Supplementary Information: Critical thermalization of a disordered dipolar spin

system in diamond

Georg Kucsko,∗ Soonwon Choi,∗ Joonhee Choi,∗ Peter C. Maurer, Hengyun Zhou, Renate Landig, Hitoshi
Sumiya, Shinobu Onoda, Junichi Isoya, Fedor Jelezko, Eugene Demler, Norman Y. Yao, and Mikhail D. Lukin
(Dated: January 27, 2018)

CONTENTS

Materials and Methods 1

Sample Fabrication 1
Optical Setup 2
Microwave Setup 2
Magnetic Field Setup 2

Characterization of Experimental System 3

On-site Potential Disorder 3
Estimation of NV Density and Dipolar Interaction Strength 5
Inhomogeneity of the Microwave Field 7

Effective Hamiltonian of a Driven System 7

Resonance Counting Theory 10

Disorder-dependent Power-law Decay 10
Interplay between Dimensionality and long-range Interaction 11
Time-dependent Disorder 12

Detailed Analysis of Thermalization Experiments 13

Effective Disorder Control under Spin-locking Conditions 13
Effects of Incoherent Dynamics 14
Dependence of Thermalization Dynamics on Spin Bath Polarization 15

References 19

MATERIALS AND METHODS

Sample Fabrication

The diamond sample used in this work (type-Ib, ∼4 mm in diameter) was grown via high pressure and high
temperature (HPHT), at 5.5 GPa and 1350 o C, using a Fe-Co alloy as a solvent. The main source of paramagnetic
impurities was provided by substitutional nitrogen atoms in the neutral charge state (P1 centers) at a concentration
of ∼100 ppm. A diamond plate of thickness ∼1 mm was obtained via laser cutting and polishing. To obtain NV
centers, high energy electron irradiation was performed at ∼2 MeV with a flux of 1.3-1.4·1013 e·cm−2 ·s−1 and in-situ
annealing at 700-800 o C up to a total fluence of 1.4·1019 cm−2 (total time of 285 hrs). Additional annealing at
1000 o C for 2 hrs in vacuum was performed after half as well as after the full irradiation time. This process resulted
in the diamond with NV centers of a concentration ∼45 ppm, corresponding to ∼5 nm of average separation and
∼ (2π) 420 kHz dipole-dipole interaction strength. To control the region of optical excitation, we used angle etching
to create a beam-shaped piece of diamond, of 20 µm length and ∼300 nm width, and transferred it onto our coplanar
waveguide [1].
 2

Optical Setup

As shown in Fig. S1A, the optical setup consists of a home-built confocal microscope with a Nikon Plan Fluor
100x oil immersion objective (NA = 1.3). The sample is mounted on a xyz-piezoelectric stage in the focal plane of
the microscope. Excitation of the ensemble of NV centers is performed by illuminating a green laser (λ = 532 nm)
with average power less than 50 µW. Short laser pulses are generated by an acousto-optic modulator (AOM) from
Isomet in a double pass configuration. The λ/2-waveplate at the objective allows the control over the polarization
of excitation light. NV centers emits fluorescence into the phonon sideband (630-800 nm), which is isolated from
the excitation laser by a dichroic mirror. An additional 650 nm long-pass filter further suppresses the detection of
unwanted signal. After passing a pinhole the collection beam is then focused onto a single photon counting avalanche
photodiode (APD) to achieve detections with confocal resolution.
To probe the spin dynamics over time, we used a pulse sequence illustrated in Fig. S1B. We repeat the same
pulse sequence twice, but include an extra π-pulse right before the read-out at end of the second sequence. The
photon-count difference between the two read-outs allows us to measure the NV polarization, while being insensitive
to changes in the background fluorescence due to charge dynamics [2].

Microwave Setup

To coherently control the electronic spin states of NV centers we deliver microwaves to the sample through an
impedance-matched coplanar waveguide fabricated on a glass coverslip. An omega-shaped microstructure (with a
inner diameter 20 µm) at the center of the waveguide allows us to achieve Rabi frequencies up to ∼(2π) 100 MHz. In
Fig. S2, we illustrate the schematic diagram of the microwave control system. In order to have full control over two
groups of NV centers with different transition frequencies, we employ two independent microwave circuits. In each
circuit, a RF signal generator (Rohde & Schwarz SMIQ06B) produces the main driving frequency; an IQ mixer (Marki
IQ1545LMP) generates pulsed signals; a low-pass microwave filter (Mini-Circuits VLF-3000+) suppresses unwanted
higher-order harmonics of fundamental frequencies; and a DC block (Picosecond 5501a) additionally isolates the
signals from low-frequency noises. After separately amplified (ZHL-16W-43+), two RF signals are then combined by
a power combiner (Mini-Circuits ZFRSC-42-S+) and delivered to our sample. The inset of Fig. S2 depicts the detailed
configuration of analog inputs (AI) connected to the IQ mixers. An arbitrary waveform generator (The Tektronix
AWG7052) defines the duration and the phase of the pulses with a temporal resolution of 1 ns. For fine tuning of the
voltage offset on the I and Q ports, a DC voltage is applied to the AWG signal. The addition of a 10-dB attenuator
between the voltage source and the combiner suppresses unwanted reflections (see inset of Fig. S2).

Magnetic Field Setup

For an external magnetic field, we use three water-cooled electromagnetic (EM) coils, which can provide a B-field
up to ∼300 Gauss in an arbitrary orientation (see Fig. S1A and S3A).
As shown in Fig. S3B, we calibrate the magnetic field by recording electron spin resonance (ESR) spectra at various
values of currents in the coils; since the Hamiltonian of a NV center in the presence of a magnetic field is known, the
magnetic field at the position of our sample can be extracted from transitions frequencies of NV centers. For this
process we utilize all four groups {A, B, C, D} of NV centers oriented in different crystallographic axes of diamond
lattice, e.g., A = [111], B = [1̄1̄1], C = [11̄1̄], and D = [1̄11̄] (see Fig. S3C). Fig. S3D shows an ESR spectrum when
the B-field is aligned along [111] direction; group A exhibits the largest Zeeman splitting, while the other groups B,
C, and D become degenerate. In Fig. S3E, the direction of an external B-field is perpendicular to the sample surface,
i.e., B k [001], resulting in four degenerate groups.
 3

A
Laser

AOM
APD

LP 650 Pinhole

DM
PBS

Coplanar waveguide (top view)

λ/2

MW NV sample
EM coil 100x
NA 1.3

B I II III IV
Green
Readout

Microwave A
π-pulse
Microwave B

FIG. S1. Schematic of the Optical Setup and the Pulse Sequence. (A) Green and red lines indicate the optical paths
(excitation: green, collection: red). An acousto-optic modulator (AOM) is used to control green laser duration. A dichroic
mirror (DM) spectrally filters out the fluorescence from NV centers for electronic spin state readout. A 650 nm long pass filter
additionally helps to filter fluorescence emission, corresponding to the phonon sideband (PSB) of NV centers. A 5-µm pinhole
is used in combination with a single photon counting avalanche photodiode (APD) to achieve confocal detection. A polarizing
beam splitter (PBS) is used to polarize the excitation beam. With the addition of a λ/2 waveplate we get control over the
incident green polarization onto the diamond sample. The sample is placed ontop of a coplanar microwave (MW) structure
in the shape of an omega (inset). Three electromagnetic coils are used to create a static magnetic field up to ∼300 Gauss
in an arbitrary direction. (B) Typical experimental sequence used to measure NV dynamics. I: charge equilibration; II: spin
polarization; III: experimental sequence; IV: spin readout.

