)$ with the Borel family of measures
$\{\eta_x\}_{x\in X}$;
$\mathrm{pr}_i$ the $i$-th projection map $\mathrm{pr}_i(x_1,\dots,x_N)=x_i$;
$\Pi(\mu,\nu)$ the set of admissible transport plans from $\mu$ to $\nu$;
$\Pi_o(\mu,\nu)$ the set of optimal transport plans from $\mu$ to $\nu$;
$W_p(\mu,\nu)$ the $p$-Wasserstein distance between $\mu$ and $\nu$;
$\mathscr P_p(X)$ the subset of the elements $\mathscr P(X)$ with finite $p$-moment,
endowed with the $p$-Wasserstein distance;
$\mathscr L^d$ the Lebesgue measure on $\mathbb R^d$;
$\dfrac{\nu}{\mu}$ the Radon-Nikodym derivative of the measure $\nu$ w.r.t. the measure $\mu$;
$\mathrm{Lip}(f)$ the Lipschitz constant of a function $f$.
--------------------------------------- -------------------------------------------------------------------------------------------------------------
In this section we give some preliminaries and fix the notation. Our main reference for this part is [@AGS].
Given Banach spaces $X,Y$, we denote by $\mathscr P(X)$ the set of Borel probability measures on $X$ endowed with the weak$^*$ topology induced by the duality with the Banach space $C^0_b(X)$ of the real-valued continuous bounded functions on $X$ with the uniform convergence norm. For any $p\ge 1$, the $p$-moment of $\mu\in\mathscr P(X)$ is defined by $\displaystyle\mathrm{m}_p(\mu)=\int_{X}\|x\|_X^p\,d\mu(x)$, and we set $\mathscr P_p(X)=\{\mu\in\mathscr P(X):\, \mathrm{m}_p(\mu)<+\infty\}$. For any Borel map $r:X\to Y$ and $\mu\in\mathscr P(X)$, we define the *push forward measure* $r\sharp\mu\in\mathscr P(Y)$ by setting $r\sharp\mu(B)=\mu(r^{-1}(B))$ for any Borel set $B$ of $Y$.
Let $X,Y$ be Banach spaces, and denote by $\mathscr M(X;Y)$ the set of $Y$-valued Borel measures defined on $X$. The total variation measure of $\nu\in\mathscr M(X;Y)$ is defined for every Borel set $B\subseteq X$ as $$|\nu|(B):=\sup_{\{B_i\}_{i\in\mathbb N}}\left\{\sum \|\nu(B_i)\|_Y\right\}$$ where the sup ranges on the set of countable collections $\{B_i\}_{i\in\mathbb N}$ of pairwise disjoint Borel sets such that $\displaystyle\bigcup_{i\in\mathbb N}B_i=B$.
For the following result see [@AGS Theorem 5.3.1].
Given a measure $\mu\in\mathscr P(\mathbb X)$ and a Borel map $r:\mathbb X\to X$, there exists a family of probability measures $\{\mu_x\}_{x\in X}\subseteq \mathscr P(\mathbb X)$, uniquely defined for $r\sharp \mu$-a.e. $x\in X$, such that $\mu_x(\mathbb X\setminus r^{-1}(x))=0$ for $r\sharp \mu$-a.e. $x\in X$, and for any Borel map $\varphi:\mathbb X\to [0,+\infty]$ we have $$\int_{\mathbb X}\varphi(z)\,d\mu(z)=\int_X \left[\int_{r^{-1}(x)}\varphi(z)\,d\mu_x(z)\right]d(r\sharp \mu)(x).$$ We will write $\mu=(r\sharp \mu)\otimes \mu_x$. If $\mathbb X=X\times Y$ and $r^{-1}(x)\subseteq\{x\}\times Y$ for all $x\in X$, we can identify each measure $\mu_x\in\mathscr P(X\times Y)$ with a measure on $Y$.
Let $X$ be a complete separable Banach space, $\mu_1,\mu_2\in\mathscr P(X)$. We define the set of *admissible transport plans* between $\mu_1$ and $\mu_2$ by setting $$\Pi(\mu_1,\mu_2)=\{\boldsymbol{\pi}\in\mathscr P(X\times X): \mathrm{pr}_i\sharp\boldsymbol\pi=\mu_i,\,i=1,2\},$$ where for $i=1,2$, we defined $\mathrm{pr}_i:\mathbb R^d\times\mathbb R^d\to\mathbb R^d$ by $\mathrm{pr}_i(x_1,x_2)=x_i$. The *inverse* $\boldsymbol\pi^{-1}$ of a transport plan $\boldsymbol\pi\in \Pi(\mu,\nu)$ is defined by $\boldsymbol\pi^{-1}=i\sharp \boldsymbol\pi\in \Pi(\nu,\mu)$, where $i(x,y)=(y,x)$ for all $x,y\in X$. The *$p$-Wasserstein distance* between $\mu_1$ and $\mu_2$ is $$W_p^p(\mu_1,\mu_2)=\inf_{\boldsymbol\pi\in\Pi(\mu_1,\mu_2)}\int_{X\times X}|x_1-x_2|^p\,d\boldsymbol\pi(x_1,x_2).$$ If $\mu_1,\mu_2\in\mathscr P_p(X)$ then the above infimum is actually a minimum, and we define $$\Pi^p_o(\mu_1,\mu_2)=\left\{\boldsymbol\pi\in\Pi(\mu_1,\mu_2): W_p^p(\mu_1,\mu_2)= \int_{X\times X}|x_1-x_2|^p\,d\boldsymbol\pi(x_1,x_2)\right\}.$$ The space $\mathscr P_p(X)$ endowed with the $W_p$-Wasserstein distance is a complete separable metric space, moreover for all $\mu\in \mathscr P_p(X)$ there exists a sequence $\{\mu^N\}_{N\in\mathbb N}\subseteq\mathrm{co}\{\delta_x:\,x\in\mathrm{supp}\,\mu\}$ such that $W_p(\mu^N,\mu)\to 0$ as $N\to +\infty$.
Recalling formula (5.2.12) in [@AGS], we have $$W_p(\delta_0,\mu)=m_{p}^{1/p}(\mu)=\left(\int_{\mathbb R^d}|x|^p\,d\mu(x)\right)^{1/p}$$ for all $\mu\in \mathscr P_p(\mathbb R^d)$. In particular, if $t\mapsto \mu_t$ is $W_p$-continuous, then $t\mapsto \mathrm{m}_p^{1/p}(\mu_t)$ is continuous.
Let $X,Y$ be sets. A *set-valued* map $F$ from $X$ to $Y$ is a map associating to each $x\in X$ a (possible empty) subset $F(x)$ of $Y$. We will write $F:X\rightrightarrows Y$ to denote a set-valued map from $X$ to $Y$. The *graph* of a set-valued map $F$ is $$\mathrm{graph}\,F:=\{(x,y)\in X\times Y:\,y\in F(x)\}\subseteq X\times Y,$$ while the *domain* of $F$ is $\mathrm{dom}\,F:=\{x\in X:\, F(x)\ne\emptyset\}\subseteq X$. A *selection* of $F$ is a map $f:\mathrm{dom}\,F\to Y$ such that $f(x)\in F(x)$ for all $x\in\mathrm{dom}\,F$. When $X,Y$ are topological spaces, we say that
- $F$ has *closed images* if $F(x)$ is closed in $Y$ for every $x\in X$,
- $F$ has *closed graph* if $\mathrm{graph}\,F$ is closed in $X\times Y$,
- $F$ is *compact valued* (or that it has *compact images*) if $F(x)$ is compact for every $x\in X$,
- $F$ is *upper semicontinuous* at $x\in X$ if for every open set $V\subseteq Y$ such that $V\supseteq F(x)$ there exists an open neighborhood $U\subseteq X$ of $x$ such that $F(z)\subseteq V$ for all $z\in U$.
- $F$ is *lower semicontinuous* at $x\in X$ if for every open set $V\subseteq Y$ such that $V\cap F(x)\ne \emptyset$ there exists an open neighborhood $U\subseteq X$ of $x$ such that $F(z)\cap V\ne\emptyset$ for all $z\in U$.
- $F$ is continuous at $x\in X$ if it is both lower and upper semicontinuous at $x$.
- $F$ will be called continuous (resp. lower semicontinuous, upper semicontinuous) if it is continuous (resp. lower semicontinuous, upper semicontinuous) at every $x\in X$.
When $Y$ is a vector space, $F$ is convex valued (or it has *convex images*) if $F(x)$ is convex for every $x\in X$. When $X,Y$ are measurable spaces, we say that $F$ is measurable if $\mathrm{graph}\,F$ is measurable in $X\times Y$ endowed with the product of $\sigma$-algebrae on $X$ and $Y$. When $(X,d)$ is a metric space and $Y$ is a normed space, given $L>0$ we say that $F$ is Lipschitz continuous with constant $L$ if for all $x_1,x_2\in X$ $$F(x_2)\subseteq F(x_1)+L\cdot d(x_1,x_2)\overline{B_Y(0,1)},$$ where the sum and the product of sets are in the Minkowski sense: $A+B=\{a+b:\,a\in A,\,b\in B\}$ and $\lambda A=\{\lambda a:\,a\in A\}$ for every $A,B\subseteq Y$, $\lambda\in\mathbb R$.
Admissible trajectories {#sec:admtraj}
=======================
Given a collection $\boldsymbol\mu=\{\mu_h\}_{h\in I}\subseteq\mathscr P(X)$ of Borel measures on the measure space $X$ indexed by a parameter $h\in I$, by a slight abuse of notation we will denote by $\boldsymbol\mu$ both the set $\boldsymbol\mu=\{\mu_h\}_{h\in I}\subseteq\mathscr P(X)$ and the function $h\mapsto \mu_h$. In each occurrence, the context will clarify what we are referring to.
Let $I=[a,b]$ be a compact real interval, $\boldsymbol\mu=\{\mu_t\}_{t\in I}\subseteq \mathscr P_p(\mathbb R^d)$, $\boldsymbol\nu=\{\nu_t\}_{t\in I}\subseteq \mathscr M(\mathbb R^d;\mathbb R^d)$, $F:\mathscr P_p(\mathbb R^d)\times\mathbb R^d \rightrightarrows \mathbb R^d$ be a set-valued map.
We say that $\boldsymbol\mu$ is an *admissible trajectory driven by* $\boldsymbol\nu$ defined on $I$ with underlying dynamics $F$ if
- $|\nu_t|\ll \mu_t$ for a.e. $t\in I$;
- $v_t(x):=\dfrac{\nu_t}{\mu_t}(x)\in F(\mu_t,x)$ for a.e. $t\in I$ and $\mu_t$-a.e. $x\in \mathbb R^d$;
- $\partial_t\mu_t+\mathrm{div}\,\nu_t=0$ in the sense of distributions on $[0,T]\times\mathbb R^d$, equivalently $$\dfrac{d}{dt}\int_{\mathbb R^d}\varphi(x)\,d\mu_t(x)=\int_{\mathbb R^d}\langle\nabla\varphi(x),v_t(x)\rangle\,d\mu_t(x),$$ for a.e. $t\in [0,T]$ and all $\varphi\in C^1_c(\mathbb R^d)$.
Define $\mathcal A^p_I:\mathscr P_p(\mathbb R^d)\rightrightarrows C^0([0,T];\mathscr P_p(\mathbb R^d))$ by $$\mathcal A^p_I(\mu):=\left\{\boldsymbol\mu=\{\mu_t\}_{t\in I}:\, \begin{array}{l}\boldsymbol\mu\textrm{ is an admissible trajectory}\\ \textrm{with }\mu_a=\mu\end{array}\right\}.$$ When $I$ and $p$ are clear by the context, we will omit them.
The following gluing lemma will be also used.
\[lemma:glue\] Let $a_1\le b_1=a_2\le b_2$ be given. For $i=1,2$, assume that $\boldsymbol\mu^i=\{\mu^i_t\}_{t\in[a_i,b_i]}$ are narrowly continuous families of probability measures on $\mathbb R^d$, and $v^i:[a_i,b_i]\times\mathbb R^d\to\mathbb R^d$ are Borel maps such that $\mu^1_{|t=b_1}=\mu^2_{|t=a_2}$ and $$\begin{cases}
&\partial_t\mu^i_t+\mathrm{div}(v^i_t\mu^i_t)=0,\\ \\
&\displaystyle\int_{a_i}^{b_i}\int_{\mathbb R^d}|v^i_t(x)|\,d\mu^i_t(x)\,dt<+\infty,
\end{cases}
\qquad\qquad i=1,2\,.$$ Then if we set $$(\mu_t,v_t)=\begin{cases}(\mu^1_t,v^1_t),&\textrm{ for }a_1\le t\le b_1,\\ (\mu^2_{t},v^2_{t}),&\textrm{ for }a_2\le t\le b_2,\end{cases}$$ we have that $\boldsymbol\mu:=\{\mu_t\}_{t\in[a_1,b_2]}$ solves the continuity equation $\partial_t\mu_t+\mathrm{div}(v_t\mu_t)=0$.
See [@DNS Lemma 4.4].
Let $I_i=[a_i,b_i]$, $\boldsymbol\mu^{(i)}=\{\mu^{(i)}_t\}_{t\in I_i}\in\mathcal A^p_{I_i}(\mu^{(i)}_{a_i})$, $i=1,2$, be satisfying $\mu^{(1)}_{b_1}=\mu^{(2)}_{a_2}$. We define $I_3=[a_1,b_1+b_2-a_2]$ and $\boldsymbol\mu^{(3)}=\{\mu^{(3)}_t\}_{t\in I_3}$ by setting $\mu^{(3)}_t=\mu^{(1)}_t$ for $t\in I_1$ and $\mu^{(3)}_{t}=\mu^{(2)}_{t+a_2-b_1}$ for $t\in I_3\setminus I_1$. The curve $\boldsymbol\mu^{(3)}$ will be called the *concatenation* of $\boldsymbol\mu^{(1)}$ and $\boldsymbol\mu^{(2)}$ and will be denoted by $\boldsymbol\mu^{(3)}=\boldsymbol\mu^{(1)}\odot\boldsymbol\mu^{(2)}$. By Lemma \[lemma:glue\], we have $\boldsymbol\mu^{(3)}\in\mathcal A_{I_3}^p(\mu^{(1)}_{a_1})$.
Let $I=[a,b]$, $J=[a',b']$ with $J\subseteq I$, and $\boldsymbol\mu=\{\mu_t\}_{t\in I}\in\mathcal A^p_{I}(\mu_a)$. The *restriction* $\boldsymbol\mu_{|J}=\{\hat\mu_t\}_{t\in J}$ of $\boldsymbol\mu$ to $J$ is defined by taking $\hat\mu_t=\mu_t$ for all $t\in J$ and we have $\boldsymbol\mu_{|J}\in\mathcal A^p_{J}(\mu_{a'})$.
Let $\boldsymbol\mu^{(i)}\in\mathcal A^p_{I_i}(\mu^{(i)})$, $i=1,2$. We say that $\boldsymbol\mu^{(2)}$ is an *extension* of $\boldsymbol\mu^{(1)}$ if $I_2\supseteq I_1$ and $\boldsymbol\mu^{(2)}_{|I_1}=\boldsymbol\mu^{(1)}$.
Throughout the paper, we will assume the following
*Assumption $\boldsymbol{(F)}$*: The set-valued map $F:\mathscr P_p(\mathbb R^d)\times\mathbb R^d\to \mathbb R^d$ has nonempty, convex and compact images, moreover it is Lipschitz continuous with constant $L>0$ with respect to the metric $$d_{\mathscr P_p(\mathbb R^d)\times\mathbb R^d}((\mu_1,x_1),(\mu_2,x_2))=W_p(\mu_1,\mu_2)+|x_1-x_2|$$ on $\mathscr P_p(\mathbb R^d)\times\mathbb R^d$. We set $\displaystyle K_F:=\max_{v\in F(\delta_0,0)}\{|v|\}$.
\[def:bas\] Assume $\boldsymbol{(F)}$. Let $\boldsymbol\theta=\{\theta_t\}_{t\in [0,T]}$ be a $W_p$-continuous curve in $\mathscr P_p(\mathbb R^d)$, $\mu\in\mathscr P_p(\mathbb R^d)$. Denote by $\Upsilon_F(\mu,\boldsymbol\theta)$ the set of $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\subseteq\mathscr P_p(\mathbb R^d)$ satisfying the following property: there exists $\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_T)$ such that
- $\mu_t=e_t\sharp\boldsymbol\eta$ for all $t\in [0,T]$, $\mu_0=\mu$;
- for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ and a.e. $t\in[0,T]$ it holds $\gamma\in AC([0,T])$, $\gamma(0)=x$, $\dot\gamma(t)\in F(\theta_t,\gamma(t))$ .
We set $\displaystyle M_{\boldsymbol\theta}:=\sup_{\tau\in[0,T]}\mathrm{m}_p^{1/p}(\theta_\tau)$ and $$\Xi(\mu,\boldsymbol\theta):=\{\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_T):\, \{e_t\sharp\boldsymbol\eta\}_{t\in[0,T]}\in \Upsilon_F(\mu,\boldsymbol\theta)\}.$$ On the set $X:=\mathscr P_p(\mathbb R^d)\times C^0([0,T];\mathscr P_p(\mathbb R^d))$ we define the metric $$d_{X}\left(\left(\mu^{(1)},\boldsymbol{\theta^{(1)}}\right),\left(\mu^{(2)},\boldsymbol{\theta^{(2)}}\right)\right)
=W_p\left(\mu^{(1)},\mu^{(2)}\right)+\sup_{t\in[0,T]}W_p\left(\theta^{(1)}_t,\theta^{(2)}_t\right),$$ where $\boldsymbol{\theta^{(i)}}=\{\theta^{(i)}_t\}_{t\in[0,T]}$, $i=1,2$.
Finally, we define the set-valued map $S^{\boldsymbol\theta}:\mathbb R^d\rightrightarrows \Gamma_T$ by setting for all $y\in\mathbb R^d$ $$\begin{aligned}
S^{\boldsymbol\theta}(y):=&\{\xi\in AC([0,T]):\, \dot\xi\in F(\theta_t,\xi(t))\textrm{ for a.e. }t\in[0,T],\,\xi(0)=y\}.\end{aligned}$$ Since $t\mapsto F(\theta_t,x)$ is continuous for all $x\in\mathbb R^d$ and $x\mapsto F(\theta_t,x)$ is Lipschitz continuous by assumption $\boldsymbol{(F)}$, with constant $L$ for all $t\in[0,T]$, the set-valued map $S^{\boldsymbol\theta}(\cdot)$ is Lipschitz continuous by [@AuF Corollary 10.4.2].
\[lemma:moment\] Assume $\boldsymbol{(F)}$. Let $\boldsymbol\theta=\{\theta_t\}_{t\in [0,T]}$ be a $W_p$-continuous curve in $\mathscr P_p(\mathbb R^d)$, $\mu\in\mathscr P_p(\mathbb R^d)$, $\boldsymbol \eta\in \Xi(\mu,\boldsymbol\theta)$ and $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\in \Upsilon_F(\mu,\boldsymbol\theta)$ such that $\mu_t=e_t\sharp\boldsymbol\eta$ for all $t\in [0,T]$. Then for all $t,s\in[0,T]$ $$\begin{aligned}
\mathrm{m}_p^{1/p}(\mu_t)\le&e^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+K_FT+LTM_{\boldsymbol\theta}\right),\\
W_p(\mu_t,\mu_s)\le&\left(K_F+LM_{\boldsymbol\theta}+Le^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+K_FT+LTM_{\boldsymbol\theta}\right)\right)\cdot|t-s|,\\
\int_{\mathbb R^d\times\Gamma_T}\|\dot\gamma\|^p_{L^{\infty}([0,1])}&\,d\boldsymbol\eta(x,\gamma)\le \left[K_F+L\left(e^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+K_FT+LTM_{\boldsymbol\theta}\right)+M_{\boldsymbol\theta}\right)\right]^p.\end{aligned}$$
Set $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in\Upsilon_F(\mu,\boldsymbol\theta)$. For $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ and a.e. $t\in[0,T]$ we have $$\dot \gamma(t)\in F(\theta_t,\gamma(t))\subseteq F(\delta_0,0)+L(|\gamma(t)|+\mathrm{m}_p^{1/p}(\theta_t))\overline{B(0,1)}$$ Thus for all $s,t\in [0,T]$ and a.e. $\tau\in[0,T]$ $$\begin{aligned}
|\dot\gamma(\tau)|\le &K_F+L(|\gamma(\tau)|+\mathrm{m}_p^{1/p}(\theta_\tau))\le K_F+LM_{\boldsymbol\theta}+L|\gamma(\tau)|,\\
|\gamma(t)|-|\gamma(0)|\le&\int_0^t|\dot\gamma(\tau)|\,d\tau\le (K_F+LM_{\boldsymbol\theta})T+L\int_0^t|\gamma(\tau)|\,d\tau,\\
|\gamma(t)-\gamma(s)|\le&\left|\int_s^t|\dot\gamma(\tau)|\,d\tau\right|\le (K_F+LM_{\boldsymbol\theta})|t-s|+L\left|\int_s^t|\gamma(\tau)|\,d\tau\right|.\end{aligned}$$ By Grönwall lemma, this implies for all $0\le s\le t\in [0,T]$ $$\begin{aligned}
\nonumber
|\gamma(t)|\le&e^{Lt}\left(|\gamma(0)|+(K_F+LM_{\boldsymbol\theta})T\right),\\ \label{eq:Lip}
\|\dot\gamma\|_{L^{\infty}([0,1])}\le &K_F+L\left(M_{\boldsymbol\theta}+e^{LT}\left(|\gamma(0)|+(K_F+LM_{\boldsymbol\theta})T\right)\right),\\ \nonumber
|\gamma(t)-\gamma(s)|\le& \left(K_F+LM_{\boldsymbol\theta}+Le^{LT}\left(|\gamma(0)|+(K_F+LM_{\boldsymbol\theta})T\right)\right)\cdot |t-s|,\end{aligned}$$ recalling that $\dot\gamma(s)\in F(\mu_s,\gamma(s))$ for a.e. $s$.
We conclude by taking the $L^p_{\boldsymbol\eta}$ norm of the above inequalities and using the triangular inequality.
\[prop:uscsol\] Set $X:=\mathscr P_p(\mathbb R^d)\times C^0([0,T];\mathscr P_p(\mathbb R^d))$.
The set-valued map $\Upsilon_F:X\rightrightarrows C^0([0,T];\mathscr P_p(\mathbb R^d))$ is upper semicontinuous with compact nonempty images.
We prove first that $\Upsilon_F(\mu,\boldsymbol\theta)\ne\emptyset$ for all $(\mu,\boldsymbol\theta)\in X$. Consider now the set-valued map $S^{\boldsymbol\theta}(\cdot)$ defined as in Definition \[def:bas\]. Since it is Lipschitz continuous, it has a Borel selection. Thus let $h_0:\mathbb R^d\to AC([0,T])$ be a Borel map such that $h_0(x)\in S^{\boldsymbol\theta}(x)$ for every $x\in\mathbb R^d$. Define $\boldsymbol\eta=\mu\otimes\delta_{h_0(x)}$, $\mu_t=e_t\sharp\boldsymbol\eta$, $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}$. Then, by construction, we have $\boldsymbol\mu\in\Upsilon_F(\mu,\boldsymbol\theta)$.
Let now $\left\{\left(\mu^{(n)},\boldsymbol\theta^{(n)}\right)\right\}_{n\in\mathbb N}\subseteq X$ be a sequence $d_X$-converging to $(\mu,\boldsymbol\theta)\in X$, and $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}\subseteq C^0([0,T];\mathscr P_p(\mathbb R^d))$, $\{\boldsymbol\eta^{(n)}\}_{n\in\mathbb N}\subseteq \mathscr P(\mathbb R^d\times\Gamma_T)$ be such that
- $\boldsymbol\theta^{(n)}=\{\theta^{(n)}_t\}_{t\in[0,T]}$, $\boldsymbol\theta=\{\theta_t\}_{t\in[0,T]}$;
- $\boldsymbol\mu^{(n)}\in \Upsilon_F(\mu^{(n)},\boldsymbol\theta^{(n)})$ and $\boldsymbol\eta^{(n)}\in\Xi(\mu^{(n)},\boldsymbol\theta^{(n)})$ for all $n\in\mathbb N$;
- $\boldsymbol\mu^{(n)}=\{\mu^{(n)}_t\}_{t\in[0,T]}$ with $\mu^{(n)}_t=e_t\sharp\boldsymbol\eta^{(n)}$ for all $t\in[0,T]$ and $n\in\mathbb N$.
We prove that the sequence $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ has always cluster points, and all the cluster points are contained in $\Upsilon_F(\mu,\boldsymbol{\theta})$. This will imply in particular that $\Upsilon_F(\cdot)$ has compact images (by taking constant sequences $(\mu^{(n)},\boldsymbol\theta^{(n)})\equiv(\mu,\boldsymbol\theta)$).
For $n$ sufficiently large, we have $\mathrm{m}_p^{1/p}(\mu^{(n)})\le \mathrm{m}_p^{1/p}(\mu)+1$ and $M_{\boldsymbol\theta^{(n)}}\le M_{\boldsymbol{\theta}}+1$, recalling the definition of the convergence in $X$. Thus, by applying the estimates of Lemma \[lemma:moment\], we have $$\begin{aligned}
&\mathrm{m}_p^{1/p}(\mu^{(n)}_t)\le e^{LT}\left(m_p^{1/p}(\mu)+1+K_FT+LTM_{\boldsymbol\theta}+LT\right),\\
&W_p(\mu^{(n)}_t,\mu^{(n)}_s)\le\\
&\le\left(K_F+LM_{\boldsymbol\theta}+L+Le^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+1+K_FT+LTM_{\boldsymbol\theta}+LT\right)\right)\cdot|t-s|,\\
&\int_{\mathbb R^d\times\Gamma_T}\|\dot\gamma\|^p_{L^{\infty}([0,1])}\,d\boldsymbol\eta^{(n)}(x,\gamma)\le\\
&\le\left[K_F+L\left(e^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+1+K_F T+LTM_{\boldsymbol\theta}+LT\right)+M_{\boldsymbol\theta}+1\right)\right]^p.\end{aligned}$$ In particular
- $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ is equicontinuous;
- for all $t\in [0,T]$, we have that $\{\mu^{(n)}_t\}_{n\in\mathbb N}$ is relatively compact in $\mathscr P_p(\mathbb R^d)$, since it has $p$-moment uniformly bounded.
Thus $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ is relatively compact in $C^0([0,T];\mathscr P_p(\mathbb R^d))$ by Ascoli-Arzelà theorem. Up to a passing to a subsequence, we may assume that there exists $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in C^0([0,T];\mathscr P_p(\mathbb R^d))$ such that $$\lim_{n\to +\infty}\sup_{t\in [0,T]}W_p(\mu_t,\mu^{(n)}_t)=0.$$ We notice also that the functional $\Psi:\mathbb R^d\times\Gamma_T\to\mathbb R\cup\{+\infty\}$ $$\Psi(x,\gamma):=\begin{cases}(|x|+|\gamma(0)|+\|\dot\gamma\|_{L^\infty})^p,\textrm{ if }\gamma\in\mathrm{Lip}([0,T];\mathbb R^d),\\ \\ +\infty,\textrm{ otherwise},\end{cases}$$ has compact sublevels in $\mathbb R^d\times\Gamma_T$ by Ascoli-Arzelà theorem. Since $$\begin{gathered}
\sup_{n\in\mathbb N}\int_{\mathbb R^d\times\Gamma_T}\Psi(x,\gamma)\,d\boldsymbol\eta^{(n)}(x,\gamma)\le\\
\le 2^{p-1}\sup_{n\in\mathbb N}\left[2\mathrm{m}_p(\mu^{(n)})+\int_{\mathbb R^d\times\Gamma_T}\|\dot\gamma\|^p_{L^{\infty}([0,1])}\,d\boldsymbol\eta^{(n)}(x,\gamma)\right]<+\infty,\end{gathered}$$ we have that $\{\boldsymbol\eta^{(n)}\}_{n\in\mathbb N}$ is tight in $\mathscr P(\mathbb R^d\times\Gamma_T)$. Thus, up to passing to a subsequence, we may assume also that there exists $\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_T)$ such that $\boldsymbol\eta^{(n)}\rightharpoonup \boldsymbol\eta$ narrowly. By the continuity of $e_t:\mathbb R^d\times\Gamma_T\to\mathbb R^d$, we have $\mu_t=e_t\sharp\boldsymbol\eta$. By [@AGS Proposition 5.1.8], for $\boldsymbol\eta$-a.e. $(x,\gamma)\in \mathbb R^d\times\Gamma_T$ there exists a sequence $\{(x_n,\gamma_n)\}_{n\in\mathbb N}$ such that $x_n=\gamma_n(0)$, $\gamma_n\in AC([0,T])$, $\dot\gamma_n(t)\in F(\theta_t^{(n)},\gamma_n(t))$ for a.e. $t\in [0,T]$ and for all $n\in\mathbb N$ with $x_n\to x$, $\|\gamma_n-\gamma\|_{\infty}\to 0$ as $n\to+\infty$.
By and recalling that $M_{\boldsymbol\theta^{(n)}}\le M_{\boldsymbol\theta}+1$ and $|x_n|\le |x|+1$ for $n$ sufficiently large, we have $n\in\mathbb N$, $$\|\dot\gamma_n\|_{L^{\infty}([0,1])}\le K_F+L\left(M_{\boldsymbol\theta}+1+e^{LT}\left(|x|+1+(K_F+LM_{\boldsymbol\theta}+L)T\right)\right).$$ In particular, by Ascoli-Arzelà Theorem, we have that $\gamma$ is Lipschitz continuous. For a.e. $t,\tau\in[0,T]$ we have also $$\begin{aligned}
F&(\theta^{(n)}_\tau,\gamma_n(\tau))\subseteq F(\theta_t,\gamma(t))+\\
&+L\left(W_p(\theta^{(n)}_\tau,\theta_\tau)+W_p(\theta_{\tau},\theta_t)+|\gamma_n(\tau)-\gamma(\tau)|+|\gamma(t)-\gamma(\tau)|\right)\overline{B(0,1)}\\
\subseteq&F(\theta_t,\gamma(t))+L\left(\sup_{\tau\in[0,T]}W_p(\theta^{(n)}_\tau,\theta_\tau)+W_p(\theta_\tau,\theta_t)+\|\gamma_n-\gamma\|_{\infty}+\mathrm{Lip}(\gamma)\cdot |t-\tau|\right)\overline{B(0,1)}\end{aligned}$$ For every $\varepsilon>0$ there is $n_{\varepsilon}\in\mathbb N$ such that if $n>n_\varepsilon$ we have for a.e. $t,\tau\in[0,T]$ $$\dot\gamma_n(\tau)\in F(\theta_t,\gamma(t))+L(\varepsilon+W_p(\theta_\tau,\theta_t)+\mathrm{Lip}(\gamma)|t-\tau|).$$ In particular, let $t\in [0,T]$ be a differentiability point of $\gamma_n$. We have for all $z\in\mathbb R^d$, $s\in[0,T]$, $s\ne t$, and $n>n_\varepsilon$ $$\begin{aligned}
\langle \dfrac{\gamma_n(s)-\gamma_n(t)}{s-t},z\rangle=&\dfrac{1}{s-t}\int_t^s \langle z,\dot\gamma_n(\tau)\rangle\,d\tau\\
\le&\sup_{v\in F(\theta_t,\gamma(t))}\langle z,v\rangle+L\varepsilon|z|+L|z|\mathrm{Lip}(\gamma)\dfrac{1}{s-t}\int_t^s|t-\tau|\,d\tau+L|z|\frac{1}{s-t}\int_s^tW_p(\theta_\tau,\theta_t)\,d\tau\end{aligned}$$ By letting $n\to+\infty$ and $s\to t$ we conclude that $\dot\gamma(t)\in F(\theta_t,\gamma(t))$ since $F(\theta_t,\gamma(t))$ is closed and convex. Hence $\boldsymbol\mu\in \Upsilon_F(\mu,\boldsymbol\theta)$, which completes the proof.
\[prop:SP\] Let $T>0$. Assume $\boldsymbol{(F)}$. Then $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}$ is an admissible trajectory if and only if $\boldsymbol\mu\in \Upsilon_F(\mu_0,\boldsymbol\mu)$.
1. 2. *Sufficience*. Assume that $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in\Upsilon_F(\mu_0,\boldsymbol\mu)$. Then there exists $\boldsymbol\eta\in \Xi(\mu_0,\boldsymbol\mu)$ such that $\mu_t=e_t\sharp\boldsymbol\eta$. Set $$\mathscr N:=\left\{(t,x,\gamma)\in [0,T]\times\mathbb R^d\times\Gamma_T:\, \nexists\dot\gamma(t)\textrm{ or }\dot\gamma(t)\notin F(\mu_t,\gamma(t))\textrm{ or }\gamma(0)\ne x\right\}.$$ Since $\mathscr L^1\otimes\boldsymbol\eta\left(\mathscr N\right)=0$, for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ and a.e. $t\in [0,T]$ we have that $\dot\gamma(t)$ exists and belongs to $F(\mu_t,\gamma(t))$, and $\gamma(0)=x$. Given $\varphi\in C^{1}_c(\mathbb R^d)$, we have $$\left|\int_{\mathbb R^d}\varphi(x)\,d\mu_t(x)-\int_{\mathbb R^d}\varphi(x)\,d\mu_s(x)\right|\le \|\nabla\varphi\|_{\infty}\int_{\mathbb R^d\times\Gamma_T}|\gamma(t)-\gamma(s)|\,d\boldsymbol\eta(x,\gamma).$$ According to , this implies that $$t\mapsto \int_{\mathbb R^d}\varphi(x)\,d\mu_t(x)$$ is Lipschitz continuous. Hence its distributional derivative is in $L^{\infty}$ and coincides with the pointwise derivative almost everywhere. Thus, in the sense of distributions in $]0,T[$, we obtain for all $\varphi\in C^1_c(\mathbb R^d)$ $$\begin{aligned}
\dfrac{d}{dt}\int_{\mathbb R^d}\varphi(x)\,d\mu_t(x)
=&\dfrac{d}{dt}\int_{\mathbb R^d\times\Gamma_T}\varphi(\gamma(t))\,d\boldsymbol\eta(x,\gamma)\\
=&\int_{\mathbb R^d\times\Gamma_T}\langle \nabla\varphi(\gamma(t)),\dot\gamma(t)\rangle\,d\boldsymbol\eta(x,\gamma)\\
=&\int_{\mathbb R^d}\langle\nabla\varphi(y), \int_{e^{-1}_t(y)}\dot\gamma(t)\,d\eta_{t,y}(x,\gamma)\rangle\,d\mu_t(y),\end{aligned}$$ where we disintegrated $\boldsymbol\eta$ w.r.t. $e_t$ obtaining $\boldsymbol\eta=\mu_t\otimes\eta_{t,y}$ and used the fact that $\|\nabla\varphi\|_{\infty}$ is bounded, and that the map $\gamma\mapsto\|\dot\gamma\|_{L^{\infty}}$ is in $L^1_{\boldsymbol\eta}$ due to the uniform bound on the moments. By Jensen’s inequality, we have $$d_{F(\mu_t,y)}\left(\int_{e^{-1}_t(y)}\dot\gamma(t)\,d\eta_{t,y}(x,\gamma)\right)\le \int_{e^{-1}_t(y)}d_{F(\mu_t,y)}\left(\dot\gamma(t)\right)\,d\eta_{t,y}(x,\gamma)=0,$$ and so for $\mu_t$-a.e. $y\in\mathbb R^d$ and a.e. $t\in [0,T]$ we have $$v_t(y):= \int_{e^{-1}_t(y)}\dot\gamma(t)\,d\eta_{t,y}(x,\gamma)\in F(\mu_t,y),$$ hence $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}$ is an admissible trajectory driven by $\boldsymbol\nu=\{\nu_t\}_{t\in [0,T]}$ with $\nu_t=v_t\mu_t$ for a.e. $t\in [0,T]$.
3. 4. *Necessity*. Assume that $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}$ is an admissible trajectory driven by $\boldsymbol\nu=\{\nu_t\}_{t\in[0,T]}$. Set $v_t(x)=\dfrac{\nu_t}{\mu_t}(x)\in F(\mu_t,x)$ for $\mu_t$-a.e. $x\in\mathbb R^d$ and a.e. $t\in [0,T]$. Filippov’s Theorem (see e.g. [@AuF Theorem 8.2.10]) implies that there exists a Borel selection $\xi(\cdot)$ of $F(\delta_0,0)$, such that $$|v_t(x)-\xi(x)|=d_{F(\delta_0,0)}(v_t(x))$$ for all $x\in\mathbb R^d$, and so we have $$\begin{aligned}
\left(\int_0^T\int_{\mathbb R^d}|v_t(x)|^p\,d\mu_t(x)\,dt\right)^{1/p}\le&\int_0^T\left(\int_{\mathbb R^d}|v_t(x)-\xi(x)|^p\,d\mu_t(x)\right)^{1/p}\,dt+\\
&+\int_0^T\left(\int_{\mathbb R^d}|\xi(x)|^p\,d\mu_t(x)\right)^{1/p}\,dt\\
=&\int_0^T\left(\int_{\mathbb R^d}d^p_{F(\delta_0,0)}(v_t(x))\,d\mu_t(x)\right)^{1/p}\,dt+TK_F\\
\le&2^{p-1}L\int_0^T\left[W_p(\delta_0,\mu_t)+\mathrm{m}_p^{1/p}(\mu_t)\right]\,dt+TK_F\\
\le&2^pL\int_0^T\mathrm{m}^{1/p}_p(\mu_t)\,dt+TK_F<+\infty.\end{aligned}$$ By [@AGS Theorem 8.2.1], there exists $\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_T)$ such that $\mu_t=e_t\sharp\boldsymbol\eta$ for all $t\in [0,T]$ and $\boldsymbol\eta$ is concentrated on $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ with $\gamma\in AC([0,T])$, $\dot\gamma(t)=v_t(\gamma(t))\in F(\mu_t,\gamma(t))$ for a.e. $t\in [0,T]$ and $\gamma(0)=x$. Thus $\boldsymbol\mu\in \Upsilon_F(\mu_0,\boldsymbol\mu)$.
\[cor:estmomad\] Assume $\boldsymbol{(F)}$. Let $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in \mathcal A^p_{[0,T]}(\mu_0)$. Then $$\begin{aligned}
W_p(\mu_t,\mu_0)&\le \Big[K_F+2\frac{Le^{LT}}{1-LTe^{LT}}\left(\mathrm{m}^{1/p}_p(\mu_0)+K_F T\right)\Big]\cdot t\\
&\le\Big[K_F+2L\left(\mathrm{m}^{1/p}_p(\mu_0)+K_F T\right)\Big]\cdot t.\end{aligned}$$
The proof comes from direct computations by implementing the estimates in Lemma \[lemma:moment\] and considering the result of Proposition \[prop:SP\].
\[cor:exist\] Let $T>0$. Assume $\boldsymbol{(F)}$. The set-valued map $\mathcal A:\mathscr P_p(\mathbb R^d)\rightrightarrows C^0([0,T];\mathscr P_p(\mathbb R^d))$ is upper semicontinuous with nonempty compact images.
Let $\mu\in\mathscr P_p(\mathbb R^d)$. Given $R>0$, define $$\begin{gathered}
\mathcal C(R):=\Big\{\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\subseteq\mathscr P_p(\mathbb R^d):\mu_0=\mu,\textrm{ and for all $t,s\in [0,T]$ }\mathrm{m}_p^{1/p}(\mu_t)\le R, \\
W_p(\mu_t,\mu_s)\le\left(K_F+LR+Le^{LT}\left(\mathrm{m}_p^{1/p}(\mu)+K_FT+LTR\right)\right)\cdot|t-s|\Big\}.\end{gathered}$$ Recalling that the concatenation of solutions of the continuity equation is again a solution of the continuity equation driven by the time concatenation of the vector fields (see [@DNS Lemma 4.4]), in order to prove that $\mathcal A(\mu)\ne \emptyset$ it is not restrictive to assume $LT<1/2$. In particular, we have $1-e^{LT}LT>0$. Define $$R:=\dfrac{e^{LT}(\mathrm{m}^{1/p}_p(\mu)+K_FT)}{1-e^{LT}LT}\ge \mathrm{m}^{1/p}_p(\mu).$$ Notice that $\mathcal C(R)\ne\emptyset$, since it contains the constant curve $\mu_t\equiv\mu$ for all $t\in[0,T]$, it is convex, and it is compact in $C^0([0,T];\mathscr P_p(\mathbb R^d))$ by Ascoli-Arzelà theorem. Moreover, $\Upsilon_F(\mu,\mathcal C(R))\subseteq \mathcal C(R)$ by Lemma \[lemma:moment\] and the choice of $R$. By Kakutani-Ky Fan Theorem (see e.g. [@KF Theorem 1]) we have that there exists $\boldsymbol\mu\in \mathcal C(R)$ such that $\boldsymbol\mu\in\Upsilon_F(\mu,\boldsymbol\mu)$, i.e., by Proposition \[prop:SP\], $\boldsymbol\mu$ is an admissible trajectory starting from $\mu$. All the other properties of $\mathcal A(\cdot)$ trivially follows from the fact that $\Upsilon_F(\cdot)$ is upper semicontinuous with nonempty compact images.
An alternative proof of existence of admissible trajectories, i.e. $\mathcal A(\mu)\neq\emptyset$, can be found for example in [@orrieri Theorem 6.1] where the author provides sufficient conditions in order to ensure existence (and uniqueness) of solutions of a continuity equation for some given non-local vector field.
\[thm:filippov\] Let $T>0$, $\mu^{(A)},\mu^{(B)}\in\mathscr P_p(\mathbb R^d)$ be given. Assume $\boldsymbol{(F)}$.
Let $\boldsymbol{\mu^{(A)}}=\{\mu^{(A)}_t\}_{t\in [0,T]}$ be an admissible trajectory satisfying $\mu^{(A)}_0=\mu^{(A)}$. Then there exists an admissible trajectory $\boldsymbol{\mu^{(B)}}=\{\mu^{(B)}_t\}_{t\in [0,T]}$ satisfying $\mu^{(B)}_0=\mu^{(B)}$ such that $$W_p(\mu^{(A)}_t,\mu^{(B)}_t)\le 2^{\frac{p-1}{p}}e^{L(2+Le^{LT})T}\cdot W_p(\mu^{(A)},\mu^{(B)})\quad \textrm{for all }t\in [0,T].$$ In particular, the set-valued map $\mathcal A:\mathscr P_p(\mathbb R^d)\rightrightarrows C^0([0,T];\mathscr P_p(\mathbb R^d))$ is Lipschitz continuous.
Let $\boldsymbol\pi\in\Pi^p_o(\mu^{(A)},\mu^{(B)})$ be an optimal transport plan between $\mu^{(A)}$ and $\mu^{(B)}$ for the $p$-Wasserstein distance. By disintegrating $\boldsymbol\pi$ w.r.t. $\mathrm{pr}_1:\mathbb R^d\times\mathbb R^d\to\mathbb R^d$, defined by $\mathrm{pr}_1(x,y)=x$, we have a Borel collection of measures $\{\pi_x\}_{x\in\mathbb R^d}\subseteq\mathscr P(\mathbb R^d\times\mathbb R^d)$, uniquely defined for $\mu^{(A)}$-a.e. $x\in\mathbb R^d$, such that $\boldsymbol\pi=\mu^{(A)}\otimes \pi_x$.
According to Proposition \[prop:SP\], there exists $\boldsymbol{\eta^{(A)}}\in\mathscr P(\mathbb R^d\times\Gamma_T)$ concentrated on pairs $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ with $\gamma\in AC([0,T])$, $\dot\gamma(t)\in F(e_t\sharp\boldsymbol{\eta^{(A)}},\gamma(t))$ for a.e. $t\in [0,T]$ and $\gamma(0)=x$ such that $\mu_t^{(A)}=e_t\sharp\boldsymbol{\eta^{(A)}}$.
Let $\boldsymbol\theta\in C^0([0,T];\mathscr P_p(\mathbb R^d))$, and define the set-valued map $S^{\boldsymbol\theta}(\cdot)$ as in Definition \[def:bas\]. Define the set-valued map $R^{\boldsymbol\theta}:\mathbb R^d\times\mathrm{supp}\,\boldsymbol{\eta^{(A)}}\rightrightarrows\Gamma_T$ by $$\begin{aligned}
R^{\boldsymbol\theta}(y,x,\gamma):=\left\{\xi\in S^{\boldsymbol\theta}(y):\begin{array}{l}|\gamma(t)-\xi(t)|\le e^{LT}|\gamma(0)-\xi(0)|+L(e^{LT}+1)\int_0^t W_p(\mu_\tau^{(A)},\theta_\tau)\,d\tau\\ \textrm{ for all }t\in[0,T]\end{array}\right\}.\end{aligned}$$ Notice that this map has closed domain, closed graph, and compact values since $R^{\boldsymbol\theta}(y,x,\gamma)\subseteq S^{\boldsymbol\theta}(y)$, thus it is upper semicontinuous, hence Borel measurable.
We prove that it has nonempty images. Given a point $(y,x,\gamma)\in \mathbb R^d\times\mathrm{supp}\,\boldsymbol{\eta^{(A)}}$, there are sequences $\{x_n\}_{n\in\mathbb N}$ converging to $x$ and $\{\gamma_n\}_{n\in\mathbb N}\subseteq AC([0,T])$ uniformly converging to $\gamma$ such that $x_n=\gamma_n(0)$ and $\dot \gamma_n(t)\in F(\mu_t^{(A)},\gamma_n(t))$ for a.e. $t\in[0,T]$. According to Filippov’s theorem (see [@AuF Theorem 10.4.1]), for every $n\in\mathbb N$ there exists $\xi_n\in S^{\boldsymbol\theta}(y)$ such that $$\begin{aligned}
|\gamma_n(t)-\xi_n(t)|\le&e^{LT}|\gamma_n(0)-\xi_n(0)|+(Le^{LT}+1)\int_0^t d_{F(\theta_t,\gamma_n(\tau))}(\dot\gamma_n(\tau))\,d\tau\\
\le&e^{LT}|\gamma_n(0)-\xi_n(0)|+L(Le^{LT}+1)\int_0^t W_p(\theta_t,\mu_t^{(A)})\,d\tau,\end{aligned}$$ recalling the Lipschitz continuity of $F(\cdot)$ and the choice of $\gamma_n$. By compactness of $S^{\boldsymbol\theta}(y)$, up to passing to a subsequence, we may assume that $\{\xi_n\}_{n\in\mathbb N}$ uniformly converges to $\xi\in S^{\boldsymbol\theta}(y)$ and, by construction, we have $\xi\in R^{\boldsymbol\theta}(y,x,\gamma)$, hence $R^{\boldsymbol\theta}(\cdot)$ is Borel measurable with closed domain and nonempty images, thus it admits a Borel selection $h_{\boldsymbol\theta}:\mathbb R^d\times\mathrm{supp}\,\boldsymbol{\eta^{(A)}}\to \Gamma_T$. We extend $h_{\boldsymbol\theta}(\cdot)$ to a Borel map defined on the whole of $\mathbb R^d\times\mathbb R^d\times\Gamma_T\to \Gamma_T$ by setting $h_{\boldsymbol\theta}(y,x,\gamma)=\gamma$ if $(x,\gamma)\notin\mathrm{supp}\,\boldsymbol{\eta^{(A)}}$.
Define $\boldsymbol{\eta^\theta}\in\mathscr P(\mathbb R^d\times\Gamma_T)$ by $$\int_{\mathbb R^d\times\Gamma_T}\varphi(y,\xi)\,d\boldsymbol{\eta^{\theta}}(y,\xi)=\int_{\mathbb R^d\times\Gamma_T}\left[\int_{\mathbb R^d\times\mathbb R^d}\varphi(x,h_{\boldsymbol\theta}(y,x,\gamma))\,d\pi_x(x,y)\right]\,d\boldsymbol{\eta^{(A)}}(x,\gamma),$$ and set $\boldsymbol{\mu^\theta}=\{\mu^\theta_t\}_{t\in [0,T]}$ where $\mu^\theta_t=e_t\sharp\boldsymbol{\eta^\theta}$ for all $t\in [0,T]$.
We have, by construction, $$\mathrm{supp}\,\boldsymbol{\eta^{\theta}}\subseteq \left\{(y,\xi)\in\mathbb R^d\times\Gamma_T:\,\xi\in S^{\boldsymbol\theta}(y)\right\}.$$ Notice that $$\begin{aligned}
\int_{\mathbb R^d}\varphi(x)\,d\mu^\theta_0(x)=&\int_{\mathbb R^d\times\Gamma_T}\int_{\mathbb R^d\times\mathbb R^d}\varphi\left(h_{\boldsymbol\theta}(y,x,\gamma)(0)\right)\,d\pi_x(x,y)\,d\boldsymbol{\eta^{(A)}}(x,\gamma)\\
=&\int_{\mathbb R^d\times\Gamma_T}\int_{\mathbb R^d\times\mathbb R^d}\varphi\left(y\right)\,d\pi_x(x,y)\,d\boldsymbol{\eta^{(A)}}(x,\gamma)\\
=&\int_{\mathbb R^d}\int_{\mathbb R^d\times\mathbb R^d}\varphi\left(y\right)\,d\pi_x(x,y)\,d\mu^{(A)}(x)\\
=&\int_{\mathbb R^d\times\mathbb R^d}\varphi\left(y\right)\,d\boldsymbol\pi(x,y)=\int_{\mathbb R^d\times\mathbb R^d}\varphi\left(y\right)\,d\mu^{(B)}(y).\end{aligned}$$ Thus $\boldsymbol{\mu^\theta}\in \Upsilon_F(\mu^{(B)},\boldsymbol\theta)$.
We have $$\begin{aligned}
W_p&(\mu_t^{(A)},\mu^{\theta}_t)\le\\
\le&\left(\int_{\mathbb R^d\times\Gamma_T}\int_{\mathbb R^d\times\mathbb R^d}|\gamma(t)-h_{\boldsymbol\theta}(y,x,\gamma)(t)|^p\,d\pi_x(x,y)\,d\boldsymbol{\eta^{(A)}}(x,\gamma)\right)^{1/p}\\
\le&\left(\int_{\mathbb R^d\times\Gamma_T}\int_{\mathbb R^d\times\mathbb R^d}\left[e^{LT}|x-y|+L(Le^{LT}+1)\int_0^t W_p(\mu_\tau^{(A)},\theta_\tau)\,d\tau\right]^p\,d\pi_x(x,y)\,d\boldsymbol{\eta^{(A)}}(x,\gamma)\right)^{1/p}\\
\le&2^{\frac{p-1}{p}}\left[e^{LT}W_p(\mu^{(A)},\mu^{(B)})+L(Le^{LT}+1)\int_0^t W_p(\mu_\tau^{(A)},\theta_\tau)\,d\tau\right].\end{aligned}$$ Thus, since $W_p(\mu_t^{(A)},\mu^{\theta}_t)\ge W_p(\delta_0,\mu^{\theta}_t)-W_p(\mu_t^{(A)},\delta_0)=\mathrm{m}_p^{1/p}(\mu^\theta_t)-\mathrm{m}_p^{1/p}(\mu_t^{(A)})$, we have $$\begin{aligned}
\mathrm{m}_p^{1/p}(\mu^\theta_t)\le&\mathrm{m}_p^{1/p}(\mu_t^{(A)})+2^{\frac{p-1}{p}}\left[e^{LT}W_p(\mu^{(A)},\mu^{(B)})+LD\int_0^t \mathrm{m}^{1/p}_p(\mu_\tau^{(A)})\,d\tau+LD\int_0^t\mathrm{m}_p^{1/p}(\theta_\tau)\,d\tau\right]\\
\le&(1+LTD)\sup_{t\in[0,T]}\mathrm{m}_p^{1/p}(\mu_t^{(A)})+2^{\frac{p-1}{p}}\left[e^{LT}W_p(\mu^{(A)},\mu^{(B)})+LD\int_0^t\mathrm{m}_p^{1/p}(\theta_\tau)\,d\tau\right],\end{aligned}$$ where we denoted with $D=Le^{LT}+1$. As in the proof of Corollary \[cor:exist\], without loss of generality we can assume that $0\le 2^{\frac{p-1}{p}}LDT<1$. The general case will follow by concatenating finitely many pieces of admissible curves defined on time-subintervals of sufficiently small length. We take $R>0$ sufficiently large such that $$R\ge \dfrac{\displaystyle(1+LTD)\sup_{t\in[0,T]}\mathrm{m}_p^{1/p}(\mu_t^{(A)})+2^{\frac{p-1}{p}}e^{LT}W_p(\mu^{(A)},\mu^{(B)})}{1-2^{\frac{p-1}{p}}LDT}\ge\mathrm{m}_p^{1/p}(\mu^{(B)}) ,$$ and such that $\mathrm{m}_p^{1/p}(\theta_t)\le R$ for all $t\in [0,T]$ and also $\mathrm{m}_p^{1/p}(\mu^\theta_t)\le R$ for all $t\in [0,T]$. Define a sequence $\{\boldsymbol\mu^{(n)}=\{\mu^{(n)}_t\}_{t\in [0,T]}\}_{n\in\mathbb N}\subseteq C^0([0,T];\mathscr P_p(\mathbb R^d))$ by setting $\boldsymbol\mu^{(0)}$ to be the constant $\mu^{(B)}$ and $\boldsymbol\mu^{(n)}$ to be equal to $\boldsymbol{\mu^{\theta}}$ with $\boldsymbol\theta=\boldsymbol\mu^{(n-1)}$. Notice that $\mu^{(n)}_0=\mu^{(B)}$ for all $n\in\mathbb N$. According to Lemma \[lemma:moment\], the family $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ is relatively compact, thus up to passing to a subsequence, we may assume that it converges to $\boldsymbol\mu^{\infty}\in C^0([0,T];\mathscr P_p(\mathbb R^d))$. Since $\boldsymbol\mu^{(n)}\in \Upsilon_F(\mu^{(B)},\boldsymbol\mu^{(n-1)})$, by recalling the u.s.c. of $\Upsilon_F(\cdot)$, we obtain that $\boldsymbol\mu^{\infty}\in \Upsilon_F(\mu^{(B)},\boldsymbol\mu^{\infty})$, i.e., $\boldsymbol{\mu^\infty}$ is an admissible trajectory, starting from $\mu^{(B)}$. Finally, by passing to the limit in $$\begin{aligned}
W_p(\mu_t^{(A)},\mu^{(n)}_t)\le& 2^{\frac{p-1}{p}}\left[e^{LT}W_p(\mu^{(A)},\mu^{(B)})+LD\int_0^t W_p(\mu_\tau^{(A)},\mu^{(n-1)}_\tau)\,d\tau\right],\end{aligned}$$ we have $$\begin{aligned}
W_p(\mu_t^{(A)},\mu^{\infty}_t)\le& 2^{\frac{p-1}{p}}\left[e^{LT}W_p(\mu^{(A)},\mu^{(B)})+LD\int_0^t W_p(\mu_\tau^{(A)},\mu^{\infty}_\tau)\,d\tau\right],\end{aligned}$$ and, by Grönwall’s Lemma, $$\begin{aligned}
W_p(\mu_t^{(A)},\mu^{\infty}_t)\le& 2^{\frac{p-1}{p}}e^{\hat D T} W_p(\mu^{(A)},\mu^{(B)}),\end{aligned}$$ as desired, where $\hat D=L(2+Le^{LT})$. The proof is concluded by setting $\boldsymbol{\mu^{(B)}}=\boldsymbol{\mu^{\infty}}$. The last assertion trivially follows.
\[lemma:invelset\] Assume $\boldsymbol{(F)}$. Let $\mu\in\mathscr P_p(\mathbb R^d)$.
1. Given any Borel selection $v_\mu:\mathbb R^d\to\mathbb R^d$ of $F(\mu,\cdot)$, there exists $\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_T)$ such that, set $\mu_t=e_t\sharp\boldsymbol\eta$ for $t\in [0,T]$, we have $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\in \mathcal A^p_{[0,T]}(\mu)$, $\boldsymbol\eta\in \Xi(\mu,\boldsymbol\mu)$ and $$\left|\dfrac{\gamma(t)-\gamma(0)}{t}-v_\mu(x)\right|\le Le^{Lt}\left[\dfrac{1}{t}\int_0^tW_p(\mu_\tau,\mu)\,d\tau+\frac{t}{2}|v_\mu(x)|\right]$$ for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$.
2. Given any admissible trajectory $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\in \mathcal A^p_{[0,T]}(\mu)$, there exists $\boldsymbol\eta\in \Xi(\mu,\boldsymbol\mu)$ such that $\mu_t=e_t\sharp\boldsymbol\eta$ and for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$ we have $$\lim_{t\to 0^+}\int_{\mathbb R^d\times\Gamma_T}d^p_{F(\mu,x)}\left(\dfrac{\gamma(t)-\gamma(0)}{t}\right)\,d\boldsymbol\eta(x,\gamma)=0.$$
$ $
**1.** Without loss of generality, we may assume $LT\le 1/2$, the general case will be obtained concatenating $\boldsymbol\mu$ with any other admissible trajectory starting from $\mu_T$. Let $v_0:\mathbb R^d\to\mathbb R^d$ be any Borel selection of $F(\mu,\cdot)$. Define $\gamma_x:[0,T]\to\mathbb R^d$ by $\gamma_x(t)=x+v_0(x)\cdot t$, and observe that $x\mapsto \gamma_x$ is a Borel map. Let $\boldsymbol\theta=\{\theta_t\}_{t\in [0,T]}\in C^0([0,T];\mathscr P_p(\mathbb R^d))$ such that $\theta_0=\mu$, and notice that $$d_{F(\theta_t,\gamma_x(t))}(\dot\gamma_x(t))\le L\left[W_p(\theta_t,\mu)+t|v_0(x)|\right].$$ Thus, by Filippov’s Theorem (see [@AuF Theorem 10.4.1]) the set-valued map $R^{\boldsymbol\theta}:\mathbb R^d\rightrightarrows \Gamma_T$ defined as $$R^{\boldsymbol\theta}(x):=\left\{\xi\in S^{\boldsymbol\theta}(x):\, \begin{array}{l}|\gamma_x(t)-\xi(t)|\le Le^{Lt}\int_0^t\left[W_p(\theta_\tau,\mu)+\tau|v_0(x)|\right]\,d\tau\\\textrm{for all }t\in[0,T]\end{array}\right\}$$ has nonempty images for every $x\in\mathbb R^d$. Notice that this set-valued map has closed images and it is Borel measurable by [@AuF Theorem 8.2.9], thus it admits a Borel selection $h_{\boldsymbol\theta}:\mathbb R^d\to \Gamma_T$. Set $\boldsymbol{\eta^\theta}=\mu\otimes\delta_{h_{\boldsymbol\theta}(x)}$ and $\boldsymbol{\mu^\theta}=\{\mu^\theta_t\}_{t\in [0,T]}$, $\mu^\theta_t=e_t\sharp\boldsymbol{\eta^\theta}$. By construction we have $\boldsymbol{\mu^\theta}\in \Upsilon_F(\mu,\boldsymbol\theta)$, moreover for all $x\in\mathbb R^d$ we have $$\begin{aligned}
\mathrm{m}_p^{1/p}(\mu^\theta_t)=&\left(\int_{\mathbb R^d}\left|h_{\boldsymbol\theta}(x)(t)\right|^p\,d\mu(x)\right)^{1/p}\\
\le&\left(\int_{\mathbb R^d}\left|h_{\boldsymbol\theta}(x)(t)-\gamma_x(t)\right|^p\,d\mu(x)\right)^{1/p}+\left(\int_{\mathbb R^d}\left|\gamma_x(t)\right|^p\,d\mu(x)\right)^{1/p}\\
\le&Le^{Lt}\left[\int_{\mathbb R^d}\left|\int_0^t\left(W_p(\theta_\tau,\mu)+\tau|v_0(x)|\right)\,d\tau\right|^p\,d\mu\right]^{1/p}+\mathrm{m}_p^{1/p}(\mu)+t\|v_0\|_{L^p_\mu}\\
\le&Le^{Lt}\int_0^tW_p(\theta_\tau,\mu)\,d\tau+\mathrm{m}_p^{1/p}(\mu)+(LTe^{Lt}+t)\|v_0\|_{L^p_\mu}\\
\le&Le^{Lt}\int_0^t\mathrm{m}_p^{1/p}(\theta_\tau)\,d\tau+(1+Le^{Lt}t)\mathrm{m}_p^{1/p}(\mu)+(LTe^{Lt}+t)\|v_0\|_{L^p_\mu}.\end{aligned}$$ Furthermore, $$\left|\dfrac{h_{\boldsymbol\theta}(x)(t)-h_{\boldsymbol\theta}(0)}{t}-v_0(x)\right|=\left|\dfrac{h_{\boldsymbol\theta}(x)(t)-\gamma_x(t)}{t}\right|\le Le^{Lt}\dfrac{1}{t}\int_0^t\left[W_p(\theta_\tau,\mu)+\tau|v_0(x)|\right]\,d\tau.$$ Choose $$R\ge\dfrac{(1+LTe^{LT})\mathrm{m}_p^{1/p}(\mu)+(LTe^{LT}+T)\|v_0\|_{L^p_\mu}}{1-LTe^{LT}}\ge \mathrm{m}_p^{1/p}(\mu),$$ and notice that if $\mathrm{m}_p^{1/p}(\theta_t)\le R$ for all $t\in [0,T]$, then $\mathrm{m}_p^{1/p}(\mu^\theta_t)\le R$ for all $t\in [0,T]$. Define sequences $\{\boldsymbol\mu^{(n)}=\{\mu^{(n)}_t\}_{t\in [0,T]}\}_{n\in\mathbb N}\subseteq C^0([0,T];\mathscr P_p(\mathbb R^d))$ and $\{\boldsymbol\eta^{(n)}\}_{n\in\mathbb N}\subseteq \mathscr P(\mathbb R^d\times\Gamma_T)$ by setting $\boldsymbol\mu^{(0)}$ to be the constant $\mu$, $\boldsymbol\mu^{(n)}$ and $\boldsymbol{\eta^{(n)}}$ to be equal to $\boldsymbol{\mu^{\theta}}$ and $\boldsymbol{\eta^{\theta}}$, respectively, with $\boldsymbol\theta=\boldsymbol\mu^{(n-1)}$ for all $n\in\mathbb N$. According to Lemma \[lemma:moment\], the families $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ and $\{\boldsymbol{\eta^{(n)}}\}_{n\in\mathbb N}$ are relatively compact, thus up to passing to a subsequence, we may assume that the sequences converge to $\boldsymbol\mu^{\infty}=\{\mu^\infty_t\}_{t\in[0,T]}\in C^0([0,T];\mathscr P_p(\mathbb R^d))$ and to $\boldsymbol{\eta^{\infty}}\in\mathscr P(\mathbb R^d\times\Gamma_T)$, with $\mu^{\infty}_t=e_t\sharp\boldsymbol{\eta^\infty}$ for all $t\in[0,T]$. Since $\boldsymbol\mu^{(n)}\in \Upsilon_F(\mu,\boldsymbol\mu^{(n-1)})$, by recalling the u.s.c. of $\Upsilon_F(\cdot)$, we obtain that $\boldsymbol\mu^{\infty}\in \Upsilon_F(\mu,\boldsymbol\mu^{\infty})$, i.e., $\boldsymbol{\mu^\infty}$ is an admissible trajectory, starting from $\mu$. Recall that for $\boldsymbol{\eta^{\infty}}$-a.e. $(x,\gamma)$ there exists a sequence $\{(x_n,\xi_n)\}_{n\in\mathbb N}\subseteq\mathbb R^d\times\Gamma_T$ converging to $(x,\gamma)$ such that $(x_n,\xi_n)\in\mathrm{supp}\,\boldsymbol{\eta^{(n)}}$. Thus, without loss of generality, we may assume for all $t\in[0,T]$ $$\left|\dfrac{\xi_n(t)-\xi_n(0)}{t}-v_0(x)\right|\le Le^{Lt}\dfrac{1}{t}\int_0^t\left[W_p(\mu^{(n-1)}_\tau,\mu)+\tau|v_0(x)|\right]\,d\tau,$$ and, by passing to the limit, $$\left|\dfrac{\gamma(t)-\gamma(0)}{t}-v_0(x)\right|\le Le^{Lt}\dfrac{1}{t}\int_0^t\left[W_p(\mu^\infty_\tau,\mu)+\tau|v_0(x)|\right]\,d\tau.$$ **2.** According to the Superposition Principle, there exists $\boldsymbol\eta\in\Xi(\mu,\boldsymbol\mu)$ such that $\mu_t=e_t\sharp\boldsymbol\eta$ for all $t\in [0,T]$. By Jensen’s inequality, for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_T$, we have $$\begin{aligned}
d_{F(\mu,x)}\left(\dfrac{\gamma(t)-\gamma(0)}{t}\right)=&d_{F(\mu,x)}\left(\dfrac{1}{t}\int_{0}^t \dot\gamma(s)\,ds\right)\le
\dfrac{1}{t}\int_{0}^t d_{F(\mu,x)}\left(\dot\gamma(s)\right)\,ds\\
\le&\dfrac{L}{t}\int_{0}^t \left(W_p(\mu_s,\mu)+|\gamma(s)-x|\right)\,ds\\
\le&\dfrac{L}{t}\int_{0}^t \left(W_p(\mu_s,\mu)+\|\dot\gamma\|_{L^{\infty}_{\boldsymbol\eta}}s\right)\,ds.\end{aligned}$$ We conclude by taking the $L^p_{\boldsymbol\eta}$-norm and using Lemma \[lemma:moment\].
Generalized targets {#sec:gentarget}
===================
In this section, we provide the generalized notion of target set in the space of probability measures, thus extending in a natural way the classical concept of target set in $\mathbb R^d$. A naive physical interpretation of the generalized target can be given as follows: to describe the state of the system, an observer chooses to measure some quantities $\phi$. The results of the measurements are the *averages* of the quantities $\phi$ with respect to the measure $\mu_t$, representing the state of the system at time $t$. Our aim is to steer the system to states where the result of such measurements is below a fixed threshold (that, without loss of generality, we assume to be $0$). The following result provides a characterization of the class of such generalized target.
Let $\tilde S\subseteq \mathscr P(\mathbb R^d)$ be nonempty. The following are equivalent
1. $\tilde S$ is $w^*$-closed and convex;
2. there exists a family $\Phi\subseteq C^0_b(\mathbb R^d)$ such that $$\tilde S:=\left\{\mu\in\mathscr P(\mathbb R^d):\,\int_{\mathbb R^d}\varphi(x)\,d\mu(x)\le 0\textrm{ for all }\varphi\in\Phi\right\}.$$
Recalling formula (5.1.7) in [@AGS Remark 5.1.2], we have that $\bar \mu\in\tilde S$ if and only if for all $\psi\in C^0_b(\mathbb R^d)$ it holds $$\int_{\mathbb R^d}\psi(x)\,d\bar\mu(x)\le \sup_{\mu\in\tilde S}\int_{\mathbb R^d}\psi(x)\,d\mu(x).$$ Given $\psi\in C^0_b(\mathbb R^d)$, set $$C_{\psi}:=\sup_{\mu\in\tilde S}\int_{\mathbb R^d}\psi(x)\,d\mu(x)\le +\infty.$$ Then we have that $\bar \mu\in\tilde S$ if and only if for all $\psi\in C^0_b(\mathbb R^d)$ such that $C_{\psi}<+\infty$ it holds $$\int_{\mathbb R^d}\left[\psi(x)-C_{\psi}\right]\,d\bar\mu(x)\le 0.$$ We set $$\Phi:=\left\{\varphi:=\psi-C_{\psi}:\, \psi\in C^0_b(\mathbb R^d)\textrm{ and }C_{\psi}<+\infty\right\},$$ to obtain the desired equivalence.
\[def:gentar\] Let $\tilde S\subseteq\mathscr P(\mathbb R^d)$ be nonempty $w^*$-closed and convex, $\Phi\subseteq C^0_b(\mathbb R^d)$. We say that $\tilde S$ is a *generalized target generated by* $\Phi$, and write $\tilde S=\tilde S^\Phi$ if $$\label{eq:gentar}\tilde S:=\left\{\mu\in\mathscr P(\mathbb R^d):\,\int_{\mathbb R^d}\varphi(x)\,d\mu(x)\le 0\textrm{ for all }\varphi\in\Phi\right\}.$$ Given $p\ge 1$ we set $\tilde S^\Phi_p=\tilde S^\Phi\cap \mathscr P_p(\mathbb R^d)$, and we define the *generalized distance* from $\tilde S_p^{\Phi}$ to be the $1$-Lipschitz continuous map given by $\displaystyle\tilde d_{\tilde S_p^\Phi}(\cdot):=\inf_{\mu\in \tilde S_p^\Phi}W_p(\cdot,\mu)$.
\[rem:smoothphi\]
- - In Definition \[def:gentar\] we can equivalently assume that $\Phi$ is a set of continuous bounded functions, or bounded Lipschitz functions, or even just l.s.c. functions bounded from below. Moreover, without loss of generality, we can always assume that $\Phi$ is convex. Indeed, assume that $\Psi$ is a set of l.s.c. functions bounded from below. For all $\psi\in \Psi$ and $k\in\mathbb N\setminus\{0\}$ we define a Lipschitz continuous bounded map $\varphi_{k}^{\psi}:\mathbb R^d\to\mathbb R$ by setting $$\varphi_{k}^{\psi}(x):=\min\left\{\inf_{y\in\mathbb R^d}\left\{\psi(y)+k|x-y|\right\},k\right\}.$$ We recall that $\{\varphi_k^{\psi}\}_{k\in\mathbb N}$ is an increasing sequence of bounded Lipschitz functions bounded from below and pointwise converging to $\psi$. Hence, by Monotone Convergence Theorem, we have $$\sup_{\psi\in \Psi}\int_{\mathbb R^d}\psi(x)\,d\mu(x)=\sup_{\psi\in \Psi}\int_{\mathbb R^d}\sup_{k\in\mathbb N}\varphi^\psi_k(x)\,d\mu(x)=
\sup_{k\in\mathbb N}\sup_{\psi\in \Psi}\int_{\mathbb R^d}\varphi^\psi_k(x)\,d\mu(x)=\sup_{\varphi\in \Phi}\int_{\mathbb R^d}\varphi(x)\,d\mu(x),$$ where $\Phi=\{\varphi^\psi_k:\,k\in\mathbb N\setminus\{0\},\,\psi\in \Psi\}$. Replacing $\Phi$ with its convex hull does not change anything due to the linearity of the integral operator.
- Since convergence in $W_p(\cdot,\cdot)$ implies $w^*$-convergence, if $\tilde S^{\Phi}$ is a generalized target, then $\tilde S^\Phi_p$ is closed and convex in $\mathscr P_p(\mathbb R^d)$ endowed with the $p$-Wasserstein metric $W_p(\cdot,\cdot)$.
- We notice that if there exists $\bar x\in \mathbb R^d$ such that $\varphi(\bar x)\le 0$ for all $\varphi\in \Phi$ then the set $\tilde S$ given by is nonempty, since $\delta_{\bar x}\in\tilde S$.
The last condition of Remark \[rem:smoothphi\] is indeed not necessary to have the nontriviality of $\tilde S$.
\[ex:nontrivial\] For every $y\in\mathbb R$, $\varepsilon>0$, define $$\varphi^\varepsilon_y(x)=\begin{cases}
-(x+y)^2+\varepsilon,\textrm{ if }|x+y|\le 1,\\ -1+\varepsilon,\textrm{ if }|x+y|\ge 1.
\end{cases}$$ and set $\Phi_{\varepsilon}:=\{\varphi_{y}^\varepsilon:\,y\in\mathbb R\}$. Clearly, we have that $\varphi^\varepsilon_y$ attains its maximum at $x=-y$ and the value of the maximum is $\varepsilon>0$. Thus the sufficient condition of the last assertion in Remark \[rem:smoothphi\] is violated. For $0<\varepsilon\le \dfrac{1}{12}$ sufficiently small we have $$\int_{-1/2}^{1/2}\varphi^\varepsilon_y(x)\,dx\le \int_{-1/2}^{1/2}\varphi^\varepsilon_0(x)\,dx=\varepsilon-\dfrac{1}{12}\le 0,$$ thus the measure $\chi_{[-1/2,1/2]}\mathscr L^1\in \tilde S$.
Indeed, by the translation invariance of the problem, we have that $\mu_a:=\chi_{[a,a+1]}\mathscr L^1\in \tilde S$ for all $a\in\mathbb R$, in particular, we have that $\tilde S$ is not tight, hence not $w^*$-compact, since for any $K\subseteq\mathbb R$ it is possible to find $a\in\mathbb R$ such that $\mu_a(\mathbb R\setminus K)=1$.
Let $\tilde S$ be a nonempty generalized target generated by the family $\Phi\subseteq C^0(\mathbb R^d)$. If there exists $\bar\phi\in\Phi$, $A,C>0$ and $p\ge 1$ such that $\bar \phi(x)\ge A|x|^p-C$, then $\tilde S^{\Phi}=\tilde S_p^{\Phi}$ is compact in the $w^*$-topology and in the $W_p$-topology.
Trivially we have that $\tilde S_p^{\Phi}\subseteq\tilde S^{\Phi}$ for any $p\ge 1$. Conversely, given $\mu\in \tilde S^{\Phi}$, we have $$A\cdot\mathrm{m}_p(\mu)-C\le\int_{\mathbb R^d}\bar\phi(x)\,d\mu\le 0,$$ hence $\mu\in\tilde S_p^{\Phi}$ and all the measures in $\tilde S_p^{\Phi}=\tilde S^{\Phi}$ have $p$-moments uniformly bounded by $C/A$. This means that the $w^*$-topology and $W_p$-topology coincide on $\tilde S^{\Phi}=\tilde S_p^{\Phi}$, which turns out to be tight, according to [@AGS Remark 5.1.5], and $w^*$-closed, hence $w^*$-compact and $W_p$-compact.
We mention the following example, which may be relevant for the applications.
Given a nonempty and closed set $S\subseteq\mathbb R^d$ and $\alpha\ge 0$, a natural choice for $\Phi$ can be for example $\Phi_\alpha=\{d_S(\cdot)-\alpha\}$. If $\alpha=0$ we have that $\tilde S^{\Phi_0}=\{\mu\in\mathscr P(\mathbb R^d):\, \mu(\mathbb R^d\setminus S)=0\}$. More generally, for all $r>0$ let $B_r(S):=\{z\in\mathbb R^d:\,d_S(z)\le r\}$. Then, if $\mu\in\tilde S^{\Phi_\alpha}$, $$r\mu(\mathbb R^d\setminus B_r(S))=\int_{\mathbb R^d\setminus B_r(S)}r\,d\mu\le \int_{\mathbb R^d\setminus B_r(S)}d_S(x)\,d\mu(x)\le \alpha,$$ thus, in particular, we must have $\mu(\mathbb R^d\setminus B_r(S))\le \min\left\{1,\dfrac{\alpha}{r}\right\}$ for all $r>0$, which, if $\alpha$ is sufficiently small can be interpreted as a relaxed version of the case $\alpha=0$.
Given a generalized target $\tilde S\subseteq\mathscr P(\mathbb R^d)$, a natural question is wheter it is possible to *localize* it, i.e., to describe it as the set of all the measures supported a certain (closed) subset of $\mathbb R^d$. Equivalently, we want to find a nonempty closed set $S\subseteq\mathbb R^d$, such that, set $\Phi=\{d_S(\cdot)\}$, we have $\tilde S=\tilde S^{\Phi}$. To this aim, we give the following definition.
\[def:cctar\] Let $\tilde S\subseteq\mathscr P(\mathbb R^d)$ be a generalized target. Given a set $S\subseteq\mathbb R^d$, we say that $S$ is a *classical counterpart of the generalized target* $\tilde S$ if $$\tilde S=\{\mu\in\mathscr P(\mathbb R^d):\,\mathrm{supp}\,\mu\subseteq S\}.$$ An analogous definition is given for the classical counterpart of $\tilde S\cap \mathscr P_p(\mathbb R^d)$, $p\ge 1$ by taking intersection of the right hand side with $\mathscr P_p(\mathbb R^d)$.
- - From the very definition of classical counterpart, if $\tilde S$ admits $S$ and $S'$ as classical counterparts, then $S=S'$.
- In general a classical counterpart may not exists: in $\mathbb R$, take $\Phi=\{\phi\}$ where $\phi:\mathbb R\to \mathbb R$, $\phi(y):=|y|-1$. Defined $\mu_0:=\dfrac12\left(\delta_{0}+\delta_{2}\right)$, we have $\mu_0\in \tilde S_p^{\Phi}$ for every $p\ge 1$. If a classical counterpart $S$ of $\tilde S^{\Phi}$ would exists, by definition it should contain the support of $\mu_0$, i.e. $0,2\in S$. However, $\delta_2\notin \tilde S^{\Phi}$ even if $\mathrm{supp}(\delta_2)\subseteq S$. So neither $\tilde S^{\Phi}$ nor $\tilde S_p^{\Phi}$ admit a classical counterpart.
- If $S$ is the classical counterpart of $\tilde S^{\Phi}$ (or $\tilde S_p^{\Phi}$), there exists a representation of $\tilde S^{\Phi}$ as $\tilde S^{\hat \Phi}$, where $\hat\Phi=\{\hat\phi\}$ and $\hat\phi(x)\ge 0$ for every $x\in\mathbb R^d$ where the inequality is strict at every $x\notin S$. In particular we can take $\hat\Phi=\{\arctan\circ d_S\}$ (resp. $\hat\Phi=\{d_S\}$), i.e., we can replace $\Phi$ with the set $\{\arctan\circ d_S\}$ (resp. $\{d_S\}$).
Our aim is now to characterize the generalized target possessing a classical counterpart.
\[prop:propcctar\] Let $\tilde S\subseteq\mathscr P(\mathbb R^d)$ be a generalized target, $S\subseteq\mathbb R^d$.
1. if $\tilde S$ admits $S$ as classical counterpart then $S$ is closed;
2. $\tilde S$ admits $S$ as classical counterpart if and only if $$\int_{\mathbb R^d}\left[\varphi(x)-\sup_{y\in S}\varphi(y)\right]\,d\mu(x)\le 0,$$ for all $\varphi\in C^0_b(\mathbb R^d)$ and $\mu\in \tilde S$;
3. if $\tilde S$ admits $S$ as classical counterpart, then $\tilde S_p$ admits $S$ as classical counterpart for all $p\ge 1$.
4. If $\tilde S=\tilde S^{\Phi}$ (resp. $\tilde S\cap \mathscr P_p(\mathbb R^d)=\tilde S_p^{\Phi}$), for a suitable $\Phi\subseteq C^0_b(\mathbb R^d)$, admits a classical counterpart $S$, then $$S=\bigcap_{\phi\in \Phi}\{x\in \mathbb R^d:\, \phi(x)\le 0\}.$$
<!-- -->
1. 2. Assume that $\tilde S$ admits $S$ as a classical counterpart and $\tilde S=\tilde S^\Phi$ for a suitable $\Phi\in C^0_b(\mathbb R^d)$. In particular, we have $\delta_x\in \tilde S$ for all $x\in S$, i.e. $\phi(x)\le 0$ for all $x\in S$. Let $\{x_n\}_{n\in\mathbb N}$ be a sequence in $S$ converging to $x\in\mathbb R^d$. Then for all $\varphi\in\Phi$ we have $\varphi(x_n)\le 0$ for all $n\in\mathbb N$, which implies $\varphi(x)\le 0$, and so $\delta_x\in \tilde S$. Since $S$ is a classical counterpart of $\tilde S$ and $\mathrm{supp}\delta_x=\{x\}$, we have that thus $x\in S$, so $S$ is closed.
3. $\tilde S$ admits $S$ as classical counterpart if and only if $\tilde S=\overline{\mathrm{co}}\{\delta_x:\,x\in S\}$, where the closure is the weak$^*$ closure in $\mathscr P(\mathbb R^d)$. Indeed, every measure supported in $S$ is $w^*$-limit of convex combinations of Dirac deltas concentrated in points of $S$, and conversely all such deltas belong to $\tilde S$ by definition of classical counterpart, and $\tilde S$ is convex and $w^*$-closed. Recalling formula (5.1.7) in [@AGS Remark 5.1.2], we have that $\mu\in\tilde S$ if and only if $$\int_{\mathbb R^d} \varphi(x)\,d\mu(x)\le \sup_{y\in S}\varphi(y),$$ for all $\varphi\in C^0_b$, as desired.
4. It is sufficient to use the same argument as in (2) but taking the intersection with $\mathscr P_p(\mathbb R^d)$ and the closure w.r.t. $W_p$ distance.
5. Trivially, if there exist $\bar x\in\mathbb R^d$ and $\varphi\in \Phi$ such that $\varphi(\bar x)>0$, then $\delta_{\bar x}\notin \tilde S$, thus $\bar x$ does not belong to the classical counterpart of $\tilde S$. Conversely, if $\varphi(\bar x)\le 0$ for all $\varphi\in \Phi$, then $\delta_{\bar x}\in \tilde S$, and so $\bar x\in S$ by definition of classical counterpart.
A useful sufficient condition can be expressed as follows.
\[cor:propcctar\] Assume that for every $\phi\in\Phi$ we have either $\phi(x)\geq 0$ or $\phi(x)\leq 0$ for all $x\in\mathbb R^d$. Then $\tilde S^{\Phi}$ (and so $\tilde S_p^{\Phi}$) admits classical counterpart.
Denote by $$S=\bigcap_{\phi\in \Phi}\{x\in \mathbb R^d:\, \phi(x)\le 0\}.$$ If for all $\phi\in\Phi$ and $x\in\mathbb R^d$ we had $\phi(x)\le 0$, then we would trivially have $S=\mathbb R^d$ and $\tilde S^\Phi=\mathscr P(\mathbb R^d)$ as desired since $\delta_x\in \tilde S^{\Phi}$ for all $x\in\mathbb R^d$, thus concluding with the thesis.
Otherwise, let $\mu\in\tilde S^{\Phi}$ and suppose by contradiction that $\mu(\mathbb R^d\setminus S)>0$. Thus there exists $y\in \mathbb R^d\setminus S$ of density $1$ w.r.t. $\mu$. In particular, there exists a neighborhood $A_y$ of $y$ contained in $\mathbb R^d\setminus S$ such that $\mu(A_y)>0$. If for all $\varphi\in \Phi$ we had $\varphi(y)\le 0$, we would have $y\in S$, contradicting the fact that $y\notin S$. So, according to the assumptions, there exists $\hat\phi\in\Phi$ such that $\hat\phi(x)\ge 0$ for all $x\in\mathbb R^d$ and such that $\hat \phi(y)>0$. Thus we have $$\sup_{\phi\in\Phi}\int_{\mathbb R^d} \phi(x)\,d\mu(x)\ge \int_{\mathbb R^d} \hat \phi(x)\,d\mu(x)\ge \int_{A_y} \hat\phi(x)\,d\mu(x)>0,$$ hence $\mu\notin \tilde S^\Phi$, leading to a contradiction. Thus $\tilde S^{\Phi}\subseteq\{\mu\in\mathscr P(\mathbb R^d):\, \mathrm{supp}\,\mu\subseteq S\}$. Since the converse inclusion is always true, equality holds.
The condition of Corollary \[cor:propcctar\] is not necessary in general. In $\mathbb R$, take $\Phi=\{\phi_1,\phi_2,\phi_3\}$ where $\phi_i:\mathbb R\to \mathbb R$, $i=1,2,3$ are defined to be $\phi_1(x)=\min\{\max\{x,0\},1\}$, $\phi_2(x)=\min\{\max\{-x,-1\},0\}$, $\phi_3(x)=\min\{\max\{x,-1\},1\}$. Then both $\tilde S_p^{\Phi}$ and $\tilde S^{\Phi}$ admits $S$ as their classical counterpart, with $S=]-\infty,0]$, but $\phi_3$ changes its sign.
We are now ready to state some comparison results between the generalized distance and the classical one.
\[prop:compclasdist\] Let $p\ge 1$, $\mu_0\in \mathscr P_p(\mathbb R^d)$, $\Phi\subseteq C^0_b(\mathbb R^d;\mathbb R)$ be such that $\tilde S_p^{\Phi}\ne\emptyset$, and define $$\label{eq:clco}S:=\bigcap_{\phi\in \Phi}\{x\in\mathbb R^d:\,\phi(x)\le 0\}.$$ Then $\tilde d_{\tilde S_p^\Phi}(\mu_0)\le \|d_S\|_{L^p_{\mu_0}}$, and equality holds if and only if the generalized target $\tilde S^\Phi_p$ admits classical counterpart. In this last case, the classical counterpart of $\tilde S^\Phi_p$ is $S$, moreover $\tilde d_{\tilde S_p^\Phi}^p:\mathscr P_p(\mathbb R^d)\to [0,+\infty[$ is convex.
If $S=\emptyset$ we have $d_S(x)\equiv +\infty$ at all $x\in\mathbb R^d$ so the statement is trivially true, thus suppose $S\ne\emptyset$. Since $S$ is closed and nonempty, [@AuF Corollary 8.2.13] implies the existence of a Borel map $g:\mathbb R^d\to S$ such that $|x-g(x)|=d_S(x)$. We have $$\mathrm{m}^{1/p}_p(g\sharp\mu_0)=\|g\|_{L^p_{\mu_0}}\le
\|\mathrm{Id}_{\mathbb R^d}-g\|_{L^p_{\mu_0}}+\|\mathrm{Id}_{\mathbb
R^d}\|_{L^p_{\mu_0}}\le \|d_S\|_{L^p}+\mathrm{m}_p^{1/p}(\mu_0)<+\infty,$$ moreover, for all $\phi\in\Phi$, we have $$\int_{\mathbb R^d}\phi(x)\,dg\sharp\mu_0(x)=\int_{\mathbb
R^d}\phi(g(y))\,d\mu_0(y)\le 0,$$ since $g(y)\in S$ for all $y\in\mathbb R^d$ and so $\phi\circ g(y)\le 0$ for all $y\in\mathbb R^d$. Therefore, $g\sharp\mu_0\in\tilde S^\Phi_p$, and so $$\tilde d^p_{\tilde S^\Phi}(\mu_0)\le W^p_p(\mu_0,g\sharp\mu_0)\le
\|\mathrm{Id}_{\mathbb R^d}-g\|^p_{L^p_{\mu_0}}=\|d_S\|^p_{L^p_{\mu_0}}.$$
Assume now that $\tilde S^\Phi_p$ admits classical counterpart. As noticed in Proposition \[prop:propcctar\], $S$ must be the classical counterpart of $\tilde S^\Phi_p$. For every $\nu_0\in \tilde S^{\Phi}_p$ we have thus $\mathrm{supp}\,\nu_0\subseteq S$ and hence $|x-y|\ge d_S(x)$ for all $\pi\in \Pi(\mu_0,\nu_0)$ and $\pi$-a.e. $(x,y)\in\mathbb R^d\times\mathbb R^d$. This leads to $$\iint_{\mathbb R^d\times\mathbb R^d}|x-y|^p\,d\pi(x,y)\ge \int_{\mathbb R^d}d_S^p(x)\,d\mu_0(x).$$ By taking the infimum on $\pi\in \Pi(\mu_0,\nu_0)$ and then on $\nu_0\in \tilde S^{\Phi}_p$, we obtain $\tilde d_{\tilde S_p^\Phi}(\mu_0)\ge \|d_S\|_{L^p_{\mu_0}}$, thus equality holds.
Without the assumption of existence of a classical counterpart for $\tilde S^{\Phi}_p$, the inequality $\tilde d_{\tilde S_p^\Phi}(\mu_0)\le \|d_S\|_{L^p_{\mu_0}}$ is strict. Indeed, since $\tilde S^{\Phi}_p$ does not admit $S$ as a classical counterpart, there exist a measure $\mu\in\tilde S_p^\Phi$ and $n\in\mathbb N$ such that $$\mu\left(\left\{z\in\mathbb R^d:\, d_S(z)>\dfrac{1}{n}\right\}\right)>0,$$ and so there exists a Borel set $A\subseteq\mathbb R^d$ and $\varepsilon>0$ such that $d^p_S(z)\ge \varepsilon$ for $\mu$-a.e. $z\in A$, $\mu(A)>0$. This implies $$0=\tilde d^p_{\tilde S^{\Phi}_p}(\mu)<\varepsilon\mu(A)\le \int_A\,d^p_S(z)\,d\mu(z)\le \int_{\mathbb R^d}d^p_S(x)\,d\mu(x).$$ Finally, the last statement is trivial, and it follows from the fact that $$\tilde d^p_{\tilde S_p^\Phi}(\mu)=\int_{\mathbb R^d}d_S^p(x)\,d\mu,$$ is linear in $\mu$.
Without the $p$-th power, the generalized distance in the case of the Proposition \[prop:compclasdist\] above may fail to be convex.
Let $p>1$. In $\mathbb R^2$, consider $P=(0,0)$, $Q_1=(1,0)$, $Q_2=\left(0,2^{1/p}\right)$. Set $S=\{P\}$, $\Phi=\{d_S(\cdot)\}$, hence $\tilde S_p^\Phi:=\left\{\delta_P\right\}$, and define $\nu_{\lambda}=\lambda\delta_{Q_1}+(1-\lambda)\delta_{Q_2}$, $\lambda\in[0,1]$. By Proposition \[prop:compclasdist\], we have $$\tilde d^p_{\tilde S_p^\Phi}(\nu_{\lambda})=W_p^p(\delta_P,\nu_{\lambda})=\lambda+2(1-\lambda)=2-\lambda,$$ whence $\tilde d_{\tilde S_p^\Phi}(\nu_{\lambda})=\sqrt[p]{2-\lambda}$, which is not convex.
In the metric space $\mathscr P_p(\mathbb R^d)$ endowed with the $W_p$-distance, another concept of convexity can be given, related more to the metric structure rather than to the linear one inherited by the set of all Borel signed measures.
Given any product space $X^N$ ($N\geq 1$), in the following we denote with $\mathrm{pr}^i\colon X^N\to X$ the projection on the $i$–th component, i.e., $\mathrm{pr}^i(x_1,\ldots, x_N)=x_i$.
Given a curve $\boldsymbol\mu=\{\mu_t\}_{t\in [0,1]}\subseteq\mathscr P_p(\mathbb R^d)$, we say that it is a *(constant speed) geodesic* if for all $0\leq s\leq t\leq 1$ we have $$W_p(\mu_s,\mu_t)=(t-s)W_p(\mu_0,\mu_1).$$ In this case, we will also say that the curve $\boldsymbol\mu$ is a *geodesic connecting $\mu_0$ and $\mu_1$*.
Let $\mu_0,\mu_1\in\mathscr P_p(\mathbb R^d)$ and let $\pi\in\Pi_o^p(\mu_0,\mu_1)$ be an optimal transport plan between $\mu_0$ and $\mu_1$, i.e. $$W_p^p(\mu_0,\mu_1)=\iint_{\mathbb R^d\times\mathbb R^d} |x_1-x_2|^p\,d\pi(x_1,x_2)\,.$$ Then the curve $\boldsymbol\mu=\{\mu_t\}_{t\in [0,1]}$ defined by $$\label{eq:charact_geodesic}
\mu_t:=\big((1-t)\,\mathrm{pr}^1+t\,\mathrm{pr}^2\big)\sharp \pi~\in~\mathscr P_p(\mathbb R^d),$$ is a (constant speed) geodesic connecting $\mu_0$ and $\mu_1$.
Conversely, any (constant speed) geodesic $\boldsymbol\mu=\{\mu_t\}_{t\in [0,1]}$ connecting $\mu_0$ and $\mu_1$ admits the representation for a suitable plan $\pi\in\Pi_o^p(\mu_0,\mu_1)$.
See [@AGS Theorem 7.2.2].
A subset $A\subseteq \mathscr P_p(\mathbb R^d)$ is said to be
1. *geodesically convex* if for every pair of measures $\mu_0,\mu_1$ in $A$, there exists a geodesic connecting $\mu_0$ and $\mu_1$ which is contained in $A$.
2. *strongly geodesically convex* if for every pair of measures $\mu_0,\mu_1$ in $A$ and for every admissible transport plan $\pi\in\Pi(\mu_0,\mu_1)$, the curve $t\mapsto\mu_t$ defined by is contained in $A$.
The interest in this alternative concept of convexity comes from the fact that, in many problems, functionals defined on probability measures are convex along geodesics (a notion related to geodesically convex sets) and not convex with respect to the linear structure in the usual sense. We refer to [@AGS Section 9.1] for further details.
Notice that, even if the notations do not highlight this fact, the notions of *geodesic* and *geodesical convexity* depend on the exponent $p$ which has been fixed.
\[Strong geodesic convexity of $\tilde S_p^\Phi$\] Let $p\ge 1$, $\Phi$ as in Definition \[def:gentar\] and assume that all the elements of $\Phi$ are also convex. Then the generalized target $\tilde S_p^{\Phi}$ is strongly geodesically convex.
Let $\mu_0,\mu_1\in\tilde S_p^\Phi$ and let $\pi\in\Pi(\mu_0,\mu_1)$ be an admissible transport plan between $\mu_0$ and $\mu_1$. Consider the corresponding curve $\boldsymbol\mu=\{\mu_t\}_{t\in [0,1]}$ defined by , and fix $t\in [0,1]$. We have for every $\phi(\cdot)\in\Phi$ $$\begin{aligned}
\int_{\mathbb R^d} & \phi(x)\,d\mu_t(x)\le\\
&\leq (1-t)\iint_{\mathbb R^d\times\mathbb R^d} \phi\left(\mathrm{pr}^1(\xi,\eta)\right)\,d\pi(\xi,\eta)
+t\iint_{\mathbb R^d\times\mathbb R^d} \phi\left(\mathrm{pr}^2(\xi,\eta)\right)\,d\pi(\xi,\eta)\\
&=(1-t)\int_{\mathbb R^d} \phi(x)\,d\mu_0(x)
+t\int_{\mathbb R^d} \phi(y)\,d\mu_1(y)\le 0\,,\end{aligned}$$ since $\mathrm{pr}^i\sharp\pi$ are the marginal measures of $\pi$, which belong to $\tilde S_p^\Phi$. The conclusion follows from the arbitrariness of $\phi(\cdot)\in \Phi$.
In particular, considering also the first item in Remark \[rem:smoothphi\], the above result holds for $\Phi:=\{d_S(\cdot)-\alpha\}$ when $S$ is nonempty, closed and convex, and $\alpha\in[0,1]$. In this case, since in the above proof we use only the convexity property of $d_S(\cdot)$, the statement holds also if we equip $\mathbb R^d$ with a different norm than the Euclidean one.
We conclude this section by investigating the semiconcavity properties of the generalized distance along geodesics. The case $p=2$ is particularly easy thanks to the geometric structure of the metric space $\mathscr P_2(\mathbb R^d)$.
\[prop:semiconc\_2\] Let $\tilde S_2^\Phi$ be a generalized target in $\mathscr P_2(\mathbb R^d)$. Then the square of the generalized distance satisfies the following *global semiconcavity inequality:* for every $\mu_0,\mu_1\in \mathscr P_2(\mathbb R^d)$ and every $t\in[0,1]$ $$\tilde d^{\,2}_{\tilde S_2^\Phi}(\mu_t)\ge (1-t)\,\tilde d^{\,2}_{\tilde S_2^\Phi}(\mu_0)+t\,\tilde d^{\,2}_{\tilde S_2^\Phi}(\mu_1)-t(1-t)\,W^2_2(\mu_0,\mu_1),$$ where $\boldsymbol \mu=\{\mu_t\}_{t\in[0,1]}$ is any constant speed geodesic for $W_2$ joining $\mu_0$ and $\mu_1$.
Owing to [@AGS Theorem 7.3.2], we have that for any measure $\sigma\in\mathscr P_2(\mathbb R^d)$ the function $\mu\mapsto W_2^2(\mu,\sigma)$ is semiconcave along geodesics, with semiconcavity constant independent by $\sigma$, i.e. it satisfies for every $t\in[0,1]$ $$W_2^2(\mu_t,\sigma)+t(1-t)\,W^2_2(\mu_0,\mu_1)\ge (1-t)\,W_2^2(\mu_0,\sigma)+t\,W_2^2(\mu_1,\sigma).$$ The conclusion follows by passing to the infimum on $\sigma\in \tilde S_2^\Phi$.
In the case $p\neq 2$ we need additional requirements on $\Phi$.
\[prop:semicon\] Let $p\ge 1$, and $\tilde S_p^\Phi$ be a generalized target. Assume that $\tilde S_p^\Phi$ admits a classical counterpart $S\subseteq\mathbb R^d$. Let $K\subseteq\mathbb R^d\setminus S$ be compact and convex. Then the $p$-th power of the generalized distance $\tilde d_{\tilde S_p^\Phi}(\cdot)$ satisfies the following *local semiconcavity inequality:* there exists a constant $C=C(p,K)>0$ such that for every $\mu_0,\mu_1\in\mathscr P_p(K)$ we have $$\label{eq:gendis_semiconc}
\tilde d^{\,p}_{\tilde S_p^\Phi}(\mu_t)\ge (1-t)\,\tilde d^{\,p}_{\tilde S_p^\Phi}(\mu_0)+t\,\tilde d^{\,p}_{\tilde S_p^\Phi}(\mu_1)-Ct(1-t)\,W^{\min\{p,2\}}_p(\mu_0,\mu_1),$$ where $\boldsymbol \mu=\{\mu_t\}_{t\in[0,1]}$ is any constant speed geodesic for $W_p$ joining $\mu_0$ and $\mu_1$.
In this proof to make clearer the notation we will omit the superscript $\Phi$, since $\Phi$ is fixed. Under the above assumptions, and recalling Proposition \[prop:compclasdist\], we have $\tilde d_{\tilde S_p}(\mu_0)=\|d_S\|_{L^p_{\mu_0}}$.
Given $x_0,x_1\in K$ and $t\in [0,1]$ we set $$x_t:=(1-t)x_0+tx_1,\qquad\qquad d_t:=(1-t)d_{S}(x_0)+t d_{S}(x_1).$$ According to [@CSb Proposition 2.2.2], there exists $c=c(K)>0$ such that $d_S$ satisfies the following inequality for all $x_0,x_1\in K$: $$d_S(x_t)\ge d_t-ct(1-t)|x_0-x_1|^2,$$ By using [@CSb Proposition 2.1.12 (i)], we obtain that $$\label{eq:semiconc_dp}
d_S^p(x_t)\ge (1-t)\,d^p_S(x_0)+t\, d^p_S(x_1) -C' t(1-t)|x_0-x_1|^{\min\{p,2\}},$$ with $C'=C'(p,K)$.
