Hydrodynamics in graphene: Linear-response transport

B.N. Narozhny,1, 2 I.V. Gornyi,3, 4 M. Titov,5 M. Schütt,6 and A.D. Mirlin3, 1, 7

1
Institut für Theorie der Kondensierten Materie,
Karlsruher Institut für Technologie, 76128 Karlsruhe, Germany
2
National Research Nuclear University MEPhI (Moscow Engineering
Physics Institute), Kashirskoe shosse 31, 115409 Moscow, Russia
3
Institut für Nanotechnologie, Karlsruhe Institute of Technology, 76021 Karlsruhe, Germany
4
A.F. Ioffe Physico-Technical Institute, 194021 St. Petersburg, Russia
5
Radboud University Nijmegen, Institute for Molecules and Materials, NL-6525 AJ Nijmegen, The Netherlands
6
School of Physics and Astronomy, University of Minnesota, Minneapolis, MN 55455, USA
7
arXiv:1411.0819v1 [cond-mat.mes-hall] 4 Nov 2014

Petersburg Nuclear Physics Institute, 188350 St. Petersburg, Russia

(Dated: November 5, 2014)
We develop a hydrodynamic description of transport properties in graphene-based systems which
we derive from the quantum kinetic equation. In the interaction-dominated regime, the collinear
scattering singularity in the collision integral leads to fast unidirectional thermalization and allows
us to describe the system in terms of three macroscopic currents carrying electric charge, energy, and
quasiparticle imbalance. Within this “three-mode” approximation we evaluate transport coefficients
in monolayer graphene as well as in double-layer graphene-based structures. The resulting classical
magnetoresistance is strongly sensitive to the interplay between the sample geometry and leading
relaxation processes. In small, mesoscopic samples the macroscopic currents are inhomogeneous
which leads to linear magnetoresistance in classically strong fields. Applying our theory to double-
layer graphene-based systems, we provide microscopic foundation for phenomenological description
of giant magnetodrag at charge neutrality and find magnetodrag and Hall drag in doped graphene.

PACS numbers: 72.80.Vp, 73.23.Ad, 73.63.Bd

Traditional hydrodynamics1 describes systems at large sion integral4,5,7,10 allowing for a non-perturbative solu-
length scales (compared to the mean free path). The hy- tion to the kinetic equation. The distinct feature of this
drodynamic equations are typically formulated in terms solution is fast unidirectional thermalization10 that fa-
of currents and densities of conserved quantities and cilitates integration of the kinetic equation. The unique
can be derived from the kinetic equation using either feature of the resulting hydrodynamic description of elec-
the Chapman-Enskog2 or Grad3 procedures. Within tronic transport in graphene is inequivalence of the elec-
the leading approximation, gradients of the macroscopic tric current and total momentum of the system5,6,10 . As
physical quantities are assumed to be small, such that the the latter is equivalent to the energy current, transport
system can be characterized by the local equilibrium dis- properties of graphene are governed by a non-trivial in-
tribution function. Dissipative properties, such as elec- terplay of electric current and energy relaxation.
trical or thermal conductivity or viscosity are then deter- Two-fluid hydrodynamics in graphene was suggested
mined by small corrections to the local-equilibrium distri- in Refs. 5,8 and then extended to double-layer graphene-
bution function. Within linear response, such corrections based structures in Ref. 10, which allowed for a descrip-
are proportional to a weak external bias. tion of the Coulomb drag effect12–15 in graphene. An ex-
Recently the kinetic equation approach was applied to tension of this approach to mesoscopic (finite-size) sam-
electronic excitations in graphene4–10 . In contrast to con- ples was suggested in Ref. 15. Qualitatively, this theory
ventional metals and semiconductors, graphene is charac- can be interpreted in terms of a semiclassical two-band
terized by the linear excitation spectrum which makes the model that yields non-trivial magnetic field dependence
system explicitly non-Galilean-invariant. Consequently, of the transport coefficients and accounts for the effect of
the transport scattering time in graphene is strongly af- giant magnetodrag at the neutrality point15 . The classical
fected by electron-electron interaction11 which has to be mechanism of this effect is similar to the standard mech-
taken into account on equal footing with disorder poten- anism of magnetoresistance in multi-band systems16 .
tial. At the same time, due to the classical nature of the In this paper we rigorously derive the hydrodynamic
Coulomb interaction between charge carriers in graphene, description of electronic transport in graphene within lin-
the system is also non-Lorentz-invariant. As a result, the ear response. While we use the same collinear-scattering
standard derivation1 of the hydrodynamic equations from singularity as found in Refs. 4,5,7,10 in order to integrate
the kinetic equation has to be revisited4–10 . the quantum kinetic equation, we argue that the physics
The linearity of the quasiparticle spectrum in graphene of the system should be described in terms of three macro-
leads to an important corollary: the energy and momen- scopic currents: the electric current j, energy current Q,
tum conservation laws for Dirac quasiparticles coincide and quasiparticle imbalance8 current P .
in the special case of collinear scattering. This kinematic For general doping, the resulting theory is rather cum-
peculiarity results in a singular contribution to the colli- bersome. However, at the charge neutrality point and
 2

in the degenerate limit the equations simplify allowing to be energy independent, although in physical graphene
for an analytic solution. In the former case, we focus most of the impurity scattering processes lead to energy-
on the issue of magnetoresistance, a subject of consider- dependent relaxation rates. A corresponding generaliza-
able experimental interest17–23 . In particular, we demon- tion of our theory is straightforward26 and does not lead
strate the appearance of the linear magnetoresistance in to qualitatively new effects31 . At the same time, quanti-
moderately strong, classical magnetic fields in monolayer tative description of experimental data may greatly ben-
graphene24 . In double-layer graphene-based systems we efit from a realistic description of disorder15 .
describe negative Coulomb drag15,25 and justify the phe- Non-linear hydrodynamics of graphene will be dis-
nomenological two-band model of Ref. 15 (precisely at cussed in a separate publication32 .
the Dirac point the imbalance current is proportional to
the energy current allowing one to reduce the number of
variables). Both effects occur in narrow, mesoscopic sam- A. Linear response equations in graphene
ples in the presence of energy relaxation and quasiparticle
recombination due to electron-phonon interaction. One of the main results of this paper is the set of
In the opposite limit of very high doping (i.e. in the macroscopic equations describing electronic transport in
“Fermi-liquid regime”) all three macroscopic currents be- graphene within linear response. What makes this un-
come equivalent and the theory is reduced to the stan- usual is that the electric current j is inequivalent to the
dard Drude-like description that can be also derived by energy current Q and the quasiparticle imbalance cur-
perturbative methods26 . Here we find the leading cor- rent P . The three macroscopic currents can be found
rections to the standard picture of Coulomb drag26,27 from the following equations
yielding magnetodrag and Hall drag in doped graphene.  
The rest of the paper is organized as follows. We be- π A C
−∇Π+E+RH K×eB = R0 j+ 2 + , (1a)
gin (Section I) with the summary of our theory and re- e K τvv τvs
sults for monolayer graphene. In Section II, we present
a derivation of the macroscopic description of electronic e
transport. In Section III, we use this theory to evalu- − ∇Θ + N1 E + RH [j × eB ] = R0 Q, (1b)
ate transport coefficients in graphene such as the magne- K
toresistance at the point of charge neutrality for small,  
mesoscopic samples. In Section IV, we apply our theory µ e × eB = eR0 P + π A C
− ∇Ξ + E + RH K + .
to double-layer graphene-based systems12–15 . Conclud- K e2 K τvs τss
ing remarks can be found in Section V. Technical details (1c)
are relegated to the Appendices. Here E is the electric field, eB is the unit vector in the
direction of the magnetic field B = BeB , K is the mean
quasiparticle kinetic energy in graphene15 (with T being
I. MACROSCOPIC DESCRIPTION OF the temperature and µ the chemical potential):
TRANSPORT IN MONOLAYER GRAPHENE

µ , T µ
 
µ/T
K = 2T ln 1 + e −µ→ , (2)
In this Section, we describe transport properties of 2T ln 2 , T  µ
monolayer graphene. Neglecting all quantum effects28–30 ,
we base our considerations on the set of macroscopic the dimensionless quantity N1 = 2n0 /(ν0 K) represents
transport equations which essentially generalize the usual the equilibrium charge density n0 (here ν0 = ∂n0 /∂µ),
Ohm’s law to the case of collision-dominated transport and the two coefficients R0 and RH are
in graphene. These equations can be derived from the ki- π πωB
netic equation (see Section II below) in the interaction- R0 (µ, τ, T ) = , RH (µ, B, T ) = , (3)
e2 Kτ e2 K
dominated regime, where the transport scattering time
due to electron-electron interaction τee is much smaller where e the electron charge, the frequency ωB is
than the disorder mean-free time τ
evg2 B
τee  τ. ωB = , (4)
cK
We limit ourselves to the discussion of a steady state. and vg the quasiparticle velocity.
The latter is typically established by means of disorder In graphene, the energy current Q is equivalent to the
scattering. A notable exception is neutral graphene in total momentum of electrons, which cannot be relaxed by
the absence of magnetic field, where the steady state ex- electron-electron interaction respecting momentum con-
ist due to electron-electron interaction alone. However, servation. Therefore, the transport scattering rates due
in the presence of the field, even at the Dirac point the to electron-electron interaction appear only in Eqs. (1a)
steady state cannot be reached without disorder. There- and (1c). The three scattering times τvv , τvs , and τss
fore, we have to keep the weak disorder in the prob- describe mutual scattering of the velocity and imbalance
lem. For simplicity, we assume the mean-free time τ modes respecting Onsager reciprocity.
 3

