Quantum Charging Advantage Cannot Be Extensive Without Global Operations

Ju-Yeon Gyhm,1, 2, ∗ Dominik Šafránek,1, † and Dario Rosa1, ‡

1
Center for Theoretical Physics of Complex Systems,
Institute for Basic Science (IBS), Daejeon 34126, Republic of Korea
2
Department of Physics and Astronomy, Seoul National University, 1 Gwanak-ro, Seoul 08826, Korea
(Dated: April 11, 2022)
Quantum batteries are devices made from quantum states, which store and release energy in a
fast and efficient manner, thus offering numerous possibilities in future technological applications.
They offer a significant charging speedup when compared to classical batteries, due to the possibility
of using entangling charging operations. We show that the maximal speedup that can be achieved
is extensive in the number of cells, thus offering at most quadratic scaling in the charging power
over the classically achievable linear scaling. To reach such a scaling, a global charging protocol,
charging all the cells collectively, needs to be employed. This concludes the quest on the limits
arXiv:2108.02491v2 [quant-ph] 8 Apr 2022

of charging power of quantum batteries and adds to other results in which quantum methods are
known to provide at most quadratic scaling over their classical counterparts.

Introduction.— In recent years tremendous efforts have increase the performance of a quantum battery has been
been devoted to developing quantum technologies, which heavily studied. These studies address several figures of
are now coming to fruition in several fields of prac- merit, such as work extraction [11, 12], energy storage
tical use. Among the largest successes is quantum [13–17], charging stability [18–21], available energy
metrology [1], which led to the detection of gravitational [22–24] (with the notion of ergotropy [25]) and charging
waves [2], quantum cryptography [3], which finds appli- power [26–37], the last being the actual focus of this
cations in communicating sensitive data [4, 5], quantum letter.
computing, which promises to revolutionize chemistry [6] It has been shown [27] that quantum effects lead to
as well as to speed up or solve important problems in a speedup in the charging power of a quantum battery.
optimization, cybersecurity and data analysis [7], and The source of this quantum speedup lies in the use of en-
nanoscale thermodynamic devices, which offer unprece- tangling operations, in which the cells are charged collec-
dented precision in thermometry [8]. At large, society tively as a whole. Those operations, where the number of
is moving toward quantum technologies, because they cells that are being entangled together collectively scales
promise to offer faster, smaller, and more precise devices. with the system size (i.e., creating multi-partite entangle-
All of these achievements require an efficient way of ment), are called global operations. In contrast, classical
storing and using energy, as well as fast charging and dis- batteries are charged in parallel, meaning that each cell
charging. The necessity of charging and discharging goes is charged independently of each other. The advantage
well beyond the quantum world. Examples are electric of this collective versus parallel charging is measured by
vehicles where the charging time is one of the main bot- the ratio Γ, called the quantum charging advantage [27].
tlenecks in preventing the widespread use of such tech- However, it is still not known how large the quantum ad-
nology, or future fusion power plants, in which a large vantage is in general. To this end, the best known result
amount of energy needs to be pumped in a short amount is [27]
of time and discharged in an instant to start the reac-
tion. In the quantum world, nanoscale devices will re- Γ < γ(k 2 (m − 1) + k), (1)
quire nanoscale batteries, with no energy to spare.
Outstanding successes of quantum technologies in which γ is a model-dependent constant, k is the maxi-
prompt a question whether quantum effects can also mum number of cells that are collectively charged, while
improve the energy storage to satisfy current and future m (called participation number ) is the maximum num-
demands. This leads to the notion of quantum battery, ber of parallel charging operations in which a single cell
which is a quantum mechanical system acting as an appears.
energy storage, and in which quantum effects are ex- In principle, this bound allows for a super-extensive
pected to provide significant advantages over its classical scaling of the quantum advantage, meaning that the ad-
counterpart (see [9, 10] for reviews). Starting from the vantage can scale more than linearly with the number
work of Alicki and Fannes [11], the possibility of using L of cells. For example, consider a charging protocol
quantum effects (like coherence and entanglement) to that has a finite and fixed value of k but having all-to-all
couplings. In such a case, the participation number of
a given cell is of order m = (L−1
k−1
) ≈ (L − 1)k−1 /(k − 1)!,
∗ E-mail: [email protected] leading to a quantum advantage of order Lk−1 .
† E-mail: [email protected] This prompted a race toward finding the best possible
‡ E-mail: [email protected]; scaling — the authors of [27] found that the scaling is of
DŠ and DR contributed equally order L at most through an extensive numerical search,
 2

and proposed a conjecture that this extensive scaling can- Hamiltonian Ĥ. Taking the spectral decomposition of
not be surpassed. The search for scaling advantages con- the initial Hamiltonian to be Ĥ = ∑j Ej ∣Ej ⟩⟨Ej ∣, where
tinued in [13, 14, 28, 29, 32, 38], which also showed at we assume the energy levels Ej being ordered, we express
most extensive scaling, but it was later shown that some the driving Hamiltonian as V̂ = ∑N j,m=1 Vjm ∣Ej ⟩⟨Em ∣.
of these advantages were not caused by genuine quan- The limiting property is formalized as follows: we de-
tum effects [39]. A genuine, extensive, quantum advan- fine ∆E as the minimum value such that for all j and
tage was found in [33], in a setup including both global m,
charging operations and all-to-all couplings. The con-
jecture still held, but remained unproven, together with when ∣Ej − Em ∣ > ∆E, then Vjm = 0. (5)
uncertain role as to which all-to-all interactions play in
determining the quantum advantage. Thus, it is natural to look for a more precise bound
In this letter, we prove this conjecture, showing that than eq. (4), taking this common property into account.
a quantum battery provides at most extensive advantage Main result.— In the conditions of eq. (5), we now
over classical batteries. Furthermore, we show that this show that a more stringent bound can be derived.
scaling is achievable only via global charging operations,
i.e., we show that all-to-all interactions, and more gen- Theorem 1. For driving that couples energy levels with
erally, the participation number, does not provide any at most ∆E energy difference, as expressed by eq. (5),
scaling advantage. the instantaneous charging power is bounded as
We first provide a general bound (Theorem 1), con-
straining the maximum charging power achievable with ∣P (t)∣ ≤ ∆E ∥V̂ (t) − vmin (t)∥ /2, (6)
a general quantum battery with any general Hamilto-
nian, not necessarily realized by L identical cells, thus
where vmin (t) is the smallest eigenvalue of V̂ (t), and ∥ ∥
including also more general cases described in the liter-
denotes the operator norm.
ature [28]. The conjecture is proven as a consequence of
this theorem (Corollary 1), applied to the battery made Hence, the operator norm of the initial Hamiltonian,
of identical cells. Together with examples showing ex- central in inequality (4), is not the relevant figure of
tensive advantage [26, 33], already found and discussed merit. Instead, the crucial quantity is the maximal value
in the literature, this result concludes the quest for the of energy (as measured by Ĥ) that can be transferred
best possible scaling which can be obtained by quantum
by V̂ in a single time step. While non-trivial to prove,
batteries.
this result is very intuitive. The charging power is the
Setup.— We consider quantum batteries made out of
amount of the energy stored in the battery in a single
a time independent initial Hamiltonian, Ĥ, having dis-
time step. Thus, this change in energy must be bounded
crete spectrum. At time t = 0 a possibly time-dependent
by the maximum amount of energy that the driving term
driving Hamiltonian, V̂ (t), is turned on and the initial
can transfer to the system during that time.
state ρ̂0 is evolved according to the quench
The fact that the bound is not given by the operator
dρ̂t
= −i[V̂ (t), ρ̂t ]. (2) norm of Ĥ has a far reaching consequence. As outlined in
dt
the Introduction, it has been a matter of active research
The energy stored in the battery, measured by the initial which combination of initial and driving Hamiltonians
Hamiltonian, changes from E(0) = tr(Ĥ ρ̂0 ) to E(t) = can reach a charging power scaling with ∥Ĥ∥. Theorem
tr(Ĥ ρ̂t ) during time evolution. Charging the battery 1 shows that to reach this scaling one needs to consider
means reaching large values of E(t) − E(0). driving terms having non vanishing matrix elements be-
An important figure of merit is the instantaneous tween the ground state and the highly excited states of
charging power of the battery. It is defined as the instan- Ĥ. This latter property defines the so-called global charg-
taneous change in the energy stored per unit of time: ing operations which we discuss more extensively later.
Another point of this bound is that it applies to any
P (t) = tr(Ĥ dρ̂t
dt
). (3) Hamiltonian Ĥ, even with interacting cells.
Sketch of the proof.— The full proof of the theorem
where we used that Ĥ is time-independent. is technically involved. Here we sketch the main idea,
Generally, the instantaneous power is bounded in while we refer the reader to the Supplementary material
terms of the commutator between Ĥ and the driving term for details.
We express the commutator between Ĥ and V̂ as an
∣P (t)∣ ≤ ∥[Ĥ, V̂ (t)]∥ ≤ 2 ∥Ĥ∥ ∥V̂ (t)∥ (4) integral of commutators which are more easily and di-
rectly bounded. In particular, we define certain operator
through the operator norm [40]. functions ĥ(e) and v̂(e), depending on a continuous pa-
However, the driving is often limited in realistic situ- rameter e and satisfying
ations. For example, in lattice systems, the interaction
couples only nearby sites and therefore V̂ transfers energy ∆E
only between not-too-distant energy levels of the initial [Ĥ, V̂ ] = ∫ [ĥ(e), v̂(e)] de, (7)
e=0
 3

