Optical Kinetic Theory of Nonlinear Multi-mode Photonic Networks

Arkady Kurnosov,1 Lucas J. Fernández-Alcázar,2, 3 Alba Ramos,2, 3 Boris Shapiro,4 and Tsampikos Kottos1
1
Wave Transport in Complex Systems Lab, Department of Physics,
Wesleyan University, Middletown, CT-06459, USA
2
Institute for Modeling and Innovative Technology,
IMIT (CONICET - UNNE), Corrientes W3404AAS, Argentina
3
Physics Department, Natural and Exact Science Faculty,
Northeastern University of Argentina, Corrientes W3404AAS, Argentina
4
Technion – Israel Institute of Technology, Technion City, Haifa, 3200, Israel
(Dated: )
Recent experimental developments in multimode nonlinear photonic circuits (MMNPC), have
arXiv:2310.16784v1 [physics.optics] 25 Oct 2023

motivated the development of an optical thermodynamic theory that describes the equilibrium
properties of an initial beam excitation. However, a non- equilibrium transport theory for these
systems, when they are in contact with thermal reservoirs, is still terra incognita. Here, by combining
Landauer and kinematics formalisms we develop an one-parameter scaling theory that describes
the transport in one-dimensional MMNPCs from a ballistic to a diffusive regime. We also derive a
photonic version of the Wiedemann -Franz law that connects the thermal and power conductivities.
Our work paves the way towards a fundamental understanding of the transport properties of MMNPC
and may be useful for the design of all-optical cooling protocols.

Introduction –Quite recently, we have witnessed a surge
in understanding and harnessing the convoluted behavior
of light propagation in multimode nonlinear photonic cir-
cuits (MMNPC). These experimental platforms provide an
excellent testbed not only to investigate but also unravel
new horizons and possibilities with both fundamental and
technological ramifications. For example, they have been
used for exploring exotic optical phase transitions [1–4]
and beam self- cleaning phenomena [5–7], spatiotemporal
mode locking [8], multimode solitons [9], etc. In paral-
lel, their implementation in fiber-optical communications
might resolve urgent technological needs associated with
the looming information “capacity crunch” [10, 11] or the
quest for new platforms of high-power light sources [8].
Nonlinearities lead to multi-wave mixing processes
or photon-photon “collisions” through which the many
modes can exchange energy via a multitude of possible
pathways, often numbering in the trillions even in the pres-
ence of one hundred modes or so. Evidently, modeling, pre-
dicting, and harnessing the response of such exceedingly
complex configurations is practically impossible using con-
ventional brute-force time-consuming computations that
obscure the underlying physical laws. Fortunately, an en-
tirely new universal approach inspired by concepts from
statistical thermodynamics emerged recently [12], that
self-consistently describes the utterly complex processes
of energy/power exchange in MMNPC. Under thermal
equilibrium conditions, the methodology has identified
intrinsic variables (optical temperature T and chemical
potential µ) that play the role of optical thermodynamic
forces leading an initial beam excitation to a Rayleigh
Jeans (RJ) thermal state [4, 12, 13] – a key tenet of this
theory that has been confirmed using multimode optical
fibers and time-synthetic photonic lattices [14–16].
Here, we develop a transport theory that describes the
nonequilibrium steady states generated in a MMNPC (a
photonic junction) when is in contact with two optical
reservoirs at different optical temperature and/or chemi-
cal potentials. The proposed optical kinetics framework,
allows us to define and calculate various kinetic coeffi-
cients like optical power and thermal conductivities in full
analogy with physical kinetics in condensed matter [17].
The presence of two conserved quantities (total optical
power and electrodynamic momentum flow, referred be-
low as internal energy) requires to consider the coupling
between thermal and power currents mediated by nonlin-
ear interactions – a complication that is not present in
phonon heat transport in solids. Using a combination of
Landauer (ballistic limit) and Boltzmann (diffusive limit)
transport theories, we established a one-parameter scaling
theory that describes the crossover from a ballistic to a
diffusive limit as the size of the photonic junction increases
beyond a characteristic length-scale lT that incorporates
information about the thermalization processes occurring
at such scales in the junction. Finally, we analyze the
interdependence of optical power and heat transport by
deriving the photonic analogue of Wiedemann-Franz (WF)
law that connects the thermal and power conductances.
Their ratio is inversely proportional to temperature – as
opposed to the linear temperature enhancement occurring
in typical thermoelectric devices – which is a signature of
relaxation scales separation between the energy and power
currents. Our results set the basis for the development of
novel thermo-photonic devices and paves the way for the
design of novel photonic refrigerators or engines.
Physical setting – We consider the set-up shown in
Fig. 1. It consists of two optical reservoirs and a junc-
tion which facilitates thermal and power transport be-
tween the reservoirs. The left (L) and right (R) reservoirs
consist of arrays of (weakly) nonlinear coupled optical
 2

vant) large paraxial length-scales. Rather, we focus our

investigation at a physically relevant intermediate (but
still large) length scales ztr < z ≪ zGT , where (after a
transient paraxial length ztr ) the currents through the
junction acquire (quasi-)steady-state values. Our goal is
to develop a non-equilibrium transport theory that de-
scribes such power and heat transfer at these intermediate
length scales.
Mathematical modeling – The beam dynamics at the
junction and the left/right reservoirs is described by a
temporal coupled mode theory (TCMT),