The above three modes form the three-mode Ansatz for electron-phonon collisions, we find the following continu-
the non-equilibrium correction to the electronic distribu- ity equations (see Appendix C for details):
tion function [see Eq. (19) below and Appendix A]
b c
e∇ · P = − + , (9b)
2v n  o τIb τIc
h= A + B + Csign() , (5)
eν0 T vg2 K
e b c
∇·Q= − . (9c)
where the vectors A, B, and C are linear combinations K τEb τEc
of the three macroscopic currents that are introduced for
Here the auxiliary quantities b and c are linear combi-
brevity [see Eq. (A1) for details]. The absence of the vec-
nations of inhomogeneous parts of the charge, energy,
tor B in the right-hand side of Eqs. (1) is due to momen-
and imbalance densities with the same coefficients as the
tum conservation. However, all three auxiliary vectors
vectors B and C, see Eq. (A2). Physically, imbalance re-
enter the Lorentz terms in the following combinations
laxation (described by τIb and τIc ) is due to inter-band
µ processes only and thus is expected to be slower than
K = A tanh + B + C, (6) energy relaxation (described by τEb and τEc ).
2T
The macroscopic equations (1) simplify at the neutral-
ity point and in the degenerate (or Fermi-liquid) limit.
e = A + B µ + C tanh µ .
K (7) We now turn to the discussion of the solutions to Eqs. (1)
K 2T in these cases, which clarify the structure of our theory.
The quantity Π represents the inhomogeneous part of
the flux density of the electric current (cf. the usual B. Transport in the degenerate limit
momentum flux density or the “stress-tensor”) and is
given by a linear combination of the inhomogeneous den-
sities corresponding to the three modes in the system: At high doping (or at low temperatures), the electronic
the charge δn, energy δu, and imbalance δρ. Similarly, system in graphene becomes degenerate. In the limit
the quantities Θ and Ξ describe the flux densities for the µ  T , all three macroscopic currents become equivalent
energy and imbalance currents, see Eqs. (34) below. e
j(µ  T ) ≈ Q(µ  T ) ≈ eP (µ  T ). (10)
In finite-size samples the equations (1) have to be sup- µ
plemented by the corresponding continuity equations and
Maxwell’s equations, since inhomogeneous charge density The additional vectors introduced in Eqs. (1) simplify to
fluctuations give rise to electromagnetic fields. Therefore A(µ  T ) ≈ K(µ  T ) ≈ K(µ
e  T ) ≈ j,
the electric field E in Eqs. (1) comprises the externally
applied and self-consistent (Vlasov-like1 ) fields. The self-
consistency amounts to solving the electrostatic problem C(µ  T ) = 0.
described by the Maxwell’s equations1 In this regime, the Galilean invariance is effectively re-
4π stored and all relaxation rates due to electron-electron
∇ · E = 4πδnδ(z), ∇ × E = 0, ∇×B = j. (8) interaction vanish. Consequently, the three equations (1)
c become equivalent to the Ohm’s law
While charge carriers are confined within the graphene ∇δn
sheet, the electromagnetic fields are not, hence the factor − + E + RH j × eB = R0 j, (11)
e2 ν0
of δ(z) in Eq. (8). At the same time, we assume that the
uniform charge density n0 is controlled by an external where
gate. Consequently, only the non-uniform part of the π
charge density δn is taken into account in Eq. (8). R0 = R0 (µ  T ) = , (12a)
e2 µτ
The continuity equations can be obtained by integrat-
ing the kinetic equation in the usual fashion1 . In the and
steady state, charge conservation requires πvg2 B
RH = RH (µ  T ) = , (12b)
ecµ2
∇ · j = 0. (9a)
are the usual longitudinal and Hall resistances.
Similar equations can be derived for the energy and im- Physically, the above simplification is related to the
balance density. Since both of them are conserved by fact, that in the degenerate regime inter-band processes
electron-electron interactions, the collision integral in are exponentially suppressed. Effectively only one band
Eq. (18) does not contribute to the continuity equations. participates in transport and therefore the textbook
At the same time, electron-phonon interaction (that we results apply; in particular there is no magnetoresis-
have so far neglected) may lead to energy and imbalance tance. For leading corrections to this behavior see Sec-
relaxation processes8,15,35–41 . Taking into account the tion IV A 2.
 4

C. Transport at the neutrality point At this point, the essential role of disorder becomes
self-evident. Indeed, in the absence of disorder R0 = 0
At the charge neutrality point µ = 0, the auxiliary and then Eq. (14b) becomes senseless, at least when the
vectors in Eqs. (1) have the form system is subjected to external magnetic field. Physi-
cally, this means that in the absence of disorder our orig-
A(µ = 0) = K(µ
e = 0) = j, (13a) inal assumption of the steady state becomes invalid: un-
der external bias the energy current increases indefinitely.
In the absence of magnetic field the electric current
eQ N2 (0)
C(µ = 0) = γ0 − eP , is decoupled. In this case, the electrical resistivity of
2T ∆(0) ln 2 ∆(0) graphene can be read off Eq. (14a) [using Eqs. (2) and
γ0 − 1

eQ
 (3) at the neutrality point]
K(µ = 0) = − eP γ2 . π
∆(0) 2T ln 2 R(B = 0; µ = 0) = 2
 −1
τ + τee−1

(0) . (15)
2e T ln 2
Here the numerical coefficients are
If the system is subjected to an external magnetic field,
γ0 = π 2 /(12 ln2 2) ≈ 1.7119, (13b) then all three macroscopic currents are entangled. Using
Eqs. (13), (14b), and (14c), we find the following expres-
sion for the vector K that determines the Lorentz term
N2 (0) = 9ζ(3)/(8 ln3 2) ≈ 4.0607, (13c) in the equation (14a) for the electric current
K = j × eB κRH /[R0 ∆(0)],
γ2 = (N2 (0) − γ0 )/(γ0 − 1) ≈ 3.2996, (13d)
where
and
∆(0) − γ0 τ /τss
κ = γ0 − 1 + [γ0 − N2 (0)] .
∆(0) = γ02 − N2 (0) ≈ −1.1303. (13e) ∆(0) − N2 (0)τ /τss

In addition, one of the relaxation rates vanishes as well Clearly, the direction of the Lorentz term coincides with
the direction of the electric current. Hence, there is no
−1 classical Hall effect at the Dirac point (as expected from
τvs (µ = 0) = 0.
symmetry considerations)
As a result, the equations (1) simplify. Below we consider
the two limiting cases of wide and narrow samples as RH (µ = 0) = 0. (16)
determined by the interplay between the electron-phonon At charge neutrality, carriers from both bands are in-
scattering and the magnetic field24 . volved in scattering processes and the system exhibits
nonzero classical magnetoresistance (similarly to multi-
band semiconductors16 )
1. Transport coefficients in macroscopic samples
R(B; µ = 0) = R(B = 0; µ = 0) + δR(B; µ = 0),
If the sample width is the largest length scale in the
2
problem, W  `R ωB τ τee (where τee is the typical value R2H κ vg4 τ B 2
of the electron-electron transport scattering times and δR(B; µ = 0) = ∝ 2 3. (17)
R0 ∆(0) c T
`R is the typical length scale describing quasiparticle re-
combination due to electron-phonon scattering, see Sec- The sign of δR(B; µ = 0) is determined by the interplay
tion III B), the boundary effects may be neglected and the of τ , T , and τss . However, using Eqs. (13e) and (13c) we
sample behaves as if it were infinite. Then all physical find the coefficient as
quantities can be considered uniform. At charge neutral-
πκ 1.71 + 1.03 τ /τss
ity, the equations (1) take the form 3 ≈ −1.04 κ ≈ > 0.
8 ln 2∆(0) 1 + 3.59 τ /τss
πj
E + RH K × eB = R0 j + , (14a) Thus, our Eq. (17) describes positive magnetoresistance.
2e2 T τvv ln 2
Magnetoresistance in graphene was previously calcu-
lated within the two-mode approximation in Ref. 5 where
e it was found δR(B; µ = 0) = [π/9ζ(3)]vg4 τ c−2 B 2 /T 3 .
RH j × eB = R0 Q, (14b)
2T ln 2 This expression shows the same parameter dependence
as our Eq. (17) but with a numerical prefactor π/9ζ(3) ≈
πC 0.2904 which is independent of the interaction strength.
RH j × eB = eR0 P + . (14c) The electron-electron scattering time τss does not appear
2e2 T τss ln 2
in the two-mode approximation. In the “hydrodynamic”
The parameters R0 and RH are given by Eq. (3) evalu- limit τ  τss , the prefactor in Eq. (17) approaches the
ated at µ = 0. same numerical value as the result of Ref. 5.
 5

2. Transport in mesoscopic samples Here f (0) is the equilibrium Fermi-Dirac distribution with
the corresponding temperature T . In this paper we con-
In small enough samples, or in strong enough magnetic sider the steady-state transport and thus take the distri-
fields W  `R ωB 2
τ τee , boundary conditions become im- bution function to be time-independent
portant and physical quantities become inhomogeneous. ∂f
The macroscopic equations acquire gradient terms and = 0. (20)
have to be considered alongside the corresponding conti- ∂t
nuity equations as well as the Maxwell equations describ-
ing the self-consistent electromagnetic fields. In general, 2. Macroscopic currents
solution to such system of equations is a formidable com-
putational task that is best approached numerically. The
Let us now introduce macroscopic physical observ-
notable exception is the neutrality point, where the clas-
ables. The electric current is defined as
sical Hall effect is absent (due to exact electron-hole sym-
metry). In this case, the electrostatic problem is trivial
X
j=e vδf, (21a)
and we can tackle the problem analytically. Still, within
the three-mode approximation the solution is rather te- where the sum runs over all of the single-particle states.
dious, see Section III below. The main qualitative result Similarly, the energy current is defined as
is the appearance of the linear magnetoresistance in mod-
2
erately strong classical fields for `R  W  `R ωB τ τee
X
Q= vδf. (21b)
s  
vg W 1 1 1 Finally we introduce the “imbalance current” (cf. Ref. 8)
R∼B + .
ecT 2 τph τ τee (0) X
P = sign()vδf. (21c)
The result is governed by energy relaxation and quasi-
particle recombination due to electron-phonon interac- The appearance of this current reflects the independent
tion. On a qualitative level, this effect is independent conservation of the number of particles in the upper and
of details of the quasiparticle spectrum and can also be lower bands in graphene.
found in other two-component materials, such as narrow- All currents (21) vanish in equilibrium. In the degener-
band semiconductors, semi-metals, and macroscopically ate (or “Fermi-liquid”) limit, µ  T , the non-equilibrium
disordered media at the neutrality point24,33,34 . correction (19) to the distribution function contains a δ-
function1 . Thus, the above sums are dominated by the
states with energies close to the chemical potential  ∼ µ
II. FROM KINETIC EQUATION TO and all three currents become equivalent, see Eq. (10).
MACROSCOPIC DESCRIPTION

3. Non-equilibrium distribution function: infinite sample

In this Section we derive the macroscopic equations (1)
describing electronic transport in monolayer graphene in
the interaction-dominated regime. Within the standard linear response theory1 , one de-
scribes macroscopic states that are only weakly per-
turbed from equilibrium by some external probe. In this
case, the non-equilibrium correction δf to the distribu-
A. Boltzmann equation approach
tion function, or equivalently the function h, see Eq. (19)
are linear in the strength of the probe. At the same
1. Kinetic equation time, the function h (which we will hereafter refer to
as the non-equilibrium distribution function) has to be
We begin with the standard (Boltzmann) form of the proportional to the quasiparticle velocity, otherwise the
kinetic equation1,4,5,7–10 : macroscopic currents (21) will remain zero. Now, within
linear response the strength of the external probe is pro-
∂f ∂f  e  ∂f δf portional to the electric current and thus one can express
+v· + eE + v × B · =− + I, (18)
∂t ∂r c ∂p τ the non-equilibrium distribution function h as

where f is the distribution function, I is the collision in- h = A() j · v.