as well as the following properties

1 4
∥ĥ(e)∥ = , ∥v̂(e)∥ = ∥V̂ ∥ . (8)
2
We apply triangle inequality to eq. (7) to derive bound
3
∥[Ĥ, V̂ ]∥ ≤ ∆E ∥V̂ (t)∥. Since any number λ commutes P
2
with Ĥ, and V̂λ′ = V̂ −λ also satisfies eq. (5), we can make
this bound tighter by minimizing over λ,
1 ||[H, V]||
∆E Pmax
∥[Ĥ, V̂]∥=inf ∥[Ĥ, V̂λ′ ]∥≤inf
∆E∥V̂ −λ∥= ∥V̂ −vmin ∥ . Bound by Theorem 1
λ λ 2 0 4 5 6 7 8 9 10 11 12
The theorem then follows from eq. (4).
(9)
L
(a)
Lattice case.— In case of a battery made by cells, each
of them given, for example, by a qubit, Theorem 1 pro-
vides a much more stringent bound than other known 12
||[H, V]||
bounds [27]. 10 Pmax
We consider a battery composed of L identical cells, Bound by Theorem 1
having initial Hamiltonian 8 Bound by eq. (17)

L
P 6
Ĥ = ∑ Ĥ (l) , (10)
l=1 4
where Ĥ (l) = Iˆ ⊗ ⋯ ⊗ Iˆ ⊗ Ĥs ⊗ Iˆ ⊗ ⋯ ⊗ Iˆ and Ĥs is the 2
single site Hamiltonian at the l-th place. We charge this 0 2 3 4 5 6 7 8 9 10
battery by turning on the driving Hamiltonian,
L
V̂ (t) = ∑ V̂i (t), (11) (b)
i∈K(L,k)
FIG. 1. (a) Maximum power Pmax and maximum operator
where, by definition, each term in the summation couples norm of the commutator ∥[Ĥ, V̂ ]∥ as a function of L (maxi-
together at most k cells. Expressed mathematically, mized over all time and 500 realizations of disorder for each
value of L), starting from the ground state of initial Hamil-
k
K(L, k) = ⋃ C(L, n), (12) tonian Ĥ = ∑L z
l=1 hσ̂l and charged by driving Hamiltonian as
n=1 in eq. (80). For this example, we fixed h = 1. We observe a
C(L, n) = {(i1 , . . . , in )∣i1 < ⋯ < in and ij ∈ {1, . . . , L}}, slight decrease in Pmax and ∥[Ĥ, V̂ ]∥ with growing L, which is
a result of increasing dimensionality of the system, resulting
where C(L, n) is a set of all combinations of n sites, and in the lower chance of the of the initial state to be optimal.
Vi acts as an identity on the site which does not appear in While keeping the number of realizations of different driving
fixed at 500, which means that upper the bound is more diffi-
the index, i.e., for any local matrix M̂ (l) = Iˆ⊗⋯⊗ Iˆ⊗ M̂ ⊗
ˆ ˆ where M̂ is at the l-th place, if l ∈/ i = (i1 , . . . , in ), cult to reach. (b) The same as (a) but for a driving given by
I⊗⋯⊗ I, eq. (16). For this example, we additionally fixed V = 1.
then [M̂ (l) , V̂i (t)] = 0. The cases with k ∝ L are called
global operations. The corollary follows.
Corollary 1. For initial and driving Hamiltonians (62)
and (63), the instantaneous charging power is bounded To show that, we need to discuss each of these terms
as, separately: as explained above, k is the number of cells
being coupled together by the driving, and thus k ≤ L.
∣P (t)∣ ≤ k ∥Ĥs − Es min ∥ ∥V̂ (t) − vmin (t)∥ /2. (13) ∥Ĥs − Es min ∥ is a number that depends on particulars
where Es min is the single cell ground state energy. of a single cell but does not scale with L. The last
term, ∥V̂ (t) − vmin (t)∥, which we call potential (in anal-
The result is proven by showing that the maximum ogy with electric circuits), can be in principle made arbi-
energy jump ∆E in this case is given by k ∥Ĥs ∥. As trarily large. Physically, this would correspond to invest-
shown in the Supplementary Material, the corollary then ing a very large/infinite energy into the driving. With
follows directly from Theorem 1. larger driving energy, the charging is faster. Thus, to
The consequences of the bound in eq. (66) are remark- compute the quantum charging advantage, we need to
able. In particular, it rules out the possibility of having compare the parallel and quantum scaling on an equal
extensive quantum charging advantage without global footing, by assuming that the energy scale that is in-
charging operations. vested into the driving is the same in both cases. We
 4