dΨl X
i =− Jlj Ψj + χ|Ψl |2 Ψl , (1)
dz j
FIG. 1. The two photonic reservoirs (L/R), consist of a large
number of coupled nonlinear waveguides. We initially inject
two beams at each of the reservoirs with different internal where Ψl ≡ ⟨l|Ψ⟩ is the field amplitude at the l-th waveg-
energy and power. At the steady state, the power distribution uide, ϵl = −Jll is its propagation constant and χ is the
is given by a RJ with predetermined temperatures TL/R and Kerr nonlinearity coefficient. The connectivity of the
chemical potentials µL/R . After reaching a thermal state, ∗
network is dictated by the coupling coefficients Jlj = Jjl .
the two reservoirs are coupled by an optical junction which At the junction Jlj = Jδl,l±1 (ϵl = ϵ = 0). The corre-
transports heat and power currents. For intermediate, but (J)
large, length scales these currents acquire a quasi-steady-state sponding linear dispersion relation takes the form ωα =
jh , j a . −2J cos(k) where the wavevector k ∈ [−π, π]. The two
reservoirs consist of a square lattice of Nr = 40×40 = 1600
waveguides with Jlj = J. To avoid spectral degeneracies,
we consider propagation constants given by a uniform
waveguides (or multimode/multicore optical fibers) sup-
distribution ϵl ∈ [− W W
2 , 2 ] with W = 0.5. The junction-
porting a finite (but large) number γ = 1, · · · , M of
(L/R)
reservoir coupling is Jb−r = 0.2J. We have confirmed via
linear supermodes ϕγ ⟩ – all propagating along the direct dynamical simulations of the composite system, the
(L/R) existence of a (quasi-)steady-state regime during which
axial direction z with propagation constants ωγ . At
each reservoir, we launch a beam prepared at some state the temperature and chemical potential at each reservoir
P (L/R) (L/R) (L/R) remain (approximately) constant when M ≫ N .
Ψ(L/R) (z = 0)⟩ = γ Cγ ϕγ ⟩ where Cγ =
Onsager Matrix Formalism – We consider the power
⟨ϕγ |Ψ(L/R) ⟩ are the projection coefficients to their su- a(x) and energy h(x) densities at any segment (x, x + dx)
permodes. Initially, the reservoirs are decoupled from inside the junction which includes many unit cells. At
one another and, therefore, their total optical power the (quasi-)stationary regime, these segments are at local
(L/R) (L/R) 2
N L/R ({Cγ | ≡ A(L/R) and inter-
P
}) = γ |Cγ equilibrium characterized by a local temperature and
(L/R) (L/R) 2 chemical potential, i.e., T = T (x), µ = µ(x) which slowly
nal energy HL/R ({Cγ | ωγ ≡ E L/R ,
P
}) ≈ γ |Cγ
are constants of motion which are used to determine the change with the position. Under this assumption, the
optical temperature T , and chemical potential µ that de- currents are evaluated by expanding the spatial gradients
fine their RJ thermal state nL/R (ωγ
(L/R) (L/R) 2
) = ⟨|Cγ | ⟩≡ of µ, T up to the first term [20, 21]:
TL/R
[4, 12, 18, 19].
 
ωγ −µL/R Laa Laq
j = L̂f , L̂ = . (2)
Once each reservoir reaches a thermal equilibrium with Lqa Lqq
(TL , µL ) ̸= (TR , µR ), they are coupled with an optical
junction. It consists of an array of coupled (weakly) non- Above, j = (ja , jq )T where ja , jh , jq = jh − µja are the
linear single-mode waveguides supporting α = 1, · · · , N power, energy, and heat currents, L̂ is the the Onsager
(J)
linear supermodes ϕα ⟩ that propagate along the parax- matrix, while the affinities f = (−∇µ/T, ∇(1/T ))T are
(J)
the thermodynamical forces that induce the currents. For
ial distance z with propagation constants ωα . systems preserving time reversal symmetry, the Onsager
The junction facilitates heat and power transport be- reciprocity relations hold, i.e., Lqa = Laq [20, 22].
tween the reservoirs. Obviously, the characteristics of the We distinguish between two limiting cases of short
junction, i.e., coupling between waveguides, nonlinearity N ≪ lT (ballistic) and long N ≫ lT (diffusive) junctions
strength, size, etc., will determine these currents, and, in where lT = vg zT , vg is a typical group velocity of the lin-
turn, the paraxial length zGT. at which the whole system ear supermodes, while zT = zT (χ, J, T, a) is a relaxation
reaches a global thermalization. We are not interested distance that dictates the thermalization process of a
in the behavior of the system at these (practically irrele- non-equilibrium state in the isolated junction towards its

