Introduction

Recent advancements in quantum metrology1,2,3 have pushed the boundaries of thermodynamics into new territories, where small objects are cooled to ultra-low temperatures4,5. This has led to the development of a fast-growing field of quantum thermodynamics6,7,8,9. From a fundamental perspective, this provides revised definitions for some of the well-known concepts, such as work10,11 and heat12, and extending the formulation for the laws of thermodynamics13,14,15,16 and scaling relations17 into the quantum regime. From a practical viewpoint, however, measuring the thermodynamic quantities at the quantum level mandates unprecedented precision18 and advanced refrigeration19,20,21. Temperature, defined through the zeroth law of thermodynamics, is one of the key thermodynamical parameters, with its precise measurement having numerous applications from our daily life to almost any quantum experiment. Thermometry has been explored for near-term quantum simulators22,23 and various experimental platforms7, including nitrogen-vacancy (NV) centres in diamonds24,25, optical nanofibres26 and nano-photonic cavities coupled to nano-optomechanical resonators27.

At low temperatures, precise temperature measurement is extremely challenging28,29,30,31,32,33,34,35,36,37,38,39,40. To measure the temperature of a quantum system, one brings in a probe, i.e. thermometer, and interact it with the system. Then, the procedure of thermometry can be pursued in two ways: (i) non-equilibrium dynamics in which the temperature is extracted as a parameter from the state of the probe before its thermalization30,31,36,41,42,43,44,45,46,47,48,49; and (ii) equilibrium approach in which the probe is measured after reaching an equilibrium with the system and, thus, is described by a thermal state with the same temperature as the system28,38,50,51,52,53,54. In the first approach, the optimal measurement, which minimizes the uncertainty of the estimation through saturating the Cramér-Rao inequality55, generally depends on the unknown temperature of the system which makes its achievement challenging in practice. In the second approach, which is the focus of this paper, the energy measurement is known to be the optimal choice which saturates the quantum Cramér-Rao bound28 independent of temperature. Nonetheless, not all probes provide the same precision in equilibrium thermometry. Therefore, one may wonder what an optimal probe for equilibrium quantum thermometry is.

In a previous work50, the above question has been answered analytically for a general N-level quantum probe. The result shows that the probe has to be an effective two-level system with a unique ground state and N − 1 degenerate excited states. Unlike the optimal measurement basis, the energy gap of the optimal probe is temperature-dependent. This means that the design of an optimal probe needs prior information about the temperature of the system. In other words, the probe can only be used for local thermometry in which the temperature is roughly known, and the objective of thermometry is just to measure it more precisely. Moreover, this optimal probe, with enormous degeneracy in its excited state, is very difficult to be realized in practice. Based on these, there are several natural questions that one can ask: (i) How can we design a probe which operates optimally over a wide range of temperatures? (ii) What does the optimal probe look like if one takes the practical constraints into account?

In this paper, we address all of the aforementioned questions. We first quantitatively develop the concept of global thermometry over an arbitrary temperature interval and introduce the average variance as the quantity which has to be minimized to get the optimal probe. Our notion of global thermometry is distinct from existing works in literature, where the word “global” refers to system size37 and not temperature range. Using our approach, we can determine the optimal probe over a wide range of temperature intervals. We find that the structure of the optimal probe goes through different phases, each with its own energy structure, as the temperature interval increases. We then focus on spin chain probes with various constraints to get the structure of the optimal probes considering practical limitations which one may face in realistic scenarios. Our results show that for the non-uniform XYZ Hamiltonians, the Ising spin chain provides the optimal probe. By considering the non-uniform Heisenberg Hamiltonians, the dimer spin chains become optimal. Our results pave the way for global thermometry in various physical systems, including ion traps and quantum dot arrays.

