---
title: Two-Parameter Quantum Fisher Matrix
url: https://www.emergentmind.com/topics/two-parameter-quantum-fisher-information-matrix
type: topic
---

# Two-Parameter Quantum Fisher Matrix

The two-parameter quantum Fisher information matrix (QFIM) is the \(2\times 2\) local information matrix associated with a quantum statistical model \(\rho(\theta_1,\theta_2)\). In contemporary quantum metrology it is usually the symmetric-logarithmic-derivative (SLD) QFIM, although in the broader information-geometric formulation quantum Fisher information is a family of monotone metrics indexed by a standard operator monotone function [1008.2417]. In two parameters, the QFIM already exhibits the main structural issues of multiparameter estimation: diagonal sensitivities, off-diagonal parameter coupling, invertibility versus rank deficiency, measurement compatibility, and discontinuities induced by parameter-dependent rank changes [2108.05976].

## 1. Definition and metric status

For a smooth two-parameter family \(\rho(\theta_1,\theta_2)\), the SLDs \(L_i\) are defined by
\[
\partial_i\rho=\frac12(L_i\rho+\rho L_i),
\]
and the SLD-QFIM is
\[
F_{ij}=\frac12\operatorname{Tr}\!\big[\rho(L_iL_j+L_jL_i)\big],\qquad i,j\in\{1,2\}.
\]
The corresponding matrix is
\[
F=\begin{pmatrix}
F_{11} & F_{12}\\
F_{21} & F_{22}
\end{pmatrix},
\qquad F_{12}=F_{21},
\]
and enters the multiparameter quantum Cramér–Rao bound through the inverse matrix when \(F\) is nonsingular [1601.05866].

The broader mathematical setting is not unique. If monotonicity under quantum channels is taken as the defining requirement, then there is a whole family of quantum Fisher informations, each determined by a standard operator monotone function \(f\). For a multiparameter model, the corresponding matrix is
\[
J(\theta)_{ij}
=
\operatorname{Tr}\,\big(J_{\rho(\theta)}^f\big)^{-1}\!\big(\partial_i\rho(\theta)\big)\,\partial_j\rho(\theta),
\]
with generalized logarithmic derivatives
\[
L_i^f(\theta)=(J_{\rho(\theta)}^f)^{-1}\big(\partial_i\rho(\theta)\big).
\]
Within this family, the minimal monotone metric, corresponding to \(f(x)=(x+1)/2\), is the most popular in physics and reproduces the SLD form [1008.2417].

This dual role—simultaneously a statistical precision matrix and a Riemannian metric tensor—is central. In two parameters, \(F_{11}\) and \(F_{22}\) are the squared metric lengths of the tangent directions \(\partial_{\theta_1}\rho\) and \(\partial_{\theta_2}\rho\), while \(F_{12}\) is their inner product under the chosen quantum Fisher metric [1008.2417].

## 2. Geometry, fidelity, and reparameterization

For mixed states, fidelity can be used as a generating function for the two-parameter QFIM. Writing
\[
|\tilde B(\mathbf x,\mathbf x')|
=
\operatorname{Tr}\sqrt{\sqrt{\rho(\mathbf x)}\,\rho(\mathbf x')\,\sqrt{\rho(\mathbf x)}},
\]
the mixed-state quantum metric \(g_{\mu\nu}\) satisfies
\[
g_{\mu\nu}
=
(-i\partial_{\mu'})(i\partial_\nu)\ln|\tilde B(\mathbf x,\mathbf x')|\big|_{\mathbf x'=\mathbf x}
=
\frac14F_{\mu\nu},
\]
so the nearby-state expansion becomes
\[
|\tilde B(\mathbf x,\mathbf x+d\mathbf x)|
=
1-\frac18F_{\mu\nu}\,dx^\mu dx^\nu+O(dx^3).
\]
For a two-parameter family \((\lambda_1,\lambda_2)\), this reads
\[
|\tilde B|
=
1-\frac18\left(F_{11}\,d\lambda_1^2+2F_{12}\,d\lambda_1d\lambda_2+F_{22}\,d\lambda_2^2\right)+O(d\lambda^3)
\]
[2511.05260].

