---
title: Extended Canonical Ensemble
url: https://www.emergentmind.com/topics/extended-canonical-ensemble
type: topic
---

# Extended Canonical Ensemble

The term **extended canonical ensemble** is used in several non-equivalent but structurally related ways across statistical mechanics, quantum theory, gravitational thermodynamics, numerical sampling, and hybrid dynamical formalisms. In the most restrictive sense, it denotes a construction that enlarges the state space or the weighting rule so that the canonical Gibbs factor appears as a projection, limit, interpolation, or generalized equilibrium law. In the recent constrained-quantum formulation, the canonical ensemble is realized in an extended Hilbert space and obtained from the single projector $\delta(\hat C)$ by taking an imaginary clock separation, while the microcanonical ensemble is obtained from the same object in the conjugate clock-energy representation [2603.12148]. Other usages extend the canonical ensemble by conditioning principles with strong-coupling corrections [1912.11137], by Gaussian or quadratic penalties that interpolate between canonical and microcanonical or grand-canonical descriptions [1006.0708], [1706.02587], by operator-valued hybrid quantum-classical states [2004.01986], by covariance- and complex-parameter corrections beyond ensemble equivalence [1710.11422], by generalized thermodynamic conjugacies [2006.00485], or by extended phase-space and singularity-based constructions in black-hole and gravity settings [1705.05002], [2603.20571]. The expression therefore names a family of extensions rather than a single universally standardized ensemble.

## 1. Terminological scope and recurrent structure

Across the cited literature, the extension is introduced by altering one of four ingredients of the canonical formalism. First, the **state space** may be enlarged, as in the extended Hilbert space $\mathcal{H}_{ext}=\mathcal{H}_T\otimes\mathcal{H}_{QM}$ with constraint $\hat C=\hat P_T+\hat H$ [2603.12148]. Second, the **weighting rule** may be modified by an additional factor, such as the Gaussian factor
$$
\exp\!\left[-\frac{(N-N̄)^2}{2\sigma^2}\right]
$$
for particle-number control in classical fields, or $\exp[-\gamma(E-U)^2]$ in the Extended Gaussian Ensemble [1706.02587], [1006.0708]. Third, the **conditioning principle** may be generalized so that the canonical exponential factor arises as an asymptotic conditional law and may acquire an interaction-induced shift in the exponent [1912.11137]. Fourth, the **thermodynamic control variables** may be enlarged, as in generalized exponential-family ensembles with observables $x_\eta(\omega)$ and conjugate forces $X_\eta$, or in AdS black-hole thermodynamics where pressure $P=-\Lambda/(8\pi)$ enters the first law [2006.00485], [1705.05002].

A concise comparison is useful because the same expression denotes different mathematical operations.

| Usage | Defining extension | Representative object |
|---|---|---|
| Constrained quantum theory | Time promoted to auxiliary clock degree of freedom | $\mathcal{P}=\delta(\hat C)$ |
| Conditional/probabilistic | Conditioning on total function with asymptotic exponential tilt | $f_X(x)e^{-\lambda x}/Z(\lambda)$ |
| Gaussian/interpolating | Extra quadratic penalty in energy or particle number | $\exp[-\gamma(E-U)^2]$, $\exp[-(N-N̄)^2/(2\sigma^2)]$ |
| Hybrid quantum-classical | Operator-valued density on phase space | $p_\beta(\xi)=Z_{HCE}^{-1}e^{-\beta H(\xi)}$ |
| Beyond ensemble equivalence | Complex extension and covariance corrections | $\Omega_{C^*}=\int P_{can}(G^*|\theta^*+i\psi)\,d\psi$ |
| Gravity/black holes | Extended phase space or residue-controlled singular sector | $dM=T\,dS+\Phi\,dQ+\Omega\,dJ+V\,dP$ |

This comparison suggests a common pattern: the canonical weight is not discarded but embedded into a larger structural framework. A plausible implication is that “extension” often refers less to replacing the canonical ensemble than to explaining its origin, widening its domain of validity, or interpolating it with neighboring ensembles.