CHARACTERIZATION OF EXPERIMENTAL SYSTEM

On-site Potential Disorder

The ESR linewidth of an NV ensembles is influenced by multiple factors. To discuss and estimate their contributions
we introduce the ground state hamilitonian of the electronic spin state of a single NV center:
 
~ · B)
~ − d⊥ E⊥ y
H = ~∆0 + d|| E||z Sz2 + γN V (S
 x
Sx2 − Sy2 ,


(Sx Sy + Sy Sx ) + E⊥ (S1)

where Sx , Sy and Sz denote the spin-1 matrices and ~ the reduced Planck constant; ∆0 ≈ (2π)2.87 GHz, γN V =
(2π) 2.8 MHz G−1 , dk = (2π) 0.35 Hz cm V−1 and d⊥ = (2π) 17 Hz cm V−1 are the zero field splitting, the gyromagnetic
 4

AI

AI DC
AWG Combiner 10dB
voltage

Q
Signal
Low-pass
generator IQ Mixer DC block Amplifier IQ Mixer
filter
#1
I

AI

Combiner Sample

AI

Q
Signal
Low-pass
generator IQ Mixer DC block Amplifier
filter
#2
I

AI

FIG. S2. Schematic of the Microwave Control Setup. Two sets of independent microwave circuits are used to achieve full
control over two separate groups of NV centers at different transition frequencies. A 3 GHz low-pass filter suppresses unwanted
higher-order harmonics. The two microwave paths are separately amplified to avoid saturation and then combined and sent to
the diamond sample. In order to precisely control the microwave pulse length as well as phase, each path is sent through an
IQ mixer controlled by an arbitrary waveform generator (AWG) output. The inset shows the detailed configuration of analog
inputs connected to the IQ mixers used to define microwave pulse length and phase. In order to finely tune the voltage offset
of the I and Q port, to achieve high isolation, a DC voltage source is combined with the AWG signal. The addition of an
attenuator allows the suppression of unwanted reflections.

ratio, axial and perpendicular components of the ground triplet state permanent electric dipole moment of a NV center
[3]. Bk(⊥) and Ek(⊥) are projection of the effective magnetic and electric field parallel (perpendicular) to the NV axis.
To a leading order, we ignore the effect of the perpendicular magnetic field noise δB⊥ , since it influence less on the
spin coherence than the parallel one δBk , owing to the large zero field splitting.
To account for effects of the local NV environment we include in B|| and E||(⊥) on-site potential disorders originating
from randomly distributed magnetic fields due to nuclear spins (i.e. 13 C or 14 N) and paramagnetic impurities (i.e.
P1 centers) as well as fields caused by local electric fields and lattice strain. To quantify the different contributions
to the ESR linewidth, we conduct Ramsey spectroscopy in distinct basis states as listed in Fig. S4A. Since each basis
has a well defined sensitivity to different physical noise sources, our Ramsey measurements provide insight into the
local environment of the NV centers. Table SI lists the effects of magnetic and electric field noise on free induction
decay of several different basis states. Figure S4 shows the outcome of Ramsey spectroscopy in the five different bases
defined in SI .

Definition Wavefunction√ Precession rate Noise 1/T2∗

|ψ1 i (|0i + |1i)/ 2 γN V Bk + dk Ek δBk , δEk π[ΓBk + ΓEk ]
√
|ψ2 i (|0i + | − 1i)/ 2 γN V Bk + dk Ek δBk , δEk π[ΓBk + ΓEk ]
√
|ψ3 i (|1i + | − 1i)/ 2 2γN V Bk δBk 2πΓBk
√
|ψ4 i (|0i + |Di)/ 2 dk Ek + d⊥ E⊥ δEk , δE⊥ π[ΓEk + ΓE⊥ ]
√
|ψ5 i (|0i + |Bi)/ 2 dk Ek + d⊥ E⊥ δEk , δE⊥ π[ΓEk + ΓE⊥ ]
√
TABLE SI. Five different basis states√ used for characterizing the local on-site disorder. The dark (|Di ≡ (|1i − | − 1i)/ 2 ) and
bright states ( |Bi ≡ (|1i + | − 1i)/ 2) are prepared by applying an off-axis magnetic field perpendicular to an NV symmetry
axis. Γ is a noise source-dependent inhomogeneous broadening contributing to the linewidth of the ESR.
 5

A B ESR spectra
1.1
EM coil (z-axis) EM Coil
1.08 Current
EM coil
(y-axis)

Contrast [arb units]

EM coil 1.06
(x-axis)
1.04

1.02

MW driving 0.98
2.4 2.6 2.8 3 3.2 3.4
MW Frequency [GHz]

C D E
B // [1,1,1] B // [0,0,1]
A B
N C 1

Contrast [arb units]

1
Contrast [arb units]

V V
C C C N
C C 0.99 A A 0.99

C D
C C 0.98
0.98
V V
N C C C 0.97
C N B,C,D B,C,D 0.97
A,B,C,D A,B,C,D
2.75 2.8 2.85 2.9 2.95 3 2.75 2.8 2.85 2.9 2.95 3
MW Frequency [GHz] #10 9 MW Frequency [GHz] #10 9

FIG. S3. Magnetic Field Calibration and Control. (A) Three electromagnetic (EM) coils are located in the vicinity of
the diamond sample in order to provide an external magnetic field (B-field) in an arbitrary direction with an amplitude up to
∼300 Gauss. (B) To calibrate the coil’s magnetic field, electron spin resonance (ESR) spectra are recorded for different values
of coil currents. (C) The diamond lattice allows for four different crystallographic orientations of NV centers. The different
groups A, B, C, and D of NV centers are characterized by their N-V axis orientations, i.e., A = [111], B = [1̄1̄1], C = [11̄1̄],
and D = [1̄11̄]. (D) Measured ESR spectrum for the B-field aligned along the [111] direction. Group A exhibits the largest
Zeeman splitting (highest projected B|| ) because the spin quantization axis of group A is parallel to the chosen B-field. (E)
Measured ESR spectrum for the B-field aligned along the [001] direction. Due to the [100] cutting direction of the diamond,
all 4 NV groups form the same angle to the surface. With the external B-field being perpendicular to the sample surface, this
leads to groups A-D having degenerate B field projections.