Results and discussion

Local quantum thermometry

In this section, we review the concept of local quantum thermometry at equilibrium and its optimal probes. Let us consider a thermometer that weakly interacts with a system at some temperature T (see scheme in Fig. 1a). At thermal equilibrium, we can assume that the probe thermalizes at the system temperature T, and is thus described by the Gibbs state \({\rho }_{\text{th}}=\exp (-\beta {H}_{\text{p}})/{\mathcal{Z}}\), where β ≡ 1/kBT is the inverse temperature (we set kB = 1 henceforth), Hp is the probe Hamiltonian with N eigenvalues and \({\mathcal{Z}}\) is the partition function. For \({\mathcal{M}}\) measurement samples, the variance in the estimation of temperature (see Fig. 1a) satisfies the Cramér-Rao inequality:

$${({{\Delta }}T)}^{2}\ge \frac{1}{{\mathcal{M}}{{\mathcal{F}}}_{\text{th}}}$$
(1)

where \({{\mathcal{F}}}_{\text{th}}\) is the thermal quantum Fisher information (QFI) at equilibrium, which can be related to the heat capacity and hence the energy variance of the probe7:

$${{\mathcal{F}}}_{\text{th}}=\frac{{({{\Delta }}{H}_{\text{p}})}^{2}}{{T}^{4}}=\frac{\left\langle {H}_{\,\text{p}\,}^{2}\right\rangle -{\left\langle {H}_{\text{p}}\right\rangle }^{2}}{{T}^{4}}$$
(2)
Fig. 1: Global quantum thermometry.
figure 1

a Schematic of our proposal. A probe weakly interacts with the system and thus reaches an equilibrium at the system temperature \(T\in [{T}_{\min },{T}_{\max }]\). The probe equilibrium state is characterized by a certain energy spectrum with levels E0…EN−1 with N being the number of energy levels. The energy spectrum also determines the thermal quantum Fisher information \({{\mathcal{F}}}_{\text{th}}\) as a function of T. The global thermometry measure G, defined in Eq. (4) as the average uncertainty of the probe, is a property of the probe energy spectrum, and is minimized to achieve the optimal probe. Energy measurements on the probe yield a probability distribution of temperatures with an estimated temperature T* and variance (ΔT)2. The optimal energies against temperature ratio \({T}_{\max }/{T}_{\min }\) for a general N-level system are obtained for: b N = 16, with Ek>2 denoting the degenerate energy levels E3 = ⋯ = EN−1. c N = 64, with Ek>3 denoting the degenerate energy levels E4 = ⋯ = EN−1. In (b) and (c), the energies are scaled with the characteristic temperature Thm, where Thm is the harmonic mean of \({T}_{\min }\) and \({T}_{\max }\). The leftmost data point corresponds to the case of the locally optimal thermometer. The dashed lines indicate the approximate bounds for the energies (explained in main text), while the dash-dotted line shows the optimal energy gap for an effective two-level system. The different phases indicate a sudden change in the optimal energy spectrum of the probe. d Critical ratios τk against N. The critical ratios correspond to the temperature ratios \({T}_{\max }/{T}_{\min }\) at the boundaries of two phases, as labelled in (b) and (c).

Full size image

It is worth emphasizing that the Cramér-Rao bound is saturated for all temperatures if one measures the probe in the energy basis28, for which ρth is diagonal. Thus, in equilibrium thermometry, the only way to increase precision is by increasing the thermal QFI through proper engineering of the probe. However, \({{\mathcal{F}}}_{\text{th}}\) depends on the actual energy spectrum of the probe. An interesting question then arises: what is the optimal Hamiltonian spectrum for a N-level system that maximizes the thermal QFI? This question was answered in a previous work by Correa et al.50, where they proved that the optimal energy spectrum is an effective two-level system with a single ground state and a (N − 1)-fold degenerate excited state. They also showed that the energy gap ϵ between the excited state and ground state is the solution of a transcendental equation50:

$${e}^{x}=(N-1)\frac{x+2}{x-2}$$
(3)

where x ≡ ϵ/T is the dimensionless energy gap.