## 2. Constrained quantum origin and the clock-extended formulation

In the constrained-quantum construction, ordinary quantum mechanics is embedded into an extended Hilbert space by introducing an auxiliary clock Hilbert space $\mathcal{H}_T$ carrying a canonical pair $(\hat T,\hat P_T)$ satisfying $[\hat T,\hat P_T]=i$ and commuting with the system Hamiltonian $\hat H$. The extended Hilbert space is
$$
\mathcal{H}_{ext}=\mathcal{H}_T\otimes\mathcal{H}_{QM},
$$
and reparametrization invariance is encoded by the single first-class constraint
$$
\hat C=\hat P_T+\hat H.
$$
Physical states are selected by the kernel of $\hat C$, and the projector-valued distribution
$$
\mathcal{P}=\delta(\hat C)=\frac{1}{2\pi}\int_{-\infty}^{+\infty} d\alpha\, e^{i\alpha \hat C}
$$
implements the rigging map by group averaging [2603.12148].

In the clock-time representation, the physical kernel
$$
K_{phys}(T_f,q_f;T_i,q_i)=\langle T_f,q_f|\delta(\hat C)|T_i,q_i\rangle
$$
reduces to the real-time propagator,
$$
K_{phys}(T_f,q_f;T_i,q_i)=\langle q_f|e^{-i(T_f-T_i)\hat H}|q_i\rangle.
$$
If the clock interval is taken to be purely imaginary, $\Delta T:=T_f-T_i=-i\beta$ with $\beta>0$, then
$$
K_\beta(q_f,q_i)=\langle q_f|e^{-\beta \hat H}|q_i\rangle,
$$
which is the Euclidean kernel. Tracing over $\mathcal{H}_{QM}$ gives
$$
Z(\beta)=\operatorname{Tr}_{\mathcal{H}_{QM}} e^{-\beta \hat H}.
$$
In the conjugate clock-energy basis $\hat P_T|E\rangle=-E|E\rangle$, the same projector yields
$$
K(E_f,q_f;E_i,q_i)=\delta(E_f-E_i)\langle q_f|\delta(\hat H-E_i)|q_i\rangle,
$$
and the microcanonical density of states is
$$
\Omega(E)=\operatorname{Tr}_{\mathcal{H}_{QM}}\delta(\hat H-E).
$$
The central claim is therefore structural: canonical and microcanonical statistics are complementary projections of the same constrained kernel, and the usual asymmetry between Euclidean canonical construction and spectral microcanonical construction is representational rather than structural [2603.12148].

This formulation also fixes several conceptual points. The construction explicitly states that there is **no Pauli obstruction**, because $\hat T$ is conjugate only to $\hat P_T$ on the auxiliary factor and $[\hat T,\hat H]=0$. It also states conditions under which the canonical object is well-defined: $\hat H\ge E_0$ for some lower bound, $e^{-\beta\hat H}$ being trace-class for $\beta>0$, appropriate Euclidean boundary conditions, and positivity properties of the heat kernel. The paper further identifies $\beta$ as the Euclidean “length” of the auxiliary clock separation and calls the resulting viewpoint the **extended canonical ensemble**, since the statistical operator $e^{-\beta \hat H}$ is obtained from the single object $\delta(\hat C)$ living on $\mathcal{H}_{ext}$ rather than from independent thermodynamic postulates [2603.12148].

The harmonic oscillator example makes the construction explicit. For
$$
\hat H=\omega\left(a^\dagger a+\frac12\right),\qquad \omega>0,
$$
the canonical partition function is
$$
Z(\beta)=\frac{e^{-\beta\omega/2}}{1-e^{-\beta\omega}}
=\frac{1}{2\sinh\left(\frac{\beta\omega}{2}\right)},
$$
while the microcanonical density of states is
$$
\Omega(E)=\sum_{n=0}^{\infty}\delta\!\Bigl(E-\omega\left(n+\frac12\right)\Bigr).
$$
A stated generalization is the grand-canonical-like constraint
$$
\hat C_\mu=\hat P_T+\hat H-\mu \hat N,
$$
for which an imaginary clock separation yields $e^{-\beta(\hat H-\mu \hat N)}$ [2603.12148].