tegral due to Coulomb interaction, τ is the transport im-
purity scattering time (which may be energy-dependent), In an infinite sample all physical quantities are uni-
and δf is the non-equilibrium correction form. Moreover, in the degenerate regime A() → A(µ).
Such a description is completely equivalent to the stan-
  ∂f (0) dard linear response theory1 , but is more natural in sit-
δf = f − f (0) = f (0) 1 − f (0) h = −T h. (19) uations where one passes a current through a sample
∂
 6

rather than applies an electric field, for example in drag currents can be defined by Eqs. (21), while the corre-
measurements12–15 . sponding densities are determined by the equilibrium dis-
In nearly neutral graphene, the energy dependence of tribution function f (0) . As both the currents and den-
the distribution function becomes important. Taking ad- sities are independent of the coordinates and time, the
vantage of the collinear scattering singularity4,5,7,10,15 we corresponding continuity equations are trivially satisfied.
retain only those terms in the power series of the dis- Taking into account either sample geometry or local
tribution function h [or the prefactor A()] in , which perturbations leads to non-homogeneous distributions of
correspond to either zero modes of the collision integral, physical quantities. Within linear response, the nonuni-
or to its eigenmodes with non-divergent eigenvalues. In form deviations of the macroscopic densities are expected
general, there are three such terms to be small (as determined by the small driving force) and
can be accounted for by an additional term in the non-
A() = A0 + As sign() + A1 , equilibrium distribution function similar to Eq. (22), but
expressed in terms of the densities rather than currents.
where the coefficients Ai can be expressed in terms of the To a good approximation, electron and hole numbers
macroscopic currents by evaluating the sums in Eqs. (21). in graphene are conserved independently. Defined as
The resulting distribution function allows us to formulate X X
macroscopic or hydrodynamic equations describing elec- ne = f, nh = (1 − f ), (25a)
tronic transport in graphene. >0 <0
If the system is subjected to an external magnetic field,
the direction of the macroscopic currents may deviate they can be combined into the total charge density
from the driving bias. In this case, we may write the X
non-equilibrium distribution function in the form: n = e(ne − nh ), n = n0 + δn(r), δn(r) = e δf,
(25b)
2  
h= Ck () v · j + C⊥ ()v · (j × ez ) , (22) and the quasiparticle density
eν0 T vg2
X
ρ = ne + nh , ρ = ρ0 + δρ(r), δρ(r) = sign()δf.
where ν0 is the density of states
(25c)
X  ∂f (0)  NK Finally, we define the energy density
ν0 = − = , (23)
∂ 2πvg2 X X
u= f + (1 − f ), (25d)
>0 <0
with N being the degeneracy of the single-particle states
(in physical graphene N = 4).
X
u = u0 + δu(r), δu(r) = δf.
Based on the above arguments, we truncate the energy-
dependent functions Ci () as follows
Similarly to Eq. (10), all three densities become equiva-
(0) (s) (1) lent in the degenerate limit
Ci () = Ci + Ci sign() + Ci ,
e
n(µ  T ) = u(µ  T ) = eρ(µ  T ). (26)
leading to the three-mode approximation for the distri- µ
(j)
bution function. The coefficients Ci can be found by
requiring the distribution function (22) to yield the phys-
ical observables (21). The resulting expression is some- 5. Non-equilibrium distribution function: mesoscopic
what cumbersome and is given in Appendix A. For the sample
subsequent derivation of the macroscopic equations we
only need the energy dependence of the distribution func- Consider now a small, mesoscopic sample (still within
tion for which we use a short-hand notation (5) linear response). If boundary conditions are important,
then the non-equilibrium distribution function acquires
2v n  o
a non-homogeneous term that can be expressed in terms
h= 2
A + B + Csign() . (24)
eν0 T vg K of the fluctuating densities (25). Now we can write the
deviation of the distribution function (19) as follows
The vectors A, B, C are given in Eq. (A1).
∂f (0)
δf = −T (h + δh), (27)
∂
4. Macroscopic densities
where h is given by Eq. (5) and the extra term δh can be
written in a similar form
The above arguments rely on translational invariance
of the infinite system to establish the fact that all macro- 1 h  i
scopic physical quantities are homogeneous. Then the δh = a + b + c sign() . (28)
eν0 T K
 7

The coefficients a, b, and c are linear combinations of the Note, that in the general case of energy-dependent im-
inhomogeneous densities (25) [cf. Eq. (A2)]. In the de- purity scattering time τ () the numerical coefficients en-
generate limit a(µ  T ) = δn, while b(µ  T ) = c(µ  tering the equations (30) will change. This, however,
T ) = 0. At the Dirac point, these quantities simplify to does not yield any qualitatively new behavior26 . The
same applies to all of the equations derived below.
a(µ = 0) = δn, (29)

eδu γ0 2. Energy current

b(µ = 0) = − + eδρ ,
K∆(0) ∆(0)
The equation for the energy current can be obtained
eδu N2 (0) by multiplying the kinetic equation (18) by v and inte-
c(µ = 0) = γ0 − eδρ ,
K∆(0) ∆(0) grating similarly to the above. As a result we find

[cf. Eqs. (13)]. e π 0

N1 E + RH j × eB = R0 Q− I, (31a)
K eK 2
where similarly to Eq. (30c) we define
B. Macroscopic equations: infinite system
X
I0 = vI = 0. (31b)
In an infinite system physical quantities are uniform:
Physically, the latter equality follows from momentum
∂f
= 0. conservation and time-reversal properties of the scatter-
∂r ing probability. In double-layer systems, this conclusion
Substituting the distribution function (5) into the kinetic applies to the intralayer collision integral only, see below.
equation (18) and integrating over the energies and mo-
menta of the single-particle states, we find the set of lin-
ear equations describing the macroscopic currents. 3. Imbalance current

The imbalance current obeys the equation (that can be

1. Electrical current obtained by multiplying the kinetic equation by v sign()
and integrating over all single-particle states)
Multiplying Eq. (18) by ev and integrating, we find µ
the equation for the electric current (21a) e × eB = eR0 P − π I 00 ,
E + RH K (32a)
K eK
π where the counterpart of Eq. (30b) is [see Eq. (7)]
E + RH K × eB = R0 j − I, (30a)
e2 K
X δf
where the coefficients R0 and RH are given by Eq. (3) K
e = eK v . (32b)
||
and the Lorentz-force term contains the vector
X δf The integrated collision integral in Eq. (32) is given by
K = eK v , (30b)
 X A C
I 00 = v sign() I = − − , (32c)
which for the distribution function (5) has the form (6). eτvs eτss
The last term in the right-hand-side of Eq. (30) is the
see Appendix B 1 for details. In the degenerate limit
integrated collision integral
−1
X A C τss (µ  T ) → 0.
I=e vI = − − . (30c)
τvv τvs
C. Macroscopic equations in mesoscopic systems
For more details on integration of the collision integral
and the precise expressions for the relaxation rates see
Appendix B 1. In the two-mode approximation used in Is the case of relatively small, mesoscopic samples (see
Refs. 10,15 the imbalance current was not introduced and below for specific conditions) we can no longer rely on
−1 −1 translational invariance and need to determine spatial
the rate τvs did not appear. The rate τvv was previously
introduced in Ref. 10. variations of the physical quantities from Eq. (18). In
−1 other words, we need to take into account the gradient
In the degenerate limit, the relaxation rates τvv and
−1 term in the left-hand side of the kinetic equation.
τvs vanish due to the restored Galilean invariance. More-
−1 Proceeding similarly to the case of infinite systems,
over, the rate τvs vanishes at the Dirac point as well
we adopt the three-mode approximation (5) for the non-
−1
τvv (µ  T ), τs−1 (µ  T ) → 0; −1
τvs (µ = 0) = 0. equilibrium distribution function (19) and integrate the
 8

kinetic equation. This way, we arrive at the equations simplest strip geometry: we assume that our sample has
(1), which differ from the corresponding equations for in- the form of an infinite strip along the x-axis, with the
finite systems (30), (31), and (32) by the presence of the width W in the perpendicular y-direction. We will be in-
gradient terms in the left-hand side, which originate from terested in the effects of the external magnetic field that
integrating the gradient term v · ∇f in Eq. (18). This we assume to be directed along the z-axis, i.e. perpen-
yields three new macroscopic quantities, which physically dicular to the surface of the sample.
describe the flux density of the electric, energy, and im- Since the length of the strip is assumed to be very large,
balance currents. all physical quantities are independent of x. Consider
The flux density of the electric current is a tensor that the problem, where a current is being driven through the
is defined similarly to the usual momentum flux density1 strip. This fixes the average current density defined as
(which can be called flux density of the mass current)
X X (0)
ˆ
W/2
Παβ = e vα vβ f + e vα vβ (1 − f ) = Παβ + δΠαβ , 1
jx = dy jx (y).
>0 <0 W
(33) −W/2
(0)
where Παβ is the equilibrium tensor, while δΠαβ is the
inhomogeneous correction out of equilibrium. As there are no contacts along the strip, the y-component
One of the main steps in the derivation of the usual of any current must vanish at y = ±W/2:
hydrodynamics1 is to relate higher-rank tensors, such
as Παβ , to the hydrodynamic quantities such as the jy (±W/2) = Qy (±W/2) = Py (±W/2) = 0. (35a)
macroscopic currents. Depending on the degree of
approximation1–3 , one obtains various expressions for the Combining this argument with the continuity equation
higher-rank tensors which lead to various hydrodynamic (9a) yields
equations, such as the Euler or Navier-Stokes equations.
jy = 0. (35b)
In our linear-response theory the situation is simpler.
We already have the expression for the distribution func-
Finally, charge conservation requires
tion in terms of the macroscopic currents and densities,
see Eqs. (27), (5), and (28). All we need to do is to eval-
uate the expression (33) with that distribution function. ˆ
W/2

As a result, we define the quantity Π entering Eq. (1a): dy δn(y) = 0.

−W/2
eK
δΠαβ = δαβ Π.
π Our task is to find the average electric field in the strip
Similarly, we define the flux densities of the energy and
imbalance currents. Evaluating the resulting quantities ˆ
W/2

with the distribution function (28) we find 1

E= dy E(y),
W
1 h µi −W/2
Π= 2 a + bN1 + c , (34a)
e ν0 K
and hence the sheet resistance of the sample is
T 2 π2 µ2
  
1
Θ= 2 aN1 + bN2 + c 2 + 2 , (34b) Ex
e ν0 K 3 T R= . (36)
jx

T 2 π2 µ2
   
1 µ The electric field satisfies the Maxwell equations (8).
Ξ= 2 a +b 2 + 2 +c . (34c)
e ν0 K K 3 T In particular, in our geometry it follows from the second
of the equations (8) that the x-component of the electric
The macroscopic equations (1) are thus derived. Again, field is a constant
all numerical coefficients are specific to the case of energy-
independent τ . ∂Ez ∂Ex ∂Ex
− =0 ⇒ = 0,
∂x ∂z ∂z
III. FINITE-SIZE EFFECTS IN NEUTRAL
MONOLAYER GRAPHENE ∂Ey ∂Ex ∂Ex
− =0 ⇒ = 0,
∂x ∂y ∂y
A. Boundary conditions
or in other words
Solutions of the finite-size problems are largely deter-
mined by the boundary conditions. Here we consider the Ex = Ex = const. (37)
 9

B. Mesoscopic graphene sample at the Dirac point only y components. The continuity equations can be re-
written as follows
Consider the set of equations (1) at the Dirac point.   !
d q δu
f
Given the absence of the Hall effect, the charge density = −Tph e ,
b (41a)
dy p δρ
can be assumed to be uniform. In this case, we find
πj where δu
f = eδu/K and δρ e = eδρ, and [cf. Eq. (9)]
E + RH K × eB = R0 j + 2
, (38a)
2e T τvv ln 2
!
1 γ0 N2 (0) γ0
1 τ + τ − τ − τ
Tbph = − Eb Ec Ec
N2 (0)
Eb
. (41b)
∆(0) − τ1 − τγ0 τIc + τIb
γ0
Ib Ic
∇δu e
− + RH j × eB = R0 Q, (38b) Combining the continuity equations (41) with the lin-
2eν0 T ln 2 2T ln 2
ear response equations (38b) and (38c), we find
1 d2 q
     
