\documentclass[11pt,a4paper]{article} \usepackage[a4paper,margin=25mm]{geometry} \usepackage[T1]{fontenc} \usepackage{xcolor,lmodern,microtype,amsmath,amssymb,amsthm,booktabs} \usepackage[colorlinks=true,linkcolor=black,citecolor=magenta,urlcolor=black]{hyperref} \newtheorem{thm}{Theorem}[section] \newtheorem{prop}[thm]{Proposition} \newtheorem{lem}[thm]{Lemma} \newtheorem{defn}[thm]{Definition} \newtheorem{conj}[thm]{Conjecture} \newtheorem*{introconj}{Conjecture} \newtheorem{introthm}{Theorem} \renewcommand{\theintrothm}{\Alph{introthm}} \setlength{\parskip}{1pt} \linespread{0.99} \setlength{\emergencystretch}{2em} \setcounter{tocdepth}{1} \raggedbottom \title{Numerical evidence for the almost Hadamard conjecture} \author{Chat, Grok and Teo} \date{} \hypersetup{pdftitle={Numerical evidence for the almost Hadamard conjecture},pdfauthor={Chat, Grok and Teo}} \begin{document} \maketitle \begin{abstract} We report on encouraging numerical evidence for the almost Hadamard conjecture, stating that any local maximizer of the $1$-norm on $U_N$ must have all entries of modulus $1/\sqrt{N}$, or equivalently that every complex almost Hadamard matrix $H=\sqrt{N}\,U$ must be a complex Hadamard matrix, with computations at the key values $N=4,5$ arising from our attempts to both prove and disprove the conjecture, and adding to the previous abstract evidence. \end{abstract} \begingroup \color{magenta}\hypersetup{linkcolor=magenta} \setlength{\parskip}{0pt} \renewcommand{\addvspace}[1]{\vspace{3pt}} \tableofcontents \endgroup \section*{Introduction} \phantomsection\addcontentsline{toc}{section}{Introduction} A complex Hadamard matrix is a square matrix whose entries have modulus 1 and whose rows are orthogonal: \[ |H_{ij}|=1,\qquad HH^*=NI \] Its rescaling $U=H/\sqrt N$ is a flat unitary, meaning that all its entries have modulus $1/\sqrt{N}$. These matrices attain the upper bound \[ L(U)=\|U\|=\sum_{ij}|U_{ij}|\le N\sqrt N \] The following conjecture was made in \cite{BN}. \begin{introconj}[Almost Hadamard, AHC] Every local maximizer of the $1$-norm on $U_N$ attains the bound $N\sqrt N$. \end{introconj} The starting point was the real theory \cite{BCS,BNZ}. A real almost Hadamard matrix is a rescaled local maximizer on $O_N$; such matrices exist in every order and need not be Hadamard. The passage to $U_N$ allows new directions of variation. These exclude several real families coming from designs and circulants \cite{BN}, as recalled in Section~\ref{background}. \begin{samepage} In order to put this conjecture in context and illustrate its difficulty, we discuss 2 other conjectures concerning unitary groups. \begin{introconj}[Marcus--de Oliveira] For diagonal matrices $A=diag(a_1,\ldots,a_N)$ and $B=diag(b_1,\ldots,b_N)$, the inclusion \[ \det(A+UBU^*)\in conv \left\{\prod_{i=1}^N(a_i+b_{\sigma(i)}):\sigma\in S_N\right\} \] holds for every $U\in U_N$. \end{introconj} \end{samepage} This concerns sums of normal matrices with prescribed spectra \cite{M,O}: unitary diagonalization gives the displayed form. The points on the right pair the eigenvalues in every possible order. The conjecture confines the determinant to their convex hull in the complex plane. It is known through $N=3$ \cite{BMP}; $N=4$ is the first unsettled order \cite{BP,MM}. Higher-dimensional results treat self-adjoint pairs \cite{F}, essentially self-adjoint pairs \cite{BKP}, scalar multiples of unitaries \cite{D}, and certain normal dilations \cite{R}. A variational approach maximizes the projection of the determinant onto a line and studies the resulting stationarity equations on $U_N$ \cite{PB,BP}. Its successes and obstructions \cite{MM} provide a useful comparison with the AHC. The second conjecture concerns squared entry moduli. For a unitary $U$, put \[ p(U)=(|U_{ij}|^2),\qquad J_N=(1/N)_{ij} \] The matrix $p(U)$ is bistochastic: its entries are $\ge0$ and its row and column sums are 1. Bistochastic matrices arising this way are called unistochastic. \begin{introconj}[Unistochastic star-shapedness] For every unistochastic matrix $M$, the segment \[ M_t=(1-t)M+tJ_N,\qquad 0\le t\le1 \] consists of unistochastic matrices. \end{introconj} This holds at $N=3$ \cite{BEKTZ}. Rajchel, G\k{a}siorowski and \.Zyczkowski proposed the order-4 conjecture in \cite{RGZ}; the higher-dimensional version is discussed in \cite{RKKZ}. Partial results include rays supplied by robust Hadamard matrices \cite{RGZ} and the circulant subset at order 4 \cite{RKKZ}. In every order, the larger set defined by the necessary row and column polygon inequalities, called bracelet conditions, is star-shaped \cite{RKKZ}. This does not settle the unistochastic case. The connection to the AHC is the strict increase \[ h(M_t)>h(M)\quad(M\ne J_N,\ t>0),\qquad h(M)=\sum_{ij}\sqrt{M_{ij}},\qquad L=h\circ p \] To contradict a local maximum at a particular $U$, however, one needs unitaries realizing $M_t$ that approach that same $U$, up to phases. Section~\ref{orderthree} supplies this lifting argument at order 3. Reconstruction from moduli \cite{AMM}, the defect \cite{TZ}, and the critical loci of the modulus map \cite{BebN} describe the higher-order difficulty. Random unistochastic ensembles \cite{ZKSS}, volume calculations \cite{DZ}, manifold optimization \cite{AMS}, and calibrated searches for mutually unbiased bases \cite{BW} give further context. Our principal structural results concern $U_4$. \begin{introthm} If one $2\times2$ submatrix of a local maximum on $U_4$ can be made real by row and column phases, the whole unitary is flat. \end{introthm} The proof uses a block variation with positive second derivative. When no $2\times2$ block can be made real by phases, the critical-point equations instead impose restrictions on the entry moduli. \begin{introthm} A local maximum on $U_4$ is flat if every row and every column contains the same multiset of entry moduli. \end{introthm} In the remaining case, the reciprocal moduli satisfy a rank-2 condition and explicit polynomial identities. Combining these restrictions with a count of the possible modulus patterns gives another criterion. \begin{introthm} A nonflat local maximum on $U_4$ would have at least $5$ distinct entry moduli. \end{introthm} The last proof uses an exact enumeration of $27$ patterns. All 3 proofs are in Section~\ref{orderfour}, with the remaining curvature equations and numerical evidence in Section~\ref{fourcurvature}. At order 5, our explicit saddle has \[ \text{Hessian inertia}=(15,9,1),\qquad \text{all row and column pair curvatures}<0 \] The inertia lists negative, zero and positive eigenvalues. A 3-column variation gives an increase, and both unstable gradient branches converge to Fourier matrices (Section~\ref{fivegeometry}). Thus pair tests alone can miss a saddle. The original campaigns and the subsequent targeted search have the following sizes, each from a 7-hour run: \begin{center} \begin{tabular}{clrr} \toprule Order & Campaign & Ascent starts & Stationary-search starts\\ \midrule $4$ & Original & $83\,076\,896$ & $52\,237\,424$\\ $5$ & Original & $181\,164\,464$ & $22\,933\,376$\\ $5$ & Targeted & $8\,737\,840$ & $26\,213\,552$\\ \bottomrule \end{tabular} \end{center} At order 4 every ascent return is within $6.014\cdot10^{-6}$ of flat in entry moduli. Every accepted nonflat stationary return has positive curvature; the near-zero and unresolved returns lie below a necessary entry bound. At order 5 every ascent return is near flat, while the original stationary campaign left $404\,581$ unresolved returns. Among its saved returns, $452$ nonflat stationary matrices form $40$ numerical groups and have interval exclusions. Each group representative has a nearby exact critical point with certified Hessian inertia. Their norm enclosures give at least $35$ distinct nonflat critical values. All $288$ saved stalls are explained. One explanation is a certified positive local minimum of the squared stationarity residual: minimizing that residual need not find a critical point of $L$. The targeted campaign leaves no unresolved or candidate flags under the endpoint criteria. A separate replay of the first $16\,189\,008$ original starts, about $70.6\%$ of that sequence, also leaves no such flags after follow-up. Its remaining $6\,744\,368$ starts were not revisited, and the replay is not added to the table. Small-entry exclusions concern the computed endpoints, not their possible limits; matching the known invariant signatures does not prove equivalence. These are sampled searches, not coverings of $U_4$ or $U_5$. Sections~\ref{framework} and~\ref{fivecampaign} separate the variational criteria from the numerical classifications and certificates. From \texttt{code011}, with NumPy, SciPy, mpmath and SymPy, \texttt{python3 reproduce.py} rechecks the identities, certificates and retained outputs in a temporary copy. Its report \texttt{data/reproduction.json} records versions, hashes and outputs for all $27$ checks. \texttt{README.md} indexes the code, and every program is cited below. \section{Hadamard matrices and the conjecture}\label{background} \subsection{Hadamard matrices and norm maximization} The symbols $*$ and $t$ denote conjugate transpose and transpose. \begin{samepage} \begin{defn} A complex Hadamard matrix of order $N$ is a matrix $H$ satisfying \[ |H_{ij}|=1,\qquad HH^*=NI \] It is real Hadamard when its entries belong to $\{1,-1\}$. A unitary $U$ is flat when $|U_{ij}|=1/\sqrt{N}$ for every entry. \end{defn} \end{samepage} The Fourier matrices provide complex examples in every order: \[ F_N=(w^{ij})_{0\le i,j0 \] Near $W_N$ the entry signs are fixed and $L(V)=Tr(S^tV)$. For every nonzero real skew-symmetric $A$, the curve $W_Ne^{tA}$ has \[ L'(0)=0,\qquad L''(0)=Tr((S^tW_N)A^2)<0 \] Thus $W_N$ is a strict local maximum on $O_N$. The diagonal and off-diagonal moduli agree exactly when $N-2=2$. \end{proof} The same matrices fail to be complex almost Hadamard when $N\ne4$ \cite[Theorem 4.7]{BN}. This is the basic real--complex distinction: the additional unitary directions can destroy a real local maximum. \begin{conj}[Almost Hadamard conjecture]\label{ahc} Every local maximizer of the $1$-norm on $U_N$ is flat. Equivalently, every complex almost Hadamard matrix is complex Hadamard. \end{conj} Local maximality here is unrestricted and need not be strict. Row and column phases preserve $L$, and Hadamard matrices can occur in families. The real theory also includes circulant and 2-entry constructions \cite{BNZ}, other entrywise exponents \cite{BNp}, and polar factors of complementary Hadamard submatrices under suitable size assumptions \cite{BNS}. \subsection{Designs and circulants: the earlier evidence} The 2-entry constructions come from square incidence patterns of symmetric block designs. Their rows have constant weights and constant pairwise overlaps. \begin{defn} An $(a,b,c)$ pattern has $2$ symbols $x,y$ such that every pair of distinct rows has the following counts: \[ \begin{array}{c|cccc} \text{pair}&(x,x)&(x,y)&(y,x)&(y,y)\\ \hline \text{count}&a&b&b&c \end{array} \qquad N=a+2b+c \] We take $a,c\ge0$ and $b>0$. \end{defn} For a full-support unitary with this pattern, normalize $y>0$ and write $x=-\varepsilon t y$, with $t>0$ and $|\varepsilon|=1$. Orthogonality and normalization become \[ at^2-2b\,Re(\varepsilon)t+c=0,\qquad y=\big((a+b)t^2+b+c\big)^{-1/2} \] Criticality further restricts the phase $\varepsilon$. \begin{prop} A full-support critical unitary with an $(a,b,c)$ pattern is, after multiplication by a scalar of modulus $1$, real or flat. \end{prop} \begin{proof} The criticality calculation of \cite[Theorem 5.2]{BN} gives \[ (\varepsilon-\bar\varepsilon)(t-1)=0 \] If $t=1$ the 2 entry moduli agree. Otherwise $\varepsilon$ is real, and orthogonality excludes $\varepsilon=-1$, leaving $\varepsilon=1$. \end{proof} The real branches still require a second-variation test. Theorem 6.5 of \cite{BN} excludes $K_N$ for $N\ne4$, the Grassmannian matrices $I_N^{q,d}$, and the Paley biplane matrix $P_{11}$. The projective-plane examples, including the Fano plane at $q=2$, belong to the Grassmannian family at $d=1$ \cite[Proposition 5.5]{BN}. These exclusions do not cover every construction from a design. A circulant matrix has rows obtained by successive cyclic shifts. Its unitarity can be expressed in Fourier coordinates as \[ U=\frac1N F_N\,diag(q_0,\ldots,q_{N-1})F_N^*,\qquad |q_k|=1 \] For real symmetric circulants the parameters satisfy \[ q_k\in\{1,-1\},\qquad q_k=q_{-k} \] The method of \cite[Section 7]{BN} averages second derivatives over suitable circulant self-adjoint generators. A positive average supplies an increasing direction. It excludes the explicit nonflat circulant almost Hadamard examples of \cite{BNZ}; further computer checks use the examples of \cite{BNp}. A general averaged inequality is proposed in \cite[Conjecture 7.8]{BN}, for real symmetric circulants and, in an analogous form, for self-adjoint unitary circulants. It is not proved there. This earlier evidence gives families and variation methods. The small-order study below tests the full tangent space; the order-5 model shows why restricted tests can miss an increasing direction. \section{Variations and small orders}\label{framework} We now pass from norm maximization to first and second derivatives. The only nonsmooth points are matrices with zero entries. \subsection{Criticality and the Hessian} \begin{lem}\label{fullsupport} A local maximum of $L$ on $U_N$ has no zero entry. \end{lem} \begin{proof} Choose a column rotation that moves a nonzero entry into a zero position to first order. In the sum of the increments at $t$ and $-t$, nonzero entries contribute $O(t^2)$, the chosen zero contributes a positive multiple of $|t|$, and the other zeros contribute amounts $\ge0$. This contradicts local maximality. See also \cite[Theorem 1.6]{BN}. \end{proof} We can therefore work where the norm is smooth. Its derivatives are expressed in terms of the entry phases. \begin{defn} A matrix has full support when none of its entries is zero. For such a unitary $V$, its phase matrix and associated matrix are \[ S_{ij}=\frac{V_{ij}}{|V_{ij}|},\qquad X=S^*V \] The matrix $V$ is critical for $L$ when the first derivative vanishes along every unitary curve through $V$. \end{defn} We use the real Frobenius inner product and norm \[ \langle A,B\rangle=Re\,Tr(A^*B),\qquad \|A\|_F^2=\sum_{ij}|A_{ij}|^2 \] Every tangent vector at $V$ is realized by a curve \[ V(t)=Ve^{itB},\qquad V'(0)=iVB,\qquad B=B^* \] Thus self-adjoint matrices $B$ parametrize all unitary directions. \begin{lem}\label{variations} Along $V e^{itB}$ the first derivative of $L$ at zero is $-Im\,Tr(XB)$. The second derivative is \[ Q_V(B)=\sum_{ij}\frac{Re((VB)_{ij}\bar S_{ij})^2} {|V_{ij}|}-Re\,Tr(XB^2) \] Thus $V$ is critical precisely when $X$ is self-adjoint. \end{lem} \begin{proof} For one entry $z(t)$ put $s=z(0)/|z(0)|$. Differentiating $|z|^2=z\bar z$ gives $(|z|)'=Re(z'\bar s)$ at zero, then $(|z|)''=Re(z''\bar s)+Im(z'\bar s)^2/|z(0)|$. Substitute $z'=i(VB)_{ij}$ and $z''=-(VB^2)_{ij}$, and sum. Testing the first derivative against all self-adjoint $B$ gives the last assertion. \end{proof} At a critical point, $Q_V$ is the Hessian. Local maximality requires \[ Q_V(B)\le0\qquad\text{for every self-adjoint }B \] One positive value excludes a local maximum. Nonpositivity alone need not establish one, since higher-order terms can matter. \begin{prop}\label{phasequotient} At full support the row and column phase orbits have dimension $2N-1$. At a critical point their tangent directions lie in the kernel of the Hessian, so the Hessian descends to a quotient of dimension $(N-1)^2$. \end{prop} \begin{proof} The action is $V\to D_1VD_2$, with $D_1,D_2$ diagonal unitaries. At full support it fixes $V$ exactly when $D_1=zI$ and $D_2=z^{-1}I$. The effective action is free and its orbits have dimension $2N-1$. Phase invariance makes the derivative of $L$ along every phase vector field identically zero. Differentiating at a critical point shows that all Hessian terms involving a phase direction vanish. Removing these directions from the $N^2$-dimensional tangent space gives the result. \end{proof} The horizontal Hessian is the restriction to the orthogonal complement of the phase directions. The dimensions used in our computations are \begin{center} \begin{tabular}{lccc} \toprule Space & General order & $N=4$ & $N=5$\\ \midrule Full tangent space & $N^2$ & $16$ & $25$\\ Phase directions & $2N-1$ & $7$ & $9$\\ Horizontal space & $(N-1)^2$ & $9$ & $16$\\ \bottomrule \end{tabular} \end{center} A positive geodesic second derivative also excludes a local maximum without requiring an exact stationarity certificate. \subsection{Pair tests and a polar condition} For 2 columns write their entries and relative phases as \[ p_i u_i,\quad q_i v_i,\qquad p_i,q_i>0,\quad |u_i|=|v_i|=1,\qquad h_i=\bar u_i v_i \] After removing column phases there are 2 real mixing directions. Put \[ d_i=\frac{q_i^2}{p_i}+\frac{p_i^2}{q_i} \] These quantities determine the largest pair curvature. \begin{lem}\label{twocol} For $2$ columns as above, the quantity \[ M_{jk}=\frac14\left(\sum_i d_i+ \left|\sum_i d_i h_i^2\right|\right)-\frac12\sum_i(p_i+q_i) \] is the largest value of $Q_V$ on Frobenius-unit off-diagonal self-adjoint generators supported on these columns. \end{lem} \begin{proof} Set $B_{jk}=e^{i\theta}/\sqrt2$, $B_{kj}=\bar B_{jk}$. The first term of $Q_V(B)$ in Lemma~\ref{variations} becomes $\frac12\sum_i d_i\,Re(h_i e^{-i\theta})^2$. The second uses $B^2=(E_{jj}+E_{kk})/2$. Now $2Re(z)^2=1+Re(z^2)$ on the unit circle. Maximizing the resulting expression over $\theta$ gives the stated formula for $M_{jk}$. \end{proof} The same test applies to rows. Pair tests are inexpensive but do not exhaust the Hessian, as the order-5 example will show. Rank-1 generators give a further necessary condition. \begin{prop} At a local maximum on $U_N$ we have $S^*U\ge I/N$. In particular $S$ is invertible and its unitary polar factor is $U$. \end{prop} \begin{proof} Write $r_{ij}=|U_{ij}|$ and $X=S^*U$. Fix a unit vector $v$ and insert $B=vv^*$ into $Q_U(B)\le0$. For $a_{ij}=Re((Uv)_i\bar v_j\bar S_{ij})$ we have $\sum_j r_{ij}a_{ij}=|(Uv)_i|^2$. By Cauchy--Schwarz and $\sum_j r_{ij}^3\le\sum_j r_{ij}^2=1$ we obtain \[ v^*Xv\ge\sum_{ij}\frac{a_{ij}^2}{r_{ij}} \ge\sum_i\frac{|(Uv)_i|^4}{\sum_j r_{ij}^3} \ge\sum_i |(Uv)_i|^4\ge\frac1N \] Criticality gives $X=X^*=U^*S$, hence $S=UX$ is the polar decomposition. \end{proof} \subsection{The modulus map and its defect} \begin{defn} A matrix $M$ is bistochastic when \[ M_{ij}\ge0,\qquad \sum_jM_{ij}=1,\qquad\sum_iM_{ij}=1 \] It is unistochastic when $M=p(U)$ for a unitary $U$, where \[ p(U)=(|U_{ij}|^2) \] The fiber over $M$ consists of all unitaries with these squared moduli. \end{defn} The modulus map forgets phases. A tangent variation of a bistochastic matrix has zero row and column sums, giving $(N-1)^2$ independent real coordinates. The differential $dp_U$ records which variations can be realized to first order by a curve of unitaries. \begin{defn} The defect of a full-support unitary is \[ d(U)=(N-1)^2-rank(dp_U) \] It measures the missing rank of the modulus map. \end{defn} This