## 3. Conditional canonical laws, temperature baths, and strong-coupling corrections

A different use of the term arises in the probabilistic conditioning approach, where the canonical form is derived from asymptotic conditional laws rather than from Euclidean dynamics. The setup introduces a subsystem function $X\ge 0$, a reservoir function $\tilde Y_n$, and a total $\tilde Z_n:=X+\tilde Y_n$, with conditioning on $\tilde Z_n\in E_n$, where
$$
E_n=\mu_n+I/\beta_n,\qquad I=[h,h+\delta],\qquad \beta_n\to 0.
$$
Under regularity assumptions on the reservoir window probabilities, Theorem 4.1 states that the conditional density $f_{X|\tilde Z_n}(x;E_n)$ is well-approximated in KL divergence by an exponentially tilted version of the unconditional density,
$$
f_X(x)\exp\{-\beta_n\psi_n(I;\beta_n x)x\},
$$
normalized by
$$
Z_n(I)=\int_0^\infty f_X(x)\exp\{-\beta_n\psi_n(I;\beta_n x)x\}dx.
$$
The exponential parameter is tied to the derivative
$$
\psi_n(I;\beta_n x):=\partial_y\log P(Y_n\in[y,y+\delta]\mid X_n=\beta_n x)\big|_{y=h},
$$
so the canonical weight is determined by the conditioned reservoir window probability [1912.11137].

The limit theory splits into two canonical regimes. In the **smooth-limit regime**, the unique canonical limit is
$$
f_{X|\tilde Z_n}(x;E_n)\approx \frac{f_X(x)e^{-\beta_n\psi(I)x}}{Z(\beta_n\psi(I))},
$$
with
$$
\psi(I)=\partial_y\log P(Y\in[y,y+\delta])|_{y=h},
$$
and the convergence rate satisfies
$$
\lim_{n\to\infty}\frac{D_{KL}\big(f_X e^{-\beta_n\psi(I)\cdot}/Z\,\|\, f_{X|\tilde Z_n}(\cdot;E_n)\big)}{\beta_n^2}=0.
$$
In the **large-deviation regime**, if the laws of $Y_n$ satisfy an LDP with speed $1/\beta_n$ and rate function $\phi$, then the unique canonical limit is
$$
f_{X|\tilde Z_n}(x;E_n)\Rightarrow \frac{f_X(x)e^{-\varphi(I)x}}{Z(\varphi(I))},
$$
with
$$
\varphi(I)=-\phi'(y^*),\qquad y^*=\arg\min_{y\in I}\phi(y),
$$
and
$$
\lim_{n\to\infty}D_{KL}\big(f_Xe^{-\varphi(I)\cdot}/Z\,\|\,f_{X|\tilde Z_n}(\cdot;E_n)\big)=0
$$
[1912.11137].

For additive energy exchange, the standard canonical form is recovered:
$$
P(E_s)\propto g_s(E_s)e^{-\beta E_s},\qquad Z(\beta)=\int g_s(E_s)e^{-\beta E_s}dE_s.
$$
If the reservoir window probability is represented by a structure function $\Gamma(y)$,
$$
P(Y\in[y,y+\delta])=\frac{\int_y^{y+\delta}\Gamma(s)\,ds}{\int_{\mathbb{R}_+}\Gamma(s)\,ds},
$$
then
$$
\beta=\psi(I)=\partial_y\log\int_y^{y+\delta}\Gamma(s)\,ds\big|_{y=h}
=\frac{\partial S_Y(y)}{\partial y}\Big|_{y=h},
$$
where
$$
S_Y(y):=\log\int_y^{y+\delta}\Gamma(s)\,ds.
$$
A notable claim is that $\beta$ is entirely determined by reservoir macroscopic fluctuations and the conditioning window, not by the micro-composition [1912.11137].