∇δρ πC = T cR q + RH Ex Tbph 1 , (42a)
bph M
− + RH j × eB = eR0 P + 2 , (38c) e2 ν0 dy 2 p p Rj 1
eν0 2e T τss ln 2
where we have excluded the y-dependent electric current
where the vectors K and C [given in Eqs. (13) above] are
using Eq. (40). The resistance matrix M
cR is given by
 
γ0 − 1 eQ
 
K= − eP γ2 , cR = R0 − δR
M
δRγ2
, (42b)
∆(0) 2T ln 2 −Rq − δR R0 + Rq γ1 + δRγ2

and where
R2H
eQ N2 (0) δR = − (γ0 − 1) , (42c)
C = γ0 − eP , Rj ∆(0)
2T ∆(0) ln 2 ∆(0)
γ0 π
where the numerical coefficients γ0 and γ2 are given in Rq = − , (42d)
Eqs. (13b) and (13d). The parameters R0 and RH are ∆(0) 2e2 T τss ln 2
evaluated at µ = 0
N2 (0)
π γ1 = ≈ 2.3721. (42e)
R0 → R0 (µ = 0) = 2 , (38d) γ0
2e T τ ln 2
Note, that the same matrix appears in Eqs. (14), if one
writes the second and the third equations (14b) and (14c)
πvg2 B in matrix form.
RH → RH (µ = 0) = . (38e)
4ecT 2 ln2 2 The differential equation (42a) admits a formal matrix
solution. Using the hard-wall boundary conditions (35a)
The relaxation times τvv and τss are evaluated at the and averaging over the width of the sample we find
Dirac point as well.    
As we have already mentioned, in these equations all −1
 
q RH tanh L d
ph W/2
 
−1 1
quantities are independent of the coordinate x along the = Ex  − 1 MR
 c ,
p Rj −1 1
strip, such that δu = δu(y) and δρ = δρ(y). Taking into L
d W/2
ph
account Eq. (35b) we notice, that all the vectors in the (43a)
left-hand sides of Eqs. (38b) and (38c) are directed along where
the y-axis. Thus, we find that both the energy current q
−1
and imbalance current are orthogonal to the electric cur- L
d
ph = e2 ν0 Tbph M
cR . (43b)
rent and can be written in the form
Now, we use the solution (43) to determine the auxil-
e iary quantity Ky
Q = (0, q), eP = (0, p). (39)
2T ln 2
γ0 − 1
Consequently, the vector K is also pointing in the y di- Ky = (q − γ2 p) , (44)
∆(0)
rection. Therefore, the y-component of Eq. (38a) sim-
ply reads Ey = 0, as it should be. The x-component of which we then use in Eq. (40) in order to find the resis-
Eq. (38a) now reads tance of the sample:

π Rj
Ex + RH Ky = Rj jx , Rj = R0 + . (40) R=     .
2e2 T τvv ln 2 tanh L−1
ph W/2
d  
−1 1


1 − δR 1 −γ2  − 1 MR
 c
L−1
d
ph W/2
1
The remaining equations (38b) and (38c), as well as the
corresponding continuity equations (9b) and (9c) contain (45)
 10

R (kΩ) The result (17) is shown by the top curve in Fig. 1, where
400 ●
●●
●● we present magnetoresistance in graphene at charge neu-
●
●●
●●
●
●●
●●
●
●●
●●
●
●●

50 300
●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●
trality (45) for samples of different widths and for realis-
●
●
●●
●
●

200 ●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
tic sample parameters.
●●
●
●●
●
●

40 ●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●
In narrower samples the magnetoresistance (45) weak-
100 ●●
●
●●
●●
●
●●
●●
●

●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
ens, see Fig. 1. In classically strong fields, RH  Rj ,
●
●●
●
●●
●
●
●
●●
●●
●
●
2
30 5 10 15 20 one finds an intermediate regime, `R  W  `R ωB τ τee ,
where the system exhibits linear magnetoresistance
20 s
vg R j W 2
R∼B . (46)
10 B (T) c T 3 τph
0.2 0.4 0.6 0.8 1.0 1.2 1.4
The recombination length is inversely
p proportional to the
δR(B)/B 2 magnetic field `R ∼ [cT /(evg B)] τph /τee . Linear mag-
netoresistance is illustrated in the inset in Fig. 1.
15
●
●●●●●●●
●●●●●●
2.0
●●●●●●
●●●●●●
●●●●●
●●●●●
IV. TRANSPORT PROPERTIES OF
●●●●●●●

10 1.5 ●●●●
●●●
●●●
●●
●●●●
●
DOUBLE-LAYER SYSTEMS
●●●
●●●
●●
●●
●
1.0 ●●
●●
●●
●●

0.5 ●●
●●
●● Double-layer systems are often used to study transport
5 ●●
●●
●
●●
●●●
●●●●● properties of two-dimensional systems. In comparison to
0.2 0.4 0.6 0.8 1.0
single-layer devices, one can can study two additional
W (μm) phenomena: (i) the relatively weak effect of the second
50 100 150 200
layer on the single-layer transport properties, and (ii) the
strong Coulomb drag effect. The latter is due to inter-
FIG. 1: (Color online) Upper panel: magnetoresistance in
graphene at charge neutrality. The uppermost curve shows layer electron-electron scattering and is important only
the result (17) for a macroscopic sample. The lower curves in the academic case of disorder-free graphene in the de-
show the result (45) for sample widths W = 100, 20, 10, 2 µm generate limit, where it provides the only source of re-
(top to bottom). The results are calculated for realistic sistance. In all other cases, the effect is relatively small
values of parameters: T = 240 K, τ −1 = 50 K, τee = 0.2τ , due to the weakness of the interlayer interaction. On the
τph = 20τ . The inset illustrates the linear magnetoresistance other hand, the drag effect in double-layer systems12–15
for W = 0.1 µm. The dashed line is a guide to the eye. Lower is solely due to the interlayer interaction and has no
panel: curvature of the above magnetoresistance in weak fields counterpart in non-interacting systems. Given the ex-
(in units of kΩ/T2 ) as a function of W . Green line shows the tensive theoretical literature devoted to Coulomb drag
prefactor in Eq. (17). The inset shows the region W < 1 µm.
(see Refs. 10,15,25–27 and references therein), here we
focus on the two following issues. Firstly, we compute
the leading correction to the Fermi-liquid prediction for
This is the final result of this Section. Here the field- the drag coefficient in the degenerate regime µ  T .
dependent resistance δR is given in Eqs. (42c), the nu- Secondly, we discuss the drag effect at charge neutral-
merical coefficient γ2 in Eq. (13d), and the matrices Mc ity, where our theory provides microscopic justification
−1 to the phenomenological treatment of the effect of giant
and L are defined by Eqs. (42b), (43b), and (41b).
d
ph
Qualitative behavior of the result (45) is determined magnetodrag at charge neutrality given in Ref. 15.
by the interplay of sample geometry, magnetic field, and Transport properties of double-layer systems can be
electron-phonon scattering. described within the same macroscopic approach to the
In the most narrow samples (formally, in the limit Boltzmann equation as we have used above in the context
W → 0) the square bracket in Eq. (45) vanishes and the of monolayer graphene. Now we introduce the system of
resulting resistance is independent of the magnetic field two coupled kinetic equations similar to Eq. (18):
(see Fig. 1). Physically, this happens when the electron- ∂f1 ∂f1  e  ∂f
1
phonon length scale ` given by the largest eigenvalue of + v1 · + eE 1 + v 1 × B · = (47)
∂t ∂r c ∂p
the operator (43b) exceeds the sample width, W  `.
In widest samples, W  `R ωB 2
τ τee , [here `R is the re- δf1
=− + I11 (f1 ) + I12 (f1 , f2 ),
combination length given by the smallest eigenvalue of τ
the operator (43b)] the width-dependent term in Eq. (45)
can be neglected and we reproduce Eq. (17) as ∂f2 ∂f2  e  ∂f
2
+ v2 · + eE 2 + v 2 × B · =
∂t ∂r c ∂p
−1
R2H
  

−1 1 κ δf2
1 + δR 1 −γ2 M
c
R = 1 + . =− + I22 (f2 ) + I21 (f1 , f2 ).
1 R0 Rj ∆(0) τ
 11

Now the distribution functions fi carry the layer index interlayer friction by means of the corresponding scat-
i = 1, 2. The single-layer collision integrals Iii (fi ) are the tering time τD . It is well known26 , however, that the
same as one used in the above discussion of monolayer traditional Fermi-liquid theory of Coulomb drag is appli-
graphene, see Eq. (B1) and Appendix B 1 for details. The cable only for very large densities, far beyond the current
interlayer coupling is described by the inter-layer collision experimental range12–15 .
integrals I12 (f1 , f2 ) and I21 (f1 , f2 ), see Appendix B 2. Leading corrections to the Fermi-liquid results can be
described in terms of small deviations of the energy and
imbalance currents from their limiting values (10). It is
A. Infinite system intuitively clear that the imbalance current approaches
the limiting value exponentially. In contrast, the en-
Within linear response, deviations of the distribu- ergy current is expected to exhibit power law corrections.
tion functions fi from equilibrium can be described by These can be demonstrated by the following arguments.
Eq. (19). In an infinite system, we can still use the three- The drag measurement is performed by passing a cur-
mode approximation (5) for the non-equilibrium distri- rent j 1 = j1 ex through one of the layers (the active layer)
bution functions hi and measuring the induced electric field (or voltage) in
the other, passive layer. Consider for simplicity iden-
tical, macroscopic layers. In the degenerate regime, we
 
2v 
hi = Ai + B i + C i sign() . (48) may set eP 1 = j 1 (since the deviations from this equality
eνi T vg2 Ki
are exponentially small in T /µ), neglect small differences
The vectors in Eq. (48) can be read off Eq. (A1), with between various interlayer relaxation rates, disregard in-
the self-evident addition of the layer index. tralayer interaction effects, and assume interlayer ther-
malization that yields [see, e.g., Eq. (B27)]
e e
Q = Q1 − N 1 j 1 .
1. Macroscopic equations µ 2 µ
As a result, the macroscopic equations have the form
Here we would like to describe the double-layer system
similarly to the above macroscopic description of mono- E 1 + RH K1 × eB = (R0 + RD ) j 1 , (50a)
layer graphene. Integrating the kinetic equations (47) we
obtain the following equations for the macroscopic cur- N1 + 1 e
rents (here i refers to a layer, while j to the other layer) N1 E 1 + RH j 1 ×eB = R0 Q1 +N1 RD j 1 , (50b)
2 µ