Nonetheless, this optimal thermometry has a serious limitation as from Eq. (3), one can only obtain the ratio ϵ/T. Hence, for a fixed number of energy levels N, the optimal energy gap ϵ scales linearly with the temperature T. However, T is generally unknown to us. Thus, the probe cannot be optimized unless the temperature is already known within a narrow range, which limits the usefulness of the probe. Due to this limitation, this scheme is known as local thermometry.

Global quantum thermometry

In order to go beyond the limitations of local thermometry, one has to first define the paradigm of global thermometry in which the probe can operate optimally over a wide range of temperatures. We define the average variance \(G({T}_{\max },{T}_{\min })\) over some given temperature range \([{T}_{\min },{T}_{\max }]\) as

$$G({T}_{\max },{T}_{\min })\equiv \frac{1}{{T}_{\max }-{T}_{\min }}\mathop{\int}\nolimits_{{T}_{\min }}^{{T}_{\max }}\frac{dT}{{{\mathcal{F}}}_{\text{th}}(T)}$$
(4)

where the right-hand side is proportional to the average of Cramér-Rao bound in Eq. (1). Unlike the local thermometry scheme, in which the QFI is maximized, in the global thermometry one has to minimize the average variance \(G({T}_{\max },{T}_{\min })\). In the limit of \({T}_{\max }/{T}_{\min }\to 1\), \(G\to 1/{{\mathcal{F}}}_{\text{th}}\), and the minimization of G becomes equivalent to maximizing \({{\mathcal{F}}}_{\text{th}}\). This recovers the special case of local thermometry. The temperature estimation can also be done adaptively in the sense that once the temperature is estimated for a given range then one can then use a secondary probe optimized for a tighter range to get a better estimate. This procedure can be repeated until the desired precision is achieved.

For a given probe of size N, one has to optimize the N − 1 energy levels to minimize \(G({T}_{\max },{T}_{\min })\), setting the ground state energy E0 = 0 as a reference point. The aforementioned integral in Eq. (4) is very difficult to compute analytically (see Supplementary Note 1 for some analytical approximations), so we use numerical optimization to minimize G over all the N − 1 independent energy levels. We note that for large system sizes, especially for the spin chains, the optimization routine becomes infeasible due to the computational costs from diagonalizing the Hamiltonian and the large number of iterations required to ensure a global optimum. In such cases, we combine local optimization with transfer learning to speed up the convergence to optimality (see “Methods” for more details).

Figure 1b, c shows the optimized energy spectrum for N = 16 and N = 64 energy levels, for a temperature range of up to \({T}_{\max }/{T}_{\min }=150\). We use the harmonic mean temperature \({T}_{{\rm{hm}}}\equiv {(1/{T}_{\max }+1/{T}_{\min })}^{-1}\) as the characteristic temperature, and plot the energy levels normalized by Thm against the temperature ratio \({T}_{\max }/{T}_{\min }\). Since the ground state is fixed to be E0 = 0, we only plot the energy levels of the excited states, where Ek is the energy of the kth excited state. In the limiting case where \({T}_{\max }/{T}_{\min }=1\), the situation reduces to the local thermometry case50. Unsurprisingly, numerical optimization gives the energy gap E1 = E2 = …EN−1 ≈ 7.708 Thm, which is consistent with Eq. (3). This exactly describes the effective two-level solution obtained for local thermometry50. As the ratio \({T}_{\max }/{T}_{\min }\) is increased further, the optimal solution is again an effective two-level system with a (N − 1)-fold degenerate excited state. We call this regime of the optimal energy spectrum, namely an effective two-level system, Phase 1. We use the term "phase" throughout the paper to refer to a range of temperature ratio \({T}_{\max }/{T}_{\min }\) where the energy structure of the optimal probe is the same (e.g. effective two-level system). The energy gap here, of course, cannot be described by Eq. (3), as both \({T}_{\min }\) and \({T}_{\max }\) have to be taken into account. The corrections to Eq. (3) for the regime \({T}_{\max }/{T}_{\min }\) near unity can be found in Supplementary Note 1.