do that by fixing the potential of the quantum driving C is chosen such that ∥V̂ − vmin ∥ = 2, to ensure a fair com-
to be at most equal to the potential of the parallel driv- parison between different realizations (instances). (C ∝
ing, ∥V̂ − vmin ∥ ≤ ∥V̂ ∥ − vmin
∥
∥. This is the constraint C0, L−3/2 , which for the SY Hamiltonian follows from the
introduced and argued for in Ref. [27]. replica formalism [41, 42]. We numerically confirm this
Parallel charging is given by k = 1 in driving Hamil- scaling in the Supplementary material.) The results are
tonian (63), while the initial state is assumed to be a shown in Fig. 3 (a). We clearly see that the power is
product state ρ̂ = ρ̂⊗L s . Thus, in the parallel charg- bounded by the degree of k-locality and not by the par-
ing scenario the driving affects each cell independently. ticipation number. As a result, we do not find any ex-
From this, we easily calculate that both the potential tensive charging advantage for this model as expected.
∥V̂ ∥ − vmin
∥
∥ = L∥V̂s∥ − vs∥min ∥ and the charging power Interestingly, we observe that both the maximum power
as well as the maximum value of the commutator norm
P ∥ = LPs∥ scale linearly with L. ρ̂s , ∥V̂s∥ − vs∥min ∥, and
Ps∥ denote the state, potential, and charging power of a ∥[Ĥ, V̂ ]∥
single cell, respectively.
Combining eq. (66), constraint C0, and the results for slightly decreases with the system size, L. This is a fi-
parallel charging, we bound the quantum advantage as nite size effect which sensitively reduces by further in-
creasing the system size. We present an analysis of this
phenomenon in the Supplemental Material.
∣P ∣ k∥Ĥs − Es min ∥∥V̂s − vs min ∥L
∥ ∥
∣Γ∣ = ≤ = γk, (14) Does global charging always lead to an extensive quan-
∣P ∥ ∣ 2∣Ps∥ ∣L tum advantage?— The presence of a global charging term
in V̂ does not guarantee an extensive charging advantage.
where γ is L and k-independent. Thus, the quantum As an example, consider a battery composed of L
advantage scales with the maximum number k of cells qubits having initial Hamiltonian Ĥ = ∑L z
l=1 hσ̂l and
that are coupled together by V̂ . If this number does not charged via the following driving
scale with the lattice size, L, then the quantum advan-
tage cannot scale with L, as extensively foreshadowed V L
V̂ = ( ∑ σ̂lx ⊗ σ̂l+1
x
+ ⊗ σ̂lx ), (16)
in the Introduction. The extensive scaling is possible ⌊L/2⌋ + 1 l=odd l=1
only for global interactions, k ∝ L. By showing that the
only source of the quantum advantage comes from the with V being a constant. From Theorem 1, we obtain
global entangling operations, we showed that the advan- P ≤ 2Lh ∥V̂ ∥ = 2LhV (using ∥Ĥs − Es min ∥ = 2h), due to
tage comes from genuine quantum effects. This addresses the second term representing a global operation, which
the discussion of the role of quantum-ness posed in rela- couples all of the sites at the same time. Presence of this
tion to the bound on charging power found in [39]. global charging term suggests possibility of an extensive
Finally, we ask what is the maximal scaling of power charging advantage.
with L that a quantum charging protocol can achieve. However, in this case an extensive quantum advan-
Clearly, using Eq. (66), the maximum charging power tage is not reached. This is because the nearest-
is given by the product of k, and by whatever scaling neighbor terms ∑l=odd σ̂lx ⊗ σ̂l+1x
, which provide non-
can be constructed from ∥V̂ − vmin ∥ (for now leaving con- extensive advantage, dominate the interaction. They
straint C0 behind). It is possible to artificially construct contribute V /(1 + 1/⌊L/2⌋) while the global term con-
some driving Hamiltonians that scale super-extensively, tributes only V /(⌊L/2⌋ + 1) to the total norm ∥V̂ ∥. This
i.e., with higher powers of L [28, 29]. However, such sub-extensive scaling is confirmed by the following, alter-
models are unphysical [33, 39], because they would lead native, inequality, which is derived in the Supplementary
to a free energy that is super-extensive in the thermo- material,
dynamic limit. Therefore, for any physical model, con-
L
sidering extensive energy ∥V̂ − vmin ∥ ∼ L, the maximal ∣P (t)∣ ≤ ∑ k ∥V̂k ∥ ∥Ĥs − Es min ∥ , (17)
charging power scales at most quadratically P ∼ L2 , for k=1
global operations k ∝ L. (Compare with the linear scal-
ing of parallel charging.) where V̂k is the k-local part of V̂ . From inequality (17) we
As an illustrative example, we study charging of a obtain P ≤ 4∥V̂2 ∥h + 2L∥V̂L ∥h = ( 1+1/⌊L/2⌋
4
+ ⌊L/2⌋+1
2L
)V h ≈
quantum battery by means of a driving Hamiltonian ob- 8V h for the present example, which indeed confirms that
tained via a simple generalization of the celebrated SY the power does not display an extensive advantage. As
Hamiltonian [41], i.e. a random, 2-local, all-to-all Hamil- a further confirmation, we explicitly computed the max-
tonian imum charging power for this driving Hamiltonian (16),
L
reported in Fig. 3 (b). We clearly see that the charging
V̂ = C ∑ ∑
α α α
Jij σ̂i σ̂j , (15) power stays well-below the threshold given by eq. (17).
i<j α=x, y, z Discussion and conclusions.— We found a bound on
the maximum charging power which can be achieved by
α
where the coupling constants Jij are randomly extracted charging a quantum battery via an external quench pro-
from a normal distribution and the normalization factor tocol.
 5

This bound shows that the maximum charging power While charging power of classical batteries scales linearly
is not dependent on the operator norm of the battery with the number of cells (∝ L), quantum batteries pro-
Hamiltonian, by which the amount of charged energy vide at most quadratic scaling in charging power (∝ L2 ).
is measured. Instead, it is governed by the maximum This quadratic scaling cannot be reached without global
energy difference, ∆E, between two eigenstates of the operations. However, the mere presence of global charg-
battery Hamiltonian for which the driving Hamiltonian ing operations does not always guarantee an extensive
has a non-vanishing matrix element. In other words, the charging advantage, as we demonstrated on an explicit
charging power is limited by the amount of energy that example.
the driving Hamiltonian can add into the battery in a sin- This work adds to other results, in which quantum
gle step, a result which a posteriori seems very natural. systems provide at most quadratic improvement over the
This bound can be applied to a general quantum battery, known classical method, like the Heisenberg limit in sen-
described by any Hamiltonian, even those of interacting sitivity scaling in quantum metrology over the classically
quantum cells. achievable shot-noise limit [1, 43, 44], and Grover’s search
When applied to quantum batteries made of L identi- algorithm [45], which is known to be asymptotically op-
cal cells, this bound provides a limit on how fast they can timal [46].
be charged as compared to classical batteries. The max-
imum speed by which a quantum battery can be charged The bound specifies, for a given battery Hamiltonian
depends only on the number k of cells interacting to- and for a given driving, the maximum instantaneous
gether in a single term. It does not depend on the par- charging power achievable in that particular setup. It
ticipation number, which is the number of independent does not give any information about the quantum state
terms in the driving Hamiltonian in which a single cell for which such a power can be achieved. This constitutes
appears. For example, pair-wise interactions can provide an interesting question for future research.
a quantum speedup by at most a factor of two, even in Acknowledgements.—– We acknowledge the support
the case of all-to-all couplings, where every cell is con- by the Institute for Basic Science in Korea (IBS-R024-
nected to every other cell. For a speedup of a factor of k, Y2 and IBS-R024-D1). DR would like to thank M. Car-
one needs to consider k-cell interactions, while the max- rega, J. Kim, J. Murugan and J. Olle for collaboration
imal speedup of L is achieved for L-particle interactions. on related projects.