(i) (i) π π
E i +RH Ki ×eB = R0 j i − I ii − I ij , (49a) E 2 + RH K2 × eB = −RD j 1 , (50c)
e2 Ki e2 K i
2
where RD = π/(e µτD ) [see Eq. (B14)] is the standard
drag resistivity. The auxiliary vectors in the Lorentz
(i) (i) (i) e π 0
N1 E i + RH j i × eB = R0 Q − I , (49b) terms read
Ki i eKi2 ij  
e e
K1 ≈ j 1 − Q1 − N1 j 1 , K2 ≈ − Q2 .
µ µ
µi (i) e (i) π 00 π 00
E i + RH K i × eB = eR0 P i − I ii − I . Neglecting small deviations of the energy current in
Ki eKi eKi ij
(49c) the active layer from its limiting value (e/µ)Q1 = j 1 ,
Here the intralayer collision integrals I ii and I 00ii are still we find the standard drag effect (defined according to
described by Eqs. (30c) and (32c), respectively, with the Refs. 12,15)
obvious addition of the layer index. The interlayer col-
E2x
lision integrals are described in detail in Appendix B 2. Q2 = 0 ⇒ RD = = −RD , (51)
One can recast them in terms of relaxation rates and j1
rewrite the equations (49) in the form (1). The resulting which is independent of magnetic field.
equations contain a rather large number of terms. There- In contrast, taking into account a small deviation of Q1
fore, below we will discuss the most interesting limiting from its limiting value, we find that the leading correction
cases, where they can be significantly simplified. to RD depends on magnetic field
π 2 T 2 R2H R0
RD = −RD + . (52)
2. Coulomb drag in degenerate limit 6µ2 R20 + R2H
Same calculation also yields the Hall drag resistivity:
In the degenerate limit Coulomb drag can be de-
scribed be means of the generalized Ohm’s (or Drude) E2y π2 T 2 R3H
RDH = =− . (53)
equations10 with the phenomenological term describing j1 6µ R0 + R2H
2 2
 12

|RD | 3. Macroscopic theory at the neutrality point

At the charge neutrality (or double Dirac) point we

may consider the two layers to be identical. With the help
of the thermalization conditions (B28) and Eqs. (13), we
find the following macroscopic description of an infinite
double-layer system [cf. Eqs. (14)]. The first layer is
described by the equations [with the auxiliary vectors
given by Eqs. (13)]
 
π
E 1 + RH K1 × eB = Rj + 2 j , (56a)
μ 2e T τvv,12 ln 2 1

FIG. 2: (Color online) Schematic illustration of corrections to 1 e

(1)
the Fermi-liquid predictions for the drag coefficient. The blue RH j 1 × eB = R0 Q , (56b)
dashed curve represents the Fermi-liquid result RD ∼ T 2 /µ2 . 2 2T ln 2 1
The green dotted curve represents the exponential approach
to RD within the two-mode approximation that retains the −1 −1


electric and imbalance currents. The red solid line shows the 1 (1) π τss + τss,12
RH j 1 × eB = eR0 P 1 + C 1 . (56c)
result (52) of the three-mode approximation approaching the 2 2e2 T ln 2
Fermi-liquid regime as a power law in T /µ.
The energy and imbalance currents in the second layer
are determined by the thermalization conditions (B28).
The equation for the electric current is similar to
In contrast to the traditional theories of Coulomb drag, Eq. (56a). However, if we consider the typical setup for
the above results contain contributions that do not di- drag measurements, where no electric current is allowed
rectly depend on any interlayer electron-electron scatter- to flow in the second layer, then the equation simplifies
ing rate. Instead, this is the effect of interlayer thermal- to
ization.
At the same time, the presence of the second layer leads E 2 + RH K1 × eB = 0. (56d)
to the appearance of magnetoresistance in the first layer
(which vanishes in the limit µ → ∞) The solution of the equations (56) is identical to that
described in Section I. The presence of the second layer
does not significantly change transport in the first layer,
π 2 T 2 R2H R0 in particular, Eqs. (56) still predict positive magnetore-
R(µ  T ) = R0 + RD + , (54)
6µ2 R20 + R2H sistance. The Hall classical effect does not appear at
charge neutrality as it should be.
as well as a small correction to the Hall coefficient For the second layer, this theory predicts positive
Coulomb drag [defined in Eq. (51)] in agreement with
π2 T 2 R3H qualitative arguments of Ref. 15. In order to explain
RH (µ  T ) = RH − . (55) the experimentally observed negative drag12,15 the the-
6µ2 R20 + R2H
ory needs to be refined as follows: (i) the finite width W
of the system should be taken into account; the relative
The above corrections exhibit power-law dependence parameter is W/`R , where `R is the phonon-induced re-
on the small ratio T /µ. This is in contrast to the expo- laxation length, see Section III; (ii) the above interlayer
nential approach to the Fermi-liquid limit that was found thermalization procedure should be improved to take into
within the two-mode approximation in Ref. 15, as illus- account the finite interlayer electron-electron relaxation
trated in Fig. 2. Indeed, the phenomenological model of rate. This is outlined in Section IV B.
Ref. 15 included the electric and imbalance currents. The
latter approaches its limiting value eP 1 = j 1 only expo-
nentially. Hence the exponentially small magnetodrag in B. Mesoscopic system at charge neutrality
doped graphene found in Ref. 15 (in notable disagreement
with experimental data) is an artifact of neglecting the In a mesoscopic system, we need to take into account
energy current in the simplified phenomenological model. spatial inhomogeneity of the macroscopic currents and
At the same time, at charge neutrality the imbalance cur- densities. In this case, the non-equilibrium distribution
rent is proportional to the energy current while both are function acquires the additional contribution (28)
orthogonal to j. Hence the phenomenological model of  
Ref. 15 captures qualitative physics at the Dirac point, 1 
δhi = ai + bi + ci sign() . (57)
see below. eνi T Ki
 13

Similarly to the situation in monolayer graphene, macro- tegral. Under such assumptions, transport properties of
scopic equations in double-layer systems acquire gradi- graphene can be described in terms of the three macro-
ent terms. The resulting equations contain two copies of scopic currents, j, P , and Q. In small, mesoscopic
Eqs. (1) where one has to add interlayer scattering rates samples physical properties become inhomogeneous and
from the right-hand side of Eqs. (49), two copies of con- we need to introduce the inhomogeneous corrections to
tinuity equations similar to Eqs. (9) where one has to in- the corresponding charge, energy, and imbalance densi-
clude additional contributions due to interlayer electron- ties. In that case, the complete set of macroscopic equa-
electron interaction (see Appendix D), and the Maxwell tions includes three equations (1) for the currents, which
equations (8). A general solution to this system of equa- can be viewed as the generalization of the Ohm’s law,
tions is rather convoluted. Hence here we limit ourselves three continuity equations, and the Maxwell equations,
to a qualitative discussion. describing the self-consistent electromagnetic field.
Of particular interest is the drag effect at charge neu-
trality, where the experiment12,15 shows an unusually Solving the macroscopic equations, one can find tem-
strong dependence of RD on the external magnetic field, perature, density, and geometry (i.e. the system size)
i.e. giant magnetodrag. The problem of Coulomb drag in dependence of transport coefficients. For general doping
graphene at charge neutrality was previously addressed this is a formidable computational task. However, far
in Refs. 15,25 based on a two-fluid approach. As shown in away from charge neutrality (in the degenerate or “Fermi-
Section III above, the energy and imbalance currents in liquid” regime) all the three currents become equivalent
the active layer at the neutrality point are parallel to each and the theory reduces to the single-mode equation (11)
other and orthogonal to the driving current Q1 kP 1 ⊥ j 1 . with the Drude transport coefficients (12) as it should,
Excluding one of these currents from the macroscopic given that no quantum interference processes were taken
equations one effectively derives a two-fluid model. Thus into account.
our theory provides a microscopic foundation for the ear-
lier phenomenological models. The key point is that the Exactly at the Dirac point, the theory simplifies as
currents Q and P can be transferred between the layers well and allows for analytic solutions. We have shown
by means of the interlayer interaction in contrast to the that graphene at charge neutrality exhibits strong posi-
electric current, whose transfer is forbidden by the exact tive magnetoresistance (45). Specifically, the resistance
electron-hole symmetry at the Dirac point. behaves quadratically in not too strong fields, Eq. (17),
In the limit of infinitely fast interlayer thermalization and crosses over to the linear dependence (46) once the
(discussed above in Section IV A 2) the energy and imbal- field increases beyond a certain value determined by the
ance currents in the two layers have the same direction sample width and quasiparticle recombination rate due
leading to positive drag. Taking into account finiteness of to electron-phonon interaction, see Fig. 1.
the corresponding relaxation rates (Appendix D) refines
the theory in analogy with including viscous terms into Strong positive magnetoresistance in graphene was ob-
standard hydrodynamic theory1,32 . The resulting theory served in Refs. 17–19,22 at charge neutrality. Our results
contains four differential equations for the energy and im- qualitatively agree with the experimental data. More-
balance currents [cf. Eq. (42a) in the single-layer case]. over, our theory can be generalized to account for macro-
If the sample is wide enough (i.e. if the width of the sam- scopic inhomogeneities that were discussed as a possible
ple W is larger than the phonon-induced recombination source of magnetoresistance in Refs. 17,18. Further ex-
length), the energy and imbalance currents in the two perimental studies of magnetoresistance in high-mobility
layers flow in the same direction and the system exhibits graphene samples (including the dependence on the sam-
positive drag as discussed above. On the contrary, in ple width) would be of great interest.
narrow samples it is the inhomogeneous energy and im-
In double-layer systems, our theory provides the mi-
balance densities in the two layers that coincide, pushing
croscopic justification of the phenomenological treatment
the currents in the opposite directions and yielding neg-
of the giant magnetodrag problem suggested in Ref. 15.
ative drag15,25 . Similarly to the discussion in Section III,
The three-mode Ansatz allows for more precise quanti-
the magnetic field dependence of the result is quadratic
tative description of the effect. In particular, we have
in weak fields and linear in classically strong fields.
calculated the leading correction to the Fermi-liquid pre-
diction for the drag coefficient in doped graphene. Phys-
ically, the resulting magnetodrag (52), as well as Hall
V. SUMMARY drag (53) is due to interlayer thermalization. Treating
all three modes on equal footing allows us to remove the
We have developed a macroscopic (hydrodynamic-like) artifacts of two-mode approximations, see Fig. 2.
description of electronic transport in graphene. Our
approach is based on the “three-mode” Ansatz for the In this paper we have limited ourselves to linear re-
non-equilibrium distribution function in graphene. This sponse theory. A generalization of our approach to non-
Ansatz is justified in the interaction-dominated regime linear hydrodynamics in graphene will be reported in a
by the collinear scattering singularity in the collision in- subsequent publication32 .
 14

Acknowledgments work was supported by the Dutch Science Foundation

NWO/FOM 13PR3118, the EU Network Grant Inter-
We acknowledge helpful conversations with H. Weber, NoM, DFG-SPP 1459, DFG-SPP 1666, GIF, and the
U. Briskot, A. Levchenko, L. Ponomarenko, P. Alekseev, Humboldt Foundation.
V. Kachorovsky, Yu. Vasiliev, and A. Dmitriev. This

Appendix A: Non-equilibrium distribution function in the three-mode approximation

In this Appendix we give the complete expression for the non-equilibrium distribution function h in monolayer
graphene in terms of the three macroscopic currents (21) and densities (25)
2v n  o
h= 2
A + B + Csign() , (A1a)
eν0 T vg K
e µ
− N1 j µT 2 π 2 µ2 eP − K T 2 π2 µ2
     
KQ j µ
A=j+ N1 − 3 + 2 + N2 − N1 2 + 2 , (A1b)
∆ K 3 T ∆ K K 3 T
e µ
− N1 j µ2 eP − K j T 2 π2 µ2
     