For any Borel sets $A,B\subseteq\mathbb R^d$ and $\pi\in \Pi(\mu_0,\mu_1)$, we now have $\mathrm{supp}(\pi)\subseteq K\times K$. Therefore, we choose a transport plan $\pi\in \Pi_o^p(\mu_0,\mu_1)$ realizing the $p$-Wasserstein distance between $\mu_0$ and $\mu_1$, so that the representation in formula holds, and we integrate the estimate to find that $$\begin{aligned}
\int_{\mathbb R^d}d_S^p(x)\,d\mu_t=\iint_{\mathbb R^d\times \mathbb R^d}d_S^p(x_t)\,d\pi
&\geq (1-t)\int_{\mathbb R^d}d^p_S(x)\,d\mu_0+t\int_{\mathbb R^d}d^p_S(x)\,d\mu_1\\
&\hspace{.7cm}-C'\,t\,(1-t)\,\iint_{\mathbb R^d\times \mathbb R^d}|x_0-x_1|^{\min\{p,2\}}\,d\pi,\end{aligned}$$ where $\boldsymbol \mu=\{\mu_t\}_{t\in[0,1]}\subseteq\mathscr P_p(\mathbb R^d)$ is the constant speed geodesic corresponding to $\pi$. But according to Proposition \[prop:compclasdist\], there holds $$\tilde d^{\,p}_{\tilde S_p}(\mu_t)=\int_{\mathbb R^d}d_S^p(x)\,d\mu_t(x),
\qquad\mbox{and}\qquad
\tilde d^{\,p}_{\tilde S_p}(\mu_i)=\int_{\mathbb R^d}d_S^p(x)\,d\mu_i(x),\quad i=0,1,$$ and applying Jensen’s inequality to the concave map $\xi\mapsto \xi^{\gamma/p}$ on $\mathbb R^+$, with $\gamma=\min\{p,2\}$, we obtain that $$\iint_{\mathbb R^d\times \mathbb R^d}|x_0-x_1|^{\min\{p,2\}}\,d\pi
\leq \begin{cases}
\displaystyle \iint_{\mathbb R^d\times\mathbb R^d}|x_0-x_1|^p\,d\pi,&\textrm{ for }1\le p<2,\\ &\\
\left(\displaystyle\iint_{\mathbb R^d\times\mathbb R^d}|x_0-x_1|^p\,d\pi\right)^{2/p},&\textrm{ for }p\ge 2.
\end{cases}$$ We thus conclude that $$\tilde d^{\,p}_{\tilde S_p}(\mu_t)\ge (1-t)\,\tilde d^{\,p}_{\tilde S_p}(\mu_0)+t\, \tilde d^{\,p}_{\tilde S_p}(\mu_1)- C'\,t\,(1-t)\, W_p^{\min\{p,2\}}(\mu_0,\mu_1),$$ and the proof is completed.
Notice that inequality implies that, for $p\ge 2$ and under the assumption of Proposition \[prop:semicon\], the functional $-\tilde d^{\,p}_{\tilde S_p}(\cdot)\colon\mathscr P_p(K)\to \,]-\infty,0]$ is $\lambda$-geodesically convex, in the sense of [@AGS Definition 9.1.1], with $\lambda=-2C$.
The generalized minimum time function {#sec:mintime}
=====================================
Given a generalized target $\tilde S_p=\tilde S_p^\Phi$, we define the generalized minimum time function $\tilde T_p:\mathscr P_p(\mathbb R^d)\to [0,+\infty]$ by $$\tilde T_p(\mu):=\inf\{T\ge 0:\, \textrm{ there exists }\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in\mathcal A^p_{[0,T]}(\mu)\textrm{ s.t. }\mu_T\in\tilde S_p\},$$ where we set $\inf\emptyset=+\infty$ by convention. We say that $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in\mathcal A^p_{[0,T]}(\mu)$ is time optimal from $\mu$ if $\tilde T_p(\mu)\le T<+\infty$ and $\mu_{\tilde T_p(\mu)}\in\tilde S_p$.
Assume $\boldsymbol{(F)}$. Then
1. for any $\mu\in\mathscr P_p(\mathbb R^d)$ with $\tilde T_p(\mu)<+\infty$ there exists a time optimal admissible trajectory from $\mu$;
2. the function $\tilde T_p(\cdot)$ is lower semicontinuous;
3. the following Dynamic Programming Principle holds $$\label{eq:DPP}\tilde T_p(\mu)=\inf\{t+\tilde T_p(\mu_t):\, \boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in \mathcal A^p_{[0,T]}(\mu),\, T>0\}.$$ In particular, $t\mapsto t+\tilde T_p(\mu_t)$ is nondecreasing along every admissible trajectory $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in \mathcal A^p_{[0,T]}(\mu)$, and it is constant if and only if $\boldsymbol\mu=\{\mu_t\}_{t\in[0,T]}\in \mathcal A^p_{[0,T]}(\mu)$ is the restriction to $[0,T]\cap [0,\tilde T_p(\mu)]$ of an optimal trajectory.
<!-- -->
1. 2. Fix $\mu\in\mathscr P_p(\mathbb R^d)$ with $\tilde T_p(\mu)<+\infty$. For any $n\in\mathbb N\setminus\{0\}$ there exists $T_n>0$ and $\boldsymbol{\mu^{(n)}}=\{\mu^{(n)}_t\}_{t\in[0,T_n]}\in \mathcal A^p_{[0,T_n]}(\mu)$ such that $\mu^{(n)}_{T_n}\in \tilde S_p$ and $T_n\le \tilde T_p(\mu)+1/n$. We can extend each $\boldsymbol{\mu^{(n)}}$ to an admissible curve defined on $\tilde T_p(\mu)+1$ (possibly concatenating it with an element of $\mathcal A_{[T_n,\tilde T_p(\mu)+1]}(\mu^{(n)}_{T_n})$, which is nonempty for all $n\in\mathbb N\setminus\{0\}$). Thus, without loss of generality, we may assume that we have a sequence $\boldsymbol{\hat\mu^{(n)}}=\{\hat\mu^{(n)}_t\}_{t\in[0,T]}\in\mathcal A^p_{[0,T]}(\mu)$ of admissible trajectories, which are all defined in $[0,T]$ with $T=\tilde T_p(\mu)+1$, and satisfying $\hat\mu^{(n)}_{T_n}\in \tilde S_p$ where $T_n\le \tilde T_p(\mu)+1/n$. Recalling the compactness of $\mathcal A^p_{[0,T]}(\mu)$ (see Corollary \[cor:exist\]), up to passing to a subsequence, the sequence of curves $\{\boldsymbol{\hat\mu^{(n)}}\}_{n\in\mathbb N}$ uniformly converges to $\boldsymbol{\hat\mu^{\infty}}=\{\hat\mu^{\infty}_t\}_{t\in [0,T]}$ and, moreover, we have $T_n\to \ell$. In particular, $\hat\mu^{(n)}_{T_n}\to \hat\mu^{\infty}_\ell$ which, by the closedness of $\tilde S_p$, implies $\hat\mu^{\infty}_\ell\in \tilde S_p$, and so $\tilde T_p(\mu)\le \ell$. But passing to the limit in $T_n\le \tilde T_p(\mu)+1/n$ yields the reverse inequality, thus $\ell=\tilde T_p(\mu)$, hence $\boldsymbol{\hat\mu^{\infty}}$ is optimal.
3. Let $\{\mu^{(n)}\}_{n\in\mathbb N}\subseteq \mathscr P_p(\mathbb R^d)$ be a $W_p$-converging sequence satisfying $\mu^{(n)}\to \mu^{\infty}$ and\
$\displaystyle\liminf_{n\to+\infty}\tilde T_p(\mu^{(n)})=: \ell\in\mathbb R$. If $\ell=+\infty$ there is nothing to prove, so let us assume $\ell<+\infty$. As before, up to concatenation and restriction and by taking $n$ sufficiently large, this implies that there exists a sequence $\{\boldsymbol{\mu^{(n)}}\}_{n\in\mathbb N}$ such that $\boldsymbol{\mu^{(n)}}=\{\mu^{(n)}_t\}_{t\in[0,\ell+1]}\in\mathcal A_{[0,\ell+1]}(\mu^{(n)})$ and $\mu^{(n)}_{\tilde T_p(\mu^{(n)})}\in\tilde S_p$ for all $n\in\mathbb N$.
By Theorem \[thm:filippov\], there exists a sequence $\{\boldsymbol{\hat\mu^{(n)}}=\{\hat\mu^{(n)}_t\}_{t\in[0,\ell+1]}\}_{n\in\mathbb N}\subseteq\mathcal A_{[0,\ell+1]}(\mu^{\infty})$ such that $$W_p(\hat\mu^{(n)}_t,\mu^{(n)}_t)\le D\cdot W_p(\mu^{(n)},\mu^{\infty}),$$ for all $t\in [0,\ell+1]$, where $D:=2^{\frac{p-1}{p}}e^{L(2+Le^{L(\ell+1)})(\ell+1)}$. Recalling the compactness of $\mathcal A^p_{[0,\ell+1]}(\mu^\infty)$ (see Corollary \[cor:exist\]), up to a passing to a subsequence, the sequence of curves $\{\boldsymbol{\hat\mu^{(n)}}\}_{n\in\mathbb N}$ uniformly converges to $\boldsymbol{\hat\mu^{\infty}}=\{\hat\mu^{\infty}_t\}_{t\in [0,\ell+1]}$, in particular, we have that $$\begin{aligned}
W_p\left(\mu^{(n)}_{\tilde T_p(\mu^{(n)})},\hat\mu^{\infty}_{\ell}\right)
\le&W_p\left(\mu^{(n)}_{\tilde T_p(\mu^{(n)})},\hat\mu^{(n)}_{\tilde T_p(\mu^{(n)})}\right)+W_p\left(\hat\mu^{(n)}_{\tilde T_p(\mu^{(n)})},\hat\mu^{\infty}_{\tilde T_p(\mu^{(n)})}\right)+\\&+W_p\left(\hat\mu^{\infty}_{\tilde T_p(\mu^{(n)})},\hat\mu^{\infty}_{\ell}\right)\\
\le&D\cdot W_p\left(\mu^{(n)},\mu^{\infty}\right)+\sup_{t\in[0,\ell+1]}W_p\left(\hat\mu^{(n)}_{t},\hat\mu^{\infty}_{t}\right)+\\&+W_p\left(\hat\mu^{\infty}_{\tilde T_p(\mu^{(n)})},\hat\mu^{\infty}_{\ell}\right).\end{aligned}$$ By taking the limit for $n\to +\infty$, we have that $W_p\left(\mu^{(n)}_{\tilde T_p(\mu^{(n)})},\hat\mu^{\infty}_{\ell}\right)\to 0$, hence, by the closedness of $\tilde S_p$, we obtain $\hat\mu^{\infty}_\ell\in \tilde S_p$, and so $\tilde T_p(\mu^\infty)\le \ell$.
4. Let $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\in\mathcal A^p_{[0,T]}(\mu)$ and $\boldsymbol{\hat\mu^{(t)}}\in \mathcal A^p_{[0,\tilde T_p(\mu_t)]}(\mu_t)$ such that $\boldsymbol{\hat\mu^{(t)}}$ is optimal for $\mu_t$ (such an optimal trajectory exists by item (1)). For any $t\in[0,T]$, the concatenation $\boldsymbol\mu_{|[0,t]}\odot\boldsymbol{\hat\mu^{(t)}}\in \mathcal A^p_{[0,t+\tilde T_p(\mu_t)]}(\mu)$, and so $\tilde T_p(\mu)\le t+\tilde T_p(\mu_t)$ for every $t\in [0,T]$, $\boldsymbol\mu=\{\mu_t\}_{t\in [0,T]}\in\mathcal A^p_{[0,T]}(\mu)$, $t>0$, giving the first inequality in . In particular, for $0\le t\le s\le T$, we have $$\tilde T_p(\mu)\le t+\tilde T_p(\mu_t)\le t+(s-t)+\tilde T_p(\mu_s)=s+\tilde T_p(\mu_s),$$ since the restriction of $\boldsymbol\mu$ to $[t,T]$ is an admissible trajectory from $\mu_t$. Thus $t\mapsto t+\tilde T_p(\mu_t)$ is nondecreasing along all the admissible trajectories. If $\boldsymbol\mu$ is an optimal trajectory, by taking $s=\tilde T_p(\mu)$ we have $\tilde T_p(\mu_s)=0$ and so $\tilde T_p(\mu)=t+\tilde T_p(\mu_t)$ for all $t\in [0,\tilde T_p(\mu)]$, which gives equality in . Finally, assume that $t\mapsto t+\tilde T_p(\mu_t)$ is constant along an admissible trajectory $\boldsymbol\mu\in \mathcal A^p_{[0,T]}(\mu)$. By we have that $\tilde T_p(\mu)=t+\tilde T_p(\mu_t)$ for all $t\in [0,T]$. If $T\ge \tilde T_p(\mu)$, this implies that $\boldsymbol\mu$ is optimal, since by taking $t=\tilde T_p(\mu)$ we obtain $\tilde T_p(\mu_{\tilde T_p(\mu)})=0$ and so $\mu_{\tilde T_p(\mu)}\in \tilde S_p$. If $T<\tilde T_p(\mu)$ we concatenate $\boldsymbol\mu$ with an optimal trajectory $\boldsymbol{\hat\mu}=\{\hat\mu_s\}_{s\in[0,\tilde T_p(\mu_T)]}\in\mathcal A_{[0,\tilde T_p(\mu_T)]}(\mu_T)$ for $\mu_T$. Set $\boldsymbol\mu\odot\boldsymbol{\hat\mu}=\{\tilde\mu_s\}_{s\in[0,T+\tilde T_p(\mu_T)]}$. In particular, we have $\tilde T_p(\mu_T)=s+\tilde T_p(\hat\mu_s)$ for all $s\in [0,\tilde T_p(\mu_T)]$, thus $\tilde T_p(\mu)=\tau+\tilde T_p(\tilde\mu_\tau)$ for all $\tau\in [0,\tilde T_p(\mu)]$. By taking $\tau=\tilde T_p(\mu)$ we obtain $\tilde T_p(\mu_{\tilde T_p(\mu)})=0$ and so $\mu_{\tilde T_p(\mu)}\in \tilde S_p$ thus the concatenation $\boldsymbol\mu\odot\boldsymbol{\hat\mu}$ is an optimal trajectory, whose restriction to $[0,T]$ is $\boldsymbol\mu$.
The following definition of Small-Time Local Attainability (STLA) has been introduced in [@KQ] for finite-dimensional control systems, but can be easily generalized in our framework.
\[def:STLA-W\] We say that the system with generalized target $\tilde S_p$ satisfies the STLA property if
*Property *****: for any $\varepsilon>0$ and $\hat\mu\in \tilde S_p$ there exists $\delta>0$ such that $\tilde T_p(\mu)\le \varepsilon$ for any $\mu\in\mathscr P_p(\mathbb R^d)$ satisfying $W_p(\mu,\hat\mu)\le \delta$.
The link between STLA and continuity of the generalized minimum time is provided by the following result.
Let $\tilde S_p$ be a generalized target. Assume $\boldsymbol{(F)}$, and that **** holds for the system. Then $\tilde T_p:\mathscr P_p(\mathbb R^d)\to [0,+\infty]$ is continuous at every point where it is finite.
Recalling the l.s.c. of $\tilde T_p(\cdot)$, given $\mu\in\mathscr P_p(\mathbb R^d)$ with $\tilde T_p(\mu)=+\infty$, we have $\displaystyle\lim_{n\to+\infty}\tilde T_p(\mu^{(n)})=+\infty$ for every sequence $\{\mu^{(n)}\}_{n\in\mathbb N}$ converging to $\mu$ in $W_p$.
Therefore, we assume $T:=\tilde T_p(\mu)<+\infty$. Since $\tilde T_p(\cdot)$ is l.s.c., it is enough to prove that for all $\{\bar\mu^{(n)}\}_n\subseteq\mathscr P_p(\mathbb R^d)$ such that $W_p(\bar\mu^{(n)},\mu)\to0$ as $n\to+\infty$, we have $$\limsup_{n\to+\infty}\tilde T_p(\bar\mu^{(n)})\le T.$$ Fix an optimal trajectory $\boldsymbol\mu^{\infty}:=\{\mu^\infty_t\}_{t\in[0,T]}$ starting from $\mu^{\infty}_{|t=0}=\mu$. Let $\{\mu^{(n)}\}_{n\in\mathbb N}$ be a sequence converging to $\mu$ in $W_p$ and such that $\displaystyle\lim_{n\to+\infty}\tilde T_p(\mu^{(n)})$ exists. By Theorem \[thm:filippov\], there exists a sequence of admissible trajectories $\{\boldsymbol\mu^{(n)}\}_{n\in\mathbb N}$ such that
- $\boldsymbol\mu^{(n)}=\{\mu^{(n)}_t\}_{t\in [0,T]}$, $\mu^{(n)}_0=\mu^{(n)}$ for all $n\in\mathbb N$ and
- $\tilde d_{\tilde S_p}(\mu^{(n)}_T)\le W_p(\mu^{(n)}_T,\mu^{\infty}_T)\le D\cdot W_p(\mu^{(n)},\mu)$, recalling that $\mu^{\infty}_T\in\tilde S_p$,
where $D:=2^{\frac{p-1}{p}}e^{L(2+Le^{LT})T}$. In particular, by ****, given $\varepsilon>0$ there exists $n_\varepsilon\in\mathbb N$ such that for all $n>n_\varepsilon$ we have $\tilde T_p(\mu^{(n)}_T)\le \varepsilon$. By Dynamic Programming principle, we have $$\tilde T_p(\mu^{(n)})\le T+\tilde T_p(\mu^{(n)}_T)\le T+\varepsilon,$$ By letting $n\to +\infty$ and $\varepsilon\to 0$, we have $$\lim_{n\to+\infty}\tilde T_p(\mu^{(n)})\le T.$$ We conclude by the arbitrariness of the sequence $\{\mu^{(n)}\}_{n\in\mathbb N}$.
Given $\Phi\subset C_b^0(\mathbb R^d)$, $\phi\in\Phi$ and $\mu\in \mathscr P(\mathbb R^d)$ we define $$L_\phi(\mu):=\displaystyle\int_{\mathbb R^d}\phi(x)\,d\mu(x),\hspace{2cm}\sigma_\Phi(\mu):=\sup_{\phi\in\Phi}L_\phi(\mu).$$
Our aim is to provide a sufficient condition for ****, following the line of [@LM] and [@MR] for finite-dimensional systems. We recall that the l.s.c. of $\tilde T_p(\cdot)$ was already showed in [@CMNP Theorem 4] in a simplified setting, while a stronger sufficient condition was provided in [@Cav Theorem 4.1] to prove the Lipschitz continuity regularity. The continuity of $\tilde T_p(\cdot)$ was a crucial assumption also in [@CMNP Theorem 8] to prove that it solves an Hamilton-Jacobi-Bellman equation in Wasserstein space.
\[def:Sattain\] We say that the generalized target $\tilde S^\Phi_p$ is *$(r,Q)$-attainable* if there exist continuous maps $$r:[0,+\infty[\to\left[0,\min\left\{1,\frac{1}{2L}\right\}\right],\quad Q:\left[0,\min\left\{1,\frac{1}{2L}\right\}\right]\times[0,+\infty[\to\mathbb R$$ such that
1. $r(q)=0$ if and only if $q=0$;
2. $Q(r(q),q)<0$ for all $q\in]0,+\infty[$;
3. the function $q\mapsto \dfrac{r(q)}{|Q(r(q),q)|}$ is decreasing and integrable on $[0,+\infty[$.
4. for any $\mu\in\mathscr P_p(\mathbb R^d)\setminus\tilde S^\Phi_p$ there exists $\boldsymbol\mu=\{\mu_t\}_{t\in[0,r(\sigma_\Phi(\mu))]}\in\mathcal A^p_{[0,r(\sigma_\Phi(\mu))]}(\mu)$ such that $$\inf_{t\in[0,r(\sigma_\Phi(\mu))]}\{\sigma_\Phi(\mu_t)-\sigma_\Phi(\mu)\}\le 2\, Q(r(\sigma_\Phi(\mu)),\sigma_\Phi(\mu)).$$
\[prop:Tattain\] Assume $\boldsymbol{(F)}$ and that the generalized target $\tilde S^{\Phi}_p$ is $(r,Q)$-attainable. Then $$\boldsymbol T(\mu):=\int_0^{\sigma_{\Phi}(\mu)}\dfrac{r(q)\,dq}{|Q(r(q),q)|}\ge\tilde T_p(\mu).$$
Define sequences $\{\mu^{(i)}\}_{i\in\mathbb N}\subseteq \mathscr P(\mathbb R^d)$, $\{\sigma_i\}_{i\in\mathbb N}$, $\{t_i\}_{i\in\mathbb N}\subseteq [0,1]$ as follows. Set $\mu^{(0)}=\mu$. Suppose to have defined $\mu^{(i)}$, then define $\sigma_i=\sigma_\Phi(\mu^{(i)})$. We notice that, by assumption, if $\mu^{(i)}\notin\tilde S^{\Phi}_p$ we have $\sigma_i>0$, and so $Q(r(\sigma_i),\sigma_i)<0$.
By property (4) in Definition \[def:Sattain\], if $\mu^{(i)}\notin\tilde S^\Phi_p$ there exists $\boldsymbol\mu^{(i)}=\{\mu^{(i)}_t\}_{t\in[0,r(\sigma_i)]}\in\mathcal A^p_{[0,r(\sigma_i)]}(\mu^{(i)})$ such that $$\inf_{t\in[0,r(\sigma_i)]}\{\sigma_\Phi(\mu^{(i)}_t)-\sigma_i\}\le 2\, Q(r(\sigma_i),\sigma_i).$$ Thus, for any $\varepsilon>0$ there exists $t_i^\varepsilon\in[0,r(\sigma_i)]$ such that $$\sigma_\Phi(\mu^{(i)}_{t_i^\varepsilon})-\sigma_i\le 2\, Q(r(\sigma_i),\sigma_i)+\varepsilon.$$ Notice that, if we choose $\varepsilon$ sufficiently small, in particular $0<\varepsilon<-2\,Q(r(\sigma_i),\sigma_i)$, then $t_i^\varepsilon\neq0$. We thus fix $\hat\varepsilon(i)=-Q(r(\sigma_i),\sigma_i)$ and set $t_i=t_i^{\hat\varepsilon(i)}>0$ and $\mu^{(i+1)}=\mu^{(i)}_{t_i}$.
While, if $\mu^{(i)}\in\tilde S^\Phi_p$, then we set $t_i=0$ and $\mu^{(i+1)}=\mu^{(i)}_{t_i}=\mu^{(i)}$.
Thus, together with property (1) in Definition \[def:Sattain\], this implies that $\mu^{(i)}\notin\tilde S^{\Phi}_p$ if and only if $\sigma_i,\,t_i>0$.
Notice that $\sigma_{i}\ge 0$ for all $i\in\mathbb N$, moreover, if $\sigma_i=0$ then $\sigma_{m}=t_m=0$ for all $m\ge i$.
For every $i\in\mathbb N$ such that $\sigma_i\ne 0$ we have $$\label{eq:attsigma}
\sigma_{i+1}-\sigma_i\le 2\,Q(r(\sigma_i),\sigma_i)+\hat\varepsilon(i)=Q(r(\sigma_i),\sigma_i)<0,$$ by property (2) in Definition \[def:Sattain\]. Thus the sequence $\{\sigma_i\}_{i\in\mathbb N}$ is decreasing and bounded from below, and so it has a limit $\sigma_{\infty}\ge 0$. If $\sigma_i=0$ for some $i\in\mathbb N$ then $\sigma_{\infty}=0$. If $\sigma_i\ne 0$ for all $i\in\mathbb N$, by passing to the limit in we get $Q(r(\sigma_{\infty}),\sigma_\infty)=0$, by continuity of $Q(\cdot,\cdot)$ and $r(\cdot)$, which implies $\sigma_{\infty}=0$.
We have $$\boldsymbol T(\mu)\ge \sum_{\substack{i\in\mathbb N\\ \sigma_i\ne 0}}\dfrac{r(\sigma_i)(\sigma_{i}-\sigma_{i+1})}{|Q(r(\sigma_i),\sigma_i)|}\ge \sum_{i\in\mathbb N}t_i.$$ To conclude the proof, we consider two cases
- assume that $\sigma_i\ne 0$ for all $i\in\mathbb N$. Then for any $i\in\mathbb N$ there exists an admissible trajectory $\boldsymbol\mu^{(\infty)}=\{\mu^{(\infty)}_t\}_{[0,\boldsymbol T]}$ starting from $\mu$ and coinciding with $\boldsymbol\mu^{(i)}$ on $[t_{i-1},t_i]$. In particular, $\mu^{(\infty)}_{\sum t_i}\in \tilde S^{\Phi}_p$ since $\sigma_{\infty}=0$, and so $\boldsymbol T(\mu)\ge \tilde T_p(\mu)$.
- let $\hat\imath$ the minimum of the set $\{i\in\mathbb N:\,t_i=0\}$. Then there exists an admissible trajectory $\boldsymbol{\hat\mu}^{(\hat\imath)}=\{{\hat\mu}^{(\hat\imath)}_t\}_{[0,\boldsymbol T]}$ starting from $\mu$ and coinciding with $\boldsymbol\mu^{(i)}$ on $[t_{i-1},t_i]$, for all $i\ge1$. In particular, ${\hat\mu}^{(\hat\imath)}_{\sum_{i=1}^{\hat\imath-1} t_i}\in \tilde S^{\Phi}_p$, and so $$\boldsymbol T(\mu)\ge \sum_{i=1}^{\infty} t_i=\sum_{i=1}^{\hat\imath-1} t_i\ge \tilde T_p(\mu).$$
Thus in both cases we have $\boldsymbol T(\mu)\ge \tilde T_p(\mu)$, which concludes the proof.
\[thm:suffSTLA\] Assume $\boldsymbol{(F)}$ and that the generalized target $\tilde S^{\Phi}_p$ is $(r,Q)$-attainable. Assume that there exists $C>0$ and an open set $U\subseteq\mathscr P_p(\mathbb R^d)$ such that $U\supseteq\tilde S^\Phi_p$ and $\sigma_{\Phi}(\mu)\le C$ for all $\mu\in U$. Then **** holds.
Fix $\varepsilon>0$. Since $\max\{\sigma_{\Phi}(\mu),0\}\le C$ in a neighborhood of $\tilde S^\Phi_p$, we have that the convex function $\mu\mapsto\max\{\sigma_{\Phi}(\mu),0\}$ is continuous in a neighborhood of $\tilde S^\Phi_p$ and vanishes exactly on $\tilde S^\Phi_p$. Thus for any $\varepsilon>0$ there exists $\rho,\delta>0$ such that if $d_{\tilde S^\Phi_p}(\mu)\le \delta$ we have $\sigma_{\Phi}(\mu)\le \rho$ and $$\varepsilon>\int_0^{\rho}\dfrac{r(q)\,dq}{|Q(r(q),q)|}\ge\tilde T^{\Phi}_p(\mu),$$ recalling that by the integrability assumption in item (3) in Definition \[def:Sattain\], the map $$\rho\mapsto\int_0^{\rho}\dfrac{r(q)\,dq}{|Q(r(q),q)|}$$ is continuous.
In conclusion, in order to check the $(r,Q)$-attainability of a set from the data of the problem, the following result may serve the purpose.
\[cor:pre-example\] Given $\alpha\ge 0$, an interval $I\subseteq\mathbb R$, $\gamma\in AC(I;\mathbb R^d)$ and $v:\mathbb R^d\to\mathbb R^d$, we define $$\Delta^v_{\alpha,\gamma}(t):=\left|\dfrac{\gamma(t)-\gamma(0)}{t^{1+\alpha}}-v(\gamma(0))\right|.$$ Let $\mathcal D\subseteq\mathscr P_2(\mathbb R^d)$ and assume that there exist constants $C_\phi\ge 0$, $\alpha,\beta,K>0$ such that, by defining for any $\mu\in\mathcal D$ $$t_\mu:=\min\left\{1,\dfrac{1}{2L},\sigma_\Phi^{1/\beta}(\mu)\right\} \textrm{ and } I_\mu:=[0,t_\mu],$$ we have
1. $\Phi:=\{\phi\}$, where $\phi$ is semiconcave with constant $C_\phi$;
2. for all $\mu\in\mathcal D\setminus \tilde S^\Phi_2$ there exist functions $v_\mu,\xi_\mu\in L^2_\mu(\mathbb R^d;\mathbb R^d)$, $\boldsymbol\eta\in\mathscr P(\mathbb R^d\times\Gamma_{I_\mu})$, and constants $C_{2,\mu},C_{3,\mu},C_{4,\mu}>0$ satisfying
- $0\le \alpha<\beta-1$;
- $\displaystyle\boldsymbol\mu=\{e_t\sharp\boldsymbol\eta\}_{t\in I_\mu}\in\mathcal A_{I_\mu}(\mu)$, with $e_{t_\mu}\sharp\boldsymbol\eta\in\mathcal D$;
- $\xi_\mu(x)\in \partial^P\phi(x)$ for $\mu$-a.e. $x\in\mathbb R^d$;
- $\displaystyle\int_{\mathbb R^d}\langle \xi_\mu(x),v_\mu(x)\rangle\,d\mu(x)\le -C_{2,\mu}<0$;
- $\left(\displaystyle\int_{\mathbb R^d\times\Gamma_{I_\mu}} |\Delta^{v_\mu}_{\alpha,\gamma}(t_\mu)|^2\,d\boldsymbol\eta(x,\gamma)\right)^{1/2}\le C_{3,\mu}t_\mu$;
- $\|v_\mu\|_{L^2_\mu}\le C_{4,\mu}$;
- $\left(-C_{2,\mu}+C_{3,\mu}\|\xi_\mu\|_{L^2_\mu}t_\mu+2 C_\phi(C_{3,\mu}^2 t_\mu^2+C_{4,\mu}^2)t_\mu^{\alpha+1}\right)\le -2K\cdot t_\mu$.
Then **** holds in $\mathcal D$ and for all $\mu\in\mathcal D$ we have $$\tilde T^\Phi_2(\mu)\le \begin{cases}
\dfrac{\beta\sigma^{\frac{\beta-\alpha-1}{\beta}}_\Phi(\mu)}{K(\beta-\alpha-1)},&\textrm{ if }\sigma_\Phi(\mu)\le \min\{1,(2L)^{-\beta}\}\\ \\
\dfrac{\beta\,(2L)^{-\beta+\alpha+1}}{K(\beta-\alpha-1)}+\dfrac{1}{K}(2L)^{\alpha+1}\,(\sigma_\Phi(\mu)-(2L)^{-\beta}),&\textrm{ if }\sigma_\Phi(\mu)\ge (2L)^{-\beta}=\min\{1,(2L)^{-\beta}\}\\ \\
\dfrac{\beta}{K(\beta-\alpha-1)}+\dfrac{1}{K}(\sigma_\Phi(\mu)-1),&\textrm{ if }\sigma_\Phi(\mu)\ge 1=\min\{1,(2L)^{-\beta}\}.\end{cases}$$
Indeed, for $\boldsymbol\eta$-a.e. $(x,\gamma)\in\mathbb R^d\times\Gamma_{I_\mu}$ we have $$\begin{aligned}
\phi(\gamma(t_\mu))-\phi(\gamma(0))
\le&\langle \xi_\mu(\gamma(0)), \gamma(t_\mu)-\gamma(0)\rangle+C_\phi|\gamma(t_\mu)-\gamma(0)|^2\\
\le&t_\mu^{\alpha+1}\langle \xi_\mu(\gamma(0)),v_\mu(\gamma(0))\rangle+t_\mu^{\alpha+1}|\xi_\mu(\gamma(0))|\Delta^{v_\mu}_{\alpha,\gamma}(t_\mu)+\\
&+C_\phi t_\mu^{2(\alpha+1)}\left(\Delta^{v_\mu}_{\alpha,\gamma}(t_\mu)+|v_\mu(\gamma(0))|\right)^2.\end{aligned}$$ Integrating w.r.t. $\boldsymbol\eta$ and using Hölder’s inequality yields $$\label{eq:condLsigma}
\begin{split}
\sigma_{\Phi}(\mu_{t_\mu})-&\sigma_{\Phi}(\mu)\le\\
\le&-C_{2,\mu} t_\mu^{\alpha+1}+\int_{\mathbb R^d\times\Gamma_{I_\mu}}|\xi_\mu(x)|\cdot t_\mu^{\alpha+1}\Delta^{v_\mu}_{\alpha,\gamma}(t_\mu)\,d\boldsymbol\eta(x,\gamma)+2 C_\phi(C_{3,\mu}^2 t_\mu^2+C_{4,\mu}^2)t_\mu^{2(\alpha+1)}\\
\le&-C_{2,\mu} t_\mu^{\alpha+1}+t_\mu^{\alpha+2}\|\xi_\mu\|_{L^2_\mu}\cdot C_{3,\mu}+2 C_\phi(C_{3,\mu}^2 t_\mu^2+C_{4,\mu}^2)t_\mu^{2(\alpha+1)}\\
\le&t_\mu^{\alpha+1}\left(-C_{2,\mu}+C_{3,\mu}\|\xi_\mu\|_{L^2_\mu}t_\mu+2 C_\phi(C_{3,\mu}^2 t_\mu^2+C_{4,\mu}^2)t_\mu^{\alpha+1}\right)\\
\le&- 2K\cdot t_\mu^{\alpha+1}\cdot t_\mu= -2K\,t_\mu^{\alpha+2}.
\end{split}$$ Choose $$\begin{aligned}
r(q):=\min\left\{1,\frac{1}{2L},q^{1/\beta}\right\},&&Q(t,q):=-K\cdot t^{\alpha+1}\cdot r(q).\end{aligned}$$ In particular, we have $r(q)=0$ if and only if $q=0$, $Q(r(q),q)<0$ if $q\ne 0$, $$\dfrac{r(q)}{|Q(r(q),q)|}=\dfrac{1}{K\min\left\{1,\frac{1}{(2L)^{\alpha+1}},q^{\frac{\alpha+1}{\beta}}\right\}}=\dfrac{1}{K}\max\left\{1,(2L)^{\alpha+1},q^{-\frac{\alpha+1}{\beta}}\right\},$$ which is a decreasing integrable function of $q$. Furthermore, we notice that by definition $t_\mu=r(\sigma_\Phi(\mu))$ and $Q(r(\sigma_\Phi(\mu)),\sigma_\Phi(\mu))=-K\,t_\mu^{\alpha+2}$. Thus, from we get $$\sigma_{\Phi}(\mu_{r(\sigma_\Phi(\mu))})-\sigma_{\Phi}(\mu)\le2 Q(r(\sigma_\Phi(\mu)),\sigma_\Phi(\mu)),$$ and so we showed that $\tilde S^\Phi_2$ is $(r,Q)$-attainable. The result now follows from Theorem \[thm:suffSTLA\] and Proposition \[prop:Tattain\].
**Acknowledgements**: G.C. has been partially supported by Cariplo Foundation and Regione Lombardia via project *Variational Evolution Problems and Optimal Transport* and by MIUR PRIN 2015 project *Calculus of Variations*, together with FAR funds of the Department of Mathematics of the University of Pavia.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Landauer’s principle states that it costs at least ${k_\mathrm{B}}T\ln2$ of work to reset one bit in the presence of a heat bath at temperature $T$. The bound of ${k_\mathrm{B}}T\ln2$ is achieved in the unphysical infinite-time limit. Here we ask what is possible if one is restricted to finite-time protocols. We prove analytically that it is possible to reset a bit with a work cost close to ${k_\mathrm{B}}T\ln2$ in a finite time. We construct an explicit protocol that achieves this, which involves thermalising and changing the system’s Hamiltonian so as to avoid quantum coherences. Using concepts and techniques pertaining to single-shot statistical mechanics, we furthermore prove that the heat dissipated is exponentially close to the minimal amount possible not just on average, but guaranteed with high confidence in every run. Moreover we exploit the protocol to design a quantum heat engine that works near the Carnot efficiency in finite time.'
author:
- Cormac Browne
- 'Andrew J. P. Garner'
- 'Oscar C. O. Dahlsten'
- Vlatko Vedral
title: 'Guaranteed energy-efficient bit reset in finite time'
---
[**[Introduction.—]{}**]{} Landauer’s principle [@Szilard29; @Landauer61; @Bennett82] states that resetting a bit or qubit in the presence of a heat bath at temperature T costs at least kTln2 of work, which is dissipated as heat. It represents the fundamental limit to heat generation in (irreversible) computers, which is extrapolated to be reached around 2035 [@Frank05].
The principle is also a focal point of discussions concerning how thermodynamics of quantum and nano-scale systems should be formulated. Of particular interest to us here is the single-shot approach to statistical mechanics [@delRioARDV11; @Aberg11; @DahlstenRRV11; @EgloffDRV12]. This concerns statements regarding what is guaranteed to happen or not in any single run of an experiment, as opposed to what happens on average. This distinction is important for example in nano-scale computer components, in which large heat dissipations in individual runs of the protocol could cause thermal damage, even if the average dissipation is moderate.
In [@delRioARDV11] Landauer’s principle was assumed to hold in the strict sense that one can reset a uniformly random qubit at the exact work cost of $kT\ln2$ each run of an experiment. This assumption can be showed to be justified if one allows quasistatic protocols [@Aberg11; @EgloffDRV12].
Real experiments take place in finite time [@Jarzynski97; @SekimotoS97; @SerreliLKL07; @ThornSLS08; @Toyabe2010; @AurellMM11; @BerutAPCDL12; @BergliGK13; @Zulkowski13; @Koski14]. Our present Letter is motivated by the concern that fluctuations might be much greater in finite time scenarios, and that the single-shot optimality expressions to date may therefore not be physically relevant. Therefore we extend the protocol for bit reset used in [@delRioARDV11] to the finite time regime and analyse what changes. In this regime thermalisation is imperfect, and the quantum adiabatic theorem fails so that one cannot a priori assume that shifting energy levels does not change occupation probabilities. Moreover there are correlations between occupation probabilities at different times. We prove analytically that it is in fact possible to reset a qubit in finite time at a guaranteed work cost of $kT\ln2$, up to some small errors. We also make a natural extension of our results to a qubit heat engine which operates at the Carnot efficiency in finite time up to a small error. We derive bounds on the errors, proving that they fall off exponentially or doubly exponentially in the time taken for the protocol.
![ \[fig:2level\] [**‘Bit reset’ by raising $\mathbf{E_2}$ to infinity.**]{} The numbers indicate the occupation probability for the energy levels. We consider the extension of this protocol, running in finite time.](twolevel){width="45.00000%"}
[**[A quasistatic bit reset protocol.—]{}**]{} We examine a simple two-level system, with access to a heat bath and a work reservoir, which we will manipulate with a time-varying Hamiltonian in the regime as set out by [@AlickiHHH04]. The evolution of the system takes place through two mechanisms [@AlickiHHH04] (see also [@Kieu04; @EgloffDRV12]):
1. \[mech:el\_change\] Changes to the energy spectrum – identified with work cost/production $\mathrm{d}E_i$ for occupied energy level $i$, which has no effect on the occupation .
2. \[mech:therm\] Changes to probability distributions via interactions with a heat bath (thermalisation), with no changes to the energy spectrum, and thus no associated work cost/production.
Initially, the two degenerate energy levels of a random qubit are equally likely to be populated. The system is coupled to a heat bath at temperature $T$, and one energy level is quasistatically and isothermally raised to infinity, until the lower energy level is definitely populated (see Fig. \[fig:2level\]). If we want to perform bit reset (“erasure” in Landauer’s terminology), the system is then decoupled from the heat bath, and the second energy level is returned to its original value, such that the initial and final energy level configurations are the same, and only the populations have changed. In the quasistatic limit, each stage of raising the energy level has an average cost of $P_2 \mathrm{d}E$ where $P_2$ is the (thermal) population of the upper energy level. Thus raising the second level from 0 to infinity, the work cost of the entire protocol is given: $$\label{eq:quasistatic_work}
{\langle W\rangle} = \int_0^\infty P_2(E)\mathrm{d}E = {k_\mathrm{B}}T \ln2.$$
![ \[fig:therm\_pop\] [**A few steps of the protocol.**]{} Thermal (dashed) and actual (solid) upper energy level populations with respect to time. At each time, $t(n)$, the upper energy level is raised, altering the thermal state. ](thermpop){width="45.00000%"}
We now describe how these two elementary steps are impacted by the extension to finite time.
[**[Level-shifts in finite time: the negative role of coherence.—]{}**]{} A development of coherences during the level shifts would increase the amount of work that must be invested to perform a reset. For example, suppose the initial state is $\rho_i=p(a){|a\rangle}{\langle a|}+ p(b){|b\rangle}{\langle b|}$, where ${|a\rangle}$, ${|b\rangle}$ are the energy eigenstates with occupation probabilities $p(a)$, $p(b)$ respectively. Suppose the Hamiltonian changes very quickly and the new energy eigenstates are ${|a'\rangle}$, ${|b'\rangle}$. Using the scheme of [@quan08], we perform a projective measurement in the new energy eigenbasis, ${|a'\rangle}$, ${|b'\rangle}$ and define the work input of the step as the energy difference between the initial energy eigenstate and the final. The average work cost can be written as $$\begin{aligned}
\label{coherent_work}
{\langle W\rangle}\!\! &=& p(b)(p(b \to b')(E_{b}' - E_{b}) + p(b \to a')(-E_b)) \nonumber \\
&& +\hspace{0.5em} p(a)(p(a \to a')(0) + p(a \to b')(E_{b'})) \nonumber \\
&=& p(b)(E_{b}' - E_{b}) + (p(a) - p(b))p(a \to b')E_{b'},\end{aligned}$$ where $E_i$ is the energy of associated with state $i$, and $p(a \to b')$ indicates the probability of transitioning from $a$ to $b'$. (In our protocol, $E_a\!\!=\!\!E_{a'}\!\!=\!0$; the second line follows by noting that the transition probabilities are doubly-stochastic due to the Born rule). Since $E_{b'}\!>\!E_b\!\geq\!0$, and $p(a)\!\geq\!p(b)$, in order for this expression to contribute the least work cost possible, we set the transition probability $p(a \rightarrow b')$ to zero, i.e. by not causing any .
We therefore avoid coherence, using either one of two ways: (i) By allowing for the experimenter to choose a path of Hamiltonians that share the same energy eigenstates and only differ in energy eigenvalues, as in the standard model for Zeeman splitting (Appendix \[app:avoidquantum\]), (ii) If the experimenter is instead given a fixed Hamiltonian path which is not of this kind, she may actively remove the coherences by an extra unitary being applied (Appendix \[app:reversequantum\]). This method of active correction is similar to the strategy employed in super-adiabatic processes [@Berry09; @Deffner14; @DelCampo13]. In either case, at any point of the protocol the density matrix of the system will be diagonal in the instantaneous energy eigenbasis.
[**[Partial thermalisation can be represented by partial swap matrices.—]{}**]{} The stochastic ‘partial swap’ matrix is equivalent to all other classical models of thermalisation for a two level system [@ZimanSBHSG01; @ScaraniZSGB02]. In this model, the heat bath is a large ensemble of thermal states which have some probability $p$ of being swapped with the system state in a given time interval.
We model a period of extended thermalization as being composed of a series of $t$ partial swap matrices, each representing a unit time step. Multiplying these together gives a single swap matrix with probability $P_\mathrm{sw}(t)$ of swapping with the thermal state: $$\label{eq:swap_prob_stacking}
P_\mathrm{sw}(t) = 1-(1-p)^t.$$ (See Appendix \[app:partialswap\] for details.) The influence of a finite thermalisation time, $t$, manifests as a degradation of the thermalisation quality through a lowered $P_\mathrm{sw}(t)$.
[**[The average extra work cost of finite time is exponentially suppressed.—]{}**]{} It is possible to quantify to what extent the state at the end of a period of thermalisation differs from a thermal state by employing the trace-distance $\delta(\rho,\sigma) := \frac{1}{2} \mathrm{Tr} ( \sqrt{ ( \rho - \sigma)^{\dagger}(\rho - \sigma )} ) $ (see e.g. [@LevinPE09]). Throughout the protocol we monotonically raise the energy of the second level and thus monotonically decrease the occupation probability of the second level. This makes it possible to bound the trace-distance after a period of thermalisation by the trace-distance between the desired thermal state and the occupation probability with degenerate energy levels: $$\label{eq:trace_distance_bound}
\delta \leq \left(\dfrac{1}{1+\exp \left( -\beta E_2 \right)}- \dfrac{1}{2} \right) \left( 1 - p \right)^t.$$ (Full derivation in Appendix \[app:deltabound\].)
We thus see that the occupation probabilities get exponentially close to to the thermal distribution as a function of the number of time steps spent thermalizing. It follows that the average work ${\langle W\rangle}$ required to raise the upper level from zero to some maximum energy $E_\mathrm{max}$, only thermalising for $t$ time steps at each stage, is bounded by
$$\begin{aligned}
\label{eq:work_cost}
{\langle W_\mathrm{quasi}\rangle} & \leq \,{\langle W\rangle}\, \leq &
{\langle W_\mathrm{quasi}\rangle} + \nonumber \\
&& (1-p)^t \left(\dfrac{E_\mathrm{max}}{2} - {\langle W_\mathrm{quasi}\rangle}\right),\end{aligned}$$
where ${\langle W_\mathrm{quasi}\rangle}$ is the quasistatic work cost of raising the second level to $E_\mathrm{max}$: $${\langle W_\mathrm{quasi}\rangle} = {k_\mathrm{B}}T\ln\left(\dfrac{2}{1+\exp\left( -\frac{E_\mathrm{max}}{{k_\mathrm{B}}T} \right) }\right).$$ (Full derivation in Appendix \[app:workbound\].)
Eq. \[eq:work\_cost\] shows that the average extra work cost of reset decreases exponentially in the thermalisation time, $t$.
[**[Work fluctuations in a single shot reset are doubly exponentially suppressed.—]{}**]{} The preceding discussion applies to the [*average work cost*]{} over many bit resets. However, the average work cost is not always the most useful parameter if we wish to design a physical device which performs a bit reset. Another practical parameter is the guaranteed work from single shot thermodynamics, which provides an upper bound on the work required per reset, $W_\mathrm{max}$. One could envision scenarios where if this value is exceeded the device breaks: either because the work must ultimately be dissipated as heat which above a certain level could damage the device, or if the work reservoir itself has a finite capacity and would be exhausted by an attempt to draw on too large a value.
To derive such a parameter, or other similar parameters (including the average work value), it is useful to generate a [*work distribution*]{}: the probability density function for the work cost of the bit reset process.
It has been established in the quasistatic case that the work cost in any procedure involving shifting energy levels (such as bit reset) is tightly peaked around the average value [@Aberg11; @EgloffDRV12]. Applying the method of Egloff and co-authors [@EgloffDRV12], the deviation from the average work in the quasistatic case when raising the second level by $\mathcal{E}$ per step over $N$ steps is bounded by: $$\label{eq:ss_work_quasi}
P(|W-{\langle W\rangle}| \geq \omega) \hspace{0.5em} \leq \hspace{0.5em} 2\exp\left( - \dfrac{2 \, \omega^2}{N \mathcal{E}^2} \right).$$
This result arises from modelling the work input as a stochastic process, and applying the McDiarmid inequality [@McDiarmid89] which bounds the deviation from the mean value. By thermalising fully in every step, the question of which energy level is occupied at each step can be represented by a sequence of independent random variables with distributions given by the thermal population associated with the particular energy gaps. The work cost is a function of these random variables.
We note that altering the energy level at one stage in the perfectly thermal regime will only affect the total work cost by the difference associated with that one step. However with partial thermalisation, the states the system is in at two different stages are not independent of each other. Altering the energy level of one stage has a knock-on effect on the work cost for all subsequent stages. We show in Appendix \[app:fluxbound\] that this effect falls off in a manner inversely proportional to the probability of swapping, and thus derive a new finite time version of Eq. \[eq:ss\_work\_quasi\], which explicitly takes into account this sensitivity: $$\label{eq:ss_work_partial}
P(|W-{\langle W\rangle}| \geq \omega) \hspace{0.5em} \leq \hspace{0.5em} 2\exp\left( - \dfrac{2 \, \omega^2{P_\mathrm{sw}(t)}^2}{N \mathcal{E}^2} \right).$$ When $t\to\infty$ such that $P_\mathrm{sw}(t)\to1$, we recover the perfect thermalisation case. As $P_\mathrm{sw}(t)$ is itself an exponential function of time (Eq. \[eq:swap\_prob\_stacking\]), we see this limit rapidly converges on the quasistatic variant as $t$ increases. Hence, fluctuations are supressed as a double exponential function of time spent thermalising. When $P_\mathrm{sw}=0$ the bound becomes meaningless, as changing what happens at one stage in the protocol can have an unbounded effect on the final work cost.
We can combine this spread with the average result in Eq. \[eq:work\_cost\] to determine $W^\epsilon_\mathrm{max}$: the maximum work cost in a single shot, guaranteed only to be exceeded with failure probability $\epsilon$ (mathematically, $W^\epsilon_\mathrm{max}$ is defined to satisfy $\int_{-\infty}^{W^\epsilon_\mathrm{max}} P(W) \mathrm{d}W = 1-\epsilon$ for work distribution $P(W)$): $$\begin{aligned}
\label{eq:guaranteed_work}
W_\mathrm{max}^\epsilon & \leq & \left(1-(1-p)^t\right){\langle W_\mathrm{quasi}\rangle} +\dfrac{1}{2}(1-p)^t E_\mathrm{max} \nonumber \\
&&+ \hspace{1em}\dfrac{1}{1-(1-p)^t} \sqrt{\dfrac{\ln \left( 2/\epsilon \right)}{2N}} E_\mathrm{max}. \label{eq:w_max}\end{aligned}$$ (Full derivation in Appendix \[app:wemax\].) From Eqs. \[eq:work\_cost\], \[eq:ss\_work\_quasi\] & \[eq:guaranteed\_work\] we see that spending time thermalising has two-fold benefit: it lowers the average cost, and reduces the spread of possible work values around this average.
[**[Reset error is exponentially suppressed.—]{}**]{} Initially, we have complete uncertainty about the system - the energy levels are degenerate and the particle is equally likely to occupy either of them. At the end of the raising procedure, the system has been ‘reset’ to one energy level, with probability of failure, $P_{fail}$ (defined as the occupation probability of the second energy level), upper bounded by (using the probability distribution derived in Appendix \[app:deltabound\]) $$P_\mathrm{fail} \leq \left(1 + (1-p)^t \right) \dfrac{\exp \left(- \beta E_{max} \right)}{1 + \exp \left(- \beta E_{max} \right) } - \dfrac{1}{2}(1-p)^t.$$
[**[The work cost of resetting multiple bits scales favourably.—]{}**]{} In [@delRioARDV11] the general protocol involves resetting $n$ qubits in the state $\rho={\mathbbm{1}}/2\otimes{\mathbbm{1}}/2\cdot\cdot\cdot \otimes {\mathbbm{1}}/2$ and we now consider how the errors scale in this case.
Recall that for a single qubit, thermalising for time $t$ bounds the trace-distance $\delta$ between the actual state and the true thermal state according to Eq. \[eq:trace\_distance\_bound\]. This results in an additional work cost of up to $\delta E_\mathrm{max}$ across the protocol. For $n$ bits, in the very worst case each bit will be $\delta$ away from its thermal state, and so trivially the extra work cost will be bounded by $n \delta E_\mathrm{max}$. Thus the average work cost of reset scales linearly with the number of bits (as with the quasistatic case). This argument can be naturally modified to the single-shot case (more detail in Appendix \[app:manyqubits\]), and one may therefeore say that the the multi-qubit statements in [@DahlstenRRV11; @delRioARDV11] also remain relevant in the case of finite time.
[**[Work extractable in finite time.—]{}**]{} Although we have thus far concentrated on bit reset, it is also possible to consider the inverse protocol: work extraction by lowering the upper energy level in contact with a hot bath. The derivation of this proceeds exactly as before, with the limits of the integration reversed. This allows us to extract work bounded by: $$\begin{aligned}
\label{eq:work_extraction}
-{\langle W_\mathrm{quasi}\rangle} & \leq & {\langle W_{out}\rangle} \nonumber \\
& \leq & -{\langle W_\mathrm{quasi}\rangle} + \nonumber \\
&& \hspace{0.5em} (1-p)^t \left(\dfrac{E_\mathrm{max}}{2} - {\langle W_\mathrm{quasi}\rangle}\right).\end{aligned}$$
If we have access to two baths at different temperatures, we can extract net work by running a bit reset coupled to a cold bath at temperature $T_{C}$ followed by a work extraction coupled to a hot bath at temperature $T_{H}$. The engine cycle formed (shown in Fig. \[fig:TS\_cycle\]) has net work exchange per cycle bounded by $$\begin{aligned}
{\langle W_\mathrm{net}\rangle} & \leq &-P_{sw}(t){k_\mathrm{B}}\left(T_{H}-T_{C}\right)\ln 2 \nonumber \\
&& -P_{sw}(t){k_\mathrm{B}}\left(T_{C}\ln\left(Z_{C}^\mathrm{max}\right) + T_{H}\ln\left(Z_{H}^\mathrm{max}\right)\right)\nonumber \\
&& + \left(1-P_{sw}(t)\right) E_\mathrm{max}, \nonumber \\
& & \hspace{-1em}= P_{sw}(t){\langle W_\mathrm{quasi}\rangle} + \left(1-P_{sw}(t)\right) E_\mathrm{max},\end{aligned}$$ where $Z^{max}_{C/H}=1+\exp\left( -E_\mathrm{max}/{k_\mathrm{B}}T_{C/H}\right)$ and ${\langle W_{net}\rangle}$ is the difference between the work input (Eq. \[eq:work\_cost\]) and the work output (Eq. \[eq:work\_extraction\]) with the appropriate temperatures.
The final line follows because the first two lines are the same as in the quasistatic case. The final term is independent of the temperatures of the baths; but instead accounts for the effect of finite time on the degree of over-population of the upper level when raising, and under-population of the upper level when lowering.
Adjusting the energy level in finite steps has necessitated a move away from the truly isothermal process of the quasistatic regime. In a two-level system, any distribution where the lower energy level is more populated than the upper may be written as a thermal state for some temperature. During bit reset, after the upper energy level is raised, it is over-populated with respect to the thermal population associated with the cold bath temperature. We can interpret this as a rise in the system temperature appearing as the saw-tooth pattern in Fig. \[fig:TS\_cycle\].
![ \[fig:TS\_cycle\] [**Entropy-temperature diagram for the engine cycle.**]{} The block shaded area shows the net work exchanged by the system, which is less than that of the quasistatic (Carnot) cycle shown by the dashed rectangle filled with dots behind. For illustrative purposes, the number of energy level adjustments has been greatly reduced, and size of the associated shifts in temperature greatly exaggerated. ](TScycle){width="45.00000%"}
[**[There is a speed limit for positive power output.—]{}**]{} By reconciling the concept of time with these processes, we can now talk meaningfully about the [*power*]{} of the cycle, $\mathcal{P}$: the net work exchange divided by the total time taken to complete a full cycle of bit reset and extraction. The total time is found by multiplying the amount of time spent thermalising in one stage, $t$, by the number of stages in both halves of the protocol, $N = 2 E_\mathrm{max}/\mathcal{E}$. We can upper bound the power of the entire cycle as a function of the maximum energy gap $E_\mathrm{max}$ and of $t$: $$\label{eq:finite_power_short}
\mathcal{P} \leq \dfrac{P_{sw}(t){\langle W_\mathrm{quasi}\rangle} + \left(1-P_{sw}(t)\right) E_\mathrm{max}}{2 \dfrac{E_\mathrm{max}}{\mathcal{E}} t}.$$
A positive $\mathcal{P}$ indicates the system will draw energy in from its surroundings, and therefore we note that the engine does not produce work in all parameter regimes. If one attempts to operates an engine (with ${\langle W_\mathrm{quasi}\rangle} < 0$) quicker than the limit $$\label{eq:neg_work}
t < \dfrac{-1}{\ln(1-p)}\ln\left(1-\dfrac{E_\mathrm{max}}{{\langle W_\mathrm{quasi}^\mathrm{net}\rangle}}\right),$$ then the partial thermalisation can potentially contribute an excess work cost greater than the engine’s quasistatic work output.
[**[Near-Carnot efficiency can be achieved in finite time.—]{}**]{} The efficiency $\eta$ of a cycle may be defined as the net work extracted divided by the maximum work value of one bit of information ($-{k_\mathrm{B}}T_{H}\ln2$). We write this as $$\begin{aligned}
\eta_\mathrm{quasi} - \dfrac{(1-p)^t E_\mathrm{max}}{{k_\mathrm{B}}T_{H}\ln2}
& \leq \eta \leq & \eta_\mathrm{quasi},\end{aligned}$$ where $\eta_\mathrm{quasi}$ is the quasistatic efficiency of raising to $E_\mathrm{max}$ over an infinitely long time, and can be related to the Carnot efficiency $\eta_\mathrm{C}$ by $$\eta_\mathrm{quasi} = \eta_\mathrm{C}- \dfrac{\ln\left(Z_{H}^\mathrm{max}\right) - \dfrac{T_{C}}{T_{H}}\ln\left(Z_{C}^\mathrm{max}\right)}{\ln2}.$$
We see that if $t$ is small, the cost of raising the populated upper energy level takes its toll on the efficiency, potentially plunging it into negative values (work loss) when $t$ does not satisfy Eq. \[eq:neg\_work\]. Conversely, as $t\to\infty$, the process is maximally efficient, but has no power.
[**[Conclusion.—]{}**]{} We have showed that one can reset a qubit in finite time at a guaranteed work cost of kTln2 up to some errors that fall off dramatically in the time taken for the protocol. As mentioned, several key results in single-shot statistical mechanics assume that this can be achieved perfectly [@DahlstenRRV11; @delRioARDV11; @EgloffDRV12; @Aberg11; @FaistDOR12]. Our Letter accordingly shows that the optimality statements of those papers are still relevant to finite time protocols. Moreover our results naturally extended to show that the Carnot efficiency can be achieved by a qubit engine in finite time up to an error that falls of exponentially in the cycle time.
[**[Acknowledgments.—]{}**]{} We acknowledge fruitful discussions with Alexia Auffèves, Felix Binder, Dario Egloff, John Goold, Kavan Modi, Felix Pollock and Nicole Yunger Halpern. We are grateful for funding from the EPSRC (UK), the Templeton Foundation, the Leverhulme Trust, the EU Collaborative Project TherMiQ (Grant Agreement 618074), the Oxford Martin School, the National Research Foundation (Singapore), and the Ministry of Education (Singapore).
[10]{}
L. Szilard, Zeitschrift für Physik [**53**]{}, 840 (1929).
R. Landauer, IBM Journal of Research and Development [**5**]{}, 183 (1961).
C. H. Bennett, International Journal of Theoretical Physics [**21**]{}, 905 (1982).
M. P. Frank, 2013 IEEE 43rd International Symposium on Multiple-Valued Logic [ **0**]{}, 168 (2005).
L. del Rio, J. Aberg, R. Renner, O. C. O. Dahlsten, and V. Vedral, Nature [**474**]{}, 61 (2011).
J. Aberg, Nature Communications [**4**]{} (2011/2013), arXiv:1110.6121.
O. C. O. Dahlsten, R. Renner, E. Rieper, and V. Vedral, New Journal of Physics [**13**]{}, 053015 (2011).
D. Egloff, O. C. O. Dahlsten, R. Renner, and V. Vedral, (2012), arXiv:quant-ph/1207.0434.
C. Jarzynski, Phys. Rev. Lett. [**78**]{}, 2690 (1997).
K. Sekimoto and S. Sasa, Journal of the Physical Society of Japan [**66**]{}, 3326 (1997).
V. Serreli, C. Lee, E. R. Kay, and D. A. Leigh, Nature [**445**]{}, 523 (2007).
J. Thorn, E. Schoene, T. Li, and D. Steck, Physical Review Letters [**100**]{} (2008).
S. Toyabe, T. Sagawa, M. Ueda, E. Muneyuki, and M. Sano, Nature Physics [**6**]{}, 988 (2010).
E. Aurell, C. Mejia-Monasterio, and P. Muratore-Ginanneschi, Physical Review Letters [**106**]{} (2011), : 1012.2037.
A. Bérut [*et al.*]{}, Nature [**483**]{}, 187 (2012).
J. Bergli, Y. M. Galperin, and N. B. Kopnin, Physical Review E [**88**]{} (2013), : 1306.2742.
P. R. Zulkowski and M. R. DeWeese, :1310.4167 \[cond-mat\] (2013), : 1310.4167.
J. V. Koski, V. F. Maisi, J. P. Pekola, and D. V. Avering, (2014), 1402.5907.
R. Alicki, M. Horodecki, P. Horodecki, and R. Horodecki, Open Systems & Information Dynamics [**11**]{}, 205 (2004).
T. D. Kieu, Phys. Rev. Lett. [**93**]{} (2004).
H. T. Quan and H. Dong, :0812.4955 \[cond-mat\] (2008), : 0812.4955.
M. V. Berry, Journal of Physics A: Mathematical and Theoretical [**42**]{}, 365303 (2009).
S. Deffner, C. Jarzynski, and A. del Campo, Physical Review X [**4**]{} (2014), : 1401.1184.
A. del Campo, J. Goold, and M. Paternostro, :1305.3223 \[cond-mat, physics:quant-ph\] (2013), : 1305.3223.
M. Ziman [*et al.*]{}, (2001), arXiv:quant-ph/0110164.
V. Scarani, M. Ziman, P. Štelmachovič, N. Gisin, and V. Bužek, Physical Review Letters [**88**]{}, 097905 (2002), 0110088.
D. A. Levin, Y. Peres, and E. L. Wilmer, (American Mathematical Soc., 2009).
C. McDiarmid, London Mathematical Society Lecture Note Series [**141**]{} (1989).
P. [Faist]{}, F. [Dupuis]{}, J. [Oppenheim]{}, and R. [Renner]{}, ArXiv e-prints (2012), 1211.1037, arXiv:1211.1037.
Quantum coherences {#app:quantum}
==================
Quantum coherence during energy level steps can be avoided. {#app:avoidquantum}
-----------------------------------------------------------
We can avoid inducing quantum coherences when raising energy levels, by choosing to describe the energy levels by Hamiltonians which are diagonal in the same basis at each stage, written as: $$\mathcal{H}(n) = \left(\begin{array}{cc}
E_1(n) & 0 \\
0 & E_2(n)
\end{array}\right) = \left(\begin{array}{cc}
0 & 0 \\
0 & \mathcal{E} n
\end{array}\right).$$ Because they are all diagonal in the same basis, they will always commute; $[\mathcal{H}(n),\mathcal{H}(m)] = 0$ for all $n$ and $m$. This is a reasonable regime to physically implement, for example, by varying two hyperfine levels in an atom by the application of a time-dependent external magnetic field.
Defining $P_1(n)$ and $P_2(n)$ as the populations at stage $n$ of lower and upper energy levels respectively, we write the density matrix $\rho(n)$ in the energy eigenbasis as: $$\label{eq:density_rep}
\rho(n) = \left(\begin{array}{cc}
P_1(n) & 0 \\
0 & P_2(n)
\end{array}\right).$$
In order to apply the Schrödinger equation, we can interpolate into a continuous case and consider the transition from stage $n$ to stage $n+1$ as given by a time parameter $t$ that varies from $0$ to $\tau$ in each step: $$\label{eq:energy_step_H}
\mathcal{H}(n, t) = \left(\begin{array}{cc}
0 & 0 \\
0 & \mathcal{E} \left(n + t/\tau \right)
\end{array}\right),$$ such that $\mathcal{H}(n, \tau) = \mathcal{H}(n+1, 0)$.
By using the Schrödinger equation to generate an equation for unitary evolution from stage $n$ to stage $n+1$, as we have chosen a $\mathcal{H}$ which commutes with itself throughout the process, we can write down: $$U(n\to n\!+\!1) = \exp \left( -\dfrac{i}{\hbar} \int_0^\tau \mathcal{H}(t) \mathrm{d}t \right).$$ $\mathcal{H}(t)$ is always diagonal and so it follows that $\int_0^\tau \mathcal{H}(t) \mathrm{d}t$ must also be diagonal. Taking the exponent of this sum multiplied by a constant (${i}/{\hbar}$) implies that $U$ must also be a diagonal operator.
The evolution of a mixed state from time $t_i=0$ (i.e. step $n$) to time $t_f$ can thus be written as $$\label{eq:unitary_ev}
\rho(n\!+\!1) = U(n\to n\!+\!1) ~ \rho(n) ~ U(n\to n\!+\!1)^\dag.$$ Providing we always start in a diagonal state, all terms on the right hand side are diagonal, and thus $\rho(m)$ is diagonal for all $m$.
In fact, we can go further and explicitly calculate $U$ due to the simple form of Eq. \[eq:energy\_step\_H\]: $$\begin{aligned}
U(n\to n\!+\!1) & = & \exp \left( -\dfrac{i}{\hbar} \left(\begin{array}{cc}
0 & 0 \\
0 & \mathcal{E} \left(n+\frac{1}{2}\right)\tau
\end{array}\right) \right), \nonumber\\
& = & \left( \begin{array}{cc}
1 & 0 \\
0 & \exp\left( -\frac{i}{\hbar} \mathcal{E} \left(n+\frac{1}{2}\right)\tau \right)
\end{array}\right).\end{aligned}$$ This shows us that the only effect of evolution is to introduce a phase between the upper and lower energy levels, $\phi = \frac{\mathcal{E}}{\hbar} \left(n+\frac{1}{2}\right)\tau$. If the upper and lower energy levels are an incoherent mixture, then this transformation will have no effect whatsoever on the final state; rewriting Eq. \[eq:density\_rep\] as $\rho(n) = P_1(n) {|E_1(n)\rangle\!\langleE_1(n)|} + P_2(n) {|E_2(n)\rangle\!\langleE_2(n)|}$, the evolution in Eq. \[eq:unitary\_ev\] reduces to $$\begin{aligned}
\rho(n\!+\!1) \hspace{-2.5em} & & \nonumber \\
& = & U\left( P_1(n) {|E_1(n)\rangle\!\langleE_1(n)|} + P_2(n) {|E_2(n)\rangle\!\langleE_2(n)|} \right) U^\dag, \nonumber \\
& =& P_1(n) {|E_1(n)\rangle\!\langleE_1(n)|} + P_2(n) e^{-i\phi} {|E_2(n)\rangle\!\langleE_2(n)|} e^{i\phi}, \nonumber \\
& =& P_1(n) {|E_1(n)\rangle\!\langleE_1(n)|} + P_2(n) {|E_2(n)\rangle\!\langleE_2(n)|}, \nonumber \\
& =& \rho(n).\end{aligned}$$
The cost of coherent excitations can be actively reversed. {#app:reversequantum}
----------------------------------------------------------
If the different Hamiltonians are not intrinsically diagonal, we can consider an energy level step as taking us from the initial Hamiltonian $H = \sum_i E_i {|E_i\rangle\!\langleE_i|}$ to the final Hamiltonian $\tilde{H} = \sum_i \tilde{E_i} {|\tilde{E_i}\rangle\!\langle\tilde{E_i}|}$, where not only have the eigenvalues changed, but there is also a change of basis. A change in basis can always be associated with unitary $U$ such that $${|\tilde{E_i}\rangle\!\langle\tilde{E_i}|} = U{|E_i\rangle\!\langleE_i|} U^\dag,$$ and one can express $\tilde{H} = \sum_i \tilde{E_i} U {|E_i\rangle\!\langleE_i|} U^\dag$.
![ \[fig:coherent\_excitation\] [**Coherent excitation.**]{} (a) shows the state in the Bloch sphere of the initial Hamiltonian, and (b) of the final. The state $\rho$, diagonal in the initial Hamiltonian, is a [*coherently excited*]{} state in the final Hamiltonian. The same passive transformation $U$ between initial and final Hamiltonians can be applied actively on $\rho$ to map it to $\rho'$, compensating for the work cost of coherent excitation. ](coherence2){width="45.00000%"}
Consider the initial state of the system, given by a density matrix written diagonally in the initial Hamiltonian basis: $\rho = \sum_i P_i {|E_i\rangle\!\langleE_i|}$. The system energy before the change in Hamiltonian is defined as $\mathrm{Tr}\left(\rho H \right)$ and afterwards as $\mathrm{Tr}\left(\rho \tilde{H} \right)$, and so we see that the change in system energy (and thus by conservation of energy, the work cost from reservoir) is given by $$\Delta W = \mathrm{Tr}(\rho \tilde{H} ) - \mathrm{Tr}(\rho H ).$$ A visualisation that this might result in an work cost is apparent in Figure \[fig:coherent\_excitation\] that the state $\rho$ appears higher in the sphere in terms of $\tilde{H}$ than $H$. One must take care, however, to note that the energy scale along the vertical axis of each Bloch sphere is different. For bit reset, the energy gap is increasing and $\tilde{E_2} - \tilde{E_1} > E_2 - E_1$.
If we were to let the system decohere, the final work cost would be exactly this value (the density matrix in the first term is replaced by a new density matrix with the terms off-diagonal in $\tilde{H}$ removed; but as these terms only appear inside a trace, they do not contribute to the value).
However, if instead we apply an active unitary transformation on $\rho$ to bring it into a new state $\rho'$, which is diagonal in the new energy eigenbasis $\rho = \sum_i P_i {|E_i\rangle\!\langleE_i|} = \sum_i P_i U^\dag {|\tilde{E_i}\rangle\!\langle\tilde{E_i}|} U$ and so the obvious choice is to take $\rho' = U\rho U^\dag$ given as $$\rho' = \sum_i P_i {|\tilde{E_i}\rangle\!\langle\tilde{E_i}|}.$$
If we now consider the overall change in energy between the initial state before the change in Hamiltonian, and the final state after the change in Hamiltonian followed by the application of the correcting unitary: $$\begin{aligned}
\Delta W & = & \mathrm{Tr}(\rho' \tilde{H} ) - \mathrm{Tr}(\rho H ) \\
& = & \mathrm{Tr}\left(\sum_i P_i {|\tilde{E_i}\rangle\!\langle\tilde{E_i}|} \sum_j \tilde{E_j} {|\tilde{E_j}\rangle\!\langle\tilde{E_j}|} \right) \nonumber\\
& & - \mathrm{Tr}\left(\sum_i P_i {|E_i\rangle\!\langleE_i|} \sum_j E_j {|E_j\rangle\!\langleE_j|} \right) \\
& = & \sum_i P_i \left(\tilde{E_i} - E_i\right),\end{aligned}$$ and noting that $\Delta E_i = \tilde{E_i}-E_i$, this recovers the same work cost as if there had been no coherences at all: $$\Delta W = \sum_i P_i \Delta E_i.$$ We have compensated for our passive transformation $U$ on the Hamiltonian basis by actively applying the same transformation to the density matrix.
Average work cost {#app:average}
=================
Partial swap model of thermalisation {#app:partialswap}
------------------------------------
Consider now the process of thermalisation. At stage $n$, when the energy levels are split by $\Delta E = n\mathcal{E}$, the associated thermal populations are given by the Gibbs distribution: $$\label{eq:thermal_population}
\mathbf{P^\mathrm{th}} (n) = \dfrac{1}{1+ e^{ -\beta n \mathcal{E}}}
\left(\begin{array}{c}
1 \\
e^{-\beta n \mathcal{E}}
\end{array}\right)$$
Consider a stochastic transformation in which we have a probability $P_\mathrm{swap}$ of replacing the current state with the appropriate thermal state, and probability $(1-P_\mathrm{swap})$ of leaving the state alone. This is expressed as the stochastic matrix $$\label{eq:swap_matrix}
M(n) = (1 - P_\mathrm{swap}) ~ {\mathbbm{1}}+ P_\mathrm{swap} M_\mathrm{th}(n),$$ where $$\label{eq:thermal_matrix}
M_\mathrm{th}(n) = \left( \begin{array}{cc}
P_{1}^\mathrm{th}(n) & P_{1}^\mathrm{th}(n) \\
P_{2}^\mathrm{th}(n) & P_{2}^\mathrm{th}(n)
\end{array} \right)$$ and $P_{i}^\mathrm{th}(n)$ is the $i^\mathrm{th}$ component of $\mathbf{P^\mathrm{th}}(n)$. We verify that this matrix has the defining behaviour of a thermalising process by considering its eigenvectors, and noting that it evolves all probability distributions towards the thermal distribution.
Under the assumption that we raise the energy levels in such a way that the population of the levels is undisturbed, we can write a recursive relationship between the populations $\mathbf{P}{(n)}$ at the end of each stage (that is the populations having adjusted the energy level for the $n^\mathrm{th}$ time, and then allowing it to partially thermalise): $$\label{eq:recursive_relation}
\mathbf{P}{(n)} = \big(M\left(n\right)\big)^{t(n)}\ \mathbf{P}{(n-1)},$$ where $t(n)$ is the number of times we apply the partial thermalisation matrix at stage $n$.
We attempt instead to express this energy level population as the ideal thermal distribution $\mathbf{P^{(n)}_\mathrm{th}}$ perturbed by a small difference. $$\begin{aligned}
\mathbf{P}(n) & = & \mathbf{P^{th}}(n) + \boldsymbol\delta (n)
$$
We calculate the correction term $\boldsymbol\delta$ explicitly by noting that the probability of not swapping with with $\mathbf{P^{th}}(n)$ in any stage is given by $(1-P_\mathrm{swap})^{t(n)}$, and so our correction is to subtract this amount of the thermal population and add instead the same amount of the population of the previous stage: $$\label{eq:trace_distance}
\boldsymbol\delta (n) = (1-P_\mathrm{swap})^{t(n)} \left( -\mathbf{P^{th}}(n) + \mathbf{P}(n-1) \right)$$
Further note that, as $\mathbf{P^{th}}(n)$ and $\mathbf{P}(n)$ are both valid probability distributions, the sum of the components of $\boldsymbol\delta$ must sum to zero. This can specifically seen to be the case here by noting that the second component of each contributing part of $\boldsymbol\delta$ is just just one minus the first component. Indeed for a two level system, we can express $\boldsymbol\delta$ as: $$\boldsymbol\delta(n) = \pm \delta(n) \left( \begin{array}{c}
1 \\
-1
\end{array} \right),$$ where $\delta$ is a quantity known also as the variational distance (see e.g. [@LevinPE09]). The sign of $\delta$ depends on whether we are raising or lowering the upper energy level. Incomplete thermalisation after raising the upper energy level leaves a higher probability of being in the upper level than the true thermal probability. There is a lower than thermal probability when lowering the upper energy level.
Bounding the variational distance. {#app:deltabound}
----------------------------------
We can bound the variational distance $\delta(n)$ by noting that, since we increase the energy of the second level throughout the protocol, the effect of thermalization is always to increase the occupation probability of the first energy level $P_1^{(n-1)}\geq P_1^{(0)}$. This can be seen from Eq. \[eq:trace\_distance\]. Thus we have: $$\begin{aligned}
\delta(n)& =& (1-p)^t \left(\dfrac{1}{1+\exp \left(-n \beta \mathcal{E} \right) } - P_1^{(n-1)} \right) \nonumber \\
& \leq & (1-p)^t \left( \dfrac{1}{1+\exp \left(-n \beta \mathcal{E} \right) } - \dfrac{1}{2} \right).\end{aligned}$$ Having derived a bound on the variational distance we can now write the probability distribution (after some rearranging) as: $$\mathbf{P}(n) \leq \left( \begin{array}{c}
\left(1 + (1-p)^t \right) \dfrac{1}{1 + \exp \left(-n \beta \mathcal{E} \right) } - \dfrac{1}{2}(1-p)^t \\
\left(1 + (1-p)^t \right) \dfrac{\exp \left(-n \beta \mathcal{E} \right)}{1 + \exp \left(-n \beta \mathcal{E} \right) } - \dfrac{1}{2}(1-p)^t
\end{array} \right)$$
Bounding the average work cost. {#app:workbound}
-------------------------------
In the worst case scenario, at each stage before raising the energy level, we are exactly $\delta$ away from the true thermal population. Raising the energy level by $\mathcal{E}$ at stage $n$ has an associated work cost of $P_2(n)\mathcal{E} = \left(P^\mathrm{th}_2(n)+\delta\right)\mathcal{E}$. Writing out $P^\mathrm{th}_2(n)$ explicitly (using Eq. \[eq:thermal\_population\]), and summing over all $N$ stages in the bit reset, the (worst case) work cost is $$\begin{aligned}
&{\langle W\rangle}&=\sum_{n=0}^{N}\left(\dfrac{\exp\left( \dfrac{-n\mathcal{E}}{{k_\mathrm{B}}T}\right) }{1+\exp\left( \dfrac{-n\mathcal{E}}{{k_\mathrm{B}}T} \right) } \right. + \nonumber \\
&+&\left. (1-p)^t \left( \dfrac{1}{1+\exp \left( \dfrac{-n\mathcal{E}}{{k_\mathrm{B}}T} \right) } - \dfrac{1}{2} \right) \right) \mathcal{E}.\hspace{1em}\end{aligned}$$
For small $\mathcal{E}$, we can approximate this sum as an integral and, writing $\mathcal{E}\mathrm{d}n=\mathrm{d}E_{2}$, and $N\mathcal{E} = E_\mathrm{max}$:
$$\begin{aligned}
{\langle W\rangle} & \approx &
\int_{0}^{E_\mathrm{max}} \hspace{-1em}\mathrm{d}E_{2}
\left(\dfrac{\exp\left( -\frac{E_{2}}{{k_\mathrm{B}}T} \right) }{1+\exp\left( -\frac{E_{2}}{{k_\mathrm{B}}T}\right) }\hspace{1em}+\right. \nonumber \\*
&&+ \left. (1-p)^t \left( \dfrac{1}{1+\exp \left( \frac{-E_{2}}{{k_\mathrm{B}}T} \right) } - \dfrac{1}{2} \right) \right) \nonumber \\
&=& \left(1-(1-p)^t\right) {k_\mathrm{B}}T\ln\left(\dfrac{2}{1+\exp\left( -\frac{E_\mathrm{max}}{{k_\mathrm{B}}T} \right) }\right) \nonumber \\*
&&+ \dfrac{1}{2}(1-p)^t \left(E_\mathrm{max} \right)\end{aligned}$$
Note that the term in the denominator of the argument of the logarithm is just the partition function associated with an energy level splitting of $E_\mathrm{max}$. If $E_\mathrm{max}\to\infty$, then this term turns to $1$, and provided we have perfect thermalisation such that if $\delta=0$, we recover the quasistatic Landauer cost ${k_\mathrm{B}}T \ln 2$ (as derived in Eq. 1).
Single-shot work cost {#app:singleshot}
=====================
Bounding the work cost fluctuations. {#app:fluxbound}
------------------------------------
In a single shot classical regime, we assume that at the end of each stage the system is in one of the two energy levels; and we can express each choice of energy level as a sequence of random variables $\{X_i\}_{i=1\ldots N}$. Noting that at each stage, by raising the splitting of the energy levels by $\mathcal{E}$, if the system is in the upper energy state at a particular stage, then this stage contributes a work cost of $\mathcal{E}$. With this in mind, it is useful to label the two energy levels as $0$ or $1$ such that the work contribution at each stage is given by $X_i\mathcal{E}$, and thus the [*actual work cost*]{} of a bit reset is given by the function acting on the random variables: $$\label{eq:single_shot_work_X}
W(X_1, \ldots, X_N) = \mathcal{E} \sum_{i=1}^N X_i.$$
It is possible to take the average of this function over some or even all of the random variables. For example, if we take the average over all $X_1\ldots X_N$ we arrive at: $$\begin{aligned}
{\langle W(X_1, \ldots, X_N)\rangle}_{X_1 \ldots X_N} \hspace{-5em} & & \nonumber \\
& = & \mathcal{E} \sum_{i=1}^N {\langle X_i\rangle}, \nonumber \\
& = & \mathcal{E} \sum_{i=1}^N P_2(i) = {\langle W\rangle},\end{aligned}$$ where ${\langle W\rangle}$ is the value we would typically call the [*average work cost*]{}- the average work cost of the procedure calculated before we know the outcome of any of the random variables. This is the value that we have calculated in the prior sections of this article.
There is, in fact, a series of $N$ intermediate stages between $W$ and ${\langle W\rangle}$, in which given knowledge of the first $n$ values of $X_i$ (that is, the exact cost of the first $n$ steps of the procedure), we make an estimate of what the final work cost will be. This series evolves as a random walk starting at the average value ${\langle W\rangle}$ and finishing at the actual value $W$. Thus if the first $n$ steps of the protocol are in energy levels $X_1=x_1$, $X_2=x_2$, etc, then we write the series $D(n)$ as: $$\label{eq:doob_work_function}
D(n) = {\langle W(x_1,\ldots x_n,X_{n+1}, \ldots, X_N)\rangle}_{X_{N-n}\ldots X_N}$$
$D(n)$ undergoes a special type of random walk known as a Doob martingale [@Doob40]. It is a martingale because at every step $n$, the expected value of the next step is the value of the current step: $$\label{eq:martingale}
{\langle D(n+1)\rangle} = D(n).$$ For Doob’s martingale, this is true by construction as $W(n)$ is defined to be the expectation value of $W$ over future steps.
There is a statistical result known as the Azuma inequality [@Azuma67] which bounds how far a martingale random walk is likely to deviate from its initial value. When specifically applied to a Doob martingale, this gives us the special case known as the McDiarmid inequality (see 6.10 in [@McDiarmid89]) which bounds how far the actual value ($D(N) = W$) deviates from the expectation value ($D(0) = {\langle W\rangle}$): $$\label{eq:McDiarmid}
P(|W - {\langle W\rangle}| \geq \omega) \hspace{0.75em} \leq \hspace{0.75em} 2\exp\left( \dfrac{-2\,\omega^2}{\sum_{i=1}^N {|c_n|}^2} \right)$$ where $c_n$ is the maximum amount our adjustment of the work estimate will change by knowing the outcome of the random variable $X_n$.
For pedagogical purposes we calculate this expression first for the quasistatic regime of perfect thermalisation (as discussed in the appendices of Egloff and co-authors \[13\]). In this regime, each $X_i$ is an independent random variable, with a probability distribution given by the thermal populations associated with that energy level (Eq. \[eq:thermal\_population\]). Switching any particular $X_i$ from $0$ to $1$ or vice-versa will therefore have at most an effect of $\mathcal{E}$ on $W$. Hence, $c_i = \mathcal{E}$ for all $i$, and so we arrive at Eq. 7 (repeated here for clarity): $$P(|W-{\langle W\rangle}| \geq \omega) \hspace{0.5em} \leq \hspace{0.5em} \exp\left( - \dfrac{2 \, \omega^2}{N \mathcal{E}^2} \right).$$
When we enter the finite time regime, in which thermalisation is only partially achieved, $X_i$ are no longer perfectly independent. If we treat $P_\mathrm{swap}$ as the probability over the entire period of thermalisation that we exchange our system with the thermal state, then $X_{i+1}$ will take the value of $X_{i}$ with probability $(1-P_\mathrm{swap})$, and only with probability $P_\mathrm{swap}$ will it be given by the random thermal distribution.
To calculate the impact of exchanging one stage $X_n$, we must explicitly evaluate the difference between $
D(n) = {\langle W(x_1,\ldots x_{n-1}, 0, X_{n+1}, \ldots, X_N)\rangle}_{X_{N-n}\ldots X_N}$ and ${\langle W(x_1,\ldots x_{n-1}, 1, X_{n+1}, \ldots, X_N)\rangle}_{X_{N-n}\ldots X_N}$. To evaluate this, it is necessary to consider the expectation work cost at every stage of the protocol between $n$ and the end ($N$), and how this changes depending on the value of $X_n$.
We recall the stochastic matrix for evolution between a state $n$ and $n+1$ can be written as $M(n)$, given in Eq. \[eq:swap\_matrix\]. The stochastic evolution from state $n$ to $n+k$ is thus given by the left-product of matrices $$\label{eq:many_swap_matrices}
\mathcal{M}(n\to n+k) = M(n+k-1) \ldots M(n+1) M(n).$$
The form of $M_\mathrm{swap}(n)$ allows us to simplify this product. Over $k$ steps are there are $k+1$ possible outcome states of the system corresponding swapping with one of the thermal states $\mathbf{P}^\mathrm{th}(n+j)$ (where $j = 0\ldots k-1$) or doing nothing at all. Because of the nature of a swap operation (in particular because $P_\mathrm{swap}$ is independent of the state of the system), only the final swap is important- all intermediate swaps will be ‘overwritten’. This means there is always a probability of $P_\mathrm{swap}$ that the system is in the state it last had some chance of swapping with (i.e. $\mathbf{P}^\mathrm{th}(n+k-1)$). Provided it has not swapped with this state, there is then a chance $P_\mathrm{swap}$ that the system has swapped into the state before it- giving an overall probability of swapping into this state of $(1-P_\mathrm{swap}) P_\mathrm{swap}$. Working backwards with this logic, we see that the probability of the system after $k$ steps being in the state associated with swapping after $j$ steps is $(1-P_\mathrm{swap})^{k-j}P_\mathrm{swap}$. Finally, we note that the only way for the system to have not changed at all is for it to have not swapped at any of the opportunities; and this has a probability of $(1-P_\mathrm{swap})^k$. With all of this in mind, we can now write out $\mathcal{M}(n\to n+k)$ in a simple linear form: $$\begin{aligned}
\mathcal{M}(n\to n+k) & = & (1-P_\mathrm{sw})^k {\mathbbm{1}}\nonumber \\
&& \hspace{-5.5em} + ~ P_\mathrm{sw} \sum_{j=0}^{k-1} (1-P_\mathrm{sw})^{k-j-1} M_\mathrm{th}(n+j),\end{aligned}$$ where $M_\mathrm{th}(i)$ (as defined in Eq. \[eq:thermal\_matrix\]) is the matrix that perfectly exchanges any state with the thermal state $\mathbf{P}^\mathrm{th}(i)$. If $P_\mathrm{swap}=1$, then $\mathcal{M}$ reduces to the thermalising matrix for the final stage $M_\mathrm{th}(n+k-1)$.
The expected energy cost of step $n+k$ given that the state starts in the lower energy level is then given by the top right component of $\mathcal{M}$ multiplied by $\mathcal{E}$, and the expected cost if the state starts in the upper energy level is given by the lower right component multiplied by $\mathcal{E}$. We can thus express the difference between these two values as $$\mathcal{E} \left( \begin{array}{cc} 0 & 1 \end{array}\right) \mathcal{M} (n \to n+k) \left(\begin{array}{c} 1 \\ -1 \end{array} \right).$$ Again, considering the special case $P_\mathrm{swap}=1$, we note that this value is zero for all $k\geq1$; the change in expected contribution from all future steps as a result of altering the current state is zero when the future steps are independent of the current state.
Finally, we write out the predicted difference in estimated final work cost, $c_n$, as the sum: $$\label{eq:explicit_cn}
c_n = \left( \begin{array}{cc} 0 & \mathcal{E} \end{array}\right) \left({\mathbbm{1}}+ \sum_{k=1}^N \mathcal{M} (n \to n+k) \right) \left(\begin{array}{c} 1 \\ -1 \end{array} \right).$$
In this most general form, the expression is difficult to evaluate analytically. However, noting that the components of $M_\mathrm{th}$ are the thermal populations at each stage, which are in the range $[0, 1]$, we can bound the sum: $$0 \leq \sum_j (1-P_\mathrm{sw})^j P^\mathrm{th}(j) \leq \sum_j (1-P_\mathrm{sw})^j \left(\begin{array}{cc}
1 & 1 \\
1 & 1
\end{array}\right)$$ Evaluating equation \[eq:explicit\_cn\] effectively involves subtracting the bottom right element of the matrix in the middle from the bottom left. And as a maximal value of $|c_n|$ would be the most deleterious for our bound (causing the widest spread of work values from the mean) we therefore take the upper bound on this sum for the bottom right and the lower bound for the bottom left.
This allows us to bound $|c_n|$ by: $$\label{eq:workstepdif}
|c_n| \leq {\large\mathcal{E}} \dfrac{1-(1-P_\mathrm{sw})^{N-n}}{P_\mathrm{sw}}.$$ This tells us that when we are far away from the end of the protocol, such that $n \ll N$, the effect of changing one stage has an effect that scales like $1/P_\mathrm{sw}$. Closer to the end of the protocol, the effect is diminished, as there are fewer chances to swap, which thus truncates the influence.
We consider the sum of terms $\sum_n |c_n|^2$, as required for the McDiarmid inequality (Eq. \[eq:McDiarmid\]): $$\sum_n |c_n|^2 \leq \dfrac{\mathcal{E}^2}{P_\mathrm{sw}^2} \sum_n \left(1-(1-P_\mathrm{sw})^{N-n}\right)^2.$$ The leading term of the sum is $N$, and so again we can upper bound this value: $$\sum_n |c_n|^2 \leq \dfrac{N \mathcal{E}^2}{P_\mathrm{sw}^2}.$$ By taking only the first term we slightly over-estimate the importance of a change at one step. This approximation encodes the assumption that any change has the full number of chances to influence the state. This means the bound we place on the deviation away from the average value of work is not as tight as it might otherwise be. We finally substitute this value into the McDiarmid inequality to arrive at the claim we made in Eq. 8: $$P(|W-{\langle W\rangle}| \geq \omega) \hspace{0.5em} \leq \hspace{0.5em} \exp\left( - \dfrac{2 \, \omega^2{P_\mathrm{sw}}^2}{N \mathcal{E}^2} \right).$$
![ \[fig:work\_distribution\] [**An example probabilistic work distribution.**]{} Work values are always below $W_\mathrm{max}^\epsilon$, except with probability of failure $\epsilon$ given by the area of the shaded region on the right. The sum of the areas of both the shaded regions on the left and right indicates the probability of failing to be within $\omega$ of ${\langle W\rangle}$ (as given by Eq. 8). ](workdist){width="45.00000%"}
Calculating $W_\mathrm{max}^\epsilon$. {#app:wemax}
--------------------------------------
We can express the above result in the language of single-shot statistics by calculating the maximum work, except with some probability of failure $\epsilon$, defined: $$P(W > W_\mathrm{max}^\epsilon) := \epsilon$$ or equivalently $W_\mathrm{max}^\epsilon$ is defined by the integral $$\int_{-\infty}^{W_\mathrm{max}^\epsilon} P(W) \mathrm{d}W = 1- \epsilon.$$
We can re-centre this definition around the expectation value of work, ${\langle W\rangle}$, such that: $$P\big(W - {\langle W\rangle} > W_\mathrm{max}^\epsilon - {\langle W\rangle}\big) = \epsilon$$ and in this form, we note that $\epsilon$ can be bounded by Eq. 8, with $\omega = W_\mathrm{max}^\epsilon - {\langle W\rangle}$ (see Figure \[fig:work\_distribution\]): $$\epsilon \leq \hspace{0.5em} 2\exp\left( - \dfrac{2 \, \left(W_\mathrm{max}^\epsilon - {\langle W\rangle}\right)^2{P_\mathrm{sw}}^2}{N \mathcal{E}^2} \right).$$ Re-arranging, we arrive at an upper bound on $W_\mathrm{max}^\epsilon$: $$W_\mathrm{max}^\epsilon \leq {\langle W\rangle} + \dfrac{\mathcal{E}}{P_\mathrm{sw}} \sqrt{\dfrac{N\ln (2/\epsilon) }{2}},$$ or explicitly in terms of thermalising time (using Eq. 2): $$W_\mathrm{max}^\epsilon \leq {\langle W\rangle} + \dfrac{\mathcal{E}}{1-(1-p)^t} \sqrt{\dfrac{N\ln (2/\epsilon) }{2}}.$$ We combine this bound with the influence of finite time on ${\langle W\rangle}$ (from Eq. 5 substituting $E_\mathrm{max} = N \mathcal{E}$) to get: $$\begin{aligned}
W_\mathrm{max}^\epsilon & \leq & {\langle W_\mathrm{quasi}\rangle}+ \dfrac{1}{2}(1-p)^t N \mathcal{E} \nonumber \\
&&+ \hspace{1em}\dfrac{1}{1-(1-p)^t} \sqrt{\dfrac{\ln (2/\epsilon) }{2N}}N\mathcal{E}. \nonumber\end{aligned}$$
Failure probability for resetting many qubits. {#app:manyqubits}
==============================================
We note that our calculated $W_\mathrm{max}^\epsilon$ has failure probability upper bounded by $\epsilon$, such that $P(W~>~W_\mathrm{max}^\epsilon) \leq \epsilon$. For $n$ random bits, the probability that every bit resets under this limit is given by $(1-\epsilon)^n$, such that we can bound: $$P(W_\mathrm{tot} > nW_\mathrm{tot}^{\epsilon}) \leq 1 - (1-\epsilon)^n \leq n\epsilon,$$ where the final inequality can be proved by induction.
In the worst case scenario, the failure of a single bit to reset under $W_\mathrm{max}^\epsilon$ causes the entire protocol to fail. We can thus write a bound on the work cost of resetting $n$ bits as $\mathcal{W}^{\epsilon'}_\mathrm{max} \leq nW_\mathrm{max}^{\epsilon}$ where $\epsilon' = 1 - (1-\epsilon)^n$, and as $\epsilon'\leq n\epsilon$ it follows that $\mathcal{W}^{n\epsilon} \leq nW^\epsilon_\mathrm{max}$. Hence, when resetting $H_\mathrm{max}$ bits: $$\mathcal{W}^{H_\mathrm{max}\epsilon} \leq H_\mathrm{max} W^\epsilon_\mathrm{max}.$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An orthogonal product basis of a composite Hilbert space is [*genuinely nonlocal*]{} if the basis states are locally indistinguishable across every bipartition. From an operational point of view such a basis corresponds to a separable measurement that cannot be implemented by local operations and classical communication (LOCC) unless all the parties come together in a single location. In this work we classify genuinely nonlocal product bases into different categories. Our classification is based on state elimination property of the set via orthogonality-preserving measurements when all the parties are spatially separated or different subsets of the parties come together. We then study local state discrimination protocols for several such bases with additional entangled resources shared among the parties. Apart from consuming less entanglement than teleportation based schemes our protocols indicate operational significance of the proposed classification and exhibit nontrivial use of genuine entanglement in local state discrimination problem.'
author:
- Sumit Rout
- 'Ananda G. Maity'
- Amit Mukherjee
- Saronath Halder
- Manik Banik
title: 'Genuinely Nonlocal Product Bases: Classification and Entanglement Assisted Discrimination'
---
Introduction
============
Superposition principle lies at the core of quantum mechanics which leads to several no-go results in quantum information theory, such as, no-cloning [@Wootters82] and no-deleting theorem [@Pati04]. It also gives rise to the concept of nonorthogonal states for which perfect discrimination is never possible. Origin of the quantum state discrimination problem dates back to early nineteen seventies with an initial attempt to formalize information processing with optical quantum devices [@Helstrom69; @Holevo73; @Yuen75]. Although sets of mutually orthogonal states can always be perfectly distinguished by some global measurements, the situation may change dramatically for a set of such multipartite quantum states if the spatially separated parties are allowed to perform only local operations assisted with classical communication (LOCC). In a seminal work Bennett [*et al.*]{} provided examples of mutually orthogonal product states that are indistinguishable under LOCC given one copy of the state [@Bennett99]. They coined the term ‘quantum nonlocality without entanglement’ [^1] for this phenomenon as the states allow local preparation (with some preshared strategy) but prohibit perfect local discrimination. Importantly, local indistinguishability turns out to be a crucial primitive for a number of distributed quantum protocols, namely, quantum data hiding [@Terhal01; @Eggeling02] and quantum secret sharing [@Markham08; @Rahaman15; @Wang17-0].
The result of Bennett [*et al.*]{} [@Bennett99] motivates overwhelming research interest on generic local state discrimination problems — the task of optimal discrimination of multiparty states, not necessarily product, by means of LOCC [@Bennett99-1; @Walgate00; @Virmani01; @Ghosh01; @Groisman01; @Walgate02; @DiVincenzo03; @Horodecki03; @Fan04; @Ghosh04; @Rinaldis04; @Nathanson05; @Watrous05; @Niset06; @Ye07; @Fan07; @Duan07; @Bandyopadhyay09; @Feng09; @Duan10; @Yu12; @Yang13; @Childs13; @Zhang14; @Yu15; @Zhang15; @Wang15; @Chen15; @Yang15; @Zhang16; @Xu16; @Zhang16-1; @Xu16-1; @Wang17; @Zhang17; @Xu17; @Wang17-1; @Zhang17-1; @Zhang17-2; @Zhang17-3; @Corke17; @Halder18]. Very recently, Halder [*et al*]{} have introduced a nontrivial generalization of the quantum nonlocality without entanglement phenomena [@Halder19]. They have provided examples of $3$-qutrit and $3$-ququad product bases that are not distinguishable even if (any) two of the three parties come together, [*i.e.*]{}, each of these product basis can be prepared locally but to distinguish them either all the three parties need to come together or entangled resources must be shared across all bipartitions. Quite naturally, one can define such a feature as [*genuine quantum nonlocality without entanglement*]{}. Existence of the product bases of Ref. [@Halder19] has important operational consequences. While in one hand they constitute a nontrivial primitive for multipartite information theoretic protocols, on the other hand they correspond to multipartite separable measurements that require entanglement resources across all bipartitions for implementing those measurements.
In this work, we classify genuinely nonlocal product bases (GNPBs) of multipartite quantum systems. The classification is based on whether such a basis is reducible, [*i.e*]{}, allows elimination of state(s) from the set under orthogonality-preserving measurements (OPMs) when all the parties are spatially separated or some subset of the parties are allowed to come together. We then show that this way of classification has interesting operational consequences. We provide example of multipartite GNPBs that allow local elimination of some states under OPM even when all the parties are separated and subsequently they require entanglement across one bipartite cut only for perfect discrimination of the states. In other words, local elimination makes it adequate to consume less entanglement for perfect discrimination of the states in those GNPBs. We then provide examples of tripartite GNPB which is locally irreducible when all three parties are in separate location but reducible if two of the parties are together. However, it requires entanglement across every bipartite cut for perfect discrimination. Such a GNPB is weaker than the GNPB of [@Halder19] as the later does not allow any elimination of states under nontrivial OPM even when any two parties come together. We then provide different entanglement assisted protocols for perfect discrimination of several GNPBs. To the best of our knowledge, the entanglement assisted discrimination of GNPBs that require entanglement across every bipartition for perfect local discrimination is reported in the present manuscript for the very first time. The studied protocols are resource efficient as they consume less entanglement in comparison to the teleportation based protocols. Interestingly, one of our protocols exhibits nontrivial and advantageous use of genuine entanglement in local state discrimination problem.
We arrange present manuscript in the following way: in Section \[sec2\] we briefly discuss the notations and some required preliminary concepts, Sec. \[sec3\] & \[sec4\] consist of the main contributions of the present work, in Sec. \[sec5\] we summarize the results with a discussion of some open problems.
Preliminaries and notations {#sec2}
===========================
Throughout the paper we will use standard notations and terminologies that are commonly used in quantum information theory. All the systems we consider are finite dimensional and thus, the associated Hilbert spaces are isomorphic to some complex Euclidean spaces $\mathbb{C}^d$, with $d\in\mathbb{N}$ denoting the dimension of the system. Composite quantum systems are associated with tensor product of the corresponding subsystems’ Hilbert spaces – an $n$-partite quantum system is associated with the Hilbert space $\bigotimes_{i=1}^n\mathbb{C}^{d_i}$, where $\mathbb{C}^{d_i}$ corresponds to the $i^{th}$ subsystem. For our purpose we start with recalling the following definition.
\[def1\] Nonlocal product bases (NPBs):- Consider an $n$-partite quantum system with Hilbert space $\bigotimes_{i=1}^n\mathbb{C}^{d_i}$. An orthogonal product basis (OPB), $\mathbb{B}_{nl}\equiv\left\lbrace \ket{\psi}_j=\bigotimes_{i=1}^n\ket{\alpha}^i_j~|~j=1,\cdots,\Pi_{i=1}^nd_i\right\rbrace$ is called nonlocal if the states in $\mathbb{B}_{nl}$ can not be perfectly distinguished by LOCC when all the parties are spatially separated.