The extension beyond weak coupling is explicit. Corollary 4.2 introduces the corrected exponent parameter
$$
\phi_n(I):=\partial_y\log P(Y_n\in[y,y+\delta]\mid X_n=0)\big|_{y=h}
-\partial_x\log P(Y_n\in[y,y+\delta]\mid X_n=0)\big|_{y=h},
$$
and the approximating density becomes proportional to $f_X(x)e^{-\beta_n\phi_n(I)x}$. Equivalently,
$$
\phi_n(I)=\partial_y\log P(Y_n\in[y,y+\delta])|_{y=h}
+\big[\partial_y\log C(x,y)-\partial_x\log C(x,y)\big]_{(x=0,y=h)},
$$
with
$$
C(x,y):=\frac{P(Y_n\in[y,y+\delta]\mid X_n=x)}{P(Y_n\in[y,y+\delta])}.
$$
The second bracketed term is the interaction-induced correction. In the limit theorems, this shift appears through derivatives of the correlation functions $G(I;\xi)$ or $R(I;\xi)$ at $\xi=0$ [1912.11137].

The same work gives a probabilistic definition of a **temperature bath** by the subinterval invariant property. In the smooth regime, for all $I'\subset I$, subinterval invariance is equivalent to invariant temperature $\psi(I')=\psi(I)$, which is in turn equivalent to
$$
P(Y\in[h',h'+\delta'])=\alpha(\delta')e^{\psi(I)h'}
$$
for all $[h',h'+\delta']\subset I$. In the LD regime, subinterval invariance is equivalent to $\varphi(I')=\varphi(I)$ for all $I'\subset I$, and also equivalent to linearity of the rate function on $I$ [1912.11137].

This usage of “extended canonical ensemble” therefore does not enlarge the Hilbert space or add an auxiliary clock. It extends the canonical law to reservoirs known only through function-level statistics and to strongly interacting subsystem–reservoir models in which the exponent acquires an additive correction.

## 4. Interpolating and regularizing ensembles: Gaussian, particle-number, and finite-thermostat extensions

Several works use the term for ensembles that interpolate continuously between canonical and neighboring ensembles by introducing a quadratic control parameter. In the classical-field treatment of ultracold Bose gases, the canonical ensemble has fixed particle number $N$, while the grand-canonical ensemble allows fluctuating particle number controlled by chemical potential $\mu$. The extended or intermediate ensemble prescribes a Gaussian weight over particle number,
$$
w(N)\propto \exp\!\left[-\frac{(N-N̄)^2}{2\sigma^2}\right],
$$
and the resulting c-field measure can be written as
$$
P_\sigma(\psi)\propto \exp\left\{-\frac{E(\psi)-\mu N(\psi)}{k_B T}
-\frac{[N(\psi)-N̄]^2}{2\sigma^2}\right\}.
$$
The limits are explicit: $\sigma\to 0$ recovers the canonical ensemble, while $\sigma\to\infty$ recovers the grand-canonical ensemble [1706.02587].

The same paper develops a modified projected SGPE by adding the number-feedback drift term
$$
-\frac{\gamma(x)T}{\sigma^2}[N(\psi)-N̄]\psi(x,t),
$$
so that the stationary distribution is exactly $P_\sigma(\psi)$. In a uniform interacting $1D$ gas, the natural grand-canonical width is
$$
s_{GCE}=\sqrt{\frac{TV}{g\,g^{(2)}(0)}},
$$
and the intermediate ensemble width is
$$
s_\sigma=\frac{\sigma}{\sqrt{1+(\sigma^2/s_{GCE}^2)}}.
$$
The paper states the practical rule of thumb
$$
\delta N\approx 
\begin{cases}
\sigma & \text{if }\sigma\lesssim s_{GCE},\\
s_{GCE} & \text{if }\sigma\gtrsim s_{GCE}.
\end{cases}
$$
For mesoscopic $1D$ systems with target atom number $N̄=500$, temperatures $T=0.341\,T_*$ and $T=1.365\,T_*$, and interaction strength quoted in units of $1/L$, the reported crossover to ensemble equivalence occurs by $g\approx 2$ at low $T$ and by $g\approx 8$ at high $T$, where the CE and GCE distributions of $N_0$ and $N_{ex}$ become practically identical [1706.02587].

