Numerical Shadow

The web resource on numerical range and numerical shadow

User Tools

Site Tools



This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision
Previous revision
numerical-range:generalizations:c-perturbation-of-unitary-matrix-numerical-range [2019/07/24 12:38]
numerical-range:generalizations:c-perturbation-of-unitary-matrix-numerical-range [2020/02/19 08:01] (current)
Line 1: Line 1:
 ====== Perturbation of unitary matrix numerical range ====== ====== Perturbation of unitary matrix numerical range ======
-We are given an arbitrary ​unitary matrix $U$. The numerical range $W(U)$ is convex hull of eigenvalues of matrix $U$, $W(U)=\text{conv}(\lambda(U))$. ​If $V$ denotes unitary matrix arbitrary close to matrix $Uthen the numerical range $W(V)$ should differs from $W(U)$ ​slightlyIt comes from the factthat the function ​$\to \lambda(U)$, ​that for given unitary matrix ​returns vector of its eigenvalues is continuousThe question ​is; can we actually predict how small changes impact changes ​in numerical range?+In our setup we consider the space $\mathrm{L}(\mathbb{C}^d)$. Imagine that the matrices ​are points in space $\mathrm{L}(\mathbb{C}^d)$ and the distance between them is bounded by small constant $0 < c \ll 1$. We will take two unitary ​matrices - matrix $U \in \mathrm{U}(\mathbb{C}^d)$ and its perturbation $V \in \mathrm{U}(\mathbb{C}^d)$ i.e. $||U-V||_\infty \le c$ by using $\infty$-Schatten norm. We want to determine the path connecting these points given by smooth curve. ​ To do so, we fix continuous parametric (by parameter $t$) curve $U(t) \in \mathrm{U}(\mathbb{C}^d)$ for any $t \in [0,1]$ with boundary conditions $U(0) := U$ and $U(1) := V $. The most natural and also the shortest curve connecting ​$U$ and $V$ is geodesic [(:​antezana2014optimal)] given by\begin{equation} t \rightarrow ​\exp\left(t \mathrm{Log} \left(U^\dagger V\right)\right), \end{equation} where $\mathrm{Log}$ is the matrix ​function such that it changes eigenvalues ​$\lambda \in \lambda(U)$ into $\log(\lambda(U))$, where $-i\log(\lambda(U)) \subset (-\pi, \pi]$. 
 +We will study how the numerical range $W(U(t))$ will be changed depending on parameter $t$. Let $H :-i \mathrm{Log} \left( U^\dagger V \right)$. Let us see that $H \in \mathrm{Herm}(\mathbb{C}^d)$ and $W(H) \subset (-\pi, \pi]$ for any $U,V \in \mathrm{U}(\mathbb{C}^d)$.  We can also observe that \begin{equation} \begin{split} 
 + W\left(U\exp\left(itH\right)\right) &= W\left( U \exp \left( it VDV^\dagger  
 + \right) \right) = W\left( UV \exp \left( it D \right) V^\dagger  
 + \right) \\& = W\left( V^\dagger U V \exp \left( it D \right) \right) = W\left(  
 + \widetilde{U} \exp \left( it D\right) \right) 
 + \end{split} 
 + \end{equation} where $\widetilde{U} := V^\dagger U V \in \mathrm{U}(\mathbb{C}^d)$.  
 +Hence, without loss of generality we can assume that $H$ is a diagonal ​matrix. Moreover, we can assume that $D \ge 0which follows from simple calculations 
 +W \left( U \exp \left(it D \right) \right) &= W \left( U \exp \left(it D_{+} \right) \left( \exp \left( it \alpha \1 \right) \right) \right) = W \left( e^{it \alpha} U \exp \left(it D_{+} \right) \right) \\&= W \left( U \exp \left(it D_{+} \right) \right). 