[1] V. Giovannetti, S. Lloyd, and L. Maccone, Nature Pho- chines and batteries,” (2020), arXiv:2008.07889 [quant-
tonics 5, 222 (2011), arXiv:1102.2318 [quant-ph]. ph].
[2] B. P. e. A. Abbott (LIGO Scientific Collaboration and [11] R. Alicki and M. Fannes, Phys. Rev. E 87, 042123 (2013).
Virgo Collaboration), Phys. Rev. Lett. 116, 061102 [12] K. V. Hovhannisyan, M. Perarnau-Llobet, M. Hu-
(2016). ber, and A. Acín, Physical Review Letters 111 (2013),
[3] S. Pirandola, U. L. Andersen, L. Banchi, M. Berta, 10.1103/physrevlett.111.240401.
D. Bunandar, R. Colbeck, D. Englund, T. Gehring, [13] G. M. Andolina, D. Farina, A. Mari, V. Pellegrini,
C. Lupo, C. Ottaviani, J. L. Pereira, M. Razavi, J. Sham- V. Giovannetti, and M. Polini, Phys. Rev. B 98, 205423
sul Shaari, M. Tomamichel, V. C. Usenko, G. Vallone, (2018).
P. Villoresi, and P. Wallden, Advances in Optics and Pho- [14] Y.-Y. Zhang, T.-R. Yang, L. Fu, and X. Wang, Phys.
tonics 12, 1012 (2020), arXiv:1906.01645 [quant-ph]. Rev. E 99, 052106 (2019).
[4] C. Elliott, A. Colvin, D. Pearson, O. Pikalo, J. Schlafer, [15] F. Caravelli, G. Coulter-De Wit, L. P. García-Pintos,
and H. Yeh, arXiv e-prints , quant-ph/0503058 (2005), and A. Hamma, Phys. Rev. Research 2, 023095 (2020).
arXiv:quant-ph/0503058 [quant-ph]. [16] J. Q. Quach and W. J. Munro, Physical Review Applied
[5] Y.-A. Chen, Q. Zhang, T.-Y. Chen, W.-Q. Cai, S.-K. Liao, 14 (2020), 10.1103/physrevapplied.14.024092.
J. Zhang, K. Chen, J. Yin, J.-G. Ren, Z. Chen, et al., [17] A. Crescente, M. Carrega, M. Sassetti, and D. Ferraro,
Nature 589, 214 (2021). New Journal of Physics 22, 063057 (2020).
[6] Y. Cao, J. Romero, J. P. Olson, M. Degroote, P. D. [18] N. Friis and M. Huber, Quantum 2, 61 (2018).
Johnson, M. Kieferová, I. D. Kivlichan, T. Menke, [19] D. Rossini, G. M. Andolina, and M. Polini, Phys. Rev.
B. Peropadre, N. P. D. Sawaya, S. Sim, L. Veis, B 100, 115142 (2019).
and A. Aspuru-Guzik, arXiv e-prints , arXiv:1812.09976 [20] A. C. Santos, B. i. e. i. f. m. c. Çakmak, S. Campbell,
(2018), arXiv:1812.09976 [quant-ph]. and N. T. Zinner, Phys. Rev. E 100, 032107 (2019).
[7] S. Aaronson, Scientific American 298, 62 (2008). [21] D. Rosa, D. Rossini, G. M. Andolina, M. Polini, and
[8] F. Menges, P. Mensch, H. Schmid, H. Riel, A. Stem- M. Carrega, Journal of High Energy Physics 2020 (2020),
mer, and B. Gotsmann, Nature Communications 7, 10874 10.1007/jhep11(2020)067.
(2016). [22] G. M. Andolina, M. Keck, A. Mari, M. Campisi, V. Gio-
[9] F. Campaioli, F. A. Pollock, and S. Vinjanampa- vannetti, and M. Polini, Phys. Rev. Lett. 122, 047702
thy, “Quantum batteries - review chapter,” (2018), (2019).
arXiv:1805.05507 [quant-ph]. [23] F. Barra, Phys. Rev. Lett. 122, 210601 (2019).
[10] S. Bhattacharjee and A. Dutta, “Quantum thermal ma- [24] K. V. Hovhannisyan, F. Barra, and A. Imparato, Phys.
 6

Rev. Research 2, 033413 (2020).

[25] A. E. Allahverdyan, R. Balian, and T. M. Nieuwen-
huizen, Europhysics Letters (EPL) 67, 565–571 (2004).
[26] F. C. Binder, S. Vinjanampathy, K. Modi, and J. Goold,
New Journal of Physics 17, 075015 (2015).
[27] F. Campaioli, F. A. Pollock, F. C. Binder, L. Céleri,
J. Goold, S. Vinjanampathy, and K. Modi, Physical Re-
view Letters 118 (2017), 10.1103/physrevlett.118.150601.
[28] T. P. Le, J. Levinsen, K. Modi, M. M. Parish, and
F. A. Pollock, Physical Review A 97 (2018), 10.1103/phys-
reva.97.022106.
[29] D. Ferraro, M. Campisi, G. M. Andolina, V. Pellegrini,
and M. Polini, Phys. Rev. Lett. 120, 117702 (2018).
[30] G. M. Andolina, M. Keck, A. Mari, V. Giovannetti, and
M. Polini, Phys. Rev. B 99, 205437 (2019).
[31] A. Crescente, M. Carrega, M. Sassetti, and D. Fer-
raro, Physical Review B 102 (2020), 10.1103/phys-
revb.102.245407.
[32] S. Ghosh, T. Chanda, and A. Sen(De), Phys. Rev. A
101, 032115 (2020).
[33] D. Rossini, G. M. Andolina, D. Rosa, M. Carrega,
and M. Polini, Physical Review Letters 125 (2020),
10.1103/physrevlett.125.236402.
[34] S. Zakavati, F. T. Tabesh, and S. Salimi, “Bounds
on charging power of open quantum batteries,” (2020),
arXiv:2003.09814 [quant-ph].
[35] S. Ghosh, T. Chanda, S. Mal, and A. S. De, “Fast
charging of quantum battery assisted by noise,” (2020),
arXiv:2005.12859 [quant-ph].
[36] S. Seah, M. Perarnau-Llobet, G. Haack, N. Brunner, and
S. Nimmrichter, “Quantum speed-up in collisional battery
charging,” (2021), arXiv:2105.01863 [quant-ph].
[37] L. P. García-Pintos, A. Hamma, and A. del Campo,
Phys. Rev. Lett. 125, 040601 (2020).
[38] X. Zhang and M. blaauboer, “Enhanced energy transfer
in a dicke quantum battery,” (2018), arXiv:1812.10139
[quant-ph].
[39] S. Julià-Farré, T. Salamon, A. Riera, M. N. Bera,
and M. Lewenstein, Physical Review Research 2 (2020),
10.1103/physrevresearch.2.023113.
[40] It is proved by inserting Eq. (2) to Eq. (3), which
gives P (t) = tr(ρ̂[Ĥ, V̂ ]) = ∑k ρk ⟨ψk ∣[Ĥ, V̂ ]∣ψk ⟩ ≤
∥[Ĥ, V̂ ]∥ ≤ 2 ∥Ĥ∥ ∥V̂ ∥, using the spectral decomposition
ρ̂ = ∑k ρk ∣ψk ⟩⟨ψk ∣ and ∑k ρk = 1. See, for example,
Ref. [39].
[41] S. Sachdev and J. Ye, Phys. Rev. Lett. 70, 3339 (1993).
[42] A. J. Bray and M. A. Moore, Journal of Physics C: Solid
State Physics 13, L655 (1980).
[43] V. Giovannetti, S. Lloyd, and L. Maccone, Science 306,
1330 (2004), arXiv:quant-ph/0412078 [quant-ph].
[44] R. Demkowicz-Dobrzański, J. Kołodyński, and M. Guţă,
Nature Communications 3, 1063 (2012), arXiv:1201.3940
[quant-ph].
[45] L. K. Grover, in Proceedings of the twenty-eighth annual
ACM symposium on Theory of computing (1996) pp. 212–
219.
[46] C. H. Bennett, E. Bernstein, G. Brassard, and U. Vazi-
rani, SIAM journal on Computing 26, 1510 (1997).
 7