KQ µ
B= −1 + + 2 − N1 (A1c)
∆ K2 ∆ K2 3 T K
e µ
− N1 j T 2 π 2 µ2 eP − K
   
KQ µ j 2 
C= 2
+ 2
− N 1 + N1 − N2 (A1d)
∆ K 3 T K ∆

1 h  i
δh =
a + b + c sign() , (A2a)
eν0 T K
e µ
δu − N1 δn µT 2 π 2 µ2 eδρ − K T 2 π2 µ2
     
δn µ
a = δn + K N1 − 3 + 2 + N2 − N 1 2 + 2 , (A2b)
∆ K 3 T ∆ K K 3 T
e µ
− N1 δn µ2 eδρ − K δn T 2 π 2 µ2
     
K δu µ
b= − 1 + + − N 1 (A2c)
∆ K2 ∆ K2 3 T2 K
e µ
− N1 δn T 2 π 2 µ2 eδρ − K
   
K δu µ δn  2 
c= 2
+ 2
− N 1 + N1 − N 2 (A2d)
∆ K 3 T K ∆
where N1 is a dimensionless quantity proportional to the carrier density in graphene
ˆ∞ h i X  ∂f (0) 
n0 = d ν() f (0) (; µ) − f (0) (; 0) ,  − = 2n0 , 2n0 = N ν0 µ = N1 ν0 K. (A3a)
∂
−∞

This dimensionless function depends only on the ratio x = µ/T and has the following asymptotic behavior:
 h i
2 x 5x2

2 − 6 xln 2 + . . . , x  1,  ln 2 1 − 24 ln 2 + . . . , x  1,

N (x) ≈ , N1 (x) ≈ (A3b)
π2 2
1 + 3x2 + . . . , x  1,
 
1 + π2 + ..., x  1.

3x
Similarly, the dimensionless quantity N2 represents a similar sum
  
9ζ(3) 3 9ζ(3)
 (0)
  3 2 + 4 ln2 2 1− 16 ln2 2
x2 , x  1,
1 X
2 ∂f  8 ln
N2 (x) =  − ≈ , (A4)
ν0 K 2 ∂  2
1 + πx2 , x  1.


and the dimensionless quantity ∆ is

 −1.13025 + 0.9348x2 , x  1,

2
T2 π2 µ2 µ2
   
µ 2


∆= + 2 − N1 + N2 − N1 −1 ≈ . (A5)
K2 3 T K K2 2
−x
− 4π e , x  1.

3x 3
 15

Appendix B: Relaxation rates due to electron-electron interaction

1. Monolayer graphene

Within linear response the collision integral in Eq. (18) can be linearized with the help of Eq. (19) as follows
h ih i
(0) (0) (0) (0)
X
I= W12,10 20 f1 f2 1 − f10 1 − f20 [h10 + h20 − h1 − h2 ] . (B1)
1,10 ,20

a. Collision term in the equation for the electric current

Following the usual steps involving introduction of transferred energy ω and momentum q, we find for the integrated
collision integral Eq. (30c) appearing in the equation (30) for the electric current:
ˆ 2
e d qdω |U (q, ω)|2
I=− (B2)
32 (2π)3 sinh2 (ω/2T )
 
X
3 2 0 0 1 − µ 1 + ω − µ
× (2π) |λvv0 | δ(1 − 1 + ω)δ(p1 − p1 + q) tanh − tanh
0
2T 2T
1,1
 
X 2 − µ 2 − ω − µ
× (v 2 − v 02 )(2π)3 |λvv0 |2 δ(2 − 02 − ω)δ(p2 − p02 − q) tanh − tanh
0
2T 2T
2,2

× [h10 + h20 − h1 − h2 ] .

Here |λvv0 |2 are the “Dirac factors”.

Taking into account the explicit form of the distribution function (5), summations over states 1, 10 and 2, 20 factorize.
Consequently, one can evaluate them separately. The resulting expressions can be denoted as follows:
 
3
X
2 0 0 1 − µ 1 + ω − µ
(2π) |λvv0 | δ(1 − 1 + ω)δ(p1 − p1 + q) tanh − tanh = Y00 (q, ω), (B3a)
0
2T 2T
1,1

 
X 1 − µ 1 + ω − µ q
(2π) 3
(v 01 2
− v 1 ) |λvv0 | δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh = vg Y0A (q, ω), (B3b)
2T 2T q
1,10

 
X 1 − µ 1 + ω − µ q
(2π)3
(v 01 sgn 01 − v 1 sgn 1 ) |λ vv 0
2
| δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh = vg Y0C (q, ω),
2T 2T q
1,10
(B3c)
 
(v10α − v1α ) v10β − v1β |λvv0 |2 δ(1 − 01 + ω)δ(p1 − p01 + q)
X
(2π)3
1,10

qα qβ
 
1 − µ 1 + ω − µ
× tanh − tanh = 2 vg2 YAA (q, ω), (B3d)
2T 2T q

 
(v10α − v1α ) v10β sgn 01 − v1β sgn 1 |λvv0 |2 δ(1 − 01 + ω)δ(p1 − p01 + q)
X
(2π)3
1,10

qα qβ
 
1 − µ 1 + ω − µ
× tanh − tanh = 2 vg2 YAC (q, ω), (B3e)
2T 2T q

All of thus defined functions Yij (q, ω) obey the trivial symmetry property

Yij (q, ω) = −Yij (−q, −ω). (B4)

 16

Since the collision integral I has the dimension of inverse time, it is convenient to introduce the transport scattering
times due to Coulomb interaction. Given the multitude of terms in the kinetic equation, we choose to define several
interaction-related time scales. In the current equation, two such time scales appear (if the arguments of Yi (q, ω)
have their standard form we omit them for brevity):

“ “ ˆ
1 1 2

e d2 qdω |U (q, ω)|2
= Y00 YAA − Y0A , where ··· = ..., (B5)
τvv ν0 32T (2π)3 sinh2 (ω/2T )

“
1 1
= (Y00 YAC − Y0A Y0C ) . (B6)
τvs ν0

−1 −1
Both time scales τvv and τvs vanish in the Fermi-liquid limit (physically, due to the restored Galilean invariance).
−1
On the other hand, at charge neutrality τvs = 0, since

Y0A (µ = 0) = YAC (µ = 0) = 0, (B7)

−1
while τvv remains finite.
Using the above relaxation rates, we can write the integrated collision integral in equation (30) in the form (30c).

b. Collision term in the equation for the imbalance current

Treating the collision integral in Eq. (32) in the same way as Eq. (B2) above, we find:

ˆ
00 1 d2 qdω |U (q, ω)|2
I =− (B8)
32 (2π)3 sinh2 (ω/2T )
 
X 1 − µ 1 + ω − µ
× 3 2
(2π) |λvv0 | δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh
2T 2T
1,10
 
X 2 − µ 2 − ω − µ
× (v 2 sgn 2 − v 02 sgn 02 )(2π)3 |λvv0 |2 δ(2 − 02 − ω)δ(p2 − p02 − q) tanh − tanh
0
2T 2T
2,2

× [h10 + h20 − h1 − h2 ] .

Following the same line of argument as in the previous Appendix, we introduce another time scale
“
1 1 2


= Y00 YCC − Y0C , (B9)
τss ν0

where we had to introduce another quantity YCC similarly to Eqs. (B3):

 
(v10α sgn 01 − v1α sgn 1 ) v10β sgn 01 − v1β sgn 1 |λvv0 |2
X
(2π)3
1,10

qα qβ
 
1 − µ 1 + ω − µ
×δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh = 2 vg2 YCC (q, ω). (B10)
2T 2T q

As a result, the integrated collision term (B8) takes the form (32c).
 17

2. Double-layer system

a. Collision term in the equation for the electric current

The integrated inter-layer collision integral has a form, similar to Eq. (B2),

ˆ
e d2 qdω |U12 (q, ω)|2
I 12 =− (B11)
32 (2π)3 sinh2 (ω/2T )
 
X
3 2 0 0 1 − µ2 1 + ω − µ2
× (2π) |λvv0 | δ(1 − 1 + ω)δ(p1 − p1 + q) tanh − tanh
0
2T 2T
1,1
 
X
0 3 2 0 0 2 − µ1 2 − ω − µ1
× (v 2 − v 2 )(2π) |λvv0 | δ(2 − 2 − ω)δ(p2 − p2 − q) tanh − tanh
0
2T 2T
2,2

× [h2,10 + h1,20 − h2,1 − h1,2 ] ,

except than now the chemical potentials and the non-equilibrium distribution functions carry the layer index (i.e. h2,1
stands for the distribution function in layer 2 describing the state 1) and the potential U12 (q, ω) describes interlayer
interaction.
Consequently, the auxiliary functions (B3) as well as the densities of states, will now also acquire the layer index.
−1
This leads to a larger number of decay rates in comparison to τee and τs−1 . Since most of them vanish at the Dirac
point, we express the collision integral (B11) as follows:

˛ ˛ ˛
e (1) (2) e (1) (2) evg (1) (2)
I 12 = −A1 YAA Y00 + A2 Y0A Y0A − B1 qY0A Y00 (B12)
ν01 ν02 ν01 K1
˛ ˛ ˛
evg (1) (2) e (1) (2) e (1) (2)
+B2 qY0A Y00 − C 1 YAC Y00 + C 2 Y0A Y0C ,
ν02 K2 ν01 ν02

where

˛ ˆ
1 d2 qdω |U12 (q, ω)|2
··· = .... (B13)
32T (2π)3 sinh2 (ω/2T )

The first two terms are familiar from the traditional theory of Coulomb drag26 . In particular, the usual “drag rate”
−1
τD is given by the second term

˛
1 e (1) (2)
= Y0A Y0A . (B14)
τD ν02

In the degenerate regime, the relaxation rates in the first two terms become identical. The traditional theory is then
recovered by taking into account interlayer thermalization, see below.
At the neutrality point this expression simplifies significantly. Indeed, taking into account Eq. (B7) we find

˛
e A1
I 12 (µ1 = µ2 = 0) = −A1 YAA Y00 = − , (B15)
ν0 τvv,12

where the layer indices can be omitted since at the neutrality point the layers are identical to each other. On the
other hand, the new relaxation rate 1/τee,12 differs from Eq. (B5) insofar it reflects the interlayer interaction potential
U12 (q, ω).
 18

b. Collision term in the equation for the energy current

The equation for the energy current is obtained by multiplying the kinetic equation by v and integrating over all
states. Then, similarly to Eq. (B11) we find
ˆ 2
0 1 d qdω |U12 (q, ω)|2
I 12 = − (B16)
32 (2π)3 sinh2 (ω/2T )
 
X
3 2 0 0 1 − µ2 1 + ω − µ2
× (2π) |λvv0 | δ(1 − 1 + ω)δ(p1 − p1 + q) tanh − tanh
0
2T 2T
1,1
 
X
0 0 3 2 0 0 2 − µ1 2 − ω − µ1
× (2 v 2 − 2 v 2 )(2π) |λvv0 | δ(2 − 2 − ω)δ(p2 − p2 − q) tanh − tanh
0
2T 2T
2,2