 +Let us see that the numerical range of $U(t)$ for any $t \in [0,1]$ is invariant to above calculations although the trajectory of $U(t)$ ​ is changed. Therefore, we will consider the curve \begin{equation} 
 +t \rightarrow U \exp \left( it D_{+} \right), 
 +where $t \in [0,1]$ and $U \in \mathrm{U}(\mathbb{C}^d)$,​ $D_+ \in \mathrm{Diag}(\mathbb{C}^d)$ such that $D_+ \ge 0$. We will focus on the behavior of the spectrum of the unitary matrices $U(t)$, which will reveal the behavior of $W(U(t))$ for relatively small parameter $t$ ​Without loss of generality we can assume that $\tr \left( D_+ \right)=1$. Together with the fact that $D_+ \in \mathrm{Diag}(\mathbb{C}^d)$ and $D_+ \ge 0$ we can note that  $D_+ = \sum_{i=1}^{d} p_i \ket{i}\bra{i}$,​ where $p \in \mathbb{C}^d$ is a probability vector. Let us also define ​the set 
 +S_\lambda^M=\left\{\ket{x} \in \mathbb{C}^d:​  
 +(\lambda\1_d-M)\ket{x}=0,​ \|\ket{x}\|_2=1\right\} 
 +for some matrix ​$\in \mathrm{L}(\mathbb{C}^d)$ which consists of unit eigenvectors corresponding ​to the eigenvalue $\lambda$ of the matrix $M$. We denote by $k=r(\lambda)$ the multiplicity of eigenvalue $\lambda$ whereas by $I_{M,​\lambda} \in U(\mathbb{C}^k,​\mathbb{C}^d)$ we denote the isometry which columns are formed by eigenvectors corresponding to eigenvalue  
 +$\lambda$ of a such matrix $M$. Let $\lambda(t)\beta(t) \in \mathbb{C}$ ​for $t \ge 0$. We will write  
 +$\lambda(t) \approx \beta(t)$ for relatively small $t \geq 0$, whenever  
 +$\lambda(0)=\beta(0)$ and  
 +$\frac{\partial}{\partial t}\lambda(0)=\frac{\partial}{\partial t}\beta(0)$. 
 +===== Theorem ===== 
 +Let $U \in \mathrm{U}(\mathbb{C}^d)$ be a unitary matrix ​with spectral decomposition  
 +U=\sum_{j=1}^d \lambda_j \ket{x_j}\bra{x_j}. 
 +\end{equation} Assume that the eigenvalue $\lambda \in \lambda(U)$ ​is such that $r(\lambda) = k$. Let us define a matrix $V(t)$ given by \begin{equation} 
 +V(t) = \exp(itD_+)=\sum_{i=1}^{d} e^{i p_i t} \ket{i}\bra{i} \in  
 +\mathrm{DU}(\mathbb{C}^d),​ \quad t \geq 0. 
 +Let $\lambda(t):​=\lambda(UV(t)) $  and let every $\lambda_j(t) \in \lambda(t)$ corresponds to eigenvector $\ket{x_j(t)} $. Assume that $\lambda_{1}(t),​ \ldots, \lambda_{k}(t)$ are such eigenvalues that $\lambda_{j}(t)  
 +\to \lambda$, as $t \to 0$. Then:
-In the first setup we take two unitary matrices - matrix ​$\in U_d$ and its perturbation $V \in U_d$ - for example for given constant $0<​c< ​\\! <1$, we have $\| U - V \| \le c$. Then we fix continuous parametric curve $U(t) \in U_d$ for $t \in [0,1]$, that connects matrices ​$U=U(0)and $V=U(1)$. Which one should we take?+1. If $\min\limits_{\ket{x} ​\in S_\lambda^U} ​\sum\limits_{i=1}^dp_i |\braket{i}{x}|^2=0$, then $\lambdais an eigenvalue of $UV(t)$;
-We start our considerations by taking the most natural curve between $U$ and $V$, which is the shortest one - geodesicThe geodesic between unitary matrices is well-known [( :​antezana2014optimal)] and in our case it is given by formula +2If $|\{p_i: p_i>0\}|=l<k$then  