As seen in the table SI, each coherent superposition can effectively probe different types of noise components,
enabling us to quantify the relative strengths of the on-site potential disorder. Using the identity Γ = 1/πT2∗ and the
relations given in the last column of table SI, we can estimate a value for the different noise sources ΓBk , ΓEk , and
ΓE⊥ . The discrepancy in T2∗ between |ψ1 i and |ψ2 i (as well as |ψ4 i and |ψ5 i) in experimental data is presumably due
to frequency-dependent field noise. By averaging these results, we can extract the three inhomogeneous broadening
factors as ΓBk = 3.78(3) MHz, ΓEk = 2.18(8) MHz and ΓE⊥ = 4.30(13) MHz. The measured ESR linewidth Γmeas =
√
8 ln 2W ≈ 9.4 MHz (see Fig. 1D, main text) roughly agrees up to a factor of ∼1.5 with the calculated Γcalc ≈ 6.0
MHz. According to this analysis, the random on-site disorder in our sample seems to result from both electric and
magnetic fields with comparable weights.

Estimation of NV Density and Dipolar Interaction Strength

Due to the high density of NVs within our sample, the spin-echo coherence time is limited by interactions, as
discussed in the main text. In particular, using the double electron-electron resonance (DEER) sequence presented in
Fig. 2A in the main text, we verified experimentally that the additional dephasing of group A indeed originates from
interactions with group B. Fig. S5 shows a measurement result of the DEER sequence in which we probe the relative
spin-echo amplitude at a fixed time τ as a function of driving frequency of group B. It shows a clear resonance when
ω = ω0B , indicating that inter-group interactions between group A and B lead to enhanced dephasing.
 6

ms = +1 ms = +1
Bright
Dark
ms = -1 ms = -1

ms = 0 ms = 0 ms = 0 ms = 0

B *
T2 dephasing
1
|A1> T* (Ramsey) T (Spin-e
|A > 2 2
0.8 2 60
|A3> 1500
norm. coherence

|A >
4
0.6

Time [ns]

Time [ns]
|A5> 50
1000
0.4
40 500
0.2
30 0
0
|A 1> |A 2> |A 3> |A 4> |A 5> |A 1> |A 2> |A 3>
0 20 40 60 80 100
time [ns]

FIG. S4. T2∗ and T2 Measurements of different Basis States. (A) Different initial conditions used for coherence
measurements. By aligning the magnetic field parallel (perpendicular) to the NV axis, the eigenbasis for the spin state of NV
centers becomes {|ms = 0i, |ms = +1i, |ms = −1i} ({|ms = 0i, |Darki, |Brighti}), where Bright and Dark states are defined
as even and odd combination of the original bare spin states |ms = −1i and |ms = +1i. (B) Ramsey spectroscopy data and
extracted decay timescale for different initial states.
normalized coherence

1
0.9
0.8
0.7

2.6 2.65 2.7

f (GHz)

FIG. S5. Intra-group Interaction Probed via Double Electron-Electron Resonance. The relative, normalized spin
echo coherence time at a fixed time τ as a function of driving frequency of group B.

To quantitatively analyze the dependence of decoherence rate on the spin density, we study the dynamics of
interacting spins using the exact diagonalization method with the effective Hamiltonian of Eq. (S24). Comparing the
numerical result to the experimental data allows us to extract the density of NV spins in our sample.
Specifically, we simulate the time evolution of 12 NV spins under a spin echo pulse sequence protocol. The total NV
concentrations selected for simulations are 5, 20, 40, 60, 70, 80 and 100 ppm. We averaged over ∼500 realizations of
positional disorder, resulting in a single smooth coherence curve under the spin echo sequence. We fit the coherence
decay with a simple exponential function and extract the decoherence rate, γT ≡ 1/T2 . Fig. S6A summarizes the
 7

A B
T simulation NV density extraction
2
3
X simulation 0.7
2.5 data
0.6

2 0.5
1/T [MHz]

slope
0.4
1.5
2

0.3
1
0.2
0.5
0.1

0 0
1 2 3 4 0 50 100
# of resonant groups NV density [ppm]

FIG. S6. NV Density Extraction via Spin Echo Simulation. (A) Comparison of the spin echo simulation results at
different concentrations (crosses) to the measured data (circles). The total NV concentrations selected for the simulation are
5, 20, 40, 60, 70, 80 and 100 ppm. Solid lines are linear fits to the simulation to extract both γb and γ0 in the main text. (B)
The NV concentration can be extracted by comparing the slopes (γ0 ) taken from the numerical simulations to the extracted
slope of the experiment data (orange dashed line).

spin echo simulation results as a function of the number of resonant NV groups (effective density), where a linear
dependence of γT is identified for all the density values. We model the decoherence rate as γT (ν) = γb (ν) + νγ0 (ν),
where ν is the number of resonant NV groups, γb and γ0 are density-dependent, bare and dipolar interaction-induced
dephasing rates, respectively. Such linear dependence of γT on ν is also confirmed in the experiment (see Fig. 2B in
main text). By comparing γ0 between the experiment and the simulation, we estimate the NV density in our sample
to be ∼45 ppm (see Fig. S6B).

Inhomogeneity of the Microwave Field

Hartmann-Hahn resonances rely on the exact matching of Rabi frequencies of two driving fields ΩA = ΩB . Hence,
stable and precise control of the driving strength is essential in our experiments. To this end, we estimate the
inhomogeneity of our microwave driving field, by measuring the decay time of Rabi oscillations at various driving
strengths (Fig. S7).
In an ideal case, the lifetime of Rabi oscialltions generally increases due to suppression of disorder (T∗2 ). At higher
driving strength (Weff ∼ δ I ) this lifetime should saturate due to the effect of Ising interaction. In our measurements
however we observe a slight decrease in lifetime at high driving strengths, which is well explained by a 1.1% variation
in Rabi frequency. We attribute this variation to spatial inhomogeneity in the driving field. With the strongest driving
in our measurement Ω = (2π) 32 MHz, this effect leads to a spread in Rabi frequencies of ∼ (2π) 0.3 kHz. While
it is still smaller than the effective disorder ∼ (2π) 0.6 kHz, such an inhomogeniety ultimately limits the maximum
driving strength of our thermalization experiments.