For pure states, the same paper identifies the modulus of the overlap as the generator of the quantum metric and the phase as the generator of the Berry curvature. In the standard normalization used there, \(F_{\mu\nu}=4g_{\mu\nu}\), so the two-parameter QFIM is the symmetric part of the pure-state quantum geometric tensor [2511.05260].

Reparameterization is controlled by a Jacobian congruence transformation. If two parameterizations describe the same unitary family, then for a two-parameter change of variables \((\theta_1,\theta_2)\mapsto(\phi_1,\phi_2)\),
\[
F^{(\phi)}=J^{T}F^{(\theta)}J,
\qquad
J_{ia}=\frac{\partial\theta_i}{\partial\phi_a}.
\]
The same linear rule holds for the infinitesimal generators in the \(SU(2)\) unitary setting, which makes parameter changes geometrically transparent [1807.02690].

The geometric interpretation becomes subtler in nonregular models. In that regime the QFIM and the Bures metric need not coincide pointwise; the Bures metric can remain continuous even when the QFIM is discontinuous because the density matrix rank changes with the parameters [2108.05976].

## 3. Invertibility, singular directions, and nonregularity

For a two-parameter Fisher matrix
\[
F=
\begin{pmatrix}
F_{11} & F_{12}\\
F_{12} & F_{22}
\end{pmatrix},
\]
invertibility is decided by
\[
\det F = F_{11}F_{22}-F_{12}^2.
\]
If \(\det F>0\), the matrix is invertible. If \(\det F=0\), it is singular, and only one or zero independent parameter combinations are estimable with finite variance [2412.01117].

A unified Cramér–Rao formulation replaces the inverse by the Moore–Penrose pseudoinverse,
\[
\operatorname{cov}(\breve{\mathbf x})\succeq F_{\mathbf x}^{+},
\]
and for simultaneous estimation uses
\[
\operatorname{Tr}\{\operatorname{cov}(\breve{\mathbf x})\}\ge \operatorname{Tr}\{F_{\mathbf x}^{+}\}.
\]
In the rank-one case, if \(F=\lambda\,\mathbf u\mathbf u^{T}\), then
\[
F^{+}=\frac{1}{\lambda}\mathbf u\mathbf u^{T},
\]
so only the combination \(\mathbf u^{T}\mathbf x\) is estimable [2412.01117].

The physical meaning of singularity depends on its origin. One possibility is genuine probe insensitivity: the state encodes only one effective combination of \((\theta_1,\theta_2)\). Another is a coordinate singularity, where the physical state manifold is regular but the chosen chart becomes degenerate. A third is nonregularity from parameter-dependent rank changes, where the QFIM can become discontinuous and the usual QCRB need not hold in general [2108.05976].

This distinction matters because the remedies differ. If singularity is coordinate-induced, the correct response is reparameterization. If it is physical, the probe state must be changed. The paper on nonregular models explicitly warns that pseudoinverse manipulations may be overly optimistic and that changing measurement protocols can rectify a singular classical FIM but never a singular QFIM, since the latter is already measurement-optimized [2108.05976].

A paradigmatic rank-one two-parameter QFIM is
\[
\begin{pmatrix}
1 & -2\\
-2 & 4
\end{pmatrix},
\]
which arises when only \(x_1-2x_2\) is encoded. Another is
\[
\begin{pmatrix}
1 & 1\\
1 & 1
\end{pmatrix},
\]
for which only the sum \(x_1+x_2\) is estimable [2412.01117].