× [h2,10 + h1,20 − h2,1 − h1,2 ] ,

where (due to momentum conservation)
2 v 2 − 02 v 02 = vg2 q. (B17)
In contrast to monolayer graphene [see Eq. (31b)], the integrated collision integral in the double-layer system does
not vanish. Similarly to Eq. (B12) we find
˛ ˛ ˛
vg (1) (2) vg (1) (2) vg2 (1) (2)
I 012 = −A1 qY0A Y00 + A2 qY00 Y0A − B1 q 2 Y00 Y00 (B18)
ν01 ν02 ν01 K1
˛ ˛ ˛
vg2 2 (1) (2) vg (1) (2) vg (1) (2)
+B2 q Y00 Y00 − C 1 qY0C Y00 + C 2 qY00 Y0C ,
ν02 K2 ν01 ν02
At the neutrality point, the first two terms vanish [similarly to Eq. (B15)]
˛ ˛
0 vg2 2 vg
I 12 (µ1 = µ2 = 0) = − (B1 − B2 ) q Y00 Y00 − (C 1 − C 2 ) qY0C Y00 . (B19)
ν0 K(0) ν0

c. Collision term in the equation for the imbalance current

The integrated interlayer collision integral in the equation for the imbalance current takes the form
ˆ 2
1 d qdω |U12 (q, ω)|2
I 0012 = − (B20)
32 (2π)3 sinh2 (ω/2T )
 
X 1 − µ2 1 + ω − µ2
× (2π)3 |λvv0 |2 δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh
0
2T 2T
1,1
 
X 2 − µ1 2 − ω − µ1
× [v 2 sign(2 ) − v 02 sign(02 )] (2π)3 |λvv0 |2 δ(2 − 02 − ω)δ(p2 − p02 − q) tanh − tanh
0
2T 2T
2,2

× [h2,10 + h1,20 − h2,1 − h1,2 ] .

Similarly to Eqs. (B12) and (B18), we can re-write Eq. (B20) as follows
˛ ˛ ˛
1 (1) (2) 1 (1) (2) vg (1) (2)
I 0012 = −A1 YAC Y00 + A2 Y0C Y0A − B1 qY0C Y00 (B21)
ν01 ν02 ν01 K1
˛ ˛ ˛
vg (1) (2) 1 (1) (2) 1 (1) (2)
+B2 qY0C Y00 − C 1 YCC Y00 + C 2 Y0C Y0C .
ν02 K2 ν01 ν02
At the neutrality point the above expression simplifies and takes the form
˛ ˛ ˛
00 vg 1 1
I 12 (µ1 = µ2 = 0) = − (B1 − B2 ) qY0C Y00 − C 1 YCC Y00 + C 2 Y0C Y0C . (B22)
ν0 K(0) ν0 ν0
 19

d. Interlayer thermalization

The integrated collision integrals (B18) and (B21) contain formally diverging expressions
˛ ˛ ˛ ˛
2 (1) (2) (i) (j) (i) (j) (1) (2)
q Y00 Y00 , qY0C Y00 , YCC Y00 , Y0C Y0C .

(i) (i) (i)

The divergence stems from the fact that each of the functions Y00 , Y0C , and YCC diverge as |ω| → vg q

(i) 1 (i) (i) 1

Y00 ∝ q , Y0C (|ω| > vg q), YCC (|ω| > vg q) ∝ q .
ω 2 − vg2 q 2  ω 2 − vg2 q 2

The diverging part can be separated with the help of the following relations

(i) |ω| (i) (i)

Y0C (|ω| > vg q) = 2 Y (|ω| > vg q) + Ye0C (|ω| > vg q), (B23)
vg q 00

(i) ω 2 (i) (i)

YCC (|ω| > vg q) = 4 Y (|ω| > vg q) + YeCC (|ω| > vg q), (B24)
vg2 q 2 00
(i) (i)
where the new functions Ye0C and YeCC vanish at |ω| = vg q. Then the collision integral (B18) takes the form
˛ ˛
vg (1) (2) vg (1) (2)
I 012 = −A1 qY0A Y00 + A2 qY00 Y0A (B25)
ν01 ν02
˛ ˛ 
vg (1) (2) (1) (2)
−C 1 qY0C Y00 θ(|ω| < vg q) + q Ye0C Y00 θ(|ω| > vg q)
ν01
˛ ˛ 
vg (1) (2) (1) (2)
+C 2 qY00 Y0C θ(|ω| < vg q) + qY00 Ye0C θ(|ω| > vg q)
ν02
!
vg2 vg2
 
1 1
− B1 − B2 Γ0 − C 1 − C2 Γ2 ,
ν01 K1 ν02 K2 ν01 ν02

where the last line contains the diverging integrals

˛ ˛
(1) (2) (1) (2)
Γ0 = q 2 Y00 Y00 , Γ2 = 2 |ω|Y00 Y00 θ(|ω| > vg q). (B26)

The terms with these diverging rates should be excluded from the hydrodynamic equations, which reduces the
number of independent macroscopic currents. In order to do so, one has to solve the system of equations (49) for the
combinations B1 vg2 /(ν01 K1 ) − B2 vg2 (ν02 K2 ) and C 1 /ν01 − C 2 /ν02 keeping the rates Γi and then take the limit Γi → ∞.
This yields the interlayer thermalization conditions
1 1 1 1
B1 = B2 , C1 = C2 . (B27)
ν01 K1 ν02 K2 ν01 ν02
At the neutrality point these conditions simplify to
B1 (µ1 = 0) = B2 (µ2 = 0), C 1 (µ1 = 0) = C 2 (µ2 = 0). (B28)
Now the number of independent currents and correspondingly the number of macroscopic equations is reduced from
six to four. All terms that do not contain the diverging rates Γi can be straightforwardly simplified using Eqs. (B27).
More care is needed when treating the contributions of the collision integrals (B25) and (B21) where one needs to
find the limiting value of the expressions containing Γi . As a result, we find the thermalized equations (50) and (56).
The latter equations also contain the relaxation rate τss,12 is given by
˛ ˛  
−1 1 2

1 2 |ω|
τss,12 = Y00 YCC − Y0C θ(|ω| < vg q) + Y00 YeCC − Ye0C +4 Y00 Ye0C θ(|ω| > vg q), (B29)
ν0 ν0 vg q
appearing from the non-diverging difference between the last two terms in Eq. (B22).
 20

Appendix C: Relaxation rates due to electron-phonon interaction

1. Electron-phonon collision integral

Consider the standard form of electron-phonon collision integral. In graphene it has the following form1,35–41 :
X
Ie−ph = {f1 [1 − f2 ] W1→2 − f2 [1 − f1 ] W2→1 } , (C1a)
1

where
X
W1→2 = 2π |λv1 v2 |2 Wq [(1 + Nq ) δ(p2 − p1 + q)δ(2 − 1 + ωq ) + Nq δ(p2 − p1 − q)δ(2 − 1 − ωq )] . (C1b)
q

Here Nq is the phonon distribution function, ωq is the phonon dispersion, and Wq is the transition matrix element
squared. For acoustic phonons37
D2 q2
Wq = ,
2ρm ωq
where D is the screened deformation potential and ρm is the mass density of graphene. At the same time, in graphene
inelastic relaxation may occur through a combined scattering process involving both a phonon and an impurity41 .
Other possibilities include two-phonon scattering and phonon-induced intervalley scattering. For these processes the
matrix element is more involved.
We now linearize the collision integral (C1) in the standard fashion1 using Eq. (27) and the similar form of the
non-equilibrium correction to the phonon distribution function
(0)
  ∂Nq
Nq = Nq(0) + δNq , δNq = Nq(0) 1 + Nq(0) χ = −T χ.
ωq
Consider the first term in Eq. (C1b). The same δ-functions appear also in the second term in Eq. (C1) describing the
reverse process. Combining the two, one finds the following combination of distribution functions
 
f1 f2 Nq
f1 [1 − f2 ] (1 + Nq ) − f2 [1 − f1 ] Nq = [1 − f1 ] [1 − f2 ] (1 + Nq ) − .
1 − f1 1 − f 2 1 + Nq
It is straightforward to check that the expression in square brackets vanishes in equilibrium. Linearization yields (the
non-equilibrium correction (5) contains the velocity and thus does not contribute to the relaxation rates)
h i 
(0) (0)
f1 [1 − f2 ] (1 + Nq ) − f2 [1 − f1 ] Nq ≈ f1 1 − f2 1 + Nq(0) [δh1 − δh2 − χq ] . (C2)

The combination of the equilibrium distribution functions in Eq. (C2) can be further simplified as
(0)
(0)
h
(0)
i  ∂Nq  (0) (0)

f1 1 − f2 1 + Nq(0) = −T f1 − f2 .
ωq
Finally, one may write the linearized electron-phonon collision integral as a sum of the electron and phonon parts
[following Eq. (C2)]:
Ie−ph = Ie + Iph , (C3a)
where the electronic part is given by
πX Wq X
Ie = 2 |λv1 v2 |2 [δ(p2 − p1 + q)δ(2 − 1 + ωq ) − δ(p2 − p1 − q)δ(2 − 1 − ωq )]
4 q sinh (ωq /2T ) 1
 
2 − µ 1 − µ
× tanh − tanh [δh1 − δh2 ] . (C3b)
2T 2T
In this paper we consider the phonon system to be at equilibrium and therefore neglect the phonon part of the collision
integral. This means that all back-action effects, such as phonon drag, are neglected. For some physical processes,
most notably, thermoelectric effects, such processes might be important. Then one has to consider the phonon kinetic
equation on equal footing with Eq. (18).
 21

2. Energy relaxation rates

The relaxation rates are obtained by integrating the collision integral (C3). The “energy” continuity equation
is obtained by multiplying the kinetic equation by  and integrating over all states. The corresponding integrated
collision integral has the form
X πX Wq X
2 I e = 2 2 |λv1 v2 |2 [δ(p2 − p1 + q)δ(2 − 1 + ωq ) − δ(p2 − p1 − q)δ(2 − 1 − ωq )]
2
4 q
sinh (ωq /2T ) 1,2
 
2 − µ 1 − µ
× tanh − tanh [δh1 − δh2 ] . (C4)
2T 2T
The difference between the non-equilibrium distribution functions reads
 
1 1 − 2
δh1 − δh2 = b + c [sign(1 ) − sign(2 )] .
eν0 T K
Consequently, we can define two relaxation rates
X b c
2 Ie = − + . (C5)
2
τEb τEc

Specifically at the neutrality point we can use Eqs. (29) and express the integrated collision integral in terms of the
energy and imbalance densities
π2 π2
   
X eδu 1 1 eδρ N2 (0) 1
2 Ie = + − + . (C6)
2
K∆(0) τEb 12 ln2 2 τEc ∆(0) τEc 12 ln2 2 τEb

3. Imbalance relaxation rates

Similarly, we find the imbalance relaxation rates. The corresponding integrated collision integral has the form
 
X πX Wq X
2 2 − µ 1 − µ
sign(2 )Ie = sign(2 )|λv1 v2 | tanh − tanh (C7)
2
4 q sinh2 (ωq /2T ) 1,2 2T 2T

× [δ(p2 − p1 + q)δ(2 − 1 + ωq ) − δ(p2 − p1 − q)δ(2 − 1 − ωq )] [δh1 − δh2 ] .