Bipartite as well as multipartite examples of such bases were first constructed in [@Bennett99] for $(\mathbb{C}^3)^{\otimes 2}$ and $(\mathbb{C}^2)^{\otimes 3}$ Hilbert spaces. Though for the second example the states cannot be discriminated under LOCC when all of the three parties are spatially separated, but can be done perfectly when two of the parties come together. This observation leads to a stronger notion of nonlocality without entanglement.
\[def2\] Genuinely nonlocal product bases (GNPBs):- A multipartite OPB, $\mathbb{B}_{gnl}\equiv\left\lbrace \ket{\psi}_j=\bigotimes_{i=1}^n\ket{\alpha}^i_j~|~j=1,\cdots,\Pi_{i=1}^nd_i\right\rbrace \subset\bigotimes_{i=1}^n\mathbb{C}^{d_i}$ is called genuinely nonlocal if the states in $\mathbb{B}_{gnl}$ can not be perfectly distinguished by LOCC even if any $(n-1)$ parties are allowed to come together.
Ref. [@Halder19] provides examples of such product bases for $3$-qutrit and $3$-ququad quantum systems. At this point, let us discuss a bit about local protocols. These are generally multi-round protocols. For multipartite case, depending on the scenarios whether all parties are separated or some group of parties are allowed to come together, in a particular round each party individually and/or some as a group perform(s) local quantum operations and communicate(s) the classical outcomes to other parties and/or other groups. Depending on these communications other parties (and/or groups) further choose their actions and the protocol goes on. Being multi round, it is in general difficult to mathematically characterize the set of LOCC operations. Some interesting topological behaviors of the LOCC set have been identified in [@Chitambar14]. While discriminating a set of mutually orthogonal multipartite product states by such a LOCC protocol, in a given round either the given state must be identified or some of the states must be eliminated. If the later happens then for perfect discrimination the remaining post measurement states should be mutually orthogonal so that the protocol can be further carried on. This leads to the following definition.
\[def3\] Nontrivial- orthogonality preserving measurement (N-OPM):- A measurement performed to distinguish a set of mutually orthogonal quantum states is called orthogonality preserving measurement (OPM) if after the measurement the states remain mutually orthogonal. Furthermore, such a measurement is called nontrivial if all the measurement effects constituting the OPM are not proportional to identity operator, otherwise it is trivial.
Definition \[def3\] subsequently leads to the concept of ’locally irreducible set’ – a set of mutually orthogonal multipartite quantum states from which it is not possible to eliminate one or more states by orthogonality preserving local measurements. Although local irreducibility sufficiently assures locally indistinguishable but the former is not a necessary requirement for the later one. It turns out that the examples of [@Halder19] exhibit a ‘strong nonlocal’ behavior as those product bases are not only GNPBs but they are indeed locally irreducible even if any two parties come together.
While additional entangled states are supplied as resource among the parties along with their operational power LOCC then it may be possible to perfectly distinguish a GNPB. An immediate such protocol follows from quantum teleportation [@Bennett93]. If the involved parties share sufficient entanglement so that they can teleport their respective subsystems to one of the parties then she/he can perfectly discriminate the states by performing suitable measurement. Since entanglement is costly resource under the operational paradigm of LOCC, therefore any protocol consuming less entanglement is always desirable. First instance of such protocols for a class of locally indistinguishable product states was proposed by Cohen [@Cohen08]. For instance, Bennett’s $2$-qutrit NPB can be perfectly distinguished by LOCC with additional $1$-ebit entanglement whereas the teleportation based protocol requires a $2$-qutrit maximally entangled state, [*i.e.*]{}, $\log3$-ebit. Here, we use the unit ebit, so, logarithm is taken with respect to base $2$. Cohen’s result motivates further research in identifying efficient use of entanglement in local state discriminating problem [@Bandyopadhyay09-1; @Bandyopadhyay10; @Duan14; @Bandyopadhyay14; @Bandyopadhyay16; @Zhang16-2; @Bandyopadhyay18; @Zhang18; @Halder18-1; @Li19; @Zhang18; @Halder18-1; @Li19].
In this work we also study efficient state discrimination protocols for several GNPBs. Note that, in multipartite scenario different types of entangled resource may be supplied as there exist different inequivalent types of entanglement. For instance, in tripartite scenario different pairs of parties can be supplied with $2$-qubit maximally entangled state, [*i.e.*]{}, Einstein-Podolsky-Rosen (EPR) state, $\ket{\phi^+}:=\frac{1}{\sqrt{2}}(\ket{00}+\ket{11})\in(\mathbb{C}^2)^{\otimes2}$ or they can be supplied with $3$-qubit Greenberger–Horne–Zeilinger (GHZ) state, $\ket{G}:=\frac{1}{\sqrt{2}}(\ket{000}+\ket{111})\in(\mathbb{C}^2)^{\otimes3}$. Here we consider two different configurations of entanglement resources.
[**Config. (1):**]{} $\left\lbrace\left(p,\ket{\phi^+}_{\mathcal{AB}}\right);\left(q,\ket{\phi^+}_{\mathcal{BC}}\right);\left(r,\ket{\phi^+}_{\mathcal{CA}}\right)\right\rbrace$ with $p,q,$ and $r$ taking nonnegative values.
![(Color on-line) Spatial configuration of the resource state $\left\lbrace\left(p,\ket{\phi^+}_{\mathcal{AB}}\right);\left(q,\ket{\phi^+}_{\mathcal{BC}}\right);\left(r,\ket{\phi^+}_{\mathcal{CA}}\right)\right\rbrace$. The indices $a_1,a_2\in\mathcal{A}$, $b_1,b_2\in\mathcal{B}$, and $c_1,c_2\in\mathcal{C}$.[]{data-label="Fig:3epr"}](figa){width="35.00000%"}
It denotes that on average $p$ amount of $2$-qubit maximally entangled state is consumed between Alice & Bob, and similarly in other two pairs $q$ and $r$ amounts of EPR state are consumed while discriminating a GNPB. Whenever discriminating a GNPB of $(\mathbb{C}^d)^{\otimes3}$, then a successful discrimination protocol with this resource configuration will be efficient than the corresponding teleportation based protocol if $(p+q+r)<2\log d$.
[**Config. (2):**]{} $\left\lbrace\left(p,\ket{G}_{\mathcal{ABC}}\right);\left(q,\ket{\phi^+}_{\star}\right)\right\rbrace$ with $p,q$ taking nonnegative values and $\star$ be one of the pairs from $\{\mathcal{AB},\mathcal{BC},\mathcal{CA}\}$.
![(Color on-line) Spatial configuration of the resource state $\left\lbrace\left(p,\ket{G}_{\mathcal{ABC}}\right);\left(q,\ket{\phi^+}_{\mathcal{BC}}\right)\right\rbrace$. []{data-label="Fig:GHZ"}](figb){width="35.00000%"}
It denotes that while discriminating a GNPB, $p$ amount of $3$-qubit GHZ state is consumed in addition with $q$ amount of EPR state shared between one of the three pairs. Note that, to distribute a $3$-qubit GHZ state among Alice, Bob, and Charlie $2$ copies of EPR state shared among two pairs (say) Alice & Bob and Alice & Charlie are required – Alice prepares a GHZ state at her lab and teleports two subsystems to Bob and Charlie respectively. However, the process might be irreversible as there is no known local protocol via which it is possible to get back two copies of two-qubit Bell states from a GHZ state [@Bennett00].
Classification of GNPB(s) {#sec3}
=========================
The states in an $n$-partite GNPB can not be perfectly distinguished under LOCC even if any $(n-1)$ parties are allowed to come together. However, it may be possible that while some parties come together, they can eliminate some states from the set under local measurement which keeps the post measurement states orthogonal. Based on how many parties are required to come together for such elimination we can classify the GNPBs into different types. In the following we will discuss this classification with explicit examples. Though the classification can be generalized for arbitrary number of parties, we will mainly restrict our study for tripartite Hilbert spaces.
GNPB: Type-I
------------
Such a GNPB is locally reducible even when all the parties are separated, [*i.e.*]{}, some subset of states can be eliminated under nontrivial local OPM.
[**Example:**]{} Consider the quantum system with Hilbert space $(\mathbb{C}^4)^{\otimes3}$ shared among Alice, Bob, and Charlie. Computational bases for $\mathbb{C}^4$ will be denoted as $\{\ket{i}\}_{i=0}^3$. We will use the short hand notation $\ket{\alpha}_A\ket{\beta}_B\ket{\gamma}_C$ for the state $\ket{\alpha}_A\otimes\ket{\beta}_B\otimes\ket{\gamma}_C$ and will avoid the party index where possible. To construct the required GNPB, first consider the following set of states: $$\label{eq1}
\mathcal{S}\equiv\left\lbrace \ket{3}_A\ket{\beta}_{BC},~\ket{\beta}_{AB}\ket{3}_{C}\right\rbrace;$$ where $\ket{\beta}$’s are the states belonging in the $2$-qutrit NPB of Ref. [@Bennett99], [*i.e.*]{}, $\ket{\beta}\in\mathcal{B}\equiv\left\lbrace{\ket{0}\ket{\eta_\pm}},~{\ket{\eta_\pm}\ket{2}},~{\ket{2}\ket{\xi_\pm}},~{\ket{\xi_\pm}\ket{0}},~{\ket{1}\ket{1}}\right\rbrace$; where $\ket{\eta_\pm}:=(\ket{0}\pm\ket{1})/\sqrt{2}$ and $\ket{\xi_\pm}:=(\ket{1}\pm\ket{2})/\sqrt{2}$. The definitions of $\ket{\eta_\pm}$ and $\ket{\xi_\pm}$ are maintained same throughout the manuscript. Also consider the following product states: $$\begin{aligned}
\label{c}
\rotatebox[origin=c]{0}{$\mathcal{R}\equiv$}
\left\{\!\begin{aligned}
|3\rangle|0\rangle|3\rangle,~|3\rangle|1\rangle|3\rangle,~
|3\rangle|2\rangle|3\rangle,~|3\rangle|3\rangle|0\rangle,\\
|3\rangle|3\rangle|1\rangle,~|3\rangle|3\rangle|2\rangle,~
|3\rangle|3\rangle|3\rangle,~|2\rangle|0\rangle|0\rangle,\\
|2\rangle|0\rangle|1\rangle,~|2\rangle|0\rangle|2\rangle,~
|2\rangle|1\rangle|0\rangle,~|2\rangle|1\rangle|1\rangle,\\
|2\rangle|1\rangle|2\rangle,~|2\rangle|2\rangle|0\rangle,~
|2\rangle|2\rangle|1\rangle,~|2\rangle|2\rangle|2\rangle,\\
|2\rangle|3\rangle|0\rangle,~|2\rangle|3\rangle|1\rangle,~
|2\rangle|3\rangle|2\rangle,~|2\rangle|3\rangle|3\rangle,\\
|0\rangle|0\rangle|0\rangle,~|0\rangle|0\rangle|1\rangle,~
|0\rangle|0\rangle|2\rangle,~|0\rangle|1\rangle|0\rangle,\\
|0\rangle|1\rangle|1\rangle,~|0\rangle|1\rangle|2\rangle,~
|0\rangle|2\rangle|0\rangle,~|0\rangle|2\rangle|1\rangle,\\
|0\rangle|2\rangle|2\rangle,~|0\rangle|3\rangle|0\rangle,~
|0\rangle|3\rangle|1\rangle,~|0\rangle|3\rangle|2\rangle,\\
|0\rangle|3\rangle|3\rangle,~|1\rangle|0\rangle|0\rangle,~
|1\rangle|0\rangle|1\rangle,~|1\rangle|0\rangle|2\rangle,\\
|1\rangle|1\rangle|0\rangle,~|1\rangle|1\rangle|1\rangle,~
|1\rangle|1\rangle|2\rangle,~|1\rangle|2\rangle|0\rangle,\\
|1\rangle|2\rangle|1\rangle,~|1\rangle|2\rangle|2\rangle,~
|1\rangle|3\rangle|0\rangle,~|1\rangle|3\rangle|1\rangle,\\
|1\rangle|3\rangle|2\rangle,~|1\rangle|3\rangle|3\rangle.~~~~~~~~~~~~~~~~
\end{aligned}\right\}, \end{aligned}$$ These $46$ states in $\mathcal{R}$ along with the states in $\mathcal{S}$ form a OPB in $(\mathbb{C}^4)^{\otimes 3}$. Since the set $\mathcal{B}$ is a NPB in $(\mathbb{C}^3)^{\otimes2}$, therefore the set of tripartite states $\{\ket{3}_A\ket{\beta}_{BC}~|~\ket{\beta}\in\mathcal{B}\}$ are locally indistinguishable across $B|AC$ cut as well as $C|AB$ cut. Similarly the set of states $\{\ket{\beta}_{AB}\ket{3}_C~|~\ket{\beta}\in\mathcal{B}\}$ are locally indistinguishable across $A|BC$ and $B|AC$ cuts. As the considered OPB contains both theses subsets of states, thus we have the following proposition.
\[prop1\] The set of orthogonal states $\mathbb{B}_I(4,3)\equiv\mathcal{S}\cup\mathcal{R}$ is a GNPB in $(\mathbb{C}^4)^{\otimes3}$. However the set is locally reducible even when all the parties are separated.
Construction of such GNPBs are straightforward for arbitrary number of parties. At this point it is important to note that the local indistinguishability arises due to the [*twisted*]{} states $\ket{\beta}\in\mathcal{B}$, which are obtained from linear superposition of computational states. In other words, quantum superposition principle plays key role for manifestation of ‘quantum nonlocality without entanglement’ phenomenon. While discriminating the states in $\mathbb{B}_I(4,3)$, any of the parties can perform nontrivial OPM to eliminate certain states even when all of them are separated. For instance, let Alice performs local measurement $M\equiv\{\ket{3}\bra{3},~\mathbb{I}_4-\ket{3}\bra{3}\}$. If the measurement result corresponds to the projector $\ket{3}\bra{3}$, then the given state must be one of the following states: $$\begin{aligned}
\label{c12}
\rotatebox[origin=c]{0}{$\ket{3}\Rightarrow$}
\left\{\!\begin{aligned}
\ket{3}\ket{\beta},~|3\rangle|0\rangle|3\rangle,~|3\rangle|1\rangle|3\rangle,~
|3\rangle|2\rangle|3\rangle,\\
|3\rangle|3\rangle|0\rangle,~
|3\rangle|3\rangle|1\rangle,~|3\rangle|3\rangle|2\rangle,~|3\rangle|3\rangle|3\rangle
\end{aligned}\right\}. \end{aligned}$$ Otherwise it is one of the remaining states. Since the considered measurement is an OPM, so after this step entanglement assisted discrimination protocol can be carried on. The outcome provides nontrivial information in which cut they need to share bipartite entanglement. If the outcome corresponds to the projector $\ket{3}_A\bra{3}$ then entanglement is required to share between Bob & Charlie, otherwise it should be shared between Alice & Bob. In other words, local elimination makes it possible to consume entanglement across one bipartite cut only for perfect discrimination of the states.
GNPB: Type-II
-------------
Such a GNPB is locally irreducible when all the parties are separated, [*i.e.*]{}, no local elimination is possible preserving orthogonality among the post measurement states.
[**Example:**]{} From the GNPB $\mathbb{B}_I(4,3)$ every party can locally eliminate some states by performing OPM that discriminate the subspace spanned by $\ket{3}$ vs subspace spanned by $\{\ket{0},\ket{1},\ket{2}\}$. One will obtain a GNPB of Type-II if this local elimination can be stopped. For this purpose, take the states $\{|0\rangle|3\rangle|2\rangle,~|0\rangle|3\rangle|3\rangle\}\subset\mathbb{B}_I(4,3)$. Consider now a new OPB that contains the locally twisted product states $\{|0\rangle|3\rangle|\chi_\pm\rangle\}$ instead of the states $\{|0\rangle|3\rangle|2\rangle,~|0\rangle|3\rangle|3\rangle\}$, where $\ket{\chi_\pm}:=(\ket{2}\pm\ket{3})/\sqrt{2}$. As a consequence Charlie is no more able to discriminate between the subspaces spanned by $\ket{3}$ and $\{\ket{0},\ket{1},\ket{2}\}$ and thus, he cannot eliminate any state via OPM. Similarly, the twisted product states $\{|2\rangle|\chi_\pm\rangle|2\rangle\}$ stop Bob and $\{|\chi_\pm\rangle|3\rangle|1\rangle\}$ stop Alice from eliminating any state via nontrivial OPM. Thus we have the following proposition.
\[prop2\] The set of states, $$\begin{aligned}
\label{t2}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II}(4,3):=$}
\left\{\!\begin{aligned}
\mathbb{B}_{I}(4,3)\setminus\{{\ket{0}\ket{3}\ket{2}},{\ket{0}\ket{3}\ket{3}},\\{\ket{2}\ket{2}\ket{2}},{\ket{2}\ket{3}\ket{2}},{\ket{2}\ket{3}\ket{1}},\\{\ket{3}\ket{3}\ket{1}}\}\\
\cup\{{\ket{0}\ket{3}\ket{\chi_\pm}},{\ket{2}\ket{\chi_\pm}\ket{2}},{\ket{\chi_\pm}\ket{3}\ket{1}}\}
\end{aligned}\right\}, \end{aligned}$$ is a GNPB in $(\mathbb{C}^4)^{\otimes3}$. Furthermore the set is locally irreducible when all the parties are specially separated.
Like Type-I basis, in this case also it is also possible to generalize the construction for arbitrary number of parties. Here we find that the set of multipartite nonlocal product states constructed in Ref. [@Zhang17] also posses the similar feature of our Type-II basis. Clearly, the GNPB $\mathbb{B}_{II}(4,3)$ requires entanglement across every bipartition for perfect discrimination. In other words, if two of the parties come together then also the nonlocality persists. Here comes two important observations: (i) local twist plays an important role in the construction of a new class of GNPBs – it can increase the entanglement cost of distinguishing a product basis by LOCC; (ii) the examples $\mathbb{B}_{I}(4,3)$ and $\mathbb{B}_{II}(4,3)$ also exhibit operational implication of ‘local elimination via nontrivial OPM’ as perfect discrimination of the first set requires entangled resource only in one cut, the later one demands entanglement in more than one cut.
So far we have constructed GNPBs for Hilbert spaces with local subsystem dimension $4$. Naturally the question arises regarding such constructions in lower dimensional cases. Remember that all product bases in $\mathbb{C}^2\otimes\mathbb{C}^d$ are locally distinguishable [@DiVincenzo03]. Therefore for a GNPB to exist the minimum dimension is $\mathbb{C}^3\otimes\mathbb{C}^3\otimes\mathbb{C}^3$. However, the technique used above or the technique used in [@Zhang17] are not applicable to construct GNPBs in the minimum dimension. In the following we provide an example of GNPB in $(\mathbb{C}^3)^{\otimes3}$.
\[prop3\] The set of states, $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II}(3,3)\equiv$}
\left\{\!\begin{aligned}
{\ket{0}\ket{\eta_\pm}\ket{\xi_\pm}},~{\ket{\eta_\pm}\ket{2}\ket{\xi_\pm}},~\\
{\ket{2}\ket{\xi_\pm}\ket{\eta_\pm}},~{\ket{\eta_\pm}\ket{\xi_\pm}\ket{0}},~\\
{\ket{\xi_\pm}\ket{0}\ket{\eta_\pm}},~{\ket{\xi_\pm}\ket{\eta_\pm}\ket{2}},~\\
{\ket{k}\ket{k}\ket{k}}~~|~~ k\in\{0,1,2\}~~~~
\end{aligned}\right\},
\end{aligned}$$ is a GNPB of Type-II in $(\mathbb{C}^3)^{\otimes3}$.
When the three parties are spatially separated, it is not possible to eliminate any state by OPM from the above set and hence the set is locally indistinguishable. This can be easily proved by the technique described in [@Halder19; @Halder18-1]. We are yet to prove local indistinguishability of the above set across every bipartitions. For that, first note that set $\mathcal{B}$ is present between any two pair in the above construction. For instance consider the subset of states $$\begin{aligned}
\left\{\!\begin{aligned}
{\ket{\xi_\pm}\ket{0}\ket{\eta_\pm}},~{\ket{\xi_\pm}\ket{\eta_\pm}\ket{2}},~\\
{\ket{\eta_\pm}\ket{2}\ket{\xi_\pm}},~{\ket{\eta_\pm}\ket{\xi_\pm}\ket{0}},~\\
{\ket{1}\ket{1}\ket{1}}~~~~~~~~~~~~~~
\end{aligned}\right\}. \end{aligned}$$ Presence of $\mathcal{B}$ between Bob and Charlie is evident here. Furthermore, Alice’s states $\ket{\eta_\pm},~\ket{\xi_\pm},$ and $\ket{1}$ are tagged with. These tagged states are not all mutually orthogonal due to the presence of local twist. Moreover, the twisted states saturate the local dimension of Alice, [*i.e.*]{}, $\ket{\eta_\pm}$ covers the subspace spanned by $\ket{0}$ & $\ket{1}$ whereas $\ket{\xi_\pm}$ covers the subspace spanned by $\ket{1}$ & $\ket{2}$. As a result even if Alice comes together with either Bob or Charlie, it is not possible to perfectly distinguish this set of states. Similar argument holds in case of other bipartitions and consequently $\mathbb{B}_{II}(3,3)$ turns out to be a GNPB of Type-II. Note that this particular basis appears in a recent work for a different purpose [@Agrawal19]. There the aim was to construct a genuinely entangled subspace such that all density matrices supported on it are genuinely entangled. Entanglement assisted discrimination protocol of this basis is described in the next section.
As already discussed, from a Type-II GNPB no state can be eliminated under OPM when all the parties are separated. However when two party come together the power of state elimination under OPM may increase. This leads us to the further classification of the Type-II GNPBs.
[**Type-II(a):**]{} Such a GNPB is not locally reducible when all the parties are in separate location. But, when two of the parties come together then it is possible to eliminate some states through nontrivial OPM. A careful observation reveals that the GNPB in Proposition \[prop2\] is an example of such kind (evidently this follows from Remark \[remark2\] in Appendix \[appen-prop8\]).
![The set $\mathcal{GNPB}_{II(b)}$ of all Type-II(b) GNPBs is a proper subset of the set $\mathcal{GNPB}_{II}$ of all Type-II GNPBs which is again a proper subset of the set $\mathcal{GNPB}$ of all GNPBs, [*i.e.*]{}, $\mathcal{GNPB}_{II(b)}\subset \mathcal{GNPB}_{II}\subset \mathcal{GNPB}$.[]{data-label="subset"}](figc "fig:"){width="35.00000%"} \[Fig:subset\]
[**Type-II(b):**]{} Such a GNPB is locally irreducible even if any two of parties come together. Examples of such GNPBs are constructed in [@Halder19] for Hilbert spaces $(\mathbb{C}^3)^{\otimes3}$ and $(\mathbb{C}^4)^{\otimes3}$. Here we redraft the $3$-qutrit example.
\[prop4\] (Halder et al. [@Halder19]) The set of states, $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II(b)}(3,3)\equiv$}
\left\{\!\begin{aligned}
{\ket{0}\ket{1}\ket{\eta_\pm}},~{\ket{1}\ket{\eta_\pm}\ket{0}},~{\ket{\eta_\pm}\ket{0}\ket{1}},\\
{\ket{0}\ket{2}\ket{\kappa_\pm}},~{\ket{2}\ket{\kappa_\pm}\ket{0}},~{\ket{\kappa_\pm}\ket{0}\ket{2}},\\
{\ket{1}\ket{2}\ket{\eta_\pm}},~{\ket{2}\ket{\eta_\pm}\ket{1}},~{\ket{\eta_\pm}\ket{1}\ket{2}},\\
{\ket{2}\ket{1}\ket{\kappa_\pm}},~{\ket{1}\ket{\kappa_\pm}\ket{2}},~{\ket{\kappa_\pm}\ket{2}\ket{1}},\\
{\ket{k}\ket{k}\ket{k}}~~|~~k\in\{0,1,2\}~~~~
\end{aligned}\right\}, \end{aligned}$$ in a GNPB of Type-II(b) in $(\mathbb{C}^3)^{\otimes3}$; $\ket{\kappa_\pm}:=(\ket{0}\pm\ket{2})/\sqrt{2}$.
Clearly Type-II(b) is the strongest form of GNPB from the perspective of local elimination. The above classification thus introduces a hierarchical relation as depicted in Fig. \[subset\].
Entanglement assisted discrimination {#sec4}
====================================
In this section we study entanglement assisted discrimination protocols for the GNPBs discussed earlier. First we consider the $3$-qutrit GNPBs. Note that, in $(\mathbb{C}^3)^{\otimes3}$ two pairs of $2$-qutrit maximally entangled states, [*i.e.*]{}, $2\log3$ ebits, distributed between Alice & Bob and Alice & Charlie always lead to perfect discrimination for any genuinely nonlocal basis. Therefore any protocol that consumes less than $2\log3$ ebits is nontrivial and resource efficient. Following proposition constitutes such a nontrivial protocol.
\[prop5\] The entanglement resource $\left\lbrace\left(1,\ket{\phi^+(3)}_{\mathcal{AB}}\right);\left(0,\ket{\phi^+}_{\mathcal{BC}}\right);\left(1,\ket{\phi^+}_{\mathcal{CA}}\right)\right\rbrace$ is sufficient for local discrimination of the GNPBs $\mathbb{B}_{II}(3,3)$ and $\mathbb{B}_{II(b)}(3,3)$, where $\ket{\phi^+(3)}:=(\ket{00}+\ket{11}+\ket{22})/\sqrt{3}\in\mathbb{C}^3\otimes\mathbb{C}^3$.
Using $2$-qutrit maximally entangled state $(\ket{00}+\ket{11}+\ket{22})/\sqrt{3}$ Bob first teleports his subsystem to Alice. Entanglement consumed at this step amounts to $\log3$-ebits. After that $1$-ebit entanglement shared between Alice and Charlie suffices for perfect discrimination of these GNPBs (see Appendix \[appen-prop5\] for detailed protocol). Therefore, in total, $(\log3+1)$-ebits entanglement are consumed in this protocol which is strictly less than the amount consumed in the protocol using teleportation in both arms. However, in this protocol teleportation scheme is used in one arm. We now show that even more efficient protocols are possible to discriminate these GNPBs.
\[prop6\] The entanglement resource $\left\lbrace\left(1,\ket{\phi^+}_{\mathcal{AB}}\right);\left(0,\ket{\phi^+}_{\mathcal{BC}}\right);\left(1,\ket{\phi^+}_{\mathcal{CA}}\right)\right\rbrace$ sufficiently discriminates the GNPB $\mathbb{B}_{II}(3,3)$ when all the parties are separated.
See Appendix \[appen-prop6\] for the protocol. Clearly the entanglement consumed in this protocol is strictly less than $(1+\log3)$-ebits. The resource state used in the protocol, [*i.e.*]{}, the state $\ket{\phi^+}_{\mathcal{AB}}\otimes\ket{\phi^+}_{\mathcal{CA}}$, lives in the Hilbert space $\mathbb{C}^4\otimes\mathbb{C}^2\otimes\mathbb{C}^2$. Naturally the question arises weather a lower dimensional resource from the Hilbert space $\mathbb{C}^2\otimes\mathbb{C}^2\otimes\mathbb{C}^2$ will suffice for perfect discrimination. At this point we observe that a similar protocol like Proposition \[prop6\] that uses a $3$-qubit GHZ state $\ket{G}$ or a $3$-qubit W-state $\ket{W}:=(\ket{001}+\ket{010}+\ket{100})/\sqrt{3}$ does not lead to perfect discrimination of the basis $\mathbb{B}_{II}(3,3)$.
Here we want to point out some important observations. For discriminating the NPB of $(\mathbb{C}^3)^{\otimes2}$, Cohen pointed out that in his protocol $2$-qubit maximally entangled state is the necessary resource. If instead of the $2$-qubit maximally entangled resource a partially entangled state $\lambda_0\ket{00}+\lambda_1\ket{11}$, with $\lambda_0\neq\lambda_1$ is provided as resource then at some stage the protocol leads to non orthogonal states and consequently the protocol does not succeed perfectly. Furthermore, he also gave an impression that for any successful protocol $2$-qubit maximally entangled state may be the necessary resource. To the best of our knowledge, however, the assertion is yet to be proven. If it indeed turns out to be the case then the $3$-qubit W-state can not be a sufficient resource for perfect discrimination of the basis $\mathbb{B}_{II}(3,3)$. On the other hand, if there exist no local protocol that simultaneously generates two EPR pairs shared between Alice & Bob and Alice & Charlie respectively from a $3$-qubit GHZ state, then it can also not be the sufficient resource for perfect discrimination of $\mathbb{B}_{II}(3,3)$. Therefore, at this point the question remains open whether the resource mentioned in Proposition \[prop6\] is indeed the necessary resource for perfect discrimination of the set $\mathbb{B}_{II}(3,3)$.
Using the same resource as of Proposition \[prop6\] we then proceed to discriminate the set $\mathbb{B}_{II(b)}(3,3)$ following analogous protocol. We find that with this amount of resource though some state can be eliminated, but after a certain stage the protocol can not be further extended up to perfect discrimination. However, if further entanglement resource is provided then a perfect discrimination protocol is possible as stated in the following proposition (detailed protocol provided in Appendix \[appen-prop7\]).
\[prop7\] The entanglement resource $\left\lbrace\left(1,\ket{\phi^+}_{\mathcal{AB}}\right);\left(\frac{8}{27},\ket{\phi^+}_{\star}\right);\left(1,\ket{\phi^+}_{\mathcal{CA}}\right)\right\rbrace$ is sufficient for perfect local discrimination of the GNPB $\mathbb{B}_{II(b)}(3,3)$, where $\star$ be the one of the pairs from $\{\mathcal{AB},\mathcal{BC},\mathcal{CA}\}$.
The average entanglement used in this protocol is therefore $(1+1+\frac{8}{27})\cong2.296$-ebits which is less than $(1+\log3)\cong2.585$-ebits consumed in Proposition \[prop5\]. Note that the resource state lives either in the Hilbert space $\mathbb{C}^8\otimes\mathbb{C}^4\otimes\mathbb{C}^2$ (if the addition entanglement in Proposition \[prop7\] is shared between Alice & Bob) or in the Hilbert space $\mathbb{C}^4\otimes\mathbb{C}^4\otimes\mathbb{C}^4$ (if the addition entanglement is shared between Bob & Charlie). At this point the question remains open whether a lower dimensional resource state from $\mathbb{C}^4\otimes\mathbb{C}^2\otimes\mathbb{C}^2$ will result in a successful discrimination protocol for $\mathbb{B}_{II(b)}(3,3)$ or the resource in Proposition \[prop7\] is necessary.
Let us now consider entanglement assisted discrimination protocol for the GNPBs in $(\mathbb{C}^4)^{\otimes3}$. As already discussed, for the GNPB $\mathbb{B}_I(4,3)$ in Proposition \[prop1\] entanglement resource only in one cut is sufficient for perfect discrimination. Of-course in which cut the entangled resource needs to be shared that is determined after the first elimination step under OPM. Furthermore in this case, since the genuine indistinguishability arises due to the presence of $(\mathbb{C}^3)^{\otimes2}$ NPB between Alice & Bob ([*i.e.*]{}, the set of states $\left\lbrace \ket{\beta}_{AB}\ket{3}_{C}\right\rbrace$) and between Bob & Charlie ([*i.e.*]{} the set of states$\left\lbrace \ket{3}_A\ket{\beta}_{BC}\right\rbrace$), thus Cohen’s protocol [@Cohen08] assures that a $2$-qubit maximally entangled state shared between $AB$ or $BC$ (decided accordingly after the first elimination step) is sufficient for prefect discrimination even though the local dimension for each party is four.
Consider now the GNPB $\mathbb{B}_{II}(4,3)$ in Proposition \[prop2\]. Since this one is a GNPB of Type-II, therefore no state can be eliminated under OPM while all the parties are spatially separated and consequently entangled resource across every bipartition is required in this case. Interestingly, here we find that if the three parties share a $3$-qubit GHZ state then they can start the discrimination protocol. However, as stated in the following proposition the perfect discrimination protocol we obtain requires additional EPR pair along with the GHZ resource.
\[prop8\] The entanglement resource $\left\lbrace\left(1,\ket{G}_{\mathcal{ABC}}\right);\left(\frac{1}{8},\ket{\phi^+}_{\star}\right)\right\rbrace$ is sufficient for perfect local discrimination of the GNPB $\mathbb{B}_{II}(4,3)$ in Proposition \[prop2\], where $\star$ be the one of the pairs from $\{\mathcal{AB},\mathcal{BC},\mathcal{CA}\}$.
See Appendix \[appen-prop8\] for detailed protocol. The above protocol exhibits nontrivial use of multipartite entanglement in local state discrimination protocol. Moreover, we observe that instead of GHZ state if Alice and Bob start the protocol with sharing a EPR state then for perfect discrimination further $\frac{11}{16}$-ebits is required to be shared between Bob and Charlie (see the Remark \[remark2\] in Appendix \[appen-prop8\]). This indicates advantage of genuine entangled resource over its bipartite counterpart in state discrimination problem. However, a conclusive proof of this assertion requires establishing the necessary requirement of entangled resources in different such protocols which we leave here as an open question for future research.
Summary and Open problems {#sec5}
=========================
The phenomena ‘strong quantum nonlocality without entanglement’ introduced in Ref. [@Halder19] motivates us to look for other techniques to construct new GNPBs. As a consequence, in this work we have classified GNPBs into different categories. Via this classification we have addressed an important question regarding the requirement of multipartite entangled resource state for perfect discrimination of a GNPB. Interestingly, we have found that elimination of certain states from the original set by performing orthogonality-preserving measurements may help to reduce the entanglement consumption for perfect discrimination. We have also presented entanglement assisted local discrimination protocols of several GNPBs. These protocols are resource efficient as they consume less entanglement than a teleportation based protocol. We have also addressed an open problem raised in Ref. [@Halder19]. The authors there left open the possibility of existence of a cheaper resource than that of a teleportation based scheme for perfect discrimination of a strong nonlocal basis. One of our protocols provides affirmative answer to this question. Moreover, we have discrimination protocols for GNPBs with different types and configurations of entangled resources. Interestingly, we find strong indication of genuine entanglement advantage over bipartite entanglement for discrimination of some GNPBs.
Our study also raises few important questions. First of all, the question of optimality of the entangled resources used in our discrimination protocols remains open. Clearly, this study will shed light on the optimal resource requirement for the implementation of the separable measurement corresponding to these GNPBs. Furthermore, it is intriguing to study whether the classification of GNPBs induces a hierarchy among the corresponding separable measurements.
SR acknowledges the support through the VASP programme of S. N. Bose National Center for Basic Sciences. AM acknowledges his visit at S. N. Bose National Center for Basic Sciences. MB acknowledges support through an INSPIRE-faculty position at S. N. Bose National Center for Basic Sciences by the Department of Science and Technology, Government of India.
Appendix {#appendix .unnumbered}
========
Proof of Proposition \[prop5\] {#appen-prop5}
==============================
[**Discrimination of $\mathbb{B}_{II}(3,3)$:**]{} The GNPB of Proposition \[prop3\] is given by, $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II}(3,3)\equiv$}
\left\{\!\begin{aligned}
\ket{\psi(\pm,\pm)}_1:=\ket{0}_A\ket{\eta_\pm}_B\ket{\xi_\pm}_C,~\ket{\psi(\pm,\pm)}_2:=\ket{\eta_\pm}_A\ket{2}_B\ket{\xi_\pm}_C,~\\
\ket{\psi(\pm,\pm)}_3:=\ket{2}_A\ket{\xi_\pm}_B\ket{\eta_\pm}_C,~\ket{\psi(\pm,\pm)}_4:=\ket{\eta_\pm}_A\ket{\xi_\pm}_B\ket{0}_C,~\\
\ket{\psi(\pm,\pm)}_5:=\ket{\xi_\pm}_A\ket{0}_B\ket{\eta_\pm}_C,~\ket{\psi(\pm,\pm)}_6:=\ket{\xi_\pm}_A\ket{\eta_\pm}_B\ket{2}_C,~\\
\ket{\phi(k)}:=\ket{k}_A\ket{k}_B\ket{k}_C~~|~~ k\in\{0,1,2\}~~~~~~~~~~~~~~~~
\end{aligned}\right\}. \end{aligned}$$ Using the entanglement resource $\ket{\phi^+(3)}$ Bob teleports his subsystem to Alice. Thus without loss of generality after this step we can think that they are in same lab and we will use the subindex $\tilde{A}$ for this joint part. While Alice and Bob are together the states in $\mathbb{B}_{II}(3,3)$ have the following tile structure.
![Tile structure of the GNPB $\mathbb{B}_{II}(3,3)$ in AB|C cut. This particular tile structure is similar to that of $\mathbb{C}^3\otimes\mathbb{C}^3$ tile UPB.[]{data-label="tile"}](fig1){width="60.00000%"}
For discriminating the state Charlie shares $\ket{\phi^+}$ with $\tilde{A}$. Therefore the initial state is $$\ket{\psi}_{\tilde{A}C}\otimes\ket{\phi^+}_{ac},$$ where $\ket{\psi}_{\tilde{A}C}$ is one of the states from $\mathbb{B}_{II}(3,3)$ and this allows representation as in Fig. \[tile1\] (see [@Cohen08] for details of this representation).
![While Alice and Bob are together, the state $\ket{\psi}_{\tilde{A}C}\otimes\ket{\phi^+}_{ac}$ lives in $\mathbb{C}^{18}\otimes\mathbb{C}^6$. Curly brace on right hand side denotes the measurement effect $N:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_C;\ket{0}_{c}\right]+\mathbb{P}\left[\ket{2}_C;\ket{1}_{c}\right]$ in Step-1.[]{data-label="tile1"}](fig2){width=".8\textwidth"}
For short hand notation we will denote $\ket{ij}\mapsto \ket{{\bf 3i+j}}$. Now the discrimination protocol proceeds as follows.
[**Step-1:**]{} Charlie performs the measurement $$\mathcal{N}\equiv\left\lbrace N:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_C;\ket{0}_{c}\right]+\mathbb{P}\left[\ket{2}_C;\ket{1}_{c}\right],\overline{N}:=\mathbb{I}-N\right\rbrace,$$ where, $\mathbb{P}\left[\left(\ket{i},\ket{j}\right)_{\$};\left(\ket{k},\ket{l}\right)_{\#}\right]:=\left( \ket{i}\bra{i}+\ket{j}\bra{j}\right)_{\$}\otimes\left( \ket{k}\bra{k}+\ket{l}\bra{l}\right)_{\#}$, and this definition is applicable for all the protocols. Suppose outcome corresponding to $N$ clicks.
[**Step-2:**]{} Alice performs the measurement $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathcal{K}\equiv$}
\left\{\!\begin{aligned}
K_1&:=\mathbb{P}\left[\left(\ket{\mathbf{3}},\ket{\mathbf{6}},\ket{\mathbf{7}},\ket{\mathbf{8}}\right)_{\tilde{A}};\ket{0}_{a}\right],\\
K_2&:=\mathbb{P}\left[\left(\ket{\mathbf{3}},\ket{\mathbf{4}},\ket{\mathbf{6}},\ket{\mathbf{7}},\ket{\mathbf{8}}\right)_{\tilde{A}};\ket{1}_{a}\right],\\
K_3&:=\mathbb{I}-K_1-K_2.~~~~~~~~~~
\end{aligned}\right\}. \end{aligned}$$ If $K_1$ clicks the given state is one of $\{\ket{\psi(\pm,\pm)}_3,\ket{\psi(\pm,\pm)}_5\}$ and this set of states are perfectly LOCC distinguishable. If $K_1$ clicks the state is one of $\{\ket{\psi(\pm,\pm)}_6,\ket{\phi(2)}\}$, again a LOCC distinguishable set. Otherwise it is one of the remaining $14$ states.
[**Step-3:**]{} Charlie performs the measurement $\mathcal{N}^\prime\equiv\left\lbrace N^\prime:=\mathbb{P}\left[\ket{0}_C;\mathbb{I}_{c}\right],\overline{N}^\prime:=\mathbb{I}-N^\prime\right\rbrace$. If $N^\prime$ clicks the state is one of $\{\ket{\psi(\pm,\pm)}_4,\ket{\phi(0)}\}$ (LOCC distinguishable set), else it is one of remaining $9$ states.
[**Step-4:**]{} Alice performs the measurement $\mathcal{K}^\prime\equiv\left\lbrace K^\prime:=\mathbb{P}\left[\ket{\mathbf{4}}_{\tilde{A}};\mathbb{I}_{a}\right],\overline{K}^\prime:=\mathbb{I}-K^\prime\right\rbrace$. If $K^\prime$ clicks the state is $\ket{\phi(1)}$, otherwise it is one of $\{\ket{\psi(\pm,\pm)}_1,\ket{\psi(\pm,\pm)}_2\}$, a LOCC distinguishable set. If in [**Step-1**]{} $\overline{N}$ clicks then also a similar protocol follows.
[**Discrimination of $\mathbb{B}_{II(b)}(3,3)$:**]{} The GNPB of Proposition \[prop3\] is given by, $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II(b)}(3,3)\equiv$}
\left\{\!\begin{aligned}
\ket{\alpha(\pm)}_1:={\ket{0}_A\ket{1}_B\ket{\eta_\pm}_C},~~\ket{\alpha(\pm)}_2:={\ket{0}_A\ket{2}_B\ket{\kappa_\pm}_C},\\
\ket{\alpha(\pm)}_3:={\ket{1}_A\ket{2}_B\ket{\eta_\pm}_C},~~\ket{\alpha(\pm)}_4:={\ket{2}_A\ket{1}_B\ket{\kappa_\pm}_C},\\
\ket{\beta(\pm)}_1:={\ket{1}_A\ket{\eta_\pm}_B\ket{0}_C},~~\ket{\beta(\pm)}_2:={\ket{2}_A\ket{\kappa_\pm}_B\ket{0}_C},\\
\ket{\beta(\pm)}_3:={\ket{2}_A\ket{\eta_\pm}_B\ket{1}_C},~~\ket{\beta(\pm)}_4:={\ket{1}_A\ket{\kappa_\pm}_B\ket{2}_C},\\
\ket{\gamma(\pm)}_1:={\ket{\eta_\pm}_A\ket{0}_B\ket{1}_C},~~\ket{\gamma(\pm)}_2:={\ket{\kappa_\pm}_A\ket{0}_B\ket{2}_C},\\
\ket{\gamma(\pm)}_3:={\ket{\eta_\pm}_A\ket{1}_B\ket{2}_C},~~\ket{\gamma(\pm)}_4:={\ket{\kappa_\pm}_A\ket{2}_B\ket{1}_C},\\
\ket{\phi(k)}:={\ket{k}_A\ket{k}_B\ket{k}_C}~~|~~k\in\{0,1,2\}~.~~~~~~~~~~
\end{aligned}\right\}. \end{aligned}$$ While Alice and Bob are together (after Bob teleports his part to Alice using $\log3$-ebits) the states in $\mathbb{B}_{II(b)}(3,3)$ have the following tile structure.
![Tile structure of the GNPB $\mathbb{B}_{II(b)}(3,3)$ in AB|C cut. The star (diamond) shaped tiles contain the states $\ket{\gamma(\pm)}_1~(\ket{\gamma(\pm)}_2)$.[]{data-label="tile2"}](fig3){width=".6\textwidth"}
[**Step-1:**]{} Charlie performs the measurement $$\mathcal{N}\equiv\left\lbrace N:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_C;\ket{0}_{c}\right]+\mathbb{P}\left[\ket{2}_C;\ket{1}_{c}\right],\overline{N}:=\mathbb{I}-N\right\rbrace.$$ Suppose $N$ clicks.
[**Step-2:**]{} Alice’s measurement and the states corresponding to different outcomes are shown. $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathcal{K}\equiv$}
\left\{\!\begin{aligned}
K_1&:=\mathbb{P}\left[\left(\ket{\mathbf{0}},\ket{\mathbf{3}},\ket{\mathbf{4}}\right)_{\tilde{A}};\ket{0}_{a}\right]\Rightarrow\{\ket{\beta(\pm)}_1,\ket{\gamma(\pm)}_1,\ket{\phi(0)},\ket{\phi(1)}\}\\
K_2&:=\mathbb{P}\left[\ket{\mathbf{1}}_{\tilde{A}};\ket{0}_{a}\right]\Rightarrow\{\ket{\alpha(\pm)}_1\},~~~~~~~~~~~~~~
K_3:=\mathbb{P}\left[\ket{\mathbf{5}}_{\tilde{A}};\ket{0}_{a}\right]\Rightarrow\{\ket{\alpha(\pm)}_3\},\\
K_4&:=\mathbb{P}\left[\left(\ket{\mathbf{0}},\ket{\mathbf{6}}\right)_{\tilde{A}};\ket{1}_{a}\right]\Rightarrow\{\ket{\gamma(\pm)}_2\},~~~~
K_5:=\mathbb{P}\left[\left(\ket{\mathbf{1}},\ket{\mathbf{4}}\right)_{\tilde{A}};\ket{1}_{a}\right]\Rightarrow\{\ket{\gamma(\pm)}_3\},\\
K_6&:=\mathbb{P}\left[\left(\ket{\mathbf{3}},\ket{\mathbf{5}}\right)_{\tilde{A}};\ket{1}_{a}\right]\Rightarrow\{\ket{\beta(\pm)}_4\},~~~~
K_7:=\mathbb{P}\left[\ket{\mathbf{8}}_{\tilde{A}};\ket{1}_{a}\right]\Rightarrow\{\ket{\phi(2)}\},\\
K_8&:=\mathbb{I}-\sum_{i=1}^7K_i\Rightarrow\{\ket{\alpha(\pm)}_2,\ket{\alpha(\pm)}_4,\ket{\beta(\pm)}_2,\ket{\beta(\pm)}_3,\ket{\gamma(\pm)}_4\}.
\end{aligned}\right\}. \end{aligned}$$
[**Step-3:**]{} If $K_1$ clicks, Charlie performs the measurement $\mathcal{N}^\prime\equiv\left\lbrace N^\prime:=\mathbb{P}\left[\ket{0}_C;\mathbb{I}_{c}\right],\overline{N}^\prime:=\mathbb{I}-N^\prime\right\rbrace$. If $N^\prime$ clicks then Alice performs the measurement $\mathcal{K}^\prime\equiv\left\lbrace K^\prime:=\mathbb{P}\left[\ket{\mathbf{0}}_{\tilde{A}};\mathbb{I}_{a}\right],\overline{K}^\prime:=\mathbb{I}-K^\prime\right\rbrace$, else she performs $\mathcal{K}^\prime\equiv\left\lbrace K^\prime:=\mathbb{P}\left[\ket{\mathbf{4}}_{\tilde{A}};\mathbb{I}_{a}\right],\overline{K}^\prime:=\mathbb{I}-K^\prime\right\rbrace$. The states corresponding to the outcomes are listed below: $$\begin{aligned}
\left\{\!\begin{aligned}
N^\prime~\&~K^\prime&\Rightarrow\ket{\phi(0)},~~
N^\prime~\&~\overline{K}^\prime\Rightarrow\{\ket{\beta(\pm)}_1\},\\
\overline{N}^\prime~\&~ K^\prime&\Rightarrow\ket{\phi(1)},~~
\overline{N}^\prime~\&~\overline{K}^\prime\Rightarrow\{\ket{\gamma(\pm)}_1\}.
\end{aligned}\right\}. \end{aligned}$$
If $K_8$ clicks, Charlie performs the measurement $\mathcal{N}^\prime\equiv\left\lbrace N^\prime:=\mathbb{P}\left[\ket{1}_C;\mathbb{I}_{c}\right],\overline{N}^\prime:=\mathbb{I}-N^\prime\right\rbrace$. If $N^\prime$ clicks then Alice performs the measurement $\mathcal{K}^\prime\equiv\left\lbrace K^\prime:=\mathbb{P}\left[\left(\ket{\mathbf{6}},\ket{\mathbf{7}}\right) _{\tilde{A}};\mathbb{I}_{a}\right],\overline{K}^\prime:=\mathbb{I}-K^\prime\right\rbrace$, else she performs $\mathcal{K}^\prime\equiv\left\lbrace K_1^\prime:=\mathbb{P}\left[\ket{\mathbf{7}}_{\tilde{A}};\mathbb{I}_{a}\right],K_2^\prime:=\mathbb{P}\left[\ket{\mathbf{2}}_{\tilde{A}};\mathbb{I}_{a}\right] ,K_3^\prime:=\mathbb{I}-K_1^\prime-K_2^\prime\right\rbrace $. The states corresponding to the outcomes are listed below: $$\begin{aligned}
\left\{\!\begin{aligned}
N^\prime~\&~K^\prime\Rightarrow\{\ket{\beta(\pm)}_3\},~~
N^\prime~\&~\overline{K}^\prime\Rightarrow\{\ket{\gamma(\pm)}_4\},\\
\overline{N}^\prime~\&~ K_1^\prime\Rightarrow\{\ket{\alpha(\pm)}_4\},~~
\overline{N}^\prime~\&~K_2^\prime\Rightarrow\{\ket{\alpha(\pm)}_2\},\\
\overline{N}^\prime~\&~K_3^\prime\Rightarrow\{\ket{\beta(\pm)}_2\}.~~~~~~~~~~~~~~~~
\end{aligned}\right\}. \end{aligned}$$ In [**Step-1**]{} if $\overline{N}$ clicks instead of $N$ then also a similar protocol follows.
Proof of Proposition \[prop6\] {#appen-prop6}
==============================
We need to discriminate the basis $\mathbb{B}_{II}(3,3)$. Let the EPR state between Alice & Bob be denoted as $\ket{\phi^+}_{a_1b_1}$ and that shared between Alice & Charlie be denoted as $\ket{\phi^+}_{a_2c_1}$. Therefore the initial shared states among them is, $$\ket{\psi}_{ABC}\otimes\ket{\phi^+}_{a_1b_1}\otimes\ket{\phi^+}_{a_2c_1},$$ where $\ket{\psi}_{ABC}$ is one of the state from the set $\mathbb{B}_{II}(3,3)$.
[**Step-1:**]{} Bob performs a measurement $$\mathcal{M}\equiv\left\lbrace M:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_B;\ket{0}_{b_1}\right]+\mathbb{P}\left[\ket{2}_B;\ket{1}_{b_1}\right],\overline{M}:=\mathbb{I}-M\right\rbrace,$$ and Charlie performs measurement, $$\mathcal{N}\equiv\left\lbrace N:=\mathbb{P}\left[\left(\ket{1},\ket{2}\right)_C;\ket{0}_{c_1}\right]+\mathbb{P}\left[\ket{0}_C;\ket{1}_{c_1}\right],\overline{N}:=\mathbb{I}-N\right\rbrace.$$ Suppose the outcomes corresponding to $M$ and $N$ click. The resulting post measurement state is therefore, $$\begin{aligned}
\label{5s1}
\left\{\!\begin{aligned}
\ket{\psi(\pm,\pm)}_1&\longrightarrow\ket{0}_A\ket{\eta_\pm}_B\ket{\xi_\pm}_C\ket{00}_{a_1b_1}\ket{00}_{a_2c_1},\\
\ket{\psi(\pm,\pm)}_2&\longrightarrow\ket{\eta_\pm}_A\ket{2}_B\ket{\xi_\pm}_C\ket{11}_{a_1b_1}\ket{00}_{a_2c_1},\\
\ket{\psi(\pm,\pm)}_3&\longrightarrow\ket{2}_A\left(\ket{1}_B\ket{00}_{a_1b_1}\pm\ket{2}_B\ket{11}_{a_1b_1} \right)\left(\ket{0}_C\ket{11}_{a_2c_1}\pm\ket{1}_C\ket{00}_{a_2c_1}\right),\\
\ket{\psi(\pm,\pm)}_4&\longrightarrow\ket{\eta_\pm}_A\left(\ket{1}_B\ket{00}_{a_1b_1}\pm\ket{2}_B\ket{11}_{a_1b_1} \right)\ket{0}_C\ket{11}_{a_2c_1},\\
\ket{\psi(\pm,\pm)}_5&\longrightarrow\ket{\xi_\pm}_A\ket{0}_B\ket{00}_{a_1b_1}\left(\ket{0}_C\ket{11}_{a_2c_1}\pm\ket{1}_C\ket{00}_{a_2c_1}\right),\\
\ket{\psi(\pm,\pm)}_6&\longrightarrow\ket{\xi_\pm}_A\ket{\eta_\pm}_B\ket{2}_C\ket{00}_{a_1b_1}\ket{00}_{a_2c_1},\\
\ket{\phi(0)}&\longrightarrow{\ket{0}_A\ket{0}_B\ket{0}_C}\ket{00}_{a_1b_1}\ket{11}_{a_2c_1},\\
\ket{\phi(1)}&\longrightarrow{\ket{1}_A\ket{1}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1},\\
\ket{\phi(2)}&\longrightarrow{\ket{2}_A\ket{2}_B\ket{2}_C}\ket{11}_{a_1b_1}\ket{00}_{a_2c_1}.
\end{aligned}\right\}. \end{aligned}$$
[**Step-2:**]{} Alice performs the measurement, $$\mathcal{K}\equiv\left\lbrace K_1:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_A;\ket{1}_{a_1};\ket{0}_{a_2}\right],K_2:=\mathbb{P}\left[\ket{0}_A;\ket{0}_{a_1};\ket{0}_{a_2}\right],K_3:=\mathbb{I}-K_1-K_2\right\rbrace.$$ If $K_1$ clicks, the given state is from the set $\{\ket{\psi(\pm,\pm)}_2\}$ which is LOCC distinguishable. If, $K_2$ clicks, the state is one of $\{\ket{\psi(\pm,\pm)}_1\}$ (LOCC distinguishable set). Else, the state is one of the remaining $19$ states.
[**Step-3:**]{} Charlie performs the measurement, $\mathcal{N}^\prime\equiv\left\lbrace N^\prime:=\mathbb{P}\left[\ket{2}_C;\mathbb{I}_{c_1}\right],\overline{N}^\prime:=\mathbb{I}-N^\prime\right\rbrace.$ If $N^\prime$ clicks, the state is one of $\{\ket{\psi(\pm,\pm)}_6,\ket{\phi(2)}\}$ which is perfectly LOCC distinguishable.
[**Step-4:**]{} Bob performs the measurement, $\mathcal{M}^\prime\equiv\left\lbrace M^\prime:=\mathbb{P}\left[\ket{0}_B;\mathbb{I}_{b_1}\right],\overline{M}^\prime:=\mathbb{I}-M^\prime\right\rbrace.$ If $M^\prime$ clicks, the state is one of $\{\ket{\psi(\pm,\pm)}_5,\ket{\phi(0)}\}$ which is again perfectly LOCC distinguishable.
[**Step-5:**]{} Alice performs the measurement, $\mathcal{K}^\prime\equiv\left\lbrace K^\prime:=\mathbb{P}\left[\ket{2}_A;\mathbb{I}_{a_1};\mathbb{I}_{a_2}\right],\overline{K}^\prime:=\mathbb{I}-K^\prime\right\rbrace.$ If $K^\prime$ clicks, the state is one of $\{\ket{\psi(\pm,\pm)}_3\}$ (LOCC distinguishable set) else it is from the remaining set of LOCC distinguishable states $\{\ket{\psi(\pm,\pm)}_4,\ket{\phi(1)}\}$.
After the [**Step-1**]{}, only one case (corresponding to the outcomes $M$ and $N$) is discussed. For all other cases a similar protocol follows.
Proof of Proposition \[prop7\] {#appen-prop7}
==============================
We need to discriminate the set $\mathbb{B}_{II(b)}(3,3)$. EPR state shared between Alice & Bob be denoted as $\ket{\phi^+}_{a_1b_1}$, shared between Alice & Charlie as $\ket{\phi^+}_{a_2c_1}$, and between Bob & Charlie as $\ket{\phi^+}_{b_2c_2}$. Therefore the initial shared states among them is, $$\ket{\psi}_{ABC}\otimes\ket{\phi^+}_{a_1b_1}\otimes\ket{\phi^+}_{a_2c_1}\otimes\ket{\phi^+}_{b_2c_2},$$ where $\ket{\psi}_{ABC}$ is one of the state from the set $\mathbb{B}_{II(b)}(3,3)$ which they want to identify by LOCC.
[**Step-1:**]{} Bob performs a measurement $$\mathcal{M}\equiv\left\lbrace M:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_B;\ket{0}_{b_1}\right]+\mathbb{P}\left[\ket{2}_B;\ket{1}_{b_1}\right] ,~\overline{M}:=\mathbb{I}-M\right\rbrace.$$ Charlie performs a measurement $$\mathcal{N}\equiv\left\lbrace N:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_C;\ket{0}_{c_1}\right]+\mathbb{P}\left[\ket{2}_C;\ket{1}_{c_1}\right] ,~\overline{N}:=\mathbb{I}-N\right\rbrace,$$ Suppose that the outcomes corresponding to $M$ and $N$ click. The resulting post measurement state is therefore,
$$\begin{aligned}
\label{6s1}
\left\{\!\begin{aligned}
\ket{\alpha(\pm)}_1&\longrightarrow\ket{0}_A\ket{1}_B\ket{\eta_\pm}_C\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\alpha(\pm)}_2&\longrightarrow\ket{0}_A\ket{2}_B\ket{11}_{a_1b_1}\left(\ket{0}_C\ket{00}_{a_2c_1}\pm\ket{2}_C\ket{11}_{a_2c_1}\right) \ket{\phi^+}_{b_2c_2},\\
\ket{\alpha(\pm)}_3&\longrightarrow\ket{1}_A\ket{2}_B\ket{\eta_\pm}_C\ket{11}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\alpha(\pm)}_4&\longrightarrow\ket{2}_A\ket{1}_B\ket{00}_{a_1b_1}\left(\ket{0}_C\ket{00}_{a_2c_1}\pm\ket{2}_C\ket{11}_{a_2c_1}\right)\ket{\phi^+}_{b_2c_2},\\
\ket{\beta(\pm)}_1&\longrightarrow{\ket{1}_A\ket{\eta_\pm}_B\ket{0}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\beta(\pm)}_2&\longrightarrow\ket{2}_A\left(\ket{0}_B\ket{00}_{a_1b_1}\pm\ket{2}_B\ket{11}_{a_1b_1}\right) \ket{0}_C\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\beta(\pm)}_3&\longrightarrow{\ket{2}_A\ket{\eta_\pm}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\beta(\pm)}_4&\longrightarrow\ket{1}_A\left(\ket{0}_B\ket{00}_{a_1b_1}\pm\ket{2}_B\ket{11}_{a_1b_1}\right) \ket{2}_C\ket{11}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\gamma(\pm)}_1&\longrightarrow{\ket{\eta_\pm}_A\ket{0}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\gamma(\pm)}_2&\longrightarrow{\ket{\kappa_\pm}_A\ket{0}_B\ket{2}_C}\ket{00}_{a_1b_1}\ket{11}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\gamma(\pm)}_3&\longrightarrow{\ket{\eta_\pm}_A\ket{1}_B\ket{2}_C}\ket{00}_{a_1b_1}\ket{11}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\gamma(\pm)}_4&\longrightarrow{\ket{\kappa_\pm}_A\ket{2}_B\ket{1}_C}\ket{11}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\phi(0)}&\longrightarrow{\ket{0}_A\ket{0}_B\ket{0}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\phi(1)}&\longrightarrow{\ket{1}_A\ket{1}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\phi(2)}&\longrightarrow{\ket{2}_A\ket{2}_B\ket{2}_C}\ket{11}_{a_1b_1}\ket{11}_{a_2c_1}\ket{\phi^+}_{b_2c_2}.
\end{aligned}\right\}. \end{aligned}$$
[**Step-2:**]{} Alice performs the measurement $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathcal{K}\equiv$}
\left\{\!\begin{aligned}
K_1&:=\mathbb{P}\left[\ket{1}_A;\ket{1}_{a_1};\ket{0}_{a_2}\right],~~~~~~~~~~~~~
K_2:=\mathbb{P}\left[\ket{2}_A;\ket{1}_{a_1};\ket{1}_{a_2}\right],\\
K_3&:=\mathbb{P}\left[\left( \ket{0}_A,\ket{1}_A\right) ;\ket{0}_{a_1};\ket{0}_{a_2}\right],~~
K_4:=\mathbb{I}-K_1-K_2-K_3.
\end{aligned}\right\}. \end{aligned}$$ If $K_1$ clicks the given state $\ket{\psi}_{ABC}$ is one of $\{\ket{\alpha(\pm)}_3\}$ which are LOCC discriminable; if $K_2$ clicks the state is $\ket{\phi(2)}$; if $K_3$ clicks the state is one of the following set of states, $$\begin{aligned}
\label{7s7}
\left\{\!\begin{aligned}
\ket{\alpha(\pm)}_1&\rightarrow{\ket{0}_A\ket{1}_B\ket{\eta_\pm}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\beta(\pm)}_1&\rightarrow{\ket{1}_A\ket{\eta_\pm}_B\ket{0}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\gamma(\pm)}_1&\rightarrow{\ket{\eta_\pm}_A\ket{0}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\phi(0)}&\rightarrow{\ket{0}_A\ket{0}_B\ket{0}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2},\\
\ket{\phi(1)}&\rightarrow{\ket{1}_A\ket{1}_B\ket{1}_C}\ket{00}_{a_1b_1}\ket{00}_{a_2c_1}\ket{\phi^+}_{b_2c_2}.
\end{aligned}\right\}. \end{aligned}$$ Since all states of the ancillary systems $a_1b_1$ and $a_2c_1$ are identical in all the case so they provide no further advantage in discrimination and therefore redundant. Detaching theses ancillary systems along with $\ket{\phi^+}_{b_2c_2}$ the remaining states are the complete product bases corresponding to the shift UPB of $\mathbb{C}^2\otimes\mathbb{C}^2\otimes{C}^2$ [@Bennett99-1]. These states cannot be further discriminated under LOCC. However, the additional resource state $\ket{\phi^+}_{b_2c_2}$ make it possible to discriminate the above states perfectly.
If $K_4$ clicks the given state is one of the remaining $27-(2+1+8)=16$ states. For these states discriminating protocol goes as follows.
[**Step-3:**]{} Charlie performs $\mathcal{N}^\prime\equiv\left\lbrace N^\prime:=\mathbb{P}\left[\ket{1}_C;\mathbb{I}_{c_1}\right],~\overline{N}^\prime:=\mathbb{I}-N^\prime\right\rbrace$. If $N^\prime$ clicks the state is given from $\{\ket{\beta(\pm)}_3,\ket{\gamma(\pm)}_4\}$; else it is one of the remaining $12$ states.
[**Step-4:**]{} Bob performs $\mathcal{M}^\prime\equiv\left\lbrace M^\prime:=\mathbb{P}\left[\ket{1}_B;\mathbb{I}_{b_1}\right],~\overline{M}^\prime:=\mathbb{I}-M^\prime\right\rbrace$. If $M^\prime$ clicks the state is given from $\{\ket{\alpha(\pm)}_4,\ket{\gamma(\pm)}_3\}$; else it is one of the remaining $8$ states.
[**Step-5:**]{} Alice performs the measurement $$\begin{aligned}
\rotatebox[origin=c]{0}{$\mathcal{K}^\prime\equiv$}
\left\{\!\begin{aligned}
K_1^\prime&:=\mathbb{P}\left[\ket{0}_A;\ket{1}_{a_1};\mathbb{I}_{a_2}\right],~~~
K_2^\prime:=\mathbb{P}\left[\ket{2}_A;\mathbb{I}_{a_1};\ket{0}_{a_2}\right],\\
K_3^\prime&:=\mathbb{P}\left[\ket{1}_A;\mathbb{I}_{a_1};\mathbb{I}_{a_2}\right],~~~~~~~~
K_4^\prime:=\mathbb{P}\left[\left(\ket{0}_A,\ket{2}_A\right) ;\ket{0}_{a_1};\ket{1}_{a_2}\right]
\end{aligned}\right\}. \end{aligned}$$ If $K_1^\prime$ clicks the state is one of $\{\ket{\alpha(\pm)}_2\}$, if $K_2^\prime$ clicks the state is one of $\{\ket{\beta(\pm)}_2\}$, if $K_3^\prime$ clicks the state is one of $\{\ket{\beta(\pm)}_4\}$, else the state is one of $\{\ket{\gamma(\pm)}_2\}$.
Since the given state is chosen randomly from the set $\mathbb{B}_{II(b)}(3,3)$, therefore the average entanglement consumption in the above protocol is $(1+1+\frac{8}{27})$-ebits. Here, remember that after the [**Step-1**]{}, only one outcome is discussed. Other outcomes are also equally likely and hence the entanglement consumption is actually the average.
Proof of Proposition \[prop8\] {#appen-prop8}
==============================
The set of states needs to be discriminated is given by, $$\begin{aligned}
\label{c1}
\rotatebox[origin=c]{0}{$\mathbb{B}_{II}(4,3):=$}
\left\{\!\begin{aligned}
\ket{3}\ket{\beta},~\ket{\beta}\ket{3},~{\ket{0}\ket{3}\ket{\chi_\pm}},~{\ket{2}\ket{\chi_\pm}\ket{2}},~{\ket{\chi_\pm}\ket{3}\ket{1}},~~~~~~~~~~~~~\\
|3\rangle|0\rangle|3\rangle,~|3\rangle|1\rangle|3\rangle,~|3\rangle|2\rangle|3\rangle,~|3\rangle|3\rangle|0\rangle,~|3\rangle|3\rangle|2\rangle,~
|3\rangle|3\rangle|3\rangle,~|2\rangle|0\rangle|0\rangle,\\
|2\rangle|0\rangle|1\rangle,~|2\rangle|0\rangle|2\rangle,~|2\rangle|1\rangle|0\rangle,~|2\rangle|1\rangle|1\rangle,~|2\rangle|1\rangle|2\rangle,~
|2\rangle|2\rangle|0\rangle,~|2\rangle|2\rangle|1\rangle,\\
|2\rangle|3\rangle|0\rangle,~|2\rangle|3\rangle|3\rangle,~|0\rangle|0\rangle|0\rangle,~|0\rangle|0\rangle|1\rangle,~|0\rangle|0\rangle|2\rangle,~
|0\rangle|1\rangle|0\rangle,~|0\rangle|1\rangle|1\rangle,\\
|0\rangle|1\rangle|2\rangle,~|0\rangle|2\rangle|0\rangle,~|0\rangle|2\rangle|1\rangle,~|0\rangle|2\rangle|2\rangle,~|0\rangle|3\rangle|0\rangle,~
|0\rangle|3\rangle|1\rangle,~|1\rangle|0\rangle|0\rangle,\\
|1\rangle|0\rangle|1\rangle,~|1\rangle|0\rangle|2\rangle,~|1\rangle|1\rangle|0\rangle,~|1\rangle|1\rangle|1\rangle,~|1\rangle|1\rangle|2\rangle,~
|1\rangle|2\rangle|0\rangle,~|1\rangle|2\rangle|1\rangle,\\
|1\rangle|2\rangle|2\rangle,~|1\rangle|3\rangle|0\rangle,~|1\rangle|3\rangle|1\rangle,~|1\rangle|3\rangle|2\rangle,~|1\rangle|3\rangle|3\rangle.~~~~~~~~~~~~~~~
\end{aligned}\right\}, \end{aligned}$$ where $\ket{\beta}\in\mathcal{B}\equiv\left\lbrace{\ket{0}\ket{\eta_\pm}},~{\ket{\eta_\pm}\ket{2}},~{\ket{2}\ket{\xi_\pm}},~{\ket{\xi_\pm}\ket{0}},~{\ket{1}\ket{1}}\right\rbrace$. Suppose that they share the resource state $(\ket{000}_{abc}+\ket{111}_{abc})/\sqrt{2}$ among them.
[**Step-1:**]{} Bob performs the measurement $$\mathcal{M}\equiv\left\lbrace M:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_B;\ket{0}_{b}\right]+\mathbb{P}\left[\left(\ket{2},\ket{3}\right)_B;\ket{1}_{b}\right] ,~\overline{M}:=\mathbb{I}-M\right\rbrace.$$ Suppose $M$ clicks. The set of states tagged only with $\ket{000}_{abc}$, only with $\ket{111}_{abc}$, or in entangled form of these tags are listed below: $$\begin{aligned}
\label{c11}
\rotatebox[origin=c]{0}{$\ket{000}_{abc}\Rightarrow$}
\left\{\!\begin{aligned}
\ket{3}{\ket{0}\ket{\eta_\pm}},~\ket{3}{\ket{\eta_\pm}\ket{2}},~\ket{3}{\ket{1}\ket{1}},~{\ket{0}\ket{\eta_\pm}}\ket{3},~{\ket{\xi_\pm}\ket{0}}\ket{3},~{\ket{1}\ket{1}}\ket{3},\\
|3\rangle|0\rangle|3\rangle,~|3\rangle|1\rangle|3\rangle,~|2\rangle|0\rangle|0\rangle,~|2\rangle|0\rangle|1\rangle,~|2\rangle|0\rangle|2\rangle,
~|2\rangle|1\rangle|0\rangle,~|2\rangle|1\rangle|1\rangle,~~~~\\
|2\rangle|1\rangle|2\rangle,~|0\rangle|0\rangle|0\rangle,~|0\rangle|0\rangle|1\rangle,~|0\rangle|0\rangle|2\rangle,~|0\rangle|1\rangle|0\rangle,
~|0\rangle|1\rangle|1\rangle,~|0\rangle|1\rangle|2\rangle,~~~~\\
|1\rangle|0\rangle|0\rangle,~|1\rangle|0\rangle|1\rangle,~|1\rangle|0\rangle|2\rangle,~|1\rangle|1\rangle|0\rangle,~
|1\rangle|1\rangle|1\rangle,~|1\rangle|1\rangle|2\rangle~~~~~~~~~~~~
\end{aligned}\right\},\\\nonumber\\
\rotatebox[origin=c]{0}{$\ket{111}_{abc}\Rightarrow$}
\left\{\!\begin{aligned}
\ket{3}\ket{2}\ket{\xi_\pm},~\ket{\eta_\pm}\ket{2}\ket{3},~{\ket{0}\ket{3}\ket{\chi_\pm}},~{\ket{2}\ket{\chi_\pm}\ket{2}},~{\ket{\chi_\pm}\ket{3}\ket{1}},~~~\\
|3\rangle|2\rangle|3\rangle,~|3\rangle|3\rangle|0\rangle,~|3\rangle|3\rangle|2\rangle,~|3\rangle|3\rangle|3\rangle,~
|2\rangle|2\rangle|0\rangle,~|2\rangle|2\rangle|1\rangle,~|2\rangle|3\rangle|0\rangle,\\
|2\rangle|3\rangle|3\rangle,~|0\rangle|2\rangle|0\rangle,~|0\rangle|2\rangle|1\rangle,~|0\rangle|2\rangle|2\rangle,~
|0\rangle|3\rangle|0\rangle,~|0\rangle|3\rangle|1\rangle,~|1\rangle|2\rangle|0\rangle,\\
|1\rangle|2\rangle|1\rangle,~|1\rangle|2\rangle|2\rangle,~|1\rangle|3\rangle|0\rangle,~|1\rangle|3\rangle|1\rangle,~|1\rangle|3\rangle|2\rangle,~|1\rangle|3\rangle|3\rangle~~~~~~~~~~~~
\end{aligned}\right\}~~~~~, \\\nonumber\\
\rotatebox[origin=c]{0}{$\mbox{Entangled}\Rightarrow$}
\left\{\!\begin{aligned}
\ket{3}_A\left(\ket{1}_B\ket{000}_{abc}\pm\ket{2}_B\ket{111}_{abc}\right)\ket{0}_C,\\
\ket{2}_A\left(\ket{1}_B\ket{000}_{abc}\pm\ket{2}_B\ket{111}_{abc}\right)\ket{3}_C
\end{aligned}\right\}.~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\end{aligned}$$
[**Step-2:**]{} Alice performs the measurement $$\mathcal{K}\equiv\left\lbrace K_1:=\mathbb{P}\left[\ket{0}_A;\ket{0}_{a}\right],~K_2:=\mathbb{P}\left[\left(\ket{0},\ket{1}\right)_A;\ket{1}_{a}\right] ,~K_3:=\mathbb{I}-K_1-K_2\right\rbrace .$$ States corresponding to the outcomes $K_1$ and $K_2$ are listed below: $$\begin{aligned}
\label{c13}
\rotatebox[origin=c]{0}{$K_1\Rightarrow$}
\left\{\!\begin{aligned}
{\ket{0}\ket{\eta_\pm}}\ket{3},~|0\rangle|0\rangle|0\rangle,~|0\rangle|0\rangle|1\rangle,~|0\rangle|0\rangle|2\rangle,~|0\rangle|1\rangle|0\rangle,
~|0\rangle|1\rangle|1\rangle,~|0\rangle|1\rangle|2\rangle
\end{aligned}\right\},~~~~~~~\\\nonumber\\
\rotatebox[origin=c]{0}{$K_2\Rightarrow$}
\left\{\!\begin{aligned}
\ket{\eta_\pm}\ket{2}\ket{3},~{\ket{0}\ket{3}\ket{\chi_\pm}},~|0\rangle|2\rangle|0\rangle,~|0\rangle|2\rangle|1\rangle,~|0\rangle|2\rangle|2\rangle,~
|0\rangle|3\rangle|0\rangle,~|0\rangle|3\rangle|1\rangle,\\
|1\rangle|2\rangle|0\rangle,~|1\rangle|2\rangle|1\rangle,~|1\rangle|2\rangle|2\rangle,~|1\rangle|3\rangle|0\rangle,~|1\rangle|3\rangle|1\rangle,~|1\rangle|3\rangle|2\rangle,~|1\rangle|3\rangle|3\rangle~~~~~~
\end{aligned}\right\}.\end{aligned}$$ Both of these two sets are LOCC distinguishable. If $K_3$ clicks the state is one of the remaining $40$ states.
[**Step-3:**]{} Charlie performs the measurement $$\mathcal{N}\equiv\left\lbrace N_1:=\mathbb{P}\left[\ket{2}_C;\ket{0}_{c}\right],~N_2:=\mathbb{P}\left[\left(\ket{1},\ket{2}\right)_C;\ket{1}_{c}\right] ,~N_3:=\mathbb{I}-N_1-N_2\right\rbrace .$$ States corresponding to the outcomes $N_1$ and $N_2$ are listed below: $$\begin{aligned}
\label{c14}
\left\{\!\begin{aligned}
N_1\Rightarrow&\left\lbrace \ket{3}{\ket{\eta_\pm}\ket{2}},~|2\rangle|0\rangle|2\rangle,~|2\rangle|1\rangle|2\rangle,~|1\rangle|0\rangle|2\rangle,~|1\rangle|1\rangle|2\rangle\right\rbrace,\\
N_2\Rightarrow&\left\lbrace \ket{3}\ket{2}\ket{\xi_\pm},~{\ket{2}\ket{\chi_\pm}\ket{2}},~{\ket{\chi_\pm}\ket{3}\ket{1}},~|3\rangle|3\rangle|2\rangle,~|2\rangle|2\rangle|1\rangle\right\rbrace
\end{aligned}\right\}.\end{aligned}$$ The states corresponding to outcome $N_1$ are LOCC distinguishable. LOCC distinguishability of the set of states corresponding to $N_2$ is discussed later (see Remark \[remark1\]). If $N_3$ clicks the given state is one of the remaining $26$ states.
[**Step-4:**]{} Bob performs the measurement $$\mathcal{M}^\prime\equiv\left\lbrace M^\prime_1:=\mathbb{P}\left[\ket{0}_B;\mathbb{I}_{b}\right],~M^\prime_2:=\mathbb{P}\left[\ket{3}_B;\mathbb{I}_{b}\right] ,~M^\prime_3:=\mathbb{I}-M^\prime_1-M^\prime_2\right\rbrace .$$ States corresponding to the outcomes $M^\prime_1$ and $M^\prime_2$ are listed below: $$\begin{aligned}
\label{c15}
\left\{\!\begin{aligned}
M^\prime_1\Rightarrow&\left\lbrace {\ket{3}\ket{0}\ket{\eta_\pm}},~{\ket{\xi_\pm}\ket{0}\ket{3}},~{\ket{3}\ket{0}\ket{3}},~{\ket{2}\ket{0}\ket{0}},~{\ket{2}\ket{0}\ket{1}},~{\ket{1}\ket{0}\ket{0}},~{\ket{1}\ket{0}\ket{1}}\right\rbrace,\\
M^\prime_2\Rightarrow&\left\lbrace {\ket{3}\ket{3}\ket{0}},~{\ket{3}\ket{3}\ket{3}},~{\ket{2}\ket{3}\ket{0}},~{\ket{2}\ket{3}\ket{3}}\right\rbrace
\end{aligned}\right\}.\end{aligned}$$ Evidently, these two sets are LOCC distinguishable. If $M^\prime_3$ clicks, given is one of the remaining $13$ states.
[**Step-5:**]{} Alice performs the measurement $$\mathcal{K}^\prime\equiv\left\lbrace K^\prime_1:=\mathbb{P}\left[\ket{1}_A;\mathbb{I}_{a}\right],~K^\prime_2:=\mathbb{P}\left[\ket{2}_A;\mathbb{I}_{a}\right] ,~K^\prime_3:=\mathbb{I}-K^\prime_1-K^\prime_2\right\rbrace .$$ States corresponding to the outcomes are listed below: $$\begin{aligned}
\label{c16}
\left\{\!\begin{aligned}
K^\prime_1\Rightarrow&\left\lbrace {\ket{1}\ket{1}\ket{3}},~{\ket{1}\ket{1}\ket{0}},~{\ket{1}\ket{1}\ket{1}}\right\rbrace,\\
K^\prime_2\Rightarrow&\left\lbrace {\ket{2}\ket{1}\ket{0}},~{\ket{2}\ket{1}\ket{1}},~{\ket{2}\ket{2}\ket{0}},~\ket{2}_A\left(\ket{1}_B\ket{000}_{abc}\pm\ket{2}_B\ket{111}_{abc}\pm\right)\ket{3}_C\right\rbrace ,\\
K^\prime_3\Rightarrow&\left\lbrace {\ket{3}\ket{1}\ket{1}},~{\ket{3}\ket{1}\ket{3}},~{\ket{3}\ket{2}\ket{3}},~\ket{3}_A\left(\ket{1}_B\ket{000}_{abc}\pm\ket{2}_B\ket{111}_{abc}\pm\right)\ket{0}_C\right\rbrace
\end{aligned}\right\}.\end{aligned}$$ Evidently all these three sets are LOCC distinguishable.
\[remark1\] If $N_2$ clicks in [**Step-3**]{} then the given state is one of $$\left\lbrace \ket{3}\ket{2}\ket{\xi_\pm},~{\ket{2}\ket{\chi_\pm}\ket{2}},~{\ket{\chi_\pm}\ket{3}\ket{1}},~|3\rangle|3\rangle|2\rangle,~|2\rangle|2\rangle|1\rangle\right\rbrace.$$ Considering the relabeling $2\mapsto1$, $3\mapsto0$ for Alice & Bob and $1\mapsto1$, $2\mapsto0$ for Charlie, the above set reads as, $$\left\lbrace {\ket{0}\ket{1}\ket{0\pm 1}},~{\ket{1}\ket{0\pm 1}\ket{0}},~{\ket{0\pm 1}\ket{0}\ket{1}},~{\ket{0}\ket{0}\ket{0}},~{\ket{1}\ket{1}\ket{1}}\right\rbrace.$$ It is the OPB corresponding to the Shift UPB of $(\mathbb{C}^2)^{\otimes3}$ [@Bennett99], and this set can be perfectly distinguished under LOCC if a $2$-qubit maximally entangled state is shared between any two parties. Since the unknown state is chosen at random ([*i.e*]{} with uniform probability) from the set $\mathbb{B}_{II}(4,3)$, therefore total entanglement consumed in this protocol are $1$ $\mbox{GHZ}$ and $\frac{1}{8}$ $\mbox{EPR}$ .
\[remark2\] Instead of GHZ resource, consider that Alice and Bob share a $2$-qubit maximally entangled state. After [**Step-1**]{}, the tag is shared between Alice and Bob only. As already discussed in [**Step-2**]{} Alice can discriminate $20$ state corresponding to the outcomes $K_1$ and $K_2$. However, if $K_3$ outcome occurs, the discrimination protocol can not be further proceeded if no more entangled resource is used. But if an entangled state $\ket{\phi^+}_{b^\prime c}$ is provided between Bob and Charlie then a perfect discrimination protocol is possible. For that Bob starts with a twist-breaking measurement $\{\mathbb{P}[(\ket{1},\ket{2})_B;\ket{0}_{b^\prime}]+\mathbb{P}[(\ket{0},\ket{3})_B;\ket{1}_{b^\prime}],~\mathbb{I}-\mathbb{P}\}$. Then an analogous protocol follows that discriminate all the remaining $44$ states. Therefore total entanglement consumption in this protocol is $\left\lbrace(1,\ket{\phi^+}_{\mathcal{AB}});(\frac{11}{16},\ket{\phi^+}_{\mathcal{BC}}) \right\rbrace $.
[99]{} W. K. Wootters and W. H. Zurek; A single quantum cannot be cloned, [Nature [**299**]{}, 802 (1982)](https://www.nature.com/articles/299802a0).
A. K. Pati and S. L. Braunstein; Impossibility of deleting an unknown quantum state, [Nature [**404**]{}, 164 (2000)](https://www.nature.com/articles/35004532).
C. W. Helstrom, Quantum Detection and Estimation Theory, [J. Stat. Phys. [**1**]{}, 231–252 (1969)](https://link.springer.com/article/10.1007/BF01007479).
A. S. Holevo; Statistical decision theory for quantum systems, [J. Multivar. Anal. [**3**]{}, 337 (1973)](https://doi.org/10.1016/0047-259X(73)90028-6).
H. Yuen, R. Kennedy, and M. Lax; Optimum testing of multiple hypotheses in quantum detection theory, [IEEE Trans. Inf. Theory [**21**]{}, 125 (1975)](https://ieeexplore.ieee.org/document/1055351).
C. H. Bennett, D. P. DiVincenzo, C. A. Fuchs, T. Mor, E. Rains, P. W. Shor, J. A. Smolin, and W. K. Wootters, Quantum nonlocality without entanglement, [Phys. Rev. A [**59**]{}, 1070 (1999)](http://dx.doi.org/10.1103/PhysRevA.59.1070).
J. S. Bell; On the Einstein Podolsky Rosen paradox, [Physics Physique Fizika [**1**]{}, 195 (1964)](https://doi.org/10.1103/PhysicsPhysiqueFizika.1.195).
N. Brunner, D. Cavalcanti, S. Pironio, V. Scarani, and S. Wehner; Bell nonlocality, [Rev. Mod. Phys. [**86**]{}, 419 (2014)](https://doi.org/10.1103/RevModPhys.86.419).
B. M. Terhal, D. P. DiVincenzo, and D. W. Leung; Hiding Bits in Bell States, [Phys. Rev. Lett. [**86**]{}, 5807 (2001)](https://doi.org/10.1103/PhysRevLett.86.5807).
T. Eggeling and R. F. Werner; Hiding classical data in multipartite quantum states, [Phys. Rev. Lett. [**89**]{}, 097905 (2002)](https://doi.org/10.1103/PhysRevLett.89.097905).
D. Markham and B. C. Sanders; Graph states for quantum secret sharing, [Phys. Rev. A [**78**]{}, 042309 (2008)](https://doi.org/10.1103/PhysRevA.78.042309).
R. Rahaman and M. G. Parker; Quantum scheme for secret sharing based on local distinguishability, [Phys. Rev. A [**91**]{}, 022330 (2015)](https://doi.org/10.1103/PhysRevA.91.022330).
J. Wang, L. Li, H. Peng, and Y. Yang; Quantum-secret-sharing scheme based on local distinguishability of orthogonal multiqudit entangled states, [Phys. Rev. A [**95**]{}, 022320 (2017)](https://doi.org/10.1103/PhysRevA.95.022320).
C. H. Bennett, D. P. DiVincenzo, T. Mor, P. W. Shor, J. A. Smolin, and B. M. Terhal; Unextendible Product Bases and Bound Entanglement, [Phys. Rev. Lett. [**82**]{}, 5385 (1999)](http://dx.doi.org/10.1103/PhysRevLett.82.5385).
J. Walgate, A. J. Short, L. Hardy, and V. Vedral; Local Distinguishability of Multipartite Orthogonal Quantum States, [Phys. Rev. Lett. [**85**]{}, 4972 (2000)](http://dx.doi.org/10.1103/PhysRevLett.85.4972).
S. Virmani, M. F. Sacchi, M. B. Plenio, and D. Markham; Optimal local discrimination of two multipartite pure states, [Phys. Lett. A. [**288**]{}, 62 (2001)](https://doi.org/10.1016/S0375-9601(01)00484-4).
S. Ghosh, G. Kar, A. Roy, A. Sen (De), and U. Sen; Distinguishability of Bell States, [Phys. Rev. Lett. [**87**]{}, 277902 (2001)](http://dx.doi.org/10.1103/PhysRevLett.87.277902).
B. Groisman and L. Vaidman; Nonlocal variables with product-state eigenstates, [J. Phys. A: Math. Gen. [**34**]{}, 6881 (2001)](http://dx.doi.org/10.1088/0305-4470/34/35/313).
J. Walgate and L. Hardy; Nonlocality, asymmetry, and distinguishing bipartite states, [Phys. Rev. Lett. [**89**]{}, 147901 (2002)](http://dx.doi.org/10.1103/PhysRevLett.89.147901).
D. P. DiVincenzo, T. Mor, P. W. Shor, J. A. Smolin, and B. M. Terhal, Unextendible product bases, uncompletable product bases and bound entanglement, [Commun. Math. Phys. [**238**]{}, 379 (2003)](http://dx.doi.org/10.1007/s00220-003-0877-6).
M. Horodecki, A. Sen (De), U. Sen, and K. Horodecki; Local Indistinguishability: More Nonlocality with Less Entanglement, [Phys. Rev. Lett. [**90**]{}, 047902 (2003)](http://dx.doi.org/10.1103/PhysRevLett.90.047902).
H. Fan; Distinguishability and Indistinguishability by Local Operations and Classical Communication, [Phys. Rev. Lett. [**92**]{}, 177905 (2004)](http://dx.doi.org/10.1103/PhysRevLett.92.177905).
S. Ghosh, G. Kar, A. Roy, and D. Sarkar; Distinguishability of maximally entangled states, [Phys. Rev. A [**70**]{}, 022304 (2004)](http://dx.doi.org/10.1103/PhysRevA.70.022304).
S. De Rinaldis; Distinguishability of complete and unextendible product bases, [Phys. Rev. A [**70**]{}, 022309 (2004)](http://dx.doi.org/10.1103/PhysRevA.70.022309).
M. Nathanson; Distinguishing bipartite orthogonal states by LOCC: best and worst cases, [J. Math. Phys. [**46**]{}, 062103 (2005)](https://doi.org/10.1063/1.1914731).
J. Watrous; Bipartite subspaces having no bases distinguishable by local operations and classical communication, [Phys. Rev. Lett. [**95**]{}, 080505 (2005)](https://doi.org/10.1103/PhysRevLett.95.080505).
J. Niset and N. J. Cerf; Multipartite nonlocality without entanglement in many dimensions, [Phys. Rev. A [**74**]{}, 052103 (2006)](http://dx.doi.org/10.1103/PhysRevA.74.052103).
M.-Y. Ye, W. Jiang, P.-X. Chen, Y.-S. Zhang, Z.-W. Zhou, and G.-C. Guo; Local distinguishability of orthogonal quantum states and generators of SU(N), [Phys. Rev. A [**76**]{}, 032329 (2007)](http://dx.doi.org/10.1103/PhysRevA.76.032329).
H. Fan; Distinguishing bipartite states by local operations and classical communication, [Phys. Rev. A [**75**]{}, 014305 (2007)](http://dx.doi.org/10.1103/PhysRevA.75.014305).
R. Duan, Y. Feng, Z. Ji, and M. Ying; Distinguishing Arbitrary Multipartite Basis Unambiguously Using Local Operations and Classical Communication, [Phys. Rev. Lett. [**98**]{}, 230502 (2007)](http://dx.doi.org/10.1103/PhysRevLett.98.230502).
S. Bandyopadhyay and J. Walgate; Local distinguishability of any three quantum states, [J. Phys. A: Math. Theor. [**42**]{}, 072002 (2009)](http://iopscience.iop.org/1751-8121/42/7/072002).
Y. Feng and Y.-Y. Shi; Characterizing locally indistinguishable orthogonal product states, [IEEE Trans. Inf. Theory [**55**]{}, 2799 (2009)](http://dx.doi.org/10.1109/TIT.2009.2018330).
R. Duan, Y. Xin, and M. Ying; Locally indistinguishable subspaces spanned by three-qubit unextendible product bases, [Phys. Rev. A [**81**]{}, 032329 (2010)](https://doi.org/10.1103/PhysRevA.81.032329).
N. Yu, R. Duan, and M. Ying; Four Locally Indistinguishable Ququad-Ququad Orthogonal Maximally Entangled States, [Phys. Rev. Lett. [**109**]{}, 020506 (2012)](http://dx.doi.org/10.1103/PhysRevLett.109.020506).
Y.-H. Yang, F. Gao, G.-J. Tian, T.-Q. Cao, and Q.-Y. Wen; Local distinguishability of orthogonal quantum states in a $2\otimes2\otimes2$ system, [Phys. Rev. A [**88**]{}, 024301 (2013)](http://dx.doi.org/10.1103/PhysRevA.88.024301).
A. M. Childs, D. Leung, L. Mančinska, and M. Ozols; A framework for bounding nonlocality of state discrimination, [Commun. Math. Phys. [**323**]{}, 1121 (2013)](http://dx.doi.org/10.1007/s00220-013-1784-0).
Z.-C. Zhang, F. Gao, G.-J. Tian, T.-Q. Cao, and Q.-Y. Wen; Nonlocality of orthogonal product basis quantum states, [Phys. Rev. A [**90**]{}, 022313 (2014)](http://dx.doi.org/10.1103/PhysRevA.90.022313).
S. X. Yu and C. H. Oh; Detecting the local indistinguishability of maximally entangled states, [arXiv:1502.01274 \[quant-ph\] (2015)](https://arxiv.org/abs/1502.01274).
Z.-C. Zhang, F. Gao, S.-J. Qin, Y.-H. Yang, and Q.-Y. Wen; Nonlocality of orthogonal product states, [Phys. Rev. A [**92**]{}, 012332 (2015)](http://dx.doi.org/10.1103/PhysRevA.92.012332).
Y.-L. Wang, M.-S. Li, Z.-J. Zheng, and S.-M. Fei; Nonlocality of orthogonal product-basis quantum states, [Phys. Rev. A [**92**]{}, 032313 (2015)](http://dx.doi.org/10.1103/PhysRevA.92.032313).
J. Chen and N. Johnston; The Minimum Size of Unextendible Product Bases in the Bipartite Case (and Some Multipartite Cases), [Commun. Math. Phys. [**333**]{}, 351 (2015)](https://doi.org/10.1007/s00220-014-2186-7).
Y.-H. Yang, F. Gao, G.-B. Xu, H.-J. Zuo, Z.-C. Zhang, and Q.-Y. Wen; Characterizing unextendible product bases in qutrit-ququad system, [Sci. Rep. [**5**]{}, 11963 (2015)](https://dx.doi.org/10.1038/srep11963).
Z.-C. Zhang, F. Gao, Y. Cao, S.-J. Qin, and Q.-Y. Wen; Local indistinguishability of orthogonal product states, [Phys. Rev. A [**93**]{}, 012314 (2016)](http://dx.doi.org/10.1103/PhysRevA.93.012314).
G.-B. Xu, Q.-Y. Wen, S.-J. Qin, Y.-H. Yang, and F. Gao; Quantum nonlocality of multipartite orthogonal product states; [Phys. Rev. A [**93**]{}, 032341 (2016)](http://dx.doi.org/10.1103/PhysRevA.93.032341).
X. Zhang, X. Tan, J. Weng, and Y. Li; LOCC indistinguishable orthogonal product quantum states, [Sci. Rep. [**6**]{}, 28864 (2016)](https://dx.doi.org/10.1038/srep28864).
G.-B. Xu, Y.-H. Yang, Q.-Y. Wen, S.-J. Qin, and F. Gao; Locally indistinguishable orthogonal product bases in arbitrary bipartite quantum system, [Sci. Rep. [**6**]{}, 31048 (2016)](https://dx.doi.org/10.1038/srep31048).
Y.-L. Wang, M.-S. Li, S.-M. Fei, and Z.-J. Zheng; Constructing unextendible product bases from the old ones, [arXiv:1703.06542 \[quant-ph\] (2017)](https://arxiv.org/abs/1703.06542).
Z.-C. Zhang, K.-J. Zhang, F. Gao, Q.-Y. Wen, and C. H. Oh; Construction of nonlocal multipartite quantum states, [Phys. Rev. A [**95**]{}, 052344 (2017)](https://doi.org/10.1103/PhysRevA.95.052344).
G.-B. Xu, Q.-Y. Wen, F. Gao, S.-J. Qin, and H.-J. Zuo; Local indistinguishability of multipartite orthogonal product bases, [Quantum Inf. Processing [**16**]{}, 276 (2017)](https://doi.org/10.1007/s11128-017-1725-5).
Y.-L. Wang, M.-S. Li, Z.-J. Zheng, and S.-M. Fei; The local indistinguishability of multipartite product states, [Quantum Inf. Processing [**16**]{}, 5 (2017)](https://doi.org/10.1007/s11128-016-1477-7).
X. Zhang, J. Weng, X. Tan, and W. Luo; Indistinguishability of pure orthogonal product states by LOCC, [Quantum Inf. Processing [**16**]{}, 168 (2017)](https://doi.org/10.1007/s11128-017-1616-9).
X. Zhang, J. Weng, Z. Zhang, X. Li, W. Luo, and X. Tan; Locally distinguishable bipartite orthogonal quantum states in a $d\otimes n$ system, [arXiv:1712.08830 \[quant-ph\] (2017)](https://arxiv.org/abs/1712.08830).
X. Zhang, J. Weng, W. Luo, and X. Tan; Local distinguishability of quantum states in multipartite System, [arXiv:1712.08970 \[quant-ph\] (2017)](https://arxiv.org/abs/1712.08970).
S. Croke and S. M. Barnett; Difficulty of distinguishing product states locally, [Phys. Rev. A [**95**]{}, 012337 (2017)](https://doi.org/10.1103/PhysRevA.95.012337).
S. Halder, M. Banik, and S. Ghosh; New family of bound entangled states residing on the boundary of Peres set, [arXiv:1801.00405 \[quant-ph\] (2018)](https://arxiv.org/abs/1801.00405).
S. Halder, M. Banik, S. Agrawal, and S. Bandyopadhyay, Strong Quantum Nonlocality without Entanglement, [Phys. Rev. Lett. [**122**]{}, 040403 (2019)](https://doi.org/10.1103/PhysRevLett.122.040403).
E. Chitambar, D. Leung, L. Mancinska, M. Ozols, and A. Winter; Everything You Always Wanted to Know About LOCC (But Were Afraid to Ask), [Commun. Math. Phys. [**328**]{}, 303 (2014)](https://doi.org/10.1007/s00220-014-1953-9).
C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wootters; Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels, [Phys. Rev. Lett. [**70**]{}, 1895 (1993)](https://doi.org/10.1103/PhysRevLett.70.1895).
S. M. Cohen; Understanding entanglement as resource: Locally distinguishing unextendible product bases, [Phys. Rev. A [**77**]{}, 012304 (2008)](http://dx.doi.org/10.1103/PhysRevA.77.012304).
S. Bandyopadhyay, G. Brassard, S. Kimmel, and W. K. Wootters; Entanglement cost of nonlocal measurements, [Phys. Rev. A [**80**]{}, 012313 (2009)](http://dx.doi.org/10.1103/PhysRevA.80.012313).
S. Bandyopadhyay, R. Rahaman, and W. K. Wootters; Entanglement cost of two-qubit orthogonal measurements, [J. Phys. A: Math. Theor. [**43**]{} 455303 (2010)](http://dx.doi.org/10.1088/1751-8113/43/45/455303).
N. Yu, R. Duan, and M. Ying; Distinguishability of Quantum States by Positive Operator-Valued Measures with Positive Partial Transpose, [IEEE Trans. Inf. Theory [**60**]{}, 2069 (2014)](https://doi.org/10.1109/TIT.2014.2307575).
S. Bandyopadhyay, A. Cosentino, N. Johnston, V. Russo, J. Watrous, and N. Yu; Limitations on Separable Measurements by Convex Optimization, [IEEE Trans. Inf. Theory [**61**]{}, 3593 (2014)](https://doi.org/10.1109/TIT.2015.2417755).
S. Bandyopadhyay, S. Halder, and M. Nathanson; Entanglement as a resource for local state discrimination in multipartite systems, [Phys. Rev. A [**94**]{}, 022311 (2016)](http://dx.doi.org/10.1103/PhysRevA.94.022311).
Z.-C. Zhang, F. Gao, T.-Q. Cao, S.-J. Qin, and Q.-Y. Wen; Entanglement as a resource to distinguish orthogonal product states, [Sci. Rep. [**6**]{}, 30493 (2016)](https://dx.doi.org/10.1038/srep30493).
S. Bandyopadhyay, S. Halder, and M. Nathanson; Optimal resource states for local state discrimination, [Phys. Rev. A [**97**]{}, 022314 (2018)](https://doi.org/10.1103/PhysRevA.97.022314).
Z.-C. Zhang, Y.-Q. Song, T.-T. Song, F. Gao, S.-J. Qin, and Q.-Y. Wen; Local distinguishability of orthogonal quantum states with multiple copies of $2\otimes2$ maximally entangled states, [Phys. Rev. A [**97**]{}, 022334 (2018)](https://doi.org/10.1103/PhysRevA.97.022334).
S. Halder; Several nonlocal sets of multipartite pure orthogonal product states, [Phys. Rev. A [**98**]{}, 022303 (2018)](https://doi.org/10.1103/PhysRevA.98.022303).
Lv-Jun Li, F. Gao, Zhi-Chao Zhang, and Qiao-Yan Wen; Local distinguishability of orthogonal quantum states with no more than one ebit of entanglement [Phys. Rev. A 99, 012343 (2019)](https://doi.org/10.1103/PhysRevA.99.012343).
C. H. Bennett, S. Popescu, D. Rohrlich, J. A. Smolin, and A. V. Thapliyal; Exact and asymptotic measures of multipartite pure-state entanglement, [Phys. Rev. A [**63**]{}, 012307 (2000)](https://doi.org/10.1103/PhysRevA.63.012307).
S. Agrawal, S. Halder, and M. Banik; Genuinely entangled subspace with all-encompassing distillable entanglement across every bipartition, [Phys. Rev. A [**99**]{}, 032335 (2019)](https://doi.org/10.1103/PhysRevA.99.032335).