## 4. Closed-form qubit constructions

For mixed qubits, an explicit invertibility criterion is available. In the mixed-qubit model
\[
\rho=\lambda(x,y)\,|\phi_1\rangle\langle\phi_1|+\bigl(1-\lambda(x,y)\bigr)\,|\phi_2\rangle\langle\phi_2|,
\]
with eigenvectors parameterized by a mixing angle \(h(x,y)\) and a constant phase \(\theta(x,y)=\theta_0\), the two-parameter QFIM takes the form
\[
F_{xy}=
\begin{pmatrix}
F_{11} & F_{12}\\
F_{21} & F_{22}
\end{pmatrix},
\]
with
\[
F_{11} = \frac{(\partial_x\lambda)^2}{\lambda-\lambda^2} +\Lambda(x,y,\theta_0)\,(\partial_x h)^2,
\]
\[
F_{12}=F_{21} = \frac{(\partial_x\lambda)(\partial_y\lambda)}{\lambda-\lambda^2} +\Lambda(x,y,\theta_0)\,(\partial_x h)(\partial_y h),
\]
\[
F_{22} = \frac{(\partial_y\lambda)^2}{\lambda-\lambda^2} +\Lambda(x,y,\theta_0)\,(\partial_y h)^2.
\]
Its determinant is
\[
\det(F_{xy})
=
\frac{\Lambda(x,y,\theta_0)}{\lambda-\lambda^2}
\bigl(\partial_x\lambda\,\partial_y h-\partial_y\lambda\,\partial_x h\bigr)^2.
\]
Hence the QFIM is singular exactly when
\[
\partial_x\lambda\,\partial_y h-\partial_y\lambda\,\partial_x h=0,
\]
equivalently when the Jacobian of the map \((x,y)\mapsto(\lambda,h)\) vanishes. In Bloch-vector form the same condition becomes
\[
\partial_x |\mathbf W|\,\partial_y w_3-\partial_y |\mathbf W|\,\partial_x w_3=0.
\]
The paper treats invertibility as a necessary condition for simultaneous estimability, not a sufficient one [1601.05866].

Its dissipative-qubit example makes the criterion explicit. For a decaying two-level system with parameters \(x\) and \(\gamma\), the QFIM satisfies
\[
\det(F_{x,\gamma})=0,
\]
so the original parameter pair is not jointly identifiable; after orthogonal diagonalization, only one linear combination has nonzero Fisher information [1601.05866].

For unitary single-qubit \(SU(2)\) processes, the QFIM acquires a geometric form. Writing the unitary generators as
\[
M_{\theta_i}=iU^\dagger\partial_{\theta_i}U=\boldsymbol\sigma\cdot \mathbf m_{\theta_i},
\]
and characterizing the probe by eigenvalues \(p_0,p_1\) and Bloch direction \(\hat{\mathbf n}\), the paper gives
\[
F_{\theta_i\theta_j}
=
4(p_0-p_1)^2
\left[
\mathbf m_{\theta_i}\cdot \mathbf m_{\theta_j}
-
(\hat{\mathbf n}\cdot \mathbf m_{\theta_i})(\hat{\mathbf n}\cdot \mathbf m_{\theta_j})
\right].
\]
In the two-parameter case this is a projected Gram matrix of the two generator Bloch vectors in the plane orthogonal to \(\hat{\mathbf n}\). The off-diagonal element is the projected inner product, so \(F_{12}=0\) when the effective generator directions are orthogonal in that transverse plane. The paper further emphasizes that linear independence of the bare generators is not sufficient for invertibility, because projection onto the probe-dependent plane can still produce a singular \(2\times2\) matrix [1807.02690].

## 5. Computation and measurement-induced Fisher geometry

Analytical evaluation of a two-parameter QFIM is often difficult because standard formulas require diagonalization of \(\rho\). One major alternative is vectorization. If
\[
\Lambda=\rho^{T}\otimes \mathbb I+\mathbb I\otimes \rho,
\]
then
\[
F_{ij}
=
2\,\operatorname{vec}[\partial_i\rho]^T
\Lambda^{-1}
\operatorname{vec}[\partial_j\rho],
\]
and
\[
\operatorname{vec}[L_{\theta_i}]
=
2\,\Lambda^{-1}\operatorname{vec}[\partial_i\rho].
\]
This representation is used directly in two-parameter Heisenberg \(XY\) models and in qubit Bloch-vector calculations, because it bypasses spectral decomposition [1904.07507].