-$$ t \to U \exp(itH(U^\dagger V)),$$ +$\lambdais an eigenvalue of $UV(t)$ and $r(\lambda) \ge k-l$;
-where function ​$Hfor arbitrary unitary matrix ​$U$ has the form $H(U)=-i \text{Log}(U) \in H_d$, with convection, that $\lambda(H(U)) \subset (-\pi, \pi]$.+
-We can simplyfy formulation of our problem to investigate $W(U \exp(itH))$,​ where $\|H\|_{\infty}\le \pi$Without loss of generality we can assume that $H$ is a diagonal matrix. Because, tha global phase does not matter ​($W(U) \sim W(\lambda U)$) and we put special interest on $t$ arbitrary small, we take such matrix $H$, for which $\lambda(H)$ is a probability vector. +3Each eigenvalue ​of product ​$UV(t)$ moves counterclockwise or stays in the initial position as parameter ​$t$ increases;
-===== Theorem ===== +
-Let $U \in U_dbe unitary matrix of dimension $d$ and denote $S_y^M=\{\ket{x}:​ (y\mathbb{1}_d-M)\ket{x}=0,​ \|x\|=1\}$ for some matrix $M$. Assume that $\lambda$ is eigenvalue of $U$, $p=\{p_i\}_{i=1}^d$ is a probability vector and define matrix $V(t) = \sum_{i=1}^{d} e^{i p_i t} \ket{i}\bra{i} \in U_dfor $t \geq 0$. Then:+
 +4. If $k=1$, then
 +\begin{equation*}\lambda_{1}(t) \approx \lambda \exp\left( i t \sum\limits_{i=1}^d\ p_i |\braket{i}{x_1}|^2 \right)
 +for small $t \geq 0$;
 +5. Let $Q:​=I_{U,​\lambda}^\dagger D_+ I_{U,​\lambda}$ and $\lambda_1(Q) \le \lambda_2(Q) \le \ldots \le \lambda_k(Q) $. Then we have
 +\lambda_{j}(t) \approx \lambda \exp\left( i \lambda_{j}(Q) t \right)
 +for small $t \geq 0$ and eigenvector $\ket{ x_j}$ corresponding to $\lambda_j \in \lambda(U)$ ​ is given by \begin{equation*} \ket{x_j}=I_{U,​\lambda} \ket{ v_j},​ \end{equation*} where $\ket{ v_j} \in S^{Q}_{\lambda_j(Q)}$;​
-  * Each eigenvalue of product ​$UV(t)$ moves counterclockwise as $t \rightarrow 2 \pi$ or stays in initial state $t=0$ +6. For each $j=1,\ldots,dwe have 
-  * If $\dim(S_\lambda^U)=k$ and eigenvalues of $UV(t)$, for which initial position was $\lambda$ are $\{\lambda_{t,​j}\}_{j=1}^k$then for small enough $t \geq 0$, +\begin{equation*} 
-$$\lambda_{t,1} \approx \lambda \exp\left( i t \min\limits_{\ket{x\in S_\lambda^U} \sum\limits_{i=1}^dp_i |\braket{i}{x}|^2 \right),​$$ +\frac{\partial}{\partial ​t}\lambda_j(t)=i \lambda_j(t)\sum_{i=1}^d p_i |\braket{i}{x_j(t)}|^2.\end{equation*} 
-$$\lambda_{t,k} \approx \lambda \exp\left\max\limits_{\ket{x} \in S_\lambda^U} \sum\limits_{i=1}^d\ p_i |\braket{i}{x}|^2 ​\right),$$ +Moreover, ​\begin{equation*} 
-  * Solutions $\ket{x_1}$ of $\min\limits_{\ket{x} \in S_\lambda^U} \sum\limits_{i=1}^dp_i |\braket{i}{x}|^2$ and $\ket{x_k}$ of $\max\limits_{\ket{x} \in S_\lambda^U} \sum\limits_{i=1}^d\ ​p_i |\braket{i}{x}|^2$ are orthogonal. +\sum_{j=1}^d \left|\frac{\partial}{\partial ​t}\lambda_j(t)\right|=1. 