The upper and lower dashed lines in Fig. 1b, c denotes the optimal energy gap when the probe is optimized for only \({T}_{\max }\) or \({T}_{\min }\), respectively. This gives a nice physical intuition for the optimal energy gap, as a compromise between the locally optimal probes at \({T}_{\max }\) and \({T}_{\min }\). A simple estimate of the energy gap is the average energy between the upper and lower bounds, which is equivalent to optimizing the probe at the average temperature \(({T}_{\max }+{T}_{\min })/2\). This is fairly accurate, within 3% of the actual energy gap in Phase 1 for N = 16.

As the temperature ratio \({T}_{\max }/{T}_{\min }\) is increased beyond some critical ratio τ1, the optimal energy spectrum is no longer an effective two-level system. Instead, the energy levels remarkably bifurcate into two branches creating a new regime of energy spectrum, namely Phase 2. The optimal probe in Phase 2 is an effective three-level system. The dashed-dotted lines in Fig. 1b, c show the optimization results when the probe is constrained to an effective two-level system, and, can be thought of as an extrapolation of the Phase 1 solutions.

The bifurcation of energy levels can be physically interpreted as the case where the optimal probe has to be sensitive to both ends of the temperature range. When the range is small (i.e. \({T}_{\max }/{T}_{\min }\) below the critical ratio), optimizing the probe at a temperature near the mean gives a sufficiently high \({{\mathcal{F}}}_{\text{th}}\) at both ends of the range. The optimal solution is the highly-degenerate two-level system which has been previously found to be the most sensitive probe for local thermometry50. This argument is obviously not true in the limit where the temperature range is large (i.e. \({T}_{\max }/{T}_{\min }\) above the critical ratio), and the energy gaps have to be split up to produce an effective probe over the entire temperature range. The above intuition is supported by the dashed lines, corresponding to the optimized local probes at \({T}_{\min }\) and \({T}_{\max }\), respectively, which provide approximate bounds for the actual energy levels.

The huge discrepancy between the degeneracies of E1 and E2 (1 vs N − 2) is consistent with previous findings38. This is because the population of the higher excited states are exponentially suppressed, which in turn demands a significantly higher degeneracy in order to achieve a reasonable sensitivity at higher temperatures. As the temperature ratio is increased further, we get further bifurcations of the energy levels in Phases 3 and 4 at the critical ratios τ2 and τ3 respectively in Fig. 1c, leading to an effective four-level system. The difference in the bifurcation diagram can be intuitively understood: for a larger N, more energy levels can be split up from the upper branch to ‘support’ the lower temperatures without sacrificing too much sensitivity at higher temperatures. The critical ratios are plotted against N in Fig. 1d. An interesting observation is that the critical ratio τ2 decreases sharply at N = 40, with the emergence of Phase 4 as indicated by the presence of τ3.

In short, we have discovered new phases of optimal solutions for global quantum thermometry, when the previously found effective two-level system is no longer ideal. This allows us to delineate global thermometry into various regimes. We also emphasize that our observations for the emergence of various phases for the optimal global probe are universal. In other words, it does not depend on the specific \({T}_{\min }\) and \({T}_{\max }\), and only depends on the ratio \({T}_{\max }/{T}_{\min }\) and the harmonic mean Thm.