A more general framework dispenses with both diagonalization and orthonormal bases. For a linearly independent, possibly non-orthogonal basis \(\mathcal B=\{|\psi_j\rangle\}\) spanning the support of \(\rho\), with Gram matrix \(G^{\mathcal B}\), operator multiplication and trace are represented as
\[
\boldsymbol{AB}\rightarrow \boldsymbol A^{\mathcal B}\boldsymbol G^{\mathcal B}\boldsymbol B^{\mathcal B},
\qquad
\operatorname{tr}(A)\rightarrow \operatorname{tr}(\boldsymbol A^{\mathcal B}\boldsymbol G^{\mathcal B}).
\]
The SLD Lyapunov equations are then solved by two matrix inverses, and the QFIM is obtained from
\[
H_{\mu,\nu}
=
\operatorname{tr}\!\left(
\boldsymbol L_\mu^{\mathcal B_{\mu,\nu}}
\boldsymbol G^{\mathcal B_{\mu,\nu}}
(\boldsymbol{\partial_\nu\rho})^{\mathcal B_{\mu,\nu}}
\boldsymbol G^{\mathcal B_{\mu,\nu}}
\right).
\]
This method is designed for arbitrary-rank density matrices and is particularly useful when the state is naturally expressed in non-orthogonal states, as in discrete quantum imaging or coherent-state models [2012.01572].

For pure-state variational circuits, a different route uses Stein’s identity. There the paper identifies the QFIM with four times the Fubini–Study metric tensor and uses the overlap Hessian
\[
F_{ij}(\boldsymbol\theta)
=
-\frac12\partial_i\partial_j
\left|
\langle\psi(\boldsymbol\theta')|\psi(\boldsymbol\theta)\rangle
\right|^2
\Big|_{\boldsymbol\theta'=\boldsymbol\theta}
\]
to build stochastic estimators of the full matrix. The stated motivation is that direct construction scales as \(O(d^2)\) with the number of parameters \(d\), whereas the Stein framework reduces the computational complexity to a constant; the approximation error of the estimate scales as \(O(N^{-1/2})\) with the number of perturbation samples \(N\) [2502.17231].

The relation between quantum and measurement-specific classical Fisher matrices can also be expressed geometrically. For a fixed POVM \(\mathsf E=\{E_i\}\), with SLDs \(L_1,L_2\) defined at a full-rank reference state \(\rho\), the induced classical Fisher matrix is
\[
[F_C]_{\mu\nu}
=
\sum_i
\frac{
\operatorname{Re}\operatorname{Tr}(\rho L_\mu E_i)\;
\operatorname{Re}\operatorname{Tr}(\rho L_\nu E_i)
}{
\operatorname{Tr}(\rho E_i)
},
\]
whereas the quantum one is
\[
[F_Q]_{\mu\nu}
=
\frac12\operatorname{Tr}\!\big[\rho(L_\mu L_\nu+L_\nu L_\mu)\big].
\]
Using the \(\rho\)-dependent inner product \(\langle A,B\rangle_\rho=\operatorname{Re}\operatorname{Tr}[\rho AB]\) and the frame operator \(\mathcal F\) of the POVM, these become
\[
[F_C]_{\mu\nu}=\langle L_\mu,\mathcal F(L_\nu)\rangle_\rho,
\qquad
[F_Q]_{\mu\nu}=\langle L_\mu,L_\nu\rangle_\rho.
\]
In a two-parameter model, the best and worst estimated linear combinations are the generalized eigenvectors of the pair \((F_C,F_Q)\). The paper does not establish a full multiparameter matrix inequality or analyze Holevo-type compatibility, but it does show that a fixed informationally complete measurement acts as a contraction on tangent space and therefore degrades every nontrivial parameter direction relative to the SLD optimum [2512.15428].