EFFECTIVE HAMILTONIAN OF A DRIVEN SYSTEM

In this section, we derive the effective Hamiltonian for a driven, dipolar interacting spin ensemble. The main idea
is to work in a frame that is rotating along each NV group’s quantization axis at corresponding driving frequency (ω0A
and ω0B for group A and B, respectively). If the difference between ω0A and ω0B is large compared to the interaction
strength, then one can ignore exchange interactions between spins from different groups (secular approximation). This
 8

800

Decay time (ns)

600

400

200

0
0 10 20 30 40
+/2: (MHz)

FIG. S7. Rabi Oscillation Measurement. Decay time of Rabi oscillations as a function of Rabi frequency Ω.

results in distinct forms of intra- and inter-group interactions. We project the original Hamiltonian into two-level
systems, and derive the effective Hamiltonian.
We start with the Hamiltonian for dipolar interacting NV centers
X X X
H= Hi0 + Hid (t) + dd
Hij , (S2)
i i ij

where Hi0 is a single particle Hamiltonian for a spin at site i, Hid (t) is time-dependent driving, and Hij
dd
is the magnetic
0
dipole-dipole interaction between spins at sites i and j. The first term Hi includes Zeeman coupling to an external
magnetic field, the zero field splitting of a NV center, and any other disordered potentials arising from couplings to
paramagnetic impurities as described in the main text. In our experiments, dominant contributions for Hi0 come from
the zero-field splitting ∼ (2π) 2.87GHz and Zeeman field projected along the quantization axis (a few hundred MHz),
which are two orders of magnitude larger than the rest of the couplings. Setting ~ = 1, we can write
 2  
~i + (∆B (ĉi ) + δB,i ) ĉi · S
Hi0 ≈ (∆0 + δ0,i ) ĉi · S ~i (S3)

where S ~i are spin-1 vector operators, ĉi is the unit vector along the quantization axis of the spin, ∆0 = (2π) 2.87 GHz
is the zero-field splitting, ∆B (ĉi ) is the Zeeman splitting along ĉi , and δ0,i and δB,i are on-site disorder potentials. If
the external magnetic field B ~ is oriented in a way that ∆B (ĉi ) for different groups are sufficiently separated (com-
pared to the driving strength), one can effectively address distinct groups independently. Below we assume such a
A(B)
case and consider resonant driving of two groups A and B using microwave frequencies ω0 = ∆0 − ∆B (ĉA(B) ).
The Hamiltonian for such driving is given as Hid (t) = γN V B ~ MW · S
~i cos (ω0 t), where γN V is the gyromagnetic ratio of
the NV center, andhB ~ MW is the microwave field vector. Now moving into the rotating frame with unitary transfor-
P  i
mation U (t) = exp −i ∆ 0 (ĉ i · ~i )2 + ∆B (ĉi )(ĉi · S
S ~i ) t and applying rotating wave approximations, we obtain
i
the effective single particle Hamiltonian
d
H̄i = U † (t) Hi0 + Hid (t) U (t) − iU † U
 
(S4)
dt
Ω
= (δ0,i + δB,i ) |1i h1| + (δ0,i − δB,i ) |−1i h−1| + (|−1i h0| + h.c.) , (S5)
2
where {|1i , |0i , |−1i} is the basis of spin states along its quantization axis and Ω is the Rabi frequency of the driving.
The effective interaction among spins can be obtained in a similar way as follows. We start with the dipole-dipole
interaction between spin-i and spin-j

dd J0   ~ 
~

~ ~

Hij =− 3 S i · r̂ S j · r̂ − S i · S j , (S6)
r3
 9

where J0 = (2π) 52 MHz · nm3 and ~r is the relative position between two spins. In the rotating frame, we obtain the
effective interaction by replacing S ~i 7→ U † (t)S
~i U (t). Since we are interested in the interaction in the basis of each
NV’s own quantization axis, we first explicitly rewrite S ~i in terms of (S x , S y , S z ) in a coordinate system where ẑi is
i i i
parallel to the quantization axis ĉi
h
dd
= U † (t)Hij
dd
U (t) = −J0 /r3 3 (r̂ · x̂i ) (r̂ · x̂j ) − x̂i · x̂j Six Sjx


H̄ij (S7)

y y
+ 3 (r̂ · ŷi ) (r̂ · ŷj ) − ŷi · ŷj Si Sj (S8)

x y
+ 3 (r̂ · x̂i ) (r̂ · ŷj ) − x̂i · ŷj Sj Sj (S9)

y x
+ 3 (r̂ · ŷi ) (r̂ · x̂j ) − ŷi · x̂j Si Sj (S10)

z zi
+ 3 (r̂ · ẑi ) (r̂ · ẑj ) − ẑi · ẑj Si Sj (S11)
+Hrest , (S12)

where Hrest contains all the other terms of the form S x S z , S y S z , S z S x , S z S y .

We now perform rotating wave approximations. This is very well justified because the typical strength of the
interaction is much weaker than the driving frequency J0 /r3 ∼ (2π) 0.4 MHz  ω0A,B ∼ (2π) 2.5 GHz. First, we note
that S x and S y operators are rapidly oscillating in time while S z remains invariant, [Siz , U (t)] = 0. Therefore, every
term in Hrest may be safely ignored. Then, introducing

+ 1h i
gij = 3 (r̂ · x̂i ) (r̂ · x̂j ) − x̂i · x̂j + 3 (r̂ · ŷi ) (r̂ · ŷj ) − ŷi · ŷj (S13)
2
− 1 h i
gij = 3 (r̂ · x̂i ) (r̂ · x̂j ) − x̂i · x̂j − 3 (r̂ · ŷi ) (r̂ · ŷj ) + ŷi · ŷj (S14)
2
1 h i
h+ij = 3 (r̂ · x̂i ) (r̂ · ŷj ) − x̂i · ŷj + 3 (r̂ · ŷi ) (r̂ · x̂j ) − ŷi · x̂j (S15)
2
1 h i
h−ij = 3 (r̂ · x̂i ) (r̂ · ŷj ) − x̂i · ŷj − 3 (r̂ · ŷi ) (r̂ · x̂j ) + ŷi · x̂j (S16)
2
qij =3 (r̂ · ẑi ) (r̂ · ẑj ) − ẑi · ẑj , (S17)

we can simply rewrite

(Si Sj + Siy Sjy ) + h− x y y x

dd
 + x x
H̄ij ≈ −J0 /r3 gij z z
ij (Si Sj − Si Sj ) + qij Si Sj (S18)
+ gij−
(Six Sjx − Siy Sjy ) + h+ x y y x 
ij (Si Sj + Si Sj ) . (S19)

Here, g + and h− terms correspond to “flip-flop” type transitions, exchanging one unit of spin polarization,

(Six Sjx + Siy Sjy ) = |+0i h0+| + |+−i h00| + |00i h−+| + |0−i h−0| + h.c. (S20)
(Six Sjy − Siy Sjx ) =i |+0i h0+| + |+−i h00| + |00i h−+| + |0−i h−0| + h.c.


