19 AUGUST 2016 • VOL 353 ISSUE 6301 753 SCIENCE sciencemag.org
large dimensionality allows them to directly
verify ETH predictions experimentally. Specifically, Kaufman et al. prepare two copies of
the same system, with exactly one boson on
every site. After a quantum quench, which allows particles to hop, correlations grow and
the system becomes entangled. By performing a many-body interference experiment on
the two copies, as suggested in (9) and tested
experimentally in (10), the entanglement entropy of different subsystems as well as the
entropy of the full state was measured (see
the figure). Although the system as a whole
remains pure, small subsystems are found to
become mixed after a short transient time.
Indeed, the reduced density matrices of one-and two-site subsystems become indistinguishable from those of a thermal ensemble.
This equivalence is verified by direct observation of the particle occupation distribution
and by comparing it with the equilibrium
predictions. A recent experiment in a smaller
system of three superconducting qubits (11)
verified that the full time-averaged density
matrix becomes thermal in chaotic regimes;
another direct consequence of ETH (8).
Not only does ETH validate the use of statistical mechanics; there are also many important implications of these ideas to future
science and technology. Understanding the
microscopic structure of complex systems can
provide the necessary tools and intuition for
designing systems with similar or better performance than those found in nature, which
often operate efficiently in far from ideal
conditions. Understanding the conditions
leading to the breakdown of ETH could be
important for developing new technologies
not suffering from the usual thermodynamic
limitations. Remarkably, what first appeared
to be an issue of controversy in quantum mechanics has provided an elegant solution to
the problem of thermalization. It is the existence of individual highly entangled eigenstates that allows the somewhat ambiguous
coarse-graining required in standard classical arguments to be dropped. Interestingly,
ETH can be applied to systems near the classical limit, providing a simple mathematical
framework to understand unanswered questions in classical chaotic systems. j
1. A. M. Kaufman et al., Science 353, 794 (2016).
2. M. Rigol, V. Dunjko, M. Olshanii, Nature 452, 854 (2008).
3. J.Eisert, M.Friesdorf,C.Gogolin, Nat.Phys. 11,124(2015).
4. S. Goldstein et al ., Proc. R. Soc. A 466, 3203 (2010).
5. F. Borgonovi et al. , Phys. Rep. 626, 1 (2016).
6. J. M. Deutsch, Phys.Rev.A 43, 2046 (1991).
7 . M. Srednicki, Phys. Rev. E 50, 888 (1994).
8. L. D’Alessio etal ., Adv.Phys. 65, 239 (2016).
9. A.J.Daley et al., Phys. Rev. Lett.109,020505(2012).
10. R. Islam et al. , Nature 528, 77 (2015).
11. C. Neill et al .; https://arxiv.org/abs/1601.00600 (2016).
12. M. Schreiber et al ., Science 349, 842 (2015).
13. J.-Y. Choi et al. , Science 352, 1547 (2016).
Fighting poverty with data
Machine learning algorithms measure and target poverty
By Joshua Evan Blumenstock
Policy-makers in the world’s poorest countries are often forced to make decisions based on limited data. Con- sider Angola, which recently con- ducted its first postcolonial census. In the 44 years that elapsed between the
prior census and the recent one, the country’s population grew from 5.6 million to
24.3 million, and the country experienced a
protracted civil war that displaced millions
of citizens. In situations where reliable survey data are missing or out of date, a novel
line of research offers promising alternatives.
On page 790 of this issue, Jean et al. (1) apply
recent advances in machine learning to high-resolution satellite imagery to accurately
measure regional poverty in Africa.
Traditionally, wealth and poverty are measured through surveys of household income
and consumption (2). These data provide a
critical input to the world’s most prominent
antipoverty programs, from basic cash transfer programs to multifaceted aid programs
designed to target the extreme poor (3).
However, nationally representative surveys
cost tens to hundreds of millions of dollars
to collect, and many developing countries go
for decades without updating their estimates.
Over the past few decades, researchers
have begun to develop different techniques
for estimating poverty remotely. Initial work
explored the potential of “nightlights” data:
satellite photographs taken at night that
capture light emitted from Earth’s surface.
Since such imagery first became available in
the early 1970s, it was evident that wealthy
regions tended to shine brightest (4). Recent
studies have found a strong correlation be-
tween nightlight luminosity and traditional
measures of economic productivity and
growth (5, 6). Nightlight-based measures
are now frequently used by researchers, for
instance to study the impact of sanctions on
the economy of North Korea (7), where offi-
cial statistics are dubious.
A series of studies in wealthy nations explore how data from the internet and social
media can provide proxies for economic activity (8, 9). Mining the tweets and search queries
of millions of individuals promises real-time
alternatives to more traditional methods of
data collection. However, these approaches
are less relevant to remote and developing regions, where internet infrastructure is limited
and few people use social media.
In developing countries, researchers have
found ways to measure wealth and poverty
using the digital footprints left behind in the
transaction logs of mobile phones, which are
increasingly ubiquitous even in very poor
regions. Regional patterns of mobile phone
use correlate with the regional distribution
of wealth (10). This relationship persists at
the individual level, such that machine learning algorithms can infer an individual subscriber’s socioeconomic status directly from
his or her history of mobile phone use. The
individual predictions can be aggregated into
regional measures of wealth that are about
as accurate as a 5-year-old household survey (11). Phone-based proxies for wealth are
beginning to be used in research, e.g., to understand how new technologies differentially
benefit the wealthy and the poor (12) and to
assess the creditworthiness of would-be borrowers (13).
Although promising, these nontraditional
methods have caveats. As Jean et al. show,
nightlights data are less effective at differentiating between regions at the bottom end
of the income distribution, where satellite
images appear uniformly dark. And mobile
phone data are owned by mobile phone operators and are generally not available to
policy-makers. By contrast, Jean et al. use
only publicly available data.
Taking nightlights as their starting point,
the authors have devised a clever technique
to also extract information from daytime satellite imagery. Daytime imagery is taken at
much higher resolution than nighttime imagery. It thus contains visible features—such
as paved roads and metal roofs—that make
it possible to differentiate between poor and
ultrapoor regions. Jean et al.’s insight was
to apply state-of-the-art deep learning algorithms to the daytime imagery to extract
these features. When given large quantities of
data with labeled patterns, these algorithms
“…there is exciting potential
for adapting machine
learning to fight poverty.”
School of Information, University of California, Berkeley, CA