## 6. Representative applications and interpretive issues

In quantum interferometry, the two-parameter QFIM appears when both arm phases are treated as unknown. Writing the sum and difference phases as
\[
\varphi_s=\varphi_1+\varphi_2,
\qquad
\varphi_d=\varphi_1-\varphi_2,
\]
the pure-state QFIM is
\[
\mathcal F=
\begin{bmatrix}
\mathcal F_{ss} & \mathcal F_{sd}\\
\mathcal F_{ds} & \mathcal F_{dd}
\end{bmatrix},
\]
and the effective two-parameter Fisher information relevant for the differential phase is the Schur complement
\[
\mathcal F^{(2p)}
=
\mathcal F_{dd}-\frac{\mathcal F_{sd}\mathcal F_{ds}}{\mathcal F_{ss}}.
\]
This framework is the correct one when the common phase is a nuisance parameter and no external reference is available. The same paper argues, however, that the single-parameter QFI retains physical meaning when an external phase reference is genuinely part of the setup, so the two-parameter QFIM is not a universal replacement for single-parameter analysis [2007.03265].

Two-qubit noisy systems provide explicit jointly estimable parameter pairs. In the model
\[
\rho(0)=\frac{1-r}{4}\mathcal I+r\,|\vartheta\rangle\langle\vartheta|,
\qquad
|\vartheta\rangle=\sqrt{1-p}\,|00\rangle+\sqrt p\,|11\rangle,
\]
the parameters \(p\) and \(r\) are estimated simultaneously after open-system evolution. The QFIM
\[
F=
\begin{pmatrix}
F_{pp} & F_{pr}\\
F_{rp} & F_{rr}
\end{pmatrix}
\]
has a generally nonzero cross term
\[
F_{pr}=F_{rp}
=
\frac{-4(g^2-1)(2p-1)r}
{r\left(r[16(g^2-1)(p-1)p-3]+2\right)+1},
\]
so the two parameters are statistically coupled. Yet the same model satisfies
\[
[L_r,L_p]=0,
\]
which means that the multiparameter QCRB is locally saturable. The paper further identifies the \(l_1\)-coherence
\[
C_{l_1}(\rho(t))=2r\sqrt{p(1-p)}\,|g|
\]
and the purity \(P(\rho(t))=\operatorname{Tr}[\rho^2]\) as key resources for optimal multiparameter estimation [1812.06172].

Thermal Heisenberg \(XY\) models furnish another class of exact two-parameter QFIMs. The paper studies \((T,\gamma)\) in the anisotropic model and \((T,B)\) in the isotropic model with magnetic field, computes the full matrices by vectorization, and finds that the off-diagonal terms are generally nonzero. In both cases the relevant SLDs commute, so the QCRB is saturable, and the authors conclude that simultaneous estimation is always advantageous over separate estimation for the studied thermal states [1904.07507].

Discrete quantum imaging yields especially clear two-parameter subproblems. For two incoherent point sources with relative and centroid coordinates, the lowest-order paraxial QFIM shows that relative coordinates are decoupled from the relative intensity, while unequal source brightness couples relative and centroid coordinates. For three equidistant sources with unknown intensities \(p_1,p_2\), the paper gives an explicit \(2\times2\) QFIM,
\[
\boldsymbol H(p_1,p_2)
=
\frac{\delta_x^2\operatorname{Var}(g_x)}
{(1-p_2)(4p_1+p_2)-4p_1^2}
\begin{pmatrix}
16(1-p_2) & 4(1+2p_1-p_2)\\
4(1+2p_1-p_2) & 1+8p_1
\end{pmatrix},
\]
making the intensity-coupling term explicit without diagonalizing the density matrix [2012.01572].

Across these examples, several misconceptions are corrected by the literature. Invertibility is only a necessary condition for meaningful joint estimation at the QFIM level, not a sufficient condition for attainability [1601.05866]. Vanishing off-diagonal entries imply statistical independence in the QFIM, but not by themselves full multiparameter compatibility [2012.01572]. Conversely, nonzero off-diagonal terms indicate coupling, but they do not automatically preclude saturability, since commuting or weakly commuting SLDs can still exist [1812.06172]. This suggests that the two-parameter QFIM should be read not merely as a precision matrix, but as a compact diagnostic of local quantum geometry, parameter identifiability, and measurement structure.

Source: https://www.emergentmind.com/topics/two-parameter-quantum-fisher-information-matrix