(S21)

In addition, owing to the strong anharmonic level structure, we may also ignore flip-flop transitions between levels with
large energy differences, e.g. terms such as |+−i h00|. Finally, we ignore the terms in Eq. (S19) as they correspond to
double flip-up or flip-down and rapidly oscillate in time. After these approximations, the effective interaction becomes

+ ih−
dd
 +
≈ −J0 /r3 gij z z


H̄ij ij |+0i h0+| + |0−i h−0| + h.c. + qij Si Sj . (S22)

Now we divide into two cases depending on whether spins i and j belong to the same group or to different groups. In the
former case, the quantization axes coincide, and we can simplify h− + 1 2 2
ij = 0, gij = 2 (1−3 cos θ), and qij = −(1−3 cos θ)
with cos θ ≡ ẑ · r̂. In the latter case, the flip-flop terms are again rapidly oscillating, and only the Ising interaction
Siz Sjz remains, resulting in
(  
− 0r3ij − |+0ih0+|+|0−ih−0|+h.c.
J q
dd 2 + Siz Sjz same group
H̄ij ≈ J0 qij z z
. (S23)
− r3 Si Sj different groups

These interactions as well as the single particle terms conserve the total population of spins in |+i. Therefore, once
the system is initialized into a state with no population in |+i, the dynamics remains in the manifold spanned by |−i
 10

dd
P P
and |0i. Projecting i H̄i + ij H̄ij into this manifold, we obtain the Hamiltonian for an effective two-level system.
Introducing spin-1/2 operators ~s for two levels |−i and |0i, we obtain HT = HA + HB + HAB , where
X X J0 qij x x
[(δ0,i − δB,i )szi + ΩA(B) sxi ] + si sj + syi syj − szi szj ,


HA(B) = 3 (S24)
rij
i∈A(B) i,j∈A(B)
X J0 qij zA zB
HAB = − 3 si sj , (S25)
rij
i∈A,j∈B

up to a constant.
Finally, we remark one particularly interesting aspect of this Hamiltonian in the dressed-state basis, i.e., quantization
along sxi . With sufficiently strong driving, sxi becomes a good spin polarization basis, and one can rewrite the
− −
interactions in terms of s± = sy ± isz , wherein the intra-group interaction becomes ∝ sxi sxj + (s+ +
i sj + si sj )/2 and
+ − + + + −
the inter-group interaction ∝ (si sj + si sj + h.c.). Here, we find that spin exchange terms (si sj + h.c.) are missing
in the intra-group interaction. Omitting the energy non-conserving terms such as s+ +
i sj (secular approximation with
a strong driving strength Ω), we obtain the effective Hamiltonian described in the main text.

RESONANCE COUNTING THEORY

In this section, we provide a detailed study of the single particle resonance counting theory. We will first focus on
the case of quenched on-site potential disorder, deriving the disorder-dependent power-law relaxation presented in the
main text. Then, we generalize the result to the case when disordered potentials are time-dependent.

Disorder-dependent Power-law Decay

As discussed in the main text, we estimate the survival probability of a single spin excitation based on a simple
counting argument. At time t, we compute the probability Pr(k; t) that the central spin is connected to k − 1 other
spins via a network of resonances, as defined in the main text. Assuming that the population of the excitation is
equally shared among a resonating cluster, the survival probability is given as
∞
X 1
P (t) ≈ Pr(k; t). (S26)
k
k=1

reducing our problem to the computation of Pr(k; t). Below we will show that the dominant contributions arise from
k = 1, suggesting that finding a single resonant partner is usually enough to delocalize the spin excitation over the
entire sample.
In general, the exact calculation of Pr(k; t) is difficult. This is because the connectivity of the resonance network is
correlated due to the spatial structure (d-dimensional Euclidean space) as well as a given assignment of random on-site
potentials, e.g., if spin pairs (a, b) and (b, c) are pair-wise resonant, it is likely that the pair (a, c) is also resonant, etc.
However, the qualitative behavior of Pr(k; t) can still be well-understood by ignoring these correlations. In such a
case, we may assume that the number of resonant partners ` for a spins is drawn from a probability distribution p(`)
and that this process can be iterated for each partner. We note that such a process may not terminate, in which case
the central excitation becomes delocalized over a macroscopic number of spins. We first compute p(`) as a function
of time t. For ` = 0, a spin of interest (spin-i) must not have any resonating spins at any distance from rmin to
R(t) ≡ (J0 t)1/3 , where rmin is the short-distance cut-off. Hence, p(0; t) is given as a product of probabilities:
βJ0 /r3
Y  
2
p(0; t) = 1 − 4πnr dr (S27)
Weff
rmin ≤r<R(t)
" Z #
R(t)
4πnQres
= exp − dr (S28)
rmin r

where 4πnr2 dr is the probability of finding a spin at distance r, and Qres = βJ0 /(Weff r3 ) is the probability that the
spin resonates with the spin-i. Defining λ(t) = 4πQres (ln R(t) − ln rmin ), we obtain p(0; t) = exp [−λ(t)]. Similarly, we
`
can calculate p(`; t) for ` > 0, and obtain p(`; t) = `!1 (λ(t)) e−λ(t) , which is the Poisson distribution with mean λ(t).
 11

To show that the dominant contribution of Eq. (S26) arises from the k = 1 term, we consider the probability of the
termination of the resonance finding process, Pterm . It satisfies the self-consistency equation
∞
X λ` e−λ `
Pterm = e−λ + (Pterm ) , (S29)
`!
`=1

where the first term corresponds to the case where the initial spin does not have any resonance up to time t, while
the second term implies the termination of each sub-graph generated from ` resonant spins. For sufficiently large λ,
Pterm becomes small, and its contribution is dominated by the first term (` = 0). In our case, λ(t) is a function of
time which diverges in the limit t → ∞. As we are interested in the late time dynamics, we may consider the first
term only. In terms of Pr(k; t), this corresponds to approximating Pr(k; t) ∼ 0 for k > 2. Finally, noting that that
Pr(k = 1; t) = p(0; t), we recover the expression in the main text.

1
0.8

0.6
P (t)

0.4

10 0 10 1 10 2
Time (μs)

FIG. S8. Single-particle simulation of power-law dynamics. Blue, red, and yellow curve correspond to Ω = (2π) 3, 8,
and 20 MHz, respectively. For the simulations, we use 104 spins and average over more than 103 disorder realizations.

We numerically test the analytic resonance counting that predicts the power-law decay dynamics. In the limit of
single-particle excitation, the survival probability P (t) = |hψ(t)|ψ(0)i|2 can be computed at any time t after the time
evolution of a system under Heff (See Eq. (1) in the main text). Considering physically relevant parameters used in the
experiments, we verify such power-law decay dynamics for up to 104 spins as shown in Fig. S8. Moreover, we confirm
the extracted power-law exponent is inversely proportional to effective disorder Weff (Fig. 4C in the main text), further
substantiating the thermalization mechanism based on rare resonances. The power-law exponents extracted from the
simulations are summarized in Fig. S12A.