RJ distribution [19, 23]. Strictly speaking the concept of
local equilibrium used in Eqs. (2) applies to the diffusive
regime only while in the ballistic regime, the meaningful
quantities are the temperatures TL,R and chemical po-
tentials µL,R of the two reservoirs. In this case we can
formally define ∇T ≡ (TR − TL )/N ; ∇µ ≡ (µR − µL )/N
to have unified notations for both regimes.
The various transport coefficients, can be extracted
from the elements of the Onsager matrix Eq. (2). For ex-
ample, the power conductivity is σ ≡ La,a /T , the thermal
conductivity is κ = det L̂/(T 2 Laa ), while the Seebeck and
Peltier coefficient that describe thermo-power transport
are S = Laq /(T Laa ) and Π = T · S (see Supplement for
details). Below, we analytically evaluate these matrix
elements.
(a) Ballistic regime – In this regime, the nonlinear inter-
actions are not able to enforce mixing among the linear
modes. The power and energy fluxes through the junction
are evaluated using Landauer’s theory [24]
Z the rhs of Eq. (5) becomes St δn ≈ −δn/zT (k) (“time”-
ja(h) = dω t(ω)ω s [nL (ω) − nR (ω)] , (3)
where s = 0 (s = 1) for power (energy) current. We
further assume that the transmittance t(ω) = t0 = const.
for all supermodes in the band [−2J, 2J] and zero oth-
erwise. Equation (3) can be evaluated analytically, thus
allowing us to extract the Onsager matrix elements (see
Supplementary Material)
4JT 2
Laa = t0 N 2 ; Lqq = t0 N 4JT 2
µ − 4J 2
  (4)
2 4J
Laq = Lqa = −t0 N T ln 1 − ,
−µ + 2J
where T = (TL + TR )/2, µ = (µL + µR )/2 and |∆µ| ≪
|µ|−2J and ∆T ≪ T (linear response regime). Equations
(4) are the main results of this section. They allow us
to derive exact expressions for the transport coefficients
σ, κ ∝ N . We, furthermore, conclude that when lT ≫ N ,
the Fourier’s law does not hold.
While the weak nonlinear interactions cannot enforce
sufficient mode-mode mixing, they can induce a nonlinear
(J) (J)
frequency shift ωα → ωα + 2aχ (see Supplement, and
Ref. [25]) which might affect the value of the currents.
Nevertheless, Eq. (3) still applies with the modification
that the transmittance acquires its constant value t0 inside
a shifted frequency window ω (J) ∈ [−2J + 2χa, 2J + 2χa]
while it is zero everywhere else. As a result, Eq. (4) still
applies with the substitution µL,R → µL,R − 2aχ. This
(χ)
nonlinear frequency correction ja(h) → ja(h) is insignifi-
cant in the high temperature limit, T ∼ |µ| ≫ 2aχ but it
becomes important when |µ| ∼ 2J.
(b) Diffusive transport – In the other limiting case of
N ≫ lT , the nonlinear mode-mode interactions become a
dominating mechanism of power and heat transport. They
3
are responsible for a local equilibrium within lT segments
of the junction, thus allowing us to define slowly varying
local temperatures T (x) and chemical potentials µ(x). At
the same time, the modal occupations also become a local
quantity, i.e., a function of coordinate x, wave vector k,
and propagation distance z, n = n(x, k, z). We proceed
by invoking a Kinetic Equation (KE) approach [25]
dn ∂n
= + (vg · ∇n) = St n, (5)
dz ∂z
where n(k, x)dkdx represents the power (number of parti-
cles) contained in a macroscopic volume element dkdx of
the phase space and St n is a collision integral. Next, we
consider the stationary regime ∂n/∂z = 0, and assume
that n depends on the position x in the junction via the lo-
cal optical temperature T (x) and chemical potential µ(x).
Further, we linearize Eq. (5), assuming small deviations
from the local equilibrium, n = n(0) + δn, where n(0) is
the (local) equilibrium RJ-distribution. Since Stn(0) = 0,
relaxation approximation). Then, the solution of the
linearized KE reads
n(0) h i
δn(k) ≈ −zT (k) (vg · ∇µ)n(0) + (vg · ∇T ) . (6)
T
resulting to power and heat currents
Z
dk h is
ja(q) = vg (k) ω (J) (k) − µ δn(k). (7)
(2π)
Evaluation of the above integrals [See Supplement for
details] together with Eq. (2) allows us to express the
Onsager matrix elements as
" #
2 µ
Laa = −T 1 + p zT , Lqq = 2J 2 T 2 zT
2
µ − 4J 2
h i
p
2
Laq = Lqa = −T µ + µ2 − 4J 2 zT , (8)
where we omitted the k-dependence of the relaxation
distance zT . Finally, within the linear response theory
(∇T ∼ ∆T /N ≪ T̄ ≡ (TL + TR )/2, ∇µ ∼ ∆µ/N ≪ µ̄ ≡
(µL + µR )/2), we neglect the x-dependence of tempera-
ture T (x) and chemical potential µ(x) and approximate
the Onsager elements as Lij (T (x), µ(x)) ≈ Lij (T̄ , µ̄). We
find that, contrary to the ballistic domain, ja , jq ∝ 1/N
featuring the so-called normal transport, where the trans-
port coefficients κ, σ are independent on the system size
and Fourier’s law holds.
One-parameter scaling theory – Next we have estab-
lished a one-parameter scaling theory that controls the
variation of ja as the size of the junction increases.
Specifically, we find that the rescaled power current
pN (χ, J, T̄ , µ̄) through a junction of length N which is
attached to two reservoirs with mean temperature and
 4
8

9 Fig. 2, where we report the outcome of our simulations

7

0
8

using two methods: (1) Modeling the large collections

10 7
6

6
5
of M modes in the bundles by Monte-Carlo reservoirs
5 [26] with effective thermostats at fixed (T, µ) (see filled
ja / ja���

102 N 103
4

symbols in Fig. 2). (2) Solving numerically the TCMT

ja
3

10-1
3

2
2
Eq. (1) for the whole system (reservoirs + junction) with
1

10 -2
1 reservoirs consisting of M = 1600 coupled modes forming
0
0.8
0
0.8
1

1
1.2

1.2
1.4

1.4
1.6
a square lattice (see Supplementary Material). A possible
10
-2 interpolating law that describes our data (including the
~N-1 ~N-1 crossover regime) is
10-4

10-4 10-2 100 102 1

� f (λ) =
1 + λ/λ∗
, (11)