-  * If $\min\limits_{\ket{x\in S_\lambda^U} \sum\limits_{i=1}^dp_i |\braket{i}{x}|^2=0$ then $\lambda_{t,1}= \lambda$ +\end{equation*}
-  * If $\dim(S_\lambda^U)=k$ and $|\{p_i: p_i>0\}|=l<k$, then $\lambda$ is eigenvalue of $UV(t)$ and $\dim(S_\lambda^{UV(t)}) \geq k-l.$+
 +Proof of this therorem you can see in [( :​kukulski2020 )].
 +Intuitively speaking, this theorem gives us equations which one can use to predict behaviour of 
 +$W(UV(t))$. Observe the postulate $(6)$ fully determines the movement of the 
 +spectrum. However, this is a theoretical statement and in practice determining ​
 +the function $t \mapsto \ket{x_j(t)}$ is a numerically complex task. The 
 +postulates $(1)-(5)$ play a key role in numerical calculations of $W(UV(t))$. ​
 +The most important fact comes from $(3)$ which says that all eigenvalues move 
 +the same direction or stay in the initial position. The instantaneous velocity ​
 +of a given eigenvalue in general case is given in $(5)$, while in the case of 
 +eigenvalue with multiplicity equal one, the instantaneous velocity is 
 +determined by $(4)$. We see that whenever the spectrum of the matrix $U$ is not 
 +degenerated,​ calculating these velocities is easy. What is more, when some 
 +eigenvalue is 
 +degenerated,​ the postulate $(5)$ not only gives us method to calculate the 
 +trajectory of this eigenvalue, but also determines the form of corresponding ​
 +eigenvector. It is worth noting that the postulates $(4),(5)$ give us 
 +only an
 +approximation of the velocities, so despite being useful in numerical calculations, ​
 +these expressions are valid only in the neighbourhood of $t=0$. Moreover, ​
 +sometimes we are able to precisely specify this velocities. This happens in the 
 +cases presented in $(1),(2)$. Whenever the calculated velocity is zero we know 
 +sure that this eigenvalue will stay in the initial position. According to the 
 +postulate $(2)$ the same happens when the multiplicity of the eigenvalue is 
 +greater than the number of positive elements of vector $p$.
 ===== Illustration of above theorem ===== ===== Illustration of above theorem =====
-For each eigenvalue $\lambda(t)$ of matrix $UE(t)$ we mark its instantaneous velocity given by formula $\sum\limits_{i=1}^d\ p_i |\braket{i}{x(t)}|^2 $, where $\ket{x(t)}$ is corresponding eigenvector. Red colour denotes instantaneous velocity equal to one and blue colour corresponds to the instantaneous velocity equal to zero. +For each eigenvalue $\lambda(t)$ of matrix $UV(t)$ we mark its instantaneous velocity given by formula $\sum\limits_{i=1}^d\ p_i |\braket{i}{x(t)}|^2 $, where $\ket{x(t)}$ is corresponding eigenvector. Red colour denotes instantaneous velocity equal to one and blue colour corresponds to the instantaneous velocity equal to zero. 
 === Example 1 === === Example 1 ===
 Diagonal matrix ​ Diagonal matrix ​
Line 46: Line 115:
 {{ :​numerical-range:​examples:​special_7_2.gif?​nolink |}} {{ :​numerical-range:​examples:​special_7_2.gif?​nolink |}}
 We act on two subspaces with probability $p_1 = 1/3$ and $p_2 = 2/3$. In this case the eigenvalue $\ii$ is three fold degenerated,​ so it stays in the initial position. We act on two subspaces with probability $p_1 = 1/3$ and $p_2 = 2/3$. In this case the eigenvalue $\ii$ is three fold degenerated,​ so it stays in the initial position.
 +More examples you can find in [( :​kukulski2020 )].
numerical-range/generalizations/c-perturbation-of-unitary-matrix-numerical-range.1563971925.txt.gz · Last modified: 2019/07/24 12:38 by plewandowska