[^1]: Note that this nonlocal feature is different than the concept of ‘quantum nonlocality’ as established in another seminal work by John S. Bell [@Bell64]. A multipartite input-output correlation is called nonlocal if it is not compatible with the classical description of [*local-realism*]{} (see [@Brunner14] for a review on Bell nonlocality). In quantum world such correlations can only be resulted from multipartite entangled states.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The role of the off-shell dependence of $\rho-\omega$ mixing in the charge symmetry breaking nucleon-nucleon potential is discussed. It is shown that models describing the off-shell dependence of $\rho-\omega$ mixing are not sufficient to determine the charge symmetry breaking nucleon-nucleon potential.'
address: 'Department of Physics$^a$ FM-15 and Institute For Nuclear Theory$^b$ NK-12, University of Washington, Seattle, Washington 98195, USA'
author:
- 'Thomas D. Cohen$^{a,b}$[^1], and Gerald A. Miller$^a$'
title: 'Rho-Omega Mixing Off Shell and Charge Symmetry Breaking In the N-N Potential '
---
Introduction
============
Charge symmetry breaking has been studied for a long time, see e.g. the reviews[@NS69; @H69; @SHLOMO; @HM79; @MNS90; @NMS92; @MVO94] and especially the references therein. We follow ref. [@MVO94] in summarizing a few main features. Charge independence and charge symmetry breaking is caused by the $d$-$u$ quark mass difference $m_d-m_u >0$, and electromagnetic effects. The general goal of this area of research is to find small but observable effects of the breaking of charge independence and charge symmetry. This provides significant insight into strong interaction dynamics since the underlying origin of the breaking is understood. Over the years there has been substantial experimental and theoretical progress. First, we recall the old idea that $m_d-m_u>0$ along with electromagnetic effects accounts for the observed mass differences between members of hadronic isospin multiplets. This mass difference also leads to the notion that the physical $\rho$ and $\omega$ mesons are isospin mixed superpositions of bare states of good isospin. Indeed, substantial effects of $\rho-\omega$ mixing have been observed in the $e^+e^-
\rightarrow \pi^+\pi^-$ cross section at $q^2\approx m^2_\omega$ [@Q78; @B85]. These results allow an extraction of the strong contribution to the $\rho-\omega$ mixing matrix element $<\rho|H_{str}|\omega>\approx$ - 5200 MeV$^2$. [@M94; @MVO94] Two nucleons may exchange a mixed $\rho-\omega$ meson. If one uses $<\rho|H_{str}|\omega>\approx$ - 5200 MeV$^2$ one obtains a nucleon-nucleon interaction which is consistent with the experimental value $\Delta a_{CSB} = a^N_{pp}-a^N_{nn} = 1.5\pm
0.5$ fm [@CB87]. Such a force can also consistently account most of the strong interaction contribution to the $^3H$-$^3He$ binding energy difference [@CB87] and for much of the Nolen-Schiffer anomaly[@BI87]. The TRIUMF (477 MeV [@A86] and 350 MeV[@GV369]) and IUCF (183 MeV) [@K90] experiments have compared analyzing powers of $\vec np$ and $n\vec p$ scattering and observe charge symmetry breaking at the level expected from $\pi,\gamma$ and $\rho$-$\omega$ exchange effects. The latter effects are important at 183 MeV. Thus $\rho$-$\omega$ mixing seems to describe most of the observed features or charge symmetry breaking in nuclear physics. While it is certainly true that other mechanism cannot be ruled out, $\rho$-$\omega$ mixing appears to give a consistent description of the bulk of the experimental data.
Recently, this success has been called into question. The momentum dependence of the $\rho-\omega$ mixing amplitude has been calculated using several different models[@GHT92; @PW92; @KTW93; @MTRC94]. While these models are based on quite different physical assumptions, they all share one important quality: the $\rho-\omega$ mixing at spacelike momenta in all of these models is quite different from its value at the $\omega$ pole—generally of the opposite sign and significantly reduced in magnitude. Indeed it has been shown that for a wide class of models[@OCPTW94] the mixing must go to zero at $q^2=0$ implying that amplitude changes sign. Moreover, a QCD sum-rule calculation, also apparently gives a similarly large momentum dependence of the coupling[@HHKM94]. Since the N-N potential probes the spacelike region, this appears to imply that the vector meson exchange part of the charge symmetry breaking nucleon-nucleon NN potential is very different from one based on the on-mass-shell mixing. Indeed, NN potentials have been constructed based on these momentum dependent mixing amplitudes and these are quite different from the ones used in the successful phenomenology[@GHT92]-[@MTRC94],[@IN94].
The purpose of the present paper is to study the general role of the off-shell meson propagator in NN potentials. We find that knowledge of the off-shell meson propagator is not sufficient to determine the potential. In particular, one needs the vertex functions computed from the same theory that supplied the propagator. None of the present treatments of the off-shell propagator deals with the issue of the necessary vertex functions. It is not our intent to compute these functions. Rather, we wish to clarify issues of principle. Accordingly we have included a number of simple illustrative examples. We do show, however, that the CSB induced by the $\rho-\omega$ exchange potential can account for the existing data even if the the $q^2$ dependence is exactly as specified in any of the references [@GHT92]-[@MTRC94],[@IN94],[@HHKM94]. This is done by using CSB vertex functions
We turn to an outline of this paper. In Sect. II we discuss the problem that in hadronic field theoretic models there is never a unique choice for fields, even in a renormalizable theory[@Haag], [@Ruelle], [@Borchers] and [@Coleman]. This means that the propagator and the vertex functions are not unique. We argue generally and with two explicit examples that while the propagator depends on the choice of field variables, the observables do not. Thus, knowledge of the off-shell meson propagator by itself gives no information unless one knows which field is used. One may be able to deduce which definition of the field has been used from a complete theory by studying the interactions with the other degrees of freedom in the problem. However, if the theory is incomplete and the interactions of the field with all of the other degrees of freedom is unknown, knowledge of the off-shell propagator by itself is not physically meaningful.
There is an even more serious problem. Modern meson exchange potentials are motivated by field theoretic concerns. However, there is no first principle method for obtaining the “correct” NN potential directly from either QCD, or from some hadron field theoretic model or from any experimentally accessible set of data of hadronic properties. Given this essential difficulty, we believe it is sensible to adopt the general approach used in the construction of meson exchange potentials to the case of charge symmetry breaking. This approach makes the pragmatic assumptions of including the long range features in the meson propagators and the short range features in the vertex functions. This separation is discussed in Sect. III. Such a separation may be questioned, but [*a priori*]{} these assumptions should be no worse for the case of CSB potentials then they are for the isospin conserving part of the interaction. Moreover given the lack of rigor in the construction of potential from the underlying field theory, some assumptions must be made in order to make make any connection between $\rho-\omega$ mixing and the CSB potential. Given this, it is highly desirable to make sure that the assumptions are consistent with those made elsewhere in the problem.
It is worth stressing at the outset, that in conventional treatments of meson exchange potentials the off-shell propagator plays no role. This is discussed in Sect. IV where realistic boson-exchange charge symmetric potentials are defined to be those that are consistent with the separation discussed in Sect.III. We show that for models with realistic spectral functions the momentum dependence of the meson propagator can be absorbed into that of the vertex function. An example of an unrealistic momentum dependent $\omega$ self energy is presented.
The ideas of the Sects. II-IV are applied to the CSB potential caused by $\rho-\omega $ exchange in Sect. V. We show that the influence of the momentum dependence of the $\rho-\omega $ mixing matrix element can be included by allowing the $\rho$-nucleon coupling constant to violate charge symmetry. In particular, if the model of Ref.[@HHKM94] is used one needs CSB coupling constants that are 0.8% of the standard coupling constants to reproduce the results of a potential obtained without momentum dependence in the $\rho-\omega $ mixing matrix element and without CSB in the coupling constants. We summarize the analysis in Sect. VI.
Field Redefinitions And Off-Shell Propagators \[fr\]
====================================================
It has been known for quite some time that value of an off shell propagator is completely dependent on the choice of field. This is an example of a general theorem proved by Haag[@Haag], Ruelle[@Ruelle], and Borchers[@Borchers] which has been discussed by Coleman, Wess and Zumino[@Coleman]. The off-shell propagators depend on the choice of interpolating fields, whereas all S-matrix elements are independent of this choice. Thus an off-shell propagator, taken in isolation, can have no physical meaning.
To illustrate why this is so, let us consider the simplest possible case, the field corresponding to a stable scalar particle in some nontrivial interacting field theory. The equation of motion for this system may be written as $$\Box \phi(x) + m^2 \phi(x) = -j(x).$$ This equation of motion is determined from a Lagrangean density ${\cal L}(\phi,j)$. Furthermore, let us insist on studying the renormalized field, mass and current. This means that the correlation function for $\phi$ will have a pole with residue of unity at the physical mass, $m$: $$\langle \phi, \vec{p} | \phi(x) | {\rm vac} \rangle =
e^{ip\cdot x}$$ which implies that $$\lim_{q^2 \rightarrow m^2}\, (q^2 - m^2) \, \int {\rm d}^4 x \,
e^{i q \cdot x}
\, \langle {\rm vac}| {\rm T}[\phi(x) \phi(0)] |{\rm vac} \rangle
\,
= \, i .
\label{renorm}$$
We are concentrating on the renormalized quantities because un-renormalized properties are not observable and depend on the details of the renormalization procedure. Ultimately we will be interested in the spectral decomposition of the propagator in terms of the physical states of the system and this is directly related to the [*renormalized*]{} fields and sources.
It is worth noting that the renormalization conditions put quite stringent constraints on matrix elements of the renormalized source. In particular they imply that source does not connect the vacuum to a one particle state $$\langle \phi , \vec{p} | j(0) | {\rm vac} \rangle \, = \, 0 .
\label{cond}$$ This can be seen simply: $$\langle \phi , \vec{p} | j(x) | {\rm vac}\rangle = \langle
\phi ,
\vec{p} | (\Box + m^2) \phi(x) | {\rm vac}\rangle = (-p^2 + m^2)
\langle
\phi , \vec{p} | \phi(x) | {\rm vac} \rangle$$ where $p^2$ is the square of the four momentum of the state which is $m^2$.
Now we come to the crux of the issue. There is enormous freedom in the choice of field variables, and consequently the Green’s functions. In particular, we introduce a new renormalized field and a new source current according to: $$\begin{aligned}
\phi^{\prime}(x) & = & \phi(x) + a(x), \label{def}\\
j^{\prime}(x)& = & j(x) + (\Box + m^2)\, a(x) \label{jpdef}\end{aligned}$$ where $a(x)$ is an operator such that $\langle {\rm vac} | a| \phi, \vec{p} \rangle = 0$. Thus, for example, $a(x)$ may be a multiple of the renormalized source $j(x)$ or $a(x)$ could have the form $a(x) = (\Box + m^2) b(x)$ where $b(x)$ is an arbitrary renormalized local composite operator. The new field and source satisfies an equation of motion with the same form as the original: $$(\Box + m^2) \phi^{\prime}(x) = j^{\prime}(x) \label{eomp}$$ It also satisfies the same renormalization conditions $$\begin{aligned}
\langle \phi, \vec{p} | \phi^{\prime}(x) | {\rm vac} \rangle & =
&
e^{i p\cdot x}
\\
\lim_{q^2 \rightarrow m^2} \, (q^2 - m^2) \, \int {\rm d}^4 x \,
e^{i q \cdot x}
\, \langle {\rm vac}|{\rm T}[\phi^{\prime}(x)
\phi^{\prime}(0)]|{\rm vac} \rangle \,
& = &\, i \label {renorm2}\end{aligned}$$
The field variable $\phi^{\prime}$ is as good a choice for the field variable as the original field $\phi$—its equation of motion is of the same form and it satisfies the correct renormalization conditions. It makes no difference to any [*physical*]{} amplitude whether one chooses to describe the physics in terms of the field $\phi$ or $\phi^{\prime}$. Thus, the masses of particles and possible bound states and S matrix elements for scattering states must be identical with either description. Going from one to the other amounts to nothing more than a change of variables.
While the physics clearly does not depend on which field is chosen, the propagator depends strongly on this choice: $$\begin{aligned}
\int {\rm d}^4 x \, e^{i q \cdot x}
\, \langle {\rm vac}| {\rm T}[\phi^{\prime}(x) \phi^{\prime}(0)]
|{\rm vac}\rangle \, \nonumber\\
= \int {\rm d}^4 x \, e^{i q \cdot x}
\, \langle {\rm vac}|{\rm T}[\phi(x) \phi(0)] \rangle +\int {\rm
d}^4 x \,
e^{i q \cdot x}
\, \langle {\rm T}[\phi(x) a(0)] + {\rm T}[a(x) \phi(0)] + {\rm
T}[a(x) a(0)]|{\rm vac}\rangle\end{aligned}$$ Eq.(\[renorm2\]), which picks out the pole at $q^2 = m^2$, is obtained since by construction $a$ does not connect the vacuum to the one $\phi$ state. Clearly, this is necessary since the correlation functions for $\phi$ and $\phi^{\prime}$ satisfy the renormalization conditions in eqs. (\[renorm\]) and (\[renorm2\]). Off shell, however, there is no requirement that this term vanish and the two propagators will in general differ. Moreover, since the overall scale of $a$ is arbitrary it is clear that one can make the difference between the two descriptions arbitrarily large.
Let us make these ideas explicit by considering two examples from a theory in which the current $j$ is a static external source. In this case the energy of the system is given by $$\begin{aligned}
E=\int {\rm d}^3r {1\over 2} j(\vec r)\phi(\vec r)\end{aligned}$$ or $$\begin{aligned}
E=\int {\rm d}^3r {1\over 2} j(\vec r) G(\vec r,\vec
r^{\,\prime})
j(\vec r^{\,\prime}),\end{aligned}$$ where $G(\vec r,\vec r^{\,\prime})$ is the inverse of the operator $\nabla^2-m^2$. Let us first take $a(x)$ to be a simple function of $\vec x$, which is independent of $\phi$. Then $\langle \phi,\vec p|a(x)|{\rm
vac}\rangle=0$ and the renormalization conditions of eqns. (\[renorm\],\[renorm2\]) are satisfied. One may determine a new Lagrangean density ${\cal
L}^\prime$ and a new Hamiltonian ${\cal H}^\prime$ by starting with the original ${\cal L}$ and transforming the variables. Then the new energy $E^\prime$ is given by $$\begin{aligned}
E^\prime=\int {\rm d}^3r {1\over 2} [\vec \nabla
(\phi^\prime(\vec r)+a(\vec r)\cdot(\phi^\prime(\vec r)+a(\vec
r))
+ \nonumber\\
m^2(\phi^\prime(\vec r)+a(\vec r))^2 +2
j(\vec r)(\phi^\prime(\vec r) +a(\vec r))],\end{aligned}$$ and using the equation of motion (\[eomp\]) in the static limit leads to $$\begin{aligned}
E^\prime=\int {\rm d}^3r {1\over 2}j(\vec r)(\phi^\prime(\vec r)
+a(\vec r)).\end{aligned}$$ But Eq.(\[def\]) tells us that $E^\prime=E$. Even though the current $j^\prime$ of Eq.(\[jpdef\]) is different than $j$ the energy of the system does not depend on the choice of the function $a(\vec x)$.
A more interesting example is obtained by letting $\phi=(1+f(\vec x))\phi^\prime $ (or $a(\vec x)= -{f(\vec x)\over 1+f(\vec x)} \phi(x)).$ We place the static source j at the origin and choose $f(\vec x)$ to vanish at large values of $|\vec x|$ faster than $e^{-m|\vec x|}/|\vec x|$. This maintains the original value of the renormalized coupling constant ( which is proportional to the asymptotic field) and therefore is the analog of our renormalization for problems with static sources. In this case the equation of motion is $$D\phi^\prime=-j^\prime
\label {neom}$$ where $$D\equiv
(1+f)^2 (-\nabla^2+m^2) -2(1+f)
\partial_\mu f \partial^\mu -(1+f)\nabla^2f$$ and $$j^\prime\equiv
j(1+f).$$ Clearly the Green’s function $G^\prime$ ( the inverse of $D$) and current $j^\prime$ are both fairly complicated. The use of the new Hamiltonian density ${\cal
H}^\prime $ gives $$\begin{aligned}
E^\prime=\int {\rm d}^3 r{1\over 2}[ (1+f)^2
\vec\nabla\phi^\prime\cdot\vec\nabla\phi^\prime \nonumber \\
+(\phi^\prime)^2 \vec\nabla f\cdot\vec\nabla f
+2\phi^\prime(1+f)\vec\nabla f\cdot\vec\nabla\phi^\prime+
(m\phi^\prime)^2(1+f)^2+2j\phi^\prime(1+f)].\end{aligned}$$ Integration by parts and the equation of motion Eq.(\[neom\]) allows one to obtain $$E^\prime \, = \, {1\over 2}\int {\rm d}^3 r
j(1+f)\phi^\prime \, =\,{1\over 2}\int {\rm d}^3 r j^\prime(\vec
r)G^\prime(\vec
r,\vec r^\prime)j^\prime(\vec r^\prime),$$ which is just the original energy since $(1+f)\phi^\prime=\phi$.
Thus we have seen two explicit examples in which transformations of field variables change the equation of motion, the Green’s functions and the currents without changing, the physical observable, the energy of the system.
These same arguments of Eqns. (5-10) can be used to show that the various n-point vertex functions of the also depend on the specific choice of field. The generalization of the argument to vector fields rather than scalars and to correlation functions of two different fields uses standard techniques. Again one finds that off-shell propagators and the vertex functions depend explicitly on the choice of field.
It is clear what is going on here. Neither the off-shell propagators nor the vertex functions are directly observable. From a theoretical point of view, the values of these quantities depend explicitly on which arbitrary choice of field one makes. Various combinations of the propagators and the vertex functions correspond to physical quantities and it is only these combinations which can be measured. Choosing a particular field amounts to making a bookkeeping choice—it only determines whether some bit of the physics will be found in the vertex or in the propagator.
The point we wish to stress is that knowledge, however precise, of the off-shell propagator contains no physical information unless one specifies the choice of the quantum field or equivalently unless one has knowledge of how the field couples to the rest of the system—[*i.e.*]{} knowledge of the vertex functions which arise from the same field choice. Thus, a model for the off-shell propagator in the absence of a [*consistent*]{} model for the vertex functions is not complete. The models of refs. [@GHT92]-[@MTRC94],[@IN94] present the mixed $\rho-\omega$ propagator off-shell, but do not give the necessary simultaneous consistent description of the CSB N-N-vector meson vertices.
Philosophy Of Meson Exchange Potentials
=======================================
The preceeding argument that off-shell meson propagators are not sufficient is entirely based on field theoretic considerations. Clearly, this does not help us to to compute observables, it does not address the question of how one can compute CSB (or any other) observables in nuclear physics. One typically constructs a nucleon-nucleon potential and then computes wavefunctions, hoping that the potentials capture the essential aspects of the underlying field theory. Nevertheless, there is no unambiguous way to construct potentials. Nontrivial assumptions must be made.
Here we will assume that the assumptions underlying phenomenologically successful meson exchange models are reasonable. While one can construct equally successful purely phenomenological models, the meson exchange models make a connection to the spectral properties of the underlying theory. Moreover, the entire question we are investigating—the role of $\rho-\omega$ mixing in CSB effects in nuclear physics—can only be addressed in the context of a potential model which employs vector mesons.
There is a definite philosophy underlying the construction of NN potentials from meson exchange. One principal idea is the need for a separation of momentum or length scales. One explicitly includes the exchange of mesons lighter (and hence more long-ranged) than some scale separation point. All short ranged effects are either incorporated in phenomenologically determined vertex functions or by some other purely phenomenological means. The physical picture underlying this philosophy is that the nucleon has a three-quark core which cannot be described efficiently in terms of mesons, while at longer distances the nucleon structure is dominated by a meson cloud.
To some extent, the fact that short ranged effects are handled as pure phenomenology is of little importance in most low energy nuclear physics applications. Because of repulsion at short distances, nuclear wave functions have strong short distance correlations which prevent the system from feeling the truly short range part of the potential. Moreover, at very short distances the concept of an NN potential becomes particularly inappropriate. Typically, in meson exchange potentials this scale separation point, which we will call $\Lambda_s$, is taken to be of order 1 GeV so that $\rho$ and $\omega$ mesons are explicitly included while heavier vector mesons are not. It is worth observing, however, that this does not mean that the short distance physics does not have important long range consequences. In particular, the value of the meson-nucleon coupling constant, determined by short distanced physics, plays an essential role in the potentials at long and intermediate ranges.
We believe that this general approach of treating the short range part of the NN interaction phenomenologically while explicitly including the effects of lighter mesons is reasonable. This general approach ought to be applicable to charge symmetry breaking effects.
There is another important assumption which underlies these models. It is assumed that at except at short distances the vector part of the potential is dominated by the vector mesons. Thus it is assumed that continuum two pion vector-isovector and three pion vector-isoscalar exchange contributions are small— [*i.e.*]{} that the only substantial strength arising from the two pion vector-isovector exchange is sufficiently concentrated at the $\rho$ mass as to be well described by $\rho$ exchange and analogously for three pions and the $\omega$ exchange. We note that this assumption can be questioned. In its favor we note that in $e^+ e^- \rightarrow$ pions, the $\rho$ and $\omega$ peaks do, in fact, completely dominate the low lying spectral function.
In our discussions we will adopt the Bonn potential [@BONN] strategy of incorporating all short range effects in vertex functions. In such a strategy the scale separation between long and short range is particularly easy to enforce: the phenomenological vertex functions are analytic for $q^2 < \Lambda_s^2$ while the propagators are analytic for $q^2 > \Lambda_s^2$, where $q^2$ is the square of the four momentum.
Momentum-Dependent Self-Energies in Meson-Exchange Potentials \[csc\]
======================================================================
It is probably useful to discuss an analogous, and perhaps somewhat simpler problem before discussing charge symmetry breaking. The $\rho-\omega$ mixing matrix element is an off-diagonal mass term. Models which give momentum dependence to this off-diagonal mass can also be expected to give momentum dependence to the analogous diagonal mass terms—[*i.e.*]{} to the vector meson self-energies. It is clearly useful to understand the role of the momentum dependence of the $\rho$ and $\omega$ self energies in the charge symmetry preserving potential before taking on the challenge of understanding the the momentum dependence of the $\rho-\omega$ mixing.
For simplicity we examine the one $\omega$ exchange contribution. First consider the traditional meson exchange model description with the scale separation as outlined above. The potential is given by $$V_{\omega} (q^2) = \frac{( g_\omega^{ \rm
v}(q^2)\gamma^{(1)}_{\mu} \,
+ \, g_\omega^{ \rm t}(q^2) q^\alpha \sigma^{(1)}_{\alpha \mu}
)
(g^{\mu \nu} - q^{\mu}q^{\nu}/m_{\omega}^2) (g_{\omega}^{\rm
v}(q^2)
\gamma^{(2)}_{\nu} + g_{\omega}^{\rm t}(q^2) \sigma_{\nu \beta}
q^\beta
) }{q^2 - m_\omega^2}
\label{oep}$$ where $g_{\omega}^{\rm v} (q^2)$ and $g_{\omega}^{\rm t}(q^2)$ are the vector and tensor couplings of the omega to the nucleons. The superscripts 1 and 2 label the nucleon. These couplings are analytic functions of $q^2$ for $q^2 < \Lambda_s^2$; the propagator is clearly analytic for $q^2 > \Lambda_s^2$.
In principle, we could consider a more sophisticated model consistent with the philosophy outlined above. For example, one could explicitly include the exchange of three low-energy pions (with the quantum numbers of the rho) along with an omega self-energy due to its coupling with the three pion channel and a longer-range part of the $\omega$-N vertex due to three pion exchange. In practice, one expects such effects to be small: in part they serve to simply widen the omega pole by an amount of no practical significance to the potential; other effects of coupling to the three pion channel are small because they are weakly coupled. In any event, we will stick to the conventional assumptions underlying meson exchange models and neglect such effects. In the remainder of this paper we will ignore such effects.
Let us now suppose that we had a detailed microscopic model of the $\omega$ meson which enables us to calculate a momentum dependent $\omega$ self energy, $\pi_{\omega}(q^2)$. As a matter of convention, we will include any effects of mass and wavefunction renormalizations of the $\omega$ in $\pi_{\omega}(q^2)$. This means that $\pi_{\omega}$ and its derivative vanishes at $q^2=m_\omega^2$. The omega exchange part of the N-N potential with such a model is given by $$V_{\omega}(q^2) = \frac{( \hat{g}_{\omega}^{\rm
v}(q^2)\gamma^{(1)}_{\mu} \, + \, \hat{g}_{\omega}^{\rm t}(q^2)
q^\alpha \sigma^{(1)}_{\alpha \mu} ) (g^{\mu \nu} -
q^{\mu}q^{\nu}/m_{\omega}^2) (\hat{g}_{\omega}^{\rm v}(q^2)
\gamma^{(2)}_{\nu} + \hat{g}_{\omega}^{\rm t}(q^2) \sigma_{\nu
\beta}
q^\beta ) }{q^2 - m_\omega^2 + \pi_\omega(q^2)}\;.
\label{oep2}$$ We have written the couplings as $\hat{g}_{\omega}^{\rm v,t}$ and rather than $g_{\omega}^{\rm v,t}$ to make evident the fact that the vertex functions used in the model in eq. (\[oep2\]) need not be the same as the vertex functions used in the model in eq. (\[oep\]): [*these vertex functions are phenomenological and depend on how the rest of the problem is treated*]{}.
Given that the vertex functions may differ between the two models, we note that the two models may be identical—[*i.e.*]{} they may be two equivalent ways of representing the same physics. One way for this to occur is if the vertex functions in the two models are related by $$g_\omega^{\rm v,t} \, = \, \hat{g}_\omega^{\rm v,t} \, \left
(\frac{q^2
- m_\omega^2}{q^2 - m_\omega^2 + \pi_{\omega}(q^2)} \right
)^{1/2}.
\label{condition}$$ Note that the square root factor is unity for $q^2=m_\rho^2$ due to the renormalization of $ \pi_{\omega}(q^2)$. The result (\[condition\]) is not surprising in light of the formal analysis of Sect.II. Neither the propagator off-shell nor the vertex function are separately meaningful.
Given that vertex functions are fit to some set of data, the only reason the condition in eq. (\[condition\]) would not be satisfied would be due to practical and philosophical limitations in the forms used in the fitting of the vertex functions. The practical limitation is that one must take some limited trial form for the phenomenological coupling. To the extent that meson exchange models make sense in the regime where they are used, the trial forms must be rich enough to describe the data with reasonable precision. Thus, apart from the philosophical concerns discussed below, eq. (\[condition\]) can be satisfied well enough so that any difference between the potentials of eqs. (\[oep\]) and (\[oep2\]) will have a small effect on the physics. The philosophical limitation is that the vertex functions are supposed to only contain effect of a range shorter than $\Lambda_s^{-1}$. Longer range effects are to be included by explicit dynamics of the lighter degrees of freedom in the problem.
Thus, the issue of whether the two models are equivalent comes down to whether both $g_{\rm v,t}(q^2)$ and $\hat{g}_{\rm v,t}(q^2)$ can be analytic for $q^2<\Lambda_s^2$ while satisfying eq. (\[condition\]). In effect, the question is whether $$f(q^2) = \left (\frac{q^2 - m_\omega^2}{q^2 - m_\omega^2 +
\pi_{\omega}(q^2)} \right )^{1/2}$$ is analytic for $q^2 < \Lambda_s^2$. Non-analyticity can occur when either the numerator or denominator vanishes or when $\pi_{\omega}$ is non-analytic. In fact, we should relax this restriction slightly—the non-analyticity associated with the $\omega$ coupling to three low energy pions which slightly broadens the pole and gives a small non-resonant contribution is, as discussed above, innocuous. In any event, this issue does not arise in the context of the models in refs. [@GHT92]-[@MTRC94].
Clearly, the analytic structure of $f(q^2)$ depends in detail on the choice of model. The simplest way to make the physics explicit is to make a spectral representation[@spectralref] for the propagator: $$\frac{1}{q^2 - m_\omega^2 + \pi_{\omega} (q^2)} = \int \, {\rm
d}s \,
\frac{\rho(s) }{q^2 - s + i \epsilon} + {\rm subtraction \quad
terms}.
\label{prop}$$ Different models will give rise to different spectral functions. However, if the model is realistic, the only substantial spectral strength for $q^2< \Lambda_s^2$ occurs at or near the omega pole. Accordingly any model which gives significant amounts of spectral strength below $\Lambda_s^2$ (apart from the $\omega$ pole), can be considered as unrealistic in our philosophy. If, however, all of the spectral strength is either at the $\omega$ pole or above $\Lambda_s^2$, then $f(q^2)$ is analytic for $q^2$ below $\Lambda_s^2$. The apparent non-analyticity due to the denominator vanishing at $q^2=
m_\omega^2$ is precisely canceled by a vanishing numerator. (Recall, all renormalization effects are included in $\pi_\omega$ so that the position of the $\omega$ pole does not shift.) In this case, one may re-define the vertex functions according to eq. (\[condition\]).
To see how the spectral representation constrains the allowable forms of the self energies consider consider the following simple example in which the self energy has the form
$$\pi_{\omega} (q^2)=(q^2-m_\omega^2)^2 Bq^2. \label{exam}$$
This form is motivated by the renormalization requirements, that $\pi_\omega(q^2)$ and its derivative vanish at $q^2=0$. One determines the nucleon-nucleon potential generated by the propagator of Eq.(\[prop\]) by taking $q^2$ to be space-like $q^2=-Q^2<0$; the potential is proportional to the integral $$\int {\rm d}Q Q{\sin (Qr)\over r}
\frac{1}{-Q^2 - m_\omega^2 -(Q^2+m_\omega^2)^2 BQ^2 }.$$ One does the contour integration by identifying the poles. There is always a pole at $Q^2=-m_\omega^2$ which is the standard term expected from the exchange of an $\omega^2$-meson. There are other poles at positions determined by the value of B. One finds that if $ m_\omega^4> m_\omega^4+4/B\ge0$ there will be poles with $Q^2$ real and negative. At least one of the poles must be at $|Q^2|<m_\omega^2$, which is unrealistic in our philosophy. If $m_\omega^4+4/B > m_\omega^4 $, the poles occur for $Q^2>0$ which are physically un-allowable tachyonic excitations. Similarly, if $m_\omega^4+4/B<0$, there are poles off the real axis, which violates the spectral representation and also renders the model for $ \pi_{\omega}
(q^2)$ as useless. This analysis demonstrates that a spectral function of the form in Eq.(\[exam\]) is not viable.
Let us now summarize the effects of the momentum dependence of the $\omega$ self energy on the meson exchange potential. In any realistic model, ([*i.e.*]{} any model without unphysical low $q^2$ spectral strength in the $\omega$ propagator) all of the effects of the momentum dependence of the self-energy can be re-absorbed into momentum dependence of the phenomenological vertex functions. Accordingly, there are no observable physical effects in the NN potential induced by such a momentum dependent self-energy. Moreover, including the short range part of the momentum dependence in the propagator of a meson exchange model violates the bookkeeping arrangement in which all of the short range effects are segregated into phenomenological vertices.
The Charge Symmetry Breaking NN Potential And The Momentum Dependence of $\rho -\omega$ Mixing
==============================================================================================
The preceeding section gives us a paradigm for what happens in the charge symmetry breaking part of the potential. We will show, for any realistic model of the momentum dependence of the mixing amplitude, that all of the effects of the momentum dependence can be absorbed into phenomenological short ranged charge-symmetry-breaking nucleon-vector meson couplings.
Consider the charge-symmetry-breaking potential arising from vector meson exchange. Let us begin by implementing this according to the philosophy of scale separation discussed in the previous two sections. Assuming that only the meson exchanges we need to consider are the $\rho$ and $\omega$, the charge symmetry breaking interaction potential can be written as $$\begin{aligned}
V_{\omega,\rho}^{\rm CSB} (q^2) \, =
\, \frac{[g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(1)} +
g_\omega^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(1)}] [g^{\mu \nu} -
q^{\mu}q^{\nu}/m_{\omega}^2] [g_\omega^{\rm v \, CSB}(q^2)
\tau_3^{(2)}
\gamma_{\nu}^{(2)} + g_\omega^{\rm t \, CSB}(q^2) \tau_3^{(2)}
\sigma_{\nu \beta}^{(2)} q^\beta] }{q^2 - m_\omega^2} \nonumber
\\ \nonumber \\
+ \, \frac{ [g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(2)} +
g_\omega^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(2)}] [g^{\mu \nu} -
q^{\mu}q^{\nu}/m_{\omega}^2] [g_\omega^{\rm v \, CSB}(q^2)
\tau_3^{(1)} \gamma_{\nu}^{(1)} + g_\omega^{\rm t \, CSB}(q^2)
\tau_3^{(1)} \sigma_{\nu \beta}^{(1)} q^\beta] }{q^2 -
m_\omega^2}
\nonumber \\ \nonumber \\
+ \, \frac{
[g_\rho^{\rm v} (q^2) \gamma_{\mu}^{(1)} \tau_3^{(1)} +
g_\rho^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(1)} \tau_3^{(1)}][g^{\mu
\nu} -
q^{\mu}q^{\nu}/m_{\rho}^2][g_\rho^{\rm v \, CSB}(q^2)
\gamma_{\nu}^{(2)} + g_\rho^{\rm t \, CSB}(q^2) \sigma_{\nu
\beta}^{(2)} q^\beta] }{q^2 - m_\rho^2} \nonumber \\ \nonumber
\\
+ \, \frac{
[g_\rho^{\rm v} (q^2) \tau_3^{(2)} \gamma_{\mu}^{(2)} +
g_\rho^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(2)} \tau_3^{(2)}] [g^{\mu
\nu} -
q^{\mu}q^{\nu}/m_{\rho}^2][g_\rho^{\rm v \, CSB}(q^2)
\gamma_{\nu}^{(1)} + g_\rho^{\rm t \, CSB}(q^2) \sigma_{\nu
\beta}^{(1)} q^\beta] }{q^2 - m_\rho^2} \nonumber \\ \nonumber
\\
\, + \, m^2_{\rho \omega} \, \, \frac{[g_\omega^{\rm v}(q^2)
\gamma_{\mu}^{(1)} + g_\omega^{\rm t}(q^2) q^\alpha
\sigma_{\alpha \mu}^{(1)}]
[g^{\mu \gamma} - q^{\mu}q^{\gamma}/m_{\omega}^2]
[g^{\gamma \nu} - q^{\gamma}q^{\nu}/m_{\rho}^2]
[g_\rho^{\rm v}(q^2) \tau_3^{(2)}\gamma_{\nu}^{(2)} +
g_\rho^{\rm
t}(q^2) \tau_3^{(2)} \sigma_{\nu \beta}^{(2)} q^\beta]
}{(q^2-m_\omega^2)(q^2-m_\rho^2)} \nonumber \\ \nonumber \\
+ \, m^2_{\rho \omega} \, \, \frac{
[g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(2)} + g_\omega^{\rm
t}(q^2)
q^\alpha \sigma_{\alpha \mu}^{(2)}]
[g^{\mu \gamma} - q^{\mu}q^{\gamma}/m_{\omega}^2]
[g^{\gamma \nu} - q^{\gamma}q^{\nu}/m_{\rho}^2]
[g_\rho^{\rm v}(q^2) \tau_3^{(1)}\gamma_{\nu}^{(1)} +
g_\rho^{\rm
t}(q^2) \tau_3^{(1)} \sigma_{\nu \beta}^{(1)} q^\beta] }
{(q^2-m_\omega^2)(q^2-m_\rho^2)} \nonumber\ \\ \label{CSB1}\end{aligned}$$ This form is rather general: in addition to $\rho-\omega$ mixing, it explicitly includes possible charge symmetry breaking couplings between the vector mesons and the nucleons arising from short distance effects: these couplings are labeled by the superscript CSB. The coefficient $m^2_{\rho \omega}$ is the mixing parameter which in this model is taken to be independent of $q^2$.
It should be noted that the general form of eq. (\[CSB1\]) is consistent with the general philosophy of meson exchange used here. In particular all short-ranged effects are merely parameterized, while the long ranged effects are treated dynamically in terms of the mesons. It is for this reason, that we must include the $\rho-\omega$ mixing explicitly rather than including all of the effects in terms of the charge symmetry breaking couplings.
The couplings $g_{\omega,\rho}^{\rm v,t } (q^2)$ are presumed to have been determined in fits to the charge symmetry conserving interactions. In principle, the coupling constants $g_{\omega,\rho}^{\rm v,t
\, CSB}
(q^2)$ must be determined phenomenologically from experimental data on charge symmetry breaking. In fact, in the treatments of CSB in refs. [@CB87; @BI87; @M94; @MVO94] these couplings were all taken to be zero. In that work, model assumptions and existing nucleon-nucleon and pion-nucleon scattering data were used to make [*a priori*]{} arguments that these couplings should be small and hence could be neglected. See, for example, Refs. [@MNS90] and [@MVO94] which reviews the charge-dependence of the couplings. The neglect of charge dependence in the meson-nucleon coupling constants is not invalidated by present data. In particular, descriptions of all known CSB effects do not require the inclusion of such terms. Had the data required the inclusion of such terms they could have been included without violating the spirit of a meson exchange potential model.
Now suppose, we had a detailed model for the structure of the vector mesons in which the $\rho-\omega$ mixing amplitude has a nontrivial momentum dependence. The form for the CSB potential is very similar to the form above: $$\begin{aligned}
V_{\omega,\rho}^{\rm CSB} (q^2) =
\frac{[g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(1)} +
g_\omega^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(1)}] [g^{\mu \nu} -
q^{\mu}q^{\nu}/m_{\omega}^2] [\hat{g}_\omega^{\rm v \, CSB}(q^2)
\tau_3^{(2)} \gamma_{\nu}^{(2)} + \hat{g}_\omega^{\rm t \,
CSB}(q^2)
\tau_3^{(2)} \sigma_{\nu \beta}^{(2)} q^\beta] }{q^2 -
m_\omega^2}
\nonumber \\ \nonumber \\
+ \frac{ [g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(2)} +
g_\omega^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(2)}] [g^{\mu \nu} -
q^{\mu}q^{\nu}/m_{\omega}^2] [\hat{g}_\omega^{\rm v \, CSB}(q^2)
\tau_3^{(1)} \gamma_{\nu}^{(1)} + \hat{g}_\omega^{\rm t \,
CSB}(q^2)
\tau_3^{(1)} \sigma_{\nu \beta}^{(1)} q^\beta] }{q^2 -
m_\omega^2}
\nonumber \\ \nonumber \\
+ \frac{
[g_\rho^{\rm v} (q^2) \gamma_{\mu}^{(1)} \tau_3^{(1)} +
g_\rho^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(1)} \tau_3^{(1)}][g^{\mu
\nu} -
q^{\mu}q^{\nu}/m_{\rho}^2][\hat{g}_\rho^{\rm v \, CSB}(q^2)
\gamma_{\nu}^{(2)} + \hat{g} _\rho^{\rm t \, CSB}(q^2)
\sigma_{\nu
\beta}^{(2)} q^\beta] }{q^2 - m_\rho^2} \nonumber \\ \nonumber
\\
+ \frac{
[g_\rho^{\rm v} (q^2) \tau_3^{(2)} \gamma_{\mu}^{(2)} +
g_\rho^{\rm
t}(q^2) q^\alpha \sigma_{\alpha \mu}^{(2)} \tau_3^{(2)}] [g^{\mu
\nu} -
q^{\mu}q^{\nu}/m_{\rho}^2][\hat{g}_\rho^{\rm v \, CSB}(q^2)
\gamma_{\nu}^{(1)} + \hat{g}_\rho^{\rm t \, CSB}(q^2)
\sigma_{\nu
\beta}^{(1)} q^\beta] }{q^2 - m_\rho^2} \nonumber \\ \nonumber
\\
+ \, m^2_{\rho \omega}(q^2) \, \, \frac{[g_\omega^{\rm v}(q^2)
\gamma_{\mu}^{(1)} + g_\omega^{\rm t}(q^2) q^\alpha
\sigma_{\alpha \mu}^{(1)}]
[g^{\mu \gamma} - q^{\mu}q^{\gamma}/m_{\omega}^2]
[g^{\gamma \nu} - q^{\gamma}q^{\nu}/m_{\rho}^2]
[g_\rho^{\rm v}(q^2) \tau_3^{(2)}\gamma_{\nu}^{(2)} +
g_\rho^{\rm
t}(q^2) \tau_3^{(2)} \sigma_{\nu \beta}^{(2)} q^\beta]
}{(q^2-m_\omega^2)(q^2-m_\rho^2)} \nonumber \\ \nonumber \\
+ m^2_{\rho \omega}(q^2) \, \, \frac{
[g_\omega^{\rm v}(q^2) \gamma_{\mu}^{(2)} + g_\omega^{\rm
t}(q^2)
q^\alpha \sigma_{\alpha \mu}^{(2)}]
[g^{\mu \gamma} - q^{\mu}q^{\gamma}/m_{\omega}^2]
[g^{\gamma \nu} - q^{\gamma}q^{\nu}/m_{\rho}^2]
[g_\rho^{\rm v}(q^2) \tau_3^{(1)}\gamma_{\nu}^{(1)} +
g_\rho^{\rm
t}(q^2) \tau_3^{(1)} \sigma_{\nu \beta}^{(1)} q^\beta] }
{(q^2-m_\omega^2)(q^2-m_\rho^2)}\nonumber\end{aligned}$$ $$\label{CSB2}$$ We have labeled the CSB coupling as $\hat{g}_{\omega,\rho}^{\rm v,t \,CSB}$ rather than $g_{\omega,\rho}^{\rm v,t \, CSB}$ to make explicit the fact the CSB couplings in eq. (\[CSB2\]) may be different from the CSB couplings in eq. (\[CSB1\]).
The question we wish to address is whether the model in eq. (\[CSB2\]) is equivalent to the model in eq. (\[CSB1\]). The issue comes down to whether the effects of the momentum dependence of the mixing can be entirely absorbed into differences between $\hat{g}_{\omega,\rho}^{\rm
v,t \,
CSB}$ and $g_{\omega,\rho}^{\rm v,t \, CSB}$ without introducing any unnaturally long range effects into the CSB couplings. We shall show that this can be done.
The $\rho-\omega$ mixing is measured rather accurately at the pole at $q^2=m_\omega^2$. Accordingly, it is sensible to express $$m^2_{\rho \omega}(q^2) = m^2_{\rho \omega} + \delta m^2_{\rho
\omega}(q^2)$$ with $\delta m^2_{\rho \omega}(m_{\omega}^2) = 0$. Thus, the expression $$\frac{ \delta m^2_{\rho
\omega}(q^2)}{(q^2-m_\rho^2)(q^2-m_\omega^2)}$$ has no $\omega$ pole. All effects with this term are indistinguishable from terms arising due $\rho$ exchange with a CSB vertex. In particular, if $$\begin{aligned}
\hat{g}_{\omega}^{\rm v,t CSB} & = & g_{\omega}^{\rm v,t CSB}
\label{CSB3}\\
\hat{g}_{\rho}^{\rm v,t CSB} & = & g_{\rho}^{\rm v,t CSB} -
\frac{
\delta m^2_{\rho \omega}(q^2)}{q^2-m_\omega^2} g_{\rho}^{\rm v,t}
\label{CSB4}\end{aligned}$$ then the potential in eq. (\[CSB2\]) is identical with the one of eq. (\[CSB1\]). This result can also be obtained from Feynman diagrams. Let an $\omega$ be emitted from a nucleon and then be converted via $m^2_{\rho \omega}(q^2)$ into a $\rho$. One can draw a box which includes the strong vertex and $m^2_{\rho \omega}(q^2)$. This box is the charge-dependent $\rho$-nucleon coupling constant. Alternatively one can regard the $m^2_{\rho \omega}(q^2)$ as part of the propagator. Either way, the result is the same.
We can do a specific calculation. For example, suppose $\delta m^2_{\rho
\omega}(q^2)=m^2_{\rho\omega}/m_\omega^2
(q^2-m_\omega^2)$. This is a good approximation to the $m^2_{\rho
\omega}(q^2)$ obtained in the sum rule work of Ref.[@HHKM94]. Then the difference between $\hat{g}_{\omega}^{\rm v,t CSB}$ and $ g_{\rho}^{\rm v,t CSB}$ is a simple constant $\approx -.008g_\rho^{v,t}$; if $ \hat{g}_{\rho}^{\rm v,t CSB}$ were chosen as the negative of that constant, one would obtain the standard form of the $\rho-\omega$ mixing contribution to the NN potential. See also Ref.[@G95].
Moreover, for any reasonable model of the momentum dependence of the mixing, eq. (\[CSB4\]) can be satisfied without introducing unnaturally long ranged effects into the meson-nucleon vertex functions. The issues are completely analogous to the ones raised in connection with the $\omega$ exchange potential discussed in the previous section. First, it should be noted that there is no $\omega$ pole singularity on the right hand side of eq. (\[CSB4\])—it is eliminated because $\delta
m_{\rho \omega}^2$ vanishes at the $\omega$ pole. Thus, the only source of long range contamination of the couplings is in $\delta
m_{\rho \omega}^2 (q^2)$ itself. Note, that by construction $\delta
m_{\rho \omega}^2 (q^2)$ cannot have a singularity associated with either the $\rho$ or the $\omega$. Moreover, we know that the only substantial strength in the vector channels at $q^2 < \Lambda_s^2$ is through the $\rho$ and $\omega$ mesons. Thus, any model which yields long range effects in $\delta m_{\rho \omega}^2
(q^2)$ must be regarded as unrealistic according to our philosophy.
Summary
=======
We are working in the framework of boson exchange potentials. This means that in realistic boson-exchange models long range effects are included via boson exchanges and that short range effects are included in the vertex functions. For any such realistic model of the momentum dependence of the $\rho-\omega$ mixing parameter, there are no effects in the CSB breaking potential which cannot be absorbed into a redefinition of a short ranged CSB $\rho$-N vertex. Thus, a model which provides knowledge of the momentum dependence of the mixing parameter alone, without simultaneously giving a self-consistent model for the short ranged CSB vector meson-nucleon couplings gives no information about the CSB N-N potential.
The work of refs. [@GHT92]-[@MTRC94] found major differences between the CSB potentials based on the on-shell $\rho-\omega$ mixing and models with a large momentum dependence. Our boson exchange model view is that this is because the short ranged CSB vector meson-nucleon coupling are assumed to be zero— just as in the models based on the on-shell $\rho-\omega$ mixing. However, there is no reason, [*a priori*]{} that this assumption is be true for the models under discussion. Indeed, there is [*a posteriori*]{} evidence that the assumption may be wrong: the models based on the on-shell $\rho-\omega$ mixing and negligible $g_{\rho \omega}^{\rm v,t \, CSB}$ reproduce the available data with reasonable accuracy.
One of us (TDC) thanks the Department of Physics and the Institute of Nuclear Theory at the University Of Washington for its hospitality. The authors thank the U.S. Department of Energy for supporting this work; TDC also acknowledges the financial support of the National Science Foundation.
[99]{}
J.A. Nolen and J.P. Schiffer, [*Ann. Rev. Nucl. Sci.*]{} [**19**]{} (1969) 471.
E.M. Henley, in [*Isospin in Nuclear Physics*]{}, ed. D.H. Wilkinson (North-Holland, Amsterdam, 1969) 17.
S. Shlomo, [*Rep. Prog. Phys.*]{} [**41**]{} (1978) 957.
E.M. Henley and G.A. Miller in [*Mesons in Nuclei*]{}, ed. M. Rho and D.H. Wilkinson (North-Holland, 1979) 405.
G.A. Miller, B.M.K. Nefkens, and I. $\check{S}$laus, [*Phys. Repts.*]{} [**194**]{} (1990) 1.
B. M.K. Nefkens, G.A. Miller and I. $\check{S}$laus, [*Comments on Nucl. Part. Phys.*]{} [**20**]{} (1992) 221.
W. T. H. Van Oers and G. A. Miller “Charge independence and charge symmetry", to be published in “Fundamental Symmetries" edited by W. C. Haxton and E. M. Henley.
A. Quenzer, [*Phys. Lett.*]{} [**76B**]{} (1978) 512.
L.M. Barkov [*Nucl. Phys.*]{} [**B256**]{} (1985) 365. G.A. Miller, [*Charge Independence and Charge Symmetry Breaking*]{}, talk presented at [*The First International Symposium on Symmetries in Subatomic Physics*]{}, May 1994, Taipei and U. of Washington preprint 1994 DOE/ER/40427-09-N94, NUCLTH-9406023, to be published in the conference proceedings, Chin. J. Physics. This will also appear in the proceedings of the 1994 Symposium on High Energy Spin Physics and Polarization Phenomena in Nuclear Physics.
S.A. Coon and R.C. Barrett, [*Phys. Rev.*]{} [**C36**]{} (1987) 2189.
P.G. Blunden and M.J. Iqbal, [*Phys. Lett.*]{} [**B198**]{} (1987) 14.
R. Abegg [*Phys. Rev. Lett.*]{} [**56**]{} (1986) 2571; Phys. Rev. [**D39**]{} (1989) 2464.
, spokespersons L.G. Greeniaus and W.T.H. van Oers. Also spin 94
L.D. Knutson [*Phys. Rev. Lett.*]{} [**66**]{} (1991) 1410;S.E. Vigdor [*Phys. Rev.*]{} [**C46**]{} (1992) 410.
T. Goldman, J.A. Henderson, and A.W. Thomas, [*Few Body Systems*]{} [**12**]{} (1992) 193.
J. Piekarewicz and A.G. Williams, [*Phys. Rev.*]{} [**C47**]{} (1993) R2462.
G. Krein, A.W. Thomas, and A.G. Williams, [*Phys. Lett.*]{} [**B317**]{} (1993) 293.
K.L. Mitchell, P.C. Tandy, C.D. Roberts, and R.T. Cahill, 1994 preprint, [*Phys. Lett.*]{} [**B**]{} in press.
H.B. O‘Connell, B.C. Pierce, A.W. Thomas and A.G. Williams, [*Constraints on the momentum dependence of rho-omega mixing*]{} 1994 preprint, to appear in Phys. Lett
T. Hatsuda, E.M. Henley, Th. Meissner, and G. Krein, [*Phys. Rev.*]{} [**C49**]{}(1994) 452.
M.J. Iqbal, and J.A. Niskanen, [*Phys. Lett.*]{} [**B322**]{} (1994) 7.
R. Haag, [*Phys. Rev.* ]{}[**112**]{}, 669 (1958). D. Ruelle, [*Helv. Phys. Act.*]{} [**35**]{} 34 (1962). H.J. Borchers, Nuovo Cimento [**25**]{} 270 (1960).
S. Coleman, J. Wess and B. Zumino, [*Phys. Rev.*]{} [**177**]{} 2239 (1969).
R. Machleidt, K. Holinde and Ch. Elster [*Phys. Reports*]{} [**149**]{} 1 (1987).
J. D. Bjorken and S. D. Drell [*Relativistic Quantum Fields*]{} Mc Graw Hill NY 1965
S. Gardner, C.J. Horowitz,and J. Piekarewicz, “Charge Symmetry Breaking Potentials From Isospin Violating Meson-Baryon Coupling Constants", FSU-SCRI-95-37, May 1995. Submitted to Phys. Rev. Lett. e-Print Archive: nucl-th/9505001
[^1]: Permanent address: Department of Physics and Center for Theoretical Physics, University of Maryland, College Park, MD 20742, USA
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider a family of equivalent norms (called operator *E*-norms) on the algebra $\B(\H)$ of all bounded operators on a separable Hilbert space $\H$ induced by a positive densely defined operator $G$ on $\H$. By choosing different generating operator $G$ one can obtain the operator *E*-norms producing different topologies, in particular, the strong operator topology on bounded subsets of $\B(\H)$.
We obtain a generalised version of the Kretschmann-Schlingemann-Werner theorem, which shows continuity of the Stinespring representation of CP linear maps w.r.t. the energy-constrained $cb$-norm (diamond norm) on the set of CP linear maps and the operator *E*-norm on the set of Stinespring operators.
The operator *E*-norms induced by a positive operator $G$ are well defined for linear operators relatively bounded w.r.t. the operator $\sqrt{G}$ and the linear space of such operators equipped with any of these norms is a Banach space. We obtain explicit relations between the operator *E*-norms and the standard characteristics of $\sqrt{G}$-bounded operators. The operator *E*-norms allow to obtain simple upper estimates and continuity bounds for some functions depending on $\sqrt{G}$-bounded operators used in applications.
author:
- 'M.E. Shirokov[^1]'
title: 'Operator *E*-norms and their use'
---
Introduction
============
The algebra $\B(\H)$ of all bounded linear operators on a separable Hilbert space $\H$, some its subalgebras and subspaces are basic objects in different fields of modern mathematics and mathematical physics . In particular, $\B(\H)$ appears as an algebra of observables in the theory of quantum systems while unital completely positive maps between such algebras called quantum channels play the role of dynamical maps in the Heisenberg picture [@H-SCI; @Watrous; @Wilde].
The variety of different topologies on $\B(\H)$, relations between them and their “physical” sense are well known for anybody who is interested in functional analysis, theory of operator algebras, mathematical and theoretical physics.
In this article we describe families of norms on $\B(\H)$ producing different topologies on $\B(\H)$, in particular, the strong operator topology on bounded subsets of $\B(\H)$. These norms depending on a positive densely defined operator ${G}$ and a positive parameter $E$ were introduced in [@CSR] for quantitative analysis of continuity of the Stinespring representation of a quantum channel with respect to the strong convergence of quantum channels and the strong operator convergence of Stinespring isometries.[^2]
Now we consider these norms (called the operator *E*-norms) in more general context (assuming that ${G}$ is an arbitrary positive operator). In Section 3 we consider equivalent definitions and basic properties of the operator *E*-norms. We obtain explicit relations between the operator *E*-norms and the equivalent norm on $\B(\H)$ also induced by a positive operator $G$ (which is commonly used in analysis of $\sqrt{G}$-bounded operators).
The operator *E*-norms make it possible to obtain a generalization the Kretschmann-Schlingemann-Werner theorem. The original version of this theorem presented in shows continuity of the Stinespring representation of a completely positive (CP) linear map with respect to the norm of complete boundedness ($cb$-norm in what follows)[^3] on the set of CP maps and the operator norm on the set of Stinespring operators. Our aim was to obtain a version of this theorem for other (weaker) topologies on the sets of CP maps and corresponding Stinespring operators, in particular, for the strong convergence topology on the set of CP maps and the strong operator topology on the set of Stinespring operators. By using the operator *E*-norms one can upgrade the proof of the Kretschmann-Schlingemann-Werner theorem without essential changes. The generalised version of this theorem and its corollaries are presented in Section 4.
In Section 5 the operator *E*-norms induced by a positive operator $G$ are extended to linear operators relatively bounded w.r.t. the operator $\sqrt{G}$. We prove that the linear space of such operators equipped with any of these norms is a Banach space. Its subspace consisting of all operators with zero $\sqrt{G}$-bound is the completion of $\B(\H)$ w.r.t. any of the operator *E*-norms. We obtain explicit relations between the operator*E*-norms and the standard characteristics of $\sqrt{G}$-bounded operators.
The operator *E*-norms allow to obtain simple upper estimates and continuity bounds for some functions depending on $\sqrt{G}$-bounded operators used in applications.
As a basic example we consider the operators associated with the Heisenberg Commutation Relation.
Preliminaries
=============
Let $\mathcal{H}$ be a separable infinite-dimensional Hilbert space, $\mathfrak{B}(\mathcal{H})$ – the algebra of all bounded operators on $\mathcal{H}$ with the operator norm $\|\!\cdot\!\|$ and $\mathfrak{T}(\mathcal{H})$ – the Banach space of all trace-class operators on $\mathcal{H}$ with the trace norm $\|\!\cdot\!\|_1$ (the Schatten class of order 1) . Let $\mathfrak{T}_{+}(\mathcal{H})$ be the cone of positive operators in $\mathfrak{T}(\mathcal{H})$. Trace-class operators will be usually denoted by the Greek letters $\rho $, $\sigma $, $\omega $, ... The closed convex subsets $$\T_{+,1}(\H)=\{\rho\in \T_+(\H)\,|\, {\mathrm{Tr}}\rho\leq 1\}\quad \textrm{and} \quad \S(\H)=\{\rho\in \T_+(\H)\,|\, {\mathrm{Tr}}\rho=1\}$$ of the cone $\T_+(\H)$ are complete separable metric spaces with the metric defined by the trace norm. Operators in $\S(\H)$ are called density operators or *states*, since any $\rho$ in $\S(\H)$ determines a normal state $A\mapsto {\mathrm{Tr}}A\rho$ on the algebra $\B(\H)$ . Extreme points of $\S(\H)$ are 1-rank projectors called *pure states*.
Denote by $I_{\H}$ the unit operator on a Hilbert space $\mathcal{H}$ and by ${\mathrm{Id}}_{\mathcal{\H}}$ the identity transformation of the Banach space $\mathfrak{T}(\mathcal{H})$.
We will use the Dirac notations $|\varphi\rangle$, $|\psi\rangle\langle\varphi|$,... for vectors and operators of rank 1 on a Hilbert space (in this notations the action of an operator $|\psi\rangle\langle\varphi|$ on a vector $|\chi\rangle$ gives the vector $\langle\varphi|\chi\rangle\langle\psi|$) [@H-SCI].
We will pay a special attention to the class of unbounded densely defined positive operators on $\H$ having discrete spectrum of finite multiplicity. In Dirac’s notations any such operator ${G}$ can be represented as follows $$\label{H-rep}
G=\sum_{k=0}^{+\infty} E_k|\tau_k\rangle\langle \tau_k|$$ on the domain $\mathcal{D}(G)=\{ \varphi\in\H\,|\,\sum_{k=0}^{+\infty} E^2_k|\langle\tau_k|\varphi\rangle|^2<+\infty\}$, where $\left\{\tau_k\right\}_{k=0}^{+\infty}$ is the orthonormal basis of eigenvectors of ${G}$ corresponding to the nondecreasing sequence $\left\{\smash{E_k}\right\}_{k=0}^{+\infty}$ of eigenvalues tending to $+\infty$. We will use the following (cf.[@W-EBN])
\[D-H\] An operator ${G}$ having representation (\[H-rep\]) is called *discrete*.
The set $\S(\H)$ is compact if and only if $\dim \H<+\infty$. We will use the following
\[Comp\] [@H-c-w-c] *If ${G}$ is a discrete unbounded operator on $\H$ then the set of states $\rho$ in $\S(\H)$ satisfying the inequality ${\mathrm{Tr}}{G}\rho\leq E$ is compact for any $E\geq\inf_{\|\varphi\|=1}\langle\varphi|{G}|\varphi\rangle$.*
We will also use the following simple lemma.
\[WL\] [@W-CB] *If $f$ is a concave nonnegative function on $[0,+\infty)$ then for any positive $x< y$ and any $z\geq0$ the inequality $\,xf(z/x)\leq yf(z/y)\,$ holds.*
Operator *E*-norms on $\B(\H)$
==============================
Let ${G}$ be a positive semidefinite operator on $\H$ with a dense domain $\mathcal{D}({G})$ such that $$\label{H-cond}
\inf\left\{{\hspace{1pt}}\|G\varphi\|\,|\,\varphi\in\mathcal{D}({G}),\|\varphi\|=1{\hspace{1pt}}\right\}=0.$$ We will assume that for any positive operator $\rho$ in $\T(\H)$ the value of ${\mathrm{Tr}}{G}\rho$ (finite or infinite) is defined as $\sup_n{\mathrm{Tr}}P_nG\rho$, where $P_n$ is the spectral projector of $G$ corresponding to the interval $[0,n]$.
For given $E>0$ consider the function on $\B(\H)$ defined as $$\label{ec-on-b}
\|A\|^G_E\doteq \sup_{\varphi\in\H_1, \langle\varphi|G|\varphi\rangle\leq E}\|A\varphi\|,$$ where $\H_1$ is the unit sphere in $\H$ and it is assumed that $\,\langle\varphi|G|\varphi\rangle=\|\sqrt{G}\varphi\|^2\,$ if $\varphi$ lies in $\D(\sqrt{G})$ and $\,\langle\varphi|G|\varphi\rangle=+\infty\,$ otherwise. This function can be also defined as $$\label{ec-on}
\|A\|^{G}_E\doteq \sup_{\substack{\rho\in\mathfrak{S}(\mathcal{H}):
{\mathrm{Tr}}{G}\rho\leq E}}\sqrt{{\mathrm{Tr}}A\rho A^*},$$ where the supremum is over all states $\rho$ in $\S(\H)$ satisfying the inequality ${\mathrm{Tr}}{G}\rho\leq E$.[^4]
The coincidence of the r.h.s. of (\[ec-on-b\]) and (\[ec-on\]) for any $A\in\B(\H)$ is shown in .
It is easy to see that the function $A\mapsto\|A\|^{G}_E$ is a norm on $\B(\H)$. Definition (\[ec-on-b\]) shows the sense of the norm $\|\cdot\|^G_E$ (as a constrained version of the operator norm $\|\cdot\|$) while definition (\[ec-on\]) is more convenient for studying its analytical properties. In particular, by using definition (\[ec-on\]) the following proposition is proved in [@CSR].[^5]
\[ec-on-p1\] *For any operator $\,A\in\B(\H)$ the following properties hold:*
a) *$\|A\|^{G}_E$ tends to $\|A\|$ as $E\rightarrow+\infty$;*
b) *the function $E\mapsto\left[\|A\|^{G}_E\right]^p$ is concave and nondecreasing on $\,\mathbb{R}_+$ for $p\in(0,2]$;*
c) *$\|A\varphi\|\leq K_{\varphi}\|A\|^{G}_E$ for any unit vector $\varphi$ in $\D(\sqrt{G})$, where $K_{\varphi}=\max\{1, \|\sqrt{G}\varphi\|/\sqrt{E}\}$.*
We will call the norms $\|\!\cdot\!\|^{G}_E$ the *operator E-norms* on $\B(\H)$.[^6] Property b) in Proposition \[ec-on-p1\] shows that $$\label{E-n-eq}
\|A\|^{G}_{E_1}\leq \|A\|^{G}_{E_2}\leq \sqrt{E_2/E_1}\|A\|^{G}_{E_1}\quad\textrm{ for any } E_2>E_1>0.$$ Hence for given operator ${G}$ all the norms $\|\!\cdot\!\|^{G}_{E}$, $E>0$, are equivalent on $\B(\H)$.
\[en-gen\] The definition of the operator *E*-norm is obviously generalized to operators between different Hilbert spaces $\H$ and $\K$. It is easy to see that all the above and below results concerning properties of the operator *E*-norms remain valid (with obvious modifications) for this generalization. $\square$
Since the set $\D(\sqrt{G})$ is dense in $\H$, property c) in Proposition \[ec-on-p1\] shows that the topology generated by any of the norms $\|\!\cdot\!\|^{G}_{E}$ on bounded subsets of $\B(\H)$ is not weaker than the strong operator topology. On the other hand, it is not stronger than the norm topology on $\B(\H)$. The following proposition characterizes these extreme cases.
\[ec-on-p2\] A) *The norm $\,\|\!\cdot\!\|^{G}_E$, $E>0$, is equivalent to the operator norm $\,\|\!\cdot\!\|$ on $\B(\H)$ if and only if the operator ${G}$ is bounded.*
B\) *The norm $\,\|\!\cdot\!\|^{G}_E$, $E>0$, generates the strong operator topology on bounded subsets of $\B(\H)$ if and only if $\,{G}$ is an unbounded discrete operator (Definition \[D-H\]).*
*Proof.* A) If ${G}$ is a bounded operator then $\,\|\!\cdot\!\|^{G}_E=\|\!\cdot\!\|$ for any $E\geq \|G\|$.
If ${G}$ is a unbounded operator and $P_n$ is the spectral projector of ${G}$ corresponding to the interval $[n,+\infty)$ then $\|P_n\|=1$ for all $n$. By noting that ${\mathrm{Tr}}P_n\rho\leq E/n$ for any state $\rho$ such that ${\mathrm{Tr}}{G}\rho\leq E$, it is easy to see that $\,\|P_n\|^{G}_E\rightarrow 0$ as $n\rightarrow+\infty$.
B\) The “if” part of this assertion is proved in [@CSR].
Assume there is a spectral projector of the operator ${G}$ corresponding to a finite interval $[0, E_0]$ with infinite-dimensional range $\H_0$. Since $\langle\varphi|G|\varphi\rangle\leq E_0$ for any unit vector $\varphi$ in $\H_0$, we have $\|A\|^{G}_E=\|A\|$ for any $A\in\B(\H_0)$ and $E>E_0$. So, any of the norms $\,\|\!\cdot\!\|^{G}_E$, $E>0$, generates the norm topology on $\B(\H_0)$ in this case $\square$.
Different types of operator convergence can be obtained by using the norm $\|\!\cdot\!\|_E^{G}$ induced by different operators ${G}$.
**Example.** Let $\H=\H_1\oplus\H_2$ and ${G}={G}_1\oplus {G}_2$, where ${G}_k$ is a positive densely defined operator on a separable Hilbert space $\H_k$ satisfying condition (\[H-cond\]), $k=1,2$. By using definition (\[ec-on\]) and the triangle inequality it is easy to show that $$\label{E-ineq}
\sqrt{p\!\left[\|AP_1\|^{{G}_1}_E\right]^2+(1-p)\!\left[\|AP_2\|^{{G}_2}_E\right]^2}\leq \|A\|^{G}_E\leq \|AP_1\|^{{G}_1}_E+\|AP_2\|^{{G}_2}_E$$ for any $p\in[0,1]$, where $P_k$ is the projector on the subspace $\H_k$ and $\|AP_k\|^{{G}_k}_E$, $k=1,2$, are defined in accordance with Remark \[en-gen\].
Assume that ${G}_1$ is a discrete unbounded operator (Def.\[D-H\]) and ${G}_2$ is a bounded operator. Then it follows from (\[E-ineq\]) and Proposition \[ec-on-p2\] that $$\left\{\|\!\cdot\!\|^{G}_E{\hspace{1pt}}\textrm{-}\lim_{n\rightarrow\infty} A_n=A_0\right\}\quad \Leftrightarrow\quad \left\{{s.o.}{\hspace{1pt}}\textrm{-}\lim_{n\rightarrow\infty} A_nP_1=A_0P_1\right\} \wedge \left\{\|\!\cdot\!\|{\hspace{1pt}}\textrm{-}\lim_{n\rightarrow\infty} A_nP_2=A_0P_2\right\}$$ for a bounded sequence $\{A_n\}\subset\B(\H)$, where ${s.o.}{\hspace{1pt}}\textrm{-}\lim$ denotes the limit w.r.t the strong operator topology. So, in this case the norm $\|\!\cdot\!\|^{G}_E$ generates a “hybrid” topology on bounded subsets of $\B(\H)$ – some kind of the Cartesian product of the strong operator and the norm topologies.
Equivalent definitions and equivalent norms
-------------------------------------------
Recall that $\T_{+,1}(\H)$ denotes the positive part of the unit ball in $\T(\H)$. Denote by $\H_{\leq 1}$ the unit ball in $\H$.
\[en-def\] A) *For any $A\in\B(\H)$ and $E>0$ the following expressions hold $$\label{ec-on+}
\|A\|^{G}_E=\sup_{\substack{\varphi\in\H_{\leq 1}:\langle\varphi|G|\varphi\rangle\leq E}}\|A\varphi\|=\sup_{\substack{\rho\in\T_{+,1}(\H):
{\mathrm{Tr}}{G}\rho\leq E}}\sqrt{{\mathrm{Tr}}A\rho A^*},\qquad \forall A\in \B(\H),$$ i.e. the suprema in definitions (\[ec-on-b\]) and (\[ec-on\]) can be taken, respectively, over all vectors in $\H_{\leq 1}$ satisfying the condition $\,\langle\varphi|G|\varphi\rangle\leq E$ and over all operators in $\T_{+,1}(\H)$ satisfying the condition $\,{\mathrm{Tr}}{G}\rho\leq E$.*
B\) *If the operator $G$ is unbounded then for any $A\in\B(\H)$ and $E>0$ the conditions $\,\langle\varphi|G|\varphi\rangle\leq E$ in (\[ec-on-b\]) and $\,{\mathrm{Tr}}{G}\rho\leq E$ in (\[ec-on\]) and can be replaced, respectively, by the conditions $\,\langle\varphi|G|\varphi\rangle=E$ and $\,{\mathrm{Tr}}G\rho=E$.*
C\) *If $G$ is a discrete unbounded operator (Def.\[D-H\]) then for any $A\in\B(\H)$ and $E>0$ the suprema in (\[ec-on-b\]) and (\[ec-on\]) are attainable. Moreover, if $\|A\|^{G}_{E}<\|A\|$ then the suprema in (\[ec-on-b\]) and (\[ec-on\]) are attained, respectively, at unit vector $\varphi_0$ in $\H$ such that $\,\langle\varphi_0|G|\varphi_0\rangle=E$ and at a state $\rho_0$ in $\S(\H)$ such that $\,{\mathrm{Tr}}G\rho_0=E$.*
\[en-def-r\] It is easy to see that assertion A of Proposition \[en-def\] is not valid if the operator $G$ doesn’t satisfy condition (\[H-cond\]).
*Proof of Proposition \[en-def\].* A) It suffices to show that the last expression in (\[ec-on+\]) does not exceed $\|A\|^{G}_E$. Let $\rho$ be an operator in $\T_{+,1}(\H)$ such that $\,{\mathrm{Tr}}{G}\rho\leq E$ and $r={\mathrm{Tr}}\rho$. Then $\hat{\rho}\doteq r^{-1}\rho$ is a state such that $\,{\mathrm{Tr}}{G}\hat{\rho}\leq E/r$. So, by using concavity of the function $E\rightarrow\left[\|A\|^{G}_E\right]^2$ and Lemma \[WL\] in Section 2 we obtain $${\mathrm{Tr}}A\rho A^*=r{\mathrm{Tr}}A\hat{\rho} A^*\leq r\left[\|A\|^{G}_{E/r}\right]^2\leq \left[\|A\|^{G}_{E}\right]^2.$$
B)[^7] Show first that the inequality ${\mathrm{Tr}}{G}\rho\leq E$ can be replaced by the equality ${\mathrm{Tr}}{G}\rho=E$ in (\[ec-on\]). Assume that there exist $E>0$ and $A\in\B(\H)$ such that $$\label{assum}
\sup_{\substack{\rho\in\mathfrak{S}(\mathcal{H}):
{\mathrm{Tr}}{G}\rho=E}}\sqrt{{\mathrm{Tr}}A\rho A^*}\leq \|A\|^{G}_E-\varepsilon$$ for some $\varepsilon>0$. Let $\rho_\varepsilon$ be a state such that $\sqrt{{\mathrm{Tr}}A\rho_\varepsilon A^*}>\|A\|^{G}_E-\varepsilon/2$ and ${\mathrm{Tr}}G\rho_\varepsilon<E$. For each natural $n>E$ there exist a state $\sigma_n$ and a number $p_n\in(0,1)$ such that ${\mathrm{Tr}}G\sigma_n\in(n,+\infty)$ and ${\mathrm{Tr}}G\varrho_n=E$, where $\varrho_n=(1-p_n)\rho_\varepsilon+p_n\sigma_n$. It is clear that $p_n\rightarrow 0$ as $n\rightarrow+\infty$. Hence ${\mathrm{Tr}}A\varrho_n A^*$ tends to ${\mathrm{Tr}}A\rho_\varepsilon A^*$ contradicting (\[assum\]).
For any $\varepsilon>0$ let $\rho_\varepsilon$ be a state such that ${\mathrm{Tr}}A\rho_\varepsilon A^*>[\|A\|^{G}_E-\varepsilon]^2$ and ${\mathrm{Tr}}G\rho_\varepsilon=E$. By Corollary 1 in there is a probability measure $\mu$ on $\S(\H)$ supported by pure states such that $$\label{p-s-d}
\rho_\varepsilon=\int \sigma\mu(d\sigma)\quad \textrm{and}\quad {\mathrm{Tr}}H\sigma=E\;\;\textrm{for}\;\;\mu\textrm{-almost all}\;\;\sigma.$$ Since that function $\sigma\mapsto{\mathrm{Tr}}A\sigma A^*$ is affine and continuous, we have $$\int {\mathrm{Tr}}A\sigma A^*\mu(d\sigma)={\mathrm{Tr}}A\rho_\varepsilon A^*>[\|A\|^{G}_E-\varepsilon]^2.$$ It shows existence of a pure state $\sigma_\varepsilon$ such that ${\mathrm{Tr}}A\sigma_\varepsilon A^*>[\|A\|^{G}_E-\varepsilon]^2$ and ${\mathrm{Tr}}H\sigma_\varepsilon=E$.
C\) In this case the set of states $\rho$ satisfying the condition ${\mathrm{Tr}}{G}\rho\leq E$ is compact by Lemma \[Comp\] in Section 2. Hence the supremum in (\[ec-on\]) is attained at some state $\rho_0$. The condition $\|A\|^{G}_E<\|A\|$ implies, by Proposition \[ec-on-p1\], that $\|A\|^{G}_{E'}<\|A\|^{G}_E$ for any $E'<E$. This shows that ${\mathrm{Tr}}G\rho_0=E$. By using decomposition (\[p-s-d\]) of $\rho_0$ and the arguments below it is easy to show existence of a unit vector $\varphi_0$ in $\H$ such that $\,\langle\varphi_0|G|\varphi_0\rangle=E$ at which the supremum in (\[ec-on-b\]) is attained. $\square$
Consider the following norm on $\B(\H)$ depending on a positive parameter $E$: $$\label{eq-norms-2}
{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}=\sup\left\{\|A\varphi\|\left|\, \varphi\in \D(\sqrt{G}),\, \|\varphi\|^2+\langle\varphi|G|\varphi\rangle/E\leq 1\right.\right\}.$$ This norm naturally appears in analysis of relatively bounded operators (see Section 5). We will obtain relations between the norms $\|\cdot\|^{G}_{E}$ and ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}^{G}_{E}$ assuming that $G$ is an arbitrary positive operator satisfying condition (\[H-cond\]).
By using definitions (\[ec-on-b\]) and (\[eq-norms-2\]) it is easy to obtain $$\label{one}
\sqrt{1/2}\|A\|^{G}_{E}\leq {\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}\leq\|A\|^{G}_{E}$$ for any $E>0$. This inequality and inequality (\[E-n-eq\]) imply that $$\label{E-n-eq+}
{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E_1}\leq {\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E_2}\leq \sqrt{2E_2/E_1}{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E_1}\quad\textrm{ for any } E_2>E_1>0$$ and any $A\in\B(\H)$. The above inequalities show that all the norms in the families $\{\|\cdot\|^{G}_{E}\}_{E>0}$ and $\{{\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}^{G}_{E}\}_{E>0}$ are equivalent to each other.
In fact, all the norms $\|\cdot\|^{G}_{E}$ and ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}^{G}_{E}$ are equivalent on the space of all linear operators on $\H$ relatively bounded w.r.t. the operator $\sqrt{G}$, and each of these norms makes this space a Banach space. Moreover, the functions $E\mapsto\|A\|^G_E$ and $E\mapsto{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^G_E$ are completely determined by each other (Remark \[comp-d-r+\] in Section 5). The main advantages of the operator *E*-norm $\|\cdot\|^{G}_{E}$ in comparison with the norm ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}^{G}_{E}$ are the following:
- the concavity of the function $E\mapsto\left[\|A\|^{G}_{E}\right]^p$ for any $p\in(0,2]$;[^8]
- the appearance of the norm $\|\cdot\|^{G}_{E}$ in the generalized Kretschmann-Schlingemann-Werner theorem (Section 4);
- the simple estimation of $\|\Phi(A)\|^{G}_E$ via $\|A\|^{G}_E$, where $\Phi:\B(\H)\rightarrow\B(\H)$ is any 2-positive linear map satisfying the conditions of Proposition \[bp-en-0\]E (Section 3.2).
\[nonconc\] To show that the function $E\mapsto\left[{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}\right]^p$ is not concave in general for any $p\in(0,2]$ it suffices to consider two-dimensional Hilbert space $\H=\mathbb{C}^2$ and the operators $$G=\left[\begin{array}{ll}
1 & 0\\
0 & 0
\end{array}\right]\quad \textrm{and} \quad A=\left[\begin{array}{ll}
\sqrt{2} & 0\\
0 & 1
\end{array}\right].$$ It is easy to see that $\,{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}=1\,$ if $\,E\in(0,1]\,$ and $\,{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}=\sqrt{2E/(E+1)}\,$ if $\,E>1$.
Basic properties of the operator *E*-norms
------------------------------------------
In the following proposition we collect properties of the operator *E*-norms used below.
\[bp-en-0\] *Let ${G}$ be a positive densely defined operator on a Hilbert space $\H$ satisfying condition (\[H-cond\]) and $E>0$.*
A\) *$\|A\|^{G}_E=\||A|\|^{G}_E\leq \sqrt{\|A^*A\|^{G}_E}$ for all $A\in\B(\H)$ but $\,\|A^*\|^{G}_E\neq\|A\|^{G}_E$ in general;*
B\) *For arbitrary operators $A$ and $B$ in $\B(\H)$ the following inequalities hold $$m(A)\|B\|^{G}_E\leq \|AB\|^{G}_E\leq \|A\|\|B\|^{G}_E,$$ where $m(A)$ is the infimum of the spectrum of the operator $|A|=\sqrt{A^*A}$.*
C\) *For arbitrary operators $A$ and $B$ in $\B(\H)$ such that $\,\langle A\varphi |B\varphi\rangle=0$ for any $\varphi\in\H$ the following inequalities hold* $$\max\!{\hspace{1pt}}\left\{\|A\|^{G}_{E},\|B\|^{G}_{E}{\hspace{1pt}}\right\}\leq \|A+B\|^{G}_{E}\leq \sqrt{\left[\|A\|^{G}_{E}\right]^2+\left[\|B\|^{G}_{E}\right]^2}.$$
D\) *For an operator $\rho$ in $\T_{+,1}(\H)$ with finite $E_{\rho}\doteq{\mathrm{Tr}}{G}\rho$ and any operators $A$ and $B$ in $\B(\H)$ the following inequalities hold* $$|{\mathrm{Tr}}A\rho B^*|\leq \|A\rho B^*\|_1 \leq \|A\|^{G}_{E_{\rho}}\|B\|^{G}_{E_{\rho}}.$$
E\) *For any 2-positive map $\Phi:\B(\H)\rightarrow\B(\H)$ such that $\,\Phi(I_{\H})\leq I_{\H}\,$ having the predual map[^9] $\Phi_*:\T(\H)\rightarrow\T(\H)$ with finite $\,Y_{\Phi}(E)\doteq \sup\!\left\{{\mathrm{Tr}}{G}\Phi_*(\rho)\,|\,\rho\in\mathfrak{S}(\mathcal{H}),{\mathrm{Tr}}{G}\rho\leq E{\hspace{1pt}}\right\}$ and arbitrary operator $A$ in $\B(\H)$ the following inequalities hold* [^10] $$\|\Phi(A)\|^{G}_E\leq\sqrt{\|\Phi(I_{\H})\|}\,\|A\|^{G}_{Y_{\Phi}(E)}\leq \sqrt{\|\Phi(I_{\H})\|K_{\Phi}}\,\|A\|^{G}_{E},\qquad K_{\Phi}=\max\{1, Y_{\Phi}(E)/E\}.$$
Proposition \[bp-en-0\] shows that the linear transformations $A\mapsto BA$ and $A\mapsto\Phi(A)$ of $\B(\H)$, where $B\in\B(\H)$ and $\Phi$ is a map with the properties pointed in part E, are bounded operators w.r.t. the norm $\|\!\cdot\!\|^{G}_E$ (in contrast to the transformation $A\mapsto AB$).
*Proof.* A) The equality $\|A\|^{G}_E=\||A|\|^{G}_E$ is obvious. The inequality $\|A\|^{G}_E\leq\sqrt{\|A^*A\|^{G}_E}$ follows from the operator Cauchy-Schwarz inequality $$[{\mathrm{Tr}}A^*A\rho]^2\leq [{\mathrm{Tr}}[A^*A]^2\rho][{\mathrm{Tr}}\rho].$$ To show that $\|A^*\|^{G}_E$ may not coincide with $\|A\|^{G}_E$ take any operator ${G}$ having form (\[H-rep\]). It is easy to see that $\||\tau_0\rangle\langle\tau_n|\|^{G}_E=\sqrt{E/E_n}\,$ while $\||\tau_n\rangle\langle\tau_0|\|^{G}_E=1\,$ for all $E>0$.
B\) This assertion follows directly from definition (\[ec-on\]) of the operator *E*-norm.
C\) This assertion follows directly from definition (\[ec-on-b\]) of the operator *E*-norm.
D\) The first inequality is obvious. Let $U$ be the partial isometry from the polar decomposition of $A\rho B^*$, i.e. $A\rho B^*=U|A\rho B^*|$. By using the operator Cauchy-Schwarz inequality we obtain $$\|{\mathrm{Tr}}A\rho B^*\|^2_1=[{\mathrm{Tr}}U^*\!A\rho B^*]^2\leq [{\mathrm{Tr}}UU^*\!A\rho A^*][{\mathrm{Tr}}B\rho B^*]\leq [{\mathrm{Tr}}A\rho A^*][{\mathrm{Tr}}B\rho B^*],$$ where that last inequality is due to the fact that $UU^*\leq I_{\H}$. By Proposition \[en-def\]A the right hand side of this inequality does not exceed $\left[\|A\|^{G}_{E_{\rho}}\|B\|^{G}_{E_{\rho}}\right]^2$.
E\) By Kadison’s inequality and Proposition \[en-def\]A we have $${\mathrm{Tr}}[\Phi(A)]^*\Phi(A)\rho\leq\|\Phi(I_{\H})\|{\mathrm{Tr}}\Phi(A^*A)\rho=\|\Phi(I_{\H})\|{\mathrm{Tr}}A^*A\Phi_*(\rho)\leq \|\Phi(I_{\H})\| \left[\|A\|^{G}_{Y_{\Phi}(E)}\right]^2$$ for any $A\in\B(\H)$ and any $\rho\in\S(\H)$ such that ${\mathrm{Tr}}{G}\rho\leq E$ (since the condition $\,\Phi(I_{\H})\leq I_{\H}\,$ guarantees that $\Phi_*(\rho)\in\T_{+,1}(\H)$). This implies the first inequality. The second inequality follows from (\[E-n-eq\]). $\square$
Properties of the *E*-norms related to tensor products
------------------------------------------------------
If ${G}_1$ and ${G}_2$ are positive densely defined operators on Hilbert spaces $\H_1$ and $\H_2$ satisfying condition (\[H-cond\]) then ${G}_{12}={G}_{1}\otimes I_{2}+ I_{1}\otimes {G}_{2}$ is an operator on the Hilbert space $\H_{12}=\H_1\otimes\H_2$ with the same properties.[^11][^12] The following proposition contains several estimates for the operator *E*-norms of product operators used in Section 4.
\[pen-tp\] *Let ${G}_1$, ${G}_2$ and ${G}_{12}$ be the operators described above.*
A\) *For arbitrary operator $A$ in $\B(\H_1)$ the following equalities hold* $$\|A\otimes I_{2}\|^{{G}_{12}}_E=\|A\otimes I_{2}\|^{{G}_{1}\otimes {\hspace{1pt}}I_{2}}_E=\|A\|^{{G}_1}_E$$
B\) *For arbitrary operators $A\in\B(\H_1)$ and $B\in\B(\H_2)$ the following inequalities hold* $$\label{pen-tp-1}
\sup_{x\in(0,E)}\|A\|^{{G}_1}_{x}\|B\|^{{G}_2}_{E-x}\leq \|A\otimes B\|^{{G}_{12}}_{E} \leq\sup_{x\in(0,E)} \sqrt{\|A^*A\|^{{G}_1}_{x}\|B^*B\|^{{G}_2}_{E-x}},$$ *and* $$\label{pen-tp-2}
\|A\otimes B\|^{{G}_{12}}_{E} \leq\min\left\{\|A\|^{{G}_1}_{E}\|B\|,\|A\|\|B\|^{{G}_2}_{E}\right\}.$$
Note that the lower and upper bounds in (\[pen-tp-1\]) and the r.h.s. of (\[pen-tp-2\]) tend to $\|A\|\|B\|=\|A\otimes B\|=\lim_{E\to+\infty}\|A\otimes B\|^{{G}_{12}}_{E}$ as $E\to+\infty$.
*Proof.* A) It suffices to note that ${\mathrm{Tr}}[|A|^2\otimes I_{2}]\rho_{12}={\mathrm{Tr}}|A|^2\rho_{1}$ and that ${\mathrm{Tr}}{G}_{12}\rho_{12}={\mathrm{Tr}}{G}_{1}\rho_{1}+{\mathrm{Tr}}{G}_{2}\rho_{2}$ for any state $\rho_{12}\in\S(\H_{12})$, where $\rho_1={\mathrm{Tr}}_{\H_2}\rho_{12}$ and $\rho_2={\mathrm{Tr}}_{\H_1}\rho_{12}$ are the partial states of $\rho_{12}$.
B\) For each $x\in(0,E)$ and any $\varepsilon>0$ there exist states $\rho_1$ in $\S(\H_1)$ and $\rho_2$ in $\S(\H_2)$ such that $\sqrt{{\mathrm{Tr}}|A|^2\rho_1}>\|A\|^{{G}_1}_x-\varepsilon$, ${\mathrm{Tr}}{G}_1\rho_1\leq x$, $\sqrt{{\mathrm{Tr}}|B|^2 \rho_2}>\|B\|^{{G}_2}_{E-x}-\varepsilon$ and ${\mathrm{Tr}}{G}_2\rho_2\leq E-x$. Then ${\mathrm{Tr}}{G}_{12}[\rho_1\otimes\rho_2]\leq x+E-x=E$ and $\sqrt{{\mathrm{Tr}}[|A|^2\otimes|B|^2][\rho_1\otimes\rho_2]}\geq[\|A\|^{{G}_1}_x-\varepsilon][\|B\|^{{G}_2}_{E-x}-\varepsilon]$. Since $\varepsilon$ is arbitrary, this implies the left inequality in (\[pen-tp-1\]).
By the operator Cauchy-Schwarz inequality for any state $\rho_{12}$ in $\S(\H_{12})$ we have $$\begin{array}{c}
{\mathrm{Tr}}[|A|^2\otimes |B|^2]\rho_{12}\leq \sqrt{{\mathrm{Tr}}[|A|^4\otimes I_{2}]\rho_{12}}\sqrt{{\mathrm{Tr}}[I_{1}\otimes |B|^4]\rho_{12}}\\\\=
\sqrt{{\mathrm{Tr}}|A|^4\rho_{1}}\sqrt{{\mathrm{Tr}}|B|^4\rho_{2}}\leq \||A|^2\|^{{G}_1}_{{\mathrm{Tr}}{G}_1\rho_1}\||B|^2\|^{{G}_2}_{{\mathrm{Tr}}{G}_2\rho_2}.
\end{array}$$ Since ${\mathrm{Tr}}{G}_{12}\rho_{12}={\mathrm{Tr}}{G}_{1}\rho_{1}+{\mathrm{Tr}}{G}_{2}\rho_{2}$, this implies the right inequality in (\[pen-tp-1\]).
To prove inequality (\[pen-tp-2\]) it suffices to note that $$A\otimes B=[A\otimes I_{2}][I_{1}\otimes B]=[I_{1}\otimes B][A\otimes I_{2}]$$ and to apply Proposition \[bp-en-0\]B and part A of this proposition. $\square$
The *E*-version of the Kretschmann-Schlingemann-Werner theorem
==============================================================
In this section we consider application of the operator $E$-norms to the theory of completely positive (CP) linear maps between Banach spaces of trace class operators on separable Hilbert spaces (the Schatten classes of order 1). Since $\T(\H)^*=\B(\H)$, the below results can be reformulated in terms of CP linear maps between algebras of all bounded operators on separable Hilbert spaces. Nevertheless, the use of the “predual picture” is more natural for representation of our results. The theory of CP linear maps between Banach spaces of trace class operators has important applications in mathematical physics, in particular, in the theory of open quantum systems, where CP trace-preserving linear maps called *quantum channels* play the role of dynamical maps (in the Schrodinger picture), while CP trace-non-increasing linear maps called *quantum operations* are essentially used in the theory of quantum measurements [@H-SCI; @Watrous; @Wilde].
For a CP linear map $\,\Phi:\T(\H_A)\rightarrow \T(\H_B)\,$ the Stinespring theorem (cf.[@St]) implies existence of a Hilbert space $\mathcal{H}_E$ and an operator $V_{\Phi}:\mathcal{H}_A\rightarrow\mathcal{H}_B\otimes\mathcal{H}_E$ such that $$\label{St-rep}
\Phi(\rho)=\mathrm{Tr}_{E}V_{\Phi}\rho V_{\Phi}^{*},\quad
\rho\in\mathfrak{T}(\mathcal{H}_A),$$ where ${\mathrm{Tr}}_E$ denotes the partial trace over $\H_E$. If $\Phi$ is trace-preserving (correspondingly, trace-non-increasing) then $V_{\Phi}$ is an isometry (correspondingly, contraction) [@H-SCI Ch.6].
The dual CP linear map $\,\Phi^*:\B(\H_B)\rightarrow \B(\H_A)\,$ has the corresponding representation $$\label{St-rep+}
\Phi^*(B)=V^*_{\Phi}[B\otimes I_E] V_{\Phi},\quad
B\in\mathfrak{B}(\mathcal{H}_B).$$
The norm of complete boundedness ($cb{\hspace{1pt}}$-norm in what follows) of a linear map between the algebras $\B(\H_B)$ and $\B(\H_A)$ (cf. [@Paul]) induces (by duality) the norm $$\label{d-n-def}
\|\Phi\|_{\rm cb}\doteq\sup_{\rho\in\T(\H_{AR}),\|\rho\|_1\leq 1}\|\Phi\otimes {\mathrm{Id}}_R(\rho)\|_1$$ on the set of all linear maps between Banach spaces $\T(\H_A)$ and $\T(\H_B)$, where $\H_R$ is a separable Hilbert space and $\H_{AR}=\H_{A}\otimes \H_{R}$. If $\Phi$ is a Hermitian preserving map then the supremum in (\[d-n-def\]) can be taken over the set $\S(\H_{AR})$ [@Watrous Ch.3].
The Kretschmann-Schlingemann-Werner theorem (the KSW-theorem in what follows) obtained in states that $$\frac{\|\Phi-\Psi\|_{\rm cb}}{\sqrt{\|\Phi\|_{\rm cb}}+\sqrt{\|\Psi\|_{\rm cb}}}\leq\inf_{V_{\Phi},V_{\Psi}}\|V_{\Phi}-V_{\Psi}\|\leq\sqrt{\|\Phi-\Psi\|_{\rm cb}},$$ where the infimum is over all common Stinespring representations $$\label{c-S-r}
\Phi(\rho)={\mathrm{Tr}}_E V_{\Phi}\rho V^*_{\Phi}\quad\textrm{and}\quad\Psi(\rho)={\mathrm{Tr}}_E V_{\Psi}\rho V^*_{\Psi}.$$ In the proof of the KSW theorem it is shown that the quantity $\inf_{V_{\Phi},V_{\Psi}}\|V_{\Phi}-V_{\Psi}\|$ coincides with the *Bures distance* between the maps $\Phi$ and $\Psi$ defined by the expression $$\label{b-dist+}
\beta(\Phi,\Psi)=\sup_{\rho\in\S(\H_{AR})} \beta\!\left(\Phi\otimes {\mathrm{Id}}_R(\rho),\Psi\otimes {\mathrm{Id}}_R(\rho)\right),$$ in which $\H_R$ is a separable Hilbert space and $\beta(\cdot,\cdot)$ in the r.h.s. is the Bures distance between operators in $\T_+(\H_{BR})$ defined as $$\label{B-d-s}
\beta(\rho,\sigma)=\sqrt{\|\rho\|_1+\|\sigma\|_1-2\sqrt{F(\rho,\sigma)}},$$ where $$\label{fidelity}
F(\rho,\sigma)=\|\sqrt{\rho}\sqrt{\sigma}\|^2_1$$ is the fidelity of the operators $\rho$ and $\sigma$ [@H-SCI; @Watrous; @Wilde]. The Bures distance between CP linear maps $\Phi$ and $\Psi$ is connected to the *operational fidelity* of these maps introduced in .
The KSW theorem shows continuity of the map $V_{\Phi}\mapsto\Phi$ and selective continuity of the multi-valued map $\Phi\mapsto V_{\Phi}$ with respect to the $cb{\hspace{1pt}}$-norm topology on the set $\F(A,B)$ of all CP linear maps $\Phi$ from $\T(\H_A)$ to $\T(\H_B)$ and the operator norm topology on the set of Stinespring operators $V_{\Phi}$.
The $cb{\hspace{1pt}}$-norm topology is widely used in the quantum theory, by it is too strong for description of physical perturbations of infinite-dimensional quantum channels [@SCT; @W-EBN]. Our aim is to obtain a version of the KSW theorem which would show continuity of the map $V_{\Phi}\mapsto\Phi$ and selective continuity of the multi-valued map $\Phi\mapsto V_{\Phi}$ with respect to *weaker topologies* on the sets of CP linear maps $\Phi$ and Stinespring operators $V_{\Phi}$. A natural way to do this is to use the operator *E*-norms induced by some positive operator ${G}$ on $\H_A$ (naturally generalized to operators between different separable Hilbert spaces, see Remark \[en-gen\]) and the energy-constrained $cb{\hspace{1pt}}$-norms $$\label{E-sn}
\|\Phi\|^G_{\mathrm{cb},E}\doteq\sup_{\rho\in\S(\H_{AR}):{\mathrm{Tr}}{G}\rho_A\leq E}\|\Phi\otimes {\mathrm{Id}}_R(\rho)\|_1,\quad E>0,\quad (\textrm{where}\; \rho_A\doteq{\mathrm{Tr}}_R{\hspace{1pt}}\rho)$$ on the set of Hermitian-preserving linear maps from $\T(\H_{A})$ to $\T(\H_{B})$ introduced independently in [@SCT] and [@W-EBN] (the positive operator ${G}$ is treated therein as a Hamiltonian of a quantum system $A$).[^13] If ${G}$ is a discrete unbounded operator (see Def.\[D-H\]) then the topology generated by any of the norms (\[E-sn\]) on bounded subsets of $\F(A,B)$ coincides with the strong convergence topology generated by the family of seminorms $\Phi\mapsto\|\Phi(\rho)\|_1$, $\rho\in\T(\H_A)$ [@SCT Proposition 3].[^14]
Following [@CID] introduce the *energy-constrained Bures distance* $$\label{ec-b-dist}
\beta_E^G(\Phi,\Psi)=\sup_{\rho\in\S(\H_{AR}):{\mathrm{Tr}}{G}\rho_A\leq E} \beta(\Phi\otimes {\mathrm{Id}}_R(\rho),\Psi\otimes {\mathrm{Id}}_R(\rho)), \quad E>0,$$ between CP linear maps $\Phi$ and $\Psi$ from $\T(\H_A)$ to $\T(\H_B)$, where $\beta(\cdot,\cdot)$ in the r.h.s. is the Bures distance between operators in $\T_+(\H_{BR})$ defined in (\[B-d-s\]) and $\H_R$ is a separable Hilbert space.
\[ec-b-dist-r\] The infimum in (\[ec-b-dist\]) can be taken only over pure states $\rho\in\S(\H_{AR})$. This follows from the freedom of choice of $R$, which implies possibility to purify any mixed state in $\S(\H_{AR})$ by extending system $R$. We have only to note that the Bures distance between operators in $\T_{+}(\H_{XY})$ defined in (\[B-d-s\]) does not increase under partial trace: $\beta(\rho,\sigma)\geq \beta(\rho_X,\sigma_X)$ for any $\rho$ and $\sigma$ in $\T_{+}(\H_{XY})$ [@H-SCI; @Watrous; @Wilde].
The distance $\beta_E^G(\Phi,\Psi)$ turns out to be extremely useful in quantitative continuity analysis of capacities of energy-constrained infinite-dimensional quantum channels [@CID Theorem 2]. By using the well known relations between the trace norm and the Bures distance (\[B-d-s\]) one can show that for any $E>0$ the distance $\beta_E^G(\Phi,\Psi)$ generates the same topology on bounded subsets of $\F(A,B)$ as any of the energy-constrained $cb{\hspace{1pt}}$-norms (\[E-sn\]). The results of calculation of $\beta_E^G(\Phi,\Psi)$ for real quantum channels can be found in [@Nair].
Now we can formulate the $E$-version of KSW-theorem.
\[KSW-E\] *Let ${G}$ be a positive semidefinite densely defined operator on $\H_A$ satisfying condition (\[H-cond\]) and $E>0$. Let $\|\cdot\|^G_{\mathrm{cb},E}$ and $\|\cdot\|_E^G$ be, respectively, the energy-constrained $cb{\hspace{1pt}}$-norm and the operator $E\textrm{-}$norm induced by ${G}$. For any CP linear maps $\,\Phi$ and $\,\Psi$ from $\,\T(\H_A)$ to $\,\T(\H_B)$ the following inequalities hold $$\label{KSW-rel}
\frac{\|\Phi-\Psi\|^G_{\mathrm{cb},E}}{\sqrt{\|\Phi\|^G_{\mathrm{cb},E}}+\sqrt{\|\Psi\|^G_{\mathrm{cb},E}}}\leq\inf_{V_{\Phi},V_{\Psi}}
\|V_{\Phi}-V_{\Psi}\|_E^G\leq\sqrt{\|\Phi-\Psi\|^G_{\mathrm{cb},E}},$$ where the infimum is over all common Stinespring representation (\[c-S-r\]). The quantity $\,\inf_{V_{\Phi},V_{\Psi}}\|V_{\Phi}-V_{\Psi}\|_E^G$ coincides with the energy-constrained Bures distance $\beta_E^G(\Phi,\Psi)$ defined in (\[ec-b-dist\]). The infimum in (\[KSW-rel\]) is attainable.*
*Proof.* We will follow the proof of the KSW theorem (given in ) with necessary modifications concerning the use of the energy-constrained $cb{\hspace{1pt}}$-norms and the operator $E$-norms (instead of the ordinary $cb{\hspace{1pt}}$-norm and the operator norm).
To prove the first inequality in (\[KSW-rel\]) assume that $\rho$ is a state in $\S(\H_{AR})$ such that ${\mathrm{Tr}}{G}\rho_A\leq E$. For a given common Stinespring representation (\[c-S-r\]) we have $$\begin{array}{c}
\|(\Phi-\Psi)\otimes {\mathrm{Id}}_R(\rho)\|_1\leq \|V_{\Phi}\otimes I_R \cdot\rho\cdot V^*_{\Phi}\otimes I_R-V_{\Psi}\otimes I_R \cdot\rho\cdot V^*_{\Psi}\otimes I_R\|_1\\\\
\leq\|(V_{\Phi}-V_{\Psi})\otimes I_R \cdot\rho\cdot V^*_{\Phi}\otimes I_R\|_1+\|V_{\Psi}\otimes I_R \cdot\rho\cdot(V^*_{\Phi}-V^*_{\Psi})\otimes I_R\|_1\\\\
\leq\|(V_{\Phi}-V_{\Psi})\otimes I_R\|_E^{G\otimes I_R} \|V_{\Phi}\otimes I_R\|_E^{G\otimes I_R}+\|(V_{\Phi}-V_{\Psi})\otimes I_R\|_E^{G\otimes I_R}\|V_{\Psi}\otimes I_R\|_E^{G\otimes I_R}
\\\\
\leq\|V_{\Phi}-V_{\Psi}\|_E^G \|V_{\Phi}\|_E^G+\|V_{\Phi}-V_{\Psi}\|_E^G \|V_{\Psi}\|_E^G.