FIG. 2. Normalized power current ja /ja versus the scaling

(χ) where λ∗ = 0.18 is the best fitting parameter.
parameter λ = N/lT for a 1D junction, of (transverse) size Photonic Wiedemann-Franz law and thermal current
N . The different values of the current ja for a variety of – In thermoelectric devices, the Wiedemann-Franz (WF)
parameters (χ, J, T̄ , µ̄) are shown in the inset. The black line law connects the thermal and current conductivity in a
is the interpolating function Eq. (11), while the dashed line simple manner: their ratio is proportional to the tem-
indicates a 1/N behavior and is drawn to guide the eye. perature i.e. κ/σ = LT with a proportionality constant
L which, in typical metals, takes a universal value [27].
This proportionality relation, essentially describes the fact
chemical potential T̄ , µ̄ obeys a one-parameter scaling,
that a good electrical conductor, is also an efficient heat
i.e.,
conductor, with thermal to electric conductivity ratio
∂pN proportional to temperature – a property that is rooted
= β(pN ), (9) on the fact that heat and charge currents are associated
∂lnN
to the flux of the same (quasi)particles. Deviations from
where β is a function of pN alone. The rescaled power
(χ) the WF law signify the existence of multiple transport
current is pN (χ, J, T, µ) ≡ j(χ)
a
where ja is the power mechanisms for thermal and/or electrical fluxes, that ulti-
ja
current in the ballistic regime. mately allow for the independent control of electrical and
The scaling ansatz of Eq. (9) is equivalent to postulating thermal transport [22, 28].
the existence of a function f (x) such that The equivalent of charge (particle) conductivity in pho-
 tonics, is power conductivity. It is natural, therefore,
1; λ ≪ O(1) to extend the above definition of WF law and analyze
pN = f (λ ≡ N/lT ) ∝ 1 . (10)
N ; λ ≫ O(1) the corresponding ratio of thermal conductivity to power
The scaling parameter λ ≡ N/lT , encodes all the in- conductivity. By combining Eqs. (4,8) we obtain
formation about the relaxation process towards a non- κ J2
equilibrium (quasi-)steady state current and describes the ≈ , for |µ| ≫ (2J), (12)
σ T
number of thermalized segments with length lT ∼ vg zT
contained in a junction of length N . In the ballistic limit signifying a novel form of WF law which occurs in MM-
lT ≫ N , the junction consists of a single segment and the NPC. This theoretical prediction is nicely confirmed by
nonlinear interactions are unable to enforce thermaliza- our simulations using Monte-Carlo optical reservoirs (see
tion of the modes. Therefore, the transport is essentially main panel of Fig. 3). The various values of the nonlinear
ballistic and j(χ)
a
≈ 1. On the other hand, when N ≫ lT , coefficient χ and junction size N that have been used
ja in these simulations were chosen to guarantee that we
the network consists of a number of λ ≫ 1 uncorrelated
have spanned the full transport domain from ballistic to
segments. This situation is reminiscent of the law of ad-
diffusive regimes.
ditive resistances connected in series. As in this case, the
The inverse temperature dependence of the ratio κ/σ
total current decays inversely proportional to the number
signifies the decoupling of thermal and power transfer. A
of segments ja ∝ 1/N . The h relaxation distance that de-
χ2 a 2
i similar phenomenon has been also observed in ultracold
T
fines lT is zT ∝ J tanh ζJa and has been previously atomic gases [29] implying that atom-atom interactions
evaluated in Ref. 23 (a ∼ 1 is the average value of norm affect the associated thermal and particle conductivities
per site, and ζ ≈ 8 is a best fitting parameter). in radically different ways. Specifically, it was shown that
To validate our scaling ansatz Eq. (9), we have per- there is a time-scale separation for the equilibration of
formed detailed numerical simulations for a variety of temperature and particle imbalances between the two
nonlinear coefficients χ and system sizes N for both high reservoirs. We have demonstrated the different equili-
and low temperatures T . Our results are summarized in bration times by performing dynamical simulations with

FIG. 3. The ratio of thermal to power conductivity κ/σ versus
temperature (black dots) for different values of nonlinearities
χ = 0, 0.025, 0.05, 0.1, 0.25, 0.5, 0.75, 1.0, and junction
sizes: N = 150, 300, 600. The theoretical prediction Eq. (12)
(red dashed line) indicates a ∝ 1/T behavior, signifying a
separation of relaxation processes between thermal and power
transport. In the insets we show the difference of power (upper
inset) and energy (lower inset) densities between the left and
right reservoirs for three different reservoir sizes M . As M
increases the two reservoirs maintain for longer time their
initial energy and power densities allowing us to extract the
(quasi)-steady-state values for the thermal and power currents.
small composite (junction + reservoirs) system sizes. In
these simulations we have utilized a microcanonical ap-
proach for the whole system and found different relaxation
scales for the internal energy and power differences be-
tween the two reservoirs (see insets of Fig. 3).
Let us finally point out that unlike the familiar case of
metals, in the developed optical kinetics framework, the
results for the WF law are sensitive to the definition of κ.
If, for example, we had defined the thermal conductivity
under the constraint ∇µ = 0 (as opposed to the traditional
ja = 0 used above), we will end up with a different
result for κ/σ (see Eqs. (S12,S19) of the Supplementary
material). In this case, at the high temperature limit
where n(0) ≈ 1, we get the familiar expression κ/σ = T .
Conclusion – We have developed an optical kinetics
framework that can be utilized to predict the complex re-
sponse of nonlinear heavily multimoded optical junctions
when they are brought in contact with optical reservoirs.
The transport coefficients derived here can be utilized
for the design of all-optical cooling protocols. It will be
interesting to extend our formalism to include localization
effects due to the presence of transverse disorder or the
influence of paraxial noise in the steady-state currents.
We acknowledge partial support from Simons Founda-
tion for Collaboration in MPS grant No. 733698
5
[1] G. Situ, J. W. Fleischer, Dynamics of the Berezinskii-
Kosterlitz-Thoouless transition in a photon fluid, Nat.
Photonics 14, 517 (2020)
[2] J. Klaers, J. Schmitt, F. Vewinger, M. Weitz, Bose-
Einstein condensation of photons in an optical microcavity,
Nature 468, 545 (2010)
[3] C. Sun et al., Observation of the kinetic condensation of
classical waves, Nat. Phys. 8, 471 (2012).
[4] A. Ramos, L. Fernandez-Alcazar, Tsampikos Kottos, B.
Shapiro, Optical Phase Transitions in Photonic Networks:
a Spin-System Formulation, Phys. Rev. X 10, 031024
(2020)
[5] K. Krupa, A. Tonello, B. Shalaby, M. Fabert, A.
Barthélémy, G. Millot, S. Wabnitz and V. Couderc, Spa-
tial beam self-cleaning in multimode fibres, Nat. Photonics
11, 237 (2017).
[6] Z. Liu, L. Wright, D. Christodoulides, and F. Wise, Kerr
self-cleaning of femtosecond-pulsed beams in graded-index
multimode fiber, Opt. Lett. 41,3675 (2016).
[7] A. Niang, T. Mansuryan, K. Krupa, A. Tonello, M. Fabert,
P. Leproux, D. Modotto, O. N. Egorova, A. E. Levchenko,
D. S. Lipatov, S. L. Semjonov, G. Millot, V. Couderc, and
S. Wabnitz, Spatial beam self-cleaning and supercontinuum
generation with Yb-doped multimode graded-index fiber
taper based on accelerating self-imaging and dissipative
landscape, Opt. Express 27, 24018 (2019).
[8] L. G. Wright, D. N. Christodoulides, F. W. Wise, Spa-
tiotemporal mode-locking in multimode fiber lasers, Science
358, 94 (2017).
[9] L. G. Wright, D. N. Christodoulides, F. W. Wise, Control-
lable spatiotemporal nonlinear effects in multimode fibres,
Nat. Photonics 9, 306 (2015).
[10] K-P Ho, J. M. Kahn, Mode Coupling and its Impact on
Spatially Multiplexed Systems, Optical Fiber Telecommu-
nications VIB, Elsevier (2013).
[11] D. Richardson, J. Fini, L. Nelson, Space-division multi-
plexing in optical fibres, Nat. Photonics 7, 354 (2013)
[12] F. O. Wu, A. U. Hassan, and D. N. Christodoulides,
Thermodynamic theory of highly multimoded nonlinear
optical systems, Nature Photonics 13, 776 (2019).
[13] K. G. Makris, F. O. Wu, P. S. Jung, D. N. Christodoulides,
Statistical mechanics of Weakly nonlinear optical multi-
mode gases, Opt. Lett 45, 1651 (2020)
[14] H. Pourbeyram, P. Sidorenko, F. O. Wu, N. Bender, L.
Wright, D. N. Christodoulides, and F. Wise, Direct obser-
vations of thermalization to a Rayleigh–Jeans distribution
in multimode optical fibres, Nature Phys. 18, 685 (2022).
[15] A. L. Marques-Muniz, F. O. Wu, P. S. Jung, M. Kha-
javikhan, D. N. Christodoulides, U. Peschel, Observation
of photon-photon thermodynamic processes under negative
optical temperature conditions. Science 379,1019 (2023).
[16] K. Baudin, J. Garnier, A. Fusaro, N. Berti, C. Michel, K.
Krupa, G. Millot, and A. Picozzi, Observation of Light
Thermalization to Negative-Temperature Rayleigh-Jeans
Equilibrium States in Multimode Optical Fibers, Phys.
Rev. Lett. 130, 063801 (2023).
[17] L. D. Landau, E. M. Lifshitz, Physical Kinetics, Pergamon
Pres, New York, USA (1981)
[18] M. Parto, F. O. Wu, P. S. Jung, K. Makris, and D. N.
Christodoulides, Thermodynamic conditions governing the
optical temperature and chemical potential in nonlinear
 6