A second interpolating construction is the **Extended Gaussian Ensemble** for the mean-field Blume–Capel model. Its probability weight is
$$
P_{\gamma,\alpha}(E)=\frac{\rho(E)e^{-\alpha E-\gamma(E-U)^2}}{Z_\gamma(U,\alpha)},
$$
with
$$
Z_\gamma(U,\alpha)=\sum_E \rho(E)e^{-\alpha E-\gamma(E-U)^2}.
$$
The limits are again canonical and microcanonical: $\gamma\to 0$ gives the canonical ensemble with $\alpha=\beta$, while $\gamma\to\infty$ yields microcanonical conditioning at $E=U$. The extended thermodynamic potential per site for the Blume–Capel model is
$$
\phi_\gamma(\epsilon,\alpha,m,q)
= q\ln\!\left[\frac{\sqrt{q^2-m^2}}{2(1-q)}\right]
+\frac{m}{2}\ln\!\left[\frac{q+m}{q-m}\right]
+\ln(1-q)
+\gamma\Delta^2(q-Km^2-\epsilon)^2+\alpha\Delta(q-Km^2),
$$
and the paper states that the extended entropy $s_\gamma(\epsilon)$ becomes concave for sufficiently large $\gamma$, with the sufficient condition
$$
\gamma>-\frac{1}{2\Delta T(\epsilon)^2 c(\epsilon)}.
$$
In the nonequivalence region, finite values such as $\gamma\simeq 30$ or $\gamma\simeq 4950$ are reported as sufficient to recover all microcanonical equilibrium states for particular $\Delta/J$ values, while for $\Delta/J>0.46240788$ complete recovery requires $\gamma\to\infty$ [1006.0708].

A third extension modifies the thermostat rather than the statistical weight directly. The generalized or dynamical ensemble used to extend canonical Monte Carlo adopts a linear inverse-temperature feedback
$$
\beta_\omega(E)=\beta_e+\lambda\,\delta E/N,
$$
leading to the Gaussian effective weight
$$
\omega_{GE}(E)=A\exp\left[-\beta_eE-\frac{1}{N}\lambda(E-E_e)^2\right].
$$
The generalized fluctuation–dissipation relation is
$$
C=\beta^2\langle\delta E^2\rangle+C\langle\delta\beta_\omega\delta E\rangle,
$$
and sampling of $C<0$ macrostates requires
$$
\langle\delta\beta_\omega\delta E\rangle>1.
$$
For the $2D$ $10$-state Potts model, the paper reports that the extended Swendsen–Wang method changes the decorrelation scaling from exponential to a weak power law,
$$
\tau\propto N^\alpha,\qquad \alpha\simeq 0.2,
$$
while for the $2D$, $q=7$, $L=25$ case it reports reductions such as $\tau\approx 450$ to $\approx 80$ for Metropolis and $\approx 300$ to $\approx 12.2$ for Swendsen–Wang [1002.2231].

These constructions share a precise formal feature: a quadratic control term softens a hard constraint or a fixed intensive parameter. This suggests a broad interpolating meaning of “extended canonical ensemble,” in which canonical weighting is retained but supplemented by a stiffness parameter controlling fluctuations of energy, particle number, or inverse temperature.

## 5. Hybrid, generalized, and exact-bridge formulations

The phrase also appears in frameworks that generalize what counts as a state or what counts as a canonical probability law. In hybrid quantum-classical statistical mechanics, a hybrid state is an operator-valued density $p(\xi)=f(\xi)\rho(\xi)$ on classical phase space $\Gamma$, with hybrid entropy
$$
S_h[p]=-k_B\int_\Gamma d\xi\, \operatorname{Tr}[p(\xi)\log p(\xi)].
$$
Applying MaxEnt under normalization and mean-energy constraints gives the **Hybrid Canonical Ensemble**
$$
p_\beta(\xi)=Z_{HCE}(\beta)^{-1}e^{-\beta H(\xi)},
$$
with
$$
Z_{HCE}(\beta)=\int_\Gamma d\xi\, \operatorname{Tr}[e^{-\beta H(\xi)}].
$$
The associated thermodynamic quantities are
$$
F(\beta)=-\beta^{-1}\log Z_{HCE}(\beta),\qquad
U(\beta)=-\partial_\beta \log Z_{HCE},\qquad
S_h(\beta)=k_B(\beta U+\log Z_{HCE}).
$$
The paper proves that the classical and quantum limits reproduce the usual classical and quantum canonical ensembles, and that the construction is additive for independent subsystems and invariant under measure-preserving canonical transformations and unitary rotations [2004.01986].