SUPPLEMENTAL MATERIAL FOR: ∆E ∆E

“QUANTUM CHARGING ADVANTAGE ∆E
CANNOT BE EXTENSIVE WITHOUT GLOBAL
1
OPERATIONS”
e 2
− 21 1
⋯ − 12 1
2 2
I. PROOF OF THEOREM 1
0
E1 E2 E3 E4 E5 . . . Ej . . . EN −1 EN
Let us consider an initial Hamiltonian, Ĥ, which is as-
sumed to be time independent, having spectral decompo-
FIG. 2. Values ĥj (e) (1/2 as the blue region, and −1/2 as the
sition Ĥ = ∑Nj=1 Ej ∣Ej ⟩⟨Ej ∣, where the energies Ej are or- red region) as a function of index j, which relates to eigenstate
dered, i.e. Ej ≤ Em for j < m. Let us consider a different energy Ej depicted on the x-axis, and of parameter e depicted
Hamiltonian V̂ , which we call the driving Hamiltonian, on the y-axis. Condition (18) means that for energies Ej and
which we will write in basis of the initial Hamiltonian as Em that are further than ∆E, the corresponding element of
V̂ = ∑N j,m=1 Vjm ∣Ej ⟩⟨Em ∣. ∆E is the minimum number the driving Hamiltonian Vjm is zero.
such that for all j and m,
when ∣Ej − Em ∣ > ∆E, then Vjm = 0. (18)
as well as the commutator of operator ĥ(e) with operator
We also consider the spectral decomposition of the driv- v̂(e)
ing Hamiltonian as V̂ = ∑α α∣α⟩⟨α∣. Out of Ĥ and V̂ we
define other two operators ĥ(e) and v̂(e), functions of a ∆E

continuous variable e satisfying 0 < e ≤ ∆E. Explicitly, ∫ [ĥ(e), v̂(e)]de

e=0
the definition of ĥ(e) takes the following form ∆E N
=∫ ∑ [ĥ(e), vjm (e)∣Ej ⟩⟨Em ∣]de
N e=0 j,m=0
ĥ(e) = ∑ hj (e)∣Ej ⟩⟨Ej ∣ , (19)
N ∆E
j=1
= ∑ (∫ (hj (e) − hm (e))vjm (e)de)∣Ej ⟩⟨Em ∣.
where the functions hj (e) are defined as follows j,m=0 e=0
(24)
E −e
1
for ⌊ ∆E
j
⌋ is odd
hj (e) = {2 Ej −e . (20) We prove this lemma by comparing elements
− 12 for ⌊ ∆E ⌋ is even

By definition, ĥ(e) has the same eigenvectors as Ĥ but ajm = (Ej − Em )Vjm (25)
its eigenvalues are restricted to be either − 12 or 12 . The
in eq. (23) and
eigenvalues of ĥ(e), as functions of e, are depicted in
∆E
Fig. 2. In turn, v̂(e) is defined using ĥ(e) as follows
bjm = ∫ (hj (e) − hm (e))vjm (e)de (26)
N e=0
v̂(e) = ∑ vjm (e)∣Ej ⟩⟨Em ∣. (21) eq. (24), and showing that they are equal.
j,m=1
When a pair of j, m satisfies ∣Ej − Em ∣ > ∆E, Vjm and
Vjm for (j − m)(hj (e) − hm (e)) ≥ 0 vjm (e) are zero by definition(eqs. (18) and (21)), which
vjm (e) = {
−Vjm for (j − m)(hj (e) − hm (e)) < 0 means that the corresponding elements ajm and bjm are
both zero and thus equal.
By their definition, we can deduce the following lemmas In the opposite case, i.e. when ∣Ej − Em ∣ ≤ ∆E, Vjm
for v̂(e) and ĥ(e). and vjm (e) may not be zero. In this case, we explicitly
Lemma 1. We can express the commutator between the have
initial and the driving Hamiltonians using v̂(e) and ĥ(e) (hj (e) − hm (e))vjm (e)
as follows: ⎧
⎪
⎪(hj (e) − hm (e))Vjm for j > m, hj (e) > hm (e)
∆E ⎪
⎪
⎪
[Ĥ, V̂ ] = ∫ [ĥ(e), v̂(e)]de. (22) ⎪
⎪
⎪(hm (e) − hj (e))Vjm for j > m, hj (e) < hm (e)
e=0 ⎪
= ⎨(hm (e) − hj (e))Vjm for j < m, hj (e) > hm (e)
Proof. Let us write explicitly the commutator of initial ⎪
⎪
⎪
⎪
⎪
⎪(hj (e) − hm (e))Vjm for j < m, hj (e) < hm (e)
Hamiltonian ,Ĥ, with driving Hamiltonian ,V̂ , as ⎪
⎪
⎪
⎩0 for hj (e) = hm (e)
N
[Ĥ, V̂ ] = ∑ [Ĥ, Vjm ∣Ej ⟩⟨Em ∣] ⎧
⎪ Vjm for j > m, hj (e) ≠ hm (e)
⎪
⎪
⎪
j,m=0
= ⎨−Vjm for j < m, hj (e) ≠ hm (e)
N ⎪
⎪
⎪
= ∑ (Ej − Em )Vjm ∣Ej ⟩⟨Em ∣ (23) ⎪
⎩0 for hj (e) = hm (e)
j,m=0 (27)
 8

For a fixed j, m, this equation implies that the inte- while in case 2 we obtain
grand (hj (e)−hm (e))vjm (e) is a piecewise constant func-
⎧
⎪ for e ≤ xj
tion, which is equal to either 0 and Vjm when j > m, or ⎪
⎪Vjm
⎪
0 and −Vjm when j < m. To figure out at which points (hj (e) − hm (e))vjm (e) = ⎨0 for xj < e < xm (35)
⎪
⎪
⎪
this function jumps, we need to take a look at the defi- ⎪
⎩Vjm for e ≥ xm
nition of hm (e), eq. (20). Then to calculate bjm , we just
need to integrate over this piecewise function, which is In case 1, we compute
straightforward once we know where it jumps.
In order to identify this jump, we need to determine in bjm = Vjm (xj − xm ) = Vjm (Ej − Em ) = ajm . (36)
which region, in the interval 0 ≤ e < ∆E, hj (e) and hm (e)
are not equal. First, we assume that j > m. Hence, Ej is In case 2, we compute
also bigger than Em , since we are assuming that Ej are
ordered. Let us rewrite Em as bjm = Vjm (xj + ∆E − xm ) = Vjm (Ej − Em ) = ajm . (37)

Em = nm ∆E + xm (28) In both cases, we have shown that ajm = bjm . The case of
m > j follows from identical arguments. This concludes
for the maximal integer nm such that xm is a real number the proof.
in the interval 0 ≤ x < ∆E. This shows

Em − e nm ∆E + xm − e nm for xm ≥ e Lemma 2. Operator norms of v̂(e) and V̂ are equal for

⌊ ⌋=⌊ ⌋={
∆E ∆E nm − 1 for xm < e every e, expressed mathematically as
(29)
which implies that hm is a piecewise constant function ∥v̂(e)∥ = ∥V̂ ∥ . (38)
made from two pieces with the jump at e = x.
Given that Ej and Em differ at most by ∆E, we con- Proof. We prove lemma 2 by showing that v̂(e) and V̂
clude that in the decomposition have same spectra. Then the Lemma follows directly
from the definition of the operator norm. Operator norm
Ej = nj ∆E + xj , (30) is defined as

∥Â∣ψ⟩∥
nj can be either equal to nm or nm + 1. Let us first ∥Â∥ = sup (39)
consider the case nj = nm . In this case (case 1), we have ∣ψ⟩ ∥∣ψ⟩∥

