|
| 1 | +% !TEX program = pdflatex |
| 2 | +\documentclass[aspectratio=169]{beamer} |
| 3 | + |
| 4 | +\usetheme{Madrid} |
| 5 | +\usecolortheme{default} |
| 6 | + |
| 7 | +\usepackage{amsmath,amssymb,bm,mathtools} |
| 8 | +\usepackage{physics} |
| 9 | +\usepackage{booktabs} |
| 10 | +\usepackage{tikz} |
| 11 | + |
| 12 | +\title[Parameter-shift for VQE]{For which Hamiltonians can we use the parameter-shift rule in VQE?} |
| 13 | +\author{Morten Hjorth-jensen} |
| 14 | +\date{Spring 2026} |
| 15 | + |
| 16 | +\begin{document} |
| 17 | + |
| 18 | +\begin{frame} |
| 19 | +\titlepage |
| 20 | +\end{frame} |
| 21 | + |
| 22 | +\begin{frame}{Core message} |
| 23 | +\begin{block}{Key point} |
| 24 | +The \textbf{parameter-shift rule depends on the generators of the parametrized gates in the ansatz}, \emph{not} directly on the Hamiltonian. |
| 25 | +\end{block} |
| 26 | + |
| 27 | +\begin{itemize} |
| 28 | +\item VQE objective: |
| 29 | +\[ |
| 30 | +E(\bm{\theta})=\bra{\psi(\bm{\theta})}H\ket{\psi(\bm{\theta})},\qquad \ket{\psi(\bm{\theta})}=U(\bm{\theta})\ket{0}. |
| 31 | +\] |
| 32 | +\item The Hamiltonian matters insofar as we must be able to \textbf{measure} $E(\bm{\theta})$ (typically via a Pauli decomposition). |
| 33 | +\end{itemize} |
| 34 | +\end{frame} |
| 35 | + |
| 36 | +\begin{frame}{Typical measurement form for $H$} |
| 37 | +In most applications (quantum chemistry, spin models, lattice gauge models), |
| 38 | +\[ |
| 39 | +H=\sum_{k} c_k P_k, |
| 40 | +\] |
| 41 | +where each $P_k$ is a tensor product of Pauli matrices (a ``Pauli string''): |
| 42 | +\[ |
| 43 | +P_k \in \{I,X,Y,Z\}^{\otimes n}. |
| 44 | +\] |
| 45 | +Then |
| 46 | +\[ |
| 47 | +E(\bm{\theta})=\sum_k c_k \expval{P_k}{\psi(\bm{\theta})}, |
| 48 | +\] |
| 49 | +and each $\expval{P_k}$ can be obtained from measurement statistics (after local basis changes if needed). |
| 50 | +\end{frame} |
| 51 | + |
| 52 | +\begin{frame}{Ansatz structure and generators} |
| 53 | +Assume the ansatz is a product of parametrized unitaries, |
| 54 | +\[ |
| 55 | +U(\bm{\theta}) = U_L(\theta_L)\cdots U_2(\theta_2)U_1(\theta_1), |
| 56 | +\] |
| 57 | +and each parameter enters through |
| 58 | +\[ |
| 59 | +U_j(\theta_j)=e^{-i\theta_j G_j}, |
| 60 | +\] |
| 61 | +with $G_j$ Hermitian (the generator). |
| 62 | +\medskip |
| 63 | + |
| 64 | +\begin{block}{When does (simple) parameter-shift work?} |
| 65 | +When $G_j$ has \textbf{only two distinct eigenvalues}. |
| 66 | +\end{block} |
| 67 | +\end{frame} |
| 68 | + |
| 69 | +\begin{frame}{Most common case in VQE} |
| 70 | +A very common choice is |
| 71 | +\[ |
| 72 | +G_j = \frac{1}{2}P,\qquad P\in\{I,X,Y,Z\}^{\otimes n}. |
| 73 | +\] |
| 74 | +Then the eigenvalues of $G_j$ are $\pm \frac{1}{2}$, and the standard two-term shift rule holds: |
| 75 | +\[ |
| 76 | +\frac{\partial E}{\partial \theta_j} |
| 77 | +=\frac{1}{2}\left[E(\theta_j+\tfrac{\pi}{2})-E(\theta_j-\tfrac{\pi}{2})\right], |