is the defect of \cite{TZ}. Nonflat critical points must lie in the singular part of the modulus map: \begin{prop}\label{defect} At a full-support nonflat critical point of $L$ on $U_N$, the modulus map $p(U)=(|U_{ij}|^2)$ has rank strictly less than $(N-1)^2$. \end{prop} \begin{proof} The affine space of bistochastic matrices has dimension $(N-1)^2$. Write $M=p(U)$ and $h(M)=\sum_{ij}\sqrt{M_{ij}}$. Concavity gives \[ dh_M(J_N-M)\ge h(J_N)-h(M)>0 \] At a critical point of $L=h\circ p$, the differential $dh_M$ vanishes on the image of $dp_U$. That image is therefore a proper subspace of the space of matrices with zero row and column sums. \end{proof} A direction in the kernel of $dp_U$ leaves the squared moduli fixed to first order. Along a unitary curve with this initial direction, $p'(0)=0$ gives \[ L''(0)=\frac12\sum_{ij}\frac{p''_{ij}(0)}{|U_{ij}|} \] At a critical point, this second variation depends only on the initial tangent vector. Finding a positive value in this kernel, after removing phases, is one route to the AHC. Section~\ref{fourphases} implements it at order 4. Averaging over the unit sphere of the horizontal kernel is another possible test, following \cite{BN}; a generally positive average is not established here. \subsection{Orders 1 to 3 and the lifting argument}\label{orderthree} \begin{defn} A full-support unitary is dephased when its first row and column are positive real. A path of unitaries $U(t)$ satisfying \[ p(U(t))=M(t) \] is called a lift of the modulus path $M(t)$. \end{defn} Row and column phases always give a dephased representative. The local lifting problem asks for representatives approaching the specified unitary, rather than another point of its fiber. \begin{thm}\label{threeflat} Every local maximum of $L$ on $U_N$ is flat for $N\le3$. \end{thm} \begin{proof} At order 1 the assertion is automatic. At order 2, full support and dephasing reduce the unitary to \[ \begin{pmatrix}\cos t&\sin t\\\sin t&-\cos t\end{pmatrix},\qquad 0h(M)$ yield \[ L(V_n)=h((1-t_n)M+t_nJ_3) \ge (1-t_n)h(M)+t_n h(J_3)>L(U) \] This contradicts local maximality. \end{proof} The argument needs both a segment in the image of $p$ and control of its lifts. Higher-order fibers can be more complicated: star-shapedness alone would leave the lifting question at a prescribed unitary. The direct variations below avoid this assumption. \section{Order 4: structural restrictions}\label{orderfour} We first prove Theorem A, then obtain equations for the remaining phase patterns and deduce Theorems B and C. Section~\ref{fourcurvature} then treats curvature and the numerical evidence. Throughout, $U$ has full support and $S_{ij}=U_{ij}/|U_{ij}|$. \subsection{A real submatrix forces flatness}\label{fourphases} For a column pair $p_iu_i,q_iv_i$, call the real lines through the phase quotients $h_i=\bar u_iv_i$ its phase lines; $h$ and $-h$ count as 1 line. For a full-support $2\times2$ block put \[ j(E)=Im(\bar E_{11}E_{12}E_{21}\bar E_{22}) \] Thus $j(E)=0$ exactly when phases make $E$ real. \begin{thm}\label{threecase} A local maximum on $U_4$ is flat if one column pair uses at most $3$ phase lines. Equivalently, every nonflat local maximum has $j(E)\ne0$ for all $36$ $2\times2$ submatrices. \end{thm} We separate the possibilities of 1, 2 and 3 phase lines. The basic variation is an imaginary rotation of 2 real columns. \begin{lem}\label{realup} A real unitary locally maximizing $L$ on the complex unitary group is flat. \end{lem} \begin{proof} Mix 2 columns by diagonal entries $\cos t$ and off-diagonal entries $i\sin t$. Their squared moduli $a^2,b^2$ in each row become $(1-s)a^2+sb^2,sa^2+(1-s)b^2$, where $s=\sin^2t$. Strict concavity of the square root increases their sum unless $a^2=b^2$. At a maximum every row therefore has constant moduli. \end{proof} The same rotation handles a pair whose relative phases lie on a single real line. The remaining columns determine the equality case. \begin{lem}\label{onepair} A nonflat full-support $U\in U_4$ with a collinear phase-quotient pair has an increasing arbitrarily small $2$-row or $2$-column rotation. \end{lem} \begin{proof} Make the 2 columns real. Lemma~\ref{realup}'s rotation works unless they are $(a_i),(\epsilon_i a_i)$ with $a_i>0$, $\epsilon_i=\pm1$. Each sign has squared mass $1/2$ and must occur twice: a singleton would force zeros in the remaining columns. Reorder to obtain $(a,b,c,d)^t,(a,b,-c,-d)^t$, with $a^2+b^2=c^2+d^2=1/2$. The remaining columns are $(e,f)W$, $W\in U_2$, where \[ e=\sqrt2(b,-a,0,0)^t,\qquad f=\sqrt2(0,0,d,-c)^t \] The first 2 rows are collinear in phase, as are the last 2. The same rotation increases $L$ unless $a=b$ and $c=d$, making all 4 numbers $1/2$. Orthogonality of $W$ makes the last 2 column quotients collinear. Mixing them increases $L$ unless their moduli also agree, which forces flatness by the row norms. Transposition gives the row version. \end{proof} At a critical point, 2 phase lines reduce to the preceding 1-line case through orthogonality and the first-order equations. \begin{lem}\label{twosingle} At a full-support critical point on $U_4$, a row pair with exactly $2$ phase lines forces a collinear column pair, and conversely by transposition. \end{lem} \begin{proof} The equations $\sum p_iq_ih_i=\sum(p_i-q_i)h_i=0$ split on the 2 lines, each containing 2 opposite phases. On each, products and differences agree, determining the positive moduli. Rephase the rows to \[ (a,a,b,b),\qquad(c,-c,dz,-dz),\qquad z\notin\mathbb R \] If columns 1 and 2 differ in modulus on the bottom rows, their 2 remaining differences have opposite signs. Stationarity makes their bottom phase quotients equal; orthogonality then makes these real, giving collinearity. The same applies to columns 3 and 4. Otherwise each bottom row has equal moduli within both pairs and lies in the span of $(b,b,-a,-a)$ and $(d,-d,-cz,cz)$. For its coefficients $\lambda,\mu$, the modulus equalities imply $Re(\bar\lambda\mu)=Re(z\bar\lambda\mu)=0$, so $\lambda\mu=0$. Columns 1 and 2 are again collinear. \end{proof} 3 phase lines require a block variation. We first identify 8 independent modulus coordinates for the differential of the modulus map. \begin{lem}\label{eightblocks} Write $U=\left(\begin{smallmatrix}A&B\\C&D\end{smallmatrix}\right)$ in $2\times2$ blocks. At full support, $j(B)j(C)\ne0$ makes $dp_U$ surjective onto the $4$ coordinates of $|A|^2$, the first column of $|B|^2$ and the first row of $|C|^2$. \end{lem} \begin{proof} The block $C$ is invertible, since rank 1 would give $j(C)=0$. The left skew-adjoint generator $\left(\begin{smallmatrix}0&Z\\-Z^*&0\end{smallmatrix}\right)$ gives $\dot A=ZC$, prescribing the first 4 coordinates. Keeping $A$ fixed, a last-column mixing changes the 2 entries of $|B_{i1}|^2$ by $2Re(w\bar B_{i1}B_{i2})$ up to sign. These forms are independent precisely when $j(B)\ne0$, and $C$ stays fixed. Mixing the last rows likewise prescribes the 2 $C$ coordinates. \end{proof} At a nonflat critical point, Proposition~\ref{defect} bounds the rank by 8. Under this lemma it is exactly 8, and these coordinates determine every vector in its image. Put $M=p(U)$ and $h(M)=\sum\sqrt{M_{ij}}$. \begin{lem}\label{blockcurve} Let $U$ be a full-support nonflat critical point with $j(A)=0$ and $j(B)j(C)\ne0$. Simultaneously exchange rows $1,2$ and $3,4$ and likewise columns, obtaining $\tilde M$. There is a unitary curve with $p'(0)=0$, $j(D)=0$, and \[ L''(0)=\sum_{\{x,y\}}\frac{(x-y)^2(x+y)}{xy} \] where the $8$ pairs $\{x,y\}$ are the exchanged entry moduli. \end{lem} \begin{proof} Make $A$ real. For $E^\#=\left(\begin{smallmatrix} \bar E_{22}&-\bar E_{21}\\-\bar E_{12}&\bar E_{11} \end{smallmatrix}\right)$, the operation respects products and adjoints, and $V^\#=\overline{\det V}V$ for $V\in U_2$. Polar decomposition and unitarity give fixed $V,W\in U_2$ with \[ B=(I-AA^*)^{1/2}V,\quad C=W(I-A^*A)^{1/2},\quad D=-WA^*V \] Conversely these formulas complete any strict contraction $A$ to a unitary. Use $A(t)=\cos t\,A+i\sin t\,A^\#$. The products $A(t)^*A(t),A(t)A(t)^*$ are even, since the cross products for real $A$ equal $(\det A)I$. Thus $B'(0)=C'(0)=0$, while \[ D(t)=\cos t\,D-i\sin t\,\eta D^\#,\qquad \eta=\det W\det V \] The 8 coordinates above have zero derivative; hence all of $p'(0)$ vanishes. In particular $\eta\overline{D_{11}D_{22}}$ and $\eta\overline{D_{12}D_{21}}$ are real, so $j(D)=0$. In the diagonal blocks, squared moduli follow exactly $\cos^2t\,M+\sin^2t\,\tilde M$. The difference $p''(0)-2(\tilde M-M)$ vanishes there and has zero row and column sums. Each other block is a multiple of $\left(\begin{smallmatrix}1&-1\\-1&1\end{smallmatrix}\right)$. Both variations lie in $im(dp_U)$: since $j(D)=0$, some last-column mixing fixes $|D|^2$ to first order, while $j(B)\ne0$ makes its change in $|B|^2$ nonzero. Transpose for $C$. Criticality kills these differences under $dh_M$; consequently $L''(0)=2dh_M(\tilde M-M)$, giving the formula by pairing terms. \end{proof} The curvature formula is positive unless all exchanged moduli agree. We now rule out this equality case at a nonflat critical point. \begin{lem}\label{blockequal} Under these hypotheses, $\tilde M=M$ is impossible. \end{lem} \begin{proof} The blocks $A,D$ would have equal opposite moduli and could be made real. A negative product of their entries would cancel their column inner product, forcing respectively $j(C)=0$ or $j(B)=0$. Make both blocks positive by independent phases; then $\eta=\pm1$. For squared moduli of a block set $f(X)=\sqrt{X_{11}X_{22}}-\sqrt{X_{12}X_{21}}$. Differentiate $\det D=\eta\overline{\det