Ej − e nm ∆E + xj − e nm for xj ≥ e where the supremum goes over all vectors ∣ψ⟩. In the
⌊ ⌋=⌊ ⌋={ (31) case of bounded Hermitian operators, this is equivalent
∆E ∆E nm − 1 for xj < e
to the supremum of absolutes value of eigenvalues of the
Noticing that our assumption Ej > Em implies that xj > operator Â,
xm , we have
∥Â∥ = sup ∣λ∣, (40)
⎧
⎪ for e ≤ xm
λ
⎪
⎪
⎪
0
hj (e) − hm (e) = ⎨(−1)nm +1 for xm < e ≤ xj (32) which means that the same spectrum directly imply the
⎪
⎪
⎪ same operator norm between different operators.
⎪
⎩0 for e > xj
We express each eigenvector ∣α⟩ of the driving Hamilto-
In the opposite case (case 2), i.e. when nj = nm + 1, our nian V̂ in terms of eigenvectors of the initial Hamiltonian
assumption Ej > Em implies xj < xm , and we have Ĥ as
N
⎧
⎪(−1)nm for e ≤ xj ∣α⟩ = ∑ βj ∣Ej ⟩. (41)
⎪
⎪
⎪ j=1
hj (e) − hm (e) = ⎨0 for xj < e ≤ xm (33)
⎪
⎪
⎪ By definition of an eigenstate, the following equation
⎪
⎩(−1)
nm +1
for e > xm
holds:
Then to get the full integrand, we plug in the vjm (e) N N
again, whose sole role is to transform the negative sign α ∑ βj ∣Ej ⟩ = α∣α⟩ = V̂ ∣α⟩ = ∑ Vjm βm ∣Ej ⟩. (42)
j=1 j,m=0
into the positive sign and to add Vjm . This in case 1
gives Therefore, the fact that ∣α⟩ is an eigenstate of driving
⎧ Hamiltonian V̂ is equivalent to that for all j,
⎪
⎪
⎪0 for e ≤ xm
⎪
(hj (e) − hm (e))vjm (e) = ⎨Vjm for xm < e < xj (34) N
⎪
⎪
⎪ αβj = ∑ Vjm βm . (43)
⎪
⎩0 for e ≥ xj m=1
 9

Further, for each eigenvalue of the driving Hamiltonian Because k is the only point where value h(e) changes,
α, and for any arbitrary but fixed value of e, we define it means that hj (e) = hk+1 (e) and hm (e) = hk (e), which
the state ∣γ(e)⟩ (will we omit writing the dependence on means we can rewrite the above equation as
α for clarity), which we will prove to be an eigenvector
⎧ βm
of operator v̂(e) that corresponds to the same eigenvalue γm (e) ⎪⎪ for hj (e) − hm (e) > 0
(but now of operator v̂(e)) α. Thus, by doing this we = ⎨ βj (49)
γj (e) ⎪⎪− βm for hj (e) − hm (e) < 0.
will show that V̂ and v̂(e) have the same spectrum. We ⎩ βj
define state ∣γ(e)⟩ as
If such k does not exist, then hm (e) = hm+1 (e) = ⋯ =
N hj (e), in which case γm (e)/γj (e) = βm /βj , so by includ-
∣γ(e)⟩ = ∑ γj (e)∣Ej ⟩. (44) ing this case we can generalize this the above equation
j=1 to
⎧
⎪ βm
Coefficients γj (e) are defined by a recurrence relation γm (e) ⎪ for hj (e) − hm (e) ≥ 0
= ⎨ βj (50)
γj (e) ⎪⎪− βm for hj (e) − hm (e) < 0.
γ1 (e) = β1 ⎩ βj
⎧
⎪ βj+1
The case of m > j follows from identical arguments,
⎪ γj (e) βj , for hj+1 (e) ≥ hj (e), (45)
γj+1 (e) = ⎨ which result in a change of sign. Combining cases j > m,
⎪
⎪
β
−γ (e) βj+1 , for hj+1 (e) < hj (e).
⎩ j j j < m and j = m (which is trivial) together, we finally
obtain the ratio for any j and m as
Since Vjm and vjm (e) are zero when ∣Ej − Em ∣ > ∆E (as
follows from eqs. (18) and (21)), we just need to consider ⎧ βm
pairs of j, m that satisfy ∣Ej − Em ∣ ≤ ∆E. γm (e) ⎪⎪ for (hj (e) − hm (e))(j − m) ≥ 0
= ⎨ ββjm
First, we assume j > m. We will show that there is γj (e) ⎪⎪− β for (hj (e) − hm (e))(j − m) < 0.
⎩ j
at most one index k, j > k ≥ m, for which hk+1 (e) and (51)
hk (e) are different. This statement is relatively clear Combining eq. (21), (43) and eq. (51) we obtain
from Fig. 2: for ∣Ej − Em ∣ ≤ ∆E and a fixed e, the hor-
izontal line drawn at height of e can cross the red re- N
γm (e) N
βm
gion into blue (or blue into red) at most once. To prove ∑ vjm (e) = ∑ Vjm = α. (52)
m=1 γj (e) m=1 βj
this statement mathematically (which duplicates the vi-
sual proof), for a contradiction we assume that there are Using this property, we have
two distinct indexes k and k ′ such that j > k ′ > k ≥ m,
where the sign changes, i.e., for which hk+1 (e) ≠ hk (e) N
and hk′ +1 (e) ≠ hk′ (e). By definition of hk (e) (eq. (20)), v̂(e)∣γ(e)⟩ = ∑ vjm (e)γm (e)∣Ej ⟩
j,m=1
(53)
Ek′ +1 − Ek > ∆E. (46) N
= α ∑ γj (e)∣Ej ⟩ = α∣γ(e)⟩.
However, from j > k ′ > k ≥ m and from the fact that we j=1

assumed our energy levels to be ordered, we have

This directly implies that ∣γ(e)⟩ is an eigenstate of v̂(e)
Ej − Em ≥ Ek′ +1 − Ek (47) associated with eigenvalue α. Therefore, V̂ and v̂(e) have
the same spectrum. By Eq. (40), ∥V̂ ∥ = ∥v̂(e)∥, which
which according to eq. (46) means that Ej − Em > ∆E, concludes the proof.
which is in contradiction with our assumption ∣Ej −Em ∣ ≤
∆E. Theorem 1. (Extended version of Theorem 1. shown in
So we just showed that for every combination j > m the main text) For driving that couples energy levels with
there exists at most one index k, j > k ≥ m, such that h(e) at most ∆E energy difference, as expressed by eq. (18),
changes sign at this index, i.e., hk (e) ≠ hk+1 (e). Using the following series of inequalities holds.
this knowledge, we compute a ratio that will be useful
later in showing that ∣γ(e)⟩ is an eigenstate of v̂(e). In ∥[Ĥ, V̂ ]∥ ≤ ∆E inf ∥V̂ − D̂∥ ≤ ∆E ∥V̂ − vmin ∥ /2 ≤ ∆E ∥V̂ ∥ ,
D̂
the case when k exists, we obtain ratio
(54)
where D̂ are matrices diagonal in the eigenbasis of the
γm (e) γs (e)
j−1
= ∏ initial Hamiltonian Ĥ.
γj (e) s=m γs+1 (e)
⎧
⎪
(48) Proof. By definition (20) the eigenvalues of ĥ(e) are ei-
⎪ ∏s=m βs+1 = βj for hk+1 (e) − hk (e) > 0
j−1 βs βm
=⎨ ther 12 or − 21 . Hence
⎪ − j−1 βs = − βm for hk+1 (e) − hk (e) < 0
⎩ ∏s=m βs+1
⎪ βj
1
by the definition of γ (eq. (45)). ∥ĥ(e)∥ = (55)
2
 10

holds for all e’s. Using lemma 1, ∥[Ĥ, V̂ ]∥ can be rewrit- We charge this battery by turning on the driving
ten as Hamiltonian,
∆E
∥[Ĥ, V̂ ]∥ = ∥∫ [ĥ(e), v̂(e)]de∥ . (56) V̂ (t) = ∑ V̂i (t), (63)
e=0 i∈K(L,k)