highly multimoded photonic systems, Optics Letters 44, .

3936 (2019).
[19] C. Shi, T. Kottos, B. Shapiro, Controlling optical beam
thermalization via band-gap engineering, Phys. Rev. Re-
search 3, 033219 (2021)
[20] H. B. Callen, Thermodynamics and an Introduction to
Thermostatistics, John Wiley & Sons (1985).
[21] K. Saito, G. Benenti, and G. Casati, A microscopic mech-
anism for increasing thermoelectric efficiency, Chemical
Physics 375, 508, stochastic processes in Physics and
Chemistry (in honor of Peter Hänggi) (2010).
[22] G. Benenti, S. Lepri, R. Livi, Anomalous Heat Transport
in Classical Many-Body Systems: Overview and Perspec-
tives, Frontiers in Physics 8, 00292 (2020)
[23] A. Y. Ramos, C. Shi, L. J. Fernández-Alcázar, D.
N. Christodoulides, T. Kottos, Theory of localization-
hindered thermalization in nonlinear multimode photonics,
Communications Physics 6, 189 (2023).
[24] Y. Imry and R. Landauer, Conductance viewed as trans-
mission, Rev. Mod. Phys. 71, S306 (1999).
[25] D. M. Basko, Kinetic theory of nonlinear diffusion in
a weakly disordered nonlinear Schrödinger chain in the
regime of homogeneous chaos, Phys. Rev. E 89, 022921
(2014).
[26] S. Iubini, S. Lepri, and A. Politi, Nonequilibrium discrete
nonlinear Schrödinger equation, Physical Review E 86,
011108 (2012).
[27] N. W. Ashcroft, N. D. Mermin, Solid State Physics 21
(Saunders College, Philadelphia, 1976).
[28] G. T. Craven, A. Nitzan, Wiedemann-Franz Law for
Molecular Hopping Transport, Nano Lett. 20, 989-993
(2020)
[29] M. Filippone, F. Hekking, A. Minguzzi, Violation of
the Wiedemann-Franz law for one-dimensional ultracold
atomic gases, Phys. Rev. A 93, 011602(R) (2016)
 1

Supplementary Material
COUPLED MODE THEORY MODELING

Equations of motion.-

The beam dynamics can be described by a coupled mode theory (CMT)[S1]

dψl X
i =− Jlj ψj + χ|ψl |2 ψl , (S1)
dz j

where the propagation through the z-direction plays the role of time in a DNLSE. Here, ψl is the complex field
amplitude at the l-th node of the network, ϵl = −Jll represent the optical on-site potentials (e.g. propagation constants)
in the absence of coupling or nonlinearities, and the last term dictates the non-linearity due, for instance, to a Kerr
∗
effect. The coupling coefficients Jlj = Jjl between sites l, j dictate the connectivity of the photonic network. For the
junction, Jlj = Jδl,j±1 , and ϵl = ϵ = 0.
The field dynamics given by Eq. S1 ensures that there are two constants of motion that represent the internal energy
and the total optical power, respectively,
X Xh χ i X
H({ψψ }) = − Jlj ψl∗ ψj + ϵl + |ψl |2 |ψl |2 ≈ ωα |Cα |2 ≡ E,
2 α
i̸=l l
X 2
X
2
N ({ψ
ψ }) = |ψl | = |Cα | ≡ A. (S2)
l α

Like in the main text, Cα represents the projection coefficient to the supermode α. To arrive to the rhs of Eq. (S2) we
have assumed weak nonlinearity. The presence of nonlinearities provides a mechanism for mode-mixing processes that
lead, ultimately, to thermal equilibrium. If the nonlinearity is weak, such a thermal state corresponds to a supermode
occupation n(ωα ) = |Cα |2 = ωαT−µ , i.e., given by the Rayleigh-Jeans (RJ) statistics [S1–S3], where {ωα } represents
the linear spectrum, and the Lagrange multipliers T = T (E, A), µ = µ(E, A) correspond to the optical temperature
and chemical potential.

CMT simulations.-

For our CMT simulations, we have employed two methods: (1) We conduct ’time’ simulations by integrating Eq. (S1)
for the entire system, encompassing the junction and the reservoirs. (2) Alternatively, we adopt the Monte-Carlo
thermostat approach [S4], where we replace the extensive collections of modes within the reservoirs with effective
thermostats that enforce specific temperature and chemical potential values (T, µ). Full simulations are time-consuming,
rendering the second method, reliant on Monte Carlo reservoirs, highly desirable for our calculations. Both methods
are elaborated upon below.