Realization of optimal probes

Before we study the optimization of various practical spin chain models, we first demonstrate that the optimal energy spectrum for Phase 1 thermometry, the effective two-level system, can be obtained from a generalized version of the classical (or longitudinal-field) Ising model. Consider the Hamiltonian for n spin-1/2 particles (i.e. N = 2n):

$$H=\mathop{\sum }\limits_{k=1}^{n}{H}_{k}$$
(5)

where

$${H}_{k}=\mathop{\sum}\limits_{\left\langle {i}_{1},{i}_{2},\ldots {i}_{k}\right\rangle }{J}_{k}^{{i}_{1},{i}_{2},\ldots ,{i}_{k}}{\sigma }_{z}^{{i}_{1}}{\sigma }_{z}^{{i}_{2}}\ldots {\sigma }_{z}^{{i}_{k}}$$
(6)

is the k-local Hamiltonian with coupling parameters Jk (omitting the spin indices), and \({\sigma }_{z}^{i}\) is the Pauli Z operator for the ith spin. For example, H1 can be interpreted as the interaction of individual spins with their respective external magnetic fields J1, while H2 is the usual ZZ-interaction in the Ising model. Diagrammatically, this Hamiltonian can be depicted by the union of complete k-uniform hypergraphs (k = 1, 2, …, N) where the vertices denote the spins and the hyperedges denote the couplings between them.

To simplify the model, we can set the magnitudes of all the Jk to be the same, i.e. ∣Jk∣ = J, k = 1, 2, …n. By choosing the signs appropriately, the energy spectrum can be made such that there is one ground state and (N − 1) degenerate excited states, with an energy gap of NJ. Thus, by knowing the theoretically optimal energy gap \(\epsilon ({T}_{\min },{T}_{\max })\), one can tune J = ϵ/N to obtain the optimal probe for the Phase 1 thermometry. Interestingly, our numerical evidence suggests that the optimal energy structure in other phases can also be obtained with this generalized model (more details in Supplementary Note 2).

Although one could achieve the theoretical limits of equilibrium quantum thermometry with the classical Hamiltonian in Eq. (5), this generalized model is unrealistic since it requires full control of arbitrary k-body interactions. This is clearly not feasible for a reasonably large n. In the following subsection, we discuss the optimal global thermometry in realistic spin chain setups with much fewer control parameters.

Spin chain probes

As mentioned above, the optimal probes for global thermometry demand special form of energy structures, which are difficult to engineer with physical systems. Moreover, even if such spectra can be engineered, it might be challenging to measure the probe in the energy basis which is required to saturate the Cramér-Rao bound. Hence, there is a need to design practical probes with a high thermometric performance (compared to the theoretical limit). In this section, instead of optimizing over all the (N − 1) independent energy levels, we first start with a spin chain system and minimize G over the spin chain parameters which can be controlled experimentally.

Restricting ourselves to 1D spin chains with only nearest-neighbour (two-body) interactions and periodic boundary conditions, we can write down a non-uniform Heisenberg XYZ Hamiltonian:

$${H}_{\text{XYZ}}= \, \mathop{\sum }\limits_{i=1}^{n}({J}_{x}^{i}{\sigma }_{x}^{i}{\sigma }_{x}^{i+1}+{J}_{y}^{i}{\sigma }_{y}^{i}{\sigma }_{y}^{i+1}+{J}_{z}^{i}{\sigma }_{z}^{i}{\sigma }_{z}^{i+1})\\ \,+\mathop{\sum }\limits_{i=1}^{n}({h}_{x}^{i}{\sigma }_{x}^{i}+{h}_{y}^{i}{\sigma }_{y}^{i}+{h}_{z}^{i}{\sigma }_{z}^{i})$$
(7)

with the presence of local external magnetic fields in all three directions. The couplings and fields are also allowed to vary independently throughout the spin chain. A further constraint \({J}_{x}^{i}={J}_{y}^{i}={J}_{z}^{i}={J}^{i}\) can be added, which gives the non-uniform Heisenberg XXX Hamiltonian:

$${H}_{\text{XXX}}= \, \mathop{\sum }\limits_{i=1}^{n}{J}^{i}({\sigma }_{x}^{i}{\sigma }_{x}^{i+1}+{\sigma }_{y}^{i}{\sigma }_{y}^{i+1}+{\sigma }_{z}^{i}{\sigma }_{z}^{i+1})\\ \,+\mathop{\sum }\limits_{i=1}^{n}({h}_{x}^{i}{\sigma }_{x}^{i}+{h}_{y}^{i}{\sigma }_{y}^{i}+{h}_{z}^{i}{\sigma }_{z}^{i})$$
(8)