We apply the triangle inequality of operator norm two where, by definition, each term in the summation couples
times (once for the first and once for the second inequal- together at most k cells. Expressed mathematically,
ity) and obtain
k
∆E K(L, k) = ⋃ C(L, n), (64)
∥∫ [ĥ(e), v̂(e)]de∥ n=1
e=0
C(L, n) = {(i1 , . . . , in )∣i1 < ⋯ < in and ij ∈ {1, . . . , L}},
∆E ∆E
≤∫ ∥[ĥ(e), v̂(e)]∥ de ≤ ∫ 2 ∥ĥ(e)∥ ∥v̂(e)∥ de.
e=0 e=0 where C(L, n) is a set of all combinations of n sites, and
(57) Vi acts as an identity on the site which does not appear in
the index, i.e., for any local matrix M̂ (l) = Iˆ⊗⋯⊗ Iˆ⊗ M̂ ⊗
By lemma 2 and eq. (55), the right hand side of this ˆ
I⊗⋯⊗ ˆ where M̂ is at the l-th place, if l ∈/ i = (i1 , . . . , in ),
I,
equation equals then
∆E ∆E
∫ 2 ∥ĥ(e)∥ ∥v̂(e)∥ de = ∫ ∥V̂ ∥ de = ∆E ∥V̂ ∥ , [M̂ (l) , V̂i (t)] = 0. (65)
e=0 e=0
(58)
Corollary 1. For identical cells, which have initial
Since any diagonal matrix D̂ = ∑N j=1 λj ∣Ej ⟩⟨Ej ∣, com- Hamiltonian Ĥ = ∑L l=1 Ĥ
(l)
, where Ĥ (l) = Iˆ⊗ ⋯ ⊗ Iˆ⊗ Ĥs ⊗
mutes with Ĥ, we can make this bound tighter by min- Iˆ⊗ ⋯ ⊗ Iˆ and the driving Hamiltonian is of form (63), the
imizing over D̂. The key is to realize that if we de- following inequality holds:
fine V̂ ′ = V̂ − D̂, where V̂ satisfies eq. (18), then also
V̂ ′ = ∑N j,m=1 (Vjm − δjm λj )∣Ej ⟩⟨Em ∣ satisfies the same ∣P (t)∣ ≤ k ∥Ĥs − Emin ∥ ∥V̂ (t) − vmin (t)∥ /2,
(s)
(66)
equation, with the same ∆E. Thus, equation (58) holds
also for operator V̂ ′ , and we have (s)
where Emin is the single cell ground state energy. (In the
∥[Ĥ, V̂ ]∥ = ∥[Ĥ, V̂ − D̂]∥ ≤ ∆E ∥V̂ − D̂∥ , (59) main text, we denoted Es min ≡ Emin for better readabil-
(s)

ity.)
which holds for any D̂ which is diagonal in the energy
basis. Proof. The single site Hamiltonian has spectral decom-
We can take the infimum over all such diagonal matri- j=1 Ej ∣Ej ⟩⟨Ej ∣, where Ej
position as Ĥs = ∑N s (s) (s) (s) (s)
is or-
ces, which gives
dered by increasing energy, and ∣Ej ⟩ is the single site
(s)

∥[Ĥ, V̂ ]∥ = inf ∥[Ĥ, V̂ − D̂]∥ ≤ ∆E inf ∥V̂ − D̂∥ (60) energy eigenstate. The basis of the initial Hamiltonian is
D̂ D̂ rewritten as
This represents the tightest bound obtained by the (Ns ,⋯,Ns )
present method. Ĥ = ∑ Ej ∣Ej ⟩⟨Ej ∣
(s) (s) (s)
(67)
To prove the theorem as it is written in the main text, j=(1,⋯,1)
we restrict ourselves to the case of D̂ = λI.
ˆ In this case
we can evaluate the infimum and obtain explicitly where j = (j1 , j2 , ⋯, jL ), Ej = ∑L
l=1 Ejl , and ∣Ej ⟩ =
(s) (s) (s)

∣Ej1 ⟩ ⊗ ⋯ ⊗ ∣EjL ⟩.
(s) (s)
∥[Ĥ, V̂ ]∥ ≤ ∆E inf ∥V̂ − D̂∥
D̂
(61) An element of the driving Hamiltonian given by
≤ ∆E inf ∥V̂ −λI∥
ˆ = ∆E∥V̂ −vmin ∥ /2, eq. (63), V̂i , is decomposed in the basis of the initial
λ Hamiltonian as
which proves the theorem. (Ns ,⋯,Ns )
V̂i = ∑ Vi,jm ∣Ej ⟩⟨Em ∣.
(s)(s)
(68)
j,m=(1,⋯,1)

II. LATTICE CASE To understand this complicated expression, we illus-

trate the use of indexes as follows: i denotes positions of
We consider a battery composed of L cells described all the sites which interact through V̂i (for example i =
by initial Hamiltonian (1, 2) means that only the first and the second site inter-
act), while j = (j1 , . . . , jL ) and m = (m1 , . . . , mL ) denote
L
Ĥ = ∑ Ĥ (l) , (62) energy eigenstates of each site (for example, for L = 4,
l=1 m = (1, 2, 1, 4) corresponds to tensor product of local
 11

energy eigenstates ∣E(1,2,1,4) ⟩ = ∣E1 ⟩∣E2 ⟩∣E1 ⟩∣E4 ⟩).

(s) (s) (s) (s) (s)
This holds for every i that make up V̂ , thus also
Then by definition (using additionally j = (1, 3, 2, 1)),
Vjm = 0, (75)
V(1,2),(1,3,2,1)(1,2,1,4) = ⟨E(1,3,2,1) ∣V̂(1,2) ∣E(1,2,1,4) ⟩
(s) (s)
(69)
which, by explanation below eq. (72), proves the Corol-
Many of the elements of Vi are zero. To show that, let lary.
us assume that l ∉ i. Then choosing M̂ = ∣Ejl ⟩⟨Ejl ∣, from
eq. (65) we obtain Finally, we prove eq. (17) in the main text.
0 = ⟨Ej ∣[M̂ (l)
, V̂i (t)]∣Em ⟩ = ⟨Ej ∣V̂i ∣Em ⟩(1 − δjl ml ) (70) Corollary 2. Let us consider identical cells, which have
initial Hamiltonian Ĥ = ∑Ll=1 Ĥ
(l)
, where Ĥ (l) = Iˆ ⊗ ⋯ ⊗
Thus, if jl ≠ ml , then
Iˆ⊗ Ĥs ⊗ Iˆ⊗ ⋯ ⊗ I.
ˆ Where we rewrite driving Hamiltonian
Vi,jm ≡ ⟨Ej ∣V̂i ∣Em ⟩ = 0. (71) V̂ = ∑i∈K(L,k) V̂i as

To summarize, for any l ∉ i, if jl ≠ ml , then Vi,jm = 0. In N N

other words, every element Vi,jm that is non-zero must V̂ = ∑ ∑ V̂i = ∑ V̂k . (76)
have the same indexes ml = jl for every l ∈/ i. k=1 i∈C(L,k) k=1
For example, element (69) is zero, because 3 ∉ i =
(1, 2), and 2 = j3 ≠ m3 = 1. Elements that are allowed The following inequality holds:
to be non-zero are, for example, V(1,2),(1,3,1,1)(1,2,1,1) , L
V(1,2),(2,4,2,1)(1,2,2,1) , or V(1,2),(3,3,3,3)(3,2,3,3) . ∣P (t)∣ ≤ ∑ k ∥V̂k − vk,min ∥ ∥Ĥs − Emin ∥ /2.
(s)
(77)
In order to prove the Corollary, we must show that for k=1
all j and m such that ∣Ej − Em ∣ > k ∥Ĥs − Emin ∥ (where
(s)
Proof. By triangular inequality, the charging power is
Ej and Em are eigen energies of the full Hamiltonian bounded as
Ĥ = ∑l Ĥ (l) ) implies
N
Vjm = 0. (72) ∣P (t)∣ ≤ ∥[Ĥ, V̂ ]∥ ≤ ∑ ∥[Ĥ, V̂k ]∥ . (78)
k=1
If we manage to show that, because ∆E is the mini-
mum of all such numbers (see eq. (18)), it must be that Using Corollary 1, we obtain
k ∥Ĥs − Emin ∥ ≥ ∆E. Thus, the Corollary will follow from
(s)
N
Theorem 1. ∣P (t)∣ ≤ ∑ ∥[Ĥ, V̂k ]∥
k=1
Because ∥Ĥs − Emin ∥ is the difference between the
(s)
(79)
L
≤∑ k ∥V̂k − vk,min ∥ ∥Ĥs − Emin ∥ /2.
(s)
largest and the smallest eigenvalue of Ĥs , it takes more
than k sites to achieve energy difference ∣Ej − Em ∣ big- k=1

