\documentclass[11pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage{amsmath,amssymb,amsthm} \usepackage{microtype} \usepackage[margin=1in]{geometry} \usepackage[hidelinks]{hyperref} \newtheorem{theorem}{Theorem} \newtheorem{lemma}[theorem]{Lemma} \newcommand{\R}{\mathbb R} \newcommand{\E}{\mathbb E} \newcommand{\Prob}{\mathbb P} \newcommand{\supp}{\operatorname{supp}} \newcommand{\Var}{\operatorname{Var}} \newcommand{\dd}{\,\mathrm d} \title[Convex, time-decreasing functions need not preserve semimartingales]{Convex, time-decreasing functions\\ need not preserve semimartingales} \author{A. Dabrowski's AI agents} \email{alex@dabrow.cc} \date{} \hypersetup{pdfauthor={A. Dabrowski's AI agents},pdftitle={Convex, time-decreasing functions need not preserve semimartingales}} \begin{document} \begin{abstract} We construct a continuous function, strictly convex in space and strictly decreasing in time, whose composition with a bounded continuous martingale is not a semimartingale. We also construct a continuous, spatially convex, time-increasing function that is not a difference of spatially convex, time-nonincreasing functions on a rectangle. The construction combines parabolic oscillations, diffusion superposition and martingale concatenation to obtain infinite drift variation in the composition. \end{abstract} \maketitle \section{Introduction} \begin{theorem}\label{thm:stochastic} There are a continuous martingale $X$ with $X_0=0$ and $|X|\le2$ and a continuous function $G:\R\times[0,\infty)\to\R$, strictly convex in $x$ and strictly decreasing in $t$, such that $G(X_t,t)$ is not a semimartingale on $[0,1]$ in the usual natural filtration of $X$. The derivative $G_x$ is continuous; $G$ is smooth for $t<1$ and equals $x^2/2-t$ for $t\ge1$. \end{theorem} \begin{theorem}\label{thm:analytic} There is a continuous $f:\R^2\to\R$, strictly convex in $x$ and strictly increasing in $t$, whose restriction to $Q=(-1,1)^2$ is not $g-h$ for any $g,h:Q\to\R$ convex in $x$ and nonincreasing in $t$. Moreover, $f-(x^2/2+t)$ is compactly supported in $Q$, $f_x$ is continuous, and the distributional derivatives satisfy $f_{xx},f_t\in L^1_{\mathrm{loc}}(\R^2)$ but $f_t\notin L^\infty(Q)$. \end{theorem} Theorem~\ref{thm:stochastic} answers Lowther's question \cite{LowtherMO} by completing the construction from \cite{LowtherBlog}. Parabolic oscillations produce smooth, spatially convex pairs $F,P$ with $F_t>0>P_t$ and arbitrarily large $\int_Q|P_{xx}F_t+F_{xx}P_t|$. Convex combinations give Theorem~\ref{thm:analytic}; diffusion superposition and concatenation turn this divergence into infinite drift variation for Theorem~\ref{thm:stochastic}. \section{Oscillation} Put $F_0(x,t)=x^2/2+t$, $P_0(x,t)=x^2/2-t$ and $K=(-1/2,1/2)^2$. Integrals over rectangles are with respect to $\dd x\dd t$. \begin{lemma}\label{lem:oscillation} There are $F_n,P_n\in C^\infty(\R^2)$, equal to $F_0,P_0$ off a fixed compact $S\subset(-2/3,2/3)^2$, with $\|F_n-F_0\|_\infty,\|P_n-P_0\|_\infty\le1/16$ and positive derivatives \[ (A_n,B_n,C_n,D_n)=((F_n)_{xx},(F_n)_t,(P_n)_{xx},-(P_n)_t). \] They satisfy \begin{equation}\label{eq:interactions} J_n:=\int_Q|C_nB_n-A_nD_n|\longrightarrow\infty, \qquad I_n:=\int_Q A_nD_n=\int_Q C_nB_n\ge\tfrac12J_n, \end{equation} and \begin{equation}\label{eq:bounds} \int_Q A_n=\int_Q B_n=\int_Q C_n=\int_Q D_n=4, \qquad \|(F_n)_x-x\|_\infty,\ \|(P_n)_x-x\|_\infty\le2. \end{equation} \end{lemma} \begin{proof} Choose $\chi\in C_c^\infty((-2/3,2/3)^2)$ with $0\le\chi\le1$ and $\chi=1$ on $K$, and set $S=\supp\chi$. For the current pair $F,P$, write $(A,B,C,D)=(F_{xx},F_t,P_{xx},-P_t)$ and put \[ a=\frac{\chi}{2(A^{-1}+B^{-1}+C^{-1}+D^{-1})},\qquad \theta=\lambda x+\lambda^2t. \] Replace $F,P$ by $F'=F-\lambda^{-2}a\cos\theta$ and $P'=P-\lambda^{-2}a\sin\theta$. For fixed $F,P$, uniformly as $\lambda\to\infty$, \[ (A',B',C',D')=(A,B,C,D) +a(\cos\theta,\sin\theta,\sin\theta,\cos\theta) +O(\lambda^{-1}). \] Since $a\le\frac12\min(A,B,C,D)$ and the derivatives equal $1$ off $S$, large frequencies preserve positivity. The increments of $A-D$ and $B-C$ tend uniformly to zero. Write $\langle w\rangle=\int_K w$ and $E=\langle A^2+B^2\rangle$. Integration by parts in $x$, with boundary terms $O(\lambda^{-1})$, gives \[ \Delta(\langle A\rangle,\langle B\rangle,\langle AB\rangle)=o(1), \qquad \Delta E=\langle a^2\rangle+o(1). \] Starting from $F_0,P_0$, choose the frequency at step $n\ge0$ so that the averaging errors, mismatch increments and function perturbations are each at most $\delta2^{-n-1}$, where $\delta=1/16$. Then \begin{equation}\label{eq:controls} \begin{gathered} \|A_n-D_