A more abstract mathematical generalization treats the canonical ensemble as a special case of a general exponential family. For observables $x_\eta(\omega)$ with conjugate forces $X_\eta$, Theorem 1 states that the only distribution consistent with thermodynamics under the stated assumptions is
$$
p(\omega)=Z^{-1}\exp\!\Big(\beta\big[\sum_\eta X_\eta x_\eta(\omega)-E(\omega)\big]\Big),
$$
with
$$
Z(T,X,y)=\sum_{\omega\in\Omega}\exp\!\Big(\beta\big[\sum_\eta X_\eta x_\eta(\omega)-E(\omega)\big]\Big).
$$
The canonical, grand-canonical, and isothermal-isobaric ensembles are then special cases. In this sense, the extended canonical ensemble is simply the canonical distribution augmented by additional linear couplings to extensive observables, while retaining the exponential family form and the Legendre structure generated by $B=\ln Z$ and $J=k_BT\ln Z$ [2006.00485].

Another extension appears in the exact bridge between microcanonical enumeration and canonical probabilities beyond ensemble equivalence. For hard constraints $C(G)=C^*$, the microcanonical count satisfies the exact finite-size identity
$$
\Omega_{C^*}=\int_{[-\pi,+\pi]^K}\frac{d^K\psi}{(2\pi)^K}\,
P_{can}(G^*|\theta^*+i\psi).
$$
The relative entropy between microcanonical and canonical ensembles is
$$
\Delta=\ln\!\left[\frac{\Omega_{C^*}^{-1}}{P_{can}(G^*)}\right],
$$
and the saddle-point expansion gives
$$
\Omega_{C^*}=\frac{e^{S_{can}}}{\sqrt{\det(2\pi\tilde\Sigma^*)}}\cdot T^*,
$$
with $\tilde\Sigma^*$ the reduced covariance matrix of the constraints. Here the extension is twofold: a complex continuation $\theta\mapsto \theta+i\psi$ and a covariance-informed correction to canonical enumeration [1710.11422].

These three usages have a common technical theme. They do not merely alter parameters inside a pre-existing Gibbs distribution; they enlarge the admissible objects—operator-valued phase-space densities, generalized conjugate observables, or complexified canonical parameters—while preserving canonical structure at the level of exponential weighting or partition functions.

## 6. Gravitational, black-hole, and extended-dynamics usages

In black-hole thermodynamics, the expression often refers to the **canonical ensemble in extended phase space**, where the cosmological constant is treated as pressure,
$$
P=-\Lambda/(8\pi),
$$
and the first law becomes
$$
dM=T\,dS+\Phi\,dQ+\Omega\,dJ+V\,dP.
$$
For four-dimensional RN-AdS black holes, the thermodynamic volume is
$$
V=\frac{4}{3}\pi r_+^3,
$$
the mass is identified with enthalpy,
$$
M=\frac{3Q^2+3r_+^2+8\pi P r_+^4}{6r_+},
$$
the temperature is
$$
T=\frac{-Q^2+r_+^2+8\pi P r_+^4}{4\pi r_+^3},
$$
and the equation of state is
$$
P=\frac{T}{2r_+}-\frac{1}{8\pi r_+^2}+\frac{Q^2}{8\pi r_+^4}.
$$
For Kerr-AdS black holes, the paper analyzes the canonical ensemble at fixed $J$ in the same extended phase-space setting and uses the thermodynamic metric
$$
g_{ij}=\frac{1}{T}\frac{\partial^2 F}{\partial x^i\partial x^j},
$$
with $x^\mu=(T,\Phi)$ for RN-AdS and $x^\mu=(T,\Omega)$ for Kerr-AdS. The scalar curvature $R$ is used through the $R$-Crossing Method: for subcritical control parameter values, the branches of $R(T)$ cross at the first-order transition temperature, and $R$ diverges at criticality [1705.05002].