Full CMT simulations.-

Our CMT simulations have been performed by integrating Eq. (S1) using an 8-th order three-part split symplectic
integrator algorithm [S1, S2, S5, S6]. Such an algorithm guarantees the conservation of both the total optical power A
and internal energy E during the whole evolution with a numerical accuracy ensuring conservation up to O(10−8 ).
The lattice connectivity is such that the reservoirs consist of Nr = 40 × 40 = 1600 sites forming a square lattice
with coordination number z = 2 and a uniform nearest-neighbor coupling constant Jr = 1. To prevent spectral
degeneracies, we set the optical on-site potentials to be random numbers drawn from a uniform distribution within the
range [− W W
2 , 2 ] with W = 0.5. We have verified that a reservoir size Nr = 1600 guarantees a (quasi-)steady state of
the currents.
On the other hand, the junction corresponds to a 1D chain with coordination number z0 = 1, and it is connected to
the reservoirs at the corners with a junction-reservoir coupling Jj−r = 0.2J.
 2

In our simulations, we have prepared the initial states of the reservoir L(R) at energy and optical power (EL(R) , AL(R) ),
that after thermalization, will converge toward a RJ distribution with optical temperature and chemical potential
(TL(R) , µL(R) ). Following this protocol, we have generated an ensemble of M ≳ 50 realizations by setting random
phases for the supermode complex amplitudes Cα and integrated Eq. (S1) for times t ≳ 5 × 104 in units of J −1 .

Monte Carlo thermostats.-

This method, inspired in Ref. S4, enables us to replace the large number of modes in the reservoirs with an effective
thermostat, allowing us to concentrate solely on the field dynamics through the junction. In turn, this results in
more efficient simulations. Then, in the present study, the majority of our numerical results were obtained using this
approach.
At the Monte-Carlo thermostat, the field amplitudes can undergo a random perturbation (at random times)
ψm → ψm + δψm at sites m connected to the reservoirs. Such perturbations are accepted or denied based on the
standard Metropolis algorithm, which evaluates the cost function
 
∆H − µ∆N
ξ(δψm ) = exp − ,
T
where T , µ are temperature and chemical potential of the corresponding thermostat. ∆H and ∆N correspond to the
variation of the energy and optical power of the chain due to the perturbation, i.e., ∆H ≡ H(ψ δψ) − H(ψ
ψ + δψ ψ ) and
∆N ≡ N (ψ δψ) − N (ψ
ψ + δψ ψ ).

Evaluation of the currents.-

After neglecting transients, we can extract the optical power current ja and the internal energy current jh through
the junction via
∗
ja ≡ ⟨ja,n ⟩n,z = iJ ψn ψn−1 − ψn−1 ψn∗ z ,
D
∗
E (S3)
jh ≡ ⟨jh,n ⟩n,z = J ψ̇n ψn−1 + ψn−1 ψ̇n∗ ,
z

∂ja ∂jh
which follow from the continuity equations da dh
dz + ∂x = 0, and dz + ∂x = 0, respectively[S4]. The values of ja and jh do
not depend on the position, and an average over different sites n and on ’time’ z help in reducing fluctuations. However,
an alternative approach consists in extracting the currents by monitoring the variations of the optical power and
internal energy at the reservoirs. We have numerically verified that both approaches give the same results, nevertheless,
the latter is more efficient in reducing the number of realizations in the average. These methodologies can be used
with both the full CMT simulations or the Monte Carlo thermostats.

LINEAR RESPONSE THEORY

In the (quasi-)stationary regime, the power density a ≡ N /N and energy density h ≡ H/N at any segment dx inside
the junction (of length N ≫ dx) satisfy the continuity equations[S7, S8]
da dh
+ ∇ja = 0, + ∇jh = 0, (S4)
dz dz
where we assumed that the segment dx includes many unit cells while a(x) and h(x) are the values of these densities at
a position (site) x. Assuming that each segment (x, x + dx) is at local equilibrium, we can define a local temperature
and chemical potential, T = T (x), µ = µ(x), and thus study the transport using the framework of linear response
theory following the Onsager matrix formalism [S7, S8]:
 µ  
′ ′ 1
ja = Laa ∇ − + Lah ∇ ,
T T
 µ   (S5)
′ ′ 1
jh = Lha ∇ − + Lhh ∇ ,
T T
 3

where ∇(−µ/T ), ∇(1/T ) are so-called thermodynamical forces or affinities. An alternative formulation of the Onsager
matrix results from introducing the heat current jq = jh − µja ,
   
−1 1
ja = Laa ∇µ + Laq ∇ ,
T T
    (S6)
−1 1
jq = Lqa ∇µ + Lqq ∇ ,
T T
where now the new affinities read ∇(−µ)/T , ∇(1/T ). Both formulations are equivalent, thus, the relation among the
coefficients of the Onsager matrices L and L′ are given by
Laa = L′aa , Laq = L′ah − µL′aa ,
(S7)
Lqa = L′ha − µL′aa , Lqq = µ2 L′aa + L′hh − 2µL′ah ,
or
L′aa = Laa , L′ah = Laq + µLaa ,
(S8)
L′ha = Lqa + µLaa , L′hh = Lqq + 2µLaq + µ2 Laa .
′
As it follows from the Onsager theorem, det L̂ = det L̂ ⩾ 0, Laa ⩾ 0, Lqq ⩾ 0, L′aa ⩾ 0, L′hh ⩾ 0, Laq = Lqa ,
L′ah = L′ha . The latter are the well-known Onsager reciprocity relations.
All coupled transport properties are characterized by the Onsager matrix or, equivalently, by the transport coefficients,
which include the optical power conductivity, σ, thermal conductivity, κ, and the Seebeck coefficient, S, defined as
follows:
Laa
ja = −σ∇µ =⇒ σ = , (S9a)
∇T =0 T

det L̂
jq = −κ∇T =⇒ κ = , (S9b)
ja =0 T 2 Laa

1 Laq
∇µind = −S∇T =⇒ S = , (S9c)
ja =0 T Laa
where ∇µind stands for the chemical potential gradient induced by a temperature gradient without optical power
exchange between the thermostats.