In practice, it is difficult to engineer exchange couplings with varying signs. Therefore, we impose the constraint that the spin chain probe must either be antiferromagnetic (all J > 0) or ferromagnetic (all J < 0), while the signs of the local magnetic fields can be arbitrarily varied. In the following, we numerically optimize HXYZ and HXXX for global quantum thermometry.

Special case: XYZ probe

As before, we minimize the average variance \(G({T}_{\min },{T}_{\max })\) for various \({T}_{\min }\) and \({T}_{\max }\) over the parameters in HXYZ. Interestingly, the optimization yields the classical (longitudinal-field) Ising model with Jx = Jy = hx = hy = 0 (omitting the spin indices). Having the Ising probe as the optimal solution also means that the same results would apply for the Heisenberg XXZ model as well. Figure 2a, b shows the optimal Ising couplings and magnetic fields obtained from HXYZ for n = 4 spins. For thermometry in Phase 1, it turns out that the optimal solution is a uniform ferromagnetic Ising model with \({J}_{z}^{i}=-J\,<\, 0\) and \({h}_{z}^{i}=h\,> \, 0\) (see Supplementary Note 3). We note that the solution is not unique, and an antiferromagnetic Ising model with \({J}_{z}^{i}=J\,> \, 0\) is also possible with a different \({h}_{z}^{i}\). Interestingly, at the critical ratios indicated by the dashed lines, the uniformity of the optimal parameters breaks down, and the optimal parameters show a more sophisticated pattern.

Fig. 2: Optimal probe for the non-uniform XYZ Hamiltonian.
figure 2

Optimal parameters against temperature ratio \({T}_{\max }/{T}_{\min }\) for an optimized HXYZ (reduced to an Ising model) for n = 4 spins. a ZZ-coupling strengths Jk between the kth and (k+1)th spin (k = 1, 2, 3, 4). For illustration purposes, we plot − Jk shifted positively by 1 unit. b External magnetic fields, with strength hk acting on the kth spin. The parameters are scaled with the characteristic temperature Thm, where Thm is the harmonic mean of \({T}_{\min }\) and \({T}_{\max }\). The dashed lines indicate the onset of the splitting of optimal parameters.

Full size image

Special case: XXX probe

For the XXX probe, for the sake of simplicity, we first impose another constraint on HXXX by setting a homogenous magnetic field in each direction, i.e. \({h}_{x}^{i}={h}_{x},{h}_{y}^{i}={h}_{y},{h}_{z}^{i}={h}_{z}\ \forall i\). Surprisingly, the optimal probe in this case turns out to be a fully dimerized spin chain with Ji > 0 for odd i and Ji = 0 otherwise. The optimal solutions require no external magnetic field (hx = hy = hz = 0). Figure 3 shows the optimal couplings for the dimer chain obtained from HXXX, for n = 6 spins. This gives us three independent dimers with coupling parameters J1, J2 and J3. In Phase 1, the dimers are all identical to one another. Again, at the critical ratios denoted by the dashed lines, the coupling parameters split into multiple branches.

Fig. 3: Optimal probe for the non-uniform XXX Hamiltonian.
figure 3

Optimal coupling parameters Jk within the kth dimer (k = 1, 2, 3) against temperature ratio \({T}_{\max }/{T}_{\min }\) for an optimized HXXX (reduced to an dimerized chain) for n = 6 spins. The coupling parameters are scaled with the characteristic temperature Thm, where Thm is the harmonic mean of \({T}_{\min }\) and \({T}_{\max }\). The dashed lines indicate the onset of the splitting of optimal parameters.