A}$ along an arbitrary unitary variation. Since $\eta'$ is imaginary, real parts give \[ df_{M_D}(\dot M_D)-\eta\,df_{M_A}(\dot M_A)=0 \] This functional is nonzero on the 9-dimensional bistochastic tangent space: its value on the checkerboard variation in $A$ is $-2\eta$. Its kernel is therefore exactly $im(dp_U)$. On blocks with equal opposite entries, the segment $M\to J_4$ scales both $f$ values by $1-t$. Thus $J_4-M$ belongs to this kernel, contradicting $dh_M(J_4-M)>0$. \end{proof} \begin{proof}[Proof of Theorem~\ref{threecase}] Lemmas~\ref{onepair} and~\ref{twosingle} settle 1 or 2 lines. For 3 lines place the repeated pair in $A$, so $j(A)=0$ and $j(C)\ne0$. If $j(B)=0$, the first 2 rows use at most 2 lines. Otherwise Lemmas~\ref{blockcurve} and~\ref{blockequal} give positive second derivative. Transposition treats row pairs. \end{proof} The block direction can mix all 4 columns; existence of an increasing single pair at every nonflat critical point remains open. For a 3-line column pair with singleton rows $a,b$, orthogonality and stationarity give $p_a^{-1}-q_a^{-1}=p_b^{-1}-q_b^{-1}$: project both equations onto a linear functional killing the repeated line. If every column pair uses at most 2 lines, any local maximum is flat by Theorem~\ref{threecase}. As a separate finite check, \texttt{phase\_grid.py} tests all $4^9=262144$ dephased fourth-root phase patterns: $192$ pass the necessary sign-balance test, comprising $168$ real non-Hadamard patterns, $6$ real and $18$ nonreal Hadamard patterns. Passing this test does not prove unitary realizability. \subsection{The equations for 4 distinct phase lines}\label{fourlinessection} We now assume that all $36$ quantities $j(E)$ from Theorem~\ref{threecase} are nonzero. The first-order equations already restrict the moduli in this case. Write $e=(1,1,1,1)^t$ and put \[ R_{ia}=\frac1{|U_{ia}|},\qquad P=I-\frac14ee^t, \qquad G=PRP \] Thus $G$ is obtained from the reciprocal moduli by subtracting the row and column means and adding the overall mean. \begin{lem}\label{distinctdifferences} At a nonflat critical point under our assumption, the $4$ numbers $R_{ia}-R_{ib}$ are pairwise distinct for every pair of columns $a,b$. The same holds for rows. \end{lem} \begin{proof} Put $z_i=\bar U_{ia}U_{ib}$ and $d_i=R_{ia}-R_{ib}$. Orthogonality and criticality give, respectively, \[ \sum_i z_i=0,\qquad \sum_i d_i z_i=0 \] Indeed, the second equation is the $(a,b)$ entry of $S^*U-U^*S=0$. If $d_1=d_2$, subtract $d_1$ times the first equation from the second. The vectors $z_3,z_4$ are linearly independent over $\mathbb R$, so all 4 differences agree. Since the 2 columns have equal squared norms and $r\to r^{-2}$ is strictly decreasing, their common difference is zero. In particular, for every 2 rows $i,j$, the row difference $R_{il}-R_{jl}$ agrees at $l=a,b$. Apply the same argument to $U^t$, which is also critical. All row differences are zero, and column normalization now gives $|U_{ia}|=1/2$ everywhere. \end{proof} The distinct reciprocal differences determine 2-dimensional planes for each column pair. Compatibility between these planes gives a rank restriction on the whole matrix. \begin{thm}\label{reciprocalrank} At a nonflat critical point of $L$ on $U_4$ with $4$ distinct phase lines for every pair, the centered reciprocal modulus matrix $G$ has rank exactly $2$. \end{thm} \begin{proof} Fix $a,b$ and put $J_{kl,ab}=Im(\bar z_k z_l)$. The 2 columns $Re(z),Im(z)$ span the orthogonal complement in $\mathbb R^4$ of $e,d$. Both planes have dimension 2 by our phase assumption and Lemma~\ref{distinctdifferences}. The oriented area forms of complementary 2-planes are Hodge dual up to a nonzero scalar. In coordinates their $2\times2$ minors therefore satisfy \[ J_{kl,ab}=c_{ab}\,\epsilon_{ijkl}(d_i-d_j) \] where $i,j,k,l$ are distinct and $\epsilon_{ijkl}$ is the sign of this permutation. The real number $c_{ab}$ is independent of the row indices. Exchanging $a,b$ negates both $J$ and $d$, hence $c_{ba}=c_{ab}$. For fixed $k,l,a$ we have $\sum_{b\ne a}J_{kl,ab}=0$, by orthogonality of rows $k,l$. For the complementary rows $i,j$, put $v_a=R_{ia}-R_{ja}$. The preceding identity gives \[ \sum_{b\ne a} c_{ab}(v_a-v_b)=0 \] Let $H$ have off-diagonal entries $-c_{ab}$ and diagonal entries $\sum_{b\ne a}c_{ab}$. This nonzero real symmetric matrix kills $e$ and every row difference of $R$, viewed as a column vector. It therefore kills their projections by $P$, which span the row space of $G$. This space is orthogonal to $e$, so $rank(G)\le2$. Rank zero contradicts Lemma~\ref{distinctdifferences}. In rank 1 we could write \[ R_{ia}=u_i+v_a+x_i y_a \] with all parameters real, by restoring the row and column means to $G$. The lemma forces the $x_i$ to be pairwise distinct, and likewise the $y_a$. Criticality says that $iS$ is tangent at $U$, since $U^*S$ is self-adjoint. Subtracting the phase tangents $i\,diag(u)U+iU\,diag(v)$ leaves $i\,diag(x)U\,diag(y)$ tangent. Thus $U^*\,diag(x)U$ commutes with $diag(y)$ and must be diagonal. Every column of $U$ would be an eigenvector of $diag(x)$. Its eigenvalues are simple, whereas every column has 4 nonzero entries. \end{proof} Interpolation at the 4 distinct reciprocal differences gives a polynomial identity involving only the moduli. \begin{prop}\label{pairequation} Write $p_i=|U_{ia}|$, $q_i=|U_{ib}|$, $d_i=p_i^{-1}-q_i^{-1}$ and $D_i=\prod_{j\ne i}(d_i-d_j)$. Under the assumptions of Theorem~\ref{reciprocalrank}, every column pair satisfies \[ \sum_{i=1}^4 p_i^2 q_i^2 D_i=0 \] The $6$ row pairs satisfy the same identity. \end{prop} \begin{proof} For $k=0,1,2$, interpolation of $t^k$ at the 4 distinct nodes $d_i$ gives $\sum_i d_i^k/D_i=0$ by comparison of cubic coefficients. The solutions of $\sum z_i=\sum d_i z_i=0$ therefore have the form \[ z_i=(A+B d_i)/D_i,\qquad A,B\in\mathbb C \] Indeed these vectors form a 2-dimensional complex space. Now $p_i^2q_i^2D_i^2=|A+B d_i|^2$ is the value of a quadratic at $d_i$. The coefficient of degree 3 in its interpolating polynomial is $\sum_i p_i^2q_i^2D_i$, giving the formula. \end{proof} The same interpolation also reconstructs the relative phases. It leaves only finitely many conjugation choices to check. \begin{prop}\label{phaseuniqueness} In the $4$-line case, prescribed moduli determine at most $1$ dephased nonflat critical unitary, up to complex conjugation. There is an explicit reconstruction test. \end{prop} \begin{proof} For each pair interpolate $p_i^2q_i^2D_i^2$ at $d_i$. By Proposition~\ref{pairequation} this gives a quadratic $q(t)=a_2t^2+a_1t+a_0$. 4 distinct phase lines require and, for that pair, are equivalent to \[ a_2>0,\qquad 4a_0a_2>a_1^2 \] Take $B=\sqrt{a_2}$ and $A=a_1/(2B)+i\sqrt{a_0-a_1^2/(4a_2)}$ in $z_i=(A+Bd_i)/D_i$. This gives all pair solutions up to common phase and conjugation. Pair column 1 with each other column and make their first entries positive. The 3 conjugation choices leave at most 4 candidates up to simultaneous conjugation; test their remaining orthogonality and stationarity equations. In fact at most 1 candidate survives. Let $v$ span $\ker G\cap e^\perp$. The symmetric matrix $H$ in the proof of Theorem~\ref{reciprocalrank} has this line as its image, so $c_{ab}=\lambda v_av_b$ with all factors nonzero. Choose the conjugate representative with $\lambda>0$. For each $b>1$, the sign \[ sgn(J_{12,1b}) =sgn\big(v_1v_b(R_{31}-R_{3b}-R_{41}+R_{4b})\big) \] fixes the conjugation choice. This proves uniqueness. In particular, symmetric moduli force $U^t=U$: the alternative $U^t=\bar U$ would make $U$ self-adjoint and give a real principal $2\times2$ phase product. \end{proof} \subsection{Common profiles and at most 4 moduli}\label{fewmoduli} The rank and pair equations have concrete consequences. We first handle 2 group patterns, then show that a common multiset reduces to them. \begin{thm}\label{groupmoduli} Let $U\in U_4$ be a local maximum of $L$. Suppose that after row and column permutations its moduli have the form $|U_{ia}|=f(a-i)$ on $\mathbb Z_4$ or on $\mathbb Z_2^2$. Then $U$ is flat. \end{thm} \begin{proof} If $U$ were nonflat, all $j(E)$ would be nonzero by Theorem~\ref{threecase}. The reciprocal moduli have the same symmetry. In the cyclic case write their first row as $(a,b,c,d)$. On the orthogonal complement of constants the eigenvalues of the circulant are $a-b+c-d$ and the conjugate pair $a-c\pm i(b-d)$. Rank 2 forces $a+c=b+d$. Write \[ (a,b,c,d)=(m+x,m+y,m-x,m-y) \] Read 2 adjacent rows cyclically, with reciprocal moduli $(a,b,c,d)$ and $(b,c,d,a)$. Their 4 differences are $(x-y,x+y,y-x,-x-y)$. Substitution in Proposition~\ref{pairequation} and multiplication by $a^2b^2c^2d^2$ gives \[ 64m^3xy(x-y)(x+y)=0 \] For example, the left side is the expansion of $c^2d^2D_1+d^2a^2D_2+a^2b^2D_3+b^2c^2D_4$. Since $m>0$, one of the 4 other factors must vanish. If $x=y$ or $x=-y$, 2 adjacent-row differences coincide. If $x=0$ or $y=0$, 2 opposite-row differences in $(2x,2y,-2x,-2y)$ coincide. Both conclusions contradict Lemma~\ref{distinctdifferences}. For $\mathbb Z_2^2$ the 3 nonconstant eigenvalues are $a+b-c-d$, $a-b+c-d$, $a-b-c+d$. One must vanish. Relabel the 3 nonidentity elements to arrange $a+b=c+d$. The rows $(a,b,c,d)$ and $(c,d,a,b)$ have differences $(a-c,b-d,c-a,d-b)$. The first and last coincide, giving the same contradiction. \end{proof} A common row and column profile can be decomposed into permutation matrices. At order 4 these recover exactly the 2 group patterns just considered. \begin{thm}\label{multisetmoduli} Suppose that each row and each column of a local maximizer on $U_4$ contains the same multiset of entry moduli. Then the unitary is flat. \end{thm} \begin{proof} For a modulus occurring $k$ times per row and column, join row $i$ to column $a$ when the entry has this modulus. The resulting bipartite graph is $k$-regular. Any set of rows has at least as many neighbors, by counting incident edges. Hall's theorem gives a perfect matching, whose removal lowers every degree by 1. Repeating this for each modulus partitions the array into 4 permutation matrices, each carrying a constant modulus. This step allows repetitions among the 4 moduli. Permute columns to make the first matching the identity. The other 3 permutations have no fixed points. If they are all involutions, they are the 3 double transpositions, which together with the identity act regularly as the Klein 4-group. Otherwise one is a 4-cycle $s$. At each row $i$ the remaining columns are $s^2(i)$ and $s^3(i)$. These edges form a single bipartite cycle of length 8. Indeed, traversing an $s^3$ edge and returning along an $s^2$ edge advances the row by $s$. Its only perfect matchings are the 2 alternating ones. Thus the 4 permutations are $1,s,s^2,s^3$. In either case we can label rows and columns by the group so the modulus at $(i,a)$ depends only on $a-i$. Theorem~\ref{groupmoduli} applies. \end{proof} In particular, a local maximizer on $U_4$ with at most 2 distinct entry moduli is flat. Indeed, if these moduli are $r\ne s$, the number $k$ of entries of modulus $r$ on any row or column satisfies $kr^2+(4-k)s^2=1$. This fixes $k$ throughout the matrix. With 3 distinct moduli, different row profiles can still occur. Normalization and the reciprocal-difference restrictions rule them out. \begin{prop}\label{threemoduli} A local maximum on $U_4$ having at most $3$ distinct entry moduli is flat. \end{prop} \begin{proof} It remains to consider 3 distinct squared moduli, which we call symbols. Suppose that a nonflat local maximum exists. By Theorem~\ref{threecase} and Lemma~\ref{distinctdifferences}, 2 rows can have equal symbols in at most 1 column. Also, wherever one row has a repeated symbol, another row must take different values at those positions. Both rules hold for columns. A constant row is impossible, since another row would need 4 distinct symbols. Suppose that one row has the form $(x,x,x,y)$. Each other row contains all 3 symbols in its first 3 positions. Normalization forces their fourth entries to have the same squared modulus $z$. Comparing the first row with the last column gives $3x+y=3z+y$, hence $x=z$. Comparing with any other row now gives $3x+y=x+y+w+x$, where $w$ is the third symbol. This forces $w=x$. We have ruled out triples in both rows and columns. The possible multiplicities in a row are now $(2,1,1)$ and $(2,2)$. 2 rows of the second kind must use the same pair $x,y$, since their sums agree and there are only 3 symbols. Under a row $(x,x,y,y)$ the second such row must put distinct entries in the first pair and in the last pair. The rows then agree twice. Thus at most 1 row has type $(2,2)$, and likewise for columns. All rows and columns of type $(2,1,1)$ repeat the same symbol, by normalization. If there is a row $(x,x,y,y)$, comparison with a row of the other type shows that its repeated symbol is the third symbol $z$ and that $x+y=2z$. There are 6 copies of $z$ in the matrix. Consequently exactly 1 column has type $(2,2)$ and contains no $z$. Delete this column and the row $(x,x,y,y)$. Each row and column of the remaining $3\times3$ array contains 2 copies of $z$. Any 2 of its rows share a $z$ in 1 column. In the deleted column 2 of these 3 rows must also agree, since only $x,y$ occur there. This contradicts the first rule. Transposition rules out a column of type $(2,2)$ as well. All rows and columns therefore have the same multiset, and Theorem~\ref{multisetmoduli} finishes the proof. \end{proof} 4 distinct moduli require a finite classification. The count constraints reduce the problem to $27$ patterns, each excluded by explicit polynomial identities. \begin{thm}\label{fourmoduli} Every local maximum of $L$ on $U_4$ with at most $4$ distinct entry moduli is flat. \end{thm} \begin{proof} Only 4 distinct moduli and unequal profiles remain. Label symbols $0,1,2,3$ in increasing reciprocal order, with squared moduli $s_0>s_1>s_2>s_3$. For a sorted row or column profile $w$, put $n_k(w)=\#\{w_i\ge k\}$ and $u=(s_0-s_1,s_1-s_2,s_2-s_3)>0$. Its squared norm is $4s_0-n(w)\cdot u$, so all 8 profiles have the same scalar product with $u$. For enumerating profiles, $u$ can be taken integral with entries at most $32$. In rank 2, cross 2 independent differences $n(w)-n(w_0)$: the coordinates are bounded by $2\cdot4^2$ and have one sign. In rank 1 a difference $d$ has both signs; assign the sum of negative absolute entries to its positive positions, the sum of positive entries to its negative positions, and $1$ to zeros. Rank zero is the common-profile case. This imposes no arithmetic restriction on the actual moduli. There are $35$ profiles. Group them by their scalar product for $u\in\{1,\ldots,32\}^3$, retaining groups with at least 2 profiles and all 4 symbols: $157$ groups remain, $93$ maximal. Assemble 4 words from each maximal group, requiring columns from the same word set. Reject row pairs repeating an ordered symbol pair or agreeing twice, by Lemma~\ref{distinctdifferences}; reject fewer than 4 symbols or a common profile. Canonicalize by sorting rows after every column permutation, also after transposition. Exactly $27$ arrays remain. Scale the reciprocal levels to $10$. This check excludes $10$ arrays by norms or determinant and $13$ by a pair equation. In 3 more, the surviving determinant factors are $3x-y-z-1$, $6x-4y-z-1$, $6x-y-4z-1$; each also requires $x^2(1+y+z)-yz=0$. Substitution for $x$ makes this strictly positive by the same coefficient test. The last array and determinant equation are \[ \begin{pmatrix}0&2&3&3\\2&1&2&3\\3&2&1&2\\3&3&2&0\end{pmatrix}, \qquad xz-x+y^2-3yz+y+z^2=0 \] Rescale the reciprocal levels to $r_3=1$ and put $h=1-r_0$, $t=(1-r_2)/h$. The determinant gives \[ (r_0,r_1,r_2,r_3)=(1-h,1-h(t+t^2),1-ht,1), \qquad 00$, all mixed differences are nonzero, and the claimed cubic bounds hold. Moreover $C_{11}\in(-0.002962,-0.002961)$ and $C_{44}\in(-0.000017406,-0.000017404)$; their inequality excludes a critical unitary by Proposition~\ref{realcubic}. All interval operations are in \texttt{curvature.py}. \end{proof} There is nevertheless a positive result near $J_4$. Write \[ H=\begin{pmatrix}1&1&1\\1&-1&-1\\-1&1&-1\\-1&-1&1\end{pmatrix}, \qquad F=(e\ H),\qquad G=HAH^t \] The 12 pair equations mean the equations of Proposition~\ref{pairequation} with weights $M_{ia}M_{ib}$; they make sense even when differences coincide. \begin{thm}\label{regularflatlimit} Let $M(t)$ be a $C^2$ path of positive bistochastic matrices with $M(0)=J_4$, $M'(0)\ne0$. Suppose for all small $t>0$ that $rank(G(t))\le2$, the $12$ pair equations hold, and all $16$ $C_{ab}(t)$ agree. Write $G(t)=tHA_0H^t+t^2HA_1H^t+o(t^2)$. After row and column permutations and possibly transposition, \[ A_0=\begin{pmatrix}a&b&0\\0&0&0\\0&0&0\end{pmatrix},\quad a^2+b^2>0,\qquad C(t)=32(a^4+b^4)t^4+o(t^4)>0 \] No unitary realization is assumed. \end{thm} \begin{proof} Normalization gives $R=2ee^t+tG_0+O(t^2)$ and $M=J_4-tG_0/4+O(t^2)$, where $G_k=HA_kH^t$. For each anchor let $T,Z$ be the mixed differences of $G_0,G_1$. Put $u_k=\frac14\sum T_{ij}^k$. The cubic coefficient is $E=u_3-\frac32u_1u_2+\frac12u_1^3$. For the matrix of these coefficients, direct expansion gives \[ (F^tEF/16)_{k0}=-96\prod_j(A_0)_{kj},\qquad (F^tEF/16)_{0k}=-96\prod_i(A_0)_{ik}\quad(k=1,2,3) \] They vanish, so every row and column of $A_0$ contains a zero. The leading coefficient of a column-pair equation is $32\prod_i(A_0v)_i$, with $v=e_j\pm e_k$ up to sign. Thus any 2 columns $u,v$ satisfy $\prod_i(u_i+v_i)=\prod_i(u_i-v_i)=0$. They have a common zero: otherwise 2 different positions have $u_i=\pm v_i\ne0$, forcing both third entries to vanish because each column contains a zero. Their zero sets intersect pairwise. Unless all 3 share a zero, they are the 3 distinct 2-element subsets, making $A_0$ a nonsingular monomial matrix. This contradicts $\det A_0=0$. Transposition gives a zero column as well. For the remaining block $\left(\begin{smallmatrix}a&b\\c&d\end{smallmatrix}\right)$, the $(1,1),(1,2),(2,1),(2,2)$ transformed cubic coefficients are $48bcd,48acd,48abd,48abc$. At most 2 entries survive, and then $E=0$. For completeness the fourth coefficient can be recovered without any asymptotic guess: expand the same cubic expression using \[ T(t)=tT+t^2Z,\qquad M(t)=J_4-tG_0/4 \] and retain degree 4. Call the resulting anchor matrix $D$. If the 2 entries occupy different rows and columns, put $A_0=diag(a,b,0)$. Then \[ (F^tDF/16)_{33}=-48ab(a^2+b^2)\ne0 \] independently of $A_1$, contradicting equality of the anchor values. The remaining support is 1 row or column. For the displayed $A_0$ and arbitrary $A_1$, the constant coefficient is \[ (F^tDF/16)_{00}=32(a^4+b^4) \] These polynomial identities are checked exactly in \texttt{flat\_limit.py}. The $C^2$ remainders contribute $o(t^4)$, giving the conclusion. \end{proof} This result concerns regular paths in the modulus equations. It gives neither a uniform punctured neighborhood nor a global exclusion. Paths with vanishing first derivative can require fractional powers after reparametrization and are not covered. A bounded exploratory search parametrized $G=AB^t$ with $\|G\|_F<\sqrt3$, balancing $R$ by additive row and column