---
title: 'EPR* Problem: Quantum Hamiltonian Optimization'
url: https://www.emergentmind.com/topics/epr-problem
type: topic
---

# EPR* Problem: Quantum Hamiltonian Optimization

The EPR* problem is a Hamiltonian decision and optimization problem on a weighted graph \(G=(V,E,w)\) with positive edge-weights \(w_{ij}>0\). Its central object is the EPR Hamiltonian
\[
H(G)\;=\;\sum_{(i,j)\in E} w_{ij}\,h_{ij},
\qquad
h_{ij}\;=\;\tfrac12\bigl(I_iI_j+X_iX_j-\,Y_iY_j+Z_iZ_j\bigr),
\]
where \(X_i,Y_i,Z_i\) are Pauli operators on qubit \(i\). The decision version asks, for a threshold \(\lambda\), whether the maximum eigenvalue \(\lambda_{\max}(H(G))\) exceeds \(\lambda\), while the optimization version seeks
\[
\lambda_{\max}(H(G))\;=\;\max_{\ket\psi}\bra\psi\,H(G)\,\ket\psi.
\]
No polynomial-time algorithm is known; the problem lies in the class StoqMA. Recent work gives an efficient \(0.8395\)-approximation algorithm based on a level-2 quantum moment-SoS semidefinite relaxation, a depth-1 circuit ansatz, and a new nonlinear monogamy-of-entanglement bound on star graphs [2512.09896].

## 1. Formal statement of the problem

An instance of EPR* is specified by a graph \(G=(V,E,w)\) on \(n=|V|\) vertices with positive edge-weights. The associated operator \(H(G)\) is a 2-local qubit Hamiltonian obtained by summing the edge terms \(w_{ij}h_{ij}\). The optimization target is the spectral quantity \(\lambda_{\max}(H(G))\), and the decision problem asks whether this value is larger than a prescribed threshold \(\lambda\).

An \(\alpha\)-approximation algorithm for EPR* outputs in polynomial time a number \(\mathcal A(G)\) satisfying
\[
\alpha\,\lambda_{\max}(H(G))
\;\le\;
\mathcal A(G)
\;\le\;
\lambda_{\max}(H(G)).
\]
The paper establishes such a guarantee with \(\alpha=0.8395\), thereby improving the approximation ratio for the EPR Hamiltonian [2512.09896].

| Quantity | Definition | Role |
|---|---|---|
| \(H(G)\) | \(\sum_{(i,j)\in E} w_{ij} h_{ij}\) | Objective Hamiltonian |
| \(h_{ij}\) | \(\tfrac12(I_iI_j+X_iX_j-Y_iY_j+Z_iZ_j)\) | Edge interaction |
| \(\lambda_{\max}(H(G))\) | \(\max_{\ket\psi}\bra\psi H(G)\ket\psi\) | Optimum value |
| \(\mathcal A(G)\) | Polynomial-time output | Approximate value |

## 2. Semidefinite relaxation and edge parameters

The approximation framework begins with an upper bound on \(\lambda_{\max}(H(G))\) obtained from the level-2 quantum moment-SoS SDP relaxation. Its dual yields edge-dependent quantities
\[
g_{ij}
\;=\;
\tfrac{-1+L(X_iX_j)-L(Y_iY_j)+L(Z_iZ_j)}{2},
\]
where \(L\) is the pseudo-expectation produced by the SDP. These variables compress the relaxation into edge-local surrogates that subsequently parameterize the circuit ansatz.

From the dual data one forms
\[
u(G)\;=\;\sum_{(i,j)\in E} w_{ij}\bigl(1+g_{ij}\bigr),
\]
and the relaxation guarantees
\[
u(G)\;\ge\;\lambda_{\max}(H(G)).
\]
This upper bound is the benchmark against which the constructive lower bound is compared. The overall strategy is therefore not merely to produce a candidate quantum state, but to prove that its value is a uniform fraction of an efficiently computable SDP upper bound.

## 3. Depth-1 circuit ansatz and certified lower bound

After solving the level-2 SDP, the algorithm computes angles \(\theta_{ij}=\nu(g_{ij})\) via a fixed scalar function \(\nu\). It then prepares the \(n\)-qubit state
\[
\ket\psi
\;=\;
\prod_{(i,j)\in E}
\exp\!\Bigl(
i\,\theta_{ij}(X_i-Y_i)\otimes(X_j-Y_j)/4
\Bigr)\ket{0}^{\otimes n}.
\]
This is a depth-1 product ansatz over edges. The Hamiltonian value \(\bra\psi H(G)\ket\psi\) can be estimated by measurement, but the analysis also furnishes a deterministic classical lower bound [2512.09896].