ger than k ∥Ĥs − Emin ∥, meaning that at least k + 1 in-

(s)

dices (elements) in j and m must differ. To prove that

mathematically, for a contradiction we assume that there
are k ′ ≤ k cells in which j and m are different, while III. RANDOM ALL-TO-ALL HAMILTONIAN
∣Ej − Em ∣ > k ∥Ĥs − Emin ∥. We have
(s)

In this section, we estimate the normalization factor in

L k′ the SY-like Hamiltonian used as an example in the main
∣Ej − Em ∣ = ∣ ∑ Ejl − Em ∣ = ∣ ∑ Ejπ(l) − Em ∣
(s) (s) (s) (s)
l π(l)
text, and we provide an explanation of why in Fig 1. (b)
l=1 l=1 the maximum decreases slightly with growing size of the
(73)
k′ system L.
≤ ∑ ∣Ejπ(l) − Em ∣ ≤k ∥Ĥs − Emin ∥ ,
(s) (s) ′ (s)
π(l) Let us consider a battery made of a random, 2-local,
l=1
all-to-all driving Hamiltonian given by
where π(l) is a relabeling of indices to include only those
where jl and ml are different. This is in contradiction V̂ = C V̂0 , (80)
with our assumption. Thus, if ∣Ej − Em ∣ > k ∥Ĥs − Emin ∥,
(s)
where
then vectors j and m have at least k+1 different elements.
However, since vector i in eq. (71) has at most k ele- L

ments according to the assumption of the Corollary, but V̂0 = ∑ ∑

α α α
Jij σ̂i σ̂j . (81)
i<j α=x, y, z
there are at least k + 1 elements of j and m that differ,
according to that equation it must be that α
the couplings Jij are randomly extracted from a normal
Vi,jm = 0. (74) distribution and C = ∥V̂ −v2 ∥
is a normalization factor
0 0,min
 12

which ensures that ∥V̂ − vmin ∥ = 2. We consider the initial

Hamiltonian to be Ĥ = ∑L z
i=1 hσ̂i . 60 Mean
Since C depends on Jij α
, which are randomly extracted Standard Deviation
50 Trend Line
from a normal distribution, we see that C as well changes

||V0 v0, min||/2

randomly in each simulation. Our goal is to estimate the 40
scaling of the quantity ∥V̂0 − v0,min ∥ with the system size,
30
L.
To this end, we start by recalling that, for large L, the 20
eigenvalues of V̂0 , v0,i , follow a Gaussian distribution. In
10
particular, they satisfy
0
2L 2 4 6 8 10 12 14 16
⟨∑ v0,i ⟩J = ⟨Tr(V̂0 )⟩J = 0, (82) L
i=1
FIG. 3. Mean and standard deviation of ∥V̂0 − v0 min ∥ /2 for
as well as each L. Total number of realization is 10000. The√ numerical
2L
data are fitted against an Ansatz of the form α L2 (L − 1).
σ 2 2L =∶ ⟨∑ v0,i 2 ⟩J = ⟨Tr(V̂02 )⟩J The fitting parameter, α, takes the value α = 0.964111.
i=1

⎛L L
α α′ α α α′ α′ ⎞
= ⟨Tr ∑ ∑ ∑ ∑ Jij Ji′ j ′ σ̂i σ̂j σ̂i′ σ̂j ′ ⟩J ,
⎝i<j α=x, y, z i′ <j ′ α′ =x, y, z ⎠
(83) 1500 L=2
1250 L=3
L=4
where ⟨⟩J denotes ensemble averaging over Jij α
. Since L=5
1000
α
trace of σ is 0, trace of elements in (83) are 0, except in L=6
the case i = i′ , j = j ′ and α = α′ . Using that the variance 750 L=7
#
α
of Jij is 1, we obtain 500
250
⎛L α 2 ⎞ 3L(L − 1) L
σ 2 2L = ⟨Tr ∑ ∑ (Jij ) 1 ⟩J = 2 . (84) 0
⎝i<j α=x, y, z ⎠ 2 0 1 2 3 4
||[H, V]||
Hence, the eigenvalues of V̂0 , v0,i , follow a Gaussian
distribution having zero mean and variance given by (a)
σ 2 = 3L(L−1)
2
.
As last step, we need to estimate the expectation 2.5
value for the maximum among N numbers randomly ex-
tracted from a Gaussian distribution having vanishing 2.0
mean value and variance equal to σ. It can√be shown,
||[H, V]||

[1], that such a maximum value goes like σ 2 ln N for 1.5 Mean
large N . As a result, using that, in our case, N = 2L the Standard Deviation
∥V̂0 −v0,min ∥
mean value of 2
averaged α
over Jij takes the form 1.0

∥V̂0 − v0,min ∥ v0,max − v0,min 0.5

⟨ ⟩J = ⟨ ⟩J = ⟨v0,max ⟩J
2
√
2 2 3 4 5 6 7
√ 3L(L − 1) √ L
= σ 2 ln(2 ) =
L 2L ln 2 = 3L2 (L − 1) ln 2 (b)
2
(85)
FIG. 4. (a) Number of the cases x + 0.1 > ∥[V̂ − Ĥ]∥ ≥ x for
Motivated by this result, in Fig. 4 we numerically com- each L. Total number of realization is 10000. (b) Mean and
∥V̂ −v standard deviation of ∥[V̂ − Ĥ]∥ for each L. Total number of
puted ⟨ 0 20,min ⟩J for L up to 16 and we have fit-
∥
realization is 10000.
ted
√ the resulting values against an Ansatz of the form
α L2 (L − 1), with α being a fitting parameter. As it
is evident from the figure the agreement is excellent.
Thus, from the analytical estimate, Eq. (85), tested in Now we will explain why in Fig 1. (b) the maximum
Fig. 3, we
√ conclude that the normalization factor scales decreases slightly with growing size of the system L. Also
as C ∝ L2 (L − 1). the quantity ∥[Ĥ, V̂ ]∥ is determined in terms Jij
α
; there-
 13

fore, it also changes in each realization (instance) of the deviation decreases faster with L than the raising of the
simulation. Fig. 4 (a) shows the distribution (histogram) mean value. To conclude, the probability of getting large
of ∥[Ĥ, V̂ ]∥ as a function of L. As L increase, the shape values of ∥[Ĥ, V̂ ]∥ decreases with L. For this reason, the
of distribution becomes sharper and localized around the expectation value of the maximum value of ∥[Ĥ, V̂ ]∥ re-
mean value ∼ 2.5. Since Fig 1. (b) in the main text shows duces with L, if the number of realizations stays fixed for
the maxima from a finite number of 500 realizations, the every L. Since ∥[Ĥ, V̂ ]∥ is also an upper bound on maxi-
maximum drawn from this finite number also decreases. mum charging power Pmax , we expect a similar behavior
Additionally, from Fig. 4 (b) we see that the standard also for this quantity.

[1] G. Kamath, “Bounds on the expectation of the maximum

of samples from a gaussian,” (2015).