\end{array}$$ The first and the second inequalities follow from the properties of the trace norm (the non-increasing under partial trace and the triangle inequality), the third inequality follows from Proposition \[bp-en-0\]D, the last one – from Proposition \[pen-tp\]A. By noting that $[\|V_{\Phi}\|^G_E]^2=\|\Phi\|^G_{\mathrm{cb},E}$ and $[\|V_{\Psi}\|^G_E]^2=\|\Psi\|^G_{\mathrm{cb},E}$ we obtain the first inequality in (\[KSW-rel\]).
To prove the second inequality in (\[KSW-rel\]) note that $\beta_E^G(\Phi,\Psi)\leq\sqrt{\|\Phi-\Psi\|^G_{\mathrm{cb},E}}$. This follows from the inequality $\beta(\rho,\sigma)\leq\sqrt{\|\rho-\sigma\|_1}$ valid for any $\rho$ and $\sigma$ in $\T_+(\H)$, which is easily proved by using the inequality ${\mathrm{Tr}}(\sqrt{\rho}-\sqrt{\sigma})^2\leq \|\rho-\sigma\|_1$ (see the proof of Lemma 9.2.3 in [@H-SCI]). So, it suffices to show that $$\label{r-r}
\inf_{V_{\Phi},V_{\Psi}}\|V_{\Phi}-V_{\Psi}\|_E^G=\beta_E^G(\Phi,\Psi).$$
Denote by $\alpha^G_E(\Phi,\Psi)$ the l.h.s. of (\[r-r\]). Let $\C^s_{{G},E}$ be the subset of $\S(\H_A)$ determined by the inequality $\,{\mathrm{Tr}}{G}\rho\leq E\,$ and $\N(\Phi,\Psi)=\bigcup V_{\Phi}^*V_{\Psi}$, where the union is over *all* common Stinespring representations (\[c-S-r\]). Then by using definition (\[ec-on\]) we obtain $$\label{beta-e}
\alpha^G_E(\Phi,\Psi)=\inf_{N\in\N(\Phi,\Psi)} \sup_{\rho\in\C^s_{{G},E}}\sqrt{{\mathrm{Tr}}{\hspace{1pt}}\Phi(\rho)+{\mathrm{Tr}}\Psi(\rho)-2\Re{\hspace{1pt}}{\mathrm{Tr}}N\rho}.$$ Following the proof of Theorem 1 in show that $\N(\Phi,\Psi)$ coincides with the set $$\M(\Phi,\Psi)\doteq \left\{V_{\Phi} ^*(I_B\otimes C)V_{\Psi}\,|\,C\in \B(\H_E), \|C\|\leq 1\right\},$$ defined via some *fixed* common Stinespring representation (\[c-S-r\]). It will imply, in particular, that $\M(\Phi,\Psi)$ does not depend on this representation.
To show that $\M(\Phi,\Psi)\subseteq\N(\Phi,\Psi)$ it suffices to find for any contraction $C\in\B(\H_E)$ a common Stinespring representation for $\Phi$ and $\Psi$ with the operators $\tilde{V}_{\Phi}$ and $\tilde{V}_{\Psi}$ from $\H_A$ to $\H_B\otimes\H_{\tilde{E}}$ such that $\tilde{V}^*_{\Phi}\tilde{V}_{\Psi}=V_{\Phi} ^*(I_B\otimes C)V_{\Psi}$.
Let $\H_{\tilde{E}}=\H^1_E\oplus\H^2_E$, where $\H^1_E$ and $\H^2_E$ are copies of $\H_E$. For given $C$ define the operators $\tilde{V}_{\Phi}$ and $\tilde{V}^C_{\Psi}$ from $\H_A$ into $\H_B\otimes(\H_{E_1}\oplus\H_{E_2})=\H_B\otimes\H_{E_1}\oplus\H_B\otimes\H_{E_2}$ by setting $$\label{bar-v}
\tilde{V}_{\Phi}|\varphi\rangle=V_{\Phi}|\varphi\rangle\oplus|0\rangle,\quad
\tilde{V}^C_{\Psi}|\varphi\rangle=(I_B\otimes C)V_{\Psi}|\varphi\rangle\oplus \left(I_B\otimes\sqrt{I_{E}-C^*C}\right)V_{\Psi}|\varphi\rangle$$ for any $\varphi\in\H_A$, where we assume that the operators $V_{\Phi}$ and $V_{\Psi}$ act from $\H_A$ to $\H_B\otimes \H^1_E$ and $\H_B\otimes \H^2_E$ correspondingly, while the contraction $C$ acts from $\H^2_E$ to $\H^1_E$. It is easy to see that the operators $\tilde{V}_{\Phi}$ and $\tilde{V}^C_{\Psi}$ form a common Stinespring representation for the maps $\Phi$ and $\Psi$ with the required property.
To prove that $\N(\Phi,\Psi)\subseteq\M(\Phi,\Psi)$ take any common Stinespring representation for the maps $\Phi$ and $\Psi$ with the operators $\tilde{V}_{\Phi}$ and $\tilde{V}_{\Psi}$ from $\H_A$ to $\H_B\otimes\H_{\tilde{E}}$. By Theorem 6.2.2 in [@H-SCI] there exist partial isometries $W_{\Phi}$ and $W_{\Psi}$ from $\H_E$ to $\H_{\tilde{E}}$ such that $\tilde{V}_{\Phi}=(I_B\otimes W_{\Phi})V_{\Phi}$ and $\tilde{V}_{\Psi}=(I_B\otimes W_{\Psi})V_{\Psi}$. So, $\tilde{V}^*_{\Phi}\tilde{V}_{\Psi}=V^*_{\Phi}(I_B\otimes W^*_{\Phi}W_{\Psi})V_{\Psi}\in \M(\Phi,\Psi)$, since $\|W^*_{\Phi}W_{\Psi}\|\leq 1$.
Since $\N(\Phi,\Psi)=\M(\Phi,\Psi)$, the infimum in (\[beta-e\]) can be taken over the set $\M(\Phi,\Psi)$. This implies $$\label{beta-d}
\begin{array}{rl}
\alpha^G_E(\Phi,\Psi)&
=\displaystyle\inf_{C\in\B_1(\H_E)}\sup_{\rho\in\C^s_{{G},E}}\sqrt{{\mathrm{Tr}}{\hspace{1pt}}\Phi(\rho)+{\mathrm{Tr}}\Psi(\rho)-2\Re{\hspace{1pt}}{\mathrm{Tr}}V_{\Phi}^*(I_B\otimes C)V_{\Psi}\rho}\\
&=\displaystyle\sup_{\rho\in\C^s_{{G},E}}\inf_{C\in\B_1(\H_E)} \sqrt{{\mathrm{Tr}}{\hspace{1pt}}\Phi(\rho)+{\mathrm{Tr}}\Psi(\rho)-2\Re{\hspace{1pt}}{\mathrm{Tr}}V_{\Phi} ^*(I_B\otimes C)V_{\Psi}\rho}\\
&=\displaystyle\sup_{\rho\in\C^s_{{G},E}}\sqrt{{\hspace{1pt}}{\mathrm{Tr}}{\hspace{1pt}}\Phi(\rho)+{\mathrm{Tr}}\Psi(\rho)-2\sup_{C\in\B_1(\H_E)}|{\mathrm{Tr}}V_{\Phi} ^*(I_B\otimes C)V_{\Psi}\rho|},
\end{array}$$ where the possibility to change the order of the optimization follows from Ky Fan’s minimax theorem [@Simons] and the $\sigma$-weak compactness of the unit ball $\B_1(\H_E)$ of $\B(\H_E)$ . It is easy to see that $$\label{a-eq}
\sup_{C\in\B_1(\H_E)}|{\mathrm{Tr}}V_{\Phi} ^*(I_B\otimes C)V_{\Psi}\rho|=\!\!\sup_{C\in\B_1(\H_E)}|\langle V_{\Phi}\otimes I_R{\hspace{1pt}}\varphi|I_{BR}\otimes C |V_{\Psi}\otimes I_R {\hspace{1pt}}\varphi\rangle|,\!$$ where $\varphi$ is a purification of $\rho$, i.e. a vector in $\H_{A}\otimes\H_R$ such that ${\mathrm{Tr}}_R|\varphi\rangle\langle\varphi|=\rho$.
Since for *any* common Stinespring representation (\[c-S-r\]) and any purification $\varphi$ of a state $\rho$ the vectors $V_{\Phi}\otimes I_R{\hspace{1pt}}|\varphi\rangle$ and $V_{\Psi}\otimes I_R{\hspace{1pt}}|\varphi\rangle$ in $\H_{BER}$ are purifications of the operators $\Phi\otimes {\mathrm{Id}}_R(|\varphi\rangle\langle\varphi|)$ and $\Psi\otimes {\mathrm{Id}}_R(|\varphi\rangle\langle\varphi|)$ in $\T(\H_{BR})$, by using the relation $\N(\Phi,\Psi)=\M(\Phi,\Psi)$ proved before and Uhlmann’s theorem [@Uhlmann; @Wilde] it is easy to show that the square of the r.h.s. of (\[a-eq\]) coincides with the fidelity of these operators defined in (\[fidelity\]). Note also that ${\mathrm{Tr}}{\hspace{1pt}}\Phi\otimes {\mathrm{Id}}_R(\sigma)={\mathrm{Tr}}{\hspace{1pt}}\Phi(\sigma_A)$ and ${\mathrm{Tr}}\Psi\otimes {\mathrm{Id}}_R(\sigma)={\mathrm{Tr}}\Psi(\sigma_A)$ for any state $\sigma$ in $\S(\H_{AR})$. By Remark \[ec-b-dist-r\] these observations and (\[beta-d\]) imply that $\alpha^G_E(\Phi,\Psi)=\beta_E^G(\Phi,\Psi)$, i.e. that (\[r-r\]) holds.
The last assertion can be derived from the attainability of the infimum in the first line in (\[beta-d\]) which follows from the $\sigma$-weak compactness of the unit ball $\B_1(\H_E)$. $\square$
Theorem \[KSW-E\] shows continuity of the map $V_{\Phi}\mapsto\Phi$ and selective continuity of the multi-valued map $\Phi\mapsto V_{\Phi}$ with respect to the energy-constrained $cb{\hspace{1pt}}\textrm{-}$norm on the set of CP linear maps $\Phi$ and the operator *E*-norm on the set of Stinespring operators $V_{\Phi}$. Its basic assertion is the equality $$\label{ksw-rel-+}
\beta_E^G(\Phi,\Psi)=\inf_{V_{\Phi},V_{\Psi}} \|V_{\Phi}-V_{\Psi}\|_E^G.$$ Some difficulty of applying Theorem \[KSW-E\] is related to the fact that the infimum in (\[ksw-rel-+\]) is over all common Stinespring representation (\[c-S-r\]). But by using the constructions from the proof of this theorem one can obtain its versions which are more convenient for applications, in particular, for analysis of converging sequences of CP linear maps.
\[KSW-E+\] *Let ${G}$ be a positive semidefinite densely defined operator on $\H_A$ satisfying condition (\[H-cond\]), $\beta_E^G$ and $\|\!\cdot\!\|_E^G$ be, respectively, the energy-constrained Bures distance and the operator $E\textrm{-}$norm induced by ${G}$. Let $\,\Phi$ be a CP linear map from $\T(\H_A)$ to $\T(\H_B)$.*
A\) *There is a Stinespring representation of $\,\Phi$ with the operator $\,V'_{\Phi}:\H_A\rightarrow\H_B\otimes\H_{E'}$ such that $$\label{ksw-rel+}
\beta_E^G(\Phi,\Psi)=\inf_{V_{\Psi}}
\|V'_{\Phi}-V_{\Psi}\|_E^G,$$ for any CP linear map $\Psi:\T(\H_A)\rightarrow\T(\H_B)$, where the infimum is over all Stinespring representations of $\,\Psi$ with the same environment space $\,\H_{E'}$. The infimum in (\[ksw-rel+\]) is attainable.*
B\) *If $\,V_{\Phi}:\H_A\rightarrow\H_B\otimes\H_{E}$ is the operator from a given Stinespring representation of $\,\Phi$ such that $\,\dim\H_{E}=+\infty\,$ then $$\beta_E^G(\Phi,\Psi)\leq \inf_{V_{\Psi}}\|V_{\Phi}-V_{\Psi}\|_E^G\leq 2\beta_E^G(\Phi,\Psi),$$ for any CP linear map $\,\Psi:\T(\H_A)\rightarrow\T(\H_B)$, where the infimum is over all Stinespring representations of $\,\Psi$ with the same environment space $\H_{E}$.*
*Proof.* If $\,V_{\Phi}:\H_A\rightarrow\H_B\otimes\H_{E}$ is the operator from a Stinespring representation of $\,\Phi$ such that $\,\dim\H_E=+\infty\,$ then, since any separable Hilbert space can be isometrically embedded into $\H_E$, we may assume that any CP linear map $\Psi:\T(\H_A)\rightarrow\T(\H_B)$ has a Stinespring representation with the same environment space $\H_E$. Denote by $V_{\Psi}$ the Stinespring operator of $\Psi$ in this representation. Let $\tilde{V}_{\Phi}$ and $\tilde{V}^C_{\Psi}$ be the operators from $\H_A$ into $\H_B\otimes(\H^1_E\oplus\H^2_E)=(\H_B\otimes\H^1_E)\oplus(\H_B\otimes\H^2_E)$ defined by formulae (\[bar-v\]), where $\H^1_E$ and $\H^2_E$ are copies of $\H_E$ and $C$ is a contraction in $\B(\H_E)$. The arguments from the proof of Theorem \[KSW-E\] show that $\,\beta_E^G(\Psi,\Phi)=\|\tilde{V}^{C_0}_{\Psi}-\tilde{V}_{\Phi}\|_E^G$ for some $C_0\in\B(\H_E)$ depending on $\Phi$ and $\Psi$. So, to obtain assertion A it suffices to take $\tilde{V}_{\Phi}$ in the role of $V'_{\Phi}$.
To prove assertion B we will use the above operators $\tilde{V}_{\Phi}$ and $\tilde{V}^{C_0}_{\Psi}$ as follows. Assume first that the operator $C_0$ is nondegenerate, i.e. $\ker C_0=\{0\}$. Let $U$ be the isometry from the polar decomposition of $C_0$, i.e. $C_0=U|C_0|$. Since $\,\|\tilde{V}^{C_0}_{\Psi}-\tilde{V}_{\Phi}\|_E^G=\beta_E^G(\Psi,\Phi)$, it follows from Proposition \[bp-en-0\]C that $$\label{s-ub}
\!\!\|(I_B\otimes C_0)V_{\Psi}-V_{\Phi}\|_E^G\leq\beta_E^G(\Psi,\Phi)\quad\textrm{and}\quad \left\|\left(I_B\otimes\sqrt{I_{E}-|C_0|^2}\right)\!V_{\Psi}\right\|_E^G\!\leq\beta_E^G(\Psi,\Phi)\!$$ Hence the triangle inequality and Proposition \[bp-en-0\]B imply that $$\label{s-ub+}
\begin{array}{c}
\|(I_B\otimes U)V_{\Psi}-V_{\Phi}\|_E^G\leq \|(I_B\otimes C_0)V_{\Psi}-V_{\Phi}\|_E^G\\\\+\|(I_B\otimes C_0)V_{\Psi}-(I_B\otimes U)V_{\Psi}\|_E^G
\leq \beta_E^G(\Psi,\Phi)+\|I_B\otimes (I_E-|C_0|)V_{\Psi}\|_E^G.
\end{array}$$ Since $C_0$ is a contraction, by using Proposition \[bp-en-0\]B and the second inequality in (\[s-ub\]) we obtain $$\|I_B\otimes (I_E-|C_0|)V_{\Psi}\|_E^G\leq\|I_B\otimes (I_E-|C_0|^2)V_{\Psi}\|_E^G\leq\|I_B\otimes \sqrt{I_E-|C_0|^2}V_{\Psi}\|_E^G\leq\beta_E^G(\Psi,\Phi)$$ Thus, it follows from (\[s-ub+\]) that $\|(I_B\otimes U)V_{\Psi}-V_{\Phi}\|_E^G\leq2\beta_E^G(\Psi,\Phi)$. Since $U$ is an isometry, $(I_B\otimes U)V_{\Psi}$ is a Stinespring operator for $\Psi$.
To omit the assumption $\ker C_0=\{0\}$ it suffices to note that nondegenerate operators form a dense subset of $\B_1(\H_E)$ in the weak operator topology and that the expression under the square root in the first line of (\[beta-d\]) is a continuous function of $C$ in this topology. $\square$
If $\{V_n\}$ is a sequence of operators from $\H_A$ to $\H_B\otimes\H_E$ converging to an operator $V_0:\H_A\rightarrow\H_B\otimes\H_E$ w.r.t. the norm $\|\!\cdot\!\|^G_{E}$ then the first inequality in (\[KSW-rel\]) implies that the sequence of CP maps $\,\Phi_n(\rho)={\mathrm{Tr}}_E V_n\rho V^*_n\,$ converges to the map $\,\Phi_0(\rho)={\mathrm{Tr}}_E V_0\rho V^*_0\,$ w.r.t. the norm $\|\!\cdot\!\|^G_{\mathrm{cb},E}$ and for each $n$ the following inequalities hold $$\|\Phi_n-\Phi_0\|^G_{\mathrm{cb},E}\leq\beta^G_{E}(\Phi_n,\Phi_0)\!\left[\sqrt{\|\Phi_n\|^G_{\mathrm{cb},E}}+\sqrt{\|\Phi_0\|^G_{\mathrm{cb},E}}\right]
\leq\|V_{n}-V_{0}\|_E^G\!\left[\|V_{n}\|_E^G+\|V_{0}\|_E^G\right].$$
Theorem \[KSW-E+\] allows to describe all sequences of CP linear maps converging w.r.t. the energy-constrained $cb{\hspace{1pt}}\textrm{-}$norm.
\[KSW-E+c\] *Let $\,\{\Phi_n\}$ be a sequence of CP linear maps from $\,\T(\H_A)$ to $\T(\H_B)$ converging to a CP linear map $\,\Phi_0$ with respect to the norm $\|\!\cdot\!\|^G_{\mathrm{cb},E}$.*
A\) *There exist a separable Hilbert space $\H_{E'}$ and a sequence $\{V_n\}$ of operators from $\H_A$ into $\H_B\otimes\H_{E'}$ converging to an operator $V_0$ with respect to the norm $\|\!\cdot\!\|^G_{E}$ such that $\,\Phi_n(\rho)={\mathrm{Tr}}_{E'} V_n\rho V^*_n\,$ for all $\,n\geq0$ and* $$\|V_{n}-V_{0}\|_E^G=\beta_{E}^G(\Phi_n,\Phi_0)\leq \sqrt{\|\Phi_n-\Phi_0\|^G_{\mathrm{cb},E}}.$$ B) *If $\,V_0:\H_A\rightarrow\H_B\otimes\H_E$ is the operator from a given Stinespring representation of the map $\,\Phi_0$ such that $\,\dim \H_E=+\infty$, then there exists a sequence $\{V_n\}$ of operators from $\H_A$ into $\H_B\otimes\H_E$ converging to the operator $V_0$ with respect to the norm $\|\!\cdot\!\|_{E}^G$ such that $\,\Phi_n(\rho)={\mathrm{Tr}}_E V_n\rho V^*_n\,$ for all $\,n>0$ and* $$\label{2-est}
\|V_{n}-V_{0}\|_E^G\leq2\beta^G_{E}(\Phi_n,\Phi_0)\leq 2\sqrt{\|\Phi_n-\Phi_0\|^G_{\mathrm{cb},E}}.$$
Factor $"2"$ in (\[2-est\]) is a cost of the possibility to take the sequence $\{V_n\}$ of Stinespring operators representing the sequence $\{\Phi_n\}$ for *given* $\H_E$ and $V_0:\H_A\rightarrow\H_B\otimes\H_E$.
If the operator ${G}$ is discrete and unbounded (Def.\[D-H\]) then the norm $\|\!\cdot\!\|^G_{\mathrm{cb},E}$generates the strong convergence topology on bounded subsets of the set $\F(A,B)$ of all CP linear maps from $\T(\H_A)$ to $\T(\H_B)$ (by Proposition 3 in [@SCT]), while the norm $\|\!\cdot\!\|^G_{E}$ generates the strong operator topology on subsets of linear maps from $\H_A$ to $\H_B\otimes\H_E$ bounded by the operator norm (by Proposition \[ec-on-p2\]B). Thus, in this case Corollary \[KSW-E+c\] gives representation of bounded strongly converging sequences of CP linear maps via strongly converging sequence of Stinespring operators. For sequences of quantum channels such representation is obtained in [@CSR] (in the form of part A of Corollary \[KSW-E+c\]).
Operator *E*-norms for unbounded operators
==========================================
In this section we will extend the operator *E*-norms to unbounded operators. We will assume that ${G}$ is a positive semidefinite *unbounded*[^15] operator on $\H$ with dense domain satisfying condition (\[H-cond\]). The case of discrete type operator ${G}$ will be considered separately after formulations of general results.
Speaking about extension of the operator *E*-norms to unbounded operators we may restrict attention to linear operators on $\H$ relatively bounded w.r.t. the operator $\sqrt{G}$, i.e. linear operators $A$ defined on $\mathcal{D}(\sqrt{G})$ such that $$\label{rb-rel}
\|A\varphi\|^2\leq a^2\|\varphi\|^2+b^2\|\sqrt{G}\varphi\|^2,\quad \forall \varphi\in\mathcal{D}(\sqrt{G}),$$ for some nonnegative numbers $a$ and $b$ (depending on $A$ but not depending on $\varphi$)[@Kato]. Such operators are briefly called $\sqrt{G}$-*bounded*. Indeed, it is easy to see that the r.h.s. of (\[ec-on-b\]) is finite for any $\sqrt{G}$-bounded operator $A$ and all $E>0$. The following lemma contains the converse statement (in strengthened form).
\[new-l\]
*Proof.* Consider the set $\S^{\rm f}_{\!G}\doteq \{ \rho\in \S(\H)\,|\,{\mathrm{Tr}}{G}\rho<+\infty, {\mathrm{rank}}\rho<+\infty\}$. A state $\rho$ belongs to this set if and only if it has a finite decomposition $$\label{p-s-d}
\rho=\sum_i |\varphi_i\rangle\langle\varphi_i|,$$ where $\{\varphi_i\}$ is a set of vectors in $\D(\sqrt{G})$. Moreover, any such decomposition of $\rho$ consists of vectors in $\D(\sqrt{G})$.
For any state $\rho$ in $\S^{\rm f}_{\!G}$ with representation (\[p-s-d\]) define the operator $A\rho A^*$ as follows $$\label{ab-d-p}
A\rho A^*\doteq\sum_i|\alpha_i\rangle\langle\alpha_i|,\quad\textrm{where}\quad |\alpha_i\rangle=A|\varphi_i\rangle.$$ By using Schrodinger’s mixture theorem (see ) it is easy to show that the r.h.s. of (\[ab-d-p\]) does not depend on representation (\[p-s-d\]). This implies that $$\label{a-fun}
\rho\mapsto{\mathrm{Tr}}A\rho A^*\doteq\sum_i\|A\varphi_i\|^2$$ is an affine function on $\S^{\rm f}_{\!G}$.
Thus, to prove the first assertion of the lemma it suffices to show that $$\label{ec-on-f}
\|A\|^{G}_E=\sup_{\rho\in\S^{\rm f}_{\!G}:{\mathrm{Tr}}{G}\rho\leq E}\sqrt{{\mathrm{Tr}}A\rho A^*}$$ for any $E>0$. This equality means that the supremum in the r.h.s. of (\[ec-on-f\]) can be taken only over pure states $\rho$ in $\S^{\rm f}_{\!G}$ such that ${\mathrm{Tr}}{G}\rho\leq E$. Since the function (\[a-fun\]) is affine, this can be shown easily by using the fact that all the extreme points of the convex set of states $\rho$ such that ${\mathrm{Tr}}{G}\rho\leq E$ are pure states .
The second assertion of the lemma is derived from the first one, since the concavity of the function $\,E\mapsto \left[\|A\|^{G}_E\right]^2$ implies existence of numbers $a$ and $b$ such that $$[\|A\|^{G}_E]^2\leq a^2+b^2 E\quad \forall E>0.\;\square$$
Thus, in what follows we will consider the operator $E$-norm $\|\cdot\|_E^G$ defined by formula (\[ec-on-b\]) on the set of all $\sqrt{G}$-bounded operators. Below we will show that the quantity ${\mathrm{Tr}}A\rho A^*$ can be defined correctly (without using the notion of adjoint operator) for any $\sqrt{G}$-bounded operator $A$ and any state $\rho$ with finite ${\mathrm{Tr}}G\rho$ (not only for a finite rank state as in the proof of Lemma \[new-l\]). This will allows to show that the operator $E$-norm $\|A\|_E^G$ of any $\sqrt{G}$-bounded operator $A$ can be also defined by formula (\[ec-on\]).
Denote by $\Pi_{\sqrt{G}}(A)$ the set of all pairs $(a,b)$ for which (\[rb-rel\]) holds. It is easy to see that $\Pi_{\sqrt{G}}(A)$ is a closed subset of $\mathbb{\mathbb{R}}^2_+$. The $\sqrt{G}$-bound of $A$ (denoted by $b_{\sqrt{G}}(A)$ in what follows) is defined as $$b_{\sqrt{G}}(A)=\inf\left\{b\,|\,(a,b)\in\Pi_{\sqrt{G}}(A)\right\}.$$ If $\,b_{\sqrt{G}}(A)=0\,$ then $A$ is called $\sqrt{G}$-infinitesimal operator (infinitesimally bounded w.r.t. $\sqrt{G}$). These notions are widely used in the modern operator theory, in particular, in analysis of perturbations of unbounded operators on a Hilbert space .
We will use the following simple lemmas.[^16]
\[p-l-1\] *If $A$ is a $\sqrt{G}$-bounded operator on $\H$ then for any separable Hilbert space $\K$ the operator $A\otimes I_{\K}$ naturally defined on the set $\,\D(\sqrt{G})\otimes\K$ has a unique linear $\sqrt{G}\otimes I_{\K}$-bounded extension to the set $\,\D(\sqrt{G}\otimes I_{\K})$.[^17] This extension (also denoted by $A\otimes I_{\K}$) has the following property $$\label{s-prop}
A\otimes I_{\K}\!\left(\sum_{i}|\varphi_i\rangle\otimes|\psi_i\rangle\right)=\sum_{i}A|\varphi_i\rangle\otimes|\psi_i\rangle$$ for any countable sets $\{\varphi_i\}\subset\D(\sqrt{G})$ and $\{\psi_i\}\subset\K$ such that $\sum_{i}\|\sqrt{G}\varphi_i\|^2<+\infty$, $\sum_{i}\|\varphi_i\|^2<+\infty$ and $\langle\psi_i|\psi_j\rangle=\delta_{ij}$, which implies that $\,\Pi_{\sqrt{G}\otimes I_{\K}}(A\otimes I_{\K})=\Pi_{\sqrt{G}}(A)$.*
*Proof.* For any $E>0$ the linear spaces $\D(\sqrt{G})$ and $\D(\sqrt{G}\otimes I_{\K})$ equipped, respectively, with the inner products $$\langle\varphi|\psi\rangle_E^G=\langle\varphi|\psi\rangle+\langle\varphi|G|\psi\rangle/E\quad\textrm{and}\quad \langle\eta|\theta\rangle_E^{G\otimes I_{\K}}=\langle\eta|\theta\rangle+\langle\eta|G\otimes I_{\K}|\theta\rangle/E$$ are Hilbert spaces . Denote the first space by $\H^G_E$. Then it is easy to see that the second space coincides with the Hilbert space $\H^G_E\otimes\K$. Since the operator $A$ is bounded as an operator from $\H^G_E$ into $\H$ the operator $A\otimes I_{\K}$ defined on $\,\D(\sqrt{G})\otimes\K$ is uniquely extended to a bounded operator from $\H^G_E\otimes\K$ into $\H\otimes\K$. Since the linear spaces $\H^G_E\otimes\K$ and $\D(\sqrt{G}\otimes I_{\K})$ coincide, this extension is a $\sqrt{G}\otimes I_{\K}$-bounded linear operator on $\H\otimes\K$.
Property (\[s-prop\]) follows from continuity of the operator $A\otimes I_{\K}:\H^G_E\otimes\K\rightarrow \H\otimes\K$.
Any vector $\eta$ in $\D(\sqrt{G}\otimes I_{\K})$ can be represented as $|\eta\rangle=\sum_{i}|\varphi_i\rangle\otimes|\psi_i\rangle$, where $\{\varphi_i\}\subset\D(\sqrt{G})$ and $\{\psi_i\}\subset\K$ are collections of vectors such that $\sum_{i}\|\sqrt{G}\varphi_i\|^2<+\infty$, $\sum_{i}\|\varphi_i\|^2<+\infty$ and $\langle\psi_i|\psi_j\rangle=\delta_{ij}$. By using property (\[s-prop\]) we obtain $$\|A\otimes I_{\K}\eta\|^2=\sum_{i}\|A\varphi_i\|^2\leq a^2\sum_{i}\|\varphi_i\|^2+b^2\sum_{i}\|\sqrt{G}\varphi_i\|^2=a^2\|\eta\|^2+b^2\|\sqrt{G}\otimes I_{\K}\eta\|^2$$ for any $(a,b)\in\Pi_{\sqrt{G}}(A)$. This implies $\,\Pi_{\sqrt{G}}(A)\subseteq\Pi_{\sqrt{G}\otimes I_{\K}}(A\otimes I_{\K})$, and hence $\,\Pi_{\sqrt{G}}(A)=\Pi_{\sqrt{G}\otimes I_{\K}}(A\otimes I_{\K})$, since the converse inclusion is obvious. $\square$
\[w-c\] Property (\[s-prop\]) implies that $$(A\otimes I_{\K})(I_{\H}\otimes W)|\varphi\rangle=(I_{\H}\otimes W)(A\otimes I_{\K})|\varphi\rangle$$ for any $\varphi\in\D(\sqrt{G}\otimes I_{\K})$ and a partial isometry $W\in\B(\K)$ s.t. $I_{\H}\otimes W^{*}W |\varphi\rangle=|\varphi\rangle$.
\[p-l-2\] *For any $\sqrt{G}$-bounded operators $A$ and $B$ on $\H$ the affine function ${\hspace{1pt}}\rho\mapsto A\rho B^*\in\T(\H)$ is well defined on the set $\,\T^+_{G}\doteq \{\rho\in\T_{+}(\H)\,|\,{\mathrm{Tr}}{G}\rho<+\infty{\hspace{1pt}}\}\,$ by the formula [^18] $$\label{ab-d}
A\rho B^*\doteq\sum_i|\alpha_i\rangle\langle\beta_i|,\qquad |\alpha_i\rangle=A|\varphi_i\rangle,\;|\beta_i\rangle=B|\varphi_i\rangle,$$ where $\rho=\sum_i |\varphi_i\rangle\langle\varphi_i|$ is any decomposition of $\rho\in\T^+_{G}$ into $1$-rank positive operators.*
*Proof.* If $\rho=\sum_i |\varphi_i\rangle\langle\varphi_i|$ and $\{\psi_i\}$ is any set of orthogonal unit vectors in a separable Hilbert space $\K$ then $|\eta\rangle=\sum_i |\varphi_i\rangle\otimes|\psi_i\rangle$ is a vector in $\,\D(\sqrt{G}\otimes I_{\K})$ such that $\rho={\mathrm{Tr}}_{\K}|\eta\rangle\langle\eta|$. By Lemma \[p-l-1\] the operators $A\otimes I_{\K}$ and $B\otimes I_{\K}$ have unique linear $\sqrt{G}\otimes I_{\K}$-bounded extensions to the set $\,\D(\sqrt{G}\otimes I_{\K})$ satisfying (\[s-prop\]). Hence $$\label{lf}
\sum_i |A\varphi_i\rangle\langle B\varphi_i|={\mathrm{Tr}}_{\K}|A\otimes I_{\K}\eta\rangle\langle B\otimes I_{\K}\eta|.$$ So, by using the well known relation between different purifications of a given state [@H-SCI; @Wilde] and Remark \[w-c\], it is easy to show that the r.h.s. of (\[lf\]) does not depend on the representation $\rho=\sum_i |\varphi_i\rangle\langle\varphi_i|$. It follows that the r.h.s. of (\[lf\]) correctly defines an affine function $\rho\mapsto A\rho B^*$ on the set $\T^+_{{G}}$. $\square$
Lemma \[p-l-2\] implies, in particular, that $\rho\mapsto{\mathrm{Tr}}A\rho A^*$ is an affine function from $\T^+_{G}$ into $\T_+(\H)$ (well defined by formula (\[ab-d\]) with $B=A$) for any $\sqrt{G}$-bounded operator $A$. Hence the r.h.s. of (\[ec-on\]) is well defined for any such operator.
The following proposition shows that we may also define the operator *E*-norm $\|A\|_E^G$ of any $\sqrt{G}$-bounded operator $A$ by formula (\[ec-on\]).
\[v-new\] *Let $A$ be an arbitrary $\sqrt{G}$-bounded operator and $E>0$.*
A\) *The right hand sides of (\[ec-on-b\]) and (\[ec-on\]) coincide (provided that $A\rho A^*$ is defined by formula (\[ab-d\]) with $B=A$).*
B\) *The suprema in definitions (\[ec-on-b\]) and (\[ec-on\]) can be taken, respectively, over all vectors in $\H_{\leq 1}$ satisfying the condition $\,\langle\varphi|G|\varphi\rangle\leq E$ and over all operators in $\T_{+,1}(\H)$ satisfying the condition $\,{\mathrm{Tr}}{G}\rho\leq E$.*
*Proof.* A) Let $\{P_n\}$ be a sequence of finite-rank projectors strongly converging to the unit operator $I_{\H}$. Then $P_n A$ is (extented to) a bounded operator on $\H$ for each $n$. Denote by $u(A)$ the r.h.s. of (\[ec-on\]). It is easy to see that $u(A)=\sup_n \|P_nA\|^{G}_E$. So, by using the coincidence of the r.h.s. of (\[ec-on-b\]) and (\[ec-on\]) for any bounded operator $A$ we obtain $$u(A)=\sup_n \sup_{\varphi\in\H_1,\langle\varphi|G|\varphi\rangle\leq E}\|P_n A\varphi\|.$$ Since $\|A\varphi\|\geq\|P_n A\varphi\|$ for any $\varphi$, this implies that the r.h.s. of (\[ec-on-b\]) is not less than $u(A)$. The converse inequality is obvious.
B\) This assertion is proved by using concavity of the function $E\mapsto \left[\|A\|^{G}_E\right]^2$ (Lemma \[new-l\]A) and the proof of Proposition \[en-def\]A. $\square$
By Proposition \[v-new\]B for any vector $\varphi$ in $\D(\sqrt{G})$ such that $\|\varphi\|\leq 1$ we have $$\label{a-phi-est}
\|A\varphi\|\leq \|A\|^{G}_{E_{\varphi}}\leq K_{\varphi}\|A\|^{G}_E,\vspace{-5pt}$$ where $\,E_{\varphi}=\|\sqrt{G}\varphi\|^2\,$ and $\,K_{\varphi}=\max\{1, \sqrt{E_{\varphi}/E}\}\,$. This implies the following
\[vsl\] *Let $P_E$ be the spectral projector of ${G}$ corresponding to the interval $[0,E]$. For any $\sqrt{G}$-bounded operator $A$ the operator $AP_E$ is bounded and $\|AP_E\|\leq \|A\|^{G}_E$.*
*Proof.* It follows from (\[a-phi-est\]) that $\|AP_E\varphi\|\leq \|A\|^{G}_E$ for any unit vector $\varphi$ in $\H$, since $\|\sqrt{G}P_E\varphi\|^2\leq E$ and $\|P_E\varphi\|\leq 1$. $\square$
The following lemma shows that the set $\,\Pi_{\sqrt{G}}(A)$ is completely determined by the function $\,E\mapsto \left[\|A\|^{G}_E\right]^2$ and vice versa.
\[bl\] *A pair $(a,b)$ belongs to the set $\,\Pi_{\sqrt{G}}(A)$ if and only if $\|A\|^{G}_{E}\leq \sqrt{a^2+b^2 E}$ for all $\,E>0$.*
*Proof.* If $\|A\|^{G}_{E}\leq \sqrt{a^2+b^2 E}$ then it follows from (\[a-phi-est\]) that $$\|A\varphi\|\leq \|A\|^{G}_{\|\sqrt{G}\varphi\|^2}\leq \sqrt{a^2+b^2 \|\sqrt{G}\varphi\|^2}$$ for any unit vector $\varphi$ in $\D(\sqrt{G})$. Hence $(a,b)\in\Pi_{\sqrt{G}}(A)$. If $(a,b)\in\Pi_{\sqrt{G}}(A)$ then $$\sup\left\{\|A\varphi\|\,\left|\, \varphi\in\D(\sqrt{G}), \|\varphi\|\leq1, \|\sqrt{G}\varphi\|^2\leq E\right\}\right.\leq\sqrt{a^2+b^2 E}$$ for any $E>0$. So, definition (\[ec-on-b\]) implies that $\|A\|^{G}_{E}\leq \sqrt{a^2+b^2 E}$. $\square$
Denote by $\B_{\!G}(\H)$ the linear space of all $\sqrt{G}$-bounded operators equipped with the norm $\|\!\cdot\!\|^{G}_E$ defined by the equivalent expressions (\[ec-on-b\]) and (\[ec-on\]) (we identify operators coinciding on $\D(\sqrt{G})$). We will also consider the norm ${\|\hspace{-1pt}|}\!\cdot\!{\|\hspace{-1pt}|}^{G}_E$ defined in (\[eq-norms-2\]), which is commonly used on the space of $\sqrt{G}$-bounded operators.
\[comp-d\] *Let ${G}$ be a positive semidefinite unbounded densely defined operator on $\H$ satisfying condition (\[H-cond\]).*
A\) *$\B_{\!G}(\H)$ is a nonseparable Banach space. The norms $\|\cdot\|^{G}_{E}$ and ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}^{G}_{E}$ satisfy the equivalence relations (\[E-n-eq\]), (\[one\]) and (\[E-n-eq+\]) on $\,\B_{\!G}(\H)$ . For any $A\in\B_{\!G}(\H)$ and $E>0$ the following expressions hold* $${\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}=\sup_{t>0}\|A\|^{G}_{tE}/\sqrt{1+t},\qquad \|A\|^{G}_{E}=\inf_{t>0}{\|\hspace{-1pt}|}A {\|\hspace{-1pt}|}^{G}_{tE}\sqrt{1+1/t}.$$
B\) *If $A\in\B_G(\H)$ and $E>0$ then* $$\|A\|^{G}_{E}=\inf\left.\left\{\sqrt{a^2+b^2 E}\;\right|(a,b)\in\Pi_{\sqrt{G}}(A)\right\} \quad \textit{and} \quad b_{\sqrt{G}}(A)=\lim_{E\rightarrow+\infty}\|A\|^{G}_{E}/\sqrt{E}.$$
*The limit in the last formula can be replaced by the infimum over all $E>0$.*
C\) *The completion of $\B(\H)$ w.r.t. any of the norms $\|\!\cdot\!\|^{G}_E$, $E>0$, coincides with the closed subspace $\,\B^0_{\!G}(\H)$ of $\,\B_{\!G}(\H)$ consisting of all $\sqrt{G}$-infinitesimal operators, i.e. operators with the $\sqrt{G}$-bound equal to $0$. An operator $A$ belongs to $\B^0_{\!G}(\H)$ if and only if*$$\label{s-cond}
\|A\|^{G}_E=o{\hspace{1pt}}(\sqrt{E})\quad\textup{ as }\quad E\rightarrow+\infty.$$
*If $\,G$ is a discrete operator (Def.\[D-H\]) then the Banach space $\B^0_{\!G}(\H)$ is separable.*
D\) *Any ball in $\B(\H)$ is complete with respect to any of the norms $\|\!\cdot\!\|^{G}_E$, $E>0$. An operator $A$ belongs to $\B(\H)$ if and only if the function $E\mapsto \|A\|^{G}_E$ is bounded. In this case $\|A\|=\sup_{E>0}\|A\|^{G}_E=\lim_{E\rightarrow+\infty}\|A\|^{G}_E$.*
E\) *The $\sqrt{G}$-bound is a continuous seminorm on $\B_G(\H)$. Quantitatively, $$\label{b-cb}
\left|{\hspace{1pt}}b_{\sqrt{G}}(A)-b_{\sqrt{G}}(B)\right|\leq b_{\sqrt{G}}(A-B)\leq \|A-B\|^{G}_{E}/\sqrt{E}$$ for arbitrary $A,B$ in $\B_G(\H)$ and any $E>0$.*
F\) *If $\,\K$ is a separable Hilbert space then $\,\|A\otimes I_{\K}\|_E^{G\otimes I_{\K}}=\|A\|_E^{G}$ for any $A\in\B_G(\H)$*[^19]
G\) *For arbitrary $\sqrt{G}$-bounded operators $A$ and $B$ and any operator $\rho$ in $\T_{+}(\H)$ such that ${\mathrm{Tr}}\rho\leq 1$ and $\,E_{\rho}\doteq{\mathrm{Tr}}G\rho<+\infty\,$ the following inequalities hold $$|{\mathrm{Tr}}A\rho B^*|\leq \|A\rho B^*\|_1 \leq \|A\|^{G}_{E_{\rho}}\|B\|^{G}_{E_{\rho}},$$ where $A\rho B^*$ is the trace class operator defined in (\[ab-d\])*.
H\) *For any $A$ in $\B_{\!G}(\H)$ and $E>0$ the suprema in definitions (\[ec-on-b\]) and (\[ec-on\]) can be taken, respectively, over all unit vectors in $\H$ satisfying the condition $\,\langle\varphi|G|\varphi\rangle=E$ and over all states in $\S(\H)$ satisfying the condition $\,{\mathrm{Tr}}{G}\rho=E$.*
I\) *If the operator $G$ is discrete (Def.\[D-H\]) then for any $A\in\B^0_G(\H)$ and $E>0$ the suprema in (\[ec-on-b\]) and (\[ec-on\]) are attainable. Moreover, if $A$ is a unbounded operator in $\B^0_G(\H)$ or $\|A\|^{G}_{E}<\|A\|<+\infty$ then the suprema in (\[ec-on-b\]) and (\[ec-on\]) are attained, respectively, at unit vector $\varphi_0$ in $\H$ such that $\,\langle\varphi_0|G|\varphi_0\rangle=E$ and at a state $\rho_0$ in $\S(\H)$ such that $\,{\mathrm{Tr}}G\rho_0=E$.*
\[comp-d-r+\] The expressions in Theorem 3A show that the functions $E\mapsto\|A\|^G_E$ and $E\mapsto{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^G_E$ are completely determined by each other for arbitrary $\sqrt{G}$-bounded operator $A$. So, if $\|A\|^G_E=\|B\|^G_E$ for all $E>0$ for some $\sqrt{G}$-bounded operators $A$ and $B$ then ${\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^G_E={\|\hspace{-1pt}|}B{\|\hspace{-1pt}|}^G_E$ for all $E>0$ and vise versa.[^20] These expressions mean that the concave function $E\mapsto[\|A\|^G_E]^2$ and the function $E\mapsto[{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^G_E]^2$ are related by the transformations $$F[f](x)\ = \ \sup_{\, t > 0}\ \frac{f(xt)}{t+1}
\quad \textrm{and} \quad G[f](x)\ = \ \inf_{\, t > 0}\ f(xt)\, (1 \, + \, 1/t),$$ defined on the set of nonnegative functions on $(0,+\infty)$. In it is shown that $G\circ F$ maps *any* nonnegative function $f$ on $(0,+\infty)$ into its concave hull and hence $G[F[f]]=f$ for any concave nonnegative function $f$. This shows that the second expression in Theorem 3A follows from the first one and the concavity of the function $E\mapsto[\|A\|^G_E]^2$.
\[comp-d-r\] The below proof of the density of $\B(\H)$ in $\B^0_{\!G}(\H)$ shows that the $\sqrt{G}$-infinitesimality criterion (\[s-cond\]) is equivalent to the following one $$\label{s-cond+}
\lim_{n\rightarrow+\infty} \|A\bar{P}_n\|^{G}_E=0,$$ where $\bar{P}_n$ is the spectral projector of ${G}$ corresponding to the interval $\,(n,+\infty)$.
*Proof of Theorem \[comp-d\].* A) By Lemma \[new-l\]A inequalities (\[E-n-eq\]), (\[one\]) and (\[E-n-eq+\]) for any $A$ in $\B_{\!G}(\H)$ are proved by the same arguments as for a bounded operator $A$. By using definitions (\[ec-on-b\]) and (\[eq-norms-2\]) it is easy to show that $${\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}=\sup_{r\in(0,1)} \sqrt{r}\|A\|^{G}_{\frac{1-r}{r}E},\quad A\in\B_{\!G}(\H).$$ The first expression in A follows from this one by the change of variables $t=(1-r)/r$.
The second expression in A is derived from the first formula in part B proved below by noting that the infimum in that formula can be taken over all the pairs $({\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G,{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G/\sqrt{E}), E>0$. This follows from density of the set $$\left.\left\{ \left({\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G+x, {\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G/\sqrt{E}+y\right)\,\right|\, E>0,\, x,y\geq 0 \right\}$$ in $\Pi_{\sqrt{G}}(A)$, which can be proved by noting that ${\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G=\min\{ a\,|\,(a, a/\sqrt{E})\in \Pi_{\sqrt{G}}(A)\}$.
Denote by $\H_E^G$ the Hilbert space obtained by equipping the linear space $\D(\sqrt{G})$ with the inner product $$\langle\varphi|\psi\rangle_E^G=\langle\varphi|\psi\rangle+\langle\varphi|G|\psi\rangle/E.$$
Since the norm ${\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}_E^G$ of any $\sqrt{G}$-bounded operator $A$ is the operator norm of $A$ treated as a bounded operator from $\H_E^G$ into $\H$, the linear space of all $\sqrt{G}$-bounded operators equipped with the norm ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}_E^G$ is a nonseparable Banach space . Hence, the equivalence of the norms ${\|\hspace{-1pt}|}\cdot{\|\hspace{-1pt}|}_E^G$ and $\|\cdot\|_E^G$ implies that $\B_{\!G}(\H)$ is a nonseparable Banach space.
B\) Since $E\mapsto [\|A\|^{G}_{E}]^2$ is a concave nonnegative function on $\mathbb{R}_+$, it coincides with the infimum of all linear functions $E\mapsto a^2+b^2 E$ such that $[\|A\|^{G}_{E}]^2\leq a^2+b^2 E$ for all $E>0$. The concavity of the function $E\mapsto [\|A\|^{G}_{E}]^2$ implies that the function $E\mapsto [\|A\|^{G}_{E}]^2/E$ is non-increasing. So, both formulae in part B follow from Lemma \[bl\].
C\) The continuity and the seminorm properties of the function $A\mapsto b_{\sqrt{G}}(A)$ stated in part E proved below show that $\B^0_{\!G}(\H)=b^{-1}_{\sqrt{G}}(0)$ is a closed subspace of $\B_{\!G}(\H)$. The characterizing property (\[s-cond\]) follows from the second formula in part B.
To prove density of $\B(\H)$ in $\B^0_{\!G}(\H)$ it suffices, by Lemma \[vsl\], to show that for any $A\in\B^0_{\!G}(\H)$ the sequence $\{AP_n\}$, where $P_n$ is the spectral projector of $G$ corresponding to the interval $[0,n]$, converges to $A$ with respect to the norm $\|\!\cdot\!\|^{G}_E$. For given $P_n$ let $\varphi$ be any unit vector such that $\langle\varphi|G|\varphi\rangle\leq E$ and $x_n=\langle\varphi|\bar{P}_n|\varphi\rangle>0$, where $\bar{P}_n=I_{\H}-P_n$. Let $|\varphi_n\rangle=x^{-1/2}_n \bar{P}_n|\varphi\rangle$. We have $$\|A\bar{P}_n \varphi\|^2=x_n\|A\varphi_n\|^2\leq x_n\left[\|A\|^{G}_{E/x_n}\right]^2\leq (E/n)\left[\|A\|^{G}_{n}\right]^2.$$ The first inequality follows from definition (\[ec-on-b\]) of the operator *E*-norm and the inequality $\langle\varphi_n|G|\varphi_n\rangle\leq E/x_n$, the second one follows from concavity of the function $E\mapsto \left[\|A\|^{G}_E\right]^2$, Lemma \[WL\] and the inequality $x_n\leq E/n$ (which holds, since $\langle\varphi|G|\varphi\rangle\leq E$). The above estimate implies that $$\|A-AP_n\|^{G}_{E}\doteq \sup_{\varphi\in\mathcal{H}_1:
\langle\varphi|G|\varphi\rangle\leq E}\|A\bar{P}_n \varphi\|
\leq \sqrt{E/n}\|A\|^{G}_{n}.$$ So, condition (\[s-cond\]) guarantees that $\|A-AP_n\|^{G}_{E}$ tends to zero as $\,n\rightarrow+\infty$.
The above arguments and Lemma \[vsl\] imply that (\[s-cond\]) is equivalent to (\[s-cond+\]).
If $G$ is a discrete operator then the separability of $\B^0_{\!G}(\H)$ follows from separability of $\B(\H)$ w.r.t. any of the operator *E*-norms, which can be easily shown by using Proposition \[ec-on-p2\]B and separability of $\B(\H)$ w.r.t. the strong operator topology.
D\) We begin with the second assertion. The “only if” part of this assertion and the expression $\|A\|=\sup_{E>0}\|A\|^{G}_E$ follow from Proposition \[ec-on-p1\]. If $\|A\|^{G}_E\leq M<+\infty$ for all $E>0$ then it follows from (\[a-phi-est\]) that $\|A\varphi\|\leq M$ for any unit vector $\varphi$ in $\D(\sqrt{{G}})$. Since $\D(\sqrt{{G}})$ is dense in $\H$, this implies that $A\in\B(\H)$.
To prove the first assertion assume that $\{A_n\}$ is a sequence in $\B(\H)$ converging to an operator $A_0\in\B^0_{\!G}(\H)$ such that $\|A_n\|\leq M<+\infty$ for all $n$. Since $\,\|A_n\|^{G}_E\leq\|A_n\|\leq M$ for all $n$ and $E>0$ and the right hand side of the inequality $$\left|\|A_n\|^{G}_E-\|A_0\|^{G}_E\right|\leq\|A_n-A_0\|^{G}_E$$ tends to zero as $n\rightarrow+\infty$ for any $E>0$, it is easy to see that $\|A_0\|^{G}_E\leq M$ for all $E$. Thus, $\|A_0\|\leq M$ by the assertion proved before.
E\) The seminorm properites of $b_{\sqrt{G}}(\cdot)$ follow from the second formula in part B of the theorem. So, since the function $E\mapsto [\|A\|^{G}_{E}]^2/E$ is non-increasing for any given $A\in\B_{\!G}(\H)$, the inequality (\[b-cb\]) follows from the triangle inequality for $b_{\sqrt{G}}(\cdot)$.
F\) This assertion follows from Lemma \[p-l-1\] and the first formula in part B of the theorem.
G\) Let $\rho=\sum_i |\varphi_i\rangle\langle\varphi_i|$ be a decomposition into $1$-rank positive operators and $\{\psi_i\}$ a set of orthogonal unit vectors in a separable Hilbert space $\K$ then $|\eta\rangle=\sum_i |\varphi_i\rangle\otimes|\psi_i\rangle$ is a vector in $\,\D(\sqrt{G}\otimes I_{\K})$ such that $\rho={\mathrm{Tr}}_{\K}|\eta\rangle\langle\eta|$. By Lemma \[p-l-1\] the operators $A\otimes I_{\K}$ and $B\otimes I_{\K}$ have unique $\sqrt{G}\otimes I_{\K}$-bounded linear extensions to the set $\,\D(\sqrt{G}\otimes I_{\K})$ satisfying (\[s-prop\]). By the monotonicity of the trace norm we have $$\|A\rho B^*\|_1\leq\||A\otimes I_{\K}\eta\rangle\langle B\otimes I_{\K}\eta|\|_1\leq\|A\otimes I_{\K}\eta\|\| B\otimes I_{\K}\eta\|\leq\|A\otimes I_{\K}\|^{{G}\otimes I_{\K}}_{E_{\rho}}\|B\otimes I_{\K}\|^{{G}\otimes I_{\K}}_{E_{\rho}}.$$ By part F of the theorem the r.h.s. of this inequality is equal to $\|A\|^{{G}}_{E_{\rho}}\|B\|^{{G}}_{E_{\rho}}$.
H\) If $A$ is a bounded operator then the possibility to take the suprema in (\[ec-on-b\]) and (\[ec-on\]) over all unit vectors in $\H$ satisfying the condition $\,\langle\varphi|G|\varphi\rangle=E$ and over all states in $\S(\H)$ satisfying the condition $\,{\mathrm{Tr}}{G}\rho=E$ correspondingly follows Proposition \[en-def\]B. If $A$ is a unbounded operator then this possibility can be easily shown by noting that the function $E\mapsto\|A\|^{G}_E$ is strictly increasing on $\mathbb{R}_+$ (since it is concave on $\mathbb{R}_+$ and tends to $+\infty$ as $E\rightarrow+\infty$).
I\) If $A\in\B_{\!G}^0(\H)$ then $\rho\mapsto {\mathrm{Tr}}A\rho A^*$ is an affine continuous function on the set $\mathfrak{C}_{{G},E}\doteq \{\rho\in\T_{+}(\H)\,|\,{\mathrm{Tr}}\rho\leq 1, {\mathrm{Tr}}{G}\rho\leq E{\hspace{1pt}}\}$ for any $E>0$ by Corollary \[vbl-c\] below (proved independently). So, both assertions are proved by repeating the arguments from the proof of Proposition \[en-def\]C. $\square$
**Example: the operators associated with the Heisenberg Commutation Relation**
Let $\H = L_2(\mathbb{R})$ and $S(\mathbb{R})$ be the set of infinitely differentiable rapidly decreasing functions with all the derivatives tending to zero quicker than any degree of $|x|$ when $|x|\rightarrow+\infty$. Consider the operators $q$ and $p$ defined on the set $S(\mathbb{R})$ by setting $$(q\varphi)(x) = x\varphi(x)\quad\textrm{and}\quad (p{\hspace{1pt}}\varphi)(x) = \frac{1}{i}\frac{d}{dx}\varphi(x).$$ These operators are essentially self-adjoint. They represent (sharp) real observables of position and momentum of a quantum particle in the system of units where Planck’s constant $\hbar$ equals to $1$ [@H-SCI Ch.12]. On the domain $S(\mathbb{R})$ these operators satisfy the Heisenberg commutation relation $$\label{H-cr}
[q, p{\hspace{1pt}}] = i I_{\H}.$$
For given $\omega>0$ consider the operators $$\label{a-oper-d}
a=(\omega q+ip)/\sqrt{2\omega}\quad \textrm{and} \quad a^{\dagger}=(\omega q-ip)/\sqrt{2\omega}$$ defined on $S(\mathbb{R})$. Via these operators the commutation relation (\[H-cr\]) can be rewritten as $
[a,a^{\dagger}] = I_{\H}.
$ The operator $$\label{a-N}
N=a^{\dagger}a=aa^{\dagger}-I_{\H}$$ is positive and essentially self-adjoint. It represents (sharp) real observable of the number of quanta of the harmonic oscillator with the frequency $\omega$. The selfadjoint extension of $N$ has the form (\[H-rep\]) with $E_n=n$ and the basic $\{\tau_n\}$ of eigenvectors of $N$ which can be described as follows $$\tau_0(x)=\sqrt[4]{\frac{\omega}{\pi}}\exp\left[-\frac{\omega x^2}{2}\right],\quad |\tau_n\rangle=\frac{1}{\sqrt{n!}}\,[a^{\dagger}]^n|\tau_{0}\rangle,\; n\geq 1.$$ So, $N$ is a positive unbounded discrete (Def.\[D-H\]) operator satisfying condition (\[H-cond\]).
The operators $a$ and $a^{\dagger}=a^*$ are called *annihilation* and *creation* operators correspondingly, since $$\label{a-oper}
a|\tau_0\rangle=0,\quad a|\tau_n\rangle=\sqrt{n}|\tau_{n-1}\rangle\quad\textrm{and}\quad a^{\dagger}|\tau_n\rangle=\sqrt{n+1}|\tau_{n+1}\rangle.$$ So, the operators $a$ and $a^{\dagger}$ are correctly extended to the set $$\D(\sqrt{N})=\left\{\varphi\in\H\,\left|\,\sum_{n=0}^{\infty}n |\langle\varphi|\tau_n\rangle|^2<+\infty\right.\right\}$$ By using relations (\[a-oper-d\]) the operators $p$ and $q$ are also extended to the set $\D(\sqrt{N})$.
We will estimate the operator *E*-norm of the operators $q$, $p$, $a$ and $a^{\dagger}$ induced by the operator $N$ (which up to the constant summand coincides with the Hamiltonian of a quantum oscillator). By using (\[a-N\]) it is easy to show that $\|a\|_E^N=\|\sqrt{N}\|_E^N=\sqrt{E}$ and $\|a^{\dagger}\|_E^N=\sqrt{E+1}$ for any $E>0$. For the operators $\,q=(a^{\dagger}+a)/\sqrt{2\omega}\,$ and $\,p=i\sqrt{\omega/2}(a^{\dagger}-a)\,$ one can obtain the following estimates $$\label{QP-en}
\sqrt{\frac{2E+1/2}{\omega}}< \|q\|^{N}_E\leq\sqrt{\frac{2E+1}{\omega}},\quad \sqrt{(2E+1/2)\omega}< \|p\|^N_E\leq\sqrt{(2E+1)\omega}$$ (the *E*-norms of $q$ and $p$ depend on $\omega$, since the operator $N$ depends on $\omega$). The right inequalities in (\[QP-en\]) directly follow from the triangle inequality and the above expressions for $\|a\|_E^N$ and $\|a^{\dagger}\|_E^N$. To prove the left inequalities in (\[QP-en\]) it suffices to show that $$\sup_{\substack{\|\varphi\|=1,{\hspace{1pt}}\langle\varphi|N|\varphi\rangle\leq E}}\|(a^{\dagger}\pm a)\varphi\|>\sqrt{4E+1}.$$ This can be easily done by using the unit vectors $|\varphi_{\pm}\rangle= \sqrt{1-r}\sum_{n=0}^{+\infty} c^{\pm}_n r^{n/2}|\tau_n\rangle$, where $r=E/(E+1)$, $c_n^{-}=e^{i\pi n/2}$ and $c^+_n=1$ for all $n$.
By using the first expression in Theorem \[comp-d\]A and the above estimates of the norms $\|a\|_E^N$, $\|a^{\dagger}\|_E^N$, $\|p\|_E^N$ and $\|q\|_E^N$ we obtain ${\|\hspace{-1pt}|}a{\|\hspace{-1pt}|}_E^N=\max\{1,\sqrt{E}\}$, ${\|\hspace{-1pt}|}a^{\dagger}{\|\hspace{-1pt}|}_E^N=\sqrt{E}$, $$\sqrt{\frac{l(E)}{\omega}}< {\|\hspace{-1pt}|}q{\|\hspace{-1pt}|}^{N}_E\leq\sqrt{\frac{u(E)}{\omega}}\quad \textrm{and} \quad \sqrt{l(E)\omega}< {\|\hspace{-1pt}|}p{\|\hspace{-1pt}|}^N_E\leq\sqrt{u(E)\omega},$$ where $l(E)=\max\{1/2, 2E\}$ and $u(E)=\max\{1, 2E\}$.
The second formula in Theorem \[comp-d\]B and the above estimates of the norms $\|a\|_E^N,$ $\|a^{\dagger}\|_E^N$, $\|p\|_E^N$ and $\|q\|_E^N$ imply that $$b_{\sqrt{N}}(a)=b_{\sqrt{N}}(a^{\dagger})=1,\qquad b_{\sqrt{N}}(q)=\sqrt{2/\omega}\quad \textrm{and}\quad b_{\sqrt{N}}(p)=\sqrt{2\omega}.$$
So, the operators $q$, $p$, $a$ and $a^{\dagger}$ belong to the Banach space $\B_{\!N}(\H)$ but not lie in the completion $\B^0_{\!N}(\H)$ of $\B(\H)$ w.r.t. the norm $\|\!\cdot\!\|^N_E$.
For any $t<1$ let $a_t$ and $a_t^{\dagger}$ be the operators defined on the set $\D(\sqrt{N})$ by settings $$\label{a-oper}
a_t|\tau_0\rangle=0,\quad a_t|\tau_n\rangle=n^{t/2}|\tau_{n-1}\rangle\quad\textrm{and}\quad a_t^{\dagger}|\tau_n\rangle=(n+1)^{t/2}|\tau_{n+1}\rangle.$$ It is easy to show that $$\label{a-soc}
\lim_{t\rightarrow 1-0} a_t|\varphi\rangle=a|\varphi\rangle\quad \textrm{and} \quad
\lim_{t\rightarrow 1-0} a_t^{\dagger}|\varphi\rangle=a^{\dagger}|\varphi\rangle\quad \textrm{for any}\;\; \varphi\in\D(\sqrt{N}).$$
Since $a_t^{\dagger}a_t=N^t$ and $a_ta_t^{\dagger}=(N+I_{\H})^t$, by using concavity of the function $x\mapsto x^{t}$, we obtain $$\|a_t\|_E^N\leq \sqrt{\sup_{{\mathrm{Tr}}N\rho\leq E } [{\mathrm{Tr}}N \rho{\hspace{1pt}}]^t}=E^{t/2},\quad \|a^{\dagger}_t\|_E^N\leq \sqrt{\sup_{{\mathrm{Tr}}N\rho\leq E } [{\mathrm{Tr}}(N+I_{\H}) \rho{\hspace{1pt}}]^t}=(E+1)^{t/2}.$$
So, the operators $a_t$ and $a_t^{\dagger}$ belong to the space $\B^0_{\!N}(\H)$ for all $t<1$ (since they satisfy condition (\[s-cond\])), while the “limit” operators $a$ and $a^{\dagger}$ lie in $\B_{\!N}(\H)\setminus\B^0_{\!N}(\H)$. So, $a_t$ and $a_t^{\dagger}$ do not tend to $a$ and $a^{\dagger}$ as $t\rightarrow 1$ w.r.t. the norm $\|\!\cdot\!\|_E^N$ in spite of the strong operator convergence (\[a-soc\]).
\[thelast\] It follows from (\[a-phi-est\]) that $$\label{vlast-eq}
\|\!\cdot\!\|^{G}_E\,\textrm{-}\lim_{n\rightarrow\infty}A_n=A_0\quad\Rightarrow\quad\lim_{n\rightarrow\infty}A_n|\varphi\rangle=A_0|\varphi\rangle \quad\forall \varphi\in\D(\sqrt{G})$$ for a sequence $\{A_n\}\subset\B_{\!G}(\H)$. The above example shows that the converse implication is not valid even in the case of discrete operator $G$ (in this case $"\Leftrightarrow"$ holds in (\[vlast-eq\]) for any bounded sequence $\{A_n\}\subset\B(\H)$ by Proposition 2B).
In the last part of this section we consider properties of the Banach space $\B^0_{\!G}(\H)$.
\[vbl\] *If $A\in\B^0_{\!G}(\H)$ then the extension of $A\otimes I_{\K}$ to the set $\D(\sqrt{G}\otimes I_{\K})$ mentioned in Lemma \[p-l-1\] is uniformly continuous on the set $$\label{s-3}
\V_{E}\doteq \{{\hspace{1pt}}\eta\in\D(\sqrt{G}\otimes I_{\K})\,|\, \|\sqrt{G}\otimes I_{\K}\eta\|^2\leq E{\hspace{1pt}}\}$$ for any $E>0$. Quantitatively, $$\label{v-CB}
\|A\otimes I_{\K}(\eta-\theta)\|\leq \varepsilon\|A\|^{G}_{4E/\varepsilon^2}=o(1)\quad \textit{as} \;\; \varepsilon\to 0^+$$ for any vectors $\,\eta$ and $\,\theta$ in $\V_{E}$ such that $\|\eta-\theta\|\leq\varepsilon$.*
*If $\,A\in\B_{\!G}(\H)\setminus\B^0_{\!G}(\H)$ then the operator $A\otimes I_{\K}$ is not continuous on the set $\,\V_{E}$ for any $E>0$.*
*Proof.* By Theorem \[comp-d\]F for any unit vector $\eta$ in $\D(\sqrt{G}\otimes I_{\K})$ we have $$\label{gen-ineq}
\|A\otimes I_{\K}\, \eta\|\leq \|A\otimes I_{\K}\|^{{G}\otimes I_{\K}}_{E_{\eta}}=\|A\|^{G}_{E_{\eta}},\quad \textrm{where}\quad E_{\eta}=\|\sqrt{G}\otimes I_{\K} \eta \|^2.$$
Assume that $\eta$ and $\theta$ are vectors in $\V_E$ such that $\|\eta-\theta\|\leq\varepsilon$. Since $\|\sqrt{G}\otimes I_{\K}\eta\|^2\leq E$ and $\|\sqrt{G}\otimes I_{\K}\theta\|^2\leq E$ we have $\|\sqrt{G}\otimes I_{\K}(\eta-\theta)\|^2\leq 4E$. So, by using (\[gen-ineq\]), the concavity of the function $E\mapsto \left[\|A\|_E^{G}\right]^2$ and Lemma \[WL\] we obtain $$\!\|A\otimes I_{\K} (\eta-\theta)\|=\|\eta-\theta\|\left\|A\otimes I_{\K}\, \frac{\eta-\theta}{\|\eta-\theta\|}\right\|\leq\|\eta-\theta\|\|A\|^{G}_{4E/\|\eta-\theta\|^2}\leq\varepsilon\|A\|^{G}_{4E/\varepsilon^2}.$$ By condition (\[s-cond\]) the r.h.s. of this inequality tends to zero as $\varepsilon\rightarrow0^+$. Thus, the function $\eta\mapsto A\otimes I_{\K}|\eta\rangle$ is uniformly continuous on $\V_E$.
The last assertion of the proposition follows from the proof of the last assertion of Corollary \[vbl-c\] below, since ${\mathrm{Tr}}_{\K}|A\otimes I_{\K}\eta\rangle\langle A\otimes I_{\K}\eta|={\mathrm{Tr}}A\rho_{\eta}A^*$, where $\rho_{\eta}={\mathrm{Tr}}_{\K}|\eta\rangle\langle\eta|$, for any vector $\eta$ in $\D(\sqrt{G}\otimes I_{\K})$. $\square$
\[vbl-c\] *For any operators $A$ and $B$ in $\B_{\!G}^0(\H)$ the function ${\hspace{1pt}}\rho\mapsto A\rho B^*$ from $\T^+_{G}$ into $\T(\H)$ (defined by formula (\[ab-d\])) is uniformly continuous on the set $\mathfrak{C}_{{G},E}\doteq \{\rho\in\T_{+}(\H)\,|\,{\mathrm{Tr}}\rho\leq 1, {\mathrm{Tr}}{G}\rho\leq E{\hspace{1pt}}\}$ for any $E>0$. Quantitatively, $$\label{ab-cb}
\|A\rho B^*- A\sigma B^*\|_1\leq \sqrt{\varepsilon}\left(\|A\|_E^G \|B\|^{G}_{4E/\varepsilon}+\|B\|_E^G\|A\|^{G}_{4E/\varepsilon}\right)=o(1)\quad \textit{as} \;\; \varepsilon\to 0^+$$ for any operators $\,\rho$ and $\sigma$ in $\,\C_{{G},E}$ such that $\|\rho-\sigma\|_1\leq\varepsilon$.*
*If $\,A\in\B_{\!G}(\H)\setminus\B^0_{\!G}(\H)$ then the function $\,\rho\mapsto A\rho A^*$ is not continuous on the set $\,\mathfrak{C}_{{G},E}$ for any $E>0$.*
\[comp-d-r++\] Corollary \[vbl-c\] shows that the operators $A$ in $\B^0_{\!G}(\H)$ are characterized by continuity of the function $\rho\mapsto A\rho A^*$ on the set $\mathfrak{C}_{{G},E}$ for any given $E>0$.
*Proof.* Let $\rho$ and $\sigma$ be operators in $\mathfrak{C}_{{G},E}$ such that $\|\rho-\sigma\|_1\leq\varepsilon$. If $\K\cong\H$ then one can find vectors $\eta$ and $\theta$ in the set $\V_{E}$ (defined in (\[s-3\])) such that $\rho={\mathrm{Tr}}_{\K} |\eta\rangle\langle\eta|$, $\sigma={\mathrm{Tr}}_{\K} |\theta\rangle\langle\theta|$ and $\|\eta-\theta\|\leq \sqrt{\varepsilon}$ [@H-SCI]. By Lemma \[p-l-1\] the operators $A\otimes I_{\K}$ and $B\otimes I_{\K}$ have unique linear $\sqrt{G}\otimes I_{\K}$-bounded extensions to the set $\,\D(\sqrt{G}\otimes I_{\K})$ satisfying (\[s-prop\]). By using the monotonicity of the trace norm, the inequality $$\||\alpha\rangle\langle\beta|-|\varphi\rangle\langle\psi|\|_1\leq \|\alpha\|\|\beta-\psi\|+\|\psi\|\|\alpha-\varphi\|,$$ where $|\alpha\rangle=A\otimes I_{\K}|\eta\rangle$, $|\beta\rangle=B\otimes I_{\K}|\eta\rangle$, $|\varphi\rangle=A\otimes I_{\K}|\theta\rangle$, $|\psi\rangle=B\otimes I_{\K}|\theta\rangle$, and continuity bound (\[v-CB\]) we obtain $$\|A\rho B^*-A\sigma B^*\|_1\leq \sqrt{\varepsilon}\|B\|^{G}_{4E/\varepsilon}\|A\otimes I_{\K}{\hspace{1pt}}\eta\|+\sqrt{\varepsilon}\|A\|^{G}_{4E/\varepsilon}\|B\otimes I_{\K}{\hspace{1pt}}\theta\|.$$ By inequality (\[gen-ineq\]) this implies (\[ab-cb\]).
The r.h.s. of (\[ab-cb\]) tends to zero as $\,\varepsilon\rightarrow 0^+$, since $A$ and $B$ satisfy condition (\[s-cond\]).
If $A\in\B_{\!G}(\H)\setminus\B^0_{\!G}(\H)$ then, by Remark \[comp-d-r\], the sequence $\|A\bar{P}_n\|^{G}_E$, where $\bar{P}_n$ is the spectral projector of ${G}$ corresponding to the interval $\,(n,+\infty)$, does not tend to zero. Hence there is a sequence $\{\rho_n\}$ of states in $\mathfrak{C}_{{G},E}$ such that the sequence $\{{\mathrm{Tr}}A\bar{P}_n\rho_n\bar{P}_n A^*\}$ does not tend to zero. Since the condition ${\mathrm{Tr}}{G}\rho_n\leq E$ implies ${\mathrm{Tr}}\bar{P}_n\rho_n\leq E/n$, the sequence $\{\bar{P}_n\rho_n\bar{P}_n\}\subset\mathfrak{C}_{{G},E}$ tends to zero. This shows discontinuity of the function $\rho\mapsto A\rho A^*$ on the set $\mathfrak{C}_{{G},E}$. $\square$ Since $\B(\H)$ is dense in $\B_{\!G}^0(\H)$ by Theorem \[comp-d\]C, Proposition \[bp-en-0\]E implies the following
\[bp-en-0+\] *Let ${G}$ be a positive densely defined operator on $\H$ satisfying condition (\[H-cond\]) and $E>0$.* *Any 2-positive linear map $\Phi:\B(\H)\rightarrow\B(\H)$ such that $\,\Phi(I_{\H})\leq I_{\H}\,$ having the predual map $\Phi_*:\T(\H)\rightarrow\T(\H)$ with finite[^21] $$\label{E-phi}
Y_{\Phi}(E)\doteq \sup\left\{{\hspace{1pt}}{\mathrm{Tr}}{G}\Phi_*(\rho)\,|\,\rho\in\mathfrak{S}(\mathcal{H}),{\mathrm{Tr}}{G}\rho\leq E\,\right\}$$ is uniquely extended to the bounded linear operator $\,\Phi_{\!G}:\B^0_{\!G}(\H)\rightarrow\B_{\!G}^0(\H)$ such that* $$\label{ch-ext}
\|\Phi_{\!G}(A)\|^{G}_E\leq\sqrt{\|\Phi(I_{\H})\|}\,\|A\|^{G}_{Y_{\Phi}(E)}\leq \sqrt{\|\Phi(I_{\H})\|K_{\Phi}}\,\|A\|^{G}_{E},$$ *where* $K_{\Phi}=\max\{1, Y_{\Phi}(E)/E\}$.
The assertion of Proposition \[bp-en-0+\] can be strengthened substantially by assuming complete positivity of $\Phi$. The corresponding result is considered in [@EPM].
Different applications of the operator *E*-norms are presented in [@SPM; @QDS; @EPM]. In [@SPM] the version of the Kretschmann-Schlingemann-Werner theorem for unbounded completely positive linear maps is obtained by using the results from Section 5.
I am grateful to A.S.Holevo, G.G.Amosov, A.V.Bulinsky and M.M.Wilde for discussion and useful remarks. I am also grateful to V.Zh.Sakbaev for consultation concerning unbounded operators and to T.V.Shulman for the help and useful discussion.
Special thanks to S.Weis for the help in proving the coincidence of definitions (\[ec-on-b\]) and (\[ec-on\]).
[99]{} D.Aharonov, A.Kitaev, N.Nisan, “Quantum circuits with mixed states”, in: Proc. 30th STOC, pp. 20-30, ACM Press, 1998; arXiv:quant-ph/9806029.
V.P.Belavkin, G.M.D’Ariano, M.Raginsky, “Operational Distance and Fidelity for Quantum Channels”, J.Math.Phys. 46 (2005) 062106; arXiv:quant-ph/0408159.
I.Bengtsson, K.Zyczkowski, “Geometry of Quantum States: An Introduction to Quantum Entanglement”, 2nd edition. Cambridge: Cambridge University Press, 2017.
O.Bratteli, D.W.Robinson, “Operators algebras and quantum statistical mechanics”, vol.I, Springer Verlag, New York-Heidelberg-Berlin, 1979.
A.S.Holevo, “Classical capacities of quantum channels with constrained inputs”, Probability Theory and Applications. V.48, N.2, 359-374 (2003); arXiv:quant-ph/0211170.
A.S.Holevo, “Quantum systems, channels, information. A mathematical introduction”, Berlin, DeGruyter, 2012.
T.Kato, “Perturbation Theory for Linear Operators”, Springer-Verlag, New York-Heidelberg-Berlin, 1980.
D.Kretschmann, D.Schlingemann, R.F.Werner, “A Continuity Theorem for Stinespring’s Dilation”, J. Funct. Anal. V.255, N.8, 1889-1904 (2008); arXiv:0710.2495.
R.Nair, “Quantum-limited loss sensing: Multiparameter estimation and Bures distance between loss channels”, arXiv:1804.02211\[quant-ph\].
V.I.Paulsen, “Completely Bounded Maps and Operator Algebras”, Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, 2002.
S.Pirandola, R.Laurenza, C.Ottaviani, L. Banchi, “Fundamental Limits of Repeaterless Quantum Communications”, Nat. Comm. 8, 15043 (2017); arXiv:1510.08863.
V.Yu.Protasov, M.E.Shirokov, “On mutually inverse transforms of functions on a half-line”, Doklady Mathematics, 100:3, 560-563 (2019) (Russian version 489:5).
M.Reed, B.Simon, “Methods of Modern Mathematical Physics. Vol I. Functional Analysis”, Academic Press Inc., 1980.
M.Reed, B.Simon, “Methods of Modern Mathematical Physics. Vol II. Fourier analysis, self-adjointness”, Academic Press Inc., 1980.
M.E.Shirokov, “Continuity bounds for information characteristics of quantum channels depending on input dimension and on input energy”, arXiv:1610.08870.
M.E.Shirokov, “Strong convergence of quantum channels: continuity of the Stinespring dilation and discontinuity of the unitary dilation.”, arXiv:1712.03219.
M.E.Shirokov, “On the Energy-Constrained Diamond Norm and Its Application in Quantum Information Theory”, Problems Inform. Transmission, V.54, N.1, 20Ц33, 2018; arXiv:1706.00361.
M.E.Shirokov, “On completion of the cone of CP linear maps with respect to the energy-constrained diamond norm”, arXiv:1810.10922.
M.E.Shirokov, A.S.Holevo, “Energy-constrained diamond norms and quantum dynamical semigroups”, arXiv:1812.07447.
M.E.Shirokov, “On extension of quantum channels and operations to the space of relatively bounded operators”, arXiv:1903.06086.
B.Simon, “Operator Theory: A Comprehensive Course in Analysis”, Part IV American Mathematical Society, 2015.
S.Simons, “Minimax and Monotonicity”, Springer, Berlin, 1998.
W.F.Stinespring, “Positive functions on $C^*$-algebras”, Proc. Amer. Math. Soc. V.6, N.2, 211-216, 1955.
A.Uhlmann, “The ”Transition Probability“ in the State Space ofa $*$-Algebra”, Rep. Math. Phys. 9 (1976) 273.
J.Watrous, “The Theory of Quantum Information”, Cambridge University Press, 2018.
S.Weis, M.Shirokov, “Extreme points of the set of quantum states with bounded energy”, arXiv:2002.03969.
M.M.Wilde, “From Classical to Quantum Shannon Theory”, arXiv:1106.1445 (v.7).
A.Winter, “Tight uniform continuity bounds for quantum entropies: conditional entropy, relative entropy distance and energy constraints”, Comm. Math. Phys., V.347, N.1, 291-313 (2016); arXiv:1507.07775.
A.Winter, “Energy-constrained diamond norm with applications to the uniform continuity of continuous variable channel capacities”, arXiv:1712.10267.
[^1]: Steklov Mathematical Institute, RAS, Moscow, email:[email protected]
[^2]: Other applications of the operator *E*-norms are presented in the recent papers [@SPM; @QDS; @EPM].
[^3]: It is also called the diamond norm in the quantum information theory [@Kit; @Wilde].
[^4]: In the previous versions of this posting the coincidence of the r.h.s. of (\[ec-on-b\]) and (\[ec-on\]) was conjectured, but it was proved only under the assumption that the operator $G$ is discrete (Definition \[D-H\]).
[^5]: In [@CSR] condition (\[H-cond\]) was not assumed. We use this assumption here, since it simplifies analysis of the norms $\|\!\cdot\!\|^{G}_E$ without reduction of generality (note that $\|A\|^{{G}+\lambda I}_E=\|A\|^{G}_{E-\lambda}$ for all $A$ and $\lambda>0$).
[^6]: In [@CSR] the norms $\|\!\cdot\!\|^{G}_E$ are called energy-constrained operator norms, since $G$ is treated therein as a Hamiltonian of a quantum system described by the Hilbert space $\H$.
[^7]: If $\|A\|^{G}_E<\|A\|$ then this assertion can be derived from properties a) and b) in Proposition \[ec-on-p1\].
[^8]: The function $E\mapsto\left[{\|\hspace{-1pt}|}A{\|\hspace{-1pt}|}^{G}_{E}\right]^p$ is not concave in general for any $p\in(0,2]$ by Remark \[nonconc\] below;
[^9]: The map $\Phi_*$ is defined by the relation ${\mathrm{Tr}}\Phi(A)\rho={\mathrm{Tr}}A\Phi_*(\rho)$ for all $A\in\B(\H)$ and $\rho\in\T(\H)$. Existence of $\Phi_*$ is equivalent to normality of the map $\Phi$, which means that $\Phi(\sup_{\lambda}A_\lambda)=\sup_{\lambda}\Phi(A_{\lambda})$ for any increasing net $A_\lambda$ of positive operators in $\B(\H)$ .
[^10]: If ${G}$ is a Hamiltonian of a quantum system described by the space $\H$ and $\Phi$ is a quantum channel (in the Heisenberg picture) then $Y_{\Phi}(E)/E$ is the energy amplification factor of $\Phi$.
[^11]: If ${G}_1$ and ${G}_2$ are Hamiltonians of quantum systems $1$ and $2$ described by the spaces $\H_1$ and $\H_2$ then ${G}_{12}$ is the Hamiltonian of the composite quantum system $12$ [@H-SCI].
[^12]: Here and in what follows we write $I_X$ instead of $I_{\H_X}\!$ (where $X=1,2,A,B,..$) to simplify notations.
[^13]: Slightly different energy-constrained $cb{\hspace{1pt}}$-norm is used in [@Pir].
[^14]: This topology is a restriction to the set $\F(A,B)$ of the strong operator topology on the set of all linear maps from $\T(\H_A)$ to $\T(\H_B)$. The strong convergence of a sequence $\{\Phi_n\}\subset\F(A,B)$ to a map $\Phi_0$ means that $\lim_{n\rightarrow\infty}\Phi_n(\rho)=\Phi_0(\rho)\,\textup{ for all }\rho\in\T(\H_A)$.
[^15]: If ${G}$ is a bounded operator then the norm $\|\!\cdot\!\|^{G}_E$ is equivalent to the operator norm by Proposition 2A.
[^16]: I would be grateful for direct references to these results.
[^17]: $\D(\sqrt{G})\otimes \K$ is the linear span of all the vectors $\varphi\otimes\psi$, where $\varphi\in\D(\sqrt{G})$ and $\psi\in\K$.
[^18]: We define the operator $A\rho B^*$ in such a way to avoid the notion of adjoint operator, since we make no assumptions about closability of the operators $A$ and $B$.
[^19]: $A\otimes I_{\K}$ denotes the operator mentioned in Lemma \[p-l-1\].
[^20]: This holds for the operators $a$ and $\sqrt{N}$ in the below example.
[^21]: By concavity of the function $\,E\mapsto Y_{\Phi}(E)\,$ its finiteness for some $E>0$ implies its finiteness for all $E>0$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Electron-doped [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} has been systematically studied by high pressure investigations of the magnetic and electrical transport properties, in order to unravel the complex interplay of superconductivity and magnetism. The compound reveals an exceedingly broad re-entrant transition to the superconducting state between $T_{\rm{c,on}} = 19.8$ K and $T_{\rm{c,0}} = 5.2$ K due to a canted A-type antiferromagnetic ordering of the Eu$^{2+}$ moments at $T_{\rm{N}} = 16.6$ K and a re-entrant spin glass transition at $T_{\rm{SG}} = 14.1$ K. At ambient pressure evidences for the coexistence of superconductivity and ferromagnetism could be observed, as well as a magnetic-field-induced enhancement of the zero-resistance temperature $T_{\rm{c,0}}$ up to $7.2$ K with small magnetic fields applied parallel to the *ab*-plane of the crystal. We attribute the field-induced-enhancement of superconductivity to the suppression of the ferromagnetic component of the Eu$^{2+}$ moments along the *c*-axis, which leads to a reduction of the orbital pair breaking effect. Application of hydrostatic pressure suppresses the superconducting state around $14$ kbar along with a linear temperature dependence of the resistivity, implying that a non-Fermi liquid region is located at the boundary of the superconducting phase. At intermediate pressure, an additional feature in the resistivity curves is identified, which can be suppressed by external magnetic fields and competes with the superconducting phase. We suggest that the effect of negative pressure by the chemical Rh substitution in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} is partially reversed, leading to a re-activation of the spin density wave.'
author:
- 'A. Löhle'
- 'A. Baumgartner'
- 'S. Zapf'
- 'W. H. Jiao'
- 'G. H. Cao'
- 'M. Dressel'
title: 'Effects of pressure and magnetic field on the re-entrant superconductor [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}'
---
Introduction
============
The discovery of high-$T_c$ superconductivity in the iron-pnictide family [@Kamihara06] has been one of the most exciting recent developments in condensed matter research. While magnetism and superconductivity are traditionally competitive phenomena in conventional superconductors [@Ginzburg56; @Saint69], there is growing experimental evidence that the unconventional superconductivity of the iron-pnictides is closely linked to magnetism as also supposed for the high-$T_c$ cuprates, which are still the most intensively investigated unconventional superconductors [@Moriya90; @Monthoux92]. Among the parent compounds of iron-based superconductors the Eu containing 122-compound [@Jeevan08; @Ren08] stands out in particular due to its magnetic sub-lattice formed by the Eu$^{2+}$ ions, carrying a local moment of $S = 7/2$ [@Jiang09; @Xiao09; @Herrero09; @Zapf16]. In this case -– additional to the spin-density-wave (SDW) ordering in the FeAs layers at approximately $190$ K -– the Eu$^{2+}$ moments align ferromagnetically along the $a$-axis and antiferromagnetically along the $c$-axis in a so called A-type antiferromagnetic pattern [@Zapf13] at $T_{\rm{N}} \approx 19$ K. Application of hydrostatic or chemical pressure [@Terashima09; @Jeevan11] as well as hole [@Jeevan082] and electron [@Jiang092] doping of EuFe$_2$As$_2$ lead to a suppression of the SDW state and an emergence of superconductivity with transition temperatures up to $40$ K. In those compounds, where the energy scales of the magnetic and superconducting state are in close proximity, the competition between the two orders become very apparent as exotic effects such as a resistivity re-entrance around $T_{\rm{N}}$ [@He10] or an enhancement of $T_c$ by the application of small external magnetic fields [@Tran12]. Up to now considerable effort was put into developing a clear picture of how superconductivity can coexist or even be induced by magnetic effects. In case of the strong Eu$^{2+}$ magnetism an additional re-entrant spin glass (SG) phase [@Zapf13] seems to be a key in understanding the coexistence, whereas in general the idea of a magnetic quantum critical point lying beneath the superconducting dome has been a long-standing hypothesis [@Shibauchi14]. In this study, detailed magneto-transport measurements under hydrostatic pressure up to $18$ kbar together with systematic dc and ac magnetization measurements were employed in order to investigate this interesting entanglement of superconductivity and magnetism in the phase diagram of an nearly optimal doped member of the rare earth EuFe$_2$As$_2$ iron pnictide family.
Experiment
==========
Single crystals of Eu(Fe$_{1-x}$Rh$_x$)$_2$As$_2$ were grown via the self-flux method [@Neubauer16; @Cao09; @Cao13]. The quality and precise chemical composition of the crystals was subsequently checked by energy-dispersive x-ray spectroscopy (EDX) on several points of the crystal. The crystal used for the pressure dependent investigations \[see Fig. \[figure1\] (b)\] showed no evidences of impurity elements in the EDX spectra and the chemical composition corresponds to a $x=7$ % rhodium doping on the iron side. For determining the transport and magnetic properties of the Rh – i.e. electron doped – sample, detailed temperature and magnetic field dependent four-point dc-resistivity measurements were carried out. The resistivity was thereby determined using electric currents up to $5$ mA -– which ensures a linear response -– from room temperature down to $1.8$ K with a relative low cooling rate of $0.3$ K/min to guarantee a proper thermalization of the sample. At ambient pressure additional ac and dc magnetization measurement were performed on a Quantum Design Magnetic Property Measurement System (MPMS) to unambiguously identify the origin of the distinct features of the dc-resistivity curve and to get further information about the underlying physics.
![(Color online) (a) Technical drawing of the clamp piston pressure cell utilized for this study. The feedthrough is equipped with 6 pairs of twisted wire pairs for the electrical measurements. (b) Microscope picture of the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} sample mounted on the kapton sample stage of the pressure cell feedthrough. The magnetic field direction as well as the rotation axis for the angle dependent magneto tansport measurements are marked in red, while the Mn wire coil for the in-situ pressure determination is indicated in white.[]{data-label="figure1"}](figure1){width="0.9\columnwidth"}
The pressure dependent electrical resistivity measurements were carried out by using a clamp piston pressure cell sketched in Figure \[figure1\] with Daphne oil 7373 [@Yokogawa07] as the pressure transmitting medium. For the in-plane resistivity measurements, four stripe-like contacts were made by silver paint on the sample, connecting it to the twisted wire pairs of the pressure cell feedthrough with 10 $\mu$m thick gold wires. To provide the opportunity of magnetoresistance measurements with applied in- and out-of-plane magnetic fields up to 65 kG, the sample was mounted on the kapton sample stage of the electrical feedthrough as shown in Figure \[figure1\] (b). In this configuration the sample can be rotated together with the whole pressure cell with respect to the horizontal magnetic field \[as indicated by the red arrow in Fig. \[figure1\] (b)\]. For determining the actual pressure inside of the cell, a Mn pressure sensor in form of a wire coil as shown in Figure \[figure1\] (b) was used. In this way the pressure could be measured in-situ and therefore the pressure loss during the cooling process could be taken into account in the further analysis of the data.
Results
=======
Electronic and magnetic properties at ambient conditions
--------------------------------------------------------
Figure \[figure2\] shows the temperature dependence of the resistivity measured along the *ab*-plane from room temperature down to $2$ K. In the high temperature region (grey shaded area) above $100$ K, the resistivity shows a $T$-linear behavior as expected for a normal metal with dominant electron-phonon scattering. At low temperatures, but still above the superconducting transition, the resistivity can be fitted with the power law expression
$$\rho(T) = \rho_0 + a \cdot T^n
\label{eq:rho}$$
leading to an exponent $n=2$ which points out Fermi-liquid behavior with appreciable electron-electron interaction. Due to the missing indication of any SDW associated anomaly in the resistivity curve, the onset of superconductivity around $19.8$ K (see Fig. \[figure3\]), together with the fact that for optimal doping a $T$-linear behavior down to the lowest temperature is expected [@Tam13; @Kurita13] we suggest the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} sample to be located in the slightly under-doped region of the phase diagram (cf. $x=9$ % rhodium doping in [@Jiao16]). The residual resistivity ratio of
$$RRR=\frac{\rho_{300K}}{\rho_{0K,fit}} = 2.2
\label{eq:RRR}$$
indicates thereby a good quality of the sample.
![(Color online) Temperature dependent dc resistivity data of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} at ambient pressure. In the high-temperature regime (grey shaded area) above $100$ K, the resistivity shows a linear behavior in $T$ (green dashed line), while the low temperature region, above the superconducting transition, can be fitted by the power law term $\rho(T)=\rho_0+c\cdot T^n$ (red dashed curve) with an exponent $n = 2$.[]{data-label="figure2"}](figure2.eps){width="0.9\columnwidth"}
Panel (a) in Figure \[figure3\] depicts the low-temperature resistivity curves of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} – measured in a cooling-heating cycle – which show a clear re-entrant superconducting transition as well as a small thermal hysteresis exhibiting a distinct three-stage shape \[see Fig. \[figure3\] panel (b)\].
![(Color online) Comparison between the low temperature dc resistivity and magnetization data of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} recorded at ambient pressure. In panel (a) the dc-resistivity with its clear re-entrant superconducting behavior is shown, while panel (b) displays the thermal hysteresis of the resistivity occurring during the cooling-heating-cycle. Panels (c)-(f) illustrate the dc and ac magnetic susceptibility. The dashed lines are guides to the eyes and mark the different transitions namely the onset of superconductivity ($T_{\rm{c,on}}= 19.8$ K, dashed black line), the antiferromagnetic ordering of the Eu$^{2+}$ moments ($T_{\rm{N}}= 16,6$ K, dashed red line), the re-entrant spin glass ($T_{\rm{SG}}= 14,1$ K, dashed green line) and the approach of zero resistance ($T_{\rm{c,0}}= 5.4$ K, dashed blue line).[]{data-label="figure3"}](figure3.eps){width="0.9\columnwidth"}
To identify the underlying physical processes of the distinct features of the resistivity curve, dc and ac magnetization measurements (panels (c)-(f) in Fig. \[figure3\]) are utilized based on Ref. . The first kink in the resistivity curve coincides with the opening of the hysteresis at $T_{\rm{c,on}} = 19.6$ K; it is caused by the onset of superconductivity, as can be seen by the splitting of the zero field cooled (ZFC) and field cooled heat (FCH) susceptibility curves \[Fig. \[figure3\] (c) and (d), black dashed line\]. The local minimum of $\rho(T)$ together with the onset of the second peak of the hysteresis curve at $T_{\rm{N}} = 16.6$ K occur simultaneously with a clear kink in the dc and ac magnetization data (Fig. \[figure3\] panel (c) and (e), red dashed line) which corresponds to a canted A-type anti-ferromagnetic ordering of the local Eu$^{2+}$ moments [@Jiang092; @Zapf11]. Lowering the temperature further leads to a second magnetic transition at $T_{\rm{SG}} = 14.1$ K into a re-entrant spin glass phase, as identified by Zapf *et al.* [@Zapf13; @Zapf11] in P substituted samples. The typical features of this phase can be preferably seen in the ac magnetization data $\chi'_{ac}$($T$) and $\chi''_{ac}$($T$) (panel (e) and (f) in Fig. \[figure3\], green dashed line) and coincide with the local maximum of the re-entrant superconducting transition in the resistivity curve. At $T_{\rm{c,0}} = 5.4$ K, the superconducting transition is completed and the sample reaches zero resistance (blue dashed line in Fig. \[figure3\]), while the hysteresis closes and the magnetization shows a distinct kink in the ZFC dc susceptibility and a maximum in $\chi''$($T$).\
In the following analysis we investigate the effect of an external magnetic field and hydrostatic pressure on the features we identified in the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} sample.
Magnetic field dependences of the low temperature phases
--------------------------------------------------------
By applying an external magnetic field we can affect the magnetic order and superconducting transition temperature of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} and thus vary the interplay between this phases. Application of in- ($H\parallel ab$) and out-of-plane ($H\parallel c$) magnetic fields give here the possibility to selectively modify the different aspects of the magnetic ordering and therefore identify their basic role for the entire system.
### Magnetic ordering
Figure \[figure4\] shows low field dc susceptibility ($\chi_{dc}=M(T)/\mu_0H$) and resistivity data recorded by applied magnetic fields parallel to the *ab*-plane of the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} single crystal.
![(Color online) (a) In-plane dc magnetization data indicate that the spin glass state of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} (indicated by the green arrows) vanishes already in the presence of minor magnetic fields of around $250$ G. Comparison of the magnetization (b) data with the evolution of the thermal hysteresis in $\rho$(T) (c) for different magnetic fields applied parallel to the *ab*-plane of the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} single crystal. The red dashed arrows are guides to the eye which indicate the antiferromagnetic transition temperatures. The opening of the big hysteresis feature (labeled with 2 in Fig.F \[figure3\]) turns out to be connected with the antiferromagnetic ordering of the Eu$^{2+}$ moments at $T_{\rm{N}}$ and can therefore identify a fast suppression of the antiferromagnetic state. The black and blue arrows indicate - exemplary for two different magnetic fields - $T_{\rm{c,on}}$ and $T_{\rm{c,0}}$ as the opening and closing of the overall hysteresis.[]{data-label="figure4"}](figure4.eps){width="0.9\columnwidth"}
![(Color online) Temperature dependent out-of-plane ($H\parallel c$) magnetization of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} measured in ZFC-FCC-FCH cycles from $2$ K to $25$ K at applied magnetic fields up to $20$ kG. The missing kink from the magnetic ordering transition and the absence of a splitting in the ZFC-FCH data reveals a suppression of $T_{\rm{N}}$ between $7.5$ kG and $10$ kG down to the lowest measured temperature of $2$ K.[]{data-label="figure5"}](figure5.eps){width="0.9\columnwidth"}
In panel (a) one can see that the feature in the magnetization data connected to the SG transition (indicated by green arrows) vanishes already in the presence of minor magnetic fields of $250$ G, while it stays at the same temperature. As the glassy freezing of the in-plane spin components in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} results from a competition of antiferromagnetic Rudeman-Kittel-Kasuya-Yosida (RKKY) interlayer coupling and the ferromagnetic intralayer spin-exchange interactions [@Jeevan11], a fast suppression of this state in magnetic fields is expected. The observed behavior therefore confirms - beside the ac susceptibility measurements - the presence of a SG phase in electron-doped members of the EuFe$_2$As$_2$ family. By comparing the resistivity hysteresis with the measured susceptibility \[Fig. \[figure4\] (b) and (c)\] as introduced before in the zero-field analysis, the hysteresis proves itself as a helpful tool to locate and follow the phase transitions under applied magnetic fields. The red dashed arrows indicate the antiferromagnetic transition temperatures determined by the sharp kink in the magnetization curve. The opening of the big hysteresis feature (labeled with 2 in Fig. \[figure3\]) confirms itself to be connected with the antiferromagnetic ordering of the Eu$^{2+}$ moments at $T_{\rm{N}}$. In line with this, one can see in Figure \[figure4\] (c), that in-plane magnetic fields lead to a fast suppression of the antiferromagnetic ordering of the Eu$^{2+}$ moments till it vanishes completely at fields around $3$ kG. At higher fields, only a third weaker feature of the hysteresis - which is most likely connected to some vortex dynamics in the superconducting state - survives \[red and orange lines in Fig. \[figure4\] (c)\]. The black and blue arrows indicate - exemplary for two different magnetic fields - $T_{\rm{c,on}}$ and $T_{\rm{c,0}}$ as the opening and closing of the overall hysteresis. It is particularly striking that the magnetic field enhances the superconducting transition temperature $T_{\rm{c,0}}$, which is untypical as superconductivity and magnetism are in general antagonistic phenomena. The origin for this anomaly and the high field evolution of the superconducting transition temperatures will be discussed in detail in the next subsection. Figure \[figure5\] shows the temperature dependent dc susceptibility $\chi_{dc}$, measured perpendicular to the *ab*-plane ($H\parallel c$). In combination with the in-plane measurements depicted in Figure \[figure4\], the characteristic shape of the A-type antiferromagnetism becomes visible. While for the in-plane measurements the above discussed bump at $T_{\rm{N}} \approx 16.6$ K indicates the magnetic transition, the out-of-plane data exhibit a rather flat magnetization below $T_{\rm{N}}$. With increasing out-of-plane field, the suppression of the magnetic transition is more gradual compared to the in-plane measurements, and for fields above $7.5$ kG the kink at $T_{\rm{N}}$ gives way to a broad shoulder typical for field-induced ferromagnetism.
### Superconductivity
The layered superconducting compound [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} shows a strong anisotropy of the magnetoresistance, as illustrated in Figure \[figure6\]. In-plane magnetic fields \[panel (a)\] lead to a shift of the magnetic transition temperature $T_{\rm{N}}$ to smaller temperatures - as indicated by the red arrow - and finally to a complete suppression of the antiferromagnetic ordering at fields around $4$ kG. In the same field range, an enhancement of the zero-temperature point $T_{\rm{c,0}}$ from $5.2$ K up to $7.2$ K (indicated by the blue arrow) can be observed. In contrast, the out-of-plane measurements \[panel (b)\] just show a reduction of the antiferromagnetic impact, as can be understood by an enhanced canting of the Eu$^{2+}$ moments out of the *ab*-plane. In this case no enhancement of $T_{\rm{c,0}}$ takes place. At higher fields (around $10$ kG) finally a suppression of the superconducting state can be observed for both field directions, as expected due to spin and orbital pair-breaking [@Khim11; @Klemm75].