Full size image

One can physically explain why the fully dimerized spin chain is the optimal solution for HXXX. In the absence of external magnetic fields, the Heisenberg XXX Hamiltonian commutes with the total spin operator \({\hat{S}}^{2}\equiv {\hat{S}}_{x}^{2}+{\hat{S}}_{y}^{2}+{\hat{S}}_{z}^{2}\). Hence, the first excited state has total spin = 1 and is triply-degenerate for all even values of n. By a similar symmetry argument, the uniform dimerized chain gives a 3n/2-fold degeneracy in the first excited state (see Fig. 4a for the example of n = 2), which is the maximum attainable value for the XXX Hamiltonian. That is why the fully dimerized chain becomes the optimal probe in Phase 1. As new phases emerge, one of the couplings of the dimers changes to vary the degeneracy of the spectrum, replicating the bifurcations found for the optimal probes. It is also noted that if we relax the constraint of homogeneous magnetic fields in HXXX, we obtain a probe that is only marginally better than the dimerized chain.

Fig. 4: Comparison of global thermometers.
figure 4

Comparison of energy levels between the Ising model (blue squares, obtained from optimizing the XYZ Hamiltonian HXYZ) and the dimerized chain (in red circles, obtained from optimizing the XYZ Hamiltonian HXXX). The energy levels are normalized against Thm, the harmonic mean of \({T}_{\min }\) and \({T}_{\max }\) (the lower and upper bounds of the temperature range respectively). The number of spins is n = 4 (corresponding to N = 2n = 16 energy levels), and the optimal energies (dashed lines) are calculated for two different temperature ratios: a \({T}_{\max }/{T}_{\min }=1\) (corresponding to Phase 1 in Fig. 1b). Eopt indicates the optimal energy level for the excited state. b \({T}_{\max }/{T}_{\max }=5\) (corresponding to Phase 2 in Fig. 1b). E1,opt and E2,opt indicate the optimal energy level for the first and second excited states, respectively. The Ising energy levels are on average of closer proximity to the optimal energies. c Relative difference in the global thermometry measure G against temperature ratio \({T}_{\max }/{T}_{\min }\) for the Ising model (black solid line, obtained from optimizing HXYZ) and the dimerized chain (red dash-dotted line, obtained from optimizing HXXX). Gopt is the theoretical lower bound for the G value. The number of spins is n = 4, compared to the ideal case with N = 16 energy levels. The dashed line indicate the critical ratio for the ideal probe.

Full size image

Optimal measurements

In general, energy measurements is very difficult in practice, making the saturation of Cramér-Rao bound challenging. Remarkably, the probes obtained from optimizing HXYZ and HXXX are not just solutions that minimize G, but also have the added advantage of being easily measurable in the energy basis. In the case of HXYZ where the longitudinal-field Ising model is optimal, the energy measurements are equivalent to measuring all the spins in the computational basis, i.e. σz, which has been realized in various physical setups56,57. In the case of HXXX where the fully dimerized chain is the optimal probe, the energy measurements is reduced to a series of singlet-triplet measurements on all the dimer pairs, which is a typical measurement in quantum dot arrays58,59.

Thus, the optimization results discussed earlier can saturate the Cramér-Rao bound on top of having a low G.

Comparison with theoretical limit

It is insightful to study the energy spectra of the optimized spin chain probes for HXYZ and HXXX. For optimal probes of size n = 4, we plot in Fig. 4a, b the energy spectra for \({T}_{\max }/{T}_{\min }=1\) (Phase 1) and \({T}_{\max }/{T}_{\min }=5\) (Phase 2), respectively. For Phase 1, the Ising excited energy levels shown in Fig. 4a are on average closer to the optimal energy Eopt, denoted by the dashed line. Likewise, considering the case of Phase 2 thermometry in Fig. 4b, the Ising energy levels are closer to the respective optimal energies E1,opt (degeneracy = 1) and E2,opt (degeneracy = N − 2). In particular, we see that the Ising model has a single first excited state very close to E1,opt, whereas the first excited state of the dimer must be threefold degenerate. This suggests that the Ising probe performs better than the dimerized probe in both cases. This is not surprising since HXYZ is more general than HXXX, so we must always have GIsing ≤ Gdimer.