BALLISTIC TRANSPORT AND LANDAUER THEORY

In the linear limiting case χ ≈ 0, there are no significant mode-mixing interactions and the modes of the junction
are freely propagating waves. Thus, the transport is essentially ballistic and currents can be analyzed using the linear
scattering framework provided by Landauer’s theory [S9],
Z
ja = dω · t(ω) · [nL (ω, µL , TL ) − nR (ω, µR , TR )] ,
Z (S10)
jh = dω · t(ω) · ω · [nL (ω, µL , TL ) − nR (ω, µR , TR )] ,

where nL,R (ω, µL,R , TL,R ) ≡ TL,R /(ω − µL,R ) are the equilibrium distribution functions of the reservoirs. The factor
t(ω) corresponds to the transmittance through the junction, which in our case t(ω) ≈ t0 for ω ∈ [−2J, 2J], and t(ω) = 0
otherwise. Then, by evaluating the integrals in Eq. (S10)
"    #
−µL + 2J −µR + 2J
ja = t0 TL ln − TR ln ,
−µL − 2J −µR − 2J
"    # (S11)
−µL + 2J −µR + 2J
jh = t0 4J(TL − TR ) + µL TL ln − µR TR ln .
−µL − 2J −µR − 2J
 4

The equations above are valid for arbitrary values of the temperature and chemical potential of the reservoirs.
Nevertheless, in order to obtain the Onsager coefficients, it is useful to consider the linear response regime, where
|∆µ| ≪ −µ − 2J and ∆T ≪ T . Then,
   
4JT −µ + 2J
ja = −t0 2 ∆µ + ln ∆T ,
µ − 4J 2 −µ − 2J
   
−µ + 2J
jh = −t0 4J∆T + T ln ∆µ + µja , (S12)
−µ − 2J
   
−µ + 2J
jq = −t0 4J∆T + T ln ∆µ ,
−µ − 2J
where we introduced T ≡ (TL + TR )/2, µ ≡ (µL + µR )/2, ∆T ≡ TR − TL , ∆µ ≡ µR − µL , and we omitted terms
O[(∆µ/µ)2 ], O[(∆T /T )2 ], O[(∆µ/µ)(∆T /T )].
With the analytical expressions obtained above, it is straightforward to obtain the Onsager matrix elements
4JT 2
 
2 4J
Laa = t0 N 2 , Laq = Lqa = −t 0 N T ln 1 − , Lqq = t0 N 4JT 2 , (S13)
µ − 4J 2 −µ + 2J
where we have replaced the gradients ∇(·) ←→ ∆(·)/N . At the same time, we can write the transport coefficients in
the ballistic regime as
4JT
σ=N t0 ,
µ2− 4J 2
µ2 − 4J 2 2
  
4J
κ = 4JN 1 − ln 1 − t0 , (S14)
16J 2 −µ + 2J
µ2 − 4J 2
 
−µ − 2J
S= ln .
4JT −µ + 2J
The distinctive feature of the ballistic regime, as confirmed by Eq. (S12) and Eq. (S14), is that currents remain
independent of the system size (N), while the conductivity scales linearly with N. In the ballistic regime, all modes are
delocalized, and no scattering mechanisms are involved. Transport in this regime can be visualized as a single direct
transfer from one reservoir to the other.

Effect of a weak nonlinearity

Next, we consider the impact of weak nonlinearity in the junction on currents and transport coefficients. As
discussed in the main text, the Landauer approach (Eq. S10) remains applicable, with the modification of introducing
a transmittance that maintains a constant value, denoted as t0 within a shifted frequency window defined by
ω ∈ [−2J + 2χa, 2J + 2χa]. This transmittance is zero everywhere outside this window. Here, we provide an
explanation of the mechanisms responsible for this frequency shift.
While a weak nonlinearity cannot enforce sufficient mode-mode mixing, it can induce a nonlinear frequency shift on
the α-th supermode frequency [S10]

ω̃α(J) ≈ ωα(J) + 2aχ. (S15)

Such a shift can be seen by considering the Hamiltonian of the junction (in absence of the reservoirs) in the normal
modes representation
X χ X
H= ωα(J) |Cα |2 + Vαβγδ Cα∗ Cβ∗ Cγ Cδ , (S16)
α
2
αβγδ

where the mode mixing factor

X
Vαβγδ = ϕ∗α (n)ϕ∗β (n)ϕγ (n)ϕδ (n), (S17)
n

describes the strength of the mode-mode interactions associated with the nonlinear mixing between supermodes. There,
ϕα (n) ≡ ⟨n|ϕα ⟩ represents a projection of the onsite amplitudes on the normal modes. The so-called secular terms, for
 5

which either α = δ, γ = β, or α = γ, β = δ, are the responsible to the frequency shift by contributing to the integrable
Hamiltonian component
 
X X
H0 = |ϕα |2 ωα(J) + 2χ Vαββα |ϕβ |2  , (S18)
α β

while the remaining terms, H − H0 , representing the integrability breaking processes, produce mode-mixing interactions.
By considering that ⟨|Cα |2 ⟩ ≈ a and invoking to the orthogonality of the supermodes ϕα , we arrive to Eq. (S15).
As a result, the frequency correction Eq. (S15) produces a uniform shift of the dispersion relation through the
junction that, in turn, results in wave propagation with a constant transmittance t0 within the frequency band
described above.