![(Color online) Magnetic field dependence of the low-temperature phases of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}. In-plane resistivity data from $T=2$ K to $30$ K; when measured with applied magnetic field in- and out-of plane show a clear anisotropic behavior. The in-plane magnetic field (a) leads to a suppression of the magnetic phases (red arrow) and additionally to an enhanced superconductivity at low fields (blue arrow, inset). An applied out-of-plane field (b) directly suppresses the superconductivity (blue arrow) while the magnetic transition temperatures at low fields stays nearly constant below $4$ kG (red arrow).[]{data-label="figure6"}](figure6.eps){width="0.9\columnwidth"}
The measurements demonstrate convincingly the close interplay between the magnetic and superconducting phases. The most likely explanation for the field enhanced superconductivity with in-plane magnetic fields is a reduction of the orbital-pair-breaking effect due to a suppression of the ferromagnetic component along the *c*-axis as suggested in Ref. for Co doped EuFe$_2$As$_2$. $T_{\rm{c,on}}$ is thereby not influenced, as the magnetic transition $T_{\rm{N}}$ is at lower temperatures than the onset of superconductivity.\
To address the question whether ferromagnetism along the *c*-axis, due to some finite spin canting, coexists with superconductivity, isothermal magnetization measurements at $1.8$ K, well below $T_{\rm{N}}$ and $T_{\rm{c,0}}$, were employed. Figure \[figure7\] (a) displays the hysteresis loop for $H\parallel c$. On the first glance the initial magnetization $M_{\rm{init}}$ measured after ZFC shows no linear behavior, meaning that no lower critical field $H_{\rm{c1}}^*$ can be defined; nevertheless the magnetization due to superconducting shielding effects $M_{\rm{SC}} = \xi H/(4 \pi)$ – where $\xi$ is related to the demagnetization effect $\xi \propto 1/(1-N_d)$ – in absence of internal magnetic fields can still be defined by the incipient slope of $M_{\rm{init}}$ [@Jiao13]. Subtracting $M_{\rm{SC}}$ from the initial and loop magnetization reveals a clear additional ferromagnetic hysteresis, as shown in Figure \[figure7\] (b), indicating the coexistence of superconductivity and ferromagnetism in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}. The broad range magnetization at $1.8$ K is shown in Figure \[figure7\] (c). Here one can find saturation fields of $8$ kG and $4$ kG for $H\parallel c$ and $H \parallel ab$, respectively. The saturated magnetization $M_{\rm{sat}} = 7.5$ $\mu_B$ per formula unit is consistent with a high-field induced ferromagnetic ordering of the Eu$^{2+}$ spins with a theoretical ordered moment of $gS = 7.0$ $\mu_B$ per formula unit [@Jin15]. Moreover, one does not observe any clear magnetization jumps at the maximum (or minimum) field, excluding that pinning effects in a hard superconductor are responsible for the hysteresis loop observed at low fields.
![(Color online) Isothermal magnetization measurements on [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} at $T=1.8$ K for in- and out-of-plane applied external magnetic fields. (a) Hysteresis loop of the magnetization for $H\parallel c$ were $M_{\rm{init}}$ was measured after zero-field cooling. The green dashed line indicates the ideal initial magnetization $M_{\rm{SC}}$ in absence of internal magnetic fields. (b) Magnetization hysteresis loop after subtraction of $M_{\rm{SC}}$, indicating a ferromagnetic behavior of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} inside the superconducting phase. (c) Isothermal magnetization at high fields point out a saturation field of ca. $10$ kG for $H \parallel c$ and $4$ kG for $H\parallel ab$.[]{data-label="figure7"}](figure7.eps){width="0.9\columnwidth"}
### Magnetic-field-dependent phase diagrams of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}
Figure \[figure8\] shows the phase diagrams for magnetic fields parallel and perpendicular to the *c*-axis of the [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} crystal generated from the resistivity and magnetization data. In these, one can identify at a glance the different aspects described above and get an insight into their interconnections. One can divide the phase diagrams in two main parts, at low-fields, $H < 10$ kG, the magnetic ordering phenomena greatly influence the behavior of the system; at high-fields $H > 10$ kG, the magnetic ordering is suppressed and the magnetic moments of the Eu$^{2+}$ ions are fully aligned along the field direction. At low fields, one can see the strong anisotropic behavior of the system under applied in- and out-of-plane magnetic fields, reflecting the characteristic behavior of the canted A-type antiferromagnetic ordering, even leading to a field enhanced superconductivity for $H\parallel ab$. As pointed out before, the enhancement of $T_{\rm{c,0}}$ can be explained by a reduction of the ferromagnetic component along the *c*-axis and therefore the reduction of the orbital-pair-breaking effect. The phase diagram further supports this assumption, as the enhancement takes only place until $T_{\rm{N}}$ (red dashed line) gets completely suppressed. In the high-field region, the superconducting transition shifts downwards upon increasing magnetic fields as expected from the Ginzburg-Landau theory. The suppression is thereby slightly faster for $H\parallel c$ as known for layered systems. One can estimate by a rough linear extrapolation upper critical fields of $H_{c2}^{ab}\approx 120$ kG and $H_{c2}^c\approx 100$ kG. Compared to compounds with similar $T_c$ values containing non-magnetic Ba instead Eu, these values are strikingly lower [@Yamamoto09], which can be explained by a strong internal exchange field of the Eu sublattice of up to $300$ kG [@Nowik11].
![(Color online) Magnetic field dependent phase diagrams of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} generated by the resistivity and magnetization measurements for the external magnetic field (a) $H\parallel ab$ and $H\parallel c$. The anisotropy becomes clearly visible especially in the low field region, where in-plane fields (upper panel) lead to an enhancement of $T_{\rm{c,0}}$ in combination with a fast suppression of the magnetic ordering phenomena of the Eu$^{2+}$ moments.[]{data-label="figure8"}](figure8.eps){width="0.9\columnwidth"}
Pressure-dependent transport studies on [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}
--------------------------------------------------------------------------------
Figure \[figure9\] (a) shows an overview of all accomplished in-plane resistivity measurements on [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} up to pressures of $18$ kbar. In the three-dimensional illustration, where $\rho(T)$ is plotted over the room temperature pressure value, one can see distinct changes of the behavior at ambient as well as at low temperatures; the exact progression will be discussed in the following paragraphs.
{width="1.8\columnwidth"}
### Low temperature phases
As there is a strong interplay between the low-temperature magnetic ordering of the Eu$^{2+}$ moments and the appearance of superconductivity in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}, let us look at the peculiarities of both phenomena separately, before we combine them for understanding the overall pressure evolution. The pressure development of the superconducting state is mainly governed by two features, which get directly visible in Figure \[figure9\] (a). In the high pressure region above $14$ kbar, the superconducting transition becomes incomplete until at $16$ kbar, only a weak kink is observed, without a clear superconducting downturn. We attribute this kink to a change of electronic scattering due to the Eu$^{2+}$ order and not to superconductivity, as we will explain later in the context of the overall phase diagram.
![(Color online) Pressure dependent resistivity measurements of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} in the intermediate pressure range. (a) Normalized resistivity plotted in comparison to the ambient pressure resistivity (black curves) to illustrate the appearance of the additional feature above the superconducting phase. (b) Differences of the pressurized resistivity to the ambient pressure data.The orange arrows indicate the temperatures $T_{\rm{F}}$ where the system shows the strongest deviation from the $0$ kbar data. (c) Low temperature data before, during and after the appearance of the resistivity feature; showing a clear broadening in the presence of the feature whereas a sharp transition recovers at higher pressures restoring $T_{c,0}$.[]{data-label="figure10"}](figure10.eps){width="0.9\columnwidth"}
In the intermediate pressure region ($2.5$ kbar - $8.2$ kbar), an additional feature becomes visible in the resistivity curve $\rho(T)$, indicated by the orange arrows in Figure \[figure9\] (a). To follow the evolution of this feature, a comparison of the normalized resistivity to the ambient pressure data is depicted in Figure \[figure10\]. Starting with a small hump at around $110$ K at $2.5$ kbar, it shifts with increasing pressure down to lower temperatures; there it even leads to a strong resistivity upturn, which gets more pronounced the closer it approaches $T_{\rm{c,on}}$. This behavior has a big impact on $T_{\rm{c,0}}$ and results in a strong suppression of $T_{\rm{c,0}}$ around $5$ kbar. In contrast, $T_{\rm{c,on}}$ is very robust against pressure exposure and stays constant up to $14$ kbar, until the superconductivity gets rapidly suppressed. In order to follow the changes of the magnetic phases with hydrostatic pressure, we plot the shifted resistivity curves of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} for different values of hydrostatic pressure up to $12.5$ kbar (at room temperature) in Figure \[figure9\] (c). With increasing pressure, the resistivity dip related to the antiferromagnetic ordering of the Eu$^{2+}$ moments, shifts to higher temperatures, as observed in Ref. for isovalent substitution and electron doping. By analyzing the data with considering the pressure loss during the cooling process, we find a linear increase with a rate of $dT_{\rm{N}}/dp =0.22\pm0.01$ K/kbar. On the basis of the ambient pressure analysis above, we expect the SG transition to be located close to the local maximum of $\rho(T)$. Therefore, we conclude that the spin glass transition (at $T_{\rm{max}}$) also increases with pressure, however slightly faster than $dT_{\rm{N}}/dp$.
### Normal state properties
For the analysis of the normal state properties, the power law fitting from Equation (\[eq:rho\]) was employed for temperatures above $T_{\rm{c,on}}$, or alternatively above $T_{\rm{N}}$ after the suppression of the superconducting state. The evolution of the exponent $n$ with pressure is plotted in Figure \[figure9\] (b). Here one can see that $\rho(T)$ changes from a quadratic temperature dependency at ambient pressure to a more linear one with increasing pressure. At $p=18$ kbar, the resistivity indeed shows a linear progression in the whole measured temperature region above $T_{\rm{N}}$, as can be seen also directly by the black dashed line in Figure \[figure9\]. This indicates that we have a crossover from a Fermi-liquid, corresponding to an exponent $n=2$, at low pressures to a non-Fermi-liquid like regime ($n=1$) at high pressures. Surprisingly, it occurs at the boundary of the superconducting phase and not at maximum $T_{c}$, as observed typically in 122 iron pnictides [@Shibauchi14; @Abrahams11]. One possible explanation for the pressure induced non-Fermi-liquid behavior at the edge of the superconducting dome is given in Ref. , where they observe a similar evolution of the temperature dependency of $\rho$ by Co doping in LiFeAs. Here, low-energy spin fluctuations tuned by Fermi-surface nesting could be identified as the cause of the crossover to a non-Fermi-liquid regime. To pin down this scenario in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}, additional measurements such as nuclear magnetic resonance (NMR), are needed to identify the low-energy spin fluctuation strength under pressure.\
A second scenario is the following: in our compound, the linear temperature dependence of $\rho$ originates from a quantum critical point. However, the high Eu ordering temperature (which increases with pressure) suppresses $T_c$. Therefore, the quantum critical point is somehow masked by the Eu ordering.
### Pressure-dependent phase diagram of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}
In Figure \[figure11\] all pressure dependent measurements are summarized in a pressure-temperature-phase-diagram. The labeled pressure values are thereby corrected with regard to the pressure loss during the cooling process.
![(Color online) Pressure-dependent phase diagram of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} generated from the resistivity measurements depicted in Figure \[figure9\] (pressures are corrected to the pressure loss during the cooling process). The magnetic transition $T_{\rm{N}}$ (red squares) increases linearly with increasing pressure, while the superconducting phase (black triangles) gets completely suppressed with pressures around $14$ kbar. The local maximum of the re-entrant transition, which is probably caused by the SG state, is depicted in green and shows a bit faster increase with pressure than $T_{\rm{N}}$. The zero resistance point $T_{\rm{c,0}}$ as well as $T_{0,1}$, $T_{1}$ and $T_{10}$ (representing the temperatures were the resistivity drops below $0.1$ $\mu\Omega$cm, $1$ $\mu\Omega$cm and $10$ $\mu\Omega$cm respectively) are labeled with blue triangles, exhibit an anomaly in the intermediate pressure range connected to the appearance of the feature at higher temperatures $T_{\rm{F}}$ (orange hexagons). Additionally the temperature range were a local minimum of $\rho(T)$ appears above the superconducting transition is marked in yellow.[]{data-label="figure11"}](figure11.eps){width="0.9\columnwidth"}
In this diagram, two processes, which cause the suppression of superconductivity, become very obvious: First of all, any trace of superconductivity is suppressed at the point where the Eu ordering temperatures get in close proximity to $T_{\rm{c,on}}$. This demonstrates clearly the strong competition of superconductivity and magnetism in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}. Secondly, the high-temperature anomaly observed at intermediate pressures is strongly interacting with $T_{\rm{c,0}}$. The direct influence of the feature to the superconducting phase is not so straight forward as in the first case, instead a multifaceted picture opens up. At intermediate pressures the kink-like feature at $T_{\rm{F}}$ shifts down rapidly with pressure while $T_{\rm{c,0}}$ slightly increases up to $7$ K at $0.7$ kbar. In the pressure range from $2$ kbar to $4$ kbar (yellow shaded area), where the high temperature feature indicates its continuous presence in an upturn of $\rho(T)$ \[Fig. \[figure10\] (a), yellow arrows\], the superconducting transition gets significantly broader \[Fig. \[figure10\] (c)\], keeping the system from reaching zero resistance. Only after the influence of the high temperature feature is completely suppressed at pressures around $4$ kbar, the full superconducting transition recovers and a second dome arises which exists till superconductivity gets completely suppressed. This coupled development of $T_{\rm{F}}$ and $T_{\rm{c,0}}$ indicates a strong competition of the two phases. In order to identify its origin, additional measurements under applied external magnetic fields were employed which will be discussed in the following paragraph.
### Magnetic field dependences of pressurized [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}
Figure \[figure12\] depicts isothermal in- and out-of plane magnetoresistance measurements at $4$ K for different hydrostatic pressures up to $18$ kbar (room temperature pressure values). A clear anisotropic behavior becomes visible, as the out-of-plane field induces a purely positive slope of the resistivity curve, while the in-plane magnetic field first reduces the resistance considerably up to a critical field $H_{\rm{AFM}}$ (right panel, black arrows), where the slope changes its sign and finally a positive magnetoresistance settles at high fields. Also in the out-of-plane data, a critical field $H_{\rm{sat}}$ (left panel, black arrow) can be identified as a clear kink with $H_{\rm{sat}} >H_{\rm{AFM}}$.
![(Color online) Isothermal resistivity measurements under magnetic fields applied perpendicular (left panel) and parallel (right panel) to the *ab*-plane of [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} at different hydrostatic pressures up to $18$ kbar (room temperature pressure values). While for the out-of-plane configuration a purely positive slope can be observed, the in-plane measurements exhibit a negative slope at low fields, connected with the field induced enhancement of $T_{\rm{c,0}}$. Two critical fields manifest themselves in the data (a) as a minimum in the in-plane configuration at $H_{\rm{AFM}} \approx 4$ kG and (b) as a kink in the out-of-plane measurement at $H_{\rm{sat}} \approx 8$ kG (see the black arrows).[]{data-label="figure12"}](figure12.eps){width="0.9\columnwidth"}
The underlying physical processes that lead to the big changes in resistivity at the critical fields can be identified by comparing the critical fields with the temperature dependent magnetoresistance measurements at ambient pressure, as well as the magnetization measurements. The in plane-critical field of around $H_{\rm{AFM}} \approx 4$ kG at $3.2$ kbar coincides with the suppression of the antiferromagnetic ordering temperature, as can be seen best in the upper phase diagram of Figure \[figure8\]. Moreover, the shift of $H_{\rm{AFM}}$ to slightly higher fields with pressure ($H_{\rm{AFM}} \approx 5$kG at $18$ kbar) reflects the enhancement of $T_{\rm{N}}$ by pressure. Thus this behavior shows clearly the relationship of negative magnetoresistance and the enhancement of $T_{\rm{c,0}}$ with the magnetic ordering in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}. For the out-of-plane measurements, the critical field $H_{\rm{sat}} \approx 8$ kG corresponds to the onset of the magnetic saturation, as can be concluded from the magnetization data of Figure \[figure7\]. The large positive magnetoresistance at intermediate fields - after the suppression or saturation of all internal magnetic effects - occurs in both cases due to paramagnetic and orbital pair-breaking effects induced by the external magnetic field, as expected for any normal superconductor without magnetic ordering phenomena. At high fields, a saturation of the magnetoresistance sets in as soon as the superconducting phase is completely destroyed.\
To get an idea about the origin of the up to now unidentified resistivity feature evolving at low pressures, magnetoresistance measurements were carried out in a temperature range from $4$ K up to $60$ K at a pressure of $4.5$ kbar, where the feature is especially pronounced. Figure \[figure13\] shows the resulting resistivity curves obtained with applied magnetic fields in-plane (out-of-plane magnetic fields induce the same field dependence). As indicated by the black arrow, magnetic fields lead to a broadening and suppression of the resistivity hump. Therefore, this feature is probably of magnetic nature.\
In principle, there are two possible scenarios to explain the origin of this feature: the Kondo effect and the spin density wave. As the feature appears at low pressures already at high temperatures and shows a very similar curvature compared to the typical SDW in pressurized EuFe$_2$As$_2$ [@Kurita13] and doped BaFe$_2$As$_2$ [@Rullier-Albenque16], we suggest that external pressure re-activates the SDW in [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{}. This view is supported by the phase diagram (Fig. \[figure11\]) revealing a suppression of $T_{\rm{c,0}}$ in the same pressure range, indicating the well known competition of the two phases in the iron pnictide family. With regard to the general phase diagram of the EuFe$_2$As$_2$ parent compound, the re-activation can be caused by a competition between the effects of the applied external pressure and the electron doping which comes, in case of Rh, along with a negative internal pressure effect [@Jeevan08; @Singh09]. By taken into account that not just doping and the unit cell volume are parameters to tune the high-$T_c$ superconductivity and magnetic phases in layered iron pnictides, but also peculiarities of the crystal structure, like the tetrahedral angel play a crucial role [@Johnston10], the combination of internal and external pressure can lead to configurations not accessible by just one tuning-parameter.
![(Color online) Temperature dependent magnetoresistance measurements at $4.5$ kbar show the evolution of the pressure induced cusp in $\rho(T)$ at around $25$ K with in-plane magnetic fields. The local maximum gets broadened and shifts to higher temperatures till at fields of $40$ kG the upturn is no longer present in the resistivity curve.[]{data-label="figure13"}](figure13.eps){width="0.9\columnwidth"}
As the compound seems to be close to Fermi-surface nesting instabilities, small external pressure can counteract to the internal negative pre-pressure paving the way for the appearance of a SDW in the under-doped regime. To pin down this scenario, further investigations are needed, preferably magnetic susceptibility and neutron scattering measurements under pressure would be a powerful tool. Additionally further pressure dependent measurements on chemically substituted superconducting compounds with negative pre-pressure, such as Eu(Fe$_{1-x}$Ir$_x$)$_2$As$_2$ or other members of the Rh doped family, would be interesting to see, if significantly higher $T_c$ values can be reached in this way.
Conclusion and Outlook {#conclusion}
======================
High-pressure magneto-transport measurements up to $18$ kbar have been performed on [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} single crystals to establish a pressure-dependent phase diagram of an electron doped re-entrant superconducting member of the EuFe$_2$As$_2$ iron-pnictide family. The systematic variation of the two parameters pressure and magnetic field provides thereby the possibility to vary individual aspects of the complex system to get an insight to the interesting interplay of superconductivity and magnetism. The achieved magnetic field depended phase diagram at ambient pressure reveals the validity of the picture of a canted A-type antiferromagnetic ordering which even leads to a field-induced enhancement of $T_{\rm{c,0}}$ in the order of $2$ K. Magnetization measurements additionally confirm the existence of a spin-glass phase and uncover the coexistence of ferromagnetism and superconductivity at low temperatures. Thus, [Eu(Fe$_{0.93}$Rh$_{0.07}$)$_2$As$_2$]{} can be count to the family of “magnetic superconductors”. The thermal hysteresis in $\rho(T)$ around the superconducting transition was identified as a novel helpful tool to follow the magnetic and superconducting transition temperatures under applied magnetic fields. The pressure dependent measurements show a fast suppression of the superconducting state at a pressure around $14$ kbar where $T_{\rm{N}} \approx T_{\rm{c,on}}$, opening the view on the direct competition between the two phenomena. By contrast to other iron-based superconductors – showing non-Fermi liquid behavior above the superconducting dome connected with a hidden quantum critical point – we observe a $T$-linear development of the resistivity at the edge of the superconducting dome. Furthermore, an additional magnetic field dependent feature develops in the intermediate pressure range. We identify it as a SDW reactivated by pressure, which counteracts the negative internal pressure induced by chemical substitution. This indicates once more that not only doping influences the phase diagram; instead, small variations of the lattice by internal and external pressure plays an important role. Moreover, we open with our experiments a novel path to manipulate and eventually enhance $T_c$ in compounds under negative internal pressure by additional external pressure.
We thank C. Kamella for performing the EXD measurements at the MPI in Stuttgart, G. Untereiner for preparing the crystals and D. V. Efremov for fruitful discussions . We acknowledge support by the Deutsche Forschungsgemeinschaft (DFG) SPP 1458.
[99]{}
Y. Kamihara, H. Hiramatsu, M. Hirano, R. Kawamura, H. Yanagi, T. Kamiya and H. Hosono, J. Am. Chem. Soc. 128(31), pp 10012 (2006).
V. L. Ginzburg, Sov. Phys. JETP Vol., pp 153-160 (1956)
D. Saint-James, G. Sarma and E. J. Thomas, *Type II Superconductivity*, Elsevier Science & Technology (1969)
T. Moriya, Y. Takahashi and K. Ueda, J. Phys. Soc. Jpn. [**59**]{}, 2905 (1990)
P. Monthoux, A. V. Balatsky and D. Pines, Phys. Rev. B [**46**]{}, 14803 (1992)
H. S. Jeevan, Z. Hossain, D. Kasinathan, H. Rosner, C. Geibel and P. Gegenwart, Phys. Rev. B [**78**]{}, 052502 (2008)
Z. Ren, Z. Zhu, S. Jiang, X. Xu, Q. Tao, C. Wang, C. Feng, G. H. Cao and Z. A. Xu, Phys. Rev. B [**78**]{}, 052501 (2008)
S. Jiang, Y. Luo, Z. Ren, Z. Zhu, C. Wang, X. Xu, Q. Tao, G. Cao and Z. A. Xu, New J. Phys. [**11**]{}, 025007 (2009)
Y. Xiao, Y. Su, M. Meven, R. Mittal, C. M. N. Kumar, T. Chatterji, S. Price, J. Persson, N. Kumar, S. K. Dhar, A. Thamizhavel, and Th. Brueckel, Phys. Rev. B [**80**]{}, 174424 (2009)
J. Herrero-Martín, V. Scagnoli, C. Mazzoli, Y. Su, R. Mittal, Y. Xiao, T. Brückel, N. Kumar, S. K. Dhar, A. Thamizhavel, and L. Paolasini, Phys. Rev. B. [**80**]{}, 134411 (2009)
S. Zapf and M. Dressel, Rep. Prog. Phys. [**80**]{}, 016501 (2016)
S. Zapf, H. S. Jeevan, T. Ivec, F. Pfister, F. Klingert, S. Jiang, D. Wu, P. Gegenwart, R. K. Kremer and M. Dressel, Phys. Rev. Lett. [**110**]{}, 237002 (2013)
T. Terashima, M. Kimata, H. Satsukawa, A. Harada, K. Hazama, S. Uji, H. S. Suzuki, T. Matsumoto and K. Murata, J. Phys. Soc. [**78**]{}, 083701 (2009)
H. S. Jeevan, D. Kasinathan, H. Rosner and P. Gegenwart, Phys. Rev. B [**83**]{}, 054511 (2011)
H. S. Jeevan, Z. Hossain, D. Kasinathan, H. Rosner, C. Geibel and P. Gegenwart, Phys. Rev. B [**78**]{}, 092406 (2008)
S. Jiang, H. Xing, G. Xuan, Z. Ren, C. Wang, Z. A. Xu and G. Cao, Phys. Rev. B [**80**]{}, 184514 (2009)
Y. He, T. Wu, G. Wu, Q. Y. Zheng, Y. Z. Liu, H. Chen, J. J. Ying, R. H. Liu, X. F. Wang and Y. L. Xie, J. Phys.: Condens. Matt. [**22**]{}, 235701 (2010)
V. H. Tran, A. Zaleski, Z. Bukowski, L. M. Tran and A. J. Zaleski, Phys. Rev. B [**85**]{}, 052502 (2012)
T. Shibauchi, A. Carrington and Y. Matsuda, Annu. Rev. Condens. Matter **5**, 113 (2014)
D. Neubauer, A. V. Pronin, S. Zapf, J. Merz, H. S. Jeevan, W. H. Jiao, P. Gegenwart, G. H. Cao and M. Dressel, Phys. Stat. Sol. B [**254**]{}, 1600148 (2017)
S. Jiang, H. Xing, G. Xuan, Z. Ren, C. Wang, Z. Xu and G. H. Cao, Phys. Rev. B [**80**]{}, 184514 (2016)
W. H. Jioa, H. F. Zhai, J. K. Bao, Y. K. Lua, Q. Tao, C. M. Feng, Z. A. Xu and G. H. Cao, New J. Phys. [**15**]{}, 113002 (2013)
K. Yokogawa, K. Murata, H. Yoshino and S. Aoyama, Jap. J. Appl. Phys. [**46**]{}, 3636-3639 (2007)
G. N. Tam, B. D. Faeth, J. S. Kim and G. R. Stewart, Phys. Rev. B [**88**]{}, 134503 (2013)
N. Kurita, M. Kimata, K. Kodama, A. Harada, M. Tomita, H. S. Suzuki, T. Matsumoto, K. Murata, S. Uji and T. Terashima, Phys. Rev. B [**88**]{}, 224510 (2013)
W. H. Jiao, Y. Liu, Z. T. Tang, Y. K. Li, X. F. Xu, Z. Ren, Z. A. Xu and G. H. Cao, Supercond. Sci. Technol. [**30**]{}, 025012 (2017)
A. Baumgartner, D. Neubauer, S. Zapf, W. H. Jiao, G. H. Cao and M. Dressel, arXiv 1702.04104 (2017)\
(The same Rh doped sample – which was used for the magnetization and optical measurements in this study – was used for the pressure dependent measurements in the current paper.)
S. Zapf, D. Wu, L. Bogani, H. S. Jeevan, P. Gegenwart and M. Dressel, Phys. Rev. B [**84**]{}, 140503 (2011)
S. Khim, B. Lee, J. W. Kim, E. S. Choi, G. R. Stewart and K. H. Kim, Phys. Rev. B [**84**]{}, 104502 (2011)
R. A. Klemm, A. Luther and M. R. Beasley, Phys. Rev. B [**12**]{}, 877 (1975)
W. H. Jiao, H. F. Zhai, J. K. Bao, Y. K. Luo, Q. Tao, C. M. Feng, Z. A. Xu and G. H. Cao, New J. Phys. [**15**]{}, 113002 (2013)
W. T. Jin, W. Li, Y. Su, S. Nandi, Y. Xiao, W. H. Jiao, M. Meven, A. P. Sazonov, E. Feng, Y. Chen, C. S. Ting, G. H. Cao and Th. Bruckel, Phys. Rev. B [**91**]{}, 064506 (2015)
A. Yamamoto, J. Jaroszynski, C. Tarantini, L. Balicas, J. Jiang, A. Gurevich, D. C. Larbalestier, R. Jin, A. S. Sefat, M. A. McGuire, B. C. Sales, D. K. Christen and C. Mandrus, Appl. Phys. Lett. [**94**]{}, 062511 (2009)
I. Nowik , I. Felner, Z. Ren, G. H. Cao and Z. A. Xu, J.Phys.: Condens. Matter [**23**]{}, 065701 (2011)
E. Abrahams and Q. Si, J. Phys.: Condens. Matter [**23**]{}, 223201 (2011)
Y. M. Dai, H. Miao, L. Y. Xing, X. C. Wang, P. S. Wang, H. Xiao, T. Qian, P. Richard, X. G. Qiu, W. Yu, C. Q. Jin, Z. Wang, P. D. Johnson, C. C. Homes and H. Ding, Phys. Rev. X [**5**]{}, 031035 (2015)
F. Rullier-Albenque, C. R. Physique [**17**]{}, 164 (2016)
Y. Singh, Y. Lee, B. N. Harmon and D. C. Johnston, Phys. Rev. B [**79**]{}, 220401(R) (2009)
D. C. Johnston, Advances in Physics [**59**]{}, 803-1061 (2010)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A new neutron lifetime experiment employing frozen Fomblin has produced a result in significant disagreement with previous experiments that used liquid Fomblin near room temperature. This new experiment is subject to very few corrections, so the source of the discrepancy remains to be identified . Here we theoretically investigate several possible systematic effects for near-room-temperature storage experiments. By considering the combined effect of gravity with the usual ultracold neutron losses together with surface wave scattering loss and ultracold spectral evolution, a correction to a previous neutron lifetime experiment is found to be insignificantly small. The source of the systematic variation between experiments remains to be identified.'
author:
- 'Steve K. Lamoreaux'
date: 'Dec. 4, 2006'
title: Possible Systematic Effects in Fomblin Coated Storage Cell Neutron Lifetime Measurements
---
Introduction
============
A recent determination of the neutron lifetime by Serebrov et al. have yielded a value of $878.5\pm 0.7_{\rm stat}\pm0.3_{\rm sys}$ s [@sere], which is in significant disagreement with the PDG recommended value of $885.7\pm0.8$ s [@pdg]. The purpose of this note is to report the results of a theoretical study of the effects of surface wave scattering on the gravitational corrections to one of the moderate-accuracy experiments [@mampe] that is included in the PDG determination. These results might also have implications for the most accurate experiment because the effects described here have not been included in the modelling of the experimental system [@aruz].
The use of liquid Fomblin, a perfluorinated ether with very low nuclear absorption, to produce a nearly perfect UCN storage bottle was suggested and pioneered by C. Bates [@bates]. The highest accuracy storage experiments have employed Fomblin coated storage bottles. In order to get rid of the thermal scattering, which is a dominant loss near room temperature, [@sere] employed frozen Fomblin and achieved a loss per wall reflection of less than $5\times 10^{-6}$, a record. However, this is an order of magnitude larger than expected and is a source of concern.
Neutron lifetime experiments employing the storage of ultracold neutrons (UCN) are generally based measuring the total storage lifetime, which includes both beta decay and wall losses, as a function of volume (or mean free path between collisions) and extrapolating to infinite volume. In principle, this seems like a good idea; in practice, however, there is a distribution of neutron velocities and each velocity group has a different lifetime. To get around this problem, in [@mampe], the volume and storage times were adjusted so that the same number of wall collisions occurred for each volume. Because the times were different, the neutron lifetime could in principle be directly determined, independent of the specific loss mechanism. However, the effects of gravity spoil the simple mean free path scaling because the wall collisions are not identical on all the internal storage cell surfaces; given that the neutron potential is 1.04 neV/cm in the Earth’s gravitational field, the change in UCN energy between the top and bottom of a storage cell is about 31 neV, a significant fraction of the wall potential, 106.5 neV for the liquid form of Fomblin used in [@mampe]. Recent work has shown that surface capillary waves are a significant source of UCN loss due to inelastic upscattering beyond the wall potential[@lamgol]. The scattering will result in the change of the velocity spectrum of stored UCN. The analysis for the gravity correction that was performed in [@mampe] did not include these effects, for which the gravity correction possibly scales differently from the usual loss function.
Only the minimal published details of the experiment [@mampe] are available. The lack of details and inconsistencies between the two publications [@mampe; @nim] is astounding, and rather than adjust the model to exactly reproduce the results reported in those publications, a different approach was taken. A model analysis was performed with and without the surface wave contribution, and the differences between the extrapolated lifetimes for the two cases were used to determine new correction terms to be added to those already included in [@mampe].
In this calculation, the UCN energy will be given in cm, representing the change in energy with height of a UCN in the Earth’s gravitational field. The zero of energy will be take at the bottom of the storage cell. Also, the neutron lifetime will be taken as infinity, so the correction to [@mampe] will be represented by any residual non-zero loss rate after the scaling reduction is applied to the numerical results.
In [@mampe], the UCN storage cell was a box $H=30$ cm high, $W=40$ cm wide, with a variable length $L$ that was used to change the mean free path: $$\lambda={2\over H^{-1}+W^{-1}+L^{-1}}.$$ The UCN energy had a minimum value set by an Al window (53 cm potential) on the 6.4 cm diameter UCN guide, with a 20 cm rise to the storage cell. The UCN enter the bottle at height about 10 cm from the bottom of the bottle, so the minimum UCN energy is taken as 43 cm. There in no window on the separate guide that leads to the detector, so the minimum energy UCN detected will by 10 cm. The results were insensitive to the assumed UCN energy spectrum, so a simple initial $$n(E,0)=\sqrt{E} \ \ {\rm for}\ E>43\ {\rm cm}$$ form for the initial density spectrum delivered to the bottle entrance was assumed. However, in [@nim] it is stated that raising the storage cell by 20 cm resulted in a factor of three change in the number of stored UCN, suggesting a more significant high energy spectrum. It is likely that the large change in the number of stored UCN with height is the result of an experimental artifact that is unknown. However, assuming a spectrum of form $n(E)=E^2$ did not significantly change the results of the calculations presented here.
An electrical analog model of the experimental system shown in Fig. 1 will help in the following discussion. The storage cell is represented by a capacitor that is charged (filled with UCN) by closing the switch located after $R_{in}$, and connects the capacitor to a current source. A sliding door serves as the switch (valve), and in order for the scaling to work, the opening area is scaled (e.g., the valve is partially opened) with the storage volume so that the ratio of the impedance presented by the filling hole and guide ($R_{in}$) to the impedance representing the storage cell loss ($R_B$) is constant. In the absence of gravity, $$R_{in}(E)={4\over v(E) A_{in}}\ \ \ \ R_B(E)={4\over \mu_{tot}(E) v(E) A_B}$$ where $v(E)=\sqrt{2gE}$, ($g=980 {\rm cm/s^2}$) is the UCN velocity for energy $E$, $A_{in}$ is the input valve opening area, $\langle \mu_{tot}(E)\rangle$ is the appropriately gravity average total UCN of energy E loss probability per reflection, and $A_B$ is the bottle area ($V_B$ is defined here as the bottle volume). It is stated in [@mampe] that the fill hole opening area was varied to keep the loading of the source constant. Unfortunately, how this adjustment was performed is not given; one might assume that the sum fill hole and bottle impedances was kept constant; this requires that the fill hole be [*reduced*]{} for larger volumes. Although this keeps the spectrum constant within the guide, the effect on the neutron spectrum in the bottle itself is enormous. There is a further contradiction in [@nim] where it is stated that the filling time was scaled with $\lambda$. Changing the filling time cannot possibly work, so it is assumed here the filling area was varied. This effect is referred to as “loading the source" but it seems more profitable to think of the effect in terms of the relative impedances represented by the fill hole conductivity and the storage cell losses. It should be noted, however, that there is a gravitational correction to $R_B$ that spoils the simple scaling. It would make more sense that the fill hole to the bottle area, $A_{in}/A_B$, and that is assumed in the calculations presented here.
In operation, the storage cell was filled for five or so net storage cell lifetimes, assuring equilibrium, after which the fill switch was opened (fill valve closed). After a the storage period, the switch to the detector was closed (valve opened) and the UCN counted. It is stated in [@mampe] that the guide lifetime was 15 s, but there is no statement of varying the opening area of the valve to the detector with storage cell volume. It is stated in [@mampe] that the detector valve opening time was varied so that the number of collisions was constant over the experiment cycle. Again, details of this time shift, possibly up to 30 seconds or so, are not described in any of the publications. The numerical modelling was unstable regarding the form of the UCN detection, and shifts in apparent lifetime of up to 40 seconds were possible. To get around this problem in the model presented here, the number of UCN left in the bottle at the end of a fixed storage period was determined, and as stated earlier, the change in number with and without the surface waves was used to determine a correction to the correction presented in [@mampe] so the results presented here are independent of these complicated and unknown details, and are expected to be a reasonably small percentage effect compared to what would be determined with a full model.
The average loss function is determined by the formalism presented in [@ucn; @rp]. First, the usual UCN loss due to absorption on the walls is given by $$\mu=2f\left[{V\over E}\arcsin\left({E\over V}\right)^{1/2}-\left({V\over E}-1\right)^{1/2}\right]$$ where $\mu$ is the loss probability due to nuclear absorption and molecular upscattering per surface collision, $f=8.3\times 10^{-6}$ is the ratio of the imaginary and real parts of the surface potential, and $V$ is the wall potential, approximately 104 cm for Fomblin, at 283 K [@lamgol]. The effective volume of the bottle, in the presence of gravity, is given by $$\gamma(E)=\int_0^{\min(E,H)} \sqrt{{E-h\over E}}\ \ A(h) dh$$ where $A$ is the cross sectional area of the bottle at height $h$, constant and $A(h)=LW$ for a rectangular bottle, and we do not assume all neutrons in the storage cell can reach the roof (to allow for spectral evolution during storage). The gravity averaged loss rate for UCN of energy E is then $$\Gamma_w={1\over 4}{v(E)\over \gamma(E)}\int_0^{\min(E,H)} {E-h\over E}\ \ \mu(E-h)\ S(h)\ dh$$ and $$S(h)=LW(\delta(h)+\delta(h-H))+2W+2L$$ where $\delta(x)$ is the Dirac delta function.
The surface wave upscattering/downscattering can be treated by a similar formalism. From [@lamgol], it can be seen that the probability to scatter a single UCN with a given initial energy to a final energy bin of width $\delta E_f$ is reasonably well described by $$P(E_i\rightarrow E_f)\delta E_f=E_i(\alpha_1e^{-\beta_1 |E_i-E_f|)}+\alpha_2 e^{-\beta_2\sqrt|E_i-E_f|})\delta E_f$$ where $\alpha_1=1\times 10^{-8}/$cm, $\beta_1=.065\ {\rm cm^{-1}} $, $\alpha_2=1\times 10^{-5}/$cm, and $\beta_2=2\ {\rm cm^{-1/2}}$ for Fomblin at 283 K. Noting that the energy differences and bin widths are independent of $h$, the rate of change in spectrum can be calculated at $h=0$ using the same formalism as used with wall absorption to account for gravity, the gravity averaged rate of scattering from $E_i$ to $E_f\delta E_f$ is $$\Gamma_{sw}(E_i\rightarrow E_f)\delta E_f= {1\over 4}\ {P(E_i\rightarrow E_f)\delta E_f\over E_i}\ {v(E)\over \gamma(E)}\ \int_0^{\min(E,H)} {E_i-h\over E_i}\ S(h) \times E_i \ dh.$$ The spectral evolution from surface waves is then $$\begin{aligned}
{dn(E,t)\over dt}&=&\int_0^V {\gamma(E_i)\over \gamma(E)}n(E_i,t)\Gamma_{sw}(E_i\rightarrow E)dE_i\\ &-&n(E,t)\left[\int_0^V \Gamma_{sw}(E\rightarrow E_f)+\int_V^\infty \Gamma_{sw}(E\rightarrow E_f)dE_f\right]\\ &=&{dn_{in}(E)\over dt}-n(E,t)\left[\Gamma_{sw,out}-\Gamma_{sw,loss}\right]\\\end{aligned}$$ where the first term on the right hand side represents UCN scattered into final energy $E$ thus contributing to $n(E,t)$, the second term UCN scattered out of energy $E$, and the third term represents UCN scattered from $E$ to energies higher than $V$ and are thus lost from the system. These equations conserve the total number of UCN in the system (total remaining stored plus those lost).
The equations were numerically integrated to determine the UCN spectrum and number density as a function of time. The initial spectrum was taken as $$n(E,0)=\sqrt{E} {\Gamma_{in}(E)\over \Gamma_{in}(E)+\Gamma_{sw,loss}(E)+\Gamma_w(E)}$$ where $$\Gamma_{in}(E)={1\over R_{in}(E) V_B}$$ (the spectrum was not assumed to evolve due to the in,out scattering during filling). The spectrum was divided into bin 1 cm wide, and integrated in time over 1 s intervals: $$n(E,t+\delta t)=n(E,t)+{dn_{in}(E,t)\over dt}\delta t -n(E,t)(\Gamma_w+\Gamma_{sw,out}+\Gamma_{sw,loss})\delta t$$ and the total number of stored UCN that can be detected as a function of time is $$N(t)=\int_{10}^V \gamma(E) n(E,t) dE.$$
Results for the spectral evolution for initially monochromatic UCN of 70 cm energy, with and without gravity, are shown in Fig. 2. Results of scaled extrapolation of the neutron lifetime however indicated an effect of less than 0.5 sec on the change in the neutron lifetime. Figure 3 shows typical data that was used in this extrapolation. Numerous tests for effects of initial spectrum, bottle emptying procedure, and source loading, were performed. It might be surprising that such a large effect on the UCN spectrum does not lead to a large correction, but insofar as the gravity correction follows the scaling law as used in [@mampe], the effect should not contribute. Indeed, shown in Fig. 4 are the results for storage at time/mfp scaled calculations. It can be seen that the effects of gravity are negligible, in that there is a less than 1% difference in the spectra, and the majority of the spectrum remains with $\pm 10$ cm of the initial energy of 70 cm. Reasonable agreement with the corrections given in [@mampe] was obtained. In addition, the effects of the surface wave was determined by adjusting $f$ for the cases with and without surface wave to give approximately the same loss rate.
With that said, it was not possible to fully model the experiment described in [@mampe] with the available information. This study would be better done in regard to the experiment described in [@aruz], which appears as more complicated, so the simple electrical models used here are not applicable. However, there are enough experimental details that exist in Ph.D. dissertations and published papers to allow construction of a reasonable model of the apparatus described in [@aruz].
[99]{}
A. Serebrov et al., Phys. Lett. B [**605**]{}, 72 (2005). W.-M. Yao et al., J. Phys. G [**33**]{}, 1 (2006). W. Mampe et al., Phys. Rev. Lett. [**63**]{}, 593 (1989). S. Arzumanov et al, Phys Lett. B [**483**]{}, 15 (2000). J.C. Bates, Nucl. Instr. and Meth. A [**249**]{}, 261 (1986). S.K. Lamoreaux and R. Golub, Phys. Rev. C [**66**]{}. 044309 (2002). W. Mampe et al., Nucl. Instr. and Meth A [**284**]{}, 111 (1989). R. Golub, D.J. Richardson, and S.K. Lamoreaux, [*Ultracold Neutrons*]{} (Adam-Hilger, Bristol, 1991). J.M. Pendlebury and D.J. Richardson, Nucl. Inst. and Meth. A [**337**]{}, 504 (1994).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the effects of energy injection into the intra-group medium of the group of galaxies associated with NGC 741. The X-ray emission reveals a large bubble, which in the absence of a currently bright central radio source, may have been inflated by a previous cycle of nuclear activity . If the bubble is filled with a light, relativistic fluid we calculate that in expanding, it has done more than sufficient work to counteract the energy lost from the intra-group medium via radiative cooling; the bubble can provide this energy as it expands and rises. Using upper limits on the flux density of the plasma filling the bubble at 330 MHz and 1.4 GHz, we derive constraints on its electron energy distribution and magnetic field strength. We show that the data require the high-energy cut-off of the electron spectrum to be very low compared to the cut-offs seen in more typical radio sources if the fluid filling the bubble is a conventional relativistic plasma. This suggests that the fluid filling the bubble may not have evolved by expansion or synchrotron losses consistent with a dead radio source, leaving a puzzle as to what the origin of the bubble may be.'
author:
- |
Nazirah N. Jetha$^{1,2}$[^1], Martin J. Hardcastle$^3$, Arif Babul$^4$, Ewan O’Sullivan$^{5}$, Trevor J. Ponman$^{2}$, Somak Raychaudhury$^{2}$, Jan Vrtilek$^{5}$\
$^{1}$Laboratoire AIM, CEA/DSM - CNRS - Université Paris Diderot, DAPNIA/Service d’Astrophysique, Bât. 709, CEA-Saclay, F-91191 Gif-sur-Yvette Cédex, France\
$^{2}$School of Physics and Astronomy, University of Birmingham, Edgbaston, Birmingham B15 2TT\
$^{3}$School of Physics, Astronomy and Mathematics, University of Hertfordshire, College Lane, Hatfield, Hertfordshire AL10 9AB\
$^{4}$Department of Physics and Astronomy, University of Victoria, Victoria, BC V8P 5C2, Canada\
$^{5}$Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138
bibliography:
- 'MN-07-1395-MJ\_R2.bib'
title: The nature of the ghost cavity in the NGC 741 group
---
\[firstpage\]
X-rays:galaxies:clusters - intergalactic medium - galaxies:active - galaxies:clusters:individual(NGC 741)
Introduction
============
It is well known that there is a lack of very cool gas in the centres of galaxy groups and clusters (e.g. and ), even in systems where the cooling time is significantly shorter than the Hubble time. Why this should be so remains a mystery, and suggests that there must be some mechanism to prevent catastrophic cooling.
If a bubble inflated by an active galactic nucleus (AGN) is over-pressured with respect to the inter-galactic medium (IGM), then it will expand, doing work on the IGM [see for example @2006astro.ph..2566N]. Depending on how significantly the bubble is overpressured, the expansion may be trans or super-sonic, driving shocks into the IGM (e.g. in Centaurus A @2003ApJ...592..129K), or the expansion may be more subtle, doing gentle $PdV$ work on the gas as the bubble expands (e.g. @2001ASPC..240..363B and @2002AAS...201.0304R). Once the bubble has expanded and reached pressure equilibrium with the external medium, it will begin to rise buoyantly. As it rises, it may distort [e.g. @2001ApJ...554..261C]. The bubble will also do further work on the IGM whilst it rises. Such bubbles could remain intact until long after the AGN has switched off, with the spectrum of the bubble plasma steepening so that the emission at higher radio frequencies drops to below detectable rates.
Observationally there is much evidence for this picture. A number of clusters show radio signatures of jets terminating in lobes that, in turn, are coincident, in projection with X-ray surface brightness depressions that, on occasion are bounded by weak shocks (c.f. [e.g. @2004ApJ...607..800B] and references therein). The simplest explanation is that the inflating radio bubbles are displacing the X-ray emitting plasma.
There are, however, cavities with no detectable radio emission (see for example Abell 1795, @2002MNRAS.331..635E; Abell 2597, @2005ApJ...625..748C; HCG 62, @2004ApJ...607..800B). These so-called ‘ghost cavities’ are typically assumed to be features created in the more distant past and whose radio emission has faded over time. This interpretation finds some support in the fact that in any given cluster, the ghost cavities are typically found further away from the cluster centre than the cavities with detectable radio emission (c.f. @2004ApJ...607..800B).
Whether or not a cavity is associated with radio emission, of course depends critically on the nature of the fluid that fills the bubbles. If the bubbles are inflated by AGN, then it would be likely that the fluid is a relativistic plasma. However, the lack of associated radio emission in ghost cavity systems implies that the plasma must be in such a state that any emission is below detectable levels. The lack of observed radio emission places strong constraints on the nature of the radiating particles and the field.
In this paper we investigate a centrally located ghost cavity in the galaxy group associated with NGC 741. This is an optically identified group [@1998ApJ...496...39Z], consisting of approximately 48 galaxies within 1 Mpc, with a mean redshift of 0.019 and a velocity dispersion $\sigma =
458\pm 66~\mathrm{km\ s^{-1}}$. The BGG, identified as NGC 741, and its companion galaxy NGC 742 are separated by 45 arcsec (17 kpc), with NGC 742 lying almost exactly to the E of NGC 741, together with a much smaller third companion approximately 9 arcsec (3.4 kpc) to the NE of NGC 741 . @1998ApJ...496...73M first reported a significant extended X-ray halo in this group, extending to 19.2 arcmin from the centre of NGC 741 (approximately 440 kpc in our adopted cosmology), with a luminosity of $5.3\times 10^{41}~\mathrm{erg\ s^{-1}}$ (again in our adopted cosmology).
The organization of the rest of the paper is as follows. We present the X-ray data and our analysis in Section \[xray\]. In Section \[radio\] we present the radio data and also comment on the dynamical state of the system. In Section \[energetics\] we discuss the energetics involved and in Section \[physcond\] we place limits on the nature of the fluid filling the cavity. Our conclusions are given in Section \[conclusions\]. Throughout this paper, we use $H_0=70~\mathrm{km\ s^{-1}\ Mpc^{-1}}$, with a [$\Lambda$CDM ]{}cosmology. NGC 741 is at a redshift of 0.019, and 1 arcsec=0.38 kpc. All errors are quoted at the 1-$\sigma$ level.
X-ray data {#xray}
==========
The [*Chandra* ]{}data of NGC 741 (ObsID 2223) were obtained from the archive, and processed as part of the study of @2007MNRAS.376..193J; the data were reprocessed using [*CIAO* ]{} 3.4 and [caldb]{} 3.3.0 following the [*CIAO* ]{}threads online to apply new calibration products and a new event file was created. The data were cleaned using a 3$\sigma$ clipping algorithm to remove times of high background, leaving a 30.3 ks exposure.
In addition, we obtained and reduced the [*XMM-Newton* ]{}data (sequence number 0153030701) for the source (originally published in @2005ApJ...622..187M), using SAS version 7.0.0. The original exposure times were 8.7 ks (MOS) and 7.0 ks (PN). The data were filtered for flares following the method described in , and vignetting corrections were applied to the data following the procedure described in . After cleaning, we were left with good time intervals of 5.8 ks (MOS1), 5.9 ks (MOS2) and 3.8 ks (PN).
Spatial analysis {#spatial}
----------------
We smoothed the exposure corrected and binned [*Chandra* ]{}data in the 0.5–5.0 keV energy range using the task [csmooth]{} with a minimum smoothing signal to noise ratio of $3~\sigma$ (see Fig. \[image1\] left panel). This process revealed an elliptical cavity to the west of the host galaxy (see also Vrtilek et al (in prep)). Increasing the minimum smoothing significance to 4 $\sigma$ resulted in a somewhat smaller cavity. We use an elliptical region to approximate the boundaries of the cavity seen with 3 $\sigma$ smoothing; this region has semi-major axis, $a_b\sim$19 kpc and semi-minor axis, $b_b
\sim$14.5 kpc, (marked by the elliptical region in Fig \[image1\]). We assume a prolate ellipsoid so that the third axis $c_b=b_b$.
To check that this cavity is real and not just an artefact of adaptive smoothing we firstly created an exposure corrected image in the 0.5–5.0 keV energy range as above and smoothed it with a fixed Gaussian kernel of FWHM 5 arcsec. We then used the [*Sherpa*]{} fitting program to fit a single 2-D $\beta$-model to the [*Chandra* ]{}data to model the large-scale emission. This single-component fit is intended to model the large-scale group emission, allowing it to be subtracted from the image and revealing structures in the core. As a single-component model can only approximate the true surface brightness distribution, we do not use this model to derive physical parameters. The model centre is slightly off-centre due to the structure in the centre of the group. We then created an image of the residuals of the fit, which is shown in Fig \[image1\] (right panel). The cavity can be clearly seen, and we overlay the same ellipse as in the left panel to guide the eye. We further created a surface brightness profile in the 0.5–5.0 keV energy range which was centred on NGC 741, and covered, in azimuthal bins, the region occupied by the cavity, together with a comparison surface brightness profile in the same energy range for the entire system excluding the region we used to extract a profile of the cavity. The opening angles of the wedges used for the two profiles were 80 and 280 degrees respectively, and a local background was used for background subtraction. The resulting profiles are shown in Fig. \[sbprofC\] (left panel) overlaid on each other. There is a clear 3 $\sigma$ surface brightness drop which begins at approximately 10 kpc from NGC 741 and continues to approximately 40 kpc, implying that that the cavity seen in the images is real. At the centre of the cavity, the ratio, $\mathbf{R_{\mathrm{SB}}}$ of the surface brightness of the cavity, to that of the undisturbed gas, is approximately $\left(0.4\pm0.1\right)$. Compared to other systems, where the surface brightness contrasts are of the order of 20–30% of the surrounding surface brightness (@2004ApJ...607..800B and @2006ApJ...652..216R), the bubbles here are of a similar surface brightness contrast.
For completeness, we generated a surface brightness profile in the 0.3–2.0 keV energy range from the [*XMM-Newton* ]{}data (shown in Fig \[sbprofC\] right panel). We use this profile to provide a model for the surface brightness for use in Section \[explainmorph\]. We used the [*XMM-Newton* ]{}rather than the [*Chandra* ]{}data for this since the [*XMM-Newton* ]{}data extends out to larger radii, providing a better model for the large scale emission. The profile was fitted by a $\beta$ -model with $\beta = 0.428\pm0.005$ and core radius, ($r_c = 2.1\pm0.2$) kpc. We also smoothed the [*XMM-Newton* ]{}data with a Gaussian beam of FWHM 6 arcsec (shown in Fig. \[xmm\_smth\]) to see if the cavity was detectable in the data. Whilst the cavity is not visible in the [*XMM-Newton* ]{}data due to PSF blurring and the shorter exposure time from the deficit seen in the [*Chandra* ]{}data, we would expect the surface brightness as viewed by [*XMM-Newton* ]{}to drop from 0.35 $\mathrm{counts\ arcsec^{-2}}$to 0.14 $\mathrm{counts\
arcsec^{-2}}$. In the [*XMM-Newton* ]{}data, we observe a surface brightness of $0.2\pm0.5~\mathrm{counts\ arcsec^{-2}}$ in the region defined by the cavity, consistent at the 1-$\sigma$ level with the cavity seen in the [*Chandra* ]{}data.
Spectral analysis {#spectral}
-----------------
From the [*XMM-Newton* ]{}data, we extracted a global spectrum from a source region of 8.8 arcmin centred on the BGG. The source region was chosen as we detect emission at 3-$\sigma$ significance out to this radius. The X-ray background is subtracted using the background files of and we use the double subtraction method of to take into account both the X-ray and the particle backgrounds correctly. The spectrum was fitted with a simple absorbed [<span style="font-variant:small-caps;">MeKaL</span> ]{}model, the details of which are given in Table \[mekal\]. We find a temperature of $1.7\pm0.1$ keV, with metal abundances fixed to $0.3$ [$\mathrm{Z_{\odot}}$]{}. The fit has $\chi^2=147$ for 115 degrees of freedom. We find a luminosity for the group of $6.6_{-0.9}^{+0.4}\times
10^{41}~\mathrm{erg\ s^{-1}}$ which agrees well with the @1998ApJ...496...73M [*ROSAT* ]{}result when differing cosmologies are taken into account. Our value for temperature agrees well with that of @2005ApJ...622..187M, who use the same [*XMM-Newton* ]{}data.
Component Parameter
---------------------------------------------------------- ---------------------------------------------------- ----------------------------------------
Absorption $N_H(\times 10^{22} \mathrm{cm ^{-2}}) (Galactic)$ 0.044
Absorption $N_H(\times 10^{22} \mathrm{cm ^{-2}}) (Fitted)$ $0.042\pm0.02$
M1
[<span style="font-variant:small-caps;">MeKaL</span> ]{} $kT$ (keV) $1.7\pm 0.1$
Norm $(1.69^{+0.07}_{-0.1})\times 10^{-3}$
M2
[<span style="font-variant:small-caps;">MeKaL</span> ]{} Norm $(1.69^{+0.09}_{-0.07})\times 10^{-3}$
PN
[<span style="font-variant:small-caps;">MeKaL</span> ]{} $kT$ (keV) $1.6\pm 0.1$
Norm $(1.06^{+0.03}_{-0.08})\times 10^{-3}$
$L_X (10^{41} \mathrm{erg\ s^{-1}})$ $6.6_{-0.9}^{+0.4}$
$\chi^{2}$(d.o.f.) 147(115)
: Details of the [<span style="font-variant:small-caps;">MeKaL</span> ]{}model fitted to the [*XMM-Newton* ]{}data for all three instruments (M1, M2 and PN). The M1 and M2 temperatures were tied together, and metalicities were fixed to 0.3 [$\mathrm{Z_{\odot}}$]{}. The hydrogen column density was initially fixed to the Galactic value, but then allowed to vary. This did not produce a significant difference in the fits. We fitted an area out to 8.8 arcmin, as described in the text.[]{data-label="mekal"}
We next generated azimuthally averaged temperature profiles from both the [*Chandra* ]{}and [*XMM-Newton* ]{}data. The profiles were again centred on the BGG, and an ellipse defining the region containing the bubble was excluded from the [*Chandra* ]{}data as shown in fig. \[image1\]. For the [*Chandra* ]{}data, annular regions were defined out to 70 kpc (corresponding to 3.4 arcmin), and a blank sky background of the same size and placement on the chip as the source region was used for background subtraction. The background files were processed in the same way as the data (following the instructions in the [*CIAO*]{} threads) and normalized such that both background files and the data had the same flux in the 8–10 keV energy band. Similarly, for the [*XMM-Newton* ]{}data, annuli were defined out to 8.8 arcmin and background regions as described earlier were used. We then followed the method described in @Jethaa for the [*Chandra* ]{}and [*XMM-Newton* ]{}data separately, fitting an absorbed [<span style="font-variant:small-caps;">MeKaL</span> ]{}model to each annulus to deproject the spectra, to produce the deprojected temperature profiles shown by the open points in Fig. \[fig4\]a for [*Chandra* ]{}. We do not show the [*XMM-Newton* ]{}points here, but the [*Chandra* ]{}and [*XMM-Newton* ]{}data are in good agreement in the regions of overlap. The normalisations of the fitted [<span style="font-variant:small-caps;">MeKaL</span> ]{}models were used to obtain the electron density profile ($n_e\left(r\right)$), since: $$n_e = \left\{\frac{4\pi\left[D_A\left(1+z\right)\right]^2
N_{mek}\times 1.17}{1\times10^{-14}V}\right\}^{\frac{1}{2}},
\label{densityeq}$$ where $D_A$ is the angular size distance to the source, $z$ is the redshift of the source, $N_{mek}$ is the normalization of the [<span style="font-variant:small-caps;">MeKaL</span> ]{}component fitted to the annulus, and $V$ is the volume of the annulus from where the spectrum was extracted.
In Fig. \[fig4\], we overplot two different deprojections. The open points show the deprojection when we disregard the presence of the cavity and treat the IGM as azimuthally symmetric. The filled points show the deprojection when we take the presence of the cavity into account and exclude a region corresponding to the cavity from deprojection. As the volume excluded due to the cavity can be a significant fraction of the volume of the shell in this case, this must be taken into account when doing the deprojection. We dealt with the excluded volume in calculating $V$ in Eqn. \[densityeq\] by scaling the volume of an individual shell according to the fractional area of the annulus that was used to extract the spectrum. Thus, for an annulus towards the centre of the bubble, the volume excluded was close to the elliptical volume of the bubble that contributed to the shell, whilst for an annulus which only contained the edge of the bubble, the excluded volume may have overestimated the contribution of the bubble to the shell.
The temperature and density profiles are shown in the top two panels of Fig \[fig4\]. We further derived the pressure profile of the data (shown in the bottom left panel of Fig. \[fig4\]) using $$P(r)=2.96 \times10^{-9}n(r)T(r),
\label{pressure}$$ where the numerical factor is used to convert the temperature and density to cgs units, $n\left (r\right)$ is the electron density profile, and $T\left(r\right)$ is the temperature profile. To ensure that our findings are robust, we redid the deprojection with different regions for the spectral extraction. We found that our profiles were generally consistent with each other, regardless of the extraction regions chosen for the deprojection.
The entropy profile shown in the bottom right panel of Fig \[fig4\] shows marginal signs of disturbances caused by the bubbles in the gas surrounding them. This suggests that the bubbles may be doing work on the IGM. Comparing our profiles to those of Mahdavi et al, we find that our profiles and those of @2005ApJ...622..187M agree well within the core of the group and at large radii. However, in the region between 20–60 kpc, @2005ApJ...622..187M find a slightly lower entropy (but consistent within the errors) than what is found here. This discrepancy arises because the structure in this region is complex and may be significant enough to affect the results of any deprojection; thus the different method of deprojection used in @2005ApJ...622..187M will produce slightly different results. Our method of radial deprojection may well average over real asymmetric structure, but comparison of profiles obtained via different methods suggests that this does not change the final values too drastically.
\
Actual location of the cavity {#explainmorph}
-----------------------------
In Section \[spatial\] we calculate that the ratio of the surface brightness at the centre of the cavity compared to the undisturbed surface brightness, which we termed $\mathbf{R_{\mathrm{SB}}}$ in Section \[spatial\], is $0.4\pm0.1$. We can use this ratio, using a method similar to that of @2007ApJ...659.1153W, to estimate the three-dimensional location of the cavity within the X-ray emitting plasma. We assume that the undisturbed IGM of the NGC 741 can be described by the azimuthally averaged fit to the [*XMM-Newton* ]{}data, using the $\beta$-model as fitted in Section \[spatial\]. The radial distance of the bubble from the centre of this spherically symmetric density profile is 25 kpc as measured from the X-ray maps (Fig \[image1\]). The unknown is the distance of the cavity centre along the line of sight, which we denote $z$ (taking $z=0$ to be the mid-plane of the spherically symmetric group). For a given $z$ the measured density profile allows us to calculate the surface brightness drop produced by a cavity of the observed dimensions, at the observed radial distance.
From the data, $\mathbf{R_{\mathrm{SB}}}=0.4\pm0.1$ (Section \[spatial\]); if the cavity were at $z=0$, then we calculate that $\mathbf{R_{\mathrm{SB}}} = 0.3$, which is within the uncertainty on our measured value. Thus it is possible that the cavity is in the $z=0$ plane. We find that a cavity placed 14 kpc away from the $z=0$ plane reproduces a ratio of 0.4. Given the errors on the decrease in surface brightness, we can calculate an error in the location of the cavity; the bubble could be in the $z=0$ plane, or as far away as 22 kpc from the $z=0$ plane. Combining this with the measured radial distance of 25 kpc we find that the cavity is located at a physical distance of (29$\pm$4) kpc from the centre of the NGC 741 group. Taking this uncertainty into account, the ambient temperature and pressure of the cavity are in the range 1–2 keV and 3–10.4$\times 10^{-12}~\mathrm{dyne~cm^{-2}}$ respectively (taken from the deprojection in which the cavity was excluded: Fig \[fig4\]). This range of temperatures and pressures has implications for both the energy imparted to the IGM and the time-scale over which the energy can be imparted, which we discuss in Section \[energetics\].
Radio sources in the system {#radio}
===========================
NGC 741/2 has been known to be associated with a relatively bright radio source (4C 05.10) since early work by @1980MNRAS.192..595J and @1985ApJ...291...32B. Radio images showed two bright peaks close to the centres of the two galaxies surrounded by more diffuse emission. @1985ApJ...291...32B argued that all the radio emission, including the bright peak close to the nucleus of NGC 742, was associated with NGC 741: on this interpretation NGC 741 would have been something akin to a classical double source and the compact source near NGC 742 would have been its hotspot. It is probably because of this interpretation that NGC 741 is generally thought of as being the radio galaxy of the system. @1994ApJ...436...67V used VLA observations to show that the component close to NGC 742 was truly compact and well aligned with the centre of the galaxy, and proposed instead that the diffuse radio emission was a head-tail or narrow-angle tail (NAT) source associated with NGC 742.
To allow us to make the most definite statements possible about the nature of the radio emission, we have acquired multi-frequency VLA observations (listed in Table \[vlaobs\]) of the system from the archive and reduced them in the standard manner in [AIPS]{}. For display purposes we combine these with a [*Hubble Space Telescope*]{} images taken with the WFPC2 instrument on 1997 Jun 20, obtained from the public archive. In the images obtained from the data (e.g.Fig. \[overlay\]), the two compact sources seen in earlier observations are well aligned with the nuclei of the two galaxies. However, twin jets are clearly visible emerging from the nucleus of NGC 742 and bending westwards to merge into the larger-scale extended emission, while there is no small-scale extension of any kind associated with the nucleus of NGC 741. (A third bright radio source in the field, a small triple whose central component has a position of RA 01$^h$ 56$^m$ 28.5$^s$ Dec. $+05^\circ$ $38'$ $24''$, is coincident with a very faint point source in the [*HST*]{} image and is most likely a background quasar.) The lack of any extension in the NGC 741 nuclear source is consistent with the results of @2000AJ....120.2950X, who show that the source is unresolved even on parsec scales. In the VLA data the compact components have flat radio spectra between 1.5 and 4.9 GHz, indicating that there is self-absorbed parsec-scale structure in both.
We therefore agree with @1994ApJ...436...67V that it is entirely plausible that the extended radio emission in the NGC 741/2 system comes only from a narrow-angle tail source associated with NGC 742. The 400 km s$^{-1}$ difference in the radial velocities of the two galaxies [@2004ApJ...607..202M] is consistent with a narrow-angle tail interpretation of the radio emission, since the much more massive NGC 741 is likely to be close to the centre of mass of the group (its radial velocity is the same as the mean of the 48 galaxies within 1 Mpc – 5570 $\mathrm{km\ s^{-1}}$). The common stellar envelope of the two systems (clearly visible in Fig. \[overlay\]) implies that they are close enough to be interacting tidally, but they need not be as close as they appear in projection (18 kpc): if NGC 742 were displaced by a few tens of kpc along the line of sight with respect to NGC 741 it would help to explain the lack of correspondence between the X-ray emission (dominated by gas on smaller scales) and the observable radio emission, and in particular the absence of any X-ray cavities associated with the radio source. NGC 741 clearly hosts some form of active galaxy, but has no evidence for current large-scale jets.
We further argue that the bubble visible in the system is associated with a previous outburst of NGC 741 rather than with NGC 742 for several reasons. Firstly, NGC 742 already shows evidence of old emission in the form of a long tail; if the cavity had been caused by NGC 742, then the bubble must be older than the current large scale emission. However, the projected distance from the core of NGC 742 to the end of the tail is 160 kpc, while the bubble is located only 45 kpc away. If the bubble is from a previous outburst of NGC 742, then we would have expected it to have travelled further than the observed tail from the current outburst. Secondly, to have got to its current position relative to NGC 742, the bubble would have had to have crossed the centre of the group against the pressure gradient. Thirdly, if the radio emission from NGC 742 were removed from the picture, NGC 741 would look much more similar to some other well-studied ‘ghost cavity’ systems like NGC 4636 [@2002ApJ...567L.115J] or HCG 62 [@2006PASJ...58..719M], with a weak central radio source bearing no obvious relation to the observable cavity. Therefore, in the remainder of the paper we adopt the hypothesis that the bubble was caused by a previous outflow of NGC 741 and that the location of the large scale emission of NGC 742 is entirely coincidental.
---------- ------------- --------------- ----------- ----------------
VLA obs. Date Configuration Frequency Time on source
ID (GHz) (min)
AB920 1999 Jul 18 A 1.5 50
AB593 1991 Dec 16 B 1.5 10
C 4.9 26
AH276 1988 May 19 C 1.5 30
C 4.9 76
AS827 2005 May 07 B 373
---------- ------------- --------------- ----------- ----------------
: Radio observations used in our analysis[]{data-label="vlaobs"}
Energy imparted to the IGM {#energetics}
==========================
In Section \[introduction\], we discussed the possibility that bubbles will do mechanical work on the IGM by expanding and rising, preventing catastrophic cooling. Here, we calculate the amount of energy available to do this. The physics of AGN-inflated bubbles is discussed in length in the literature (e.g. @2001ASPC..240..363B, @2002AAS...201.0304R and @2006astro.ph..2566N), and in all cases, the initial overpressure factor is unknown. We now briefly investigate the total amount of energy available based on whether the bubble was inflated gently (as discussed in @2006MNRAS.372...21A, where the bubble is overpressured by a factor of 2 with respect to the IGM), or if the bubble was injected more violently, say if the overpressure factor was 10, as seen in some young low-power radio sources [@2003ApJ...592..129K].
Assuming that the bubble lies 29 kpc from the centre of NGC 741 and that the ambient pressure is in the range $3-10\times
10^{-12}~\mathrm{dyne\ cm^{-2}}$ (from Fig. \[fig4\]c), we calculate, (using eqn. 18 of @2006astro.ph..2566N and assuming that $\Gamma=4/3$), that if the bubble is overpressured by a factor of 2, then the energy imparted to the IGM is $E_{IGM_{2}}=\left(0.5-1.7\right)\times 10^{58}$ erg. If however, the bubble is overpressured by a factor of 10, then $E_{IGM_{10}}=\left(1-4\right)\times 10^{58}$ erg.
To compare this with the radiative losses of the IGM, i.e. the X-ray unabsorbed luminosity within the region that the bubbles could have affected, a characteristic time-scale needs to be used. This time-scale over which this is done is also important, since if the energy is injected too slowly, it will be insufficient to counteract cooling, and if the energy is released too quickly, it would cause a spike in the temperature of the IGM.
Following @2004ApJ...607..800B, we define three different time-scales, and thus potential ages for the bubble. The first ($t_c$) is based on the time it would take for the bubble to rise to its current location from the BGG at the group’s sound speed, if the bubble had been inflated by a powerful jet close to the radio core (see for example the simulations of @2004MNRAS.348.1105O). The second estimate ($t_b$) is based on the assumption that the bubble is again inflated close to the radio core but rises buoyantly rather than being propelled due to its production mechanism. The third method is to calculate how long it would take to refill the volume displaced by the bubble (@Hydra and @2002ApJ...568..163N), which is denoted $t_r$. We use eqns. 2–4 from @2004ApJ...607..800B to calculate time-scales based on the sound-speed ($t_c$), the buoyant rising time of the bubble ($t_b$) and the refilling time for the bubble ($t_r$) respectively. The time-scale and powers calculated are shown in Table \[power\] and the unabsorbed luminosity of the group within 29 kpc is 0.36$\pm0.02 \times 10^{42}~\mathrm{erg\ s^{-1}}$. Thus, it is clear that over the lifetime of the bubble, and for the foreseeable future, energy input from the bubble will dominate over radiative losses (by 1–2 orders of magnitude), providing more than sufficient energy to counteract cooling losses, even if heating is not 100% efficient.
----- -------------------- ---------- ---------- ---------- -------------------------------- -------------------------------- ------------------------ --------------------------------
$x$ $E_{\mathrm{IGM}}$ $t_c$ $t_b$ $t_r$ $P_{t_c}$ $P_{t_b}$ $P_{t_r}$ $L_X$
$10^{58}$ erg $10^7$yr $10^7$yr $10^7$yr $10^{42}~\mathrm{erg\ s^{-1}}$ $10^{42}~\mathrm{erg\ s^{-1}}$ $10^{42}~\mathrm{erg\ $10^{42}~\mathrm{erg\ s^{-1}}$
s^{-1}}$
2 0.5–1.7 4.5 8.6 6.7 4–10 2–6 4–20 0.36 $\pm$0.02
10 1–4 4.5 8.6 6.7 7–30 4–15 9–30 0.36$\pm$0.02
----- -------------------- ---------- ---------- ---------- -------------------------------- -------------------------------- ------------------------ --------------------------------
$t_c$, $P_c$ are the time-scales and powers derived assuming that the bubble was inflated by a powerful jet and rose to its current position at the sound speed of the group.\
$t_b$, $P_b$ are the time-scales and powers derived assuming that the bubble was inflated by a more gentle jet and rose to its current position buoyantly.\
$t_r$, $P_r$ are the time-scales and powers derived from refilling arguments.
Physical conditions in the bubble {#physcond}
=================================
If the observed cavities in the X-ray emission correspond to bubbles injected by an AGN outburst, then we can use the X-ray and radio data available to constrain the nature of the plasma filling the bubbles. Our starting assumption is that the pressure in the bubble must be at least equal to the external thermal pressure, while it cannot be very much greater than that pressure to avoid driving a shock into the external medium. In addition, the bubble must be light compared to the external medium: it is hard to see how an over-dense ‘bubble’ could have reached its present distance from the nucleus of NGC 741.
Content of the bubbles
----------------------
There are then two basic possibilities for the nature of the bubble material. The first is that it is very hot (i.e. $T_{\rm b} \gg T_{\rm
ext}$), low-density material (for example @2003ApJ...596..159Mand @2007arXiv0706.2768M). The second, and in our view more [*a priori*]{} likely, possibility for the nature of the bubble material is that it consists wholly or partly of a relativistic plasma, containing at least relativistic electrons and magnetic field, as found in the lobes of radio galaxies.
In the first case, the material would simply be a hot thermal plasma, and thus, a signature of this gas may be expected in the X-ray spectrum of the bubble. To investigate this, [**we used a method similar to that of @2006astro.ph.11210S**]{}; we extracted a spectrum for the region defined as the cavity in Section \[spatial\]. We fit the background subtracted (using a blank-sky background), grouped spectrum with an absorbed [<span style="font-variant:small-caps;">MeKaL</span> ]{}model with a temperature fixed to that of the gas immediately around the cavity, with the normalisation free to vary. This model provides a reasonable fit to the cavity spectrum, but cannot conclusively rule out the presence of very hot gas in the cavity. Thus, this component is frozen in the fit and a second [<span style="font-variant:small-caps;">MeKaL</span> ]{}component is added to model the possible hot phase with fixed abundance (set to 0.3[$\mathrm{Z_{\odot}}$]{}) and temperature, leaving only the normalisation of this component free to vary. We vary $kT$ by hand between the temperature ranges of 5–40 keV, and for each value of $kT$ we obtain a $3-\sigma$ upper limit on the normalisation and convert this into the maximum pressure of the hot gas at a given temperature. If this pressure is significantly greater than the pressure of the external gas, it implies that the bubble would not be stable under this configuration (as there may also be an additional pressure component from non-thermal sources). However, if the pressure in the bubble from the putative hot component is less than the external pressure, this provides a lower limit on the temperature of any hot gas component. By varying the temperature of the second component until the maximum pressure in the hot component exceeds the external pressure, we can set a limit on the temperature of any hot component present in the gas. It should be emphasised that this is not a ’fit’ to the spectrum in the usual sense as there are insufficient counts to fit a two-component model. The method described merely provides a way of testing for the possible presence of a second component.
As the bubble is in direct mechanical contact with the IGM at a range of different radii (due to the fact that it is three dimensional), using a single temperature from the deprojected temperature profile would not be an accurate reflection of the temperature around the cavity. However, using a range of temperatures would not provide a satisfactory constraint on the pressure. We thus obtain the projected temperature in an annulus at the radius of the centre of the cavity, to provide a more representative temperature and pressure constraint. Fitting the third annulus from the centre of the [*Chandra* ]{}data with a simple absorbed [<span style="font-variant:small-caps;">MeKaL</span> ]{}model in the 0.5–5.0 keV energy range and fixing the abundance to 0.3[$\mathrm{Z_{\odot}}$]{}, we obtain an external temperature of $1.41\pm 0.1$ keV. Assuming that the lower limit of the pressure of the gas is approximately $5\times 10^{-12}~\mathrm{dyne\ cm^{-2}}$, an upper limit for $T_H$ can be found. We obtain a limit of 10–20 keV for the temperature of any putative hot component; above this temperature range, the gas is over-pressured with respect to the external pressure. Thus, we cannot rule out the presence of hot tenuous gas in the bubble.
In the second case, as discussed in Section \[introduction\], the combination of electron energy spectrum and magnetic field must be such that the bubble produces no detectable radio or inverse-Compton X-ray emission. For simplicity we consider a uniform electron spectrum and magnetic field strength throughout the bubble, and take the magnetic field to be ‘tangled’ so that it has an isotropic magnetic stress tensor [@Leahy1991]. Let the electron energy spectrum be $N(E)$ such that $N(E) {\rm d}E$ is the number density of electrons with energies between $E$ and $E+{\rm d}E$, and let the energy density in non-radiating particles be some multiple $\kappa$ of the energy density in electrons, as is conventional: then the requirement of pressure balance gives us $${{B^2}\over{2\mu_0}} + (1+\kappa) \int_{E_{\rm min}}^{E_{\rm max}} E
N(E) {\rm d} E = 3p_{\rm ext}
\label{rpb}$$ where $B$ is the magnetic field strength and $E_{\rm{min}}$ and $E_{\rm{max}}$ are the limits on the electron energy densities to be considered. If we further assume some physically reasonable form for the electron energy density, such as a power law in energy ($N(E) =
N_0 E^{-p}$), then, for a given $B$, $\kappa$, $E_{\rm{min}}$ and $E_{\rm{max}}$ we can solve equation \[rpb\] for the normalization $N_0$ of the electron spectrum, and so, effectively, for the total number or energy density of the electron (particle) population.
Pressure balance has been investigated using this basic formalism for a variety of different types of [*filled*]{} cavities, with differing results. In powerful FRII [@fr1974] radio galaxies, it has been possible to measure the magnetic field strengths in the lobes of some systems by detection of inverse-Compton scattering of the cosmic microwave background (CMB): the field strengths turn out to be comparable to, though slightly below, the equipartition values for an electron/positron ($\kappa=0$) plasma only [@2005ApJ...626..733C], so that $B \approx B_{\rm eq}$, where $${{B_{\rm eq}^2}\over{2\mu_0}} = (1 + \kappa) \int_{E_{\rm min}}^{E_{\rm max}} E
N(E) {\rm d} E
\label{equip}$$ and the corresponding plasma+magnetic field pressures are, in several cases, comparable to the measured external pressures [@2002ApJ...581..948H; @2004MNRAS.353..879C], suggesting that $\kappa$ is indeed $\approx 0$ in these systems. Even where no external pressure can be measured the fact that the measured magnetic field is close to the equipartition value for $\kappa = 0$ has to be a systematic coincidence if $(1+\kappa) \gg 1$ in these systems. On the other hand it has been known for many years that the minimum internal pressure ($\kappa = 0$, $B \approx B_{\rm
eq}$) in the lobes of low-power FRI radio galaxies often falls orders of magnitude below the measured external pressure: [@2005MNRAS.364.1343D] have recently commented on this effect in a sample of cluster-centre objects. In these systems it seems that either $(1 + \kappa) \gg 1$ (as a change in the filling factor of the plasma has a similar effect to an increase in $\kappa$, we do not consider this separately) or there is a strong departure from equipartition. In some cases the lack of detectable inverse-Compton emission has been used to argue that it is not possible that $B \ll
B_{\rm eq}$. In the following section we discuss the constraints that the synchrotron and inverse-Compton limits on the NGC 741 cavity place on the nature of the cavity-filling plasma.
Constraints on the nature of the plasma
---------------------------------------
The key free parameters of eq. \[rpb\] are $B$, $\kappa$, $E_{\rm{min}}$ and $E_{\rm{max}}$, and $p$. We cannot constrain all of these from our observations. However, it seems unlikely that the plasma is magnetically dominated, and so one constraint that we can impose is that the magnetic field should be less than or equal to the equipartition value, $B \la B_{\rm eq}$. The limits on observed synchrotron and inverse-Compton emissivity place additional constraints on the parameters of the plasma: synchrotron emissivity at a given frequency constrains a combination of $N_0$, $B$ and $E_{\rm
max}$, while inverse-Compton emissivity constrains $N_0$ and to some extent $E_{\rm max}$ only. It is easiest to calculate these constraints numerically, and we do this using the code of @1998MNRAS.294..615H, assuming that the dominant photon field for inverse-Compton scattering is the cosmic microwave background radiation (i.e. ignoring the negligible synchrotron self-Compton component). From the radio data in hand we set an upper limit on the bubble flux density at 1.4 GHz of 7.4 mJy and at 330 MHz (Jetha et al. in prep) of 20 mJy. These limits were obtained from VLA observations of the source, with the 1.4 GHz data being the same as used in Section \[radio\] and the 330 MHz from new observations (Jetha et al, in prep.). For each observation, since part of the radio source NGC 742 obscures the cavity, the non-detection limits of the cavity were taken to be three times the off-source rms noise multiplied by the square root of the number of beams in the cavity at each frequency. From the X-ray observations, we impose an upper limit on the inverse-Compton flux density of 5.8 nJy at 1 keV.
We fix the power-law index $p$ at a relatively flat value (we choose $p=2.1$ based on the results of @2005ApJ...626..748Y for FRI radio sources) which reflects expectations from particle acceleration models: although there should almost certainly be a tail of particles with a steeper value of $p$ due to spectral ageing (as reflected in the steep spectral index of some observed radio lobe systems) it is very unlikely that it is correct to extrapolate this spectral index back to low energies, where the electron loss time-scales are very long, and the steep-spectrum tail makes a negligible contribution to the overall energy budget of the electrons. We conservatively take the minimum electron energy to be $E_{\rm min} = m_{\rm e}c^2$, i.e. $\gamma_{\rm min} = 1$, where $\gamma$ is the Lorentz factor of the electron). For a given value of $\kappa$, we can investigate the constraints on $E_{\rm max}$ ($\gamma_{\rm max}m_{\rm e}c^2$) and $B$ given by the observations on the assumption of pressure balance as in equation \[rpb\].
The results are plotted in Fig. \[constraint-plot\] for two representative values of $\kappa$ as defined above. We chose two values, $\kappa=0$ (i.e. relativistic electrons/positrons and field give the only contribution to the internal pressure of the plasma, as in FRII sources) and $\kappa=10$ (i.e. there is a population of non-radiating particles, such as protons, with energy density exceeding that of the electrons by an order of magnitude). We carry out the calculations using the best estimate of the external pressure ($7.7\times 10^{-12}~\mathrm{dyne\ cm^{-2}}$) and the range imposed by the uncertainties in temperature, density and cavity position (3–10.4$\times 10^{-12}~\mathrm{dyne~cm^{-2}}$). It can be seen that the radio and X-ray data already impose quite severe constraints on the possible electron energy spectrum if the pressure is close to its best-fitting value. For $\kappa = 0$ the combination of inverse-Compton and synchrotron constraints requires $\gamma_{\rm
{max}} < 1000$ for any plausible value of $B$. Even for $\kappa = 10$ the magnetic field strengths must be very low compared to their equipartition values for $\gamma_{\rm max}$ to be greater than a few $\times 10^3$: this compares to $\gamma_{\rm {max}} \ga 2 \times 10^4$ for the lobes of typical radio sources with detections at GHz frequencies (assuming fields close to equipartition). For the values of $\kappa$ investigated (0 and 10), the bubble can be filled with a plasma with $\gamma_{\rm max}\ga 2 \times 10^3$ only for the lowest external pressures [*and*]{} the lowest internal magnetic fields. It is clearly possible to choose a value of $\kappa$ such that we have no constraint on the properties of the electron population – this corresponds to a situation in which the energetics of the bubble plasma are entirely dominated by an ‘invisible’ component. However, it is interesting to consider the constraints on the electron population for a situation in which $(1 + \kappa) \approx 1$. For these low values of $\kappa$, how can we obtain the required low values of $\gamma_{\rm max}$?
Radiative losses alone (to synchrotron and inverse-Compton emission) would require a long time to reduce an initially high $\gamma_{\rm {max}}$ to values $\sim 1000$. The maximum energy in an electron population that initially extended to high energies after a time $t$ is given by @1970ranp.book.....P $$\gamma_{\rm {max}} = {3\over 4} {{m_{\rm {e}} c}\over{\sigma_{\rm {T}} U t}}$$ where $m_e$ is the mass of the electron, $\sigma_T$ is the Thomson cross-section and $U$ is the total energy density in magnetic field and microwave background photons (i.e. the total energy density implicated in loss processes). For equipartition magnetic field strengths synchrotron losses dominate, $U = 3p_{\rm ext}/2$ and the age of the bubble is $> 2 \times 10^8$ years (though note that this calculation takes no account of the fact that the magnetic field may have been higher at earlier times). For magnetic field strengths much less than equipartition inverse-Compton losses dominate, $U =
8\pi^5(kT_{\rm CMB})^4/15(hc)^3$, and the limit on the age of the bubble is closer to $> 2 \times 10^9$ years, where $t$ is the time since any high-energy particle acceleration took place in the source. These time-scales are considerably longer than the time-scales for the bubble to rise buoyantly to its current position ($8.6\times 10^7$ years, Section \[energetics\]). $\gamma_{\rm max}$ may also be reduced by adiabatic expansion. which gives $\gamma_{\rm max} \propto
V^{-(\Gamma-1)}$, where $\Gamma$ is the adiabatic index, 4/3. To take a simple example, we could suppose that the original bubble had $\gamma_{\rm max} = 2 \times 10^4$ as quoted above and has expanded adiabatically. Then the volume must have changed by a factor 8000 during the expansion; i.e. the initial radius of the bubble was around 1 kpc. But this would require the initial bubble to have been over pressured by a factor $1.6 \times 10^5$ if it is in pressure balance now, which seems unlikely. As shown by Alexander (1987), more modest expansion factors coupled with radiative losses do not greatly change the inferred ages.
We can therefore conclude that it seems hard for the bubble to have evolved from a ‘normal’ region of radio-emitting plasma – like the plumes of NGC 742 whose radio emission we see in the system at present – if the lobe external pressure is close to the best-fitting value. The present X-ray data do not allow us to constrain the pressure around the cavity well enough to rule this possibility out completely, though, and a deeper X-ray observation is required to allow us to draw firm conclusions: if the pressure is at the low end of the range we estimate, and the internal pressure of the cavity is dominated by non-radiating relativistic protons with a very low magnetic field, then the available constraints are much weaker. In a future paper we will apply the technique used above to constrain the nature of the cavity-filling plasma in other ‘ghost cavity’ systems.
Conclusions
===========
We have examined the energetics of the galaxy group associated with NGC 741. We find that if the cavity seen in the X-ray data was originally inflated by a radio-loud AGN associated with the brightest galaxy, NGC 741, then there is sufficient energy to counteract cooling. This interpretation is consistent with the fact that NGC 741 clearly has an active nucleus at the present time, and this could have been more active in the past.
We have investigated the nature of the cavity-filling plasma. We cannot rule out the possibility that the cavity is filled with a very hot thermal plasma ($kT > 10$ keV) or with a relativistic plasma that is energetically completely dominated by non-radiating particles. However, it is not at all clear that current models for AGN inflation of cavities can produce cavities of this kind, particularly when we take into account the relatively small distance (30 kpc) between the cavity and the centre of the host group. If on the other hand there is a substantial contribution to the energy density in the cavity from a plasma that contains relativistic electrons and magnetic field, as in the active lobes of radio galaxies (such as NGC 742 in the same group) then we find that the limits placed by the non-detection of radio synchrotron and X-ray inverse-Compton radiation severely constrain the maximum electron Lorentz factor. In this scenario, we find that it is very hard for the cavity to have evolved to its current state from a ‘dead’ radio lobe by a combination of the standard processes expected to be operating (synchrotron, inverse-Compton, and adiabatic expansion losses); for radiative losses only the lobe would have to be older ($\ga 2 \times 10^8$ years) than any plausible bubble inflation time (between $5 \times 10^7$ and $9 \times 10^7$ years). The extreme parameters required for the NGC 741 cavity, and for other ‘ghost cavities’ whose physical parameters have been investigated (e.g. Dunn et al 2005) and the fact that cavities are found at a wide range of radii (Birzan et al 2004) may indicate that inflation by the lobes of a standard radio-loud AGN is not a viable explanation for all ghost cavities. In the case of NGC 741, we are left with one clearly outstanding question: if the cavity in NGC 741 is due to an AGN-inflated bubble, in what form was the energy injected into the IGM, and in what form is it now?
Acknowledgements {#acknowledgements .unnumbered}
================
NNJ thanks CNES (the French Space Agency) for funding. MJH thanks the Royal Society for a research fellowship. AB is supported by NSERC (Canada) Discovery Grant program. He is grateful to the Leverhulme Trust for awarding him a Leverhulme Visiting Professorship, and thanks University of Oxford, Institute for Computational Cosmology (Durham) and University of Birmingham for their hospitality during his frequent visits. EO’S acknowledges support from NASA grant NNX06AE90G-R. The authors would like to thank Monique Arnaud for her help with the [*XMM-Newton* ]{}data reduction and for useful conversations regarding the data, and Gabriel Pratt and Etienne Pointecouteau for their help with the [*XMM-Newton* ]{}data reduction software used in this paper. The authors also thank the anonymous referee for their useful comments in improving the text.
\[lastpage\]
[^1]: E-mail:[email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the framework of the Two Higgs Doublet Model (2HDM) type III appears two charged Higgs boson and recently there are experimental reports from D0 and CDF collaborations searching a particular signature of new physics. We present a review of the analysis done in the region $M_{H^+}>m_t$ by D0 collaboration and we use the ratio $R_\sigma$ for the region $M_{H^+} < m_t$ in different scenarios of space parameter of this model.'
author:
- 'H. Cardenas'
- 'J-Alexis Rodríguez'
title: Searching the charged Higgs boson of the type III two Higgs doublet model
---
The Two Higgs Doublet Model (2HDM) is a simple extension of the standard model (SM), where an additional Higgs doublet is introduced [@higgshunter]. The two complex scalar fields correspond to eight degrees of freedom, three of them are absorbed as Goldstone bosons in order to give mass and as longitudinal components to the $W^\pm$ and $Z$ bosons. The remaining degrees of freedom are interpreted as five physical states: two neutral scalars $h^0$ and $H^0$, a pseudo-scalar $A^0$, and a pair of charged Higgs bosons $H^\pm$. Another consequence of this extension is a more generic pattern of Flavor Changing Neutral Currents (FCNC), including such processes at tree level. Because of low energy experiments these FCNC could be a problem, usually solved by imposing a discrete symmetry that allows to couple at most one Higgs doublet to each fermion. In the 2HDM called type I, one Higgs doublet gives masses to the up and down quarks, simultaneously. And, in the model type II, one Higgs doublet gives masses to the up type quarks and the other one to the down type quarks, and it is precisely the structure of the minimal supersymmetric extension of the SM (MSSM). However, the addition of these discrete symmetries is not mandatory and in that case both doublets contribute to generate the masses for up-type and down-type quarks, such model is know as the 2HDM type III [@original].
While it may be hard to distinguish any one of these neutral Higgs bosons of the 2HDM-III from that one of the SM, the charged $H^\pm$ pair carry a distinctive hall-mark of this kind of new physics. Hence the charged Higgs boson plays a very important role in the search of new physics beyond the SM.
Usually the limits on the charged Higgs mass are in the context of the type II 2HDM. Direct searches have carried out by LEP experiments, searching for pair produced charged Higgs bosons via $s$-channel exchange of a [*[Z]{}*]{}-boson or a photon. LEP collaborations yield a combined lower limit on $M_{H^\pm}$ of 78.6 GeV [@LEP] assuming $H^+ \to \tau^+ \nu_\tau(c \bar s)$ in a wide range of the ratio of the vacuum expectation values of the two Higgs doublets, $\tan \beta=v_2/v_1$. Searches in the hadron colliders consider two cases $M_{H^\pm} > m_t$ and $M_{H^\pm} < m_t$. At the Tevatron, the direct searches for charged Higgs boson are based on $p \bar p \to t \bar t$ with at least one top quark using $t \to H^+ b$, for the first case $M_{H^\pm} > m_t$. The CDF collaboration has reported a direct search for charged Higgs boson decaying into $\tau^+ \nu_\tau$, $c \bar s$ $t \bar b$ or $W^+ A^0$, which uses measurements of the top pair production cross section in the $leptons + \not E_T +jets + leptons$ channels, from data samples corresponding to an integrated luminosity of $193$ pb$^{-1}$ [@CDF1]. On the other hand, recently the D0 collaboration has presented a direct search for a charged Higgs boson produced by $q \bar q$ annihilation and decaying to $t \bar b$ final state, in the $180 \leq M_{H^+} \leq 300$ GeV mass range, it is $m_{H^+}>m_t$, using 0.93 fb$^{-1}$ of data collected at center-of-mass energy $\sqrt{s}=1.96$ TeV [@D01]. The analysis lead to upper limits on the production cross section in the 2HDM types I, II and III [@D01; @peters]. We are going to discuss different scenarios for the 2HDM type III using these experimental results. Also, searches in the $M_{H^\pm} < m_t$ region have been performed, using the production cross section of top quark pairs at the Tevatron [@CDF2; @peters], and we intend to exploit this fact to discuss the possible limits on the charged Higgs boson mass obtained from measurements of the ratio $R_\sigma=\sigma_{t \bar t}^{l+jets}/\sigma_{t \bar t}^{dileptons}$ in the framework of the 2HDM type III.
Other experimental bounds on the charged Higgs boson mass come from virtual charged Higgs boson production in $b\to s\gamma$ transitions [@bsgamma]. Finally, the search for the charged Higgs boson will continue above the top quark mass at the experiments of LHC, ATLAS and CMS. The main production mechanisms would be the processes $gg \to tb H^+$ and $gb \to t H^+$ which have been studied using simulations of the LHC detectors [@LHC].
In the 2HDM both doublets can acquire vacuum expectation value (VEV) different from zero. As mentioned before, we will focus our study to the 2HDM type-III, which owns the most general Yukawa Lagrangian (without additional symmetries imposed), given by [@ARS; @csy; @we] $$-\mathcal{L}_Y^{(III)}=\eta_{ij}^{U,0}\bar Q_{iL}^0\widetilde
\Phi_1U_{jR}^0+\eta_{ij}^{D,0}\bar Q_{iL}^0\Phi_1D_{jR}^0+\xi_{ij}^{U,0}
\bar Q_{iL}^0\widetilde\Phi_2U_{jR}^0+\xi_{ij}^{D,0}\bar Q_{iL}^0
\Phi_2D_{jR}^0+\text{h.c},$$ where $\Phi_i$ are the complex scalar fields, $\eta_{ij}^0$ and $\xi_{ij}^0$ are complex (even if later on we simplify the problem to real $\lambda$, the original coupling matrices can be also complex as in the Yukawa sector of the SM) and non-diagonal coupling matrices and the suffix “0” denotes flavor eigenstates.
![The cross section times branching fraction versus the charged Higgs boson for different scenarios in the type III 2HDM []{data-label="crossfig"}](figura_sbr.ps)
The scalar spectrum has mass eigenstates which contain two $CP$-even neutral Higgs bosons ($h^0$, $H^0$) coming from the mixing of $\Re (\phi_i^0)$ with mixing angle $\alpha$; two charged Higgs bosons ($H^\pm$) which corresponds to a mixture of the would-be Goldstone bosons $G_W^\pm$ through the mixing angle $\tan \beta=v_2/v_1$; and one $CP$-odd Higgs ($A^0$) which mix the neutral would-be Goldstone boson $G_Z^0$. In the framework of the type III 2HDM there is a global symmetry which can make a rotation of the Higgs doublets and fix one VEV equal to zero [@ARS]. In such a way, $v_1=v$ and $v_2=0$, and the mixing parameter $\tan\beta=v_2/v_1$ can be eliminated from the Lagrangian. It is worth to mention that the experimental sensitivities are usually displayed in terms of the parameters $m_A$ and $\tan \beta$ or $M_{H^\pm}$ and $\tan \beta$, that characterize the MSSM Higgs sector or the 2HDM type II. If the parameter $\tan \beta$ is eliminated from the Lagrangian, we have the usual 2HDM type III [@ARS], where the Lagrangian of the charged sector is given by $$-{\cal L}^{III}_{H^\pm ud}=H^+ \bar U [ K \xi^D P_R-\xi^U K P_L] D+h.c.$$ where $K$ is the Cabbibo-Kobayashi-Maskawa (CKM) matrix and $\xi^{U,D}$ the flavor changing matrices. In the present work we take into account the Cheng-Sher-Yuan (CSY) parameterization which is built as the geometrical mean of the Yukawa couplings of the quark fields, $\xi_{ij}\equiv\frac{\sqrt{m_im_j}}{v}\lambda_{ij}$ [@csy] .
The search of D0 collaboration [@D01] use the process $q \bar q' \to H^+ \to t \bar b$ followed by $t \to b (W^+ \to l \nu)$, which correspond to a single top production, with a final signature 2-$b$-tagged $+ lepton +$ missing energy. We notice there are two vertices type $H^+ q \bar q'$ where appear the parameters $\xi_{q q'}$ and then different considerations can be done, taking into account that appears terms like $\sum_j (K_{qj} \xi_{j q'})$. In the annihilation vertex $q \bar q' H^+$, it is possible in the framework of the 2HDM type III: $c \bar b H^+$ and $c \bar s H^+$; about this point H. He and C. P. Yuan [@yuan] showed that it is possible to enhanced the production cross section if the vertex $c \bar b H^+$ is important, and it is possible if the parameter $\lambda_{tc}$ is bigger than one. We have evaluated the option of the vertex $c \bar s H^+$ in the framework of the 2HDM-III and it is smaller two or three orders of magnitude than the vertex $c \bar b H^+$. Now in the second vertex, the $H^+$ decay vertex, we have several couplings regarding the 2HDM-III allowed processes: $H^+ \to t \bar s$, $H^+ \to c \bar b$ and $H^+ \to t \bar b$. The first option, into $t \bar s$, does not have a quark $b$ in the final state, and the experiment is looking for at least two $b$-tagged , one coming from the $H^+$ decay and the second one coming from the top quark decay. The second option is into $c \bar b$ but it is doubly suppressed for the factor $\xi_{tc}$ and for the Cabbibo-Kobayashi-Maskawa $K_{cb}$ factor. Finally, we will have the same main channel production in the 2HDM type III that in the types I and II, $c \bar c \to H^+ \to t \bar b$ followed by $t \to (W^+ \to l \nu) b$.
Using the parameterization CYS [@csy] and the numerical values of the CKM matrix $K_{ij}$, the flavor changing couplings, $\sum_j K_{qj} \xi_{jq'}$, of the charged Higgs boson to the fermions are shown in Tab. \[acoples2\]. These factors have been simplified taken into account the numerical values of quark masses and assuming that the parameters $\lambda_{ij}$ involving the first and second generations are smaller than parameters involving the third generation. Finally, bounds and restrictions on the $\lambda_{ij}$ for the quark sector and $\xi_{ij}$ for the leptonic sector can be found in literature [@ARS; @radiaz].