Interplay between Dimensionality and long-range Interaction

The critical nature of a disordered dipolar spin ensemble in three dimensions originates from the interplay between
long-range interactions and dimensionality. To see this, we can generalize the resonance counting analysis for a
situation in which a single particle excitation is located in a d-dimensional spin system with long-range coupling
decaying as 1/rα . In such a setting, the survival probability P (t) can be expressed as,
" Z #
R(t) α
βJ0 /r
P (t) = exp − nSd rd−1 dr (S30)
rmin Weff
" #
nSd βJ0 R(t) d−α−1
Z
= exp − r dr , (S31)
Weff rmin

R R(t)
where Sd is the surface area of the d-dimensional volume. In fact, the argument of P (t), rmin rd−α−1 dr, is associated
with the probability of finding a resonance up to the distance R(t) reachable at time t. Hence, when the dimensionality
 12

d is larger (smaller) than the interaction strength α, the above integral diverges (converges) as R(t) becomes large,
which implies delocalization (localization) of the single particle excitation. In the critical case where d is equal to
α, the resonance probability increases at a slow logarithmic rate, resulting in the power-law relaxation of the initial
spin state as derived in Eq. (2) in the main text. In the limit of single particle excitations we therefore associate our
system dynamics to such criticality behavior. However, due to the presence of many-spin excitations, much richer
dynamics may appear at longer times. We attribute the deviation of power-law dynamics at late times observed in
our experiments to this effect.

Time-dependent Disorder

Now we consider the case of time-dependent disorder. For concreteness, we assume that the on-site potential
disorder is given as a sum of a static and a dynamical disorder potential, δ̃i (t) = δ̃is + δ̃id (t), where the static part δ̃is
(dynamical part δ̃id (t)) is random with zero mean and standard deviation Ws (Wd ). While δ̃is is time-independent,
the dynamical component δ̃id (t) changes over time by uncorrelated jumps at a rate Γ = 1/τd . Here, we focus on an
experimentally relevant regime where Ws  Wd & nJ0 > 1/τd .
As already mentioned in the main text, we modify the resonance criteria as follows. Two spins at sites i and j are on
resonance at time t if: (1) at any point in time t0 < t, their energy mismatch is smaller than their dipolar interaction
strength, |δ̃i (t0 ) − δ̃j (t0 )| < βJij /rij
3 3
, and (2) the interaction occurs within the time-scale t, Jij /rij > 1/t. While the
second part of the condition is unchanged, the first part now captures that a pair may be brought into resonance
by spectral jumps. Under the hierarchy of Ws  Wd & nJ0 > 1/τd , the condition (1) can be approximated by two
independent events: (a) the static energy mismatch is small enough, |δ̃is − δ̃js | < Wd , and (b) the dynamical energy
mismatch is smaller than the coupling strength, |δ̃id (t0 ) − δ̃jd (t0 )| < βJij /rij3
at some time t0 < t. In combination, the
condition (1) is satisfied with the probability

βJ0 /r3
  
Wd βJ /r 3 t
− 0
Pres (r, t) ≈ 1 − e Wd τd 1 − (S32)
Ws Wd

which is the product of probabilities for conditions (a) and (b). For the second factor, we used the probability that
3
the initial configuration is off-resonant, (1 − βJW0 /r
d
), and the probability that none of the subsequent spectral jumps
βJ0 /r 3 t
−
brings them into resonance e Wd τd . We note that, in practice, one should use max(0, 1 − βJ0 /Wd r3 ) instead of
(1 − βJ0 /Wd r3 ) since a probability cannot be less than zero. Finally, the survival probability is obtained by requiring
no resonance at every distance r up to R(t) = (J0 t)1/3
" Z #
R(t)
2
P (t) = exp − 4πnr Pres (r, t)dr , (S33)
r=r0

where r0 is the short distance cut-off of the NV separations. We use the cut-off distance r0 ∼ 1.4 nm, at which
the corresponding dipole-dipole interaction is J0 /r03 ∼ (2π) 20 MHz. Due to dipole blockade, a pair of NV centers
closer than r0 cannot be addressed by microwave driving of Rabi frequency Ω ∼ (2π) 20 MHz, which we use for
initial preparations of spin states. Those spins do not participate in the spin exchange dynamics due to large energy
mismatch. We note that, limΓ→0 Pres (r, t) → Qres (r) = βJ0 /(Ws r3 ) and the Eq. (S33) correctly reduces to the
disorder-dependent power-law decay. In the presence of a small but finite Γ = 1/τd , integrating Eq. (S33) using
Eq. (S32) yields,

P (t) = P1 (t)P0 (t), (S34)

where
 o
4πn Wd n − W βτ − tt W βτ
P1 (t) = exp − J0 (t − t0 ) − J0 (te d d − t0 e 0 d d ) (S35)
3 Ws
  
4πnJ0 β β t β
P0 (t) = exp − (1 + t/τd )G[0, ] − (1 + t/τd )G[0, ] . (S36)
3Ws Wd τd t0 Wd τd
 13

Here G is an incomplete Gamma function. In the limit of the hierarchy Ws  Wd & nJ0 > 1/τd , we can simplify:
 
4πn Wd  − β
P1 (t) ≈ C1 exp − J0 t(1 − e Wd τd ) (S37)
3 Ws
 
4πnJ0 β t
≈ C1 exp − (S38)
3Ws τd
≡ C1 exp[−t/T ∗ ], (S39)

where C1 is a time-independent prefactor and

3Ws τd Ws τd
T∗ = ∝ . (S40)
4πnJ0 β nJ0

Similarly, P0 (t) can also be simplified as,

  
4πnJ0 β β t β
P0 (t) ≈ C2 exp − G[0, ] − G[0, ] (S41)
3Ws Wd τd t0 Wd τd
 
4πnJ0 β
≈ C20 exp − ln(t/t0 ) (S42)
3Ws
4πnJ0 β
= C20 (t/t0 )− 3Ws . (S43)

Here we used the approximation G(0, z) ≈ − ln(z) + γ + O(z) for z  1. Once again, we rediscover the power-law
decay (Eq. (S43)) predicted in the main text, but now only up to a finite time T ∗ :
∗ 4πnJ0 β
P (t) = P1 (t)P0 (t) ∝ e−t/T t− 3Ws . (S44)

Therefore, according to the Eq. (S44), the weak time-dependent disorder results in a multiplicative exponential cor-
rection to the power-law decay up to t < T ∗ , beyond which the thermalization accelerates substantially. Furthermore,
our theory model predicts that T ∗ is linearly proportional to the static disorder strength Ws , which is consistent with
our observations (See Fig. 3D and Fig. 4D in the main text).