KINETIC EQUATION AND DIFFUSIVE TRANSPORT

In this section, we evaluate the integrals in Eq. (8) of the main text to obtain the currents and Onsager coefficients.
To that purpose, we introduce the linearized Kinetic Equation Eq. (7) of the main text into Eq. (8) and we obtain

Z+π   Z+π
∇µ dk 2 h
(0)
i2 1 dk
ja = − v (k) n (k) zT (k) + ∇ T vg2 (k) n(0) (k) zT (k),
T 2π g T 2π
−π −π
(S19)
Z+π   Z+π
∇µ dk 2 (0) 1 dk 2 2
jq = − T vg (k) n (k) zT (k) + ∇ T vg (k) zT (k).
T 2π T 2π
−π −π

The relaxation length zT (k) generally depends on k. Nevertheless, here we move forward by relying on a rough
h i
2 2
approximation where we assume zT (k) ≈ zT = cons., and the order of magnitude estimate of zT ∝ χ Ja tanh ζJa
T

has been previously evaluated in Ref. S11 (a = A/N ∼ 1 is the average value of norm per site inside the junction, and
(J)
ζ ≈ 8 is a best fitting parameter). Using ωα = −2J cos(kα ), we derive the Onsager matrix elements,

Zπ " #
2 1
2 sin2 k 2 µ
Laa = 4J T zT dk 2 = −T 1+ p zT ,
π [−2J cos k − µ] µ2 − 4J 2
0
Zπ
2 1
2 sin2 k h p i
Laq = Lqa = 4J T zT dk = −T 2 µ + µ2 − 4J 2 zT , (S20)
π −2J cos k − µ
0
Zπ
1
Lqq = 4J 2 T 2 zT dk sin2 k = 2J 2 T 2 zT .
π
0

Finally, the transport coefficients in the diffusive regime read

 
 1 
σ=T 2 − 1 zT ,

r 
2J
1− µ
 
(S21)
 2 s  2
1 2J 2J
κ = µ2 1 − − 1−  zT ,
2 µ µ
p
µ2 − 4J 2
S= .
T
 6

WIEDEMANN-FRANZ LAW

Here, we provide analytical expressions for the Wiedemann-Franz Law (WFL) in both the ballistic and diffusive
regime. For this purpose, we invoke the results that we have obtained in the previous sections, and the WFL reads

µ2 − 4J 2 2 1 |µ|≫2J J 2
  
4J

2 2



 µ − 4J 1 − 2
ln 1 − −−−−−→ (ballistic),


 16J −µ + 2J T T
κ
= (S22)
s 
2
 2 "  2 #  2
σ  µ 2J 1 2J 1 2J 1 |µ|≫2J J 2


2
 1− − 1− +  −−−−−→ (diffusive).
µ 2 µ 2 µ T T


Notice that this ratio does not depend on t0 nor zT . In both the ballistic or the diffusive regime when |µ| ≫ 2J, the
2
WFL takes the asymptotic form κσ ≈ JT that corresponds to the Eq. (12) discussed in the main text.
To confirm this result numerically, we calculate the ratio κ/σ from the elements of the Onsager matrix. The latter
can be numerically obtained by extracting the energy and power currents under appropriate preparations of the
corresponding affinities, as we describe next. First, we consider the situation where the temperature of the reservoirs
are the same, TL = TR , resulting in a vanishing thermal force ∇(1/T ) = 0. We denote the associated currents as
(T ) (T )
ja = L′aa · ∇ −µ and jh = L′ha · ∇ −µ



T T , which are extracted from various numerical simulations (with J = 1) at
different temperature values ranging from T = 0.5 to T = 6 and chemical potential with mean value µ = −5.0, and
∆µ/µ = 0.05. Next, we address the case where the other affinity vanishes, which requires  a preciserelation between the
−µ ∇µ
temperature and chemical potential of the reservoirs ∇ T = − T + T 2 ∇T = N T − ∆µ
µ µ ∆T


µ + T = 0. Our numerical
calculations consider various mean temperature values T = 0.5, . . . , 6.0 and a mean chemical potential µ = −5, such
(µ/T )
= L′ah · ∇ T1


that ∆T /T = −0.05 and ∆µ/µ = −0.05, from where we extract the power and energy currents ja
(µ/T )
= L′hh · ∇ T1 .


and jh
Finally, by using the definitions of the transport coefficients Eq. (S9), we calculate the WF ratio
(µ/T ) (µ/T ) (T )
!
1 L′hh L′ah L′ha
 
κ |µ| jh ja · jh
= 1− ′ ′ = 1 − (T ) (µ/T ) , (S23)
σ T L′aa Lhh Laa T ja(T ) ja · j h

where we have used ∆µ/|µ| = ∆T /T . We have used this protocol and Eq. (S23) to calculate all data points in Fig. (3)
of the main text.

[S1] A. Ramos, L. Fernández-Alcázar, T. Kottos, and B. Shapiro, Optical Phase Transitions in Photonic Networks: a
Spin-System Formulation, Physical Review X 10 (2020).
[S2] C. Shi, T. Kottos, B. Shapiro, Controlling optical beam thermalization via band-gap engineering, Phys. Rev. Research 3,
033219 (2021)
[S3] F. O. Wu, A. U. Hassan, and D. N. Christodoulides, Thermodynamic theory of highly multimoded nonlinear optical systems,
Nature Photonics 13, 776 (2019).
[S4] S. Iubini, S. Lepri, and A. Politi, Nonequilibrium discrete nonlinear Schrödinger equation, Physical Review E 86 (2012).
[S5] E. Gerlach, J. Meichsner, and C. Skokos, On the Symplectic Integration of the Discrete Nonlinear Schr̈odinger Equation
with Disorder, Eur. Phys J. Special Topics 225, 1103 (2016).
[S6] P. J. Channell and F. R. Neri, An Introduction to Symplectic Integrators, Field Inst. Commun. 10, 45 (1996).
[S7] H. B. Callen, Thermodynamics and an Introduction to Thermostatistics, John Wiley & Sons (1985).
[S8] K. Saito, G. Benenti, and G. Casati, A microscopic mechanism for increasing thermoelectric efficiency, Chemical Physics
375, 508, stochastic processes in Physics and Chemistry (in honor of Peter Hänggi) (2010).
[S9] Y. Imry and R. Landauer, Conductance viewed as transmission, Rev. Mod. Phys. 71, S306 (1999).
[S10] D. M. Basko, Kinetic theory of nonlinear diffusion in a weakly disordered nonlinear Schrödinger chain in the regime of
homogeneous chaos, Phys. Rev. E 89, 022921 (2014).
[S11] A. Y. Ramos, C. Shi, L. J. Fernández-Alcázar, D. N. Christodoulides, T. Kottos, Theory of localization-hindered
thermalization in nonlinear multimode photonics, Communications Physics 6, 189 (2023).