![Branching ratios in the 2HDM-III for the parameters defined in each figure.[]{data-label="branchings"}](figura_br5.ps "fig:") ![Branching ratios in the 2HDM-III for the parameters defined in each figure.[]{data-label="branchings"}](figura_br28.ps "fig:") ![Branching ratios in the 2HDM-III for the parameters defined in each figure.[]{data-label="branchings"}](figura_br110.ps "fig:")
----------------------------------------------------------------------------------------------------
$(ij)$ $(\xi^U K)_{ij}$ $(K \xi^D)_{ij}$
----------------- -------------------------------------------------- -------------------------------
$c\overline{s}$ $K_{ts}\frac{\lambda _{ct}\sqrt{m_{c}m_{t}}}{v}$
$c\overline{b}$ $K_{tt}\frac{\lambda _{ct}\sqrt{m_{c}m_{t}}}{v}$ $K_{cb}
\frac{\lambda _{bb}m_{b}}{v}$
$t\overline{s}$ $K_{ts}\frac{\lambda _{tt}m_{t}}{v}$
$t\overline{b}$ $K_{tb}\frac{\lambda _{tt}m_{t}}{v}$ $K_{tb}\frac{
\lambda _{bb}m_{b}}{v}$
----------------------------------------------------------------------------------------------------
: The couplings of the vertices $H^+ q \bar q'$ are proportional to $\sum_j (K_{qj} \xi_{j q'})$ but they can be simplified taking into account the CSY parameterization and the numerical values of the $K_{CKM}$ matrix as it is shown.[]{data-label="acoples2"}
The analysis presented by D0 collaboration [@D01] for the type III 2HDM has followed the analysis done in reference [@yuan] assuming that the parton level is important to enhance the cross section if $\lambda_{tc}$ is bigger than one, the experimental analysis have used $\lambda_{tc}=5$ [@peters]. Further, they assumed the parameter $\lambda_{tt}$ in the charged Higgs decay vertex, to be equal to $\lambda_{tc}$. About this last point, we should mention that Atwood, [*et. al.*]{} in reference [@ARS] have already shown that the assumption $\lambda_{ij}=\lambda$ is not in agreement with the low energy phenomenology and on the other hand, it has been shown [@rozo] that perturbation theory validity requires that $\lambda_{tc} \leq 2.8$. From this point we aim to explore scenarios allowed in the 2HDM-III space parameters, with the additional simplification that in this model $\tan\beta$ is a spurious parameter. The experimental D0 collaboration report the observed limits on the production cross section ($pb$) times the branching fraction, $\sigma(q \bar q' \to H^+) \times B(H^+ \to t \bar b)$, these limits are shown in Fig. \[crossfig\]. We have used the program CompHEP [@comphep] to simulate charged Higgs boson production and decays, $q \bar q' \to H^+ \to t \bar b \to W^+ b \bar b \to l^+ \nu b \bar b$ where $l$ represents an electron or muon. The expected limits using the same values of the $\lambda_{ij}$ in the charged Higgs boson mass range 180 to 300 GeV are plotted in Fig. \[crossfig\]. In addition, here can be seen the predictions taken $\lambda_{tc}=2.8, 1$, $\lambda_{tt}=2.8, 1$ and $\lambda_{bb}=0,10$. We can conclude that restrictions on the parameter space of the type III 2HDM are not too stronger. We should notice that in the case $\lambda_{ij}=5$ only charged Higgs masses above around 264 GeV are allowed and for values $\lambda_{tt}=2.8$ the charged Higgs boson mass should be bigger than 230 GeV. For completeness the branching fraction ratios in the different scenarios considered are shown in Fig. \[branchings\], where we have use the parameters $\xi_{\tau \tau}=2.2 \times 10^{-2}$, $\xi_{\tau \mu}=2.1 \times 10^{-1}$, and $\xi_{\mu \mu}=0.12$ according to the allowed regions cited in the literature [@ARS; @radiaz].
The discussion presented until now is for charged Higgs boson mass bigger than the top quark mass, but we can also ask for the contrary case $M_{H^+}<m_t$. In this case the decay $t \to H^+ b$ compete with the standard decay channel $t \to W^+ b$ and the experimental analysis has been done using the cross section of the top quark pair production [@peters; @CDF2]. In order to estimate upper limits on the branching fraction $B(t \to H^+ b)$ is useful to use the ratio $R_\sigma=\sigma(p\bar p \to t \bar t)_{l+jets}/\sigma(p \bar p \to t \bar t)_{dilepton}$. The ratio $R_{\sigma}$ should be consistent with unity if the dilepton and lepton+jets analyses are in the framework of the standard model $t \bar t$ production. The ratio $R_\sigma$ will have smaller systematic uncertainties than individual cross section measurements and also some common factors will cancel. And of course, $R_\sigma$ is independent of any theory prediction for $\sigma(p\bar p\to t \bar t)$, so it is a better place to look for new physics by looking for a deviation in the ratio, rather than by comparing a measured cross section to a theory prediction. $R_\sigma$ is sensitive to decays such as $t \to H^+ b$, it is possible to give an interpretation in terms of the branching fraction and a measurement of $R_\sigma$ can be translate into an upper limit on $B(t \to H^+ b)$. For instance if we consider that $B(H^+ \to cs)=100 \% $ then we can get $B(t \to H^+ b)=0.13 \pm 0.12$ in the type II 2HDM. In general this ratio can be written according to [@peters; @quadt] $$R_\sigma = 1 + \frac{B(t \to H^+ b)}{[1-B(t \to H^+ b)] B(W^+ \to q \bar q')} \, .$$ Experimentally this ratio has been measured by D0 collaboration $R_\sigma=1.21^{+0.27}_{-0.26}$ [@peters] and by the CDF collaboration $1/R_\sigma =1.45^{+0.83}_{-0.55}$ [@CDF2]. We have plotted the ratio $R_\sigma$ in figure \[rsigma\] in the framework of the 2HDM type III for the scenarios that we used so far.
![The ratio $R_\sigma$ with the experimental limits from D0 collaboration and different values of the space parmeter in the framework of the type III 2HDM []{data-label="rsigma"}](figura_rsigma.ps)
As a conclusion we have reviewed the analysis presented by D0 collaboration [@D01] in the first search for a charged Higgs boson directly produced by quark-antiquark annihilation and decaying into the $t \bar b$ final state, in the $180 \leq M_{H^+} \leq 300$ GeV, analyzing 0.9 fb$^{-1}$ of data from $p \bar p$ collisions at $\sqrt{s}=1.96$ TeV, in the framework of the 2HDM type III. We have explored different scenarios of the space parameter in the region $M_{H^+} > m_t$ and in the region $M_{H^+} < m_t$ we have used the measurements of the ratio $R_\sigma$ [@peters; @CDF2]. Results are shown in figures \[crossfig\] and \[rsigma\].
We would like to acknowledge to Y. Peters for useful discussions about the experimental results from her talk presented in SUSY08 conference (Seoul, Korea, June 2008) and to C. Jimenez for his valuable help with data manipulation. Also we want to acknowledge to D. Milanes for his careful reading of the manuscript and useful suggestions.
[99]{} J. Gunion, [*et. al., The Higgs Hunter’s Guide*]{}, Frontiers in Physics (Addison-Wesley, Redwood City, CA, 1990).
S. Glashow and S. Weinberg, Phys. Rev. D [**15**]{}, 1958 (1977);W. S. Hou, Phys. Lett. B [**296**]{}, 179 (1992); D. Chang, W. S. Hou and W. Y. Keung, Phys. Rev. D [**48**]{}, 217 (1993) \[arXiv:hep-ph/9302267\].
LEP Higgs Working Group for Higgs boson searches, arXiv:hep-ex/0107031; P. Achard, [*et. al.,*]{}\[L3 Collaboration\], Phys. Lett. B [**575**]{}, 208 (2003)\[arXiv:hep-ex/0309056\].; A. Heister, [*et. al.,*]{} \[ALEPH Collaboration\], Phys. Lett. B [**543**]{}, 1 (2002) \[arXiv:hep-ex/0207054\]; J. Abdallah, [*et. al.,*]{}\[DELPHI Collaboration\],Eur. Phys. J. C [**34**]{}, 399 (2004) \[arXiv:hep-ex/0404012\].
A. Abulencia, [*et. al*]{} \[CDF Collaboration\], Phys. Rev. Lett. [**96**]{}, 042003 (2006) \[arXiv:hep-ex/0510065\].
V. M. Abazov, [*et. al*]{}\[D0 Collaboration\], arXiv:hep-ex/0807.0859.
Y. Peters, Search for charged Higgs boson at D0, Talk presented in SUSY08 conference, 16-21 June 2008, Seoul, Korea, \[arxiv:hep-ex/0810.2078\]
K. Bloom, arXiv:hep-ex/0405020.
F. M. Borzumati and C. Greub, Phys. Rev. D [**58**]{}, 074004 (1998) \[arXiv:hep-ph/9802391\]; Z. j. Xiao and L. Guo, Phys. Rev. D [**69**]{}, 014002 (2004) \[arXiv:hep-ph/0309103\].
M. Carena, S. Heinemeyer, C. Wagner and G. Weiglein, Eur. J. Phys. C [**45**]{}, 797 (2006 \[arXiv:hep-ph/0511023\]; K. Assamagan, Y. Coadou and A. Deandrea, Eur. J. Phys. C [**4**]{}, 9 (2002); K. Assamagan and N. Gollub, Eur. J. Phys. C [**39S2**]{}, 25 (2005) \[arXiv:hep-ph/0406013\]; M. Hashemi, S. Heinemeyer, R. Kinnunen, A. Nikitenko and G. Weiglein, arXiv:hep-ph/0804.1228.
D. Atwood, L. Reina and A. Soni,Phys. Rev. D [**55**]{}, 3156 (1997) \[arXiv:hep-ph/9609279\] and references there in.
T. P. Cheng and M. Sher. *Phys. Rev.* **D35**, 3484 (1987).
R. A. Diaz, R. Martinez and J-Alexis Rodriguez, *Phys. Rev.* **D63**, 095007 (2001), \[arXiv:hep-ph/0010149\].
H. He and C. P. Yuan, Phys. Rev. Lett. [**83**]{}, 28 (1999).
R. A. Diaz, R. Martinez and C. E. Sandoval, \[arXiv:hep-ph/0311201\] (2003); R. A. Diaz, R. Martinez and C. E. Sandoval, \[arXiv:hep-ph/0406265\] (2004); R. A. Diaz, R. Martinez and J-Alexis Rodriguez, \[arXiv:hep-ph/0103050\] (2001); R. A. Diaz, R. Martinez and J-Alexis Rodriguez, *Phys. Rev.* **D63**, 095007 (2001), \[arXiv:hep-ph/0010149\].
R. Martinez, J. Alexis Rodriguez and M. Rozo, Phys. Rev. D [**68**]{}, 035001 (2003) \[arXiv:hep-ph/0212236\].
E. Boos [*et al.*]{} \[CompHEP Collaboration\], Nucl. Instrum. Meth. A [**534**]{}, 250 (2004) \[arXiv:hep-ph/0403113\].
A. Quadt, Eur. J. Physc. C [**48**]{}, 835 (2006).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- '**Hannes Stuke** [^1]'
bibliography:
- 'bibliography.bib'
title: 'Complex time blow-up of the nonlinear heat equation'
---
[^1]: Free University of Berlin, Berlin, Germany; email: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Lu Lu
title: Survey of Cognitive Radio Techniques in Wireless Network
---
Wireless access networks were originally developed for different communication scenarios. Wireless networks aim to provide ubiquitous, flexible communications mainly in community areas [@5; @8; @10; @11]. In wireless network, there always exists the need for the communication services at higher data rates and quality of service (QoS) [@1; @2; @3; @4]. Thus lots of different techniques have been explored to improve modern wireless networks [@12; @13] to satisfy different requirements and combat different problems. For example, the rapid growth in the ubiquitous wireless services has imposed increasing stress on the fixed and limited radio spectrum. Thus allocating a fixed frequency band to each wireless service, which is the current frequency allocation policies, is a necessary to eliminate interference between different wireless services [@6; @8; @10]. As a result, innovative techniques that can offer new ways of exploiting the available spectrum are needed [@1; @17; @18]. Hence, cognitive radio [@1; @7] arises as a feasible solution to the aforementioned spectral resource allocation problem by introducing the opportunistic usage of the frequency bands that are not heavily occupied by licensed users. Currently there are many different methods for spectrum sensing have been proposed, such as the matched filtering approach [@15], the feature detection approach [@16] and the energy detection approach [@17]. For the matched filtering method, it can maximize the SNR inherently. However it is difficult to do detection without signal information such as pilot and frame structure. And for feature detection method which is basically performed based on cyclostationarity, it also must have information about received signal sufficiently. However, in practice, cognitive radio system can not know about primary signal¡¯s structure and information. For the energy detection method, although it doesn’t need any information about the signal to be detected, it is prone to false detections since it is only based on the signal power [@18]. When the signal is heavily fluctuated or noise uncertainty is big, it becomes difficult to discriminate between the absence and the presence of the signal. In addition, the energy detection is not optimal for detecting the correlated (colored) signals, which are often found in practice. Thus we can see that different methods have different advantages and disadvantages, one can choose to use these different methods according to the different requirements for specific problems.
[1]{}
H. Su and X. Zhang, *“Cross-Layer Based Opportunistic MAC Protocols for QoS Provisionings Over Cognitive Radio Wireless Networks,”* IEEE Journal on Selected Areas in Communications (J-SAC), Vol. 26, No. 1, pp. 118–129, January 2008. J. Tang and X. Zhang, *“Quality-of-Service Driven Power and Rate Adaptation Over Wireless Links,”* IEEE Transactions on Wireless Communications, Vol. 6, No. 8, pp. 3058–3068, August 2007. X. Zhang, J. Tang, H. H. Chen, S. Ci, and M. Guizani, *“Cross-Layer Based Modeling for Quality of Service Guarantees in Mobile Wireless Networks,”* IEEE Communications Magazine, Vol. 44, No. 1, pp. 100-106, January, 2006. J. Tang and X. Zhang, *“Cross-Layer-Model Based Adaptive Resource Allocation for Statistical QoS Guarantees in Mobile Wireless Networks,”* IEEE Transactions on Wireless Communications, Vol. 7, No. 6, pp. 2318–2328, June 2008. X. Zhang, K. G. Shin, D. Saha, and D. Kandlur, *“Scalable Flow Control for Multicast ABR Services in ATM Networks,”* IEEE/ACM Transactions on Networking, Vol. 10, No. 1, pp. 67–85, February 2002. J. Tang and X. Zhang, *“Cross-Layer Resource Allocation Over Wireless Relay Networks for Quality of Service Provisioning,”* IEEE Journal on Selected Areas in Communications (J-SAC), Vol. 25, No. 4, pp. 645–656, May 2007. H. Su and X. Zhang, *“Opportunistic MAC Protocols for Cognitive Radio Based Wireless Networks,”* IEEE Information Theory Society, the 41st Conference on Information Sciences and Systems (CISS 2007), John Hopkings University, Baltimore, MD, USA, March 14–16, 2007. X. Zhang and K. G. Shin, *“Markov-Chain Modeling for Multicast Signaling Delay Analysis,”* IEEE/ACM Transactions on Networking, Vol. 12, No. 4, pp. 667–680, August 2004. X. Zhang and K. G. Shin, *“Statistical Analysis of Feedback-Synchronization Signaling Delay for Multicast Flow Control,”* (IEEE INFOCOM ’2001 Review Scores: 5,5,5,4, and Reviews) (Received the NSF Grant Award of IEEE INFOCOM ’2001 for Recognition of High-Quality Paper), Presented at IEEE INFOCOM ’2001, Anchorage, Alaska, USA, April 22 - 26, 2001. H. Su and X. Zhang, *“CREAM-MAC: An Efficient Cognitive Radio-EnAbled Multi-Channel MAC Protocol for Wireless Networks,”* IEEE International Symposium on a World of Wireless, Mobile, and Multimedia Networks (WoWMoM 2008), Newport Beach, California, USA, June 23-27, 2008. J. Tang and X. Zhang, *“Cross-Layer Design of Dynamic Resource Allocation of Diverse QoS Guarantees for MIMO-OFDM Wireless Networks,”* IEEE International Symposium on a World of Wireless, Mobile, and Multimedia Networks (WoWMoM 2005), Taormina, Italy, June 14-16, 2005. J. Tang and X. Zhang, *“Cross-Layer Modeling for Quality of Service Guarantees Over Wireless Links,”* IEEE Transactions on Wireless Communications, Vol. 6, No. 12, pp. 4504–4512, December 2007. H. Su and X. Zhang, *“Clustering-Based Multi-Channel MAC Protocols for QoS-Provisionings Over Vehicular Ad Hoc Networks,”* IEEE Transactions on Vehicular Technology, Vol. 56, No. 6, pp. 3309–3323, November 2007. J. Tang and X. Zhang, *“Transmit Selection Diversity with Maximal-Ratio Combining for Multicarrier DS-CDMA Wireless Networks Over Nakagami-m Fading Channels,”* IEEE Journal on Selected Areas in Communications (J-SAC), Vol. 24, No. 1, pp. 104–112, January, 2006. D. Cabric, A. Tkachenko and R. W. Brodersen, *“Spectrum Sensing Measurements of Pilot, Energy, and Collaborative Detection,”* Proceedings of IEEE Military Communications Conference, pp. 1–7, October, 2006. Y. Lin and C. He *“Subsection-average cyclostationary feature detection in cognitive radio,”* Proceedings of International Conference on Neural Networks and Signal Processing, Vol. 24, No. 1, pp. 604–608, July, 2008. S. J. Shellhammer, S. Shankar, R. Tandra and J. Tomcik, *“Performance of power detector sensors of DTV signals in IEEE 802.22 WRANs,”* Proceedings of the Fist International Workshop on Technology and Policy for Accessing Spectrum (TAPAS), August, 2006. A. Sonnenschein and P. M. Fishman, *“Radiometric detection of spreadspectrum signals in noise of uncertainty power,”* IEEE Transactions on Aerospace and Electronic Systems, Vol. 28, No. 3, pp. 654–660, July, 1992.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we consider a quantum version of Pascal’s triangle. Pascal’s triangle is a well-known triangular array of numbers and when these numbers are plotted modulo 2, a fractal known as the Sierpinski triangle appears. We first prove the appearance of more general fractals when Pascal’s triangle is considered modulo prime powers. The numbers in Pascal’s triangle can be obtained by scaling the probabilities of the simple symmetric random walk on the line. In this paper we consider a quantum version of Pascal’s triangle by replacing the random walk by the quantum walk known as the Hadamard walk. We show that when the amplitudes of the Hadamard walk are scaled to become integers and plotted modulo three, a fractal known as the Sierpinski carpet emerges and we provide a proof of this using Lucas’s theorem. We furthermore give a general class of quantum walks for which this phenomenon occurs.'
author:
- 'Tom Bannink[^1]\'
- 'Harry Buhrman[^2]\'
bibliography:
- 'references.bib'
date: August 2017
title: 'Quantum Pascal’s Triangle and Sierpinski’s carpet'
---
Introduction
============
Pascals’s triangle, shown in Figure \[fig:pascal1\], exhibits many interesting properties one of which is the appearance of a fractal when the numbers are considered modulo a prime $p$ [@Wolfram1984; @Stewart1995]. This is shown in Figure \[fig:pascal3\], and for $p=2$ the fractal is known as the Sierpinski triangle or Sierpinski gasket. One way of obtaining the numbers in Pascal’s triangle is through a random walk on a line as will be explained in Section \[sec:pascal\]. This paper explores the results of considering a similar triangle of numbers that is obtained when the 1-dimensional random walk is replaced by a 1-dimensional quantum walk. This also yields the Sierpinski triangle when the probabilities associated to the quantum walk are considered modulo 2, but more interestingly one can find another fractal known as the Sierpinski carpet hidden in the amplitudes modulo 3 which is not present in Pascal’s triangle. When these quantum walk numbers are plotted modulo $p$, more general fractals appear.
This paper starts with Pascal’s triangle and shows how it is related to the Sierpinski triangle when the numbers are taken modulo a prime. We then provide a proof of the appearance of a more general version of the Sierpinski triangle when instead we take prime *powers*. Then, quantum walks are introduced with an emphasis on a walk that is commonly known as the Hadamard walk. We derive an expression for the probabilities of these walks and then the appearance of both the Sierpinski triangle and Sierpinski carpet is shown as well as some other properties. Table \[tab:summary\] provides a summarising overview of the different fractals that are obtained from these different sources.
Pascal’s triangle {#sec:pascal}
=================
Pascal’s triangle is a set of integers arranged in a triangle, where the $k$’th value in the $n$’th row (both $n$ and $k$ start at zero) is given by the binomial coefficient $\binom{n}{k}$. It is shown in Figure \[fig:pascal1\], and can also be constructed by using $\binom{n}{k}=\binom{n-1}{k-1}+\binom{n-1}{k}$, i.e. every number is the sum of its two neighbours in the row above. Alternatively it can be thought of as ‘scaled’ probabilities of a random walk on $\mathbb{Z}$, in the following way. Assume the random walk starts in the origin, and goes left or right with probability $\frac{1}{2}$. The probability of being at location $l\in\mathbb{Z}$ after $n$ steps, with $-n\leq l \leq n$ is given by $\frac{1}{2^n}\binom{n}{(n+l)/2}$ if $n+l$ is even and 0 if $n+l$ is odd. When considering only the $n+1$ non-zero probabilities after $n$ steps, the $k$’th non-zero value corresponds to position $l=-n+2k$ of the line, where $0\leq k\leq n$. The $k$’th non-zero probability is given by $\frac{1}{2^n}\binom{n}{k}$. Removing the factor $\frac{1}{2^n}$ yields the integer numbers in Pascal’s triangle.
;
at ( 0\*,1.5\*) [(a)]{};
at ( 0\*, 0\*) [$1$]{}; at (-1\*, -1\*) [$1$]{}; at (+1\*, -1\*) [$1$]{};
at (-2\*, -2\*) [$1$]{}; at ( 0\*, -2\*) [$2$]{}; at (+2\*, -2\*) [$1$]{};
at (-3\*, -3\*) [$1$]{}; at (-1\*, -3\*) [$3$]{}; at (+1\*, -3\*) [$3$]{}; at (+3\*, -3\*) [$1$]{};
at (-4\*, -4\*) [$1$]{}; at (-2\*, -4\*) [$4$]{}; at ( 0\*, -4\*) [$6$]{}; at (+2\*, -4\*) [$4$]{}; at (+4\*, -4\*) [$1$]{};
;
at ( 0\*,1.5\*) [(b)]{};
at ( 0\*, 0\*) [$1$]{}; at (-1\*, -1\*) [$\frac{1}{2}$]{}; at (+1\*, -1\*) [$\frac{1}{2}$]{};
at (-2\*, -2\*) [$\frac{1}{4}$]{}; at ( 0\*, -2\*) [$\frac{2}{4}$]{}; at (+2\*, -2\*) [$\frac{1}{4}$]{};
at (-3\*, -3\*) [$\frac{1}{8}$]{}; at (-1\*, -3\*) [$\frac{3}{8}$]{}; at (+1\*, -3\*) [$\frac{3}{8}$]{}; at (+3\*, -3\*) [$\frac{1}{8}$]{};
at (-4\*, -4\*) [$\frac{1}{16}$]{}; at (-2\*, -4\*) [$\frac{4}{16}$]{}; at ( 0\*, -4\*) [$\frac{6}{16}$]{}; at (+2\*, -4\*) [$\frac{4}{16}$]{}; at (+4\*, -4\*) [$\frac{1}{16}$]{};
in [0,...,4]{} [ in [-4,...,4]{} [ ]{} ]{} at ( 0\*, -1\*) [$0$]{}; at (-1\*, -2\*) [$0$]{}; at (+1\*, -2\*) [$0$]{}; at (-2\*, -3\*) [$0$]{}; at ( 0\*, -3\*) [$0$]{}; at (+2\*, -3\*) [$0$]{}; at (-3\*, -4\*) [$0$]{}; at (-1\*, -4\*) [$0$]{}; at ( 0\*, -4\*) [$0$]{}; at (+1\*, -4\*) [$0$]{}; at (+3\*, -4\*) [$0$]{};
Pascal’s triangle modulo two
----------------------------
Pascal’s triangle has many interesting properties and one interesting feature comes from considering all numbers modulo two [@Wolfram1984]. This ‘binary’ triangle is shown in Figure \[fig:pascal2\] where black and white pixels are used to represent the values modulo two. The figure that appears looks very much like the Sierpinski triangle (also known as the Sierpinski gasket). Indeed, in the limit of an infinite number of rows, Pascal’s triangle modulo two is the Sierpinski gasket.
![\[fig:triangle\_construction\] First few iterations of constructing the Sierpinski triangle. In each iteration the previous shape is shrunk to half its size and three copies are put in the corners of a triangle so that the shapes are touching.](sierpinski_triangle_evolution_square.png){width="80.00000%"}
To prove that Pascal’s triangle modulo two converges to the Sierpinski triangle, a definition of the Sierpinski triangle is needed. There are different ways to construct it, and one of them is by shrinking and duplication [@Barnsley2003]. This process is shown in Figure \[fig:triangle\_construction\]. Start with any shape (a closed bounded region) in the plane, like shown in the first image. Shrink the shape to half its size (both height and width) and make three copies of it. Place these copies in the corners of an equilateral triangle such that the shapes touch as shown in the second image. Repeat this with the new shape. The rightmost image shows the result after four iterations, and after an infinite number of iterations one obtains the fractal.
![\[fig:pascal\_row\_col\] The first $2^{n+1}$ rows of Pascal’s triangle modulo two ($P_{n+1}$) contain three copies of the first $2^n$ rows of the triangle ($P_n$). A point at row $r$ and column $c$ in the original triangle is copied to row $r+2^n$ and columns $c$ and $c+2^n$.](pascal_row_col_tikz.pdf)
To show that Pascal’s triangle modulo two converges to the fractal, one can show that for all $n$, the shape given by the first $2^{n+1}$ rows contains exactly three copies of the first $2^n$ rows, and nothing more than those copies, i.e. with ‘white’ in-between, see Figure \[fig:pascal\_row\_col\]. If we index Pascal’s triangle by row $r$ and column $c$, with $r\geq 0$ and $0\leq c \leq r$ then we need to show that for $0\leq r < 2^n$ and $0\leq c \leq r$, $$\begin{aligned}
\label{eq:pascal1}
\binom{r}{c} \equiv \binom{r+2^n}{c} \equiv \binom{r+2^n}{c+2^n} \mod 2.\end{aligned}$$ Here $\binom{r+2^n}{c}$ corresponds to a ‘pixel’ in the copy at the bottom left, and $\binom{r+2^n}{c+2^n}$ corresponds to the copy at the bottom right, as shown in Figure \[fig:pascal\_row\_col\]. Furthermore, we need to show that there is nothing more than these copies, i.e. the triangular area in-between the copies is empty. This means that the values in the corresponding region are even, and it is sufficient to show that the top row of the region is even with odd endpoints. The top row of this region is at $r=2^n$, and the endpoints of the row, $c=0$ and $c=2^n$ are odd since $\binom{r}{0}=\binom{r}{r}=1$. When the remainder of this row is even, i.e. for $1\leq c \leq 2^n - 1$, $$\begin{aligned}
\label{eq:pascal2}
\binom{2^n}{c} &\equiv 0 \mod 2,\end{aligned}$$ then the complete region will be even. This is because a value in Pascal’s triangle is the sum of its two neighbours in the row above. Adding two even numbers results in an even number, so the next row ($r=2^{n}+1$) will also have even values except near the boundary where an even number is added to an odd number. In particular, at every row the size of the ‘even part’ in the middle decreases by one, resulting in an empty up-side-down triangle ending at $r=2^n + 2^n - 1$ where the two endpoints have joined. By proving that the top row is even it follows that all entries in the up-side-down triangle are even. We conclude that it is sufficient to prove (\[eq:pascal1\]) and (\[eq:pascal2\]). For this, one can use Lucas’s theorem.
It is convenient to introduce notation for representing a number by it’s base-$p$ digits for a prime $p$. We will write $$\begin{aligned}
n = [n_m n_{m-1}\cdots n_0]_p = \sum_{j=0}^{m} n_j \; p^j \qquad \text{with } 0 \leq n_i < p \text{ for each } i,\end{aligned}$$ where the $n_i$ are the base-$p$ digits of $n$
Let $p$ be prime and $n,k$ non-negative integers. Let $n=[n_m n_{m-1}\cdots n_0]_p$ and $k=[k_m k_{m-1}\cdots k_0]_p$. Then $$\begin{aligned}
\label{eq:lucas}
\binom{n}{k} \equiv \binom{n_m}{k_m}\binom{n_{m-1}}{k_{m-1}} \cdots \binom{n_0}{k_0} \mod p,\end{aligned}$$ where we define $\binom{n}{k}=0$ if $k>n$.
There are many extensions and generalisations of Lucas’s theorem [@Mestrovic2014], that include versions for prime powers or similar congruences for generalised binomial coefficients, but they are not needed here.
\[cor:zero\] For any prime $p$, $$\begin{aligned}
\binom{n}{k} \equiv 0 \mod p \quad \iff \quad \exists i \; : \; k_i > n_i\end{aligned}$$
If there is an $i$ such that $k_i > n_i$ then $\binom{n_i}{k_i}=0$ and by Lucas’s theorem, $\binom{n}{k}\equiv 0 \mod p$. Conversely, if $k_i \leq n_i$ for all $i$ then $\binom{n_i}{k_i}=\frac{n_i!}{k_i!(n_i-k_i)!}$. Since $n_i <p$ and $p$ is prime, we have that $p$ is not a factor of $n_i!$ and also not a factor of $\frac{n_i!}{k_i!(n_i-k_i)!}$. Since $p$ is not a divisor of $\binom{n_i}{k_i}$ and $p$ is prime, it also does not divide the product $\prod_i \binom{n_i}{k_i}$ which concludes the proof.
The following corollary considers adding extra digits to $n$ and $k$ and is also known as Anton’s Lemma:
\[cor:addpower\] If $n,k < p^{m}$ and then for all $l,q\geq 0$, $$\begin{aligned}
\binom{l\cdot p^m + n}{q\cdot p^m + k} \equiv \binom{l}{q}\binom{n}{k} \mod p. \label{eq:trick2}\end{aligned}$$
When $n,k<p^m$ then $n_i=k_i=0$ for $i\geq m$, so that $l\cdot p^m + n = [l_M l_{M-1}\cdots l_0 n_{m-1} n_{m-2}\cdots n_0]_p$ and $q\cdot p^m + k = [q_M q_{M-1}\cdots q_0 k_{m-1} k_{m-2}\cdots k_0]_p$. Therefore, by Lucas’s theorem $$\begin{aligned}
\binom{l\cdot p^{m}+n}{q\cdot p^m + k} &\equiv \binom{l_M}{q_M}\binom{l_{M-1}}{q_{M-1}}\cdots \binom{l_0}{q_0}\binom{n_{m-1}}{k_{m-1}} \cdots \binom{n_0}{k_0} \mod p\\
&\equiv \binom{l}{q}\binom{n}{k} \mod p.\end{aligned}$$
Note that vice versa, Corollary \[cor:addpower\] implies Lucas’s theorem by induction on the number of digits.
We can now prove (\[eq:pascal1\]) and (\[eq:pascal2\]). By Corollary \[cor:addpower\] for $p=2$ we have $$\begin{aligned}
\binom{r+2^n}{c} \equiv \binom{1}{0}\binom{r}{c} \mod 2,\\
\binom{r+2^n}{c+2^n} \equiv \binom{1}{1}\binom{r}{c} \mod 2,\end{aligned}$$ so (\[eq:pascal1\]) follows from $\binom{1}{0}=\binom{1}{1}=1$. To show (\[eq:pascal2\]), note that since $1\leq c \leq 2^n -1$ there is a digit $c_i$ that is nonzero for $i<n$, whereas all digits of $r=2^n$ are zero except for the $r_{n}$, so (\[eq:pascal2\]) follows form Corollary \[cor:zero\].
Pascals triangle modulo general $n$
-----------------------------------
In a similar fashion one can consider Pascal’s triangle modulo general $n$. Figure \[fig:pascal3\] shows this for $n\in\{2,3,4,5,6,7\}$. We can distinguish cases for primes, prime powers and other numbers.
### Pascals triangle modulo a prime
When $n$ is a prime (2,3,5,7 in the figure), one obtains a generalisation of Sierpinski’s triangle. This generalisation for primes $p$ can also be constructed using the shrinking and duplication method. When using the shrinking and duplication construction, one can start with an arbitrary shape, shrink it and create $p(p+1)/2$ copies. These copies have to be arranged into a larger triangle where all the copies are touching. The proof is a generalisation of the one given in the previous section. One has to show that $$\begin{aligned}
\binom{r}{c} \equiv \binom{r+l\cdot p^n}{c+q\cdot p^n} \mod p,\end{aligned}$$ where $0\leq l < p$ and $0\leq q \leq l$. Each value of $(l,q)$ corresponds to one of the $p(p+1)/2$ copies. This equivalence follows from Corollary \[cor:addpower\] and $\binom{l}{q}\not\equiv 0 \mod p$. The empty up-side-down triangles correspond to $\binom{r+l\cdot p^m}{c+q\cdot p^m}$ but where the range of $c$ is now $r<c<p^m$ as opposed to $0\leq c \leq r$. This case is also included in Corollary \[cor:addpower\], and as $\binom{r}{c}=0$ for $r<c$ this finishes the proof.
### Pascals triangle modulo a composite number
When $n$ is composite (mod 6 in the figure) then the resulting shape is the union of the shapes obtained of its factors (prime powers) albeit with different colours. For example, at $n=6$, shown in Figure \[fig:pascal3\] one can see the union of the shapes of $p=2$ and $p=3$. This is simply because when $n=p_1^{k_1}\cdots p_m^{k_m}$ then $x\equiv 0 \mod n$ if and only if for all $i$ : $x \equiv 0 \mod p_i^{k_i}$.
### Pascals triangle modulo a prime power
When $n$ is a prime power then the pattern becomes slightly more complicated. One can see in Figure \[fig:pascal3\] that for $n=4$, the image is the same as for $n=2$ but with extra triangles in the places that used to be empty. When one would consider $n=8$, this idea is repeated and the holes in the $n=4$ shape are filled with additional triangles.
Figure \[fig:pascal3powers\] shows what happens when the triangle is plotted modulo powers of $3$. From the Figure we can conjecture the general pattern: Start with the ‘mod $p$ triangle’ and in every empty region, add $p(p-1)/2$ copies of the ‘mod $p$ triangle’. This yields the pattern for the ‘mod $p^2$ triangle’. To go to the ‘mod $p^3$ triangle’, again add $p(p-1)/2$ copies of the ‘mod $p$ triangle’ to the empty regions of the ‘mod $p^2$ triangle’. This process can be iterated to find the shape corresponding to $p^k$ for any $k$. To make this statement more concrete, we need the following definition.
The $p$-adic valuation $\nu_p(n)$ of a number $n$ is the largest power of $p$ that divides $n$.
The statements that we want to prove are most easily explained with a picture, shown in Figure \[fig:valuations\]. We will show that at each recursion level of the triangle, the $p$-adic valuation of the numbers $\binom{r'}{c'}$ in a copy (meaning $r'=l p^n + r$ and $c' = q p^n + c$) is the same as that of the corresponding number $\binom{r}{c}$ in the original region. Furthermore, we show that the regions that were empty in the ‘mod $p$ triangle’ have particular $p$-adic valuations that are 1 or 2 or $k$ higher than the original, as depicten in Figure \[fig:valuations\]. Every time you fill a previously-empty region with triangles, the $p$-adic valuation increases by one.
We will now prove this by using Kummer’s theorem.
\
Let $p$ be prime and $n,k$ non-negative integers, $n\geq k$. Then the $p$-adic valuation $\nu_p(\binom{n}{k})$ of $\binom{n}{k}$ is equal to the number of “carries” when $k$ and $n-k$ are added in base-$p$ arithmetic.
One way to find the number of carries that occur when $k$ is added to $n-k$ in base-$p$ is by considering the base-$p$ digits of $n$ and $k$, defining $c^{n,k}_{-1}=0$ and $$\begin{aligned}
c^{n,k}_{i} = \begin{cases} 1 & n_i < k_i\\ 0 & n_i > k_i\\ c^{n,k}_{i-1} & n_i = k_i \end{cases}.\end{aligned}$$ The number of carries is then equal to $\sum_{i\geq 0} c^{n,k}_{i}$. Kummer’s theorem can therefore be written as $\nu_p(\binom{n}{k}) = \sum_{i\geq 0} c^{n,k}_{i}$.
The following claim shows what happens to $\nu_p(\binom{n}{k})$ when a digit $l$ is added to $n$ and a digit $q$ is added to $k$:
\[claim:valuationdigits\] Let $p$ be prime and $n,k,q,l,m$ non-negative integers with $0\leq k \leq n < p^m$ and $0 \leq q \leq l < p$. Then $$\begin{aligned}
\nu_p(\binom{l\cdot p^m + n}{q\cdot p^m + k}) = \nu_p(\binom{n}{k})
\end{aligned}$$
Define $n'=l\cdot p^m +n$ and $k'=q\cdot p^m + k$. Note that $n,k$ have at most $m-1$ digits when expressed in base $p$ and $l,q$ are the $m$-th digits of $n'$ and $k'$. Note that we have $c^{n,k}_{i}=c^{n',k'}_{i}$ for $i<m$ since the first $m-1$ digits are the same. For the $m$-th digits we have $$\begin{aligned}
c^{n',k'}_{m}=\begin{cases} 1 & l < q\\ 0 & l > q\\ c^{n,k}_{m-1} & l = q \end{cases} .
\end{aligned}$$ By Kummer’s theorem the difference between $\nu_p(\binom{n'}{k'})$ and $\nu_p(\binom{n}{k})$ is equal to $c^{n',k'}_{m}$, so it remains to show that $c^{n',k'}_{m}=0$. By assumption we know $q\leq l$ and if $q<l$ we have $c^{n',k'}_{m} = 0$ by definition. Consider the case $l=q$ where we have $c^{n',k'}_{m} = c^{n,k}_{m-1}$. If $n=k$ then all the $c^{n,k}_{i}$ are zero so we are done. If $n \neq k$ then the consider the most significant digit where $n$ and $k$ differ, i.e. take the highest $i$ for which $n_i \neq k_i$ and call it $i^*$. Since $k<n$ by assumption, it must be true that $k_{i^*} < n_{i^*}$ and therefore $c^{n,k}_{i^*} = 0$. For all $i>i_*$ we have $n_i=k_i$ so $c^{n,k}_{i}=c^{n,k}_{i-1}$. So $c^{n',k'}_{m} = c^{n,k}_{i_*} = 0$.
\[claim:valuationmod\] If $\nu_p(n) = \nu_p(m)$ then for any $k$ $$n\equiv 0 \mod p^k \iff m\equiv 0 \mod p^k.$$
It follows from the fact that $n\equiv 0 \mod p^k$ if and only if $\nu_p(n) \geq k$.
Consider Figure \[fig:pascal3powers\]. The size of the recursion levels in the ‘mod $p^k$ triangle’ is the same as in the mod $p$ triangle, meaning powers of $p$ and not powers of $p^k$. Repeating what we did before for the mod $p$ triangle, we can see that at recursion level $n$, the “copies” and “empty regions” correspond to the following binomial coefficients of Pascal’s triangle: $$\begin{aligned}
\binom{l\cdot p^n+r}{q\cdot p^n + c} & &
\begin{array}{rl}
\text{``copies''} \to & 0 \leq q \leq l < p \; , \; 0 \leq c \leq r < p^n \\
\text{``empty''} \to & 0 \leq q < l < p \; , \; 0 \leq r < c < p^n
\end{array}
.\end{aligned}$$ Here $n$ is the recursion level and $(l,q)$ index the different copies or empty regions whereas $r$ and $c$ index points within those regions. By claim \[claim:valuationdigits\], the $p$-adic valuation of a number $\binom{r'}{c'}$ in the copy is the same as that of the corresponding original number $\binom{r}{c}$. By Claim \[claim:valuationmod\], we have now shown that for any $k$, the copies in the ‘mod $p^k$ triangle’ are indeed all the same.
What is left to show is how the empty regions of the ‘mod $p$ triangle’ are filled. We will first consider $k=2$. The *new* triangles in the mod $p^2$ shape can be indexed as follows: $$\begin{aligned}
\binom{l\cdot p^n + s\cdot p^{n-1} + r}{q\cdot p^n + t \cdot p^{n-1} + c} & \text{ with }
\begin{array}{l}
0 \leq q < l < p \; ,\\
0 \leq s < t < p \; ,\\
0 \leq c \leq r < p^{n-1}.
\end{array} \end{aligned}$$ These lie within the *empty* regions of the mod $p$ triangle. Similar to the proof of Claim \[claim:valuationdigits\] we can let the carries $c^{r',c'}_{i}$ be defined as before, for the numbers $r'=[l\;s\;r_{n-2}r_{n-3}...r_0]_p$ and $c'=[q\;t\;c_{n-2}...c_0]_p$. We have $q<l$ hence $c^{r',c'}_{n}=0$ and $s<t$ so $c^{r',c'}_{n-1}=1$. This means there is exactly one extra carry compared to $r$ and $c$ and hence by Kummer’s theorem $$\begin{aligned}
\nu_p(\binom{l\cdot p^n + s\cdot p^{n-1} + r}{q\cdot p^n + t \cdot p^{n-1} + c}) = \nu_p(\binom{r}{c}) + 1\end{aligned}$$ for these values of $l,q,s,t,r,c$. This is what we need, because it implies $$\begin{aligned}
\binom{l\cdot p^n + s\cdot p^{n-1} + r}{q\cdot p^n + t \cdot p^{n-1} + c} \equiv 0 \mod p^{k+1} \iff \binom{r}{c} \equiv 0 \mod p^{k}\end{aligned}$$ i.e. in the mod $p^{k+1}$ shape there is a copy of the (smaller) mod $p^k$ shape. This proves what was drawn as a size-$p^n$ triangle in Figure \[fig:valuations\]. If we continue to the triangles that are newly added in the ‘mod $3^3$ triangle’, we find that they correspond to $$\begin{aligned}
&\binom{r'_n\cdot p^n + r'_{n-1}\cdot p^{n-1} + r'_{n-2}\cdot p^{n-2} + r}{c'_{n}\cdot p^n + c'_{n-1}\cdot p^{n-1} + c'_{n-2}\cdot p^{n-2} + c} \\
&\qquad\qquad \text{ with }
\begin{array}{l}
0 \leq c'_n < r'_n < p \; , \\
0 \leq r'_{n-1} \leq c'_{n-1} < p \; , \\
0 \leq r'_{n-2} < c'_{n-2} < p \; , \\
0 \leq c \leq r < p^{n-2}.
\end{array} \end{aligned}$$ Define $r'=[r'_n r'_{n-1} r'_{n-2} r_{n-3} \cdots r_0]_p$ and $c'=[c'_n c'_{n-1} c'_{n-2} c_{n-3} \cdots c_0]_p$. Then by the same reasoning as before we can apply Kummer’s theorem to obtain $\nu_p(\binom{r'}{c'}) = \nu_p(\binom{r}{c})+ c^{r',c'}_{n-2} + c^{r',c'}_{n-1} + c^{r',c'}_{n}$. Looking at the constraints for digits $n-2$ up to $n$ we see that $c^{r',c'}_{n} = 0$, and $c^{r',c'}_{n-1}$ is $1$ or equal to $c^{r',c'}_{n-2}$ which is always $1$. We conclude: $\nu_3(\binom{r'}{c'}) = \nu_3(\binom{r}{c})+2$. We can continue the pattern, and we find that in the ‘mod $p^{k+1}$ triangle’, the newly added triangles correspond to the following constraints on the digits of $r',c'$ with the following carries: $$\begin{aligned}
0 \leq c'_n &< r'_n < p & c^{r',c'}_{n} = 0 \\
0 \leq r'_{n-1} &\leq c'_{n-1} < p & c^{r',c'}_{n-1} = 1 \text{ or } c^{r',c'}_{n-1} = c^{r',c'}_{n-2} \\
&\vdots \\
0 \leq r'_{n-k+1} &\leq c'_{n-k+1} < p & c^{r',c'}_{n-k+1} = 1 \text{ or } c^{r',c'}_{n-k+1} = c^{r',c'}_{n-k} \\
0 \leq r'_{n-k} &< c'_{n-k} < p & c^{r',c'}_{n-k} = 1 \\
0 \leq c &\leq r < p^{n-k} & \nu_3\binom{r}{c} .\end{aligned}$$ We see that $\nu_3\binom{r'}{c'} = \nu_3\binom{r}{c} + k$ as required. We still have to show that the empty regions in the ‘mod $p^{k+1}$ triangle’ are empty. They correspond to the same indices as above except for $0\leq r'_{n-k} \leq c'_{n-k} < p$ and $0 \leq r < c < p^{n-k}$. We can apply the same idea as in the proof of Claim \[claim:valuationdigits\] by noting that the first digit where $r$ and $c$ differ will satisfy $r_{i^*} < c_{i^*}$ and hence all the carries $c^{r',c'}_i$ are $1$ for $i \geq i^*$. This gives $\nu_3 \binom{r'}{c'} \geq k+1$, meaning that $\binom{r'}{c'}\equiv 0 \mod p^{k+1}$ so the region is indeed empty.
Since the numbers in Pascal’s triangle can be thought of as scaled probabilities of a random walk, one could imagine writing down probabilities of a quantum walk, scaled to become integer, and show them modulo two. The next section will introduce a specific quantum walk and apply this idea with $p=2$ and $p=3$.
Hadamard Walk
=============
Quantum walks are simple models for a quantum particle moving through some system. This paper is only concerned with the probability distribution that emerges from one particular quantum walk, and therefore the physical aspects of it are left out. Here we only provide a short overview of the relevant concepts, and we refer the reader to [@Nielsen] for a complete introduction to the field of quantum information. For the purposes of this paper we only need to know that the state of a particle is described by a unit vector in a complex Hilbert space, and quantum mechanics dictates that time evolution is limited to applying unitary operators to this vector. We will denote such state vectors using the commonly used ‘bra-ket’ notation, writing ${\vert\psi\rangle}$ for a vector as opposed to $\vec{\psi}$. We can write the vector as a linear combination of orthonormal basis states, ${\vert\psi\rangle}=\sum_i \alpha_i {\vertx_i\rangle}$ where $\alpha_i\in\mathbb{C}$ and $\sum_i |\alpha_i|^2 = 1$. The ${\vertx_i\rangle}$ are the standard basis vectors of the Hilbert space and the coefficients $\alpha_i$ are known as *amplitudes*. Inner products are denoted by $\langle\phi\vert\psi\rangle$ for two vectors ${\vert\phi\rangle}$ and ${\vert\psi\rangle}$. One of the axioms of quantum mechanics states that one can observe (measure) the system ${\vert\psi\rangle}$ in a chosen basis and the result can be any of the basis states, where state ${\vertx_i\rangle}$ has probability $|\alpha_i|^2$ of appearing.
A simple example of a quantum walk on a one-dimensional line is the so-called Hadamard walk [@Ambainis01]. It can be thought of as a quantum particle moving on $\mathbb{Z}$, the discrete line. The particle has an internal degree of freedom other than its position (a spin-$\frac{1}{2}$ degree of freedom, for physicists). The internal state is sometimes referred to as the *coin state* of the particle with associated Hilbert space $\mathcal{H}_\mathrm{coin}=\mathbb{C}^2$ and basis states ${|\!\uparrow\rangle}$ and ${|\!\downarrow\rangle}$. The Hilbert space associated with the complete quantum system is $\mathcal{H}=\mathcal{H}_\mathrm{pos}\otimes\mathcal{H}_\mathrm{coin}$ where $\mathcal{H}_\mathrm{pos}=\mathrm{span}\{ {\vertn\rangle} \;|\; n\in\mathbb{Z} \}$. So the most general state of the particle is $${\vert\psi\rangle} = \sum_{n\in\mathbb{Z}} \left(\alpha_{n,\uparrow}{\vertn,\uparrow\rangle}+\alpha_{n,\downarrow}{\vertn,\downarrow\rangle}\right) ,$$ with normalization $\sum_{n\in\mathbb{Z}}\left(|\alpha_{n,\uparrow}|^2+|\alpha_{n,\downarrow}|^2\right) = 1$ and where we used the notation ${\verta,b\rangle}\equiv {\verta\rangle}\otimes{\vertb\rangle}$. The dynamics of the particle are given by repeated application of a unitary operator $U$ that consists of two steps. The first step is a unitary only applied to the internal state and is sometimes considered the quantum analogue of ‘flipping a coin’. The second step updates the position of the particle conditioned on the outcome of the coin. In the specific case of the Hadamard walk, the unitary in the first step is the Hadamard operator $H$, defined as $$\begin{aligned}
H = \frac{1}{\sqrt{2}}\begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix}, \text{ where } {|\!\uparrow\rangle}= \begin{pmatrix}1\\ 0\end{pmatrix} \text{ and } {|\!\downarrow\rangle}= \begin{pmatrix}0\\ 1\end{pmatrix}.\end{aligned}$$ The time evolution operator $U$ is then given by $$\begin{aligned}
U = S\cdot(\mathrm{Id}_\mathrm{pos} \otimes H),\end{aligned}$$ where $\mathrm{Id}_\mathrm{pos}$ is the identity on $\mathcal{H}_\mathrm{pos}$ and $S$ is called the *shift*, given by $$\begin{aligned}
S{\vertn,\uparrow\rangle} = {\vertn+1,\uparrow\rangle},\qquad S{\vertn,\downarrow\rangle}={\vertn-1,\downarrow\rangle}.\end{aligned}$$ It can be thought of as updating the position of the particle conditioned on the outcome of the coin flip. A full step $U$ of the Hadamard walk, acting on the basis state ${\vertn,\downarrow\rangle}$ for example, is given by $$U{\vertn,\downarrow\rangle} \overset{\mathrm{coin}}{=} S\;\left(\frac{1}{\sqrt{2}}{\vertn,\uparrow\rangle} - \frac{1}{\sqrt{2}}{\vertn,\downarrow\rangle}\right) \overset{\mathrm{shift}}{=} \frac{1}{\sqrt{2}} {\vert{n+1},\uparrow\rangle} - \frac{1}{\sqrt{2}} {\vertn-1,\downarrow\rangle} .$$ If we now were to measure the system, the result would be either ${\vertn+1,\uparrow\rangle}$ or ${\vertn-1,\downarrow\rangle}$, both with probability $|{\pm1}/\sqrt{2}|^2=1/2$.
in [0]{} [ (, 1) – node\[above\] [$\frac{ 1}{\sqrt{2}}$]{} (+1, 1); (, 1) – node\[left\] [$\frac{ 1}{\sqrt{2}}$]{} (-1, 0); (, 0) – node\[right\] [$\frac{ 1}{\sqrt{2}}$]{} (+1, 1); (, 0) – node\[below\] [$\frac{-1}{\sqrt{2}}$]{} (-1, 0); ]{} in [-3,...,3]{} [ (, 0) circle (0.03); (, 1) circle (0.03); at (,2.3) [$|{\x}\rangle_\mathrm{p}$]{}; ]{} at (4.0,1) [${|\!\uparrow\rangle}_\mathrm{c}$]{}; at (4.0,0) [${|\!\downarrow\rangle}_\mathrm{c}$]{};
Figure \[fig:diagram1\] shows a schematic representation of one step $U$ of the Hadamard walk.
With these definitions, one can now consider the following process. Select a starting state, say ${\vert\psi_s\rangle}={\vert0,\uparrow\rangle}$, evolve it with $U$ for $t$ steps and then measure the position. For example, starting in ${\vert0,\uparrow\rangle}$, the state of the system after three steps is given by $$\begin{aligned}
U^3 {\vert0,\uparrow\rangle} = \frac{1}{(\sqrt{2})^3}\Big(
{\vert{-3},\downarrow\rangle}
- {\vert{-1},\uparrow\rangle}
+ 2 {\vert1,\uparrow\rangle} + {\vert1,\downarrow\rangle}
+ {\vert3,\uparrow\rangle}
\Big).\end{aligned}$$ Now measuring the system will result in finding some position $X$ with probabilities $\mathbb{P}[X{=}{-3}]=\frac{1}{8}$, $\mathbb{P}[X{=}{-1}]=\frac{1}{8}$, $\mathbb{P}[X{=}1]=\frac{5}{8}$ and $\mathbb{P}[X{=}3]=\frac{1}{8}$. The amplitudes of the first five steps are also displayed in Figure \[fig:qpascal2\].
Hadamard triangle
-----------------
The numbers in Pascal’s triangle can be thought of as scaled probabilities of a random walk, and carrying this idea over to the Hadamard walk, one could consider the amplitudes or probabilities of the Hadamard walk, but scaled by a factor of $\sqrt{2^n}$ so that all numbers involved become integer. Another way to view this is instead of applying $H$, use $\sqrt{2}H$, a matrix with only integer coefficients. Note that we could either use the *amplitudes* or the *probabilities* which are simply their squares. However, since we are primarily interested in whether or not they are divisible by some prime $p$, squaring the amplitudes does not make a difference. We therefore continue with the (unsquared) amplitudes. Figure \[fig:qpascal2\] shows the start of the Hadamard triangle.
; ;
at (-5\*, 0\*) [$\frac{1}{\sqrt{2^0}}$]{}; at (-5\*, -1\*) [$\frac{1}{\sqrt{2^1}}$]{}; at (-5\*, -2\*) [$\frac{1}{\sqrt{2^2}}$]{}; at (-5\*, -3\*) [$\frac{1}{\sqrt{2^3}}$]{}; at (-5\*, -4\*) [$\frac{1}{\sqrt{2^4}}$]{};
at ( 0\*, 0\*) [${({ 1}\;,\;{ 0})}$]{}; at (-1\*, -1\*) [${({ 0}\;,\;{ \color{blue}{1}})}$]{}; at (+1\*, -1\*) [${({ \color{red}{1}}\;,\;{ 0})}$]{};
at (-2\*, -2\*) [${({ 0}\;,\;{\color{blue}{-1}})}$]{}; at ( 0\*, -2\*) [${({\color{red}{1}}\;,\;{\color{blue}{ 1}})}$]{}; at (+2\*, -2\*) [${({\color{red}{1}}\;,\;{ 0})}$]{};
at (-3\*, -3\*) [${({ 0}\;,\;{\color{blue}{1}})}$]{}; at (-1\*, -3\*) [${({\color{red}{-1}}\;,\;{\color{blue}{0}})}$]{}; at (+1\*, -3\*) [${({\color{red}{ 2}}\;,\;{\color{blue}{1}})}$]{}; at (+3\*, -3\*) [${({\color{red}{ 1}}\;,\;{ 0})}$]{};
at (-4\*, -4\*) [${({ 0}\;,\;{\color{blue}{-1}})}$]{}; at (-2\*, -4\*) [${({\color{red}{ 1}}\;,\;{\color{blue}{-1}})}$]{}; at ( 0\*, -4\*) [${({\color{red}{-1}}\;,\;{\color{blue}{ 1}})}$]{}; at (+2\*, -4\*) [${({\color{red}{ 3}}\;,\;{\color{blue}{ 1}})}$]{}; at (+4\*, -4\*) [${({\color{red}{ 1}}\;,\;{ 0})}$]{};
in [0]{} [ in [0]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.20\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.20\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [1]{} [ in [-1, 1]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.20\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.20\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [2]{} [ in [-2, 0, 2]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.20\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.20\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [3]{} [ in [-3, -1, 1, 3]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.20\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.20\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{}
We will now derive expressions for these amplitudes when starting in ${\vert0,\uparrow\rangle}$.
Expressions for amplitudes
--------------------------
Meyer [@Meyer96] gave explicit expressions for the amplitudes encountered in the Hadamard walk. Let $\psi_\uparrow(n,t)$ be the amplitude at ${\vertn,\uparrow\rangle}$ after $t$ steps when starting in ${\vert0,\uparrow\rangle}$, i.e. $\psi_\uparrow(n,t):=\langle n,\uparrow\!|U^t|0,\uparrow\rangle$. Similarly, let $\psi_\downarrow(n,t):=\langle n,\downarrow\!|U^t|0,\uparrow\rangle$. Then we have
When $t+n$ is odd or when $|n|>t$ we have $\psi_\uparrow(n,t)=\psi_\downarrow(n,t)=0$. Otherwise the amplitudes are given by $$\begin{aligned}
\psi_\uparrow (n,t) &=
\begin{cases}
\frac{1}{\sqrt{2^t}} & n = t\\
\frac{1}{\sqrt{2^t}} \sum_{k\geq 1} \binom{(t-n)/2-1}{k-1}\binom{(t+n)/2}{k}(-1)^{(t-n)/2-k} & n < t
\end{cases}\\
\psi_\downarrow(n,t) &= \frac{1}{\sqrt{2^t}} \sum_{k\geq 0} \binom{(t-n)/2-1}{k}\binom{(t+n)/2}{k}(-1)^{(t-n)/2-k-1}\end{aligned}$$
We will give an alternative and slightly shorter proof of this for a general coin operator. This proof also allows us to make another observation stated in the following claim. Let $C$ be any unitary 2x2 matrix. Any such matrix can be written as follows $$\begin{aligned}
C = \begin{pmatrix}
c_r & c_u\\
c_d & c_l
\end{pmatrix}
= \begin{pmatrix}
\sqrt{p} \; e^{i \alpha} & \sqrt{1-p} \; e^{i\beta} \\
- \sqrt{1-p} \; e^{i \gamma} & \sqrt{p} \; e^{i (\gamma + \beta - \alpha)}
\end{pmatrix} \quad , \quad \text{with } 0 \leq p \leq 1.\end{aligned}$$
\[claim:amplitudeexpressions\] Let $\psi_\uparrow(n,t)$ and $\psi_\downarrow(n,t)$ be the up and down amplitudes at position $n$ at time $t$ but for the general coin operator $C$. When $t+n$ is odd or when $|n|>t$ we have $\psi_\uparrow(n,t)=\psi_\downarrow(n,t)=0$. Otherwise the amplitudes are given by $$\begin{aligned}
\psi_\uparrow (n,t) &=
\begin{cases}
e^{i\alpha n} \sqrt{p^t} & n = t\\
e^{i\left(\alpha n + (\gamma+\beta)(t-n)/2\right)} \sqrt{p^t} \sum_{k\geq 1} \binom{(t+n)/2}{k}\binom{(t-n)/2-1}{k-1} \left(-\frac{1-p}{p}\right)^k & n < t
\end{cases}\\
\psi_\downarrow(n,t) &= -e^{i(\alpha n +(\gamma+\beta)(t-n)/2 -\beta)} \\
&\qquad \qquad \times \quad \sqrt{(1-p)p^{t-1}} \sum_{k\geq 0} \binom{(t+n)/2}{k} \binom{(t-n)/2-1}{k} \left(-\frac{1-p}{p}\right)^k
\end{aligned}$$ The probabilities $|\psi_\uparrow(n,t)|^2$ and $|\psi_\downarrow(n,t)|^2$ associated to these amplitudes are independent of the complex phases $\alpha,\beta,\gamma$ of the coin operator.
Note that the lemma follows directly from the claim by setting $p=1/2$, $\alpha=\beta=0$ and $\gamma=\pi$ to obtain the Hadamard coin matrix. Furthermore note that for a more general starting state, not equal to ${\vert0,\uparrow\rangle}$, the probabilities *do* depend on the complex phases present in the coin operator.
; in [-2,-1,0,1,2]{} [ (, 1) – node\[above\] [$c_r$]{} (+1, 1); (-0.05, 1) – node\[left\] [$c_d$]{} (+-1.05, 0); (++0.05, 0) – node\[right\] [$c_u$]{} (+1.05, 1); (+, 0) – node\[below\] [$c_l$]{} (+-1, 0); ]{} in [-3,...,3]{} [ (+, 0) circle (0.06); (, 1) circle (0.06); at (,1.5) [$\x$]{}; at (+,-0.5) [$\x$]{}; ]{} at (5.0,1) [${|\!\uparrow\rangle}$]{}; at (5.0,0) [${|\!\downarrow\rangle}$]{};
We want to have an expression for both the up and down component at position $n$ after $t$ steps, i.e. we want to know $\langle n,\uparrow\!|U^t|0,\uparrow\rangle$ and $\langle n,\downarrow\!|U^t|0,\uparrow\rangle$. These two cases will be handled separately. We will use the path counting technique, where one counts all possible paths starting at ${\vert0,\uparrow\rangle}$ and end at the desired state. Every path gets a certain (possibly negative) amplitude and these have to be added.
### Up component
We want to find all possible paths from ${\vert0,\uparrow\rangle}$ to ${\vertn,\uparrow\rangle}$ using $t$ steps. If $n=t$ there is exactly one path. Otherwise, we have $-t< n < t$. Figure \[fig:diagram2\] shows the directed graph on which we consider possible paths. Consider a single path and let $r,l,u,d$ be the number of times the path uses the right, left, up and down arrows respectively. To end at ${\vertn,\uparrow\rangle}$ we then have $$\begin{aligned}
r+l+u+d&=t &\text{total number of steps}\\
r-l &=n &\text{ending column}\\
u &=d &\text{start up and end up}\end{aligned}$$ Let $k=u=d$, then we have $r=\frac{t+n}{2} - k$ and $l=\frac{t-n}{2} - k$. We have $k\geq 1$ (we need to go down and up at least once) and $k\leq\frac{t-n}{2},\frac{t+n}{2}$.\
For a specific set of values $(r,l,u,d)$ the path will arrive with an amplitude $(c_r)^r(c_l)^l(c_u)^u(c_d)^d$. So we sum over all possible values for $(r,l,u,d)$ and count how many paths there are for a specific set of values $(r,l,u,d)$. We can construct such paths as follows. Construct a sequence of choices to make if the walker is in the top layer and another sequence of choices to make if the walker is in the bottom layer. The walker is in the ${|\!\uparrow\rangle}$ state (top layer) $k+r=\frac{t+n}{2}$ times, out of which $r$ times it goes right and $k$ times it goes down. There are $\binom{(t+n)/2}{k}$ possible ways to do this. Likewise, the particle is in ${|\!\downarrow\rangle}$ (bottom layer) $l+k=\frac{t-n}{2}$ times and has to choose between left and up. The *last* of these choices should always be up, so this gives $\binom{(t-n)/2-1}{k-1}$ possibilities. To construct the full path, start with the top-layer choices, and whenever the choice is ‘down’, continue with the bottom-layer choices and so on. Therefore $$\begin{aligned}
{\langlen,\uparrow\vert}U^t{\vert0,\uparrow\rangle} &= \begin{cases} (c_r)^t & n = t\\
\sum_{k\geq 1} \binom{(t+n)/2}{k}\binom{(t-n)/2-1}{k-1} c_r^{(t+n)/2-k} c_l^{(t-n)/2-k} c_u^k c_d^k & n < t \end{cases}.
$$ Now rewrite the last sum for $n<t$ and group the factors that do not depend on $k$: $$\begin{aligned}
c_r^{(t+n)/2} c_l^{(t-n)/2} \sum_{k\geq 1} \binom{(t+n)/2}{k}\binom{(t-n)/2-1}{k-1} \left(\frac{c_u c_d}{c_r c_l}\right)^k\end{aligned}$$ Note that $\frac{c_u c_d}{c_r c_l} = -\frac{1-p}{p}$ so this fraction is always a real (negative) number, regardless of the complex phases present in the entries of the coin matrix. The sum above in terms of $p$ and $\alpha,\beta,\gamma$ is equal to $$\begin{aligned}
e^{i\left(\alpha n + (\gamma+\beta)(t-n)/2\right)} \sqrt{p^t} \sum_{k\geq 1} \binom{(t+n)/2}{k}\binom{(t-n)/2-1}{k-1} \left(-\frac{1-p}{p}\right)^k ,\end{aligned}$$ as claimed. The probability $|\psi_\uparrow(n,t)|^2$ of being at ${\vertn,\uparrow\rangle}$ after $t$ steps when starting in ${\vert0,\uparrow\rangle}$ is independent of $\alpha,\beta,\gamma$ since the only dependence on these variables is in the prefactor $e^{i\left(\alpha n + (\gamma+\beta)(t-n)/2\right)}$ which always has norm 1.
### Down component
For the down component, the equations are similar: $$\begin{aligned}
r+l+u+d&=t &\text{total number of steps}\\
r-l &=n+1 &\text{ending column (tilted)}\\
u+1 &=d &\text{start up and end down}\end{aligned}$$ Let $u=k$, then the equations give $r=(t+n)/2-k$ and $l=(t-n)/2-k-1$. The argument is the same as before, but now the last choice in the top layer has to be ‘down’ with no restrictions on the last choice in the bottom layer. We are in ${|\!\uparrow\rangle}$ $r+d=(t+n)/2+1$ times. The last choice has to be down, so this gives $\binom{(t+n)/2}{k}$. We are in ${|\!\downarrow\rangle}$ $l+u=(t-n)/2-1$ times which gives $\binom{(t-n)/2-1}{k}$. The expression is therefore given by $$\begin{aligned}
\psi_\downarrow(n,t) &= \sum_{k\geq 0} \binom{(t+n)/2}{k} \binom{(t-n)/2-1}{k} c_u^k c_d^{k+1} c_l^{(t-n)/2-k-1} c_r^{(t+n)/2-k} .\end{aligned}$$ Rewriting this in terms of $p,\alpha,\beta,\gamma$ gives the expression given in the claim. Again the only dependency on the complex phases is in the prefactor which has norm 1 so the probability $|\psi_\downarrow(n,t)|^2$ only depends on $p$.
Hadamard walk modulo 2 - Sierpinski triangle {#sec:hadamardmod2}
--------------------------------------------
When the amplitudes of the Hadamard walk are plotted modulo two, the Sierpinski triangle appears in a similar fashion to Pascal’s triangle. To see why this is the case, we note that to find the amplitudes at some time $t$ modulo two it is enough to consider a process where every single time-step is done modulo two. The scaled Hadamard operator becomes $$\begin{aligned}
\sqrt{2}H \equiv \begin{pmatrix} 1 & 1 \\ 1 & 1 \end{pmatrix} \mod 2,\end{aligned}$$ and we can immediately see that the amplitude sent to the right is the same as the amplitude sent to the left. More precisely, after any time-step the amplitude at ${\vertn-1,\downarrow\rangle}$ is the same as the amplitude at ${\vertn+1,\uparrow\rangle}$ modulo two, for all $n$.
; ;
(0,-1\*) ellipse ([0.92\*]{} and [0.5\*]{}); (-1\*,-2\*) ellipse ([0.92\*]{} and [0.5\*]{}); ( 1\*,-2\*) ellipse ([0.92\*]{} and [0.5\*]{}); (-2\*,-3\*) ellipse ([0.92\*]{} and [0.5\*]{}); ( 0\*,-3\*) ellipse ([0.92\*]{} and [0.5\*]{}); ( 2\*,-3\*) ellipse ([0.92\*]{} and [0.5\*]{}); (-3\*,-4\*) ellipse ([0.92\*]{} and [0.5\*]{}); (-1\*,-4\*) ellipse ([0.92\*]{} and [0.5\*]{}); ( 1\*,-4\*) ellipse ([0.92\*]{} and [0.5\*]{}); ( 3\*,-4\*) ellipse ([0.92\*]{} and [0.5\*]{});
at ( 0\*, 0\*) [${({ 1}\;,\;{ 0})}$]{}; at (-1\*, -1\*) [${({ 0}\;,\;{ 1})}$]{}; at (+1\*, -1\*) [${({ 1}\;,\;{ 0})}$]{};
at (-2\*, -2\*) [${({ 0}\;,\;{ 1})}$]{}; at ( 0\*, -2\*) [${({ 1}\;,\;{ 1})}$]{}; at (+2\*, -2\*) [${({ 1}\;,\;{ 0})}$]{};
at (-3\*, -3\*) [${({ 0}\;,\;{ 1})}$]{}; at (-1\*, -3\*) [${({ 1}\;,\;{ 0})}$]{}; at (+1\*, -3\*) [${({ 0}\;,\;{ 1})}$]{}; at (+3\*, -3\*) [${({ 1}\;,\;{ 0})}$]{};
at (-4\*, -4\*) [${({ 0}\;,\;{ 1})}$]{}; at (-2\*, -4\*) [${({ 1}\;,\;{ 1})}$]{}; at ( 0\*, -4\*) [${({ 1}\;,\;{ 1})}$]{}; at (+2\*, -4\*) [${({ 1}\;,\;{ 1})}$]{}; at (+4\*, -4\*) [${({ 1}\;,\;{ 0})}$]{};
in [0]{} [ in [0]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.15\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.15\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [1]{} [ in [-1, 1]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.15\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.15\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [2]{} [ in [-2, 0, 2]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.15\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.15\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{} in [3]{} [ in [-3, -1, 1, 3]{} [ (-0.15\*,--0.2) – (- 0.75\*,--0.8\*); (+0.15\*,--0.2) – (- 0.65\*,--0.8\*); (-0.15\*,--0.2) – (+ 0.65\*,--0.8\*); (+0.15\*,--0.2) – (+ 0.75\*,--0.8\*); ]{} ]{}
This idea is shown in Figure \[fig:qpascal3\] which is similar to Figure \[fig:qpascal2\] but modulo two. An ellipse is drawn around the pairs of amplitudes of states ${\vertn-1,\downarrow\rangle}$ and ${\vertn+1,\uparrow\rangle}$. The figure shows that the two values in each ellipse are equal, and are the sum of the values in the two neighbouring ellipses above it. This is the same rule with which Pascal’s triangle can be constructed. Indeed, taking one value out of every ellipse, the Sierpinski triangle can be obtained. These are the either the red or the blue values shown in Figure \[fig:qpascal2\].
Hadamard walk modulo 3 - Sierpinski carpet {#sec:hadamardmod3}
------------------------------------------
![\[fig:carpet1\]The start of the Sierpinski carpet resulting from colouring the scaled Hadamard walk amplitudes modulo 3. The horizontal direction is position and the vertical direction is time. The shape drawn at each point is a diamond, i.e. a rotated square instead of a square, because this gives a better visualisation of the $x,y$ coordinates.](hadamard_carpet1.pdf){width="90.00000%"}
We will now show that the ${|\!\downarrow\rangle}$ components of the scaled walk, modulo three, give rise to the Sierpinski carpet. In particular, we colour a square white if the amplitude is divisible by $3$, and give it a different colour otherwise. Figure \[fig:carpet1\] shows the start of the resulting carpet. The top of the carpet is at $t=1$ and $n={-1}$ indicated by the *blue* values in Figure \[fig:qpascal2\]. Considering only these values, indexed by a row $R$ and column $C$ we are interested in amplitudes at $t=R+1$ and $n=2C-R-1$. Define $H_\mathrm{blue}(R,C):=\psi_\downarrow(2C-R-1,R+1)$, then $$\begin{aligned}
H_\mathrm{blue}(R,C) &= (-1)^{R-C} \sum_{k=0}^{\min(C,R-C)} \binom{C}{k}\binom{R-C}{k} (-1)^{k}. \label{eq:Hblue}\end{aligned}$$ For the structure of the Sierpinski carpet, however, it is more convenient to consider coordinates $x,y$ that are aligned with the square structure of the carpet. The choice of these directions is indicated in Figure \[fig:carpet1\], and we have $R=x+y$ and $C=x$. As a function of these coordinates we define $\Phi(x,y) = H_\mathrm{blue}(x+y,x)$, so $$\begin{aligned}
\Phi(x,y) = (-1)^y \sum_{k=0}^{\min(x,y)} \binom{x}{k} \binom{y}{k} (-1)^{k}. \label{eq:phidef}\end{aligned}$$ A pixel at coordinates $x,y$ is now coloured white if $\Phi(x,y)\equiv 0 \mod 3$ and a different colour otherwise. To show that the resulting figure is the Sierpinski carpet we have to show that for all $n\geq 1$ and $0\leq l,q \leq 2$ with $(l,q)\neq(1,1)$, $$\begin{aligned}
\Phi(x,y) \equiv \pm \Phi(l\cdot3^n + x,q\cdot3^n + y) \mod 3 \qquad \text{for all } 0\leq x,y \leq 3^n - 1 \label{eq:sim1}\end{aligned}$$ where this means that for every $x,y,l,q$ the equivalence should hold with either a plus or minus sign. For $(l,q)=(1,1)$ we require $$\begin{aligned}
\Phi(3^n+x,3^n+y) \equiv 0 \mod 3 \qquad \text{for all } 0\leq x,y \leq 3^n - 1 \label{eq:sim2}\end{aligned}$$ Figure \[fig:carpet1\] shows this graphically. The values $(l,q)=(1,1)$ corresponds to the empty square in the middle, and all other values of $(l,q)$ should be copies of the square at $(l,q)=(0,0)$, up to exchanging the non-white colours.
To show this we prove something slightly more general.
\[def:fm\] For any $m\in\mathbb{Z}$ define $f_m:\mathbb{N}\times\mathbb{N}\to\mathbb{Z}$ as $$\begin{aligned}
f_m(x,y) = \sum_{k=0}^{\min(x,y)} \binom{x}{k}\binom{y}{k} (-m)^k\end{aligned}$$
The reason for the minus sign in $(-m)^k$ is that all valid quantum walks will have $m\geq 0$ this way, as will become clear later. As a side note, this function is a special case of the so-called hypergeometric function $_2F_1(a,b;c;z)$, namely $f_m(x,y) = {_2}F_1(-x,-y,1,-m)$. The following claim could be seen as something similar to Corollary \[cor:addpower\] but for $f_m$:
\[claim:quantumlemma\] Let $p$ be a prime and let $0\leq l,q \leq p-1$. Then we have for all $m\in\mathbb{Z}$ and for all $0\leq x,y\leq p^n - 1$ $$\begin{aligned}
f_m(l\cdot p^n + x \;,\; q\cdot p^n + y) \equiv f_m(l,q)\cdot f_m(x,y) \mod p .
\end{aligned}$$
Note that any sum can be split in the following way: $$\begin{aligned}
\sum_{k=0}^{p^{n+1}-1} g(k) = \sum_{s=0}^{p-1} \sum_{k=0}^{p^n - 1} g(s\cdot p^n + k) ,
\end{aligned}$$ where $s$ takes the role of the most significant digit and $k$ takes the role of the other digits. We apply this idea to the sum in $f_m(x,y)$ where we note that $\min(lp^n + x, qp^n + y) \leq p^{n+1} -1$ but we can let the sum range all the way to $p^{n+1}-1$ because the summand is zero in this extra range. Therefore we have $$\begin{aligned}
f_m(l\cdot p^n + x \;,\; q\cdot p^n + y) &= \sum_{k=0}^{p^{n+1}-1} \binom{l\cdot p^n+x}{k}\binom{q\cdot p^n+y}{k} (-m)^k \\
&= \sum_{s=0}^{p-1}\sum_{k=0}^{p^n-1} \binom{l\cdot p^n+x}{s\cdot p^n + k}\binom{q\cdot p^n+y}{s\cdot p^n + k} (-m)^{s\cdot p^n + k}
\end{aligned}$$ Note that by Fermat’s little theorem we have $m^p \equiv m \mod p$ so $m^{s\cdot p^n} \equiv m^s \mod p$. Now we apply Corollary \[cor:addpower\] to the binomial coefficients to obtain $$\begin{aligned}
f_m(l\cdot p^n + x \;,\; q\cdot p^n + y) &\equiv \sum_{s=0}^{p-1}\sum_{k=0}^{p^n-1} \binom{l}{s}\binom{q}{s} \binom{x}{k}\binom{y}{k} (-m)^{s + k} \\
&\equiv \left( \sum_{s=0}^{p-1} \binom{l}{s}\binom{q}{s} (-m)^{s} \right ) f_m(x,y) \\
&\equiv f_m(l,q)\cdot f_m(x,y) \mod p,
\end{aligned}$$ as required.
Note that $l,q$ take the role of the most significant digits and $x,y$ are the other digits. Just as Corollary \[cor:addpower\] implies Lucas’s theorem, we can apply this claim inductively on the number of digits to arrive at a result very similar to Lucas’s theorem but now for the function $f_m$:
\[lemma:quantumlucas\] Let $p$ be prime and $x,y$ non-negative integers. Let $x=[x_n x_{n-1}\cdots x_0]_p$ and $y=[y_n y_{n-1}\cdots y_0]_p$. Then for all $m\in\mathbb{Z}$ we have $$\begin{aligned}
f_m(x,y) \equiv f_m(x_n,y_n)\; f_m(x_{n-1},y_{n-1})\cdots f_m(x_0,y_0) \mod p.
\end{aligned}$$
We can now prove and by noting that $\Phi(x,y) = (-1)^y f_{1}(x,y)$ so by Claim \[claim:quantumlemma\] we have $$\begin{aligned}
\Phi(l\cdot 3^n + x, q\cdot 3^n + y) \equiv (-1)^{q\cdot 3^n} f_{1}(l,q) \Phi(x,y) \equiv \Phi(l,q)\Phi(x,y) \mod 3.\end{aligned}$$ where we used that $(-1)^{q\cdot 3^n} = (-1)^q$. Note that $\Phi(1,1) = 0$ which proves and $\Phi(l,q) \equiv \pm 1 \mod 3$ for the other values of $l,q$ which proves .
Results for a more general quantum walk
---------------------------------------
We can generalize the results of the previous section. First of all, we can consider the same numbers modulo any prime $p$. But more generally, the Hadamard operator $H$ could be replaced by any matrix $C\in U(2)$. As stated before, we can write any unitary $2\times 2$ matrix as $$\begin{aligned}
C = \begin{pmatrix}
c_r & c_u\\
c_d & c_l
\end{pmatrix}
= \begin{pmatrix}
\sqrt{p} \; e^{i \alpha} & \sqrt{1-p} \; e^{i\beta} \\
-\sqrt{1-p} \; e^{i \gamma} & \sqrt{p} \; e^{i (\gamma + \beta - \alpha)}
\end{pmatrix} \quad , \quad \text{with } 0 \leq p \leq 1.\end{aligned}$$ The expression for the amplitudes with general coin operator is given by $\psi_\downarrow(n,t)$ as given in Claim \[claim:amplitudeexpressions\], and we can do the same substitutions as before to go to the $x,y$ coordinates: $$\begin{aligned}
\Phi_C(x,y) &= c_d c_r^x c_l^y \sum_{k\geq 0} \binom{x}{k} \binom{y}{k} \left(-\frac{1-p}{p}\right)^k \\
&= c_d c_r^x c_l^y \; f_{m}(x,y) .\end{aligned}$$ where $m=(1-p)/p$ and we extend the definition of $f_m$ for non-integer $m$. Note that $(1-p)/p\geq 0$ for any valid coin which was the reason for defining $f_m$ with a minus sign. In the previous sections we considered the *amplitudes* of the quantum walk as opposed to the *probabilities* (which are equal to the norm squared of the amplitudes). For the purpose of the Sierpinski carpet, this distinction was not important because all entries of the Hadamard matrix are real and we were only interested in whether or not an integer was zero modulo a prime. Since $x\equiv 0 \mod p \iff x^2\equiv 0 \mod p$, squaring did not matter. For a general coin, however, there could be complex amplitudes and so we consider the corresponding *probabilities* to make sure all numbers involved are real. Note that since $f_m$ is real, the imaginary component of $\Phi_C(x,y)$ comes only from $c_d c_r^x c_l^y$. We therefore consider the probabilities: $$\begin{aligned}
|\Phi_C(x,y)|^2 = |c_d c_r^x c_l^y |^2 \left( f_{m}(x,y) \right)^2 .\end{aligned}$$ In the previous sections we rescaled the Hadamard matrix by a factor of $\sqrt{2}$ so that all numbers involved became integer. For a general coin matrix, in order to consider the probabilities modulo a prime, we assume that the coin matrix is such that $m=(1-p)/p$ is integer. Note that this can not be achieved by scaling the entire matrix because $m$ is invariant under such scalings. In fact, we have $p=\frac{1}{1+m}$ and $m\geq 0$ has to be integer. Furthermore, as stated in Claim \[claim:quantumlemma\], the complex phases $\alpha,\beta,\gamma$ do not influence $|\Phi_C(x,y)|^2$. Therefore the most general form of the matrix we can consider to obtain integer probabilities is the unitary matrix $$\begin{aligned}
C_m =
\begin{pmatrix}
\sqrt{1/(1+m)} & \sqrt{m/(1+m)} \\
\sqrt{m/(1+m)} & -\sqrt{1/(1+m)}
\end{pmatrix}
\quad \text{for } m\in\mathbb{Z} ,\; m \geq 0 ,\end{aligned}$$ where have set $\alpha=\beta=0$ and $\gamma=\pi$ such that $C_1=H$, but any other setting of phases would be equally valid. If we want to scale the matrix by a factor $\lambda$ such that $|c_d c_r^x c_l^y|^2$ is integer then this requires $\lambda = \sqrt{n(1+m)}$ for any integer $n\geq 1$. This gives a scaled matrix $$\begin{aligned}
\sqrt{n(1+m)}C_m = \sqrt{n}
\begin{pmatrix}
1 & \sqrt{m} \\
\sqrt{m} & -1
\end{pmatrix} , \label{eq:scaledgeneralcoin}\end{aligned}$$ and for this scaled matrix, $|c_d c_r^x c_l^y|^2 = m n^{x+y+1}$. By Claim \[claim:quantumlemma\] we have for this scaled coin matrix that $$\begin{aligned}
|\Phi_C(l\cdot p^n+x,q\cdot p^n+y)|^2 &\equiv \frac{n^{(l+q)(p^n-1)-1}}{m} \; |\Phi_C(l,q)|^2 \; |\Phi_C(x,y)|^2 \mod p \\
&\equiv \frac{1}{mn} \; |\Phi_C(l,q)|^2 \; |\Phi_C(x,y)|^2 \mod p ,\end{aligned}$$ where we used Fermat’s little theorem in the second step. For $m=1$ and $n=1$ we recover the exact same rules as for the Hadamard matrix. This class also includes the commonly used coin $$\begin{aligned}
\frac{1}{\sqrt{2}}\begin{pmatrix}1 & i \\ i & 1\end{pmatrix}.\end{aligned}$$ To find out what kind of fractals are generated by these quantum walks, it is useful to note that we are only interested in distinguishing $|\Phi(x,y)|^2 \equiv 0 \mod p$ from $|\Phi(x,y)|^2\not\equiv 0 \mod p$. Since we have $|\Phi(x,y)|^2 = mn^{x+y+1}(f_m(x,y))^2$ we can see that if $m\equiv 0 \mod p$ or $n\equiv 0 \mod p$ then all values $|\Phi(x,y)|^2$ are zero modulo $p$ and there is no fractal since all pixels are white. Therefore, assume that both $m$ and $n$ are not zero modulo $p$. In that case we have $|\Phi(x,y)|^2 \equiv 0\mod p$ if and only if $f_m(x,y)\equiv 0 \mod p$. Now we can apply the quantum version of Lucas’ theorem. By Lemma \[lemma:quantumlucas\], $f_m(x,y)\equiv 0 \mod p$ if and only if there is an $i$ such that $f_m(x_i,y_i)\equiv 0 \mod p$, where $x_i,y_i$ are the base-$p$ digits of $x,y$.
In general, to find the fractal generated by a quantum walk with the general coin from Equation for some $n,m$ that are non-zero modulo $p$, we simply have to compute $f_m(x,y) \mod p$ for only $0\leq x,y < p$ to find what we call the *base image*. Figure \[fig:baseplots\] shows these base images for several values of $m$ and $p$. From this the fractal can be constructed in a simple recursive way, shown in Figure \[fig:constructionstages\], resulting in the fractals shown in Figure \[fig:generalfractals\]. This recursive method is valid because each recursion step corresponds to adding another digit to $x$ and $y$, and as mentioned above, a pixel will be white if and only if there are digits (i.e. a recursion step) in which the region corresponding to those digits is white. In case of the example construction (Figure \[fig:constructionstages\]), the third picture corresponds to $x,y$ values in the range $0\leq x,y < 2^3$ which can be described by three digits modulo 2. Let $x=[x_2 x_1 x_0]_2$ and $y=[y_2 y_1 y_0]_2$, then $x_2,y_2$ specify which of the four biggest quadrants the pixel is in. Likewise, $x_1,y_1$ specify which of the four subquadrants of that first quadrant it is in, and $x_0,y_0$ specify the final position within that subquadrant. By the construction in Figure \[fig:constructionstages\], the pixel will be white if and only if one of those chosen quadrants was bottom-right. This is equivalent to saying that the pixel is white if and only if there is an $i$ such that $f_m(x_i,y_i)\equiv 0\mod p$.
$\implies$ $\implies$ $\implies$
Other properties of the Hadamard triangle
-----------------------------------------
One can add the probabilities in each row of the triangle and by unitarity this sum will always be equal to one. Instead one can also consider summing all the *amplitudes* in a row. Define the column vector $\Psi(t) = (\Psi_\uparrow(t) \;\; \Psi_\downarrow(t))^T$ where $\Psi_\uparrow(t)$ is the sum of the up amplitudes at time $t$, i.e. $$\begin{aligned}
\Psi_\uparrow(t) = \sum_{n=-t}^{t} \psi_\uparrow(n,t) ,\end{aligned}$$ and similar for $\Psi_\downarrow(t)$. Alternatively, consider the linear map $$\begin{aligned}
\Sigma = \sum_{n\in\mathbb{Z}} {\langlen\vert},\end{aligned}$$ so that $\Psi(t) = \Sigma \; U^t {\vert0,\uparrow\rangle}$. Note that to go from time $t$ to $t+1$, one application of a coin and shift is performed ($U=S(\mathrm{Id}\otimes H)$), but the sums of all up or down amplitudes are invariant under the shift operation. In other words $\Sigma S = \Sigma$. Furthermore we have $\Sigma (\mathrm{Id}\otimes H) = H \Sigma$, so we have $$\begin{aligned}
\Sigma U^t = \Sigma (S(\mathrm{Id}\otimes H))^t = H^t \Sigma.\end{aligned}$$ This can also be seen by simply looking at Figure \[fig:qpascal2\], and noting that $\Psi_\uparrow(t+1)=\frac{1}{\sqrt{2}}(\Psi_\uparrow(t)+\Psi_\downarrow(t))$ and $\Psi_\downarrow(t+1)=\frac{1}{\sqrt{2}}(\Psi_\uparrow(t)-\Psi_\downarrow(t))$, or simply $$\begin{aligned}
\Psi(t+1) = H \Psi(t).\end{aligned}$$ The sum over all amplitudes, up and down, is therefore $$\begin{aligned}
\Psi_\uparrow(t) + \Psi_\downarrow(t) =
\begin{cases}
\Psi_\uparrow(0) + \Psi_\downarrow(0) & t \text{ even}\\
\sqrt{2} \Psi_\uparrow(0) & t \text{ odd}
\end{cases}\end{aligned}$$ Note that when the process is scaled so that all numbers become integer (i.e. $H'=\sqrt{2}H$), as was done for the fractals, and the starting state is ${\vert0,\uparrow\rangle}$ then the above gives $$\begin{aligned}
\Psi'_\uparrow(t) + \Psi'_\downarrow(t) =
\begin{cases}
2^{t/2} & t \text{ even},\\
2^{(t+1)/2} & t \text{ odd},
\end{cases}\end{aligned}$$ so the sum of all amplitudes in a row is always a power of two.
Pascal’s triangle has the property that summing over the so-called *shallow diagonals* yields the Fibonacci sequence. The $n$’th shallow diagonal $d_n$ ($n\geq 0$) corresponds to the sum $$\begin{aligned}
d_n = \sum_{c = 0}^{\lfloor n/2 \rfloor} \binom{n-c}{c},\end{aligned}$$ over the numbers in Pascal’s triangle and is equal to the Fibonacci number $F_{n+1}$, where $F_{1}=F_{2}=1$ and $F_{n+1}=F_n + F_{n-1}$. By the property $\binom{n}{k}=\binom{n-1}{k-1}+\binom{n-1}{k}$ it follows that $$\begin{aligned}
d_n &= \sum_{c\geq 1} \binom{(n-1)-c}{c-1} + \sum_{c \geq 0}\binom{(n-1)-c}{c}\\
&= \sum_{c\geq 0} \binom{(n-2)-c}{c} + \sum_{c \geq 0}\binom{(n-1)-c}{c} = d_{n-2} + d_{n-1}.\end{aligned}$$ We can consider the same diagonals but now in the triangle of amplitudes of the Hadamard walk. In particular we will consider the same numbers that gave rise to the Sierpinski triangle, namely the red and blue numbers of Figure \[fig:qpascal2\]. The blue numbers (down components) are given by $H_\mathrm{blue}$ as in Equation (\[eq:Hblue\]). Similarly, the red numbers (up components) are given by $$\begin{aligned}
H_\mathrm{red}(R,C) = \begin{cases} \sum_{k\geq 1} \binom{C+1}{k} \binom{R-C-1}{k-1} (-1)^{R-C-k} & C < R\\ 1 & C=R\end{cases}\end{aligned}$$ Unlike the case of Pascal’s triangle, it now matters in which direction the diagonal is considered because the triangle is no longer symmetric. We therefore consider four options, corresponding to the two triangles (red and blue) and the two possible directions for the diagonals $\diagup$ and $\diagdown$. We denote the $\diagup$ diagonals by $A_\mathrm{red}$ and $A_\mathrm{blue}$ and the $\diagdown$ diagonals by $B_\mathrm{red}$ and $B_\mathrm{blue}$. They are defined as $$\begin{aligned}
\begin{array}{rlcrl}
A_{\mathrm{blue},n} =& \sum_{c\geq 0} H_\mathrm{blue}(n-c,c) & ~ &
B_{\mathrm{blue},n} =& \sum_{c\geq 0} H_\mathrm{blue}(n-c,n-2c)\\
A_{\mathrm{red} ,n} =& \sum_{c \geq 0} H_\mathrm{red}(n-c,c) & ~ &
B_{\mathrm{red} ,n} =& \sum_{c\geq 0} H_\mathrm{red}(n-c,n-2c) .
\end{array}\end{aligned}$$ Using the same property of binomial coefficients as before, we have $$\begin{aligned}
A_{\mathrm{blue},n}
&=\sum_{c\geq 0}\sum_{k\geq 0} \binom{c}{k} \binom{n-2c}{k} (-1)^{n-k} \\
&=\sum_{c\geq 0}\sum_{k\geq 0} \binom{c}{k} \binom{n-2c-1}{k-1} (-1)^{n-k}
+ \sum_{c\geq 0}\sum_{k\geq 0} \binom{c}{k} \binom{n-2c-1}{k} (-1)^{n-k} \\
&= A_{\mathrm{red},n-2} - A_{\mathrm{blue},n-1}\end{aligned}$$ Similarly we find $$\begin{aligned}
A_{\mathrm{red},n} = A_{\mathrm{red},n-2} + A_{\mathrm{blue},n-1},\end{aligned}$$ and combining these two equations yields the same recurrence relation for both the red and blue diagonals: $$\begin{aligned}
A_{n} &= - A_{n-1} + A_{n-2} + 2 A_{n-3},\end{aligned}$$ but with different initial conditions for the red and blue sequences. For the diagonal in the other direction ($\diagdown$) we find $$\begin{aligned}
B_{\mathrm{blue},n} &= B_{\mathrm{red},n-1} - B_{\mathrm{blue},n-2}, \\
B_{\mathrm{red} ,n} &= B_{\mathrm{red},n-1} + B_{\mathrm{blue},n-2},\end{aligned}$$ which can be combined to form another recurrence relation $$\begin{aligned}
B_{n} = B_{n-1} - B_{n-2} + 2 B_{n-3},\end{aligned}$$ that holds for both the blue and red sequence but with different initial conditions.
Acknowledgements
================
The authors would like to thank Florian Speelman and Jeroen Zuiddam for useful discussions and Frank den Hollander for feedback. The work in this paper is supported by the Netherlands Organisation for Scientific Research (NWO) through Gravitation-grant NETWORKS-024.002.003.
[^1]: QuSoft and CWI Amsterdam, Science Park 123, 1098 XG Amsterdam, The Netherlands
[^2]: University of Amsterdam, Science Park 904, 1098 XH Amsterdam, The Netherlands
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Zika virus has been the primary suspect in the large increase in incidence of microcephaly in 2015-6 in Brazil. However its role is not confirmed despite individual cases in which viral infections were found in neural tissue. Recently, the disparity between the incidences in different geographic locations has led to questions about the virus’s role. Here we consider the alternative possibility that the use of the insecticide pyriproxyfen for control of mosquito populations in Brazilian drinking water is the primary cause. Pyriproxifen is a juvenile hormone analog which has been shown to correspond in mammals to a number of fat soluble regulatory molecules including retinoic acid, a metabolite of vitamin A, with which it has cross-reactivity and whose application during development has been shown to cause microcephaly. Methoprene, another juvenile hormone analog that was approved as an insecticide based upon tests performed in the 1970s, has metabolites that bind to the mammalian retinoid X receptor, and has been shown to cause developmental disorders in mammals. Isotretinoin is another example of a retinoid causing microcephaly in human babies via maternal exposure and activation of the retinoid X receptor in developing fetuses. Moreover, tests of pyriproxyfen by the manufacturer, Sumitomo, widely quoted as giving no evidence for developmental toxicity, actually found some evidence for such an effect, including low brain mass and arhinencephaly—incomplete formation of the anterior cerebral hemispheres—in exposed rat pups. Finally, the pyriproxyfen use in Brazil is unprecedented—it has never before been applied to a water supply on such a scale. Claims that it is not being used in Recife, the epicenter of microcephaly cases, do not distinguish the metropolitan area of Recife, where it is widely used, and the municipality, and have not been adequately confirmed. Given this combination of information about molecular mechanisms and toxicological evidence, we strongly recommend that the use of pyriproxyfen in Brazil be suspended until the potential causal link to microcephaly is investigated further.'
author:
- 'Yaneer Bar-Yam$^*$, H. Frederik Nijhout$^\dag$, Raphael Parens$^*$, Felipe Costa$^*$, and Alfredo J. Morales$^*$'
title: A Possible Link Between Pyriproxyfen and Microcephaly
---
Overview
========
In 2015 and early 2016 over 6,000 suspected [@MdS2016_2] cases of microcephaly and other neurodevelomental disorders were reported in Brazil, primarily in the northeast, a dramatic increase over expected numbers. By September 2016, over 2,000 were confirmed [@WHO]. The cause of these developmental disorders has been widely attributed to a Zika virus epidemic first detected in May, 2015 [@WHO2015]. While evidence exists, particularly in neurological infections of a few microcephalic cases [@Mlakar2016; @Tang2016], only $12-16\%$ of confirmed microcephaly cases have also been confirmed as having Zika infections [@MdS2016_2; @bul7272016], preventing health authorities from determining a conclusive causal link [@schuler2016; @Parens2016]. The strongest potential counter evidence is present in the geographic distribution of cases in Brazil and a significantly lower number of government reported cases in Colombia, though questions about the way cases are reported prevent definite conclusions. Reports suggest that areas outside of northeast Brazil do not have microcephaly cases commensurate with the prevalence of Zika infections [@naturebutler].
The first cases of Zika in Colombia reportedly occurred in September 2015 [@WHO2015]. During the first 12 full weeks of 2016, 34 cases of microcephaly were identified [@Poho8April2016; @WHOApril72016], compared to 32 cases expected based upon 140 annual background cases. Subsequently, cases with evidence of both Zika and microcephaly were analyzed [@August1]. Background cases of microcephaly that are coincidentally present in pregnancies that have Zika infections account for 11 cases until Nov. 1, 2016. An additional 45 cases for a total of 56 cases can be attributed to Zika infections at the end of the first trimester. These cases are quite limited in number, and much lower than expectations [@Cobb2016]. In order for a consistent conclusion that Zika is the primary cause of microcephaly, the reported cases of Zika in Colombia and Brazil would have to be in the same ratio; with 90,000 reported Zika cases in Colombia and 200,000 in Brazil, they are widely discrepant. Reports of the number of Zika infections are not reliable and there are differences in methods of microcephaly counting, making definite conclusions difficult. However, the two orders of magnitude separating the number of cases in Brazil and Colombia certainly raise questions about whether Zika is the dominant cause or whether there are other causes or cofactors that are increasing the incidence of microcephaly in Brazil. There continue to be many more cases in Northeast Brazil at a rate of approximately 100 per month. Other countries report very few cases [@WHO; @sitrep].
A group of physicians in Argentina and Brazil have suggested that widespread use of the pesticide pyriproxifen [@Pubchem] to reduce mosquito populations because of a dengue epidemic in 2014 may be the cause of microcephaly [@Physicians2016]. This suggestion has been challenged both by authorities based upon a claim of lack of evidence [@Costa2016; @Debunk] and by skeptics of conspiracy theories [@Gorski2016]. We have previously reviewed the primary evidence [@Parens2016], including evidence in rat toxicology studies performed by its manufacturer Sumitomo [@Saegusa1988], concluding that pyriproxifen is a credible candidate so that further inquiry is warranted. The Swedish Toxicology Research Center has published a review of toxicological information and potential molecular mechanisms associated to thyroid function and concluded that further studies are warranted [@swetox]. Moreover, arguments that toxicology studies are sufficient to rule it out as a cause would not rule it out as a cofactor.
Here we discuss the molecular mechanisms associated with pyriproxyfen in insect and mammalian development and summarize toxicological studies and insecticide use in Brazil [@Evans2016]. We also review available data on Zika infections. Of particular note is that pyriproxyfen, a biochemical analog of juvenile hormone in insects, is cross-reactive with the retinoic acid / vitamin A regulatory system of mammals, and that exposure to retinoic acid has been shown to cause microcephaly [@Encyclopedia]. Existing toxicological studies and experience with public use do not provide evidence to counter this causal chain. Toxicological studies, if anything, provide evidence for neurodevelopmental toxicity at a level consistent with that found for microcephaly in Brazil. Pyriproxyfen’s use in drinking water is unprecedented. The few trials using pyriproxyfen in the water supply did not investigate nor were they sufficient to observe a role in causing birth defects at the level seen in Brazil. Our review of the currently available information points to adequate reasons to suspect this insecticide as a developmental disruptor causing microcephaly. It also provides reasons to question the adequacy of the testing of pyriproxyfen’s toxicology and its use in water supplies.
This paper is organized as follows. In Section II, we describe the potential causes of microcephaly. In Section III, we discuss the molecular mechanisms of pyriproxyfen. In Section IV, we review animal toxicological studies conducted in laboratory settings. In Section V, we describe pyriproxyfen usage in drinking water supplies in Brazil. In Section VI, we discuss evidence for and against the causal relationship between Zika and microcephaly and analyze data on Zika and microcephaly from Brazil and Colombia. In Section VII, we summarize the evidence and suggest policy changes based upon the available data.
Potential causes and timeline
=============================
The World Health Organization (WHO) identifies the most common causes of microcephaly as: infections, exposure to toxic chemicals, genetic abnormalities, and severe malnutrition while in the womb. The first includes toxoplasmosis, which is caused by parasites in undercooked meat, rubella, herpes, syphilis, cytomegalovirus and HIV. The second includes exposure to heavy metals, such as arsenic and mercury, as well as alcohol, smoking and radiation exposure. The third may include Down syndrome [@WHO2016]. While the first category points to the possibility of infectious causes, the second points to the possibility of chemical ones. The existence of multiple causes of microcephaly suggests many potential origins of a general developmental disruption. Hence, it is possible that more than one source of microcephaly is present in Brazil.
The suspected association of increased microcephaly cases with Zika and/or pyriproxyfen relies first on an increase in the cause with the increase in number of cases. The outbreak of the former was recognized in May 2015, and the use of the latter in the fourth quarter of 2014, with the incidence of microcephaly cases beginning in October of 2015. Thus both Zika and pyriproxyfen satisfy the first criterion for a cause. The precise dates of the increase in cases are believed to be uncertain because of potential problems with underreporting prior to the medical alert and overreporting afterwards [@schuler2016].
Pyriproxyfen molecular mechanisms
=================================
Juvenile hormone and retinoic acid are both lipid-soluble terpenoids. They act as signaling molecules that control a broad diversity of embryonic and postembryonic developmental processes in insects and vertebrates, respectively. Juvenile hormone is best known for its role in the control of metamorphosis and reproduction in insects [@Riddiford2012; @Wheeler2003], whereas retinoic acid is involved in the development of the nervous system in vertebrates [@Rhinn2012].
These two classes of hormone-like molecules share some molecular similarities and are capable of some degree of cross-reactivity. For instance, retinoic acid is known to mimic some of the effects of juvenile hormone when injected into insects [@Nemec1993], and juvenile hormone and its analogs are known to bind to the vertebrate retinoic acid receptor [@Palli1991; @Jones1995; @Harmon1995]. It is possible therefore that pyriproxyfen, a powerful juvenile hormone analog [@Dhadialla1998], can bind to the retinoic acid receptor. When it does so it could either activate the receptor at inappropriate times in development, or act as a blocker that prevents the normal retinoic acid from binding to the receptor when needed. The retinoic acid receptor normally turns on gene expression in development, so either an inappropriate activation or an inhibition at a critical time could be expected to lead to developmental abnormalities.