DETAILED ANALYSIS OF THERMALIZATION EXPERIMENTS

Effective Disorder Control under Spin-locking Conditions

To investigate the interplay between disorder and interaction experimentally, it is required to tune both disorder
and interaction in a controlled way. In our experiments, we rely on a spin-locking technique in which both the energy
spacing and the on-site disorder of a spin ensemble can be controlled in a continuous fashion.
As discussed in the main text, spin-locking allows us to
√ prepare spins in the dressed state basis. In the new basis,
the energy eigenstates
√ are |±i ≈ (|m s = 0i±|m s = −1i)/ 2 and are split by an effective Rabi frequency of a spin-lock
field, Ωeff = Ω2 + δ 2 , where Ω is the driving strength and δ is the on-site disorder in the bare frame. Owing to a
random distribution of δ, the new level spacing Ωeff is also a random variable. Therefore, an effective disorder under
the spin-locking condition can be defined as,
p q p
Weff ≡ Var[Ωeff ] = E[Ω2 + δ 2 ] − E[ Ω2 + δ 2 ]2 , (S45)

where Var[X] and E[X] are the variance and expectation value of a random variable X. Since the disorder in the bare
frame follows a Gaussian distribution with a standard deviation W , the expectation values can be expressed as
Z +∞
1 2
/2W 2
E[Ω2 + δ 2 ] = √ dδ [Ω2 + δ 2 ]e−δ (S46)
2πW 2 −∞
Z +∞
p 1 p 2
/2W 2
E[ Ω2 + δ 2 ] = √ dδ Ω2 + δ 2 e−δ . (S47)
2πW 2 −∞
 14
p
In the case of weak driving (Ω  δ), Weff ≈ Var[δ] = W ; namely, the effective
q disorder is almost equal to that in
2
δ2 W
the bare frame. However, as the driving strength Ω increases we find Weff ≈ Var[ 2Ω ] = √2Ω . Hence, the effective
disorder Weff can be tuned by adjusting the Rabi frequency Ω in the dressed state basis.
We note that the probability distribution of Ωeff is highly asymmetric, which may lead to small corrections to
our counting argument at a quantitative level. To this end, for our numerical computations, we use an alternative
definition of Weff which is consistent with our resonance counting argument. Recall that two spins at site i and j with
separation r are defined to be on resonance when |δ˜i − δ˜j | < βJ0 /r3 and that we assumed this occurs with probability
Qres ∝ (J0 /r3 )/Weff . Therefore, the effective disorder strength Weff should be defined in the same way from the full
distribution of Ωeff . More specifically, we compute the probability q(ξ) that two √ independent random variables δ̃i
and δ̃i satisfy |δ̃i − δ̃j | < ξ for a small parameter ξ. In the limit of ξ  W 2 / 2Ω, the probability q(ξ) is linearly
proportional to ξ. Then, we define the effective disorder as Weff ≡ limξ→0 ξ/q(ξ). Fig. S9 shows the dependence of
Weff as a function of Ω. In the limit of large Ω, the effective disorder scales as Weff ∝ 1/Ω, as expected.

2.5

2
Weff / 2π (MHz)

1.5

0.5

0
0 10 20 30 40
Ω / 2π (MHz)

FIG. S9. Effective disorder under spin locking conditions. Based on the resonance counting argument, the effective
disorder Weff can be computed as a function of the Rabi frequency Ω.

Effects of Incoherent Dynamics

In our Hartmann-Hahn experiments, the spin dynamics are governed by both coherent cross-relaxation and inco-
herent depolarization. These two effects have qualitatively different dependence on the driving strength and can be
clearly distinguished in our observations. To perform a detailed analysis of the results presented in the main text, we
focus on the coherent dynamics by normalizing our data at the Hartmann-Hahn resonance ΩA = ΩB via a sufficiently
detuned case |ΩA − ΩB |  nJ0 , at which the spin relaxations are dominated by incoherent dynamics (Fig. S10, blue
line). Such normalization can be justified only if the two effects are independent and multiplicative. This is the
case if the incoherent dynamics are induced by an independent Markovian noise, which results√in an exponential and
multiplicative factor e−γt . In our experiment, however, we observe a stretched exponential e− t/T decay profile from
incoherent dynamics (Fig. 2D in the main text). Below, we explain why such incoherent decays are still factorizable.
Our incoherent dynamics can be modeled as follows. (See Ref. [2] for more details). Each spin at site i undergoes
incoherent depolarization at rate γi . This rate γi is determined by the microscopic local environment of the spin and
follows a random distribution ρ(γ; T1ρ ), such that the ensemble averaged polarization decays as a stretched exponential
√ Z ∞
− t/T1ρ
e = ρ(γ; T1ρ )e−γt dγ. (S48)
0

The analytical expression as well as the microscopic origin of the distribution ρ(γ; T1ρ ) are presented in Ref. [2]. At the
Hartmann-Hahn condition, both the incoherent process and the coherent cross-relaxation lead to depolarization (see
 15

+ = (2: ) 5 MHz + = (2: ) 20 MHz

1 1
at the HH resonance at the HH resonance
at a detuned condition at a detuned condition
0.8 0.8
population of group A

population of group A
0.6 0.6
P(t) = exp[-(t/T;1 )0.5 ] P(t) = exp[-(t/T;1 )0.5 ]
0.4 0.4

0.2 0.2

0 0
0 20 40 60 80 100 0 20 40 60 80 100
time (7s) time (7s)

FIG. S10. Unnormalized experimental Data. Two data sets with different common Rabi frequencies of Ω = (2π) [5,
20] MHz are presented at the Hartmann-Hahn resonance (red) and at the far-detuned case (blue). For the detuned signal, a
stretched exponential of power 0.5 is fitted to the data.

Fig. S10). Hence, at time t, the rate of depolarization for spin-i is given by ṗi (t) = −[γi + fi (t)]pi (t), where fi (t) is the
rate of cross-relaxation (which generally depends on the state of other spins). This cross-relaxation, once averaged
over an ensemble, leads to a power-law decay as derived in the R t previous section. The differential equation for the
0 0
polarization is exactly solvable with the solution pi (t) = e−γi t e 0 fi (t )dt , where one finds a multiplicative exponential
factor e−γi t . Crucially, this effect is still factorizable, even after ensemble averaging:
Z ∞ D Rt 0 0
E √ ρ
hpi (t)iensemble = ρ(γ; T1ρ )e−γi t dγ e 0 fi (t )dt ∝ e− t/T1 · t−η , (S49)
0 ensemble

where η is the disorder dependent exponent derived in the main text. Physically, this factorization arises because
the microscopic environment for each spin, which determines coherent as well as incoherent dynamics, is random
and independent. For this reason, in the experiment, we normalize the polarization decay at the Hartmann-Hahn
resonance (Fig. S10, red line) by the incoherent decay at the far-detuned case (Fig. S10, blue line).

