
3
sponds to the shot-noise limit, as expected. In second
place, it does not depend on the total energy of the sys-
tem |β|2¯ω, usually considered as a resource, but rather
to the spectral’s fluctuations ω2[38]. Thus, for a given
fixed field’s energy, one can freely engineer its spectrum
so as to define different time precision scales using ∆ω
while keeping the same shot-noise scaling (as done in [21–
23], for instance). This suggests that the field’s energy
should not be considered as the classical ressource, and
spectral properties should play a role. We’ll study this
issue by considering intrinsically multimode states where
the frequency variance takes a more complex form. In
this case, frequency and intensity properties are not in-
dependent and the frequency variance can be used to
modify the scaling of the QFI even in a situation where
the photon number variance vanishes.
To demonstrate this, we examine a system compris-
ing nsingle photons, each occupying a distinct ancillary
mode. Each photon has a given frequency profile (spec-
trum), and this system forms a subspace denoted Sn(see
[29] and Supplementary Material A). Hence, if photons
are prepared in a separable state, (∆ ˆ
Ωs)2=Pn
i=1(∆ωi)2,
where (∆ωi)2=Rω2|Si(ω)|2dω −(Rω|Si(ω)|2dω)2,
and Si(ω) is the spectrum of the i-th photon. We’ll sup-
pose, for simplicity, that all the single photons have the
same frequency RMS ∆ω- also called the frequency RMS
per photon -, and that the considered state is pure (our
results can be easily generalized for non-pure states and
arbitrary RMS per photon). Thus, (∆ˆ
Ωs)2=n(∆ω)2,
since we have the equivalent to nindependent probes.
This is the same scaling as the shot-noise. By com-
paring it to the coherent state scaling, we can identify
n(∆ω)2=|β|2ω2. Both expressions are proportional to
the number of photons: not surprisingly, a coherent state
represents the same resource as nindependent photons
[15, 38, 39]. Nevertheless, it is noteworthy that while for
a coherent state the scaling on the average photon num-
ber is due to the fact that (∆ˆn)2=|β|2, ∆ˆn= 0 in Sn.
In Fig. 1 (a) and (b) we show the Joint Spectral Intensity
(JSI) of separable states of two independent (separable)
photons (n= 2). Also, we can identify the frequency
dependency of the coherent state scaling to a frequency
variance centered at ω= 0.
We can now calculate the variance of the operator ˆ
Ω
for a pure non-separable state in Sn, which gives:
(∆ˆ
Ω)2=
n
X
i=1
(∆ωi)2+
N
X
i=1,i̸=j
αiαj(⟨ˆωiˆωj⟩−⟨ˆωi⟩⟨ˆωj⟩).(3)
This variance is bounded, and in the case where all the
variances of the single photons are the same we can
show that (∆ˆ
Ω)2≤n2(∆ω)2, which corresponds pre-
cisely to the Heisenberg limit. We now compare this
result to the usual computations of precision limits in
phase measurements in quantum optics, where the role
of mode variance is disregarded as a quantum resource
and the quantum metrological advantage is exclusively
due to the photon number variance: for NOON [40, 41]
or Schr¨odinger cat states [42], which saturate the Heisen-
berg limit, (∆ˆn)2∝ ⟨ˆn⟩2, where ⟨ˆn⟩is the average photon
number. For states in Sn, however, the photon number
variance is always equal to zero and the variance in the
global evolution generator ˆ
Ω explicitly depends on modal
properties only. Thus, the Heisenberg limit can only
be reached by exploiting mode (frequency) entanglement
and the associated mode/particle statistical properties
of an intrinsically multi-mode state. Interestingly, in or-
der to reach the Heisenberg limit, these variables must
behave as maximally correlated classical ones. However,
this is by no means a paradox: since the considered states
are pure and because of the single photon statistics, these
correlations effectively contribute to the QFI, leading to
the possibility to attain the Heisenberg limit.
We now discuss in detail the type of states that satu-
rate the Heisenberg limit for (3) and provide a geometri-
cal intuitive picture of the observed scaling, pointing out
the analogies and differences with respect to the quadra-
ture phase space [10, 11]. These states, also discussed in
[15], are maximally entangled in the (local) variables ωi,
and their general mathematical expression (see Supple-
mentary Material B) for αi= 1 ∀ireads:
|ψ⟩=ZdΩf(Ω)
Ω + ω0
1
Ω + ω0
2...
Ω + ω0
n,(4)
where ω0
iare constants. The spectral function thus only
depends on one variable (Ω), and |ψ⟩has a (non-physical)
spectrum that is infinitely localized in all collective vari-
ables except for Ω = Pn
i=1 ωi, the one associated to the
operator ˆ
Ω. This means that all the photons display a
collective behavior associated to a re-scaled de Broglie
wavelength λ=c/Ω [39]. As can be seen from the JSI
shown in Fig. 1(c) (for n= 2), these states are repre-
sented by diagonals, and the variance of each mode iis
the projection of these diagonals on the corresponding
frequency axis. This geometrically illustrates the role
of correlations in the scaling. This type of states with
different spectral functions is currently produced in ex-
periments for n= 2 (see [25–27, 30, 43], for instance
and Sup. Mat. sections D and E). In addition, from
Fig. 1 (c) and (d) we see that entanglement, even if nec-
essary, is not sufficient to obtain sub-shot-noise scaling,
and that the symmetry of the spectral variance plays an
important role on the state’s metrological precision. In
the case where we have nphotons in nmodes, the same
type of geometrical picture can be built, and the states
saturating the Heisenberg limit are diagonals of a ndi-
mensional hypercube.
These scaling effects can also be observed in the time
frequency phase space (TFPS). For this, we define the