That lower bound is
\[
\ell(G)=\sum_{(i,j)\in E}w_{ij}
\,\frac{1+A_{ij}A_{ji}+(A_{ij}+A_{ji})\sin\theta_{ij}}{2}
\;\le\;
\bra\psi H(G)\ket\psi,
\]
where
\[
A_{ij}
\;=\;
\prod_{k\in N(i)\setminus\{j\}} \cos\theta_{ik},
\qquad
N(i)=\{\,k:(i,k)\in E\,\}.
\]
Since \(u(G)\ge \lambda_{\max}(H(G))\), a bound of the form \(\ell(G)\ge \alpha\,u(G)\) immediately yields an \(\alpha\)-approximation. The analytical problem is thus reduced to controlling products of cosines of edge angles derived from the SDP data.

## 4. Nonlinear monogamy on stars and the \(\Theta/\Lambda\) parameterization

The key technical ingredient is a nonlinear monogamy-of-entanglement bound on star graphs. For any vertex \(i\) of degree \(d_i\ge 2\) and any neighbor \(j\), if
\[
g_{ij}=\bra\psi\,h_{ij}\,\ket\psi -1,
\qquad
g_{ik}=\bra\psi\,h_{ik}\,\ket\psi -1,
\]
then
\[
\sum_{k\in N(i)\setminus\{j\}} g_{ik}
\;\le\;
\begin{cases}
1,
& -1\le g_{ij}<-\tfrac1{d_i},\\[6pt]
\tfrac12\bigl(2-d_i-g_{ij}+\sqrt{(d_i^2-1)(1-g_{ij}^2)}\bigr),
& -\tfrac1{d_i}\le g_{ij}\le 1.
\end{cases}
\]
The same bound also holds for the level-2 SDP variables \(g_{ij}\). In positive-part form, with \(g_{ij}^+=\max\{0,g_{ij}\}\), this yields the corollary
\[
\sum_{k\in N(i)\setminus\{j\}} g_{ik}^+
\;\le\;
Q\bigl(g_{ij}^+\bigr),
\]
where
\[
Q(x)=
\begin{cases}
1-x,&0\le x\le \tfrac12,\\[3pt]
\tfrac12\bigl(\sqrt{3(1-x^2)}-x\bigr),
&\tfrac12<x\le\tfrac{\sqrt3}{2},\\[3pt]
0,&\tfrac{\sqrt3}{2}<x\le 1.
\end{cases}
\]
This star inequality is what permits local edge information to constrain all other adjacent edges [2512.09896].

To exploit the bound, the paper introduces scalar functions
\[
\Theta:[0,1]\to[0,1],
\qquad
\Lambda:[0,1]\to[0,1],
\]
with \(\Theta\) increasing, \(\Theta(0)=0\), and satisfying the product-to-sum condition
\[
\prod_i (1-\Theta(x_i))
\;\ge\;
1-\Theta\Bigl(\sum_i x_i\Bigr)
\]
for nonnegative \(x_1,\dots,x_p\) with \(\sum_i x_i\le 1\). One then defines
\[
\tilde\nu(x)
=
\begin{cases}
\Theta(x^+),&x\le\beta,\\[3pt]
\Lambda(x^+),&x>\beta,
\end{cases}
\qquad
\nu(x)=\arcsin\!\sqrt{\tilde\nu(x)}.
\]
Concretely, the chosen constants are
\[
\alpha'=0.839511,\qquad \beta=0.67,\qquad \gamma=0.049.
\]
The function \(\Theta\) is piecewise linear through
\[
(0,0),\;
(Q(\beta),\gamma),\;
\Bigl(\beta,\frac{(\gamma/2+\alpha'(1+\beta)-1)^2}{1-\gamma}\Bigr),\;
(1,2-2\alpha'),
\]
and
\[
\Lambda(y)
=
\frac{\bigl(\tfrac12\,\Theta(Q(y))+\alpha'(1+y)-1\bigr)^2}
{1-\Theta(Q(y))}.
\]