Dependence of Thermalization Dynamics on Spin Bath Polarization

In Fig. S11, the theory prediction from Eq. (S33) is compared with experimental data for various Rabi frequencies
and two different initial polarizations of group B spins. The functional profiles of the decay are consistent with power
laws for over a decade, followed by accelerated, though still sub-diffusive relaxation at late times. In the power-law
regime, we find that the power-law decay exponents depend on the initial polarization of group B spins (Fig. S12A).
This is consistent with our theory; for single-particle dynamics, we expect that the power-law exponent scales as
∼ nJ/Weff , where n is the density of oppositely polarized spins. Indeed, when group B is initially unpolarized, the
exponents are decreased by a factor of two compared to the fully polarized case, consistent with our theory at a
quantitative level.
To characterize the late-time acceleration of the polarization decay, we use the time-dependent model where the
pair-resonance counting criteria are modified as discussed in the previous section. By fitting the experimental data to
our model using a Monte-Carlo (MC) optimization, we extract the parameters of the dynamical disorder strength Wd
and spectral diffusion time τd = 1/Γ. Here, we assume Wd as a global fit parameter which is independent from Ω; this
is because we expect Wd to be predominantly determined by the mean-field interaction strength. In contrast, τd may
in principle be dependent on Ω since the fluctuations of the Ising mean-field potential depend on the thermalization
speed and hence also on the effective disorder strength tuned by Ω. To this end, we performed two independent MC
optimizations where we (i) treat Wd as global (Ω-independent) and τd as local (Ω-dependent) variables (Fig. S11A),
or instead (ii) fix both parameters as global variables (Fig. S11B). For the static effective disorder, we use the theory-
predicted values Ws ∼ W 2 /Ω (as described previously). Naturally, owing to the larger number of fit parameters, a
global Wd with local τd variation (Fig. S11A) shows better agreement than a global Wd together with with global τd
(Fig. S11B). In the latter case, extracted fit parameters Wd and τd are (2π) 46 ± 14 kHz and 43 ± 9 µs, respectively.
 16

In the former case (Fig. S11A), the extracted dynamical disorder Wd ∼ (2π) 0.5 MHz is consistent with the expected
strength of the Ising interaction, suggesting that spin-spin interactions play an important role for the time-dependent
disorder. Furthermore, τd is also consistent with the observed NV depolarization timescale, including contributions
from both coherent cross-relaxation and incoherent spin depolarization. We note that in the fully polarized case the
extracted values for τd are smaller than those in the unpolarized case (Fig. S12B). We speculate that this could be
due to faster coherent spin-exchange dynamics in the former case, giving rise to a faster fluctuation in δ I , responsible
for the accelerated thermalization dynamics at late times.
 17

A 1
19.5 MHz
1
15.6 MHz
1
12.5 MHz

0.8 0.8 0.8

Polarization

Polarization

Polarization
0.6 0.6 0.6

0.4 0.4 0.4

0 1 2 0 1 2 0 1 2
10 10 10 10 10 10 10 10 10
Time [7s] Time [7s] Time [7s]
8.8 MHz 6.7 MHz 4 MHz
1 1 1
0.8 0.8 0.8
Polarization

Polarization

Polarization
0.6 0.6 0.6

0.4 0.4 0.4

100 101 102 100 101 102 100 101 102

Time [7s] Time [7s] Time [7s]

B 1
19.5 MHz
1
15.6 MHz
1
12.5 MHz

0.8 0.8 0.8

Polarization

Polarization

Polarization
0.6 0.6 0.6

0.4 0.4 0.4

100 101 102 100 101 102 100 101 102

Time [7s] Time [7s] Time [7s]
8.8 MHz 6.7 MHz 4 MHz
1 1 1
0.8 0.8 0.8
Polarization

Polarization

Polarization
0.6 0.6 0.6

0.4 0.4 0.4

100 101 102 100 101 102 100 101 102

Time [7s] Time [7s] Time [7s]

C 1
20 MHz
1
16 MHz
1
12 MHz

0.8 0.8 0.8

Polarization

Polarization

Polarization

0.6 0.6 0.6

0.4 0.4 0.4

100 101 102 100 101 102 100 101 102

Time [7s] Time [7s] Time [7s]
8 MHz 5 MHz 3 MHz
1 1 1
0.8 0.8 0.8
Polarization

Polarization

Polarization

0.6 0.6 0.6

0.4 0.4 0.4

100 101 102 100 101 102 100 101 102

Time [7s] Time [7s] Time [7s]

FIG. S11. Polarization decay of a NV ensemble under Hartmann-Hahn conditions. An initially polarized group A
spin ensemble interacts with (A,B) unpolarized and (C) fully polarized group B. Solid lines are theoretical fits based upon
a time-dependent disorder model with extracted parameters (Wd , τd ) via a Monte-Carlo optimization. In (A,C), the spectral
diffusion time τd is dependent on the Rabi frequency, while in (B) τd is independent of the applied Rabi frequency. The
dynamical disorder strength Wd is a Ω-independent, global variable in all three cases.
 18

A A
B =d B
C Wd
0.2
80
A B unpolarized group B
0.6
Power-law exponent

polarized group B
0.15
60 0.5

Wd (MHz)
= (7s)
0.4
0.1
40
0.3

d
0.2
0.05 A B 20
0.1
0 0 0
0 0.5 1 1.5 2 0 5 10 15 20 unpolarized B polarized
PolarizedB
B
-1
Inverse effective disorder (MHz ) + / 2 : (MHz)

FIG. S12. Fitted parameters of the time-dependent disorder model extracted from a Monte-Carlo optimization.
(A) Exponents of the power-law decay of group A polarization with oppositely polarized (light blue) and unpolarized (dark blue)
group B as a function of inverse effective disorder. Solid lines correspond to numerical simulation results. (B) The extracted
τd as a function of Rabi frequency. Light and dark blue point corresponds to fully oppositely polarized and unpolarized group
B spin states, respectively. (C) The extracted dynamical disorder Wd . All errorbars are evaluated from the standard deviation
of the optimized parameter after running 10 independent Monte-Carlo runs.
 19

∗
These authors contributed equally to this work
[1] M. J. Burek, N. P. de Leon, B. J. Shields, B. J. Hausmann, Y. Chu, Q. Quan, A. S. Zibrov, H. Park, M. D. Lukin, and
M. Loncar, Nano letters 12, 6084 (2012).
[2] J. Choi, S. Choi, G. Kucsko, P. C. Maurer, B. J. Shields, H. Sumiya, S. Onoda, J. Isoya, E. Demler, F. Jelezko, N. Y. Yao,
and M. D. Lukin, Phys. Rev. Lett. 118, 093601 (2017).
[3] E. Van Oort and M. Glasbeek, Chemical Physics Letters 168, 529 (1990).