A more recent gravity-based usage defines the extended canonical ensemble as the **canonical sector A** of a thermodynamicized gravity framework. The background is selected by the entropy functional
$$
\mathcal{S}_\xi[g,\xi]=\int_{\mathcal V}\nabla_\mu\xi_\nu\,\nabla^\mu\xi^\nu\,\sqrt{-g}\,d^4x,
$$
whose variation yields
$$
\Box \xi_\nu=0,
$$
and, using the Killing identity, the null-equilibrium constraint
$$
R_{\mu\nu}\xi^\mu\xi^\nu=0.
$$
The canonical sector is defined by
$$
Z_A(\beta,N)=\operatorname{Tr}_N e^{-\beta H},
$$
with
$$
F_A(T,N)=-T\ln Z_A,\qquad
E_A=\partial_\beta(\beta F_A)|_N,\qquad
S_A=\beta(E_A-F_A).
$$
The extension lies in extracting the temperature from the residue of a simple pole of the lapse function. For
$$
ds^2=-f_A(r)dt^2+\frac{dr^2}{f_A(r)}+r^2 d\Omega_2^2,
$$
if $f_A(r)$ has a simple zero at $r=r_A$, then
$$
\beta_A=\frac{4\pi}{f_A'(r_A)}
=4\pi\,\operatorname*{Res}_{r=r_A}\frac{1}{f_A(r)}
=\frac{2}{i}\oint_{\Gamma_A}\frac{dr}{f_A(r)},
$$
and the singular action is
$$
I_A^{\mathrm{sing}}=\frac{2E_A}{i}\oint_{\Gamma_A}\frac{dr}{f_A(r)}=\beta_A E_A.
$$
Thus
$$
Z_A\sim \exp(S_A-\beta_AE_A).
$$
For the Schwarzschild-like realization
$$
f_A(r)=1-\frac{2GM}{r},\qquad r_A=2GM,
$$
the paper gives
$$
\beta_A=8\pi GM,\qquad T_A=\frac{1}{8\pi GM},\qquad
S_A=4\pi G M^2,\qquad F_A=\frac{M}{2},
$$
and
$$
C_N=-\frac{1}{8\pi G T_A^2}<0.
$$
The same residue machinery yields the grand-canonical sector by replacing $E$ with $E-\mu N$ [2603.20571].

Finally, the phrase “canonical-ensemble extended” also appears in molecular dynamics, where the extension is dynamical rather than thermodynamic. In canonical-ensemble extended Lagrangian Born–Oppenheimer molecular dynamics, the physical nuclear variables $(R,P)$ are coupled to a Nosé–Hoover chain thermostat, while auxiliary electronic variables are propagated in a time-reversible harmonic oscillator centered on the instantaneous BO ground state. The conserved pseudo-Hamiltonian is
$$
H'=K_{\mathrm{nuc}}(P)+U_{tot}(R)+\sum_{j=1}^M \frac{p_{\eta_j}^2}{2Q_j}
+gk_BT\,\eta_1+k_BT\sum_{j=2}^M \eta_j,
$$
and the paper reports that the conserved quantity remains stable with no systematic drift even in the presence of the thermostat. For bulk silicon, $50$ ps trajectories at $T=300$, $600$, and $900$ K with $\Delta t=1$ fs gave average temperatures $300.5$, $599.6$, and $898.5$ K and mean absolute deviations per atom of $H'$ equal to $1.03\times 10^{-5}$, $2.01\times 10^{-5}$, and $3.19\times 10^{-5}$ a.u., while a $3072$-atom silicon system and a $64$-atom SiC system were also reported stable under the same framework [1705.01448].

These gravitational and dynamical usages are terminologically distinct from the constrained-quantum and probabilistic meanings. They nonetheless preserve the same core idea: a canonical description is embedded into a larger formal apparatus—extended phase space, singular geometric residue calculus, or auxiliary dynamical variables—so that the canonical structure is recovered or controlled within that enlarged setting.

Source: https://www.emergentmind.com/topics/extended-canonical-ensemble