Clearly, only inter-band scattering processes contribute to this collision integral (unlike the case of the energy relax-
ation, where both inter- and intra-band processes have to be taken into account).
For general doping we define the following relaxation rates
X b c
sign(2 )Ie = − , (C8)
2
τIb τ Ic

where
τIb = τEc . (C9)
At the neutrality point this yields
π2 π2
   
X eδu 1 1 eδρ N2 (0) 1
sign(2 )Ie = − + + + . (C10)
2
K∆(0) τIb 12 ln2 2 τIc ∆(0) τIc 12 ln2 2 τIb

Combining the above electron-phonon collision integrals into the two continuity equations for the energy and
imbalance densities, we find Eqs. (41), where the matrix matrix elements of Tph combine the above relaxation rates.
−1 −1 −1
The rates τEc and τIc are determined by the interband scattering processes in contrast to the rate τEb which contains
contribution of the intraband processes as well. Therefore,
τEb  τEc 6 τIc , (C11)
such that the matrix Tph has two positive eigenvalues as it should.
 22

Appendix D: Continuity equations in double-layer systems

Electron-electron interaction does not contribute to continuity equations in monolayer graphene (9) due to the
conservation laws. In double-layer systems, only the electric charge is conserved leaving the corresponding continuity
equation trivial [cf. Eq. (9a)], while the quasiparticle energy and imbalance are affected by interlayer scattering.

1. Energy relaxation due to electron-electron interaction

The continuity equation for energy is obtained by multiplying the kinetic equation by  and integrating over all
states. Integrating the collision integral that describes interlayer electron-electron interaction we find [cf. Eq. (B16)]
ˆ 2
0 1 d qdω |U12 (q, ω)|2
I12 = − (D1)
32 (2π)3 sinh2 (ω/2T )
 
X
3 2 0 0 1 − µ2 1 + ω − µ2
× (2π) |λvv0 | δ(1 − 1 + ω)δ(p1 − p1 + q) tanh − tanh
0
2T 2T
1,1
 
X 2 − µ1 2 − ω − µ1
× (2 − 02 )(2π)3 |λvv0 |2 δ(2 − 02 − ω)δ(p2 − p02 − q) tanh − tanh
0
2T 2T
2,2

× [δh2,10 + δh1,20 − δh2,1 − δh1,2 ] .

Using the explicit form of the distribution function (28) and energy conservation we find
 0   
1 1 − 1 0 1 ω
δh2,1 − δh2,1 =
0 b2 + c2 (sign(1 ) − sign)1 )) = b2 + c2 sign(ω) ,
eν02 T K2 eν02 T K2

and similarly for the first layer. As a result

 ˛  ˛
0 b1 b2 (1) (2) c1 c2 (1) (2)
eI12 =− − ω 2 Y00 Y00 − − |ω|Y00 Y00 . (D2)
ν01 K1 ν02 K2 ν01 ν02

2. Energy relaxation due to electron-electron interaction

Similarly to the previous Section, we find the contribution of electron-electron interaction to the continuity equation
for quasiparticle imbalance [cf. Eq. (B20)]
ˆ 2
00 1 d qdω |U12 (q, ω)|2
I12 =− (D3)
32 (2π)3 sinh2 (ω/2T )
 
X 1 − µ2 1 + ω − µ2
× (2π)3 |λvv0 |2 δ(1 − 01 + ω)δ(p1 − p01 + q) tanh − tanh
0
2T 2T
1,1
 
X 2 − µ1 2 − ω − µ1
× [sign(2 ) − sign(02 )] (2π)3 |λvv0 |2 δ(2 − 02 − ω)δ(p2 − p02 − q) tanh − tanh
2T 2T
2,20

× [δh2,10 + δh1,20 − δh2,1 − δh1,2 ] .

Using the explicit form of the distribution function (28) and energy conservation we find
 ˛
00 b1 b2 (1) (2)
eI12 = − − |ω|Y00 Y00 (D4)
ν01 K1 ν02 K2
 ˛
c1 c2 (1) (2)
− − Y00 Y00 .
ν01 ν02
 23

3. Thermalization in finite-size samples mally diverging terms (D5). Since the macroscopic equa-
tions contain gradient terms, the resulting hydrodynamic
The collision integrals (D2) and (D4) contain formally equations will now contain gradients of the driving cur-
diverging expressions [similar to Eqs. (B26)]: rent j1 (y).
˛ ˛ ˛
(1) (2) (1) (2) (1) (2) On the other hand, at the phenomenological level one
ω 2 Y00 Y00 , |ω|Y00 Y00 , Y00 Y00 . (D5) may assume the interlayer interaction to be weaker than
the intralayer interaction. In that case, the latter is re-
If one assumes equal strength of intra- and inter- sponsible for forming the hydrodynamic modes, while the
layer Coulomb interaction, then one needs to perform former [where the terms (D5) are treated as finite] play
the interlayer thermalization procedure, described in Ap- the role of additional relaxation rates. This way one ob-
pendix B 2 d. In finite-size systems, this procedure has tains the phenomenological model of Ref. 15, which qual-
to include the continuity equations containing the for- itatively captures the essential physics of the system.

1 23
E.M. Lifshitz and L.P. Pitaevskii, Physical Kinetics, (Perg- J.B. Oostinga, B. Sacepe, M.F. Craciun, and A.F. Mor-
amon Press, 1981). purgo, Phys. Rev. B 81, 193408 (2010).
2 24
S. Chapman, Phil. Trans. R. Soc. Lond. A 216, 279 (1916); P.S. Alekseev, A.P. Dmitriev, I.V. Gornyi, V.Yu. Ka-
217, 115 (1918); D. Enskog, Arkiv Mat. Astr. Fys. 16, 60 chorovsky, B.N. Narozhny, M. Schütt, and M. Titov,
(1921). arXiv:1410.4982 (to be published).
3 25
H. Grad, Commun. Pure Appl. Math. 2, 331 (1949). J.C.W. Song and L.S. Levitov, Phys. Rev. Lett. 111,
4
A.B. Kashuba, Phys. Rev. B 78, 085415 (2008). 126601 (2013); J.C.W. Song, D.A. Abanin, and L.S. Levi-
5
M. Müller and S. Sachdev, Phys. Rev. B 78, 115419 (2008). tov, Nanolett. 13, 3631 (2013).
6 26
M. Müller, L. Fritz, and S. Sachdev, ibid, 115406 (2008). B.N. Narozhny, M. Titov, I.V. Gornyi, and P.M. Ostro-
7
L. Fritz, J. Schmalian, M. Müller, and S. Sachdev, Phys. vsky, Phys. Rev. B 85, 195421 (2012).
27
Rev. B 78, 085416 (2008); M. Müller, J. Schmalian, and W.K. Tse, B.Yu-K. Hu, and S. Das Sarma, Phys. Rev. B
L. Fritz, Phys. Rev. Lett. 103, 025301 (2009). 76, 081401 (2007).
8 28
M.S. Foster and I.L. Aleiner, Phys. Rev. B 79, 085415 B.L. Altshuler and A.G. Aronov, in Electron-Electron In-
(2009). teractions in Disordered Systems, eds. A.L. Efros, M. Pol-
9
D. Svintsov, V. Vyurkov, S. Yurchenko, T. Otsuji, and V. lak (North-Holland, Amsterdam, 1985).
29
Ryzhii, J. Appl. Phys. 111, 083715 (2012). G. Zala, B.N. Narozhny, and I.L. Aleiner, Phys. Rev. B
10
M. Schütt, P.M. Ostrovsky, M. Titov, I.V. Gornyi, B.N. 64, 214204 (2001).
30
Narozhny, and A.D. Mirlin, Phys. Rev. Lett. 110, 026601 V.V. Cheianov and V.I. Fal’ko, Phys. Rev. Lett. 97, 226801
(2013). (2006).
11 31
M. Schütt, P.M. Ostrovsky, I.V. Gornyi, and A.D. Mirlin, Note, that in disorder-dominated regime τ  τee energy
Phys. Rev. B 83, 155441 (2011). dependence of the impurity scattering time is important
12
R.V. Gorbachev, A.K. Geim, M.I. Katsnelson, K.S. Novo- and, in particular, close to the neutrality point it deter-
selov, T. Tudorovskiy, I.V. Grigorieva, A.H. MacDonald, mines the dependence of transport coefficients on the mag-
K. Watanabe, T. Taniguchi, L.A. Ponomarenko, Nature netic field, see P.S. Alekseev, A.P. Dmitriev, I.V. Gornyi,
Phys. 8, 896 (2012). and V.Yu. Kachorovskii, Phys. Rev. B 87, 165432 (2013).
13 32
S. Kim, I. Jo, J. Nah, Z. Yao, S.K. Banerjee, and E. Tutuc, U. Briskot, M. Schütt, I.V. Gornyi, B.N. Narozhny, and
Phys. Rev. B 83, 161401(R) (2011). A.D. Mirlin, to be published.
14 33
S. Kim and E. Tutuc, Solid State Comm. 152, 1283 (2012). V. Guttal and D. Stroud, Phys. Rev. B 71, 201304(R)
15
M. Titov, R.V. Gorbachev, B.N. Narozhny, T. Tudo- (2005).
34
rovskiy, M. Schütt, P.M. Ostrovsky, I.V. Gornyi, A.D. Mir- R. Magier and D.J. Bergman, Phys. Rev. B 74, 094423
lin, M.I. Katsnelson, K.S. Novoselov, A.K. Geim, and L.A. (2006).
35
Ponomarenko, Phys. Rev. Lett. 111, 166601 (2013). E.H. Hwang and S. Das Sarma, Phys. Rev. B 77, 115449
16
K. Seeger, Semiconductor Physics, (Springer, 2002) (2008).
17 36
J. Ping, I. Yudhistira, N. Ramakrishnan, S. Cho, S. Adam, S. Fratini and F. Guinea, Phys. Rev. B 77, 195415 (2008).
37
and M.S. Fuhrer, Phys. Rev. Lett. 113, 047206 (2014). R. Bistritzer and A.H. MacDonald, Phys. Rev. Lett. 102,
18
S. Cho and M.S. Fuhrer, Phys. Rev. B 77, 081402(R) 206410 (2009).
38
(2008). S.S. Kubakaddi, Phys. Rev. B 79, 075417 (2009).
19 39
J. Jobst, Ph.D. Thesis, Friedrich-Alexander-Universität W.K. Tse and S. Das Sarma, Phys. Rev. B 79, 235406
Erlangen-Nürnberg (2012). (2009).
20 40
J. Bai, R. Cheng, F. Xiu, L. Liao, M. Wang, A. Shailos, J.K. Viljas and T.T. Heikkilä, Phys. Rev. B 81, 245404
K.L. Wang, Y. Huang, and X. Duan, Nature Nanotech.5, (2010).
41
655 (2010). J.C.W. Song, M.Y. Reizer, and L.S. Levitov, Phys. Rev.
21
K. Gopinadhan, Y.J. Shin, I. Yudhistira, J. Niu, and H. Lett. 109, 106602 (2012).
42
Yang, Phys. Rev. B 88, 195429 (2013). M. Schütt, Ph.D. Thesis, KIT (2013).
22
Y. Zhao, P. Cadden-Zimansky, F. Ghahari, and P. Kim,
Phys. Rev. Lett. 108, 106804 (2012).