terms. The $300$ starts comprised $30$ fitting cubic consistency alone, $30$ fitting pair equations alone, and $120$ each combining cubic consistency with a negative-value target or with the pair equations. The last 2 minimum scaled residuals were $0.04015$ and $0.001853$, neither below $10^{-7}$. These are auxiliary residuals, not unitary criticality tests. \texttt{moduli\_search.py --verify-saved} rechecks every retained endpoint; these searches exclude no additional critical unitaries. It remains open whether normalization, reciprocal rank 2, the pair identities, cubic consistency and phase compatibility force positive common curvature at every nonflat 4-line critical point. A negative common cubic alone would not settle the full Hessian test. \subsection{2 exact calibration saddles}\label{fourcalibration} 2 of the stationary levels in the next subsection have explicit representatives. For the first, take $a=\cos(\pi/8)/\sqrt2$ and $b=\sin(\pi/8)/\sqrt2$. An exact representative is \[ U=\begin{pmatrix} a&b&a&b\\ a&-b&-b&-a\\ b&a&-a&b\\ b&-a&-b&a \end{pmatrix} \] The identities $2a^2+2b^2=1$ and $a^2-2ab-b^2=0$ give $UU^t=I$. Multiplication gives $S^tU$ symmetric, proving criticality. We have $L(U)=8(a+b)=4\sqrt{2+\sqrt2}$. The imaginary column rotation in Lemma~\ref{realup} increases $L$. Along a real rotation of the same columns the signs are locally constant, so the pair contributes $C\cos t+D\sin t$ with $C>0$ and $D=0$ by criticality. This gives a decrease. For the highest observed nonflat level, use a bordered circulant matrix. For $00 \] This proves the saddle assertion independently of numerical Hessian diagonalization. The 2 intermediate levels in the numerical table remain uncertified as exact critical values. The 2 representatives are matched by \texttt{models.py}; \texttt{three\_lines.py} also checks the block variation and certifies column curvature $M_{23}\in(0.6092,0.6094)$, row curvature $M_{14}\in(-0.6677,-0.6674)$ and block-curve curvature in $(1.4478,1.4480)$ on partition $12\mid34$ for this calibration. Thus the singleton rows of a 3-line pair need not themselves give an increase. \subsection{Numerical evidence and its limits}\label{foursearch} A useful bound separates potential local maxima from the nonsmooth boundary. We use Frobenius-unit generators as in Lemma~\ref{twocol}. \begin{prop}\label{entrybound} At a local maximum on $U_N$, $N\ge2$, every entry has modulus at least \[ b_N=\sqrt{N(N-1)^2+1}-(N-1)\sqrt N \] In particular, $b_4=\sqrt{37}-6$ and $b_5=9-4\sqrt5$. \end{prop} \begin{proof} For an entry of modulus $p$, choose a partner of modulus $q$ in its row with $q^2\ge(1-p^2)/(N-1)$. Choose the phase of the generator in Lemma~\ref{twocol} to agree with $h_i$ in the selected row. Its contribution to the first term of the curvature is then $(q^2/p+p^2/q)/2$. Each column has $1$-norm at most $\sqrt N$, giving \[ M_{jk}\ge\frac{q^2}{2p}-\sqrt N \ge\frac{1-p^2}{2(N-1)p}-\sqrt N \] At a local maximum $M_{jk}\le0$. Solving the resulting quadratic inequality gives the bound. \end{proof} We use the diagnostics \[ g(U)=\frac12\|S^*U-U^*S\|_F,\qquad d(U)=\max_{ij}\big||U_{ij}|-1/2\big| \] 2 Apple M3 laptops (8 GB, 3 workers each) ran 7-hour campaigns, with seeds $421701,421702$. Ascent uses Haar starts, geodesic gradient steps and Armijo backtracking, capped at $1800$ iterations. Stationary search uses damped least squares for $(S^*U-U^*S)/(2i)=0$. 4 protocols start from Haar, phase-polar, near-real and near-flat matrices; 4 first steer Haar matrices towards $L\le c$, $c\in\{7.40,7.65,7.90,7.995\}$, with at most $24$ steering and $120$ root steps. The tolerances are $d\le10^{-6}$ for flatness, $g\le10^{-12}$ for stationarity, and $10^{-8}$ for near-zero entries. At accepted nonflat returns both full and 9-dimensional horizontal Hessians are checked; top horizontal eigenvalue at most $10^{-7}$ triggers a candidate flag. \begin{center} \begin{tabular}{lrr} \toprule Endpoint status & Ascent & Stationary\\ \midrule Numerically flat &83\,009\,628&19\,409\,961\\ Nonflat with positive curvature &0&26\,672\,800\\ Near zero &0&6\,126\,915\\ Unresolved &67\,268&27\,748\\ \midrule Total &83\,076\,896&52\,237\,424\\ \bottomrule \end{tabular} \end{center} There were no candidate, unlisted-level, numerical-error or unitarity-error flags. All nonflat stationary returns had $L$ within $2\cdot10^{-6}$ of one of the following 4 values. Repeated visits to equivalent matrices are included in the counts. \begin{center} \begin{tabular}{rrr} \toprule $L$ (rounded) & Returns & Minimum top horizontal eigenvalue\\ \midrule 7.391036 &6\,764\,846 &2.2961006\\ 7.633239 &9\,497\,039 &2.8149691\\ 7.633377 &7\,875\,896 &2.6601070\\ 7.637615 &2\,535\,019 &2.8768867\\ \bottomrule \end{tabular} \end{center} Every unresolved ascent endpoint satisfies $d\le6.01386943033244\cdot10^{-6}$; not all were continued, so this does not assert convergence of every trajectory. For stationary returns, \[ \max_{\text{unresolved }U}\min_{ij}|U_{ij}| =5.6974893208468939\cdot10^{-7} \] All unresolved and near-zero stationary endpoints thus fall far below $\sqrt{37}-6=0.08276\ldots$. This comparison has a large numerical margin but is not an interval certificate for the entire campaign. The initial $8192$-start survey gave $3596$ nonflat uphill, $4002$ near-flat, $589$ near-zero and $5$ unresolved returns. The 4 nonflat groups have numerical positive indices $3,3,2,1$. For every nonflat return, interval arithmetic proves a recorded pair curvature exceeds $0.60$ throughout its coordinate box of radius $10^{-8}$. This excludes local maxima in the boxes without proving existence or uniqueness of critical points there. The other $594$ endpoints have a smallest entry at most $1.18\cdot10^{-8}$. A separate $1024$-start polar-ascent control reached numerical flatness, including $11$ final least-squares refinements. The compact archives retain all $95\,016$ unresolved overnight endpoints, aggregate counts and extrema, and representative outputs. Independent audits recheck $99\,164$ retained endpoints and all $98\,914$ stored spectrum pairs. The sources are \texttt{search.py} and \texttt{audit.py} for the initial survey and its certificates, and \texttt{ahc\_ascent.py} and \texttt{ahc\_stationary.py} for the campaigns. \texttt{compact.py} exports and verifies their archives; \texttt{audit\_compact.py} recomputes all retained diagnostics. The archives contain their exact generating sources and complete aggregate counts, but ordinary discarded endpoints are unavailable. These searches sample basins, not stationary points uniformly; repeated visits are not independent evidence of completeness. No covering of $U_4$ is claimed, and an unvisited basin remains possible. \section{Order 5: exact models and gradient geometry}\label{fivegeometry} \subsection{A saddle invisible to pair tests} The order-5 search needs the full Hessian. Its main calibration shows why pair rotations alone cannot decide local maximality. For $00$ and $F'(x)\to-\infty$ as $x\to1$, the root is unique. It is $x_0=0.5374606822272723\ldots$, with $L(U_0)=11.0911152536\ldots$. Index the border by $0$ and the inner coordinates by $1,2,3,4$. Cyclic symmetry leaves 3 column-pair types; self-adjointness gives the same row tests. Lemmas~\ref{variations} and~\ref{twocol}, evaluated with outward rounding, give \begin{center} \begin{tabular}{lc} \toprule Direction & Curvature enclosure\\ \midrule Border--inner pair & $-0.85024500$; reality of $Tr(S^*U)$ then makes $v_0$ real, and the remaining skew entry is $\sqrt{1-x^2}\,\partial_x L$. Restricted phase searches alone do not give this conclusion. \subsection{The actual unstable branches} Return to the exact order-5 model $U(x)$ from Proposition~\ref{fivecalibration}. Replace its inner row by \[ (-a+\frac i2\sin t,\ c-\frac i2\cos t, -a-\frac i2\sin t,\ c+\frac i2\cos t) \] to obtain $U(x,t)$. Its nonconstant Fourier eigenvalues are $e^{it},-1,-e^{-it}$. Put $A=(a^2+\sin^2t/4)^{1/2}$, $C=(c^2+\cos^2t/4)^{1/2}$ and $H(x,t)=x+4\sqrt{1-x^2}+8(A+C)$. This surface is the local common fixed set of $V\to PVP^{-1}$ and $V\to\bar V P^2$, where $P$ cycles the 4 inner coordinates. The first equation gives a circulant block; the second makes the border block real and imposes the 3 displayed Fourier phases. Both maps preserve $L$, so its gradient is tangent to the surface. The induced metric and gradient equations are \[ \frac{2\,dx^2}{1-x^2}+2\,dt^2,\qquad \dot x=\frac{1-x^2}{2}H_x,\quad \dot t=\frac12H_t, \quad H_t=\sin(2t)(1/A-1/C) \] A fixed-corner $t$-rotation is a different curve: its corner stays $x_0$, and it cannot become flat. Starting exactly at $U_0$ also gives no motion. The proposition concerns the actual nonconstant gradient trajectories with backward limit $U_0$. \begin{prop}\label{branches} The $2$ unstable branches through $U_0$ for gradient ascent converge to $U(1/\sqrt5,\pm t_*)$, where $t_*=\frac12\arccos(1/\sqrt5)$. Both limits are Fourier-equivalent. \end{prop} \begin{proof} The restricted Hessian has opposite signs in $x,t$, and the full horizontal Hessian has exactly 1 positive eigenvalue. The unstable manifold theorem \cite{T}, on the phase quotient, therefore identifies these 2 surface branches with the full unstable branches. Equivariance and the trivial effective phase stabilizer fix their lifts through $U_0$. Consider the positive branch in $00.0557$ and hence cannot themselves be local maxima. The unresolved returns have residual at least $1.254\cdot10^{-5}$, with smallest-entry moduli reaching $0.1539657$; the entry bound does not dismiss all of them. All $288$ retained unresolved