More specifically, another juvenile hormone analog that has been approved as an insecticide, methoprene, has also been shown to have metabolites that bind to the mammalian retinoid X receptor and has been shown to cause developmental disorders in mammals [@Harmon1995; @Unsworth1974].
Isotretinoin is a retinoid that is widely used in medicine but is contraindicated in women who are pregnant or might become pregnant. It causes microcephaly in human babies via maternal exposure and activation of the retinoid X receptor in developing fetuses [@Stem1989; @Irving1986].
The impact of retinoids on abnormal development (teratogenesis) has been demonstrated to be sensitive to genotype and developmental stage of exposure and to result in death, malformation, growth retardation, and/or functional disorder [@Collins1999]. The juvenile hormone itself and different juvenile hormone analogs have different binding properties to mammalian retinoic acid receptors, which could therefore produce different effects, including teratogenicity. These effects remain poorly understood [@Wheeler2003; @Flat2006].
Toxicological Studies
=====================
Overview of Toxicological Studies
---------------------------------
The potential link between pyriproxyfen and microcephaly has been challenged based upon the existence of toxicological studies by the manufacturer Sumitomo. However, we have reviewed their studies and find them largely restricted to analysis of the impact on adult animals. In the limited experiments on developmental toxicity, the ability to identify a link to microcephaly is weak based upon the specific tests that have been performed. Rather than measurements, in most tests visual macroscopic observation of rat and rabbit fetuses or pups/kits were used [@sumitomo], for which standards of microcephaly determination may not be sufficiently well established or comparable across species. In the most relevant experiment there is a reported test of brain weight and neurodevelopmental disorders in rat pups, and in that study, there is actual evidence for microcephaly. Sumitomo has argued that the evidence for microcephaly should not be considered because of its dose dependence, but this argument is not consistent with the statistical nature of the study, as the dependence on dosage is not statistically verified. Further, it is important to recognize that the study design adopts an “innocent until proven guilty" assumption as the null hypothesis is the absence of toxicity. It is the presence of toxicity that must pass a statistical significance test, which is not the same as the proof of its absence.
Specifically, the most relevant study for a determination of neurodevelopmental toxicity [@Saegusa1988] considered brain and behavioral effects of rat pups exposed to pyriproxyfen during days 7 to 17 of gestation, which lasts 21 days. The experimental group of 36 pregnant rats in each of four test groups was fed dosage levels of 0, 100, 300 and 1000 mg/kg/day. The pups were checked for physiological deformations and organs weighed.
From each dosage level, litters of pups were obtained. While the target was to obtain 99 pups in each group, only 78 pups were obtained in the 1000 mg/kg group due to adult deaths. For 99 pups in the 100 mg/kg and 78 pups of the 1000 mg/kg dosage groups no relevant developmental disorders were reported. Of the 99 pups in the 300 mg/kg dosage 1 (1%) had Arhinencephaly and 1 (1%) had Thyroid hypoplasia. The former would be consistent with concerns about neurodevelopmental disorders of the type of microcephaly as well as observations that a variety of neurological and other developmental disorders have been found associated with the epidemic in Brazil [@Brasil2016].
Of the resulting offspring, a number of pups were kept alive for emotional/mental testing at 4 and 6 weeks of age and their brains were subsequently weighed at 8 weeks. The report then gives values for group sizes of 13, 12, 11, and 10 for both male and female in the control and the three dosage levels, respectively. One of the groups, the males of the 300 mg/kg group, had statistically significant lower brain weight at 8 weeks. Male pups at other dosage levels were also lower in relative brain mass compared to the control.
These tests provide evidence that microcephaly may be an outcome of the application of pyriproxyfen to rats and other mammals. As discussed in the following section, the measures of low brain mass and incidence of Arhinencephaly were dismissed by Sumitomo based upon the assumption of dose dependence—higher dosages should lead to larger toxicological effects, otherwise the effects must not be toxicological. The conventional toxicology assumptions of dose dependence infer from the absence of similar observations in the 1,000 mg/kg dosage group that the observations in the 300 mg/kg group are not relevant to the toxicology of pyriproxyfen [@sumitomo].
This assumption, however was not correctly applied due to the statistical nature of the experiment: Statistical variation may lead to lower effects by random chance, which may mask increasing effects. Tests of toxicological effects must demonstrate that the direction of change is counter to a toxicological effect to an observed level of statistical certainty.
This is also a concern for individual events such as the case of Arhinencephaly. Rare events should not be assumed to be background effects when those effects may also be caused by toxicological effects that are rare. Specific events may be caused by genetic variability and physiological regulatory sensitivity. Instead, the outcomes might be interpreted as an estimated probability of incidence of 1 in 66, the total number of pups in the observed exposure groups. Considering the differences between humans and rats, the values for humans may be higher or lower than this. Note that the incidence of microcephaly in Brazil is estimated to be on the order of a few percent [@MdS2016_2]. If the rate in rats and human beings is comparable, then it would be expected that only one event would happen among the entire rat pup population of the experiment. The need for careful studies is apparent in that teratogenicity of a retinoid varies widely across mammalian species, and this variation itself varies across different retinoids [@Wilhite1986; @Irving1986; @Tzimas1994; @Nau1993; @Eckhoff1997]. Regulatory systems are remarkably sensitive to small molecular densities. Given the low incidence of human cases of microcephaly in affected populations, experiments that are designed to test for microcephaly in mammals, at the very least, should be designed to identify incidence at this level. Moreover, they should be unambiguously capable of detecting microcephaly.
Finally, there are multiple drugs that passed conventional regulatory animal testing and are now known to be linked to microcephaly, including phenytoin and methotrexate [@Homes2001; @Hyoun2012; @swetox], so assurances based solely upon conventional regulatory testing are inadequate in this case as well.
Toxicological tests are framed to require proof of toxicity rather than safety
------------------------------------------------------------------------------
Given the evidence for low brain mass, it is important to inquire why Sumitomo researchers and regulatory authorities dismiss the risks of pyriproxyfen. More importantly, why they consider the issue to be settled as to the absence of risk from pyriproxyfen.
As indicated above, this is particularly surprising given the underlying framing of the toxicological studies. The statistical test of toxicity takes as its null hypothesis that there is no toxicity. The use of statistical significance for toxicity assumes that proof of toxicity is required, not of safety. This is counter to the premise that the toxicity study provides evidence of a lack of toxicity. Absence of proof of toxicity is surely not proof of the absence of toxicity. Given this framing, it should not be surprising that many cases where toxicological tests do not find toxicity result in the approval of substances that are later found to be toxic.
In particular, the null hypothesis of the statistical study is that there is no difference between the control and the sample that has received pyriproxyfen. The type of test that is done is similar to that of a medicine which requires proof of positive effect, satisfying a statistical significance threshold, in order to be administered. In the case of toxicity, the absence of statistical significance is not a demonstration of the lack of toxicity, but rather a lack of proof of toxicity. Thus the study is not considering whether pyriproxyfen is safe, it is asking whether there is proof of toxicity beyond a statistical uncertainty. This logical framework also leads to a circumstance where low statistical power because of small sample numbers decreases the ability of the study to identify actual toxicity. In such a case, the absence of statistical significance may directly follow from poor statistics. In a commercial context where regulatory approval enables sales and profits, there is a financial incentive to use small samples that minimize the statistical power of the test, because a limited test does not show toxicity even when it is present.
In this case, however, there is actually a significant result for one of the trials. The average effect is large enough even in a context where the standard error is large because of the small sample used.
However, the report, and subsequent government regulatory approval, dismisses the result because there is not a larger effect at higher dosage. This dose dependent assumption is standard in toxicology. However, the dose dependence, which is only of two data points, itself is not subject to a statistical test. The trend is considered to be true even though it is not valid within statistical uncertainty—the opposite trend of an increasing impact could occur given statistical variation between the two dosage levels. Smaller brain mass relative to the control is present in the higher dosage case. While less severe on average, the difference between the lower and higher dosage level is well within a standard error, and therefore is within observational uncertainty. Thus it should not be used to make decisions about whether there is toxicity or not.
The study does not provide evidence for an absence of toxicity. It assumes a null effect, has poor statistical power, actually has some evidence for toxicity, and that evidence is dismissed based upon incorrect use of a trend analysis that statistics does not support.
Specifics of Sumitomo studies of pyriproxyfen developmental toxicity
--------------------------------------------------------------------
The toxicity experiments summarized above were performed by the producer of pyriproxyfen, Sumitomo, in 1988. The most relevant of these studies is a rat teratology study by administration of pyriproxyfen to pregnant rats (dams) by gavage on days 7-17 after pregnancy [@Saegusa1988]. A second study was performed on rabbit developmental toxicity by administration on days 6-18 of pregnancy [@Wilkinson1994]. In this section we provide additional detail about these studies.
### Low brain mass observations (rat teratogenicity study)
In the rat teratogenicity study, four groups were administered different amounts of pyriproxyfen, 0 mg/kg (control), 100mg/kg, 300 mg/kg, and 1000 mg/kg. Of particular importance to our analysis is the observation of a statistically low relative brain weight among males in the 300mg/kg group (see Table \[weights\] and Fig. \[datafigure\]).
[|X|X|X|X|X|X|]{} Gender & Measure & Control & 100 mg/kg & 300 mg/kg & 1,000 mg/kg\
& & (13 pups)$^\ddag$ & (12 pups)$^\ddag$ & (11 pups)$^\ddag$ & (10 pups)$^\ddag$\
Male & Body weight (g) &$232.2\pm30.2$ & $252.0\pm33.8$ & $253.4\pm20.1$&$257.8\pm40.2$\
Male & Brain weight (g) & $1.8\pm0.101$ &$1.817\pm0.079$&$1.859\pm0.041$ &$1.809\pm.097$\
Male & Brain relative weight (mg$\%^\dag$) & $799.9\pm103$& $730.8\pm130.9$ & $726.6\pm 38.6^*$ & $746.9 \pm 80.6$\
Female & Body weight (g) & $171.1\pm11.1$ & $182.1\pm16.3$ & $170.6\pm6.2$ & $168.5\pm13$\
Female & Brain weight (g) & $1.711\pm.091$ & $1.737\pm.077$ & $1.771\pm.041^*$ & $1.708 \pm .082$\
Female & Brain relative weight (mg$\%^\dag$) & $951.7\pm60.4$& $907.4\pm75.9$ & $955.6 \pm 59.1$ & $976.4\pm 80$\
Low brain weight measurements may be indicative of microcephaly. However, according to communications from Sumitomo, results were interpreted according to standard assumptions of dose dependence. According to this assumption deviations at lower dosages are not considered if they don’t also occur at higher dosage levels. Thus, the assumptions of dose dependence dismiss the findings of lower relative brain weights in lower dose groups (300 mg/kg group), as brain weights is observed in higher dose groups (1000 mg/kg) are not statistically significant [@Saegusa1988].
The assumption of dose dependence is a mathematical assumption about the distribution of observations resulting from variables $X$ that cause variations in the measured quantity unrelated to the application of pyriproxyfen, and $Y$ that cause variations due to pyriproxyfen. It is assumed that the observable (relative brain weight) $$W(p) = \sum_{X,Y} W(X,Y) P(X,Y;p),
\label{Eq_av}$$ is a monotonically varying function of $p$. However, this assumption is then applied to the observed value, which is not given by the actual average overall probabilities but by a sample: $$W_o(p_j) = \sum_{X_i,Y_i} W(X_i,Y_i),
\label{Eq_2}$$ where the sum is over the particular cases that arose in the sampling of the experiment for dosage $p_j$. The difference between observed and expected outcomes $W_o(p_j)-W(p)$ has distribution of values. The conclusion of monotonic dose dependence for a teratogenic effect would only be clearly valid if the standard deviation of this distribution is small compared to the difference of values observed. More precisely, a null hypothesis of a teratogenic effect should be falsified by the observational study. However, the standard error of the observations is larger than the difference between those observations and therefore does not support this conclusion.
More specifically, all three groups administered pyriproxyfen are low in relative brain mass compared to the control and differ from each other by statistically insignificant amounts.The 300 mg/kg group has a value of 726.6, compared to the control of 799.9, a difference of 73.3. The 100 mg/kg group has a value of 730.8, a difference relative to the control of 69.1, only 4.2 smaller. The 1000 mg/kg group value is 746.9 which is 53 less than the control. While the latter two are not statistically significant according to the analysis, the 100 mg/kg case is not far from statistical significance, and taken together, the set of three results across dosage levels provide evidence that pyriproxyfen causes low brain weight.
We note that a relatively high number of pregnant rats in the 1000 mg/kg died before pregnancy (12 of 42 pregnant rats) [@Saegusa1988]. Those who survived the test are not an unbiased sample (survivor bias), perhaps eliminating cases of low relative brain weight among pups in this class. This would follow from a circumstance in which those rat pups who were more susceptible to pyriproxyfen leading to low brain mass, would also be the ones who, at increased dosage, have dams more likely to die during pregnancy. Quantitatively, the invalid assumption by those performing the study is the assumption that maternal death is an independent variable to pup brain mass.
It is also important to note that for the case where rare events might be the cause of a deviation from control values, a large enough sample must exist for sampling to average over the rare events. Since microcephaly may occur in only a small fraction of pups, the number of samples must be large compared to the inverse of the rate of its occurrence so that there are many such events in the experiment. For example, if microcephaly were to occur in only 1 per 100 pups, the number of samples must be large compared to 100 in order for the sampled average of Eq. \[Eq\_2\] to be reliably close to that of Eq. \[Eq\_av\]. Rare events that have large effect on brain mass, or have a distribution that is broad and therefore the statistical deviation is large, would not be correctly evaluated.
For the rat toxicology experiment to be of use in understanding human toxicology, we must make the assumption that we can map the results of experiments on rats onto human beings in a way that is reliable. If we make this strong assumption, we would also take the rate of incidence of microcephaly in human beings as indicative of the rate in rats. The incidence of microcephaly in pregnancies in northeast Brazil in Pernambuco is approximately 30 per 10,000 births. We do not have information about the rate of exposure. Still, a $3\%$ rate would be consistent with a $31\%$ chance of having one case in 12 births. This suggests that only one of the dosage levels would be likely to have a single case of microcephaly. Discounting the $300$ mg/kg result is therefore inconsistent with the expectations based upon the incidence rate. The number of pups in each group is not sufficient for each of them to have affected individuals, if the effect occurs at the rate of microcephaly observed in Brazil. A single individual could occur in any of the experimental groups, not necessarily in the high dosage one. (The uncertainty that the test results are applicable to human beings further reduces the reliability of the experiment as a test of toxicity.) We also note that in this study, the number of rat pups at the 1,000 mg/kg level is reduced to 78 from the target number 99. While the specific reason is not explained for this difference, there are two possible reasons in the study description. First, there are fatalities of the pregnant dams that led to adding an additional 12 dams to the study. Second, five dams were excluded from the study after the fact because of a mistaken feeding of pyriproxyfen starting on day 6 rather than day 7: “Five animals (Nos. 2437 to 2441) in the 1000 mg/kg group were mistakenly administered the test substance from day 6 of gestation, and data of these animals were excluded from the evaluation." It is unclear at what point they were excluded from the study. While each of these protocol changes might be appropriate, the choices that are made by experimenters that undermine statistical assumptions have become an increasing reason to question the reliability of studies [@reproduce0; @reproduce1; @reproduce2].
Finally, in private communications Sumitomo also suggested that the use of relative brain weights may not be a good indicator of microcephaly compared to absolute brain weights. However, this statement is not reliably shown by prior studies [@Bailey2004; @Tamaru1988].
### Arhinencephaly
The case of Arhinencephaly in the rat experiment was suggested by a response of Sumitomo not to be indicative of microcephaly in humans. While both are malformations in the central nervous system, they may not be caused by the same mechanism. However, multiple types of neurological defects have been associated with the current epidemic of microcephaly in addition to microcephaly itself [@Brasil2016]. While these have often been claimed to be associated with Zika infection, the existence of these alternative central nervous system defects can equally be associated with another cause, i.e. pyriproxyfen.
Sumitomo also affirms that the dose dependence argument applies in this case as well, as it is observed in the 300 mg/kg group but not in the 1000 mg/kg group in the rat teratogenicity study [@Ozaki1]. However, the existence of only one case of Arhinencephaly in the rat teratogenicity experiment must be interpreted as either a low probability event that is independent of the application of pyriproxyfen (random variable $X$ in Eq. \[Eq\_2\]) or a low probability event that is due to the application of pyriproxyfen (random variable $Y$). Sumitomo’s argument assumes it is the former case rather than the latter case, without evidence for that claim. The assumption that it is not caused by pyriproxyfen is counter to the purpose of the testing which is to demonstrate that it does not cause harmful conditions. As stated before, the dose dependence argument does not apply to low probability events that occur at a rate that is low compared to the number of samples.
### Rabbit study
In a study conducted using rabbits, Sumitomo used a similar protocol to that for the rat study. Four groups were administered different amounts of pyriproxyfen, 0 mg/kg (control), 100mg/kg, 300 mg/kg, and 1,000 mg/kg. Rabbits in each group were mated, and resulting fetuses were examined for signs of toxicological damage. Examinations included skeletal variations, ossification of phalanges, and visceral anomalies [@Ozaki2].
The brain weight of fetuses was not measured in the rabbit teratology studies; instead, microcephaly was diagnosed by observation. When asked whether standards of observation have been sufficiently well established for the detection of microcephaly, Sumitomo responded that there is “reported historical control data of the rabbit teratogenicity studies, so it suggests that microcephaly can be detected with the macroscopic observation even in rabbits" [@Ozaki2].
The response that it is possible to diagnose microcephaly in rabbits, does not answer the question as to whether the criterion in rabbits would be similar to the condition in humans, nor whether the number of cases being counted would correspond to the corresponding number in humans. This problem is particularly important where the incidence of microcephaly is rare, only a few percent even in the high incidence area of Brazil. The criterion in human infants for microcephaly has been a key question in determining the number of cases that occur. The European Surveillance of Congenital Anomalies network (EUROCAT) uses more than three standard deviations below the normal head circumference. Brazil currently defines microcephaly as less than two standard deviations, although prior to November 2015, a broader definition was in use [@Yung2015]. Given difficulties defining microcephaly in humans, identifying microcephaly visually in a much smaller animal cranium may also be difficult. The identification of microcephaly is a key issue in evaluating whether observations in animal tests are useful. This is particularly relevant as the role of experimenter motivation has become increasingly recognized [@reproduce0; @reproduce1; @reproduce2]. While obtaining non-null statistically significant results is a motivation for scientific researchers, the opposite is true in a manufacturer conducted toxicology experiments.
Use of Pyriproxyfen in water supply in Brazil
=============================================
Pyriproxyfen has been used as an insecticide in northeast Brazil in response to an outbreak of dengue starting in the fourth quarter of 2014 [@Reis2016]. It is applied primarily to water storage containers used for home drinking water in areas that do not have a municipal water supply.
Pyriproxyfen has received widespread regulatory approval as an insecticide [@EPA; @WHO2004]. Its primary use is on agricultural crops [@Devillers2013], and against insects in households and on pets, e.g., on pet collars [@NRDC2000]. It is important to note that use of pyriproxyfen in drinking water on a large scale has never occurred anywhere else in the world, and was not used in Brazil prior to the fourth quarter of 2014. There have been several small studies of the use of pyriproxyfen for mosquito control in villages. However, these studies did not monitor for toxicological effects. Prior tests were pilots for mosquito control in a few small communities in Malaysia [@Invest2008], Peru [@Sihuincha2005], Colombia [@Overgaard2012], and Cambodia [@Seng2008]. These were small case studies, and only Cambodia and Colombia included a use in the drinking water supply [@Invest2008]. The research on impacts included measures of mosquito control but did not consider developmental effects in humans. Moreover, their scale was so small that birth defects occurring at a rate consistent with what is observed in Brazil would not have been observed. Human trials of pyriproxyfen in drinking water have not been carried out in a laboratory setting. Such experiments require a large participating sample in order to observe birth defects at such an incidence. Thus the conditions that are currently present in Brazil have not been replicated elsewhere, and the possibility that pyriproxyfen is the cause of the observed microcephaly cases cannot be ruled out by prior studies or experience with the use of pyriproxyfen. Prior to 2014, Brazil primarily used temefos (an organophosphate) in water supplies to reduce mosquito populations. The switch to pyriproxyfen was made due to increasing temefos resistance in mosquitoes [@PdS2014]. Some areas of Brazil do not use pyriproxyfen, using Bacillus thuringensis (BT) toxin instead.
It has been reported in the press that the Brazilian Ministry of Health [@MoHZika2016; @MoHMicro2016] claims areas not using pyriproxyfen also report cases of microcephaly. For example, there is a press report that pyriproxyfen was not being used in Recife, and two other cities of the state of Pernambuco [@Romo2016; @Bichell2016; @NoAuthor2016], though there are cases of microcephaly and Zika. This has been cited as a compelling reason for dismissing the possibility that pyriproxyfen is a cause of microcephaly [@NoAuthor2016]. According to the Dengue control office of Pernambuco [@personalCommunication], pyriproxyfen is widely used in the Recife metropolitan area. An unconfirmed report [@CES1], states that pyrixroyfen is not used in three specific municipalities (urban administrative areas)Recife, Paulista, and Jaboatao do Guararapes, where Bt toxin is substituted. In this context it is important to recognize that pyriproxyfen exposure is relevant for the urban favelas and nearby rural areas where it would be used in water storage containers. It is relevant to urban areas with municipal water supply systems where such systems are augmented by water storage containers due to frequent failures [@Frontline]. Urban hospitals attract populations from a large area [@Douglas2016] and official reporting of cases is not broken down by municipality but rather by state, so that a direct analysis of geographical distribution of exposure to pyriproxyfen is difficult in these particular cases. A recent paper studied the association of microcephaly with the areas of claimed pyriproxyfen use and reports a lack of correlation [@Albuquerque16]. If the geographic distribution of use and the reported residences of mothers are confirmed, this study would provide strong evidence against pyriproxyfen as a cause. However, the study does not provide a reliable source for the geographic use of pyriproxyfen. Information about the birth location of favela residents (as opposed to more conventional urban area residents) is also not adequately confirmed. Both are necessary. Thus, further analysis is needed to determine whether the large number of cases in the state of Pernambuco have occurred in areas where pyriproxyfen is used.
Review of Zika incidence in Brazil and Colombia
===============================================
The existence of an epidemic of Zika in Brazil, the first observed in the western hemisphere, during the early part of 2015, preceding the observation of a large number of cases of microcephaly starting toward the end of 2015, led to the natural inference of a causal relationship between them. We review here existing evidence including geographic data about the incidence of Zika and microcephaly in Brazil and other countries [@Parens2016; @NECSINov; @r1; @r2; @August1; @Bar-Yam2016]. A consistent picture suggests a low incidence of cases caused by Zika and another cause responsible for most of the microcephaly cases in Brazil. The possibility that a cofactor along with Zika infections is responsible is opposed by the low rate at which microcephaly cases are confirmed to be associated with Zika infections in Brazil. In particular, the ratio of microcephaly to Zika cases is inconsistent between government reported cases in Colombia and Brazil and among Brazilian states, where the majority of cases are confined to the northeast region. At the rate of microcephaly reported in Colombia, if all pregnancies in the Brazilian state of Pernambuco were infected by Zika, we estimate there would only be 105 cases of microcephaly in a year, whereas the number of confirmed cases is 386. The evidence against Zika as a cause implies other causes of microcephaly should be more fully considered, including pyriproxyfen.
Evidence linking Zika infection and microcephaly
------------------------------------------------
We begin by reviewing the evidence that is considered to link Zika with microcephaly, which we find is consistent with Zika being a cause of individual cases but does not associate it to the majority of microcephaly cases in Brazil.
In May of 2016, the CDC declared Zika the cause of microcephaly in Brazil [@Rasmussen2016]. The CDC highlights the three factors that contributed to their conclusion:
- [Zika virus infection at times during prenatal development...consistent with the defects observed;]{}
- [a specific, rare phenotype involving microcephaly and associated brain anomalies in fetuses or infants with presumed or confirmed congenital Zika virus infection;]{}
- [data that strongly support biologic plausibility, including the identification of Zika virus in the brain tissue of affected fetuses and infants.]{}
The cited evidence is about individual cases and the primary evidence of a connection to the large number of cases observed is the timing of the initial Zika outbreak relative to the initial outbreak of microcephaly cases in Northeastern Brazil.
A more extensive list of evidence linking Zika and microcephaly includes:
1. [Zika virus was detected in an aborted fetus with microcephaly after the mother had symptoms of infection in the 13th week of gestation. The virus was found in neurological tissue [@Mlakar2016].]{}
2. [Zika virus was observed infecting neural stem cells and affecting their growth [@Tang2016].]{}
3. [Zika virus was found in the amniotic fluid of two pregnant Brazilian women, after both showed possible signs of infection, including fever and a rash [@Calvet2016].]{}
4. [In a public release on February 12, 2016, Brazil’s health ministry reported 462 confirmed cases of microcephaly or other alterations to the central nervous system, after investigation of 1,227 of 5,079 suspected cases of microcephaly recorded from October 22, 2015 until February 6, 2016. 3,852 remained under investigation. Brazil confirmed that 41 of these cases of microcephaly are combined with “evidence of Zika infection...either in the baby or in the mother.” It is unclear from this report what the Zika infection status is for the microcephaly cases for which evidence of Zika infection is not reported [@Nebehay2016; @MdS2016], i.e. whether they were investigated and had no evidence of infection or whether they were not investigated. The low percentage of confirmation $8.9\% = 41/462$ has increased to approximately $15.6\%$ [@bul7272016] in sporadic subsequent reports. ]{}
5. [In a release reported on November 14, 2015, the French Polynesian health authorities reported 17 (possibly 18) cases of central nervous system malformations, including 8-9 cases of microcephaly [@polynesia1; @polynesia2; @Cauchemez2016]. A retrospective study shows that a model in which first trimester exposure is a cause of microcephaly in French Polynesia is a good fit for the data [@Cauchemez2016]. The best fit model implies a $1\%$ incidence of microcephaly resulting from maternal exposure in the first trimester of pregnancy, providing strong statistical evidence for Zika as a cause, though the total number of cases is only 7 during the critical period. Including only live births would yield an even lower incidence: one microcephaly case over background.]{}
6. [Zika is known to cause neurological damage in adults, typically leading to transitory paralysis, i.e. Guillian-Barre syndrome [@Cao-Lormeau2016; @Vogel2016].]{}
7. [A preliminary cohort report on outcomes of Zika infected pregnant women in Rio de Janeiro appears to provide strong evidence [@Brasil2016]. The study included 345 women, 134 Zika-infected pregnancies and 117 births. Multiple incidences of birth defects were determined either by ultrasound or at birth. Four cases of microcephaly were reported through ultrasonographic features. Two of these were associated with infections in the 5th and 6th months of pregnancy, inconsistent with both the French Polynesian analysis showing primarily early pregnancy effects and the absence of widespread cases in Colombia (below). Indeed, while 4 out of 23 pregnancies are reported as having anomalous ultrasounds when exposures happened in the first 4 months, 3 out of 9 (5 out of 9) pregnancies reported as infected in the 5th (6th or later) month have anomalies, including 2 still births in addition to the microcephalic ultrasounds [@error]. If similar exposure impacts were present in Colombia, a large number of anomalies, still births and microcephalic infants would have been reported very early during the epidemic there, and more by this time. The seemingly inconsistent two late term microcephaly cases were eventually reported as small for gestational age (SGA) rather than microcephaly at birth. While this might help resolve some of the discrepancy with other reports including Colombia, discounting the ultrasonographic findings leaves only one live microcephaly birth in the report. Finally, the Zika infected and small uninfected populations of the study are low/high income biased and therefore may also reflect environmental exposures of teratogenic agents associated with mosquito control that correlate with areas that have higher likelihood of infection. Perhaps most puzzling, the reported tracking of microcephaly through ultrasound during pregnancy in these two studies seems inconsistent with the absence of reports in other locations. Subsequent reports from Rio do not report nearly as high an incidence of microcephaly [@NECSINov].]{}
8. [A study of 32 microcephaly cases and 62 matched controls was carried out in 8 public hospitals of Recife, Pernambuco, Brazil between January and May 2016 [@Araujo16]. Immune factor and genetic (PCR) tests of Zika infection were performed. Of the microcephaly cases 13 (41%) of 32 cases and none of 62 controls had laboratory-confirmed Zika virus infection. This provides strong confirmation that Zika causes microcephaly, but provides very limited information about the fraction of microcephaly cases in Brazil that are caused by Zika. If the negative test results are considered valid, the fraction of microcephaly that is not caused by Zika is substantial and the small size of the sample makes the fraction value uncertain. The statistical error range is consistent with the $15\%$ reported in government tests (point 4 above and Table \[Brazilratio\]). Whether or not the negative results are conclusive is unclear, however this study shows that the best immunological and DNA tests available do not show that the large majority of microcephaly cases in northeast Brazil result from Zika infections.]{}
In summary, the central evidence that links Zika and microcephaly to individual cases is strong, but association to the large number of cases in Brazil is missing.
Analysis of geographic data on Zika and microcephaly
----------------------------------------------------
### Number of microcephaly cases by country
Zika infections have spread widely across Central and South America. Countries other than Brazil have reported a comparatively small number of cases ranging from a few to a few tens of cases [@WHO] (see Table \[sitrep\]). Early in the time of the spread, it was unclear what delay should be present between the incidence of Zika and microcephaly related to the month of pregnancy of exposure. The number of actual Zika infections is also difficult to identify. However, since the spread of Zika occurred widely during the fall of 2015 and winter of 2016, the number of cases by the end of 2016 should be compatible with population numbers and rates of infection. While these are difficult to obtain, it is apparent from a cursory examination of the WHO reported numbers of cases by country that there are no countries that rise to the level of the number of cases of microcephaly observed in Brazil, which has reached over $2,000$. Establishing the number of cases of Zika and the timing of those cases is needed to determine the relationship of Zika and microcephaly. We have performed a first analysis for the case of Colombia, the country that reports the second largest number of cases of microcephaly, as discussed in the following section.
Country/Territory Microcephaly Cases Probable Location of Infection
-------------------------- -------------------- --------------------------------
Brazil 2063 Brazil
Cabo Verde 9 Cabo Verde
Canada 1 Undetermined
Costa Rica 1 Costa Rica
Colombia 47 Colombia
Dominican Republic 10 Dominican Republic
El Salvador 4 El Salvador
French Guiana 10 French Guiana
French Polynesia 8 French Polynesia
Grenada 1 Grenada
Guatemala 15 Guatemala
Haiti 1 Haiti
Honduras 1 Honduras
Marshall Islands 1 Marshall Islands
Martinique 12 Martinique
Panama 5 Panama
Paraguay 2 Paraguay
Puerto Rico 2 Puerto Rico
Slovenia 1 Brazil
Spain 2 Colombia, Venezuela
Suriname 1 Suriname
Thailand 2 Thailand
United States of America 28 Undetermined
: Countries and territories that have reported microcephaly and/or central nervous system (CNS) malformation cases [@sitrep; @WHO]. While the WHO reports the microcephaly cases as associated with Zika infections, the association has not been shown in most cases.[]{data-label="sitrep"}
### Comparison of totals for Zika in Colombia and Brazil
Colombia reported a large number of Zika infections, but it has only seen a small number of Zika-associated microcephaly cases (see Fig. \[totals\]). The number of Zika infections reported in Brazil is roughly 200,000 while that in Colombia is 90,000. The number of confirmed microcephaly cases in Brazil now exceeds 2,000, while the number reported by the government in Colombia linked to Zika is only 57. Moreover, despite many cases of Zika in other parts of Brazil, the majority of microcephaly cases have been confined to the northeast region, which has a population of approximately 50 million, comparable to that of Colombia. While questions remain about reliability of tests and reporting, the extent of the inconsistencies is difficult to account for. Overall, the discrepancies suggest other causes or co-factors, rather than Zika itself, are the primary source of microcephaly in Brazil. The timing of Zika and microcephaly cases in Colombia is discussed in the following section.
![Reported cases of microcephaly and of Zika in Brazil and Colombia. A. Cumulative reported cases of Zika in Brazil and Colombia. B. Total microcephaly cases reported in Brazil and Zika associated microcephaly cases reported in Colombia. The number of Zika cases in Colombia is lower by a factor of 2, while the number of microcephaly cases is lower by a factor of 50. (Brazil reports total microcephaly numbers and does not distinguish those linked to Zika. Colombia reports only Zika-linked microcephaly cases. The historical background rate of microcephaly in Colombia is 140 per year.)[]{data-label="totals"}](zika_micro_colombia_growth_1.png){width="17"}
### Analysis of Zika caused microcephaly in Colombia
We construct a model of the time period of infection during pregnancy that results in microcephaly in Colombia based upon available data on Zika and microcephaly (Fig. \[fig:Colombia\]). We find that of the 56 confirmed microcephaly cases reported as of epidemiological week 42 [@bul], only about 45 cases can be attributed to Zika infections while 11 are due to the background effect of random cases of microcephaly that by coincidence occur in Zika infected pregnancies. Colombia is not reporting the number of microcephaly cases that are unlinked to Zika.
![Comparison of reported cases in Colombia with background and Zika causal models. Reported cases of Zika and microcephaly (red dots) are compared with expected number of background cases due to coincidence of microcephaly with Zika infections at a rate of 2 per 10,000 births (gray) and two models of Zika-induced microcephaly suggested by the study in French Polynesia [@Cauchemez2016]. The first assumes that pregnancies infected in the first trimester have a 1:100 rate of microcephaly; the second assumes that pregnancies infected in the first and second trimester have a 1:200 rate of microcephaly. The data is consistent with just background cases until the report of epidemiological week 24, ending June 18. The data in weeks 24 through 27 is reasonably consistent with the first trimester model, but not with later weeks. The overall rate is much lower than that consistent with the number of cases in northeast Brazil.[]{data-label="fig:Colombia"}](ZikaMicrocephalyColombia11012016.png){width="16"}
The Colombian outbreak of Zika began in August of 2015 and the number of infections increased rapidly in early 2016. The outbreak diminished and was declared over by July, 2016 [@ColombErad]. Many of the exposed pregnancies should result in births in the second and third quarters of 2016. We performed a quantitative analysis based upon an extensive preliminary cohort study published on June 15, 2016 in the New England Journal of Medicine [@ColombiaNEJM]. The study reports results of women infected until March 28, 2016, whose pregnancies were observed until May 2, 2016. The study tracked 1,850 women, whose date of infection with Zika is known relative to the start of the pregnancy. Of these, 532, 702 and 616 were infected in the first, second and third trimesters respectively. 16%, 29% and 93% (85, 204 and 583) of the pregnancies concluded. No cases of microcephaly were observed. The total number of pregnancies with Zika infections is much larger, with 11,944 cases with Zika symptoms being observed in clinical settings. At the time of the article, no cases of microcephaly occurred in any of nearly 12,000 pregnancies. However, the report cites 4 cases of microcephaly with Zika in the general population that did not report any Zika symptoms, these cases being reported prior to April 28. This finding implies that there are many more unreported, asymptomatic, cases of Zika infection. Since there is less than a 1 in 12,000 incidence of Zika until this point in the epidemic, there should be at least 4 times as many infected individuals that do not have symptoms in order for there to be 4 microcephaly with Zika cases—a ratio of total and symptomatic cases of 5:1—for a total of 60,000 Zika cases ($5 \times 12,000$). There is other evidence for underreporting of Zika. For example, the incidence among women is twice that reported in men. Women at home may be infected at a higher rate, or reporting is focused on women because of the concern about Zika’s maternal effects.
It is useful to have a reference model of Zika as a cause of microcephaly even though there is no consistency between different observational studies. For this purpose, we adopt as a reference a model that comes from the study reported in French Polynesia [@Cauchemez2016], which provided evidence that 1 in 100 pregnancies exposed in the first trimester, or, alternatively, 0.5 in 100 of all pregnancies exposed in the first and second trimester, resulted in microcephaly. This study was based upon a small number of cases and ultrasound detection rather than births. Seven cases of microcephaly were reported above background, but only one of them was a birth; the others were detected by ultrasound. The high rate of ultrasound detection is not consistent with observations by ultrasound of microcephaly in other countries, including Colombia. We use the model for French Polynesia as a reference model for comparison with other data in order to clarify which data is or is not consistent with that model. We note that this rate is about 100 times larger than the minimum reported background rate for microcephaly, 2 in 10,000 [@cdc1] if all pregnancies are infected. The Zika-induced cases should be considered to be in addition to the background cases, which are due to other causes. It should be emphasized that the French Polynesia model is not necessarily compatible with the observed cases in Brazil; and has not been adequately validated in any context including French Polynesia.
We construct a model of the Zika-infected pregnancies by considering each pregnancy to have a uniform probability of infection across 39 weeks. This enables us to estimate the total number of Zika-infected pregnancy births as well as the number that are born after exposure in the first trimester or in the first and second trimesters. The total number of cases should be a combination of those with Zika exposure and background cases. For background cases, any birth has a probability of 2 in 10,000 of microcephaly. If a Zika infection occurred anytime during pregnancy, it would be a Zika and microcephaly case at birth.
Fig. \[totals\] shows reported cases of microcephaly linked to Zika infections (red dots). These are compared with background cases predicted based on the number of Zika-infected pregnancies [@ColombiaNEJM] and two models of Zika as a cause of microcephaly that originate from the outbreak in French Polynesia. The cases of Zika and microcephaly reported until June 11, 2016 are consistent with the expected background rate of birth defects that would have occurred in those infected with Zika, even if Zika were not a cause [@r1]. After that date, cases initially tracked a trajectory matching a model of 1% of pregnancies infected in the first trimester developing microcephaly [@r2]. However, the number of cases subsequently plateaued and increased in stages but more slowly than expected from the model. As of epidemiological week 42, a total of 56 cases were reported [@bul], 11 of which can be explained as background cases unrelated to Zika. This is much less than the 155 total cases predicted from 1% of first trimester pregnancies along with the background rate. It is also far less than the over $2,000$ cases reported in Brazil. Based on this data, approximately $0.23\%$ of first trimester infections ($0.075\%$ of infected pregnancies) in Colombia have resulted in microcephaly. The coincidence of increased microcephaly cases with the timing of births infected in the first trimester suggests Zika is responsible for a limited number of microcephaly cases. This is consistent with other reports that link Zika with a few cases of microcephaly, as has also been shown with other viral infections, but is not consistent with Zika being the cause of the majority of cases reported in northeast Brazil. Details about the construction of the model for Zika-caused microcephaly are provided in Figs. \[Pop\] and \[Pregnancies\]. A similar model has previously been applied to a single municipality of Bahia [@Physicians2016], without discussion of comparisons with other areas. Other simulations of Zika as a cause of microcephaly also do not achieve numbers consistent with those reported [@Zhang2016].
![Population of Zika infected pregnancies. To model the Colombian Zika and microcephaly epidemic, the reported number of symptomatic cases per week until March 28 (green line) [@ColombiaNEJM] is normalized by the number of reported Zika infected pregnancies (11,944, shaded purple) and multiplied by 5 to obtain the total number of symptomatic and asymptomatic cases (red shading), due to the observation of four Zika and microcephaly cases that do not have Zika symptoms. While we don’t use it directly, the total number of Zika infections can be estimated (blue line) by similarly multiplying the number of reported cases, correcting for the bias in reporting between women and men, assuming the infection rate is comparable. Other assumptions about the total number of cases do not affect the results reported here.[]{data-label="Pop"}](ZikaColombiaPop){width="15"}
![Pregnancies during a particular week (vertical axis) present during a particular epidemiological week (horizontal axis). (A) Total number of pregnancies, (B) the number of pregnancies exposed in the first trimester, and (C) the number of pregnancies exposed in either the first or second trimester. A uniform exposure by week of pregnancy is assumed. The number of births in each category is approximately the number that crosses the 39/40 week boundary (horizontal gray line), but it is more accurately given by weighting them according to the distribution of births from week 35 through 43.[]{data-label="Pregnancies"}](MicrocephGraphs.jpeg)
### Brazil microcephaly and Zika geographic data
In addition to the differences between Colombia and Brazil, questions have been raised about the geographic distribution of microcephaly cases across Brazil. Much of the original data on the Zika infection did not include the numbers of individuals infected and only identified whether or not a state was infected by Zika. The mild symptoms of the infections in many individuals make precise counting difficult. Reporting was not required of states by the national government until 2016. However, since the beginning of 2016 better reporting is available, though the reliability of the numbers can be challenged. Nevertheless, informal reports suggest that even without precise numbers the existing data are difficult to reconcile with expectations based upon Zika as a cause of microcephaly: The spread of Zika across different states of Brazil has not been accompanied by a comparable wave of microcephaly [@naturebutler; @McNeil2016; @Phillips2016].
Quantitative comparisons are made easier by the analysis of Colombia which shows that the cases of Zika-caused microcephaly occur only due to infections in the first trimester, narrowing the window of delays between exposure and case reports of microcephaly. We can therefore compare reports of Zika infections with microcephaly reports 33 weeks later to identify the potential causal relationship between Zika and microcephaly in different states of Brazil. Zika and microcephaly cases reported in six Brazilian states in 2016 are shown in Fig. \[logscale\] along with an indicator of the time difference of 33 weeks.
We see that the ratio between Zika and microcephaly reports varies between $1$ and approximately $1/1000$. That it is possible to have a ratio of about $1$ is surprising if one views Zika as a cause of microcephaly. The number of pregnancies with Zika infections should be much less than the number of Zika cases. To estimate the number of Zika infected pregnancies in the first trimester at a particular time we would multiply the number of reported cases by an underreporting factor of 5 [@r1] to obtain an estimate of the actual number of cases, and multiply by the fraction of the population that is pregnant in the first trimester at any time, $0.37\%$ (birthrate per day times 90), so the number of susceptible pregnancy Zika infections would be about $1.8\%$ of the number of Zika cases. This number is smaller than the number of microcephaly cases in several states.
As a preliminary upper bound on the number of Zika induced microcephaly cases we might consider the incidence from Colombia of $0.075\%$ of pregnancies and calculate the number of cases that would be present if everyone was infected by Zika. For Pernambuco, with a population of $9.3$ million and an approximate birth rate of $15$ per $1,000$, the number of microcephaly cases would be approximately $105$ a year, much less than the actual number $386$.
Moreover, the inconsistency among the states is independent of any calculation of the rates of the number of susceptible pregnancies. The wide range of values suggests that Zika is not the primary cause of microcephaly. All calculations are sensitive to the possibility that reporting is poor and inconsistent across states and countries. However, it remains difficult to identify a way to reconcile the extent of the inconsistency across Brazil.
![Cumulative Zika (blue) and microcephaly (red) cases over time in five northeast Brazilian states and the state with the largest number of cases elsewhere, Rio de Janeiro (log scale). Blue and red dots are separated by 33 weeks, the expected delay between first trimester infections and expected microcephaly births caused by them. The differences in ratios in different states (Fig. \[ratio\]) suggests that Zika is not the cause of microcephaly.[]{data-label="logscale"}](zika_micro_log_1){width="16"}
Expanding the discussion to all states of Brazil, Fig. \[peaks\] shows confirmed cases of Zika (blue) and microcephaly (red) for Brazilian states. The widely-varying relative proportion of Zika and microcephaly is apparent in the multipliers used to show the microcephaly data on the same vertical scale. Fig. \[ratio\] shows the ratio between Zika and microcephaly cases as a function of time (including the 33 week delay) for all Brazilian states. As the figure shows, ratios vary widely across the country, but are higher than the small proportion of government reported cases in Colombia.
Table \[ratiomicrozika\] shows the ratio of microcephaly cases to Zika cases reported 33 weeks previously for all Brazilian states, including cases as of October, 2016, and the maximum over the year. In the northeastern states, only Bahia is reporting a number of cases consistent with Zika as a primary cause of microcephaly. Interestingly Rio de Janeiro, the only state outside the northeast that has more than 100 cases, is also consistent with Bahia (though the ratio was higher at earlier times). Other states outside the northeast have too few cases to reliably compare.
We note that a previously-published report on the total number of Zika and microcephaly cases across all states, in effect, inappropriately linked the cases of Zika reported for Bahia with the cases of microcephaly reported for Pernambuco [@Faria16].
![Weekly Zika (blue) and microcephaly (red) reports over time in each Brazilian state. Note the multipliers for the microcephaly numbers. If Zika is the cause of the cases of microcephaly a delay of about 33 weeks should be seen between peaks of the former and latter. This appears to be the case for Bahia, Ceara, and Alagoas for early peaks of Zika and later peaks of microcephaly. We note that if seasonal use of insecticides coincides with outbreaks, then the cause may also be those insecticides. A filter {0.25, 0.5, 0.25} has been applied to smooth the data.[]{data-label="peaks"}](zika_micro_lineal_diff_all){width="14"}
![Ratio of microcephaly cases to Zika cases 33 weeks earlier as a function of time for all Brazilian states. The horizontal green line represents a ratio of $1\%$ and the purple horizontal line represents a ratio of $0.05\%$. To obtain the number of microcephaly cases per first trimester Zika infected pregnancy (rather than all Zika cases) we would have to include both unreported Zika cases and multiply by the proportion of pregnancies, multiplying the ratio by a factor of 25. Differences between rates would remain.[]{data-label="ratio"}](zika_micro_ratio){width="12"}
[|X|X|X|X|X|X|X|X|]{} Region & State & Current Zika & Current Microcephaly & Current Ratio & Zika at Maximum Ratio & Microcephaly at Maximum Ratio & Maximum Ratio\
Northeast & Pernambuco & 408 & 386 & 0.95 & 5 & 334 & 67\
Northeast & Bahia & 59560 & 312 & 0.0052 & 27290 & 268 & 0.0098\
Northeast & Paraiba & 1547 & 166 & 0.11 & 140 & 155 & 1.1\
Northeast & Rio Grande do Norte & 1295 & 138 & 0.11 & 264 & 123 & 0.47\
Northeast & Maranhao & 1419 & 138 & 0.097 & 43 & 126 & 2.9\
Northeast & Ceara & 906 & 137 & 0.15 & 431 & 136 & 0.32\
Northeast & Piaui & 59 & 99 & 1.7 & 3 & 89 & 30\
Northeast & Sergipe & 384 & 123 & 0.32 & 2 & 26 & 13\
Northeast & Alagoas & 1766 & 84 & 0.048 & 504 & 79 & 0.16\
Southeast & Rio de Janeiro & 31542 & 130 & 0.0041 & 412 & 87 & 0.21\
Southeast & Espirito Santo & 1727 & 23 & 0.013 & 23 & 6 & 0.26\
Southeast & Sao Paulo & 1779 & 26 & 0.015 & 88 & 10 & 0.11\
Southeast & Minas Gerais & 7539 & 8 & 0.0011 & 12 & 3 & 0.25\
Center & Mato Grosso & 16680 & 47 & 0.0028 & 17 & 35 & 2.1\
Center & Goias & 1721 & 24 & 0.014 & 2 & 14 & 7.0\
Center & Mato Grosso do Sul & 370 & 20 & 0.054 & 1 & 2 & 2.0\
North & Tocantins & 1712 & 18 & 0.011 & 23 & 11 & 0.48\
North & Roraima & 62 & 10 & 0.16 & 21 & 10 & 0.48\
North & Amapa & 59 & 9 & 0.15 & 1 & 8 & 8\
North & Rondonia & 599 & 7 & 0.012 & 2 & 4 & 2.0\
North & Para & 1139 & 1 & 0.00088 & 47 & 1 & 0.021\
South & Rio Grande do Sul & 123 & 10 & 0.081 & 4 & 5 & 1.3\
South & Parana & 1014 & 4 & 0.0039 & 10 & 4 & 0.40\
DF & Distrito Federal & 201 & 8 & 0.040 & 11 & 6 & 0.55\
### Analysis of Bahia, Brazil
Compared to other states in northeast Brazil, Bahia reports the lowest proportion of microcephaly cases relative to Zika cases and has the second highest number of microcephaly cases. We might speculate that the reported number of Zika cases may be more reliable and a higher proportion of cases may be caused by Zika. In August 2016, we analyzed the data available for the state of Bahia \[Fig. \[micro\]\]. We observed that Bahia has a peak of microcephaly whose timing relative to a peak in Zika appears to be consistent with Zika as a cause. Nevertheless, our analysis suggests that the incidence of microcephaly is much higher than in either French Polynesia or Colombia. We compare three numbers: the number of confirmed cases, 263, the number of suspected cases, 1154, and the number confirmed as having both Zika and microcephaly, 41. The first two are reported from Bahia, while the last is taken from the national ratio of confirmed Zika and microcephaly cases as a fraction of confirmed microcephaly cases, $15.6\%$.
![Comparison of Zika and microcephaly cases in Bahia, Brazil. We compare reported microcephaly cases with or without symptoms of Zika (red dots) with the projected number of cases for susceptibility up to weeks 8, 10, 12, 14, 16 (shades of darker blue) with microcephaly in 84%, 71%, 63%, 56%, 49% of the Zika pregnancies (optimized fit), based upon the reported number of Zika infections (green shows estimated Zika infected pregnancies per week of gestation). The left green peak gives rise to the central red/blue peak due to the delay between first trimester infection and the births. The inset shows the goodness of fit for different numbers of weeks in which infections cause microcephaly. The best fit overall occurs for susceptibility up to 16 weeks, but the initial rise is more consistent with 8 weeks, suggesting that the Zika-induced microcephaly cases are likely due to infections approximately during the second month of pregnancy rather than the first trimester as a whole.[]{data-label="micro"}](zikvBANew.png){width="14"}
According to the data that are available, the fraction of first trimester pregnancies exposed to Zika that have confirmed microcephaly is $63\%$, the number of suspected cases is $289\%$ of first trimester Zika-exposed pregnancies (which in principle is not inconsistent with a first trimester model of confirmed microcephaly cases), and the fraction of pregnancies exposed to Zika that are confirmed to have both Zika and microcephaly is $9.8\%$. All of these values are significantly larger than the $1\%$ rate obtained from French Polynesia and used above to model Colombia. We note that the population of Bahia is about 10 million, or 1/5 of the total population of Colombia. The relative reliability of Zika case reporting is unclear. In order for the microcephaly fraction to correspond to the $1\%$ first trimester model of French Polynesia, the number of Zika cases would have to be underreported in Bahia by a factor of 63. A large discrepancy in the Zika reporting rates would be necessary if Zika is consistently causing a certain percentage of microcephaly. Our results are dominated by the analysis of Zika reporting in 2015, yet reporting only became mandatory in 2016. In any case, the analysis shown in Figs. \[states\] and \[states1\] suggests that the second peak of the Zika epidemic would be expected to give rise to a new set of microcephaly cases.
The seemingly large discrepancy between microcephaly counts in Colombia and Bahia echoes discrepancies between different parts of Brazil that led to questions about whether there are additional factors that affect the microcephaly rates [@naturebutler].
![Plots of Zika and microcephaly cases in Brazil. Plots (log scale) are shown for eight states of Brazil: Pernambuco (PE), Bahia (BA), Paraiba (PB), Ceará (CE), São Paulo (SP), Rio de Janeiro (RJ), Mato Grosso (MT), and Rio Grande do Norte (RN). Bahia has the highest counts of Zika relative to microcephaly six months later when births of pregnancies exposed in the first trimester are expected.[]{data-label="states"}](zika_micro_lineal_log.png){width="16"}
![Same as Fig. \[states\], but using a linear scale.[]{data-label="states1"}](zika_micro_lineal_1.png){width="16"}
[|X|X|X|X|]{} State & Confirmed Zika & microcephaly & Confirmed microcephaly & Suspected microcephaly\
Bahia (BA) & 9.8% & 63% & 289%\
Rio de Janeiro (RJ) & 318% & >1000% & >1000%\
Ceará (CE) & 349% & >1000% & >1000%\
Rio Grande do Norte (RN) & 644% & >1000% & >1000%\
Paraíba (PB) & >1000% & >1000% & >1000%\
Pernambuco (PE) & >1000% & >1000% & >1000%\
São Paulo (SP) & >1000% & >1000% & >1000%\
Mato Grosso (MT) & >1000% & >1000% & >1000%\
### Ultrasound versus birth identification
As part of the analysis of cases in Colombia and Brazil and other locations, it is worth noting that there has been an inconsistency among medical reports about the rate at which ultrasound detects Zika induced birth defects including microcephaly. There may be many reasons for such inconsistencies but an easy resolution of them is lacking. The French Polynesia study was based primarily upon cases detected by ultrasound. This is inconsistent with the reports from Colombia where only one case has been officially reported [@ColombiaNEJM]. A preliminary cohort study report in Rio [@Brasil2016] reported a $29\%$ rate of fetal abnormalities among Zika infected pregnancies including $10\%$ microcephaly cases based upon ultrasound findings of all exposed pregnancies including those in the third trimester. It is hard to reconcile this study with either the French Polynesia study, with only approximately $1\%$ microcephaly for first trimester exposures, or reports from Colombia. Indeed it is hard to understand how this study can be made consistent with other general findings given the rate of birth defects, still births and microcephaly. Most essentially, it is unclear why ultrasound-based observations of microcephaly in Colombia have not been more widely-reported given the 12,000 confirmed Zika cases there as of March 28, 2016.
### Confirmed cases of Zika and microcephaly in Brazil
One of the central questions about the role of Zika in Brazil is the absence of confirmation of cases of microcephaly as having Zika infections in government reporting (see Table \[Brazilratio\]). The tests being performed are not sufficiently specified, and their reliability is unknown. Reports for Brazil have not been provided in recent months. While DNA tests performed sufficiently long after an infection might not be positive due to clearing of virus, it is unclear why immunological tests would not yield higher rates of confirmation. The immunological tests are likely to have false positives due an inability to fully distinguish Zika from Dengue and, perhaps, Chikungunya infections. On the other hand if it is assumed that Zika is the cause, then the $15.6\%$ confirmation rate implies there are many false negative results, the reason for which are not discussed. Absence of evidence of Zika infection may therefore be an indication that Zika is not the cause. What has not been observed are high rates of confirmation of Zika infection that would yield strong confidence that Zika is the cause of most cases.
The absence of confirmation of Zika infections in Brazil is particularly challenging given the results of the French Polynesia study [@Cauchemez2016]. In that study, the population had seropositive results in $66\pm4\%$ ($95\%$ confidence interval) at the end of the epidemic. Thus there is very limited possibility of undetected cases there, suggesting the serological tests have high enough reliability to consider the negative results in Brazil as indicating that those cases are not the result of Zika infections. (We note that to the limited extent that the actual Zika infection rate is higher in French Polynesia, the incidence rate of microcephaly per maternal infection would have to be reduced.)
[0.8 ]{}[|X|X|X|X|]{} Date & Confirmed Zika and microcephaly & Total microcephaly & Ratio\
February 12, 2016 & 41 & 462 & 8.9%\
March 1, 2016 & 82 & 641 & 12.8%\
March 29, 2016 & 130 & 944 & 13.8%\
July 20, 2016 & 267 & 1709 & 15.6%\
July 27, 2016 & 272 & 1749 & 15.6%\
### Additional unconfirmed microcephaly cases in Colombia
When considering the implication of the Brazilian microcephaly count for the Colombian ones, we note that the models in Fig. \[micro\] consider all Zika-induced microcephaly cases as being confirmed to have Zika infections. Ignoring the serological tests in French Polynesia, we might consider speculatively that there are many more microcephaly cases whose Zika infections are undetected as they were not in Brazil. If the confirmation rate is similar to the $15.6\%$ found in Brazil, we would multiply the number of confirmed Zika induced microcephaly cases by a factor of $6.4 = 1/15.6\%$ to obtain the total number of Zika induced microcephaly cases. We would then have to increase the rate of microcephaly by this factor, which would make the Colombian results have a total of approximately $6.4*45=288$ cases, which would be a higher rate than the French Polynesian results (raising the question as to why many more cases were not observed in French Polynesia), but somewhat more consistent with the much higher microcephaly rates in Bahia. The additional cases (of order 300) should also be observed in microcephaly cases above background, but without evidence of Zika infection.
Alternatively, if we use the estimated rate of microcephaly induced by Zika from French Polynesia of 1 in 100 pregnancies exposed in the first trimester [@Cauchemez2016], there should be 200 microcephaly cases arising from Zika exposure of pregnancies infected until March 28. However, only about 1 in 6.4 of these, or 30, would also have confirmed Zika infections, if this fraction follows the pattern in Brazil. Reports of confirmed microcephaly cases not linked to Zika are not being published by the Colombian authorities, though we might consider the possibility that they have been included in the cases that were “discarded" or “under investigation" [@bul].
On the other hand, if we use the estimates of rates from Bahia, then the number of cases of Zika and microcephaly should rise to 400 or 2,400 (considering symptomatic or total number of exposed pregnancies respectively), and the number of total microcephaly cases to 2,520 or 12,600, respectively. Since this is the lowest rate in northeast Brazil, it is apparent that there is need for additional studies that can resolve the discrepancies between French Polynesian, Colombian, and Brazilian rates.
### CDC report on Colombia Dec 16, 2016
As this paper was being completed, the CDC released a new report on microcephaly in Colombia [@Cuevas2016]. New data is provided for the incidence of microcephaly, Zika and the combination of microcephaly and Zika. The results are inconsistent with those reported during the year by the Colombian Ministry of Health. There is an increase of confirmed Zika and microcephaly cases to 147 until epidemiological week 45 which is much higher than the 58 reported by the government until the same week [@bul]. While the government reports did not include overall microcephaly rates, the CDC study states there are 476 cases from weeks 5 through 45, a substantial increase from the previous year’s 110. The analysis primarily compares 2016 results with 2015 microcephaly background data, but does not consider the potential increase that might result from changes in methodology and scrutiny as a result of the public health emergency due to the epidemic in Brazil. The study acknowledges the difficulties in definition of microcephaly and their potential impact on the counts. Unfortunately, the report does not address the inconsistencies or methodological differences that are present between their numbers and the previously reported values. Since the study is authored in part by physicians associated to the government health ministry, the reasons for the discrepancies are unclear. Moreover, the focus of the paper is on showing that Zika causes microcephaly, and not on the rate at which it does so per maternal Zika infection. The latter is critical for a comparison with Brazil to determine if Zika could be the cause of the majority of cases there. In particular the analysis reported is on a per capita basis and not related to the number of Zika cases. The increased numbers of microcephaly and Zika and microcephaly cases imply a higher rate of Zika induced microcephaly than previous numbers. Hence, the new study increases the possible range of Zika caused microcephaly rates. However, it is unclear whether they are consistent with the high rates of microcephaly per Zika infected pregnancy in northeast Brazil, or with those of French Polynesia. Moreover, absent a study with similar methods in Brazil, it is unclear how to compare government reports in Brazil with the methods that were applied by this study.
The study provides information that is relevant to other aspects of the discussion of microcephaly cases in Colombia. It includes in its count of cases spontaneous abortions, pregnancy terminations and still births that are linked to microcephaly and reports that all of them constitute 44 of the 476 cases reported. This indicates that intentional pregnancy terminations cannot account for a substantial number of missing microcephaly cases as has been suggested in a number of publications, e.g. [@McNeil2016], and as reported for French Polynesia [@Cauchemez2016]; moreover the rate of still births is widely discrepant from the cohort study in Rio [@Brasil2016].
One of the interesting results that is provided is a count of microcephaly cases by department (subnational political units) in Colombia. It is stated that the rates are higher than previous year numbers in areas that do not have propagating Zika infections, above 2,000 meters in altitude. For example the capital region and largest city, Bogota, has a rate of 5.5 per 10,000 births, compared to the previous year country value of 2.1, an increase by 3.4, for a factor of 2.75 increase. A suggestion is made in the paper that the increase might arise from travel or sexual transmission (presumably from traveling male partners). However, in order for this effect to quantitatively account for the observed increase, a large percentage of fathers and pregnant mothers would have to travel to high Zika prevalence areas of the country and over a substantial fraction of the first trimester of pregnancy. The coastal tourist area of Cartagena has a rate of 10 per 10,000 births—an increase above background of $7.9$. We can calculate the fraction of population and the duration of visits residents of Bogota would have to visit Cartagena (or equivalent) for the observed increase in Bogota assuming a consistent rate of transmission for tourists and people who live there. The product of population fraction traveling and duration of their stays would have to be 0.43. So, if 50% of the pregnant couples of Bogota went to Cartagena they would have to be there for 86% of their first trimester or all but 12 days. If only the partner went there, the rate would have to be multiplied by the rate of sexual transmission. Such a travel schedule is highly improbable even under normal circumstances and even more so after the of Zika infection outbreak and its location became known. We note that the 2015 rate of microcephaly is not reported by department, nor is the rate of Zika infection of microcephaly cases in 2016. We use the country rate as a reference throughout, though the background rate might have been higher in specific departments.
It is more reasonable to suggest that the increase in Bogota and other parts of the country at high altitude (especially Cundinamarca, the area surrounding Bogota, together constituting 21% of all Colombia births), resulted from increased scrutiny and methods that brought background rates close to the median 6.6 prevalence in 17 U.S. states [@Cragan2014]. Taking the combined rate in Bogota and Cundinamarca as the reference background would give a 5.7 per 10,000 rate, see Fig. \[ColombiaElev\]. This would give 283 total background microcephaly cases for the country. Subtracting this from the reported value, 476, yields a number of Zika caused microcephaly cases of 193. This is not far from the rate obtained by taking the 147 cases that are positive for both Zika and Microcephaly and multiplying by the proportion of 476 untested to 306 tested microcephaly cases, an estimated 228 Zika positive microcephaly cases. This value has to be corrected for background cases that coincidentally have Zika infections due to high prevalence of Zika in the population. The difference between 228 and 193 suggests that about 35 of the cases are due to coincidence which is $35/283= 12\%$ prevalence, or 60,000 infected pregnancies. This is 3 times the number of reported Zika infected pregnancies in the report [@Cuevas2016]. Previous estimates have suggested that due to the low rate of reporting of Zika symptoms to Zika infections the rate is 5 times as much, which would increase the coincidence number to 58. Taking this as a range we can summarize the results as suggesting between 160 and 193 Zika caused microcephaly in Colombia. This is about four times the estimates that were obtained from the government reports.
![Microcephaly versus elevation by Colombian department. Each point represents the reported rate of microcephaly per capita (not per Zika infection) among departments of Colombia versus the elevation of their most populace city. Only departments with more than 10,000 births are shown and color of points reflect the total number of births in the department (scale on right). According to the report [@Cuevas2016] high elevation cities (above 2,000 m) do not have self-propagating Zika infections. This includes Bogota and its suburbs which constitute two departments, Bogota and Cundinamarca, at $2,640$ m. The observations suggest a background (non-Zika associated) rate of $5.7$ per $10,000$ births in $2016$ given the study’s methodology. The only department consistent with the previous year’s background rate of $2.1$ is that of Nariño, whose most populous city Pasto is at $2,527$ m. The rate observed, $2.2$, results from only 3 cases in $14,000$ births, so that statistics are very limited compared to the 60 observed in the Bogota area.[]{data-label="ColombiaElev"}](ColombiaElev_1.png){width="12"}
Such an increase would bring the numbers in Colombia in better agreement with northeast Brazil. Still, the per capita numbers even in the high infection rate regions of Colombia (see figure) do not reach those in northeast Brazil. Absent a comparison that includes Zika infection rates, the CDC study provides new evidence that brings the rate in Colombia closer to those in Brazil but additional information is needed to demonstrate that it is indeed the primary cause. Discrepancies between different measures and positive / negative test results must still be reconciled.
Finally, we note that the CDC report makes several other hard to understand suggestions in their discussion: (1) In considering the number of microcephaly cases that are reported, the paper suggests that the numbers should be considered in relation to a possible reduction in birthrate due to recommendation to delay pregnancy by authorities. However, the reduction in births from 2015 to 2016 is 18,000 or 3.5%, which would only have an impact on overall microcephaly numbers by such a percentage, about 4 cases; (2) The paper tries to explain the relative increase of number of cases in Brazil and Colombia compared to background multiplicatively (i.e. the ratio of incidence per 10,000 is different both for background observation and for 2016 observations leading to different ratios). However, the effect of Zika on microcephaly rates should be additive (after multiplying by the rate of Zika infections) rather than multiplicative.
To determine the rate of microcephaly per infected pregnancy the number of Zika infected pregnancies must be identified. How many of the microcephaly cases were originally among those identified as Zika infected pregnancies is not reported in the study. Prior studies by the same group reported four cases of microcephaly that were not originally identified as Zika infected pregnancies [@ColombiaNEJM]. If we consider the ratio of total and symptomatic cases to be 5:1, and assume 1 in 3 of pregnancies were exposed in the first trimester, and that all pregnancies have terminated, then the number of additional reported microcephaly cases over background is $366$ in $33,000$ or $1.1\%$. Perhaps coincidentally, this is consistent with the reported rate of microcephaly in French Polynesia. Among the significant differences between the studies were the large proportion of abortions relative to live births in the former, which was not present in Colombia. Were we to ignore the asymptomatic cases, then the rate of microcephaly to Zika infection in Colombia would be five fold the number of cases in French Polynesia, again raising the question as to why other cases were not observed there. If we restrict the number of microcephaly cases to those with positive tests for Zika, allowing the rest to be increased reporting due to enhanced scrutiny, we have a rate of $0.44\%$ per first trimester infection.
More importantly for our discussion, the improvement of agreement between Colombia and French Polynesia, and the increase in number of Zika confirmed microcephaly cases from approximately $50$ to $147$, reduces but does not eliminate the discrepancy between the rate of microcephaly between Colombia and northeast Brazil and among different states of Brazil. Indeed, the reported confirmation of about $50\%$ of tested cases of microcephaly as having evidence of Zika infection stands in contrast to the approximately $15\%$ reported in July in Brazil (see Table \[Brazilratio\])—consistent with the suggestion that Zika is a cause of only a small fraction of the total cases. We note that if there is another cause of microcephaly in Brazil, even the cases that have evidence of Zika infections would include coincidental co-occurence rather than a necessarily causal relationship, which is not considered in the CDC or other reports.
In summary, without clarity about differences of methodological aspects of the studies, it is hard to determine what is actually known about Zika and microcephaly in Colombia, leading to highly uncertain conclusions. Increased uncertainty should not be taken to be evidence for Zika as a cause, but rather a need for additional studies that can identify what is the cause. The question is not whether Zika is a cause (which it is, at some rate), but whether it is the cause of the large number of cases in Brazil. A conclusive answer is not yet available and inconsistencies are unlikely to be resolved until this question is directly addressed.
### Summary of evidence on Zika and microcephaly
Since the high incidence of microcephaly in Brazil and the spread of Zika virus to other countries, the international community has been waiting to see if an accompanying spread of microcephaly would also occur. After months of uncertainty, the government published data strongly indicates that Colombia will not be seeing a comparably large number of Zika-related birth defects. A new CDC report just released on Colombia implies a larger rate of microcephaly in Colombia due to Zika infections than previously reported, but does not demonstrate consistency with the rate in Brazil due to the absence of reliable Zika infection numbers. Either definite discrepancy or persistent uncertainty suggests a need to reexamine conditions in Brazil, particularly in the northeastern states which saw the majority of microcephaly cases. If Zika alone is not enough to cause large numbers of birth defects, some other factor or factors unique to Brazil are present. Recently several reports have suggested that co-factors are responsible [@naturebutler; @McNeil2016; @Phillips2016]. A cofactor would be one in which Zika is the cause, but the presence of some other environmentally present substance increases the susceptibility. It is also possible that even without Zika infections another cause is responsible if its presence began at about the same time. One possibility is the pesticide pyriproxyfen, which has been largely dismissed as a potential cause or cofactor despite being insufficiently studied [@Parens2016]. While authorities claim toxicology studies are sufficient to rule it out as a cause, they surely would not rule it out as a cofactor. Thus the absense of engaging in the possibility of its contribution to the microcephaly by health authorities is inconsistent with the scientific evidence.
While the total number of microcephaly cases remains low in Colombia and other countries, cases in Brazil continue to rise at the rate of 100 affected births per month (see Fig. \[totals\]). Since pyriproxyfen may be playing a role in Brazil’s disproportionate increase of birth defects, rapid policy action is needed to replace its use as a pesticide until its effects can be more thoroughly studied.
Recommended further studies
===========================
From our research and analysis, we have found many gaps in the available scientific and public health literature. Thus, we suggest areas of study that would contribute to our understanding of the complex public health issues associated with pyriproxyfen and Zika as causes of microcephaly:
- [Microcephaly definition and counts. One of the central challenges of studies and country data is the definition of microcephaly and the counting of its cases. In order to achieve consistency among studies it is necessary to go beyond considering a single definition and number. What is needed is a reporting of how the number of cases changes as the threshold cranial circumference is varied. This distribution is a much more robust measure than a single number.]{}
- [Analysis of the geographic and time dependence of the use of pyriproxyfen in Brazil in relation to microcephaly occurrence.]{}
- [Observations of the concentration of pyriproxyfen and its breakdown products in containers of drinking water, and in tapped water, over time after administration.]{}
- [Animal studies of pyriproxyfen and its metabolites. The existing studies upon which the approval of pyrirproxyfen are based are statistically weak and assumptions used are not well justified. Moreover, additional studies should be done on breakdown products of pyriproxyfen in water, with and without sunlight, which may have distinct effects.]{}
- [Retinoid X receptor binding/ligand tests for pyriproxyfen and its biological and environmental metabolites.]{}
- [A more systematic understanding of the effects of juvenile hormone analogs and their toxicity across mammalian models, including the effects of genetic variation and multiple potential mechanisms of action (retionid X receptor, thyroid mechanisms, etc).]{}
- [Tests of reliability of traditional toxicological test assumptions for hormones and other regulatory molecules.]{}
Conclusion
==========
This paper analyzes the potential causal connection between the pesticide pyriproxyfen and microcephaly, as an alternative to Zika. Pyriproxyfen is a juvenile hormone analog, which has been shown to be cross reactive with retinoic acid, part of the mammalian regulatory system for neurological development, whose application during development causes microcephaly. This causal chain provides ample justification for pursuing a careful research effort on the role of pyriproxyfen in neurodevelopmental disorders. Counter to stated claims, existing studies of neurodevelopmental toxicity by Sumitomo, its manufacturer, provide some supportive evidence for neurodevelopmental toxicity including low brain weight in rat pups. The large-scale use of pyriproxyfen in Brazil and its coincidental timing with an increase in microcephaly cases should provide additional motivation. We believe that this evidence is strong enough to warrant an immediate cessation of pyriproxyfen application to Brazilian water supplies until additional research can be carried out on its neurodevelopmental toxicity. Alternative hypotheses about causes or factors affecting the incidence of microcephaly should be considered [@naturebutler; @McNeil2016; @Phillips2016]. Where insecticides are considered essential, Bt toxin is considered a safe alternative that has been claimed to be used in water systems in Recife since 2002 [@BtToxin]. Other vector control methods ranging from better sealed water containers, to ovitraps may also be used [@Regis2013; @Regis2008; @Bar-Yam2016].
**Acknowledgements:** We thank Dan Evans and Audi Byrne for helpful discussions, and Keisuke Ozaki for helpful communications about Sumitomo toxicology tests.
[20]{}
Saúde investiga 4.231 casos de microcefalia, Instituto Nacional de Salud: Boletim (March 9, 2016), <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/22554-saude-investiga-4-231-casos-de-microcefalia>.
Zika cases and congenital syndrome associated with Zika virus reported by countries and territories in the Americas, 2015-2016, Cumulative cases, World Health Organization (October 20, 2016), <http://www.paho.org/hq/index.php?option=com_docman&task=doc_view&Itemid=270&gid=36621&lang=en>.
Zika virus infection, Brazil and Colombia, Disease outbreak news, World Health Organization (October 21, 2015), <http://www.who.int/csr/don/21-october-2015-zika/en/>.
J. Mlakar, M. Korva, N. Tul, M. Popović, M. Poljšak-Prijatelj, [*et al.*]{}, Zika virus associated with microcephaly, The New England Journal of Medicine doi: 10.1056/NEJMoa1600651 (2016), <http://www.nejm.org/doi/full/10.1056/NEJMoa1600651>.
H. Tang, C. Hammack, S.C. Ogden, Z. Wen, X. Qian, [*et al.*]{} Zika virus infects human cortical neural progenitors and attenuates their growth, Cell Stem Cell doi: 10.1016/j.stem.2016.02.016 (2016).
Microcefalia: 1.749 casos confirmados no Brasil, Boletin Epidemiologico Semanal (BES), Instituto Nacional de Salud (July 27, 2016) <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/24769-microcefalia-1-749-casos-confirmados-no-brasil>.
L. Schuler-Faccini, E.M. Ribeiro, I.M.L. Feitosa, D.D.G. Horovitz, D.P. Cavalcanti, [*et al.*]{} Brazilian Medical Genetics Society—Zika Embryopathy Task Force, Possible association between Zika virus infection and microcephaly—Brazil, 2015, Morbidity and Mortality Weekly Report 65(3):59-62 (2016), <http://www.cdc.gov/mmwr/volumes/65/wr/mm6503e2.htm>.
R. Parens, Y. Bar-Yam, Zika and other potential causes of microcephaly in Brazil: Status March 16, 2016, New England Complex Systems Institute (March 16, 2016), <http://necsi.edu/research/social/pandemics/zikacauses.pdf>.
D. Butler, Brazil asks whether Zika acts alone to cause birth defects, Nature (July 25, 2016), <http://www.nature.com/news/brazil-asks-whether-zika-acts-alone-to-cause-birth-defects-1.20309#/>.
Zika - Epidemiological update, Pan American Health Organization, World Health Organization Regional Office for the Americas (April 8, 2016), <http://www.paho.org/hq/index.php?option=com_docman&task=doc_view&Itemid=270&gid=34144&lang=en>.
Zika virus microcephaly and Guillain-Barré syndrome situation report, World Health Organization (March 31, 2016), <http://apps.who.int/iris/bitstream/10665/204961/1/zikasitrep_7Apr2016_eng.pdf>.
Y. Bar-Yam, F. Costa, R. Parens, A.J. Morales, F. Nijhout, Determining the rate and week of infection of Zika caused microcephaly from Colombian and Brazilian data; Status Report August 1, 2016, New England Complex Systems Institute (August 1, 2016), <http://necsi.edu/research/social/pandemics/statusreport3>.
J.S. Cobb, L.J. Acosta, Colombia’s forecast on Zika-linked birth defect may be too high: minister, Reuters (February 17, 2016), <http://www.reuters.com/article/us-health-zika-colombia-idUSKCN0VQ2AB>.
Zika situation report, World Health Organization (October 27, 2016), <http://www.who.int/emergencies/zika-virus/situation-report/27-october-2016/en/>.
Pyriproxyfen: Compound Summary for CID 91753, Pubchem (December 17, 2016), <https://pubchem.ncbi.nlm.nih.gov/compound/Pyriproxyfen>.
Report from physicians in the crop-sprayed villages regarding Dengue-Zika, microcephaly, and mass-spraying with chemical poisons, Red Universitaria de Ambiente y Salud (February 3, 2016), <http://www.reduas.com.ar/wp-content/uploads/downloads/2016/02/Informe-Zika-de-Reduas_TRAD.pdf>.
C. Costa, Entidade diz ter sido mal interpretada e nega ligação entre microcefalia e larvicida, BBC Brasil (February 15, 2016), <http://www.bbc.com/portuguese/noticias/2016/02/160215_zika_larvicida_cc.shtml>.
A. Welch, Experts debunk claim blaming larvicide, not Zika, for microcephaly, CBS News (February 16, 2016), <http://www.cbsnews.com/news/health-experts-dismiss-claims-larvicide-linked-to-microcephaly/>.
D. Gorski, Zika not to blame for microcephaly? George Takei takes conspiracy bait, Genetic Literacy Project Science Blogs (February 16, 2016), <https://www.geneticliteracyproject.org/2016/02/16/zika-not-blame-microcephaly-george-takei-takes-conspiracy-bait/>.
T. Saegusa, SUMILARV - Study of S-31183 by oral administration during the period of fetal organogenesis in rats, Sumitomo (not publicly released; as public record obtained free by request from California Department of Pesticide Regulation and from Sumitomo) (1988).
I. Cotgreave, A.A. Ghavanini, E. Alfaro-Moreno, . Bergman, K. Cederbrant, [*et al.*]{} Pyriproxifen and microcephaly: An investigation of potential ties to the ongoing “Zika epidemic”, The Swedish Center for Toxicology Sciences (March, 2016), <http://swetox.se/wp-content/uploads/2016/03/ppf-zika.pdf>.
D. Evans, F. Nijhout, R. Parens, A. Morales, Y. Bar-Yam, A Possible Link Between Pyriproxyfen and Microcephaly, arXiv (April 13, 2016), <https://arxiv.org/abs/1604.03834>.
C. Tantibanchachai, Retinoids as teratogens, The Embryo Project Encyclopedia ISSN: 1940-5030 (February 28, 2014), <https://embryo.asu.edu/pages/retinoids-teratogens>.
Microcephaly: Fact sheet, World Health Organization (October 4, 2016), <http://www.who.int/mediacentre/factsheets/microcephaly/en/>.
L.M. Riddiford, How does juvenile hormone control insect metamorphosis and reproduction?, General and Comparative Endocrinology 179(3):477-484 (June 20, 2012).
D.E. Wheeler, H.F. Nijhout, A perspective for understanding the modes of juvenile hormone action as a lipid signaling system, Bioessays 25(10):994-1001 doi: 10.1002/bies.10337 (2003), <http://cals.arizona.edu/ento/sites/cals.arizona.edu.ento/files/wheelerSite/documents/bioessay03.pdf>.
M. Rhinn, P. Dollé, Retinoic acid signalling during development, Development 139:843-858 doi: 10.1242/dev.065938 (2012) <http://dev.biologists.org/content/139/5/843>.
V. Němec, D. Kodrík, S. Matolín, H. Laufer, Juvenile hormone-like effects of retinoic acid in insect metamorphosis, embryogenesis and reproduction, Journal of Insect Physiology 39:1083-1093 doi: 10.1016/0022-1910(93)90132-B (1993).
S.R. Palli, L.M. Riddiford, K. Hiruma, Juvenile hormone and “retinoic acid” receptors in Manduca epidermis, Insect Biochemistry 21:7-15 doi: 10.1016/0020-1790(91)90059-N (1991).
G. Jones, Molecular mechanisms of action of juvenile hormone, Annual Review of Entomology 40:147-169 doi: 10.1146/annurev.en.40.010195.001051 (January 1995).
M.A. Harmon, M.F. Boehm, R.A. Heyman, D.J. Mangelsdorf, Activation of mammalian retinoid X receptors by the insect growth regulator methoprene, Proceedings of the National Academy of Sciences U.S.A. 92:6157-6160 (1995), <http://www.pnas.org/content/92/13/6157.long>.
T.S. Dhadialla, G.R. Carlson, D.P. Le, New insecticides with ecdysteroidal and juvenile hormone activity, Annual Review of Entomology 43:545-569 doi: 10.1146/annurev.ento.43.1.545 (1998).
B. Unsworth, S. Hennen, A. Krishnakumaran, P. Ting, N. Hoffman, Teratogenic evaluation of terpenoid derivatives, Life Sciences 15(9):1649-1655 (November 1, 1974).
R.S. Stem, When a uniquely effective drug is teratogenic, New England Journal of Medicine 320:1007-1009 doi: 10.1056/NEJM198904133201510 (1989) <http://www.nejm.org/doi/full/10.1056/NEJM198904133201510>.
D.W. Irving, C.C. Willhite, D.T. Burk, Morphogenesis of isotretinoin-induced microcephaly and micrognathia studied by scanning electron microscopy, Teratology 34(2):141-153 (1986).
M.D. Collins, G.E. Mao, Teratology of retinoids, Annual Review of Pharmacology and Toxicology, 39(1):399 doi: 10.1146/annurev.pharmtox.39.1.399 (1999).
T. Flatt, L.L. Moroz, M. Tatar, A. Heyland, Comparing thyroid and insect hormone signaling, Integrative and Comparative Biology, 46(6):777-794 (2006), <https://icb.oxfordjournals.org/content/46/6/777.full>.
Private communications from regulatory affairs & chemical safety dept., Environmental Health Division, Sumitomo Chemical Co., Ltd.
P. Brasil, J.P. Pereira Jr., C.R. Gabaglia, L. Damasceno, M. Wakimoto, [*et al.*]{} Zika virus infection in pregnant women in Rio de Janeiro—Preliminary report, New England Journal of Medicine (March 5, 2016), <http://www.nejm.org/doi/full/10.1056/NEJMoa1602412>.
C.C. Willhite, R.M. Hill, D.W. Irving, Isotretinoin-induced craniofacial malformations in humans and hamsters, Journal of Craniofacial Genetics and Developmental Biology Supplement 2:193-209 (1986), <http://www.ncbi.nlm.nih.gov/pubmed/3491113>.
G. Tzimas, H. Bürgin, M.D. Collins, H. Hummler, H. Nau, The high sensitivity of the rabbit to the teratogenic effects of 13-cis-retinoic acid (isotretinoin) is a consequence of prolonged exposure of the embryo to 13-cis-retinoic acid and 13-cis-4-oxo-retinoic acid, and not of isomerization to all-trans-retinoic acid, Archives of Toxicology 68(2):119-128 (1994), <http://www.ncbi.nlm.nih.gov/pubmed/8179481>.
H. Nau, Embryotoxicity and teratogenicity of topical retinoic acid, Skin Pharmacology 6(Suppl 1):35-44 (1993), <http://www.ncbi.nlm.nih.gov/pubmed/8142110>.
C. Eckhoff, C.C. Willhite, Embryonic delivered dose of isotretinoin (13-cis-retinoic acid) and its metabolites in hamsters, Toxicology and Applied Pharmacology 146(1):79-87 doi: 10.1006/taap.1997.8220 (September 1997), <http://www.unboundmedicine.com/medline/citation/9299599/Embryonic_delivered_dose_of_isotretinoin__13_cis_retinoic_acid__and_its_metabolites_in_hamsters_>.
L.B. Holmes, E.A. Harvey, B.A. Coull, K.B. Huntington, S. Khoshbin, [*et al.*]{} The teratogenicity of anticonvulsant drugs, New England Journal of Medicine 344(15):1132-1138 (2001).
S.C. Hyoun, S.G. Običan, A.R. Scialli, Teratogen update: Methotrexate, Birth Defects Research Part A: Clinical and Molecular Teratology 94(4):187-207 doi: 10.1002/bdra.23003 (March 20, 2012), <http://onlinelibrary.wiley.com/doi/10.1002/bdra.23003/full>.
C.F. Wilkinson, J. Driver, G. Burin, G. Whitmyre, C. Dragula, Response to EPA Review of Teratogenicity/Developmental Toxicity, Study of SUMILARV (S-31183) in Rabbits, Sumitomo (not publicly released; as public record obtained free by request from California Department of Pesticide Regulation and from Sumitomo) (1994).
T.R. Elliott, Registering randomized clinical trials and the case for CONSORT, Experimental and Clinical Psychopharmacology 15(6):511-518 (December 2007) doi: 10.1037/1064-1297.15.6.511.
Estimating the reproducibility of psychological science, Open Science Collaboration Science 349(6251) doi: 10.1126/science.aac4716 (August 18, 2015) <http://science.sciencemag.org/content/349/6251/aac4716>.
M. Baker, 1,500 scientists lift the lid on reproducibility, Nature 533(7604):452-454 doi: 10.1038/533452a (July 28, 2016) <http://www.nature.com/polopoly_fs/1.19970!/menu/main/topColumns/topLeftColumn/pdf/533452a.pdf>.
S.A. Bailey, R.H. Zidell, R.W. Perry, Relationships Between Organ Weight and Body/Brain Weight in the Rat: What Is the Best Analytical Endpoint? Toxicologic Pathology 32:448-466 doi: 10.1080/01926230490465874 (2004).
M. Tamaru, Y. Hirata, M. Nagayoshi, T. Matsutani, Brain changes in rats induced by prenatal injection of methylazoxymethanol, Teratology 37(2):149-57 doi: 10.1002/tera.1420370208 (1988).
K. Ozaki, Reply to Yaneer Bar-Yam, Sumitomo (April 27, 2016).
K. Ozaki, Reply to Yaneer Bar-Yam, Sumitomo (April 5, 2016).
C.F. Yung, K.T. Yeo, G.SH. Yeo, K.C. Thoon, Zika virus and microcephaly: Case ascertainment and quantifying risk, BMJ: Response (February 8, 2016), <http://www.bmj.com/content/352/bmj.i781/rr>.
V. Reis, Nota técnica sobre microcefalia e doenças vetoriais relacionadas ao Aedes aegypti: Os perigos das abordagens com larvicidas e nebulizações químicas - fumacê, ABRASCO (February 2, 2016), <https://www.abrasco.org.br/site/2016/02/nota-tecnica-sobre-microcefalia-e-doencas-vetoriais-relacionadas-ao-aedes-aegypti-os-perigos-das-abordagens-com-larvicidas-e-nebulizacoes-quimicas-fumace/#_ftn2>.
Pyriproxyfen, EPA Office of Pesticide Programs (2016), <https://iaspub.epa.gov/apex/pesticides/f?p=CHEMICALSEARCH:3:::NO:1,3,31,7,12,25:P3_XCHEMICAL_ID:3680>.
Pyriproxyfen in drinking-water: Background document for development of WHO guidelines for drinking-water quality, World Health Organization (2004), <http://www.who.int/water_sanitation_health/dwq/chemicals/pyriproxyfen.pdf>.
J. Devillers, Ed., Juvenile Hormones and Juvenoids: Modeling biological effects and environmental fate, ISBN: 9781466513228 (CRC Press, Boca Raton, FL, 2013), pp. 108-126.
K. Smith-Janssen, Nontoxic Ways to Protect Your Pet, National Resources Defense Council (January 22, 2016), <https://www.nrdc.org/stories/nontoxic-ways-protect-your-pet>.
J.F. Invest, J.R. Lucas, Pyriproxyfen as a mosquito larvicide, Proceedings of the Sixth International Conference on Urban Pests (2008), <http://www.icup.org.uk/reports%5CICUP886.pdf>.
M. Sihuincha, E. Zamora-Perea, W. Orellana-Rios, J.D. Stancil, V. López-Sifuentes, [*et al.*]{} Potential use of pyriproxyfen for control of Aedes aegypti (Diptera: Culicidae) in Iquitos, Peru, Journal of Medical Entomology (July 1, 2005), <http://jme.oxfordjournals.org/content/42/4/620.long>.
H.J. Overgaard, N. Alexander, M.I. M[á]{}tiz, J.F. Jaramillo, V.A. Olano, [*et al.*]{} Diarrhea and dengue control in rural primary schools in Colombia: Study protocol for a randomized controlled trial, Trials 13:182 (October 3, 2012), <http://trialsjournal.biomedcentral.com/articles/10.1186/1745-6215-13-182>.
C.M. Seng, T. Setha, J. Nealon, D. Socheat, M.B. Nathan, Six months of Aedes aegypti control with a novel controlled-release formulation of pyriproxyfen in domestic water storage containers in Cambodia, Southeast Asian Journal of Tropical Medicine and Public Health 39(5):822-826 (September 2008), <http://www.ncbi.nlm.nih.gov/pubmed/19058575>.
Orientações técnica para utilização do larvicida pyriproxyfen (0,5 G) no controle de Aedes aegypti, Miniserio da Saúde (May 30, 2014), <http://portalsaude.saude.gov.br/index.php/o-ministerio/principal/leia-mais-o-ministerio/632-secretaria-svs/vigilancia-de-a-a-z/controle-de-vetores-inseticidas-e-larvicidas/13059-orientacoes-tecnica-para-utilizacao-do-larvicida-pyriproxyfen-0-5-g-no-controle-de-aedes-aegypti>.
\(2016) Boletim Epidemiologico: Monitoramento dos casos de dengue, febre de chikungunya e febre pelo vírus Zika ate a Semana Epidemiologica 37, 2016, Ministerio da Saúde (September 17, 2016), <http://portalsaude.saude.gov.br/images/pdf/2016/outubro/18/2016-029-Dengue-publicacao-n-34.pdf>.
Informe Epidemioł[l]{}õgico No 32: Semana Epidemioł[l]{}õgica (SE) 25/2016 (19/06 A 25/06/2016): Monitoramento Dos Casos de Microcefalia No Brasil, Miniserio da Saúde (June 25, 2016), <http://combateaedes.saude.gov.br/images/pdf/informe_microcefalia_epidemiologico_32.pdf>.
R. Romo, Brazil, world health officials deny link between pesticide and microcephaly, CNN (February 18, 2016), <http://www.cnn.com/2016/02/17/health/brazil-who-pesticide-microcephaly-zika/>.
R.E. Bichell, Did a pesticide cause microcephaly In Brazil? Unlikely, say experts, NPR (February 20, 2016), <http://www.npr.org/sections/goatsandsoda/2016/02/18/467138913/did-a-pesticide-cause-microcephaly-in-brazil-unlikely-say-experts>.
Secretário de Saúde afirma que Recife não utiliza o larvicida pyriproxyfen, Jornal de Commercio (February 15, 2016), <http://jconline.ne10.uol.com.br/canal/cidades/saude/noticia/2016/02/15/secretario-de-saude-afirma-que-recife-nao-utiliza-o-larvicida-pyriproxyfen-221043.php>.
Dengue Control Office of Pernambuco, Personal Communication with Yaneer Bar-Yam and Alfredo Morales (2016).
Grupo de trabalho se reúne para discutir o uso do pyriproxyfen no combate ao Aedes Aegypti, CES - Conselho Estadual de Saúde de Pernambuco (April 25, 2016), <http://www.ces.saude.pe.gov.br/grupo-de-trabalho-se-reune-para-discutir-o-uso-do-pyriproxyfen-no-combate-ao-aedesaegypti/>.
K. Worth, Zika Uncontained: How Brazil became a breeding ground for an outbreak, PBS: Frontline <http://apps.frontline.org/zika-water/?linkId=27848916>.
B. Douglas, City at centre of Brazil’s Zika epidemic reeling from disease’s insidious effects, Guardian (January 25, 2016), <http://www.theguardian.com/global-development/2016/jan/25/zika-virus-mosquitoes-countries-affected-pregnant-women-children-microcephaly>.
M. de Albuquerque, W. de Souza, A. Mendes, T. Lyra, R. Ximenes, [*et al.*]{} Pyriproxyfen and the microcephaly epidemic in Brazil - an ecological approach to explore the hypothesis of their association, Memórias do Instituto Oswaldo Cruz (October 31, 2016) <http://www.scielo.br/scielo.php?pid=S0074-02762016005022104&script=sci_arttext>.
Yaneer Bar-Yam, Raphael Parens, Alfredo J. Morales, Fred Nijhout, Is Zika the cause of Microcephaly? Status Report November 4, 2016, New England Complex Systems Institute (November 2, 2016) <http://necsi.edu/research/social/pandemics/statusreport4.pdf>.
Y. Bar-Yam, D. Evans, R. Parens, A.J. Morales, F. Nijhout, Is Zika the cause of Microcephaly? Status Report June 22, 2016, New England Complex Systems Institute (June 22, 2016) <http://necsi.edu/research/social/pandemics/statusreport>.
Yaneer Bar-Yam, Raphael Parens, Alfredo J. Morales, Fred Nijhout, Is Zika the cause of Microcephaly? Status Report June 27, 2016, New England Complex Systems Institute (June 27, 2016) <http://necsi.edu/research/social/pandemics/statusreport2>.
Y. Bar-Yam, R. Menapace, DRAFT Zika virus community response, New England Complex Systems Institute (2016), <http://necsi.edu/research/social/pandemics/ZikaResponse2.pdf>.
S. Rasmussen, D. Jamieson, M. Honein, L. Petersen, Zika Virus and Birth Defects ? Reviewing the Evidence for Causality, New England Journal of Medicine (May 19, 2016), <http://www.nejm.org/doi/full/10.1056/NEJMsr1604338#t=article>.
G. Calvet, R.S. Aguiar, A.S.O. Melo, S.A. Sampaio, I. de Filippis, [*et al.*]{} Detection and sequencing of Zika virus from amniotic fluid of fetuses with microcephaly in Brazil: a case study, The Lancet Infectious Diseases 16(6):653-660 (June 2016), <http://www.thelancet.com/journals/laninf/article/PIIS1473-3099(16)00095-5/abstract>.
S. Nebehay, B. Hirschler, Zika link to birth defects could be proven within weeks: WHO, Reuters Life (2016) <http://www.reuters.com/article/us-health-zika-idUSKCN0VL0TQ>.
Ministério da Saúde investiga 3.852 casos suspeitos de microcefalia no país, Miniserio da Saúde (2016) <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/22145-ministerio-da-saude-investiga-3-852-casos-suspeitos-de-microcefalia-no-pais>.
M. Besnard, S. Lastère, A. Teissier, V. Cao-Lormeau, D. Musso, Evidence of Perinatal Transmission of Zika virus, French Polynesia, December 2013 and February 2014, Eurosurveillance 19(13):20751 doi: 10.2807/1560-7917.ES2014.19.13.20751 (2014), <http://www.eurosurveillance.org/ViewArticle.aspx?ArticleId=20751>.
D. Musso, T. Nhan, E. Robin, C. Roche, D. Bierlaire, [*et al.*]{} Potential for Zika virus transmission through blood transfusion demonstrated during an outbreak in French Polynesia, November 2013 to February 2014, Eurosurveillance 19(14):20761 doi: 10.2807/1560-7917.ES2014.19.14.20761 (2014), [http://www. eurosurveillance.org/ViewArticle.aspx?ArticleId=20761](http://www. eurosurveillance.org/ViewArticle.aspx?ArticleId=20761).
S. Cauchemez, M. Besnard, P. Bompard, T. Dub, P. Guillemette-Artur, [*et al.*]{} Association between Zika virus and microcephaly in French Polynesia, 2013-15: A retrospective study, Lancet (March 15, 2016), <http://www.thelancet.com/journals/lancet/article/PIIS0140-6736(16)00651-6/fulltext>.
V. Cao-Lormeau, A. Blake, S. Mons, S. Lastère, C. Roche, [*et al.*]{} Guillain-Barré Syndrome outbreak associated with Zika virus infection in French Polynesia: a case-control study, The Lancet 387(10027):1531-1539 (April 9 2016), <http://www.thelancet.com/journals/lancet/article/PIIS0140-6736(16)00562-6/abstract>.
G. Vogel, K. McLaughlin, Why does Zika leave some patients paralyzed?, Science (March 3, 2016), <http://www.sciencemag.org/news/2016/03/why-does-zika-leave-some-patients-paralyzed>.
There is a discrepancy between the number of reported (42) and the number of cases shown in Fig. 1 of that paper (41), the numbers we cite are those present in the figure as it is unclear which week the additional exposure is in.
T. de Araújo, L. Rodrigues, R. de Alencar Ximenes, D. de Barros Miranda-Filho, U.R. Montarroyos, [*et al.*]{} Association between Zika virus infection and microcephaly in Brazil, January to May, 2016: preliminary report of a case-control study, The Lancet (16)12:1356-1363 (December 2016), <http://www.thelancet.com/pdfs/journals/laninf/PIIS1473-3099(16)30318-8.pdf>.
Boletin Epidemiologico Semanal (BES), Instituto Nacional de Salud (October 22, 2016), <http://www.ins.gov.co/boletin-epidemiologico/Paginas/default.aspx>.
Colombia, primer país del continente que cierra epidemia de zika, Ministerio de Salud y Protección Social: Boletin de Prensa No 155 de 2016 (July 25, 2016) <https://www.minsalud.gov.co/Paginas/Colombia-primer-pais-del-continente-que-cierra-epidemia-de-zika.aspx>.
O. Pacheco, M. Beltrán, C.A. Nelson, D. Valencia, N. Tolosa, [*et al.*]{}, Zika virus disease in Colombia - Preliminary report, NEJM (June 15, 2016) <http://www.nejm.org/doi/full/10.1056/NEJMoa1604037#t=article>. All Countries & Territories with Active Zika Virus Transmission, Centers for Disease Control and Prevention (October 5, 1016), <http://www.cdc.gov/zika/geo/active-countries.html>.
Q. Zhang, K. Sun, M. Chinazzi, A. Pastore-Piontti, N. Dean, [*et al.*]{} Projected spread of Zika virus in the Americas, BioRxiv (July 28, 2016) doi: 10.1101/066456, <http://biorxiv.org/content/biorxiv/early/2016/12/12/066456.full.pdf>.
D. McNeil Jr., J. Cobb, Colombia Is Hit Hard by Zika, but Not by Microcephaly, The New York Times (October 31, 2016), <http://www.nytimes.com/2016/11/01/health/colombia-zika-microcephaly.html?_r=0>.
D. Phillips, N. Miroff, Scientists are bewildered by Zika’s path across Latin America, The Washington Post (October 25, 2016), <https://www.washingtonpost.com/world/the_americas/scientists-are-bewildered-by-zikas-path-across-latin-america/2016/10/25/5e3a992c-9614-11e6-9cae-2a3574e296a6_story.html>.
N.R. Faria, [*et al.*]{} Zika virus in the Americas: Early epidemiological and genetic findings, Science 352(6283):345-349 (2016).
Saúde investiga 4.222 casos suspeitos de microcefalia no país, Miniserio da Saúde (March 1, 2016) <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/22396-saude-investiga-4-222-casos-suspeitos-de-microcefalia-no-pais>.
Microcefalia: Ministério da Saúde investiga 4.291 casos suspeitos no país, Boletin Epidemiologico Semanal (BES), Instituto Nacional de Salud (March 29, 2016) <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/22869-microcefalia-ministerio-da-saude-investiga-4-291-casos-suspeitos-no-pais>.
Ministério da Saúde confirma 1.709 casos de microcefalia, Boletin Epidemiologico Semanal (BES), Instituto Nacional de Salud (July 20, 2016) <http://portalsaude.saude.gov.br/index.php/cidadao/principal/agencia-saude/24625-ministerio-da-saude-confirma-1-709-casos-de-microcefalia>.
E.L . Cuevas, V.T. Tong, N. Rozo, D. Valencia, O. Pacheco, [*et al.*]{} Preliminary Report of Microcephaly Potentially Associated with Zika Virus Infection During Pregnancy - Colombia, January-November 2016, Morbidity and Mortality Weekly Report 65(49):1409-1413 (December 16, 2016), <https://www.cdc.gov/mmwr/volumes/65/wr/mm6549e1.htm>.
J.D. Cragan, J.L. Isenburg, S.E. Parker, C.J. Alverson, R.E. Meyer, E.B. Stallings, [*et. al.*]{} Population-based microcephaly surveillance in the United States, 2009 to 2013: An analysis of potential sources of variation, Birth Defects Research Part A: Clinical and Molecular Teratology 106(11):972-982 (2016).
C. de Oliveira, Brazil health official defends controversial decision to halt use of pesticide in fight against Zika, Latin America News Dispatch (February 22, 2016), <http://latindispatch.com/2016/02/22/brazil-health-official-defends-controversial-decision-to-halt-use-of-pesticide-in-fight-against-zika/>.
L.N. Regis, R.V. Acioli, J.C. Silveira Jr., M.A.V. Melo-Santos, W.V. Souza, [*et al.*]{} Sustained reduction of the Dengue vector population resulting from an integrated control strategy applied in two Brazilian cities, PLOS One Open Access 8(7) (July, 2013).
L. Regis, A.M. Monteiro, M.A.V. de Melo-Santos, J.C. Silveira Jr., A.F. Furtado, [*et al.*]{} Developing new approaches for detecting and preventing Aedes aegypti population outbreaks: Basis for surveillance, alert and control system, Memórias do Instituto Oswaldo Cruz 103(1):50-59 (February 2008).
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.