| 78 | +\] |
| 79 | +where $E(\theta_j\pm \tfrac{\pi}{2})$ means evaluating the VQE energy with all other parameters fixed and shifting only $\theta_j$. |
| 80 | +\end{frame} |
| 81 | + |
| 82 | +\begin{frame}{Formal theorem (two-eigenvalue generators)} |
| 83 | +\begin{theorem}[Two-point parameter-shift rule] |
| 84 | +Let |
| 85 | +\[ |
| 86 | +E(\theta)=\bra{\psi}U(\theta)^\dagger H U(\theta)\ket{\psi},\qquad |
| 87 | +U(\theta)=e^{-i\theta G}, |
| 88 | +\] |
| 89 | +where $H$ and $G$ are Hermitian and $G$ has exactly two eigenvalues $\pm r$ (with $r>0$). |
| 90 | +Then |
| 91 | +\[ |
| 92 | +\boxed{ |
| 93 | +\frac{dE}{d\theta} |
| 94 | += |
| 95 | +r\left[E\!\left(\theta+\frac{\pi}{4r}\right)-E\!\left(\theta-\frac{\pi}{4r}\right)\right]. |
| 96 | +} |
| 97 | +\] |
| 98 | +Equivalently, |
| 99 | +\[ |
| 100 | +\frac{dE}{d\theta} |
| 101 | += |
| 102 | +\frac{1}{2s}\left[E(\theta+s)-E(\theta-s)\right], |
| 103 | +\quad s=\frac{\pi}{4r}. |
| 104 | +\] |
| 105 | +\end{theorem} |
| 106 | +\end{frame} |
| 107 | + |
| 108 | +\begin{frame}{Proof (spectral two-eigenvalue structure)} |
| 109 | +\begin{proof} |
| 110 | +Since $G$ has eigenvalues $\pm r$, define the projectors onto the eigenspaces, |
| 111 | +\[ |
| 112 | +\Pi_\pm=\frac{1}{2}\left(I\pm \frac{G}{r}\right), |
| 113 | +\qquad |
| 114 | +G=r(\Pi_+-\Pi_-),\quad \Pi_\pm^2=\Pi_\pm,\quad \Pi_+\Pi_-=0. |
| 115 | +\] |
| 116 | +Then |
| 117 | +\[ |
| 118 | +U(\theta)=e^{-i\theta G} |
| 119 | += e^{-i\theta r}\Pi_+ + e^{+i\theta r}\Pi_-. |
| 120 | +\] |
| 121 | +Hence $E(\theta)$ is a trigonometric polynomial with frequency $2r$: |
| 122 | +\[ |
| 123 | +E(\theta)=A + B\cos(2r\theta) + C\sin(2r\theta), |
| 124 | +\] |
| 125 | +for real constants $A,B,C$ determined by $H,\ket{\psi},\Pi_\pm$. |
| 126 | +Differentiate: |
| 127 | +\[ |
| 128 | +E'(\theta)= -2rB\sin(2r\theta) + 2rC\cos(2r\theta). |
| 129 | +\] |
| 130 | +Now evaluate $E(\theta\pm s)$ with $s=\frac{\pi}{4r}$, so $2rs=\frac{\pi}{2}$: |
| 131 | +\[ |
| 132 | +E(\theta+s)-E(\theta-s) |
| 133 | += |
| 134 | +2\left[-B\sin(2r\theta)+C\cos(2r\theta)\right]. |
| 135 | +\] |
| 136 | +Multiplying by $r$ gives exactly $E'(\theta)$: |
| 137 | +\[ |
| 138 | +r\left[E(\theta+s)-E(\theta-s)\right] |
| 139 | += |
| 140 | +-2rB\sin(2r\theta)+2rC\cos(2r\theta)=E'(\theta). |
| 141 | +\] |
| 142 | +\end{proof} |
| 143 | +\end{frame} |
| 144 | + |
| 145 | +\begin{frame}{Interpretation and what depends on $H$} |
| 146 | +\begin{itemize} |
| 147 | +\item The theorem does \textbf{not} require any special form of $H$ beyond Hermiticity. |
| 148 | +\item What is required in practice is the ability to \textbf{estimate} $E(\theta)$ on hardware: |
| 149 | +\[ |
| 150 | +H=\sum_k c_k P_k \quad\Rightarrow\quad |
| 151 | +E(\theta)=\sum_k c_k \expval{P_k}{\psi(\theta)}. |
| 152 | +\] |
| 153 | +\item Therefore: \textbf{parameter-shift works for essentially any Hamiltonian used in VQE}, as long as you can measure its terms, \emph{and} your ansatz gates have two-eigenvalue generators. |