In order to quantify the quality of the optimal spin chain probes, one has to compare their performance with the theoretically ideal thermometers. To do so, we compute the relative difference in G with Gopt, i.e. ΔG/Gopt ≡ G/Gopt − 1 for probes with n = 4, over a range of \({T}_{\max }/{T}_{\min }\) for which two phases are observable. When ΔG is small, it means that the probes perform closer to the theoretical limit. The results are plotted in Fig. 4c. Remarkably, the Ising and dimerized probes perform closer to the theoretical limit in Phase 2 as compared to Phase 1. In addition, as expected, the Ising probe indeed outperforms the dimerized probe at all values of \({T}_{\max }/{T}_{\min }\).

In summary, we quantitatively formalize the notion of global thermometry, which operates over arbitrary large temperature ranges. This is achieved by minimizing the average of variance over a given temperature interval. The proposed approach naturally provides the energy structure of an optimal probe. Interestingly, depending on the temperature range, the optimal thermometer goes through different phases, each favouring a particular energy structure. We note that an alternative formulation for the theory of global quantum thermometry was also recently proposed by Rubio et al.60, which treats it as a scale estimation problem. However, in their example the physical system is a gas of fermions with identical energies, hence they were unable to observe the bifurcation of energy levels that we discovered.

We provide a realization of such optimal probes in generalized Ising models with multi-particle interactions. In practice, however, such interactions might be challenging to fabricate. Hence, by imposing further constraints, we provide new designs for optimal probes which are limited to two-body interactions in the form of nearest neighbours of special type of interactions. Starting from a non-uniform XYZ Hamiltonians, our optimization results in a uniform Ising model as the optimal thermometer which can be realized in ion traps. On the other hand, if we restrict ourselves to non-uniform Heisenberg models, then the optimal thermometer is a dimerized chain which can be realized in quantum dot arrays. We also observe the appearance of various phases for practically optimal probes as we vary the temperature range.

Our results unlock several fresh research avenues. For instance, how one can generalize the developed global thermometry procedure to non-equilibrium probes. Another direction to pursue is to find optimal designs for thermometers in other physical platforms, such as optomechanical systems or itinerant particles.

Methods

The numerical optimization in this work is done primarily using global optimization routines. In particular, we use the SciPy implementation61 of differential evolution62, which is an example of an evolutionary algorithm. Differential evolution is a metaheuristic search algorithm that iteratively improves the candidate solution by evolving the population which navigates the search space. It does not rely on gradient information, making it suitable for exploring large search spaces which is useful for optimizing over large number of energy levels. In all cases, we use a population of 500, and increasing the population further does not improve the quality of the optimal solution found. The results are also verified with other popular global optimization routines such as simulated annealing and basin-hopping. Hence, our results do not depend on the specific optimization algorithm used.

For spin chains with n ≥ 10, global optimization becomes infeasible due to two sources of computational costs: (i) diagonalization of the Hamiltonian (in an exponentially large Hilbert space) to obtain all the eigenenergies at each step, and (ii) large number of iterations required for global optimization. While it is possible to speed up the computation by restricting the diagonalization to only return a subset of the lowest eigenenergies, we found that the optimal results are rather inaccurate unless most of the eigenvalues are retained. Hence, the speed up from this approach is marginal. In order to proceed with the optimization for larger systems, we employ a transfer learning approach where the optimal probe for a smaller N is recursively used as an initial guess for an incrementally larger system size, and perform local optimization using the gradient-based BFGS algorithm (see Supplementary Note 3).