endpoints have numerically positive geodesic curvature, with smallest top value $3.67446\ldots$. The saved-point follow-up below explains these stalls. A subsequent partial replay addresses the original starting sequence; neither calculation asserts exhaustive coverage. The archives \texttt{code011/overnight5/ascent-mac.zip} and \texttt{stationary-mac.zip} contain the generating sources, all-run counters and extrema, and $96$ and $1076$ retained endpoints. Ordinary endpoints use bounded sampling, supplemented by priority pools for weak saddles and small unresolved residuals; every candidate flag would be retained. The 2 databases total $4\,460\,544$ bytes. \texttt{audit\_ahc5.py} rechecks source hashes, counters, starts and retained endpoints, including $452$ stored pairs of full and horizontal spectra and their uphill witnesses. Maximum spectral and witness discrepancies were below $6\cdot10^{-14}$. The audit also computes curvature at the retained unresolved points without refining them. The unretained endpoints cannot be audited individually from these archives. \subsection{Catalogue and systematic certificates} The $452$ retained nonflat stationary returns form $40$ numerical groups under row and column phases, permutations and complex conjugation, using Frobenius tolerance $10^{-6}$. The comparison checks all pairs of row and column permutations after dephasing; equal norms alone are not used as equivalence tests. The table gives each representative norm, its positive horizontal index $p$, and a lower bound $q$ for a fixed uphill direction throughout every retained rounding box in that group. Theorem~\ref{criticalcatalogue} certifies these indices and the displayed norms at exact critical points near the $40$ representatives. \begin{center} \begin{tabular}{rrrr@{\qquad}rrrr} \toprule Group & $L$ & $p$ & $q$ & Group & $L$ & $p$ & $q$\\ \midrule 1 & 9.779960 & 6 & 17.279 & 21 & 10.475603 & 6 & 3.673\\ 2 & 9.944272 & 6 & 3.572 & 22 & 10.475618 & 5 & 4.844\\ 3 & 10.204351 & 7 & 5.864 & 23 & 10.475618 & 5 & 4.844\\ 4 & 10.204611 & 6 & 4.582 & 24 & 10.475787 & 4 & 9.810\\ 5 & 10.204611 & 6 & 4.582 & 25 & 10.493467 & 4 & 8.293\\ 6 & 10.229297 & 6 & 6.638 & 26 & 10.599732 & 5 & 4.534\\ 7 & 10.231276 & 5 & 7.045 & 27 & 10.656854 & 5 & 6.094\\ 8 & 10.234788 & 5 & 3.557 & 28 & 10.658120 & 4 & 6.675\\ 9 & 10.252396 & 6 & 10.519 & 29 & 10.659153 & 4 & 7.120\\ 10 & 10.255519 & 6 & 3.761 & 30 & 10.662660 & 3 & 4.602\\ 11 & 10.264114 & 6 & 3.507 & 31 & 10.662660 & 3 & 4.602\\ 12 & 10.282845 & 6 & 4.288 & 32 & 11.000000 & 5 & 0.499\\ 13 & 10.283393 & 5 & 3.930 & 33 & 11.047758 & 4 & 0.539\\ 14 & 10.286257 & 5 & 6.297 & 34 & 11.048168 & 3 & 0.621\\ 15 & 10.290855 & 5 & 5.992 & 35 & 11.048247 & 2 & 0.652\\ 16 & 10.290855 & 5 & 2.989 & 36 & 11.061538 & 2 & 0.292\\ 17 & 10.322914 & 4 & 36.252 & 37 & 11.079258 & 3 & 0.740\\ 18 & 10.325913 & 4 & 6.172 & 38 & 11.080767 & 2 & 0.322\\ 19 & 10.334241 & 4 & 8.302 & 39 & 11.081083 & 1 & 0.206\\ 20 & 10.334241 & 4 & 8.302 & 40 & 11.091115 & 1 & 0.653\\ \bottomrule \end{tabular} \end{center} For every one of the $452$ matrices, \texttt{catalogue5.py} rounds a self-adjoint uphill generator to 12 decimals and interval-evaluates $Q_U(B)$ on the real and imaginary coordinate box of radius $10^{-8}$. Every lower bound exceeds $.20634$. These are rigorous exclusions of local maxima among unitaries in the boxes. The grouping remains numerical; these box exclusions alone do not establish critical-root existence. That additional conclusion is proved below for the $40$ representatives. No completeness of the catalogue is asserted. Representatives, matching permutations and individual certificates are in \texttt{data/catalogue5.npz} and \texttt{data/catalogue5.json}. Allowing transposition merges groups $4/5$, $19/20$, $22/23$ and $30/31$ numerically, with matching errors below $6\cdot10^{-11}$, leaving $36$ groups. Groups $15,16$ have norms agreeing to $12$ decimals, $L=10.290855136357$, but different smallest entry moduli, $0.109049\ldots$ and $0.211227\ldots$. These comparisons are also checked by \texttt{catalogue5.py}; exact equality of these 2 norms is not asserted. \subsection{Stalled stationary searches}\label{fivestalls} Write $R(U)=(S^*U-U^*S)/(2i)$ and $\Phi(U)=\|R(U)\|_F^2/2$. Restarting the residual solver on the $288$ saved points, with damping reset to $.01$ and up to $1600$ further steps, gives 3 outcomes: \begin{center} \begin{tabular}{rll} \toprule Count & Outcome & Diagnostic\\ \midrule $139$ & Stationary saddles & $8$ at $L=10.475603$, $131$ at $L=10.475787$\\ $134$ & Positive residual minimum & $L=10.2696684693$, $\|R\|_F=0.0164775744$\\ $15$ & Near the zero-entry boundary & Smallest entry below $4.865\cdot10^{-7}$\\ \bottomrule \end{tabular} \end{center} All $139$ in the first group had exhausted the original $160$ steps and missed the polishing trigger $\|R\|_F<10^{-5}$. Restarting needs only $7$--$29$ steps, giving residual below $10^{-10}$ and top curvature above $3.67$. The $15$ boundary points already had an entry below $4.269\cdot10^{-5}$ before restarting, far below $b_5$. The $134$ middle returns agree with one matrix up to phases, permutations and conjugation within $10^{-6}$. At each, replace $U$ by \[ U\exp\big(-0.6iR(U)/\|R(U)\|_F\big) \] and allow $800$ root steps. All reach the known level $L=10.2641137511\ldots$, with top curvature above $3.50$ and residual below $10^{-10}$. Perturbations of lengths $.1$ and $.3$ returned to the residual minimum. The 3 final saddle types match already retained representatives within $10^{-6}$. These are numerical observations, reproduced by \texttt{stalled5.py}, which saves the matrices, equivalences and witnesses in \texttt{data/stalled5.npz} and \texttt{data/stalled5.json}. The identity $\nabla\Phi=Hess(L)\nabla L$ explains how the residual can be stationary while $L$ is not. The positive minimum is certified in Proposition~\ref{residualminimum}. \subsection{Daytime follow-up and targeted search}\label{fivefollowup} The program \texttt{ahc5\_day\_replay.py} replays seed $521812$ with the original batches and solver. Unresolved full-support returns above $b_5-10^{-6}$ receive $320$ restart steps, then if necessary the length-$.6$ perturbation and up to $800$ root steps. Entry-bound exclusions and nonstationary positive curvature are counted separately. The complementary \texttt{ahc5\_day\_search.py}, seed $521819$, uses 8 protocols: root search near all $40$ representatives, near the 8 weakest, and along their soft directions; perturbed bordered and circulant starts; geodesic and polar ascent near weak saddles; and Haar root controls. Amplitudes range from $10^{-4}$ to $1$. Both jobs check full and horizontal Hessians and retain new signatures. Each Mac ran for 7 active hours with 2 workers and 1 BLAS thread per worker. Unresolved identifiers and diagnostics and all candidate flags are retained, with bounded pools for ordinary matrices. Data growth stops at $112$ MiB, below the $128$ MiB cap. Baseline replay, standalone launches, rollback, eigenbasis compatibility, locks and exact resumption are checked by \texttt{test\_day5.py}. The $18$-second release checks ($5600$ replay and $10\,032$ targeted starts, no unresolved or candidate returns) are prefixes, not additional starts. The completed runs give the following counts; the lower block partitions the initial stalls already included above. \begin{center} \begin{tabular}{lrr} \toprule Endpoint status & Replay & Targeted search\\ \midrule Numerically near flat &1\,696\,013&10\,068\,469\\ Nonflat stationary, positive curvature &13\,672\,742&24\,303\,512\\ Small-entry exclusion &820\,253&579\,411\\ Unresolved or candidate flags &0&0\\ \midrule Total starts &16\,189\,008&34\,951\,392\\ \midrule Initially unresolved &284\,297&235\,667\\ Restart alone gives an uphill saddle &12\,306&15\,090\\ Escape and refinement give an uphill saddle &248\,455&146\,623\\ Excluded by the entry bound &23\,536&73\,954\\ \bottomrule \end{tabular} \end{center} No uncertain-Hessian, unitarity or numerical-error flags occurred. Replay covers $70.59\%$ of the original $22\,933\,376$ starts; $6\,744\,368$ remain unvisited, so resolution of all $404\,581$ original stalls is not claimed. Its $284\,297$ stalls are those observed on replay, and neither replay nor follow-up iterates add independent trials. All $8\,737\,840$ targeted ascent returns are near flat. Every accepted nonflat stationary return matches a stored signature, using sorted moduli and horizontal spectra within $2\cdot10^{-5}$; both runs encounter all $40$, with no new signature and least top curvature $0.206350429\ldots$. Signature matching is screening, not an equivalence proof or interval certificate. The unchanged archives \texttt{code011/daytime/replay-mac.zip} and \texttt{search-mac.zip} contain $788$ and $569$ retained runs, with starting, raw and final matrices. Database sizes are $31.94,26.03$ MiB; sums of parent and worker peak memory are $128.5,157.2$ MiB. The program \texttt{audit\_day5.py} checks hashes, database integrity, counters, retained matrices and $452+356$ full/horizontal spectrum pairs and uphill witnesses. Polarization independently checks each full Hessian; all spectral and witness errors are below $6\cdot10^{-14}$. The adjacent \texttt{replay-mac-audit.json} and \texttt{search-mac-audit.json} record these checks, which do not audit unretained matrices. All $788$ saved replay starts and $496$ targeted starts regenerate within $10^{-11}$. The other $67$ soft-direction and $6$ ascent starts depend on eigenbasis choices. Recover $B=-i\log(U_0^*U)$ from the saved start $U$ and catalogue centre $U_0$. For a soft start of length $\rho$, Gaussian coefficients $c_1,c_2,c_3$ and selected eigenvalues $\lambda_1,\lambda_2,\lambda_3$, the audit checks \[ \|P_\lambda B\|_F^2 =\rho^2\frac{\sum_{a:\lambda_a=\lambda}c_a^2} {\sum_{a=1}^3c_a^2} \] Here $P_\lambda$ is the horizontal Hessian eigenspace projection; eigenvalues within $10^{-7}$ are grouped. For ascent starts it checks $B=\rho(W+.1E)/\|W+.1E\|_F$, where $E$ is the regenerated noise and $W$ a unit vector in the top eigenspace. Relative norm and projection discrepancies are below $2\cdot10^{-12}$. These are numerical compatibility checks, not identical regeneration across numerical libraries; saved starts and generating sources are unchanged. \subsection{A positive minimum of the stationarity residual} The following certificate explains the middle group in Section~\ref{fivestalls}. It establishes an obstruction to residual minimization, without producing a local maximum of $L$. \begin{prop}\label{residualminimum} On $U_5$, $\Phi$ has a strict local minimum modulo row and column phases at a full-support point $V$ satisfying \[ .01647<\|R(V)\|_F<.01649 \] In particular, this point is not stationary for $L$. \end{prop} \begin{proof} The following integer matrix specifies a rational self-adjoint matrix $A$: put $A_{jj}=10^{-12}C_{jj}$ and $A_{jk}=10^{-12}(C_{jk}+iC_{kj})$ for $j.1273$ and the displayed bounds on $\|R(V(x))\|_F$. Phase invariance proves the asserted strict local minimum modulo phases. \end{proof} \subsection{Certified critical points}\label{criticalpoints} We finish by validating the critical points behind the catalogue. The certificate concerns local existence, uniqueness and inertia; the numerical grouping is not used to infer completeness or pairwise inequivalence of all $40$ entries. \begin{thm}\label{criticalcatalogue} Each of the $40$ catalogue representatives is within $10^{-7}$ in Frobenius norm of a full-support critical point of the $1$-norm on $U_5$. This point is locally unique modulo row and column phases, and its full Hessian has inertia \[ (16-p,9,p) \] where $p$ is the positive index listed in the table. Its norm rounds to the displayed value. In particular, the $1$-norm has at least $35$ distinct nonflat critical values. For group $39$, all row and column pair curvatures are less than $-.22$. \end{thm} \begin{proof} The rational chart data are listed in \texttt{data/critical5-input.json}. For each representative an integer matrix $C$ defines a self-adjoint matrix $A$ by $A_{ii}=10^{-20}C_{ii}$ and $A_{ij}=10^{-20}(C_{ij}+iC_{ji})$ for $i.016,\qquad f(x)<11.1<5\sqrt5,\qquad \|V(x)/z-U^{(0)}\|_F<10^{-7} \] Here $U^{(0)}$ is the corresponding stored representative. Thus all points in the chart have full support and are nonflat. The derivatives used by \texttt{critical5.py} are \[ V_a=2iWB_aW,\qquad V_{ab}=-2W(B_aWB_b+B_bWB_a)W \] Writing $m_{ij}=|V_{ij}|$ and $t_{a,ij}=Re(\overline{V_{ij}}(V_a)_{ij})$, these give \[ f_a=\sum_{ij}\frac{t_{a,ij}}{m_{ij}},\qquad f_{ab}=\sum_{ij}\left( \frac{Re\big(\overline{(V_a)_{ij}}(V_b)_{ij} +\overline{V_{ij}}(V_{ab})_{ij}\big)}{m_{ij}} -\frac{t_{a,ij}t_{b,ij}}{m_{ij}^3}\right) \] Put $H(x)=D^2f(x)$. At $50$ decimal digits, interval arithmetic verifies the following bounds for all $40$ charts: \[ \|Y\nabla f(0)\|_\infty<10^{-20},\qquad \|I-YH(x)\|_\infty<.002,\qquad \|P^tH(x)P-D\|_\infty<.001 \] Here $Y,P$ are rational matrices, proposed by midpoint inversion and diagonalization and rounded to $17$ significant digits. The diagonal matrix $D$ has $16-p$ entries $-1$ and $p$ entries $1$. The first 2 bounds make $x-Y\nabla f(x)$ a contraction of the cube strictly into itself, with a unique zero of $\nabla f$. The third fixes the inertia of $H$ by congruence and a nonsingular homotopy to $D$; numerical eigenvalue signs alone are not used as proof. To pass to full criticality, form the real $25\times25$ coordinate matrix $T(x)$ of the $16$ self-adjoint tangent generators $-iV^*V_a$, the $5$ generators $E_{jj}$, and the $4$ generators $V^*E_{jj}V$ with $j<5$. Coordinates are diagonal entries followed by real and imaginary parts of upper entries. A rational preconditioner gives \[ \|I-ZT(x)\|_\infty<10^{-6} \] Thus the chart is transverse to the phase orbits. Phase invariance turns its critical point into a critical point on $U_5$, locally unique modulo phases. At that point the $9$ phase directions lie in the full Hessian kernel, while its restriction to the chart has the certified nondegenerate inertia. This proves the stated full inertia. Finally, the norm enclosure for each chart lies strictly within $5\cdot10^{-7}$ of the displayed value. The table contains $35$ different such values, whose open rounding intervals are disjoint. For group $39$, interval evaluation of Lemma~\ref{twocol} on $V(x)$ and its transpose bounds all $20$ pair curvatures above by $-.22$. The individual checks and enclosures are recorded in \texttt{data/critical5.json} and reproduced by \texttt{critical5.py}. \end{proof} \phantomsection\addcontentsline{toc}{section}{References} \begin{thebibliography}{99} \setlength{\itemsep}{2pt} \setlength{\parsep}{0pt} \bibitem{AMS} P.-A. Absil, R. Mahony and R. Sepulchre, Optimization Algorithms on Matrix Manifolds, Princeton University Press (2008). \bibitem{AMM} G. Auberson, A. Martin and G. Mennessier, On the reconstruction of a unitary matrix from its moduli, \emph{Commun. Math. Phys.} \textbf{140} (1991), 417--437. \bibitem{BCS} T. Banica, B. Collins and J.-M. Schlenker, On orthogonal matrices maximizing the $1$-norm, \emph{Indiana Univ. Math. J.} \textbf{59} (2010), 839--856. \bibitem{BNp} T. Banica and I. Nechita, Almost Hadamard matrices: the case of arbitrary exponents, \emph{Discrete Appl. Math.} \textbf{161} (2013), 2367--2379. \bibitem{BN} T. Banica and I. Nechita, Almost Hadamard matrices with complex entries, \emph{Adv. Oper. Theory} \textbf{3} (2018), 137--177. \bibitem{BNS} T. Banica, I. Nechita and J.-M. Schlenker, Submatrices of Hadamard matrices: complementation results, \emph{Electron. J. Linear Algebra} \textbf{27} (2014), 197--212. \bibitem{BNZ} T. Banica, I. Nechita and K. \.Zyczkowski, Almost Hadamard matrices: general theory and examples, \emph{Open Syst. Inf. Dyn.} \textbf{19} (2012), 1250024. \bibitem{BKP} N. Bebiano, A. Kova\v{c}ec and J. da Provid\^encia, The validity of the Marcus--de Oliveira conjecture for essentially Hermitian matrices, \emph{Linear Algebra Appl.} \textbf{197--198} (1994), 411--427. \bibitem{BMP} N. Bebiano, J.K. Merikoski and J. da Provid\^encia, On a conjecture of G.N. de Oliveira on determinants, \emph{Linear Multilinear Algebra} \textbf{20} (1987), 167--170. \bibitem{BebN} N. Bebiano and H. Nakazato, Critical loci of the $4\times4$ unistochastic map, DMUC Preprint 25-03 (2025). \bibitem{BP} N. Bebiano and J.P. da Provid\^encia, Revisiting the Marcus--de Oliveira conjecture, \emph{Mathematics} \textbf{13} (2025), 711. \bibitem{BEKTZ} I. Bengtsson, \AA. Ericsson, M. Ku\'s, W. Tadej and K. \.Zyczkowski, Birkhoff's polytope and unistochastic matrices, $N=3$ and $N=4$, \emph{Commun. Math. Phys.} \textbf{259} (2005), 307--324. \bibitem{BW} S. Brierley and S. Weigert, Maximal sets of mutually unbiased quantum states in dimension 6, \emph{Phys. Rev. A} \textbf{78} (2008), 042312. \bibitem{D} S.W. Drury, OMC for scalar multiples of unitaries, \emph{Linear Algebra Appl.} \textbf{422} (2007), 318--325. \bibitem{DZ} C. Dunkl and K. \.Zyczkowski, Volume of the set of unistochastic matrices of order 3 and the mean Jarlskog invariant, \emph{J. Math. Phys.} \textbf{50} (2009), 123521. \bibitem{F} M. Fiedler, Bounds for the determinant of the sum of Hermitian matrices, \emph{Proc. Amer. Math. Soc.} \textbf{30} (1971), 27--31. \bibitem{M} M. Marcus, Derivations, Pl\"ucker relations, and the numerical range, \emph{Indiana Univ. Math. J.} \textbf{22} (1973), 1137--1149. \bibitem{MM} J.I. Mulero-Mart\'inez, A variational framework for determinantal inequalities of normal matrices: successes and obstructions, \emph{Linear Algebra Appl.} \textbf{740} (2026), 19--38. \bibitem{O} G.N. de Oliveira, Normal matrices (research problem), \emph{Linear Multilinear Algebra} \textbf{12} (1982), 153--154. \bibitem{PB} J. da Provid\^encia and N. Bebiano, Matrices satisfying a conjecture of G.N. de Oliveira on determinants, \emph{Linear Algebra Appl.} \textbf{78} (1986), 187--198. \bibitem{RGZ} G. Rajchel, A. G\k{a}siorowski and K. \.Zyczkowski, Robust Hadamard matrices, unistochastic rays in Birkhoff polytope and equi-entangled bases in composite spaces, \emph{Math. Comput. Sci.} \textbf{12} (2018), 473--490. \bibitem{RKKZ} G. Rajchel-Mieldzio\'c, K. Korzekwa, Z. Pucha\l a and K. \.Zyczkowski, Algebraic and geometric structures inside the Birkhoff polytope, \emph{J. Math. Phys.} \textbf{63} (2022), 012202. \bibitem{R} K. Rodtes, A class of normal dilation matrices affirming the Marcus--de Oliveira conjecture, \emph{Oper. Matrices} \textbf{15} (2021), 127--130. \bibitem{TZguide} W. Tadej and K. \.Zyczkowski, A concise guide to complex Hadamard matrices, \emph{Open Syst. Inf. Dyn.} \textbf{13} (2006), 133--177. \bibitem{TZ} W. Tadej and K. \.Zyczkowski, Defect of a unitary matrix, \emph{Linear Algebra Appl.} \textbf{429} (2008), 447--481. \bibitem{T} G. Teschl, Ordinary Differential Equations and Dynamical Systems, Graduate Studies in Mathematics \textbf{140}, American Mathematical Society (2012). \bibitem{ZKSS} K. \.Zyczkowski, M. Ku\'s, W. S\l omczy\'nski and H.-J. Sommers, Random unistochastic matrices, \emph{J. Phys. A} \textbf{36} (2003), 3425--3450. \end{thebibliography} \end{document}