| 154 | +\end{itemize} |
| 155 | +\end{frame} |
| 156 | + |
| 157 | +\begin{frame}{Examples of Hamiltonians used in VQE} |
| 158 | +\begin{block}{Quantum chemistry (after mapping)} |
| 159 | +Fermionic Hamiltonians (second quantization) mapped via Jordan--Wigner / Bravyi--Kitaev become Pauli sums: |
| 160 | +\[ |
| 161 | +H=\sum_k c_k P_k. |
| 162 | +\] |
| 163 | +\end{block} |
| 164 | + |
| 165 | +\begin{block}{Spin models} |
| 166 | +Ising: |
| 167 | +\[ |
| 168 | +H=\sum_i h_i Z_i + \sum_{ij} J_{ij} Z_iZ_j. |
| 169 | +\] |
| 170 | +Heisenberg: |
| 171 | +\[ |
| 172 | +H=\sum_{ij} J_{ij}\left(X_iX_j+Y_iY_j+Z_iZ_j\right). |
| 173 | +\] |
| 174 | +\end{block} |
| 175 | +All of these are directly Pauli decompositions $\Rightarrow$ measurable energy. |
| 176 | +\end{frame} |
| 177 | + |
| 178 | +\begin{frame}{When the simple two-point rule fails} |
| 179 | +The simple two-evaluation shift rule can fail if the \emph{generator} has more than two eigenvalues. |
| 180 | +\begin{itemize} |
| 181 | +\item Example (schematic): $U(\theta)=e^{-i\theta(X+Z)}$. |
| 182 | +\item Here $G=X+Z$ has a richer spectrum (more than two distinct eigenvalues for multi-qubit generalizations). |
| 183 | +\end{itemize} |
| 184 | + |
| 185 | +\begin{block}{What then?} |
| 186 | +You may need: |
| 187 | +\begin{itemize} |
| 188 | +\item generalized multi-shift rules (more evaluations), |
| 189 | +\item linear-combination-of-unitaries constructions, |
| 190 | +\item or alternative gradient estimators (finite differences, SPSA, adjoint on simulators). |
| 191 | +\end{itemize} |
| 192 | +\end{block} |
| 193 | +\end{frame} |
| 194 | + |
| 195 | +\begin{frame}{Practical takeaway} |
| 196 | +\begin{block}{Answer to ``for which Hamiltonians?''} |
| 197 | +\textbf{For (almost) any Hamiltonian used in VQE}, provided it can be measured (typically via a Pauli decomposition), |
| 198 | +\emph{and} the ansatz gates are generated by operators with two distinct eigenvalues (e.g.\ Pauli strings). |
| 199 | +\end{block} |
| 200 | + |
| 201 | +\begin{itemize} |
| 202 | +\item Hamiltonian requirement (measurement): $H=\sum_k c_k P_k$ (or can be reduced to this). |
| 203 | +\item Ansatz requirement (parameter-shift): each $U_j(\theta_j)=e^{-i\theta_j G_j}$ with $\mathrm{spec}(G_j)=\{\pm r_j\}$. |
| 204 | +\end{itemize} |
| 205 | +\end{frame} |
| 206 | + |
| 207 | +\begin{frame}{One-slide summary} |
| 208 | +\begin{itemize} |
| 209 | +\item Parameter-shift is controlled by the \textbf{gate generators} in the ansatz, not $H$. |
| 210 | +\item If each generator has two eigenvalues $\pm r$, then a \textbf{two-point shift rule} gives exact gradients. |
| 211 | +\item In practice VQE uses Hamiltonians decomposable into Pauli strings, so energies and shifted energies are measurable. |
| 212 | +\end{itemize} |
| 213 | +\end{frame} |
| 214 | + |
| 215 | +\end{document} |
0 commit comments