## 5. Worst-case-edge analysis and the \(0.8395\) guarantee

The proof reduces the global approximation ratio to a one-edge minimization:
\[
\frac{\ell(G)}{u(G)}
\;\ge\;
\min_{(i,j)\in E:\,1+g_{ij}>0}
\frac{
1+A_{ij}A_{ji}+(A_{ij}+A_{ji})\sin\nu(g_{ij})
}{
2(1+g_{ij})
}.
\]
The paper then analyzes three exhaustive cases. In the first case, all incident \(g\)-values satisfy \(g\le \beta\), giving
\[
A_{ij}\ge \sqrt{1-\Theta(Q(g_{ij}^+))},
\qquad
\sin\nu(g_{ij})=\sqrt{\Theta(g_{ij}^+)},
\]
and hence
\[
r_1(g)
=
\frac{
2-\Theta(Q(g^+))
+2\sqrt{\Theta(g^+)\bigl(1-\Theta(Q(g^+))\bigr)}
}{
2(1+g)
},
\qquad
0\le g\le\beta.
\]

In the second case, the edge value itself satisfies \(g_{ij}>\beta\). Then
\[
A_{ij}\ge \sqrt{1-\Theta(Q(g_{ij}^+))},
\qquad
\sin\nu(g_{ij})=\sqrt{\Lambda(g_{ij}^+)},
\]
which yields
\[
r_2(g)
=
\frac{
2-\Theta(Q(g^+))
+2\sqrt{\Lambda(g^+)\bigl(1-\Theta(Q(g^+))\bigr)}
}{
2(1+g)
},
\qquad
\beta\le g\le 1.
\]

In the third case, there is a neighbor \(k\) with \(g_{ik}>\beta\). By bipartiteness and the monogamy bound, all other adjacent \(g\)-values are then at most \(\beta\), and a refined estimate produces a third one-dimensional ratio \(r_3(g)\). Claim 3.9 asserts
\[
\alpha
\;\ge\;
\min\Bigl\{
\min_{0\le g\le\beta} r_1(g),\;
\min_{\beta\le g\le 1} r_2(g),\;
\min r_3(g)
\Bigr\}.
\]
A careful numerical, but rigorously checked, evaluation shows that each of these three minima is at least \(\alpha'\approx 0.839511\). This establishes the certified \(0.8395\)-approximation ratio [2512.09896].

The proof architecture combines three ingredients: the nonlinear star monogamy bound, the product-to-sum control built into \(\Theta\), and the reduction to three one-dimensional minimizations. The full technical proofs appear in Sections 4–6 of the paper, and the underlying control problem is to bound
\[
\prod_k \cos\theta_{ik}
=
\prod_k \sqrt{1-\tilde\nu(g_{ik})}
\]
in terms of sums of positive parts \(g_{ik}^+\).

## 6. Limits of the current framework

The same work also proves that the present method is close to its internal limit. First, any analysis that both chooses \(\theta_{ij}=\nu(g_{ij})\) and lower-bounds \(\ell(G)/u(G)\) by the minimum over edges cannot exceed
\[
\max_{\nu}\min\{r_1,r_2,r_3\}\approx 0.839512.
\]
This no-go result shows that the certified \(0.8395\) ratio is essentially optimal within that proof template [2512.09896].

Second, the depth-1 product ansatz
\[
\prod_{(i,j)}
e^{i\theta_{ij}(X_i-Y_i)\otimes(X_j-Y_j)/4}
\]
cannot surpass
\[
(3+\sqrt5)/6\approx 0.8727
\]
on the 4-cycle, even with fully global angle selection as a function of all \(g_{ij}\). The paper therefore concludes that any route to approximation ratios above \(0.8727\) must use deeper circuits or completely new ideas. In the paper’s own summary, current SDP-plus-depth-1-circuit methods cannot achieve substantially better approximation ratios, and further progress will require fundamentally new techniques.

Within this landscape, EPR* occupies a precise niche: it is a graph-structured quantum Hamiltonian optimization problem for which semidefinite relaxations, edge-wise pseudo-moment data, shallow entangling circuits, and monogamy constraints can be assembled into a provable approximation algorithm. The current frontier is not merely the \(0.8395\) factor itself, but the delineation of which parts of that factor arise from the Hamiltonian’s geometry and which arise from the limitations of the analytic and variational framework presently available.

Source: https://www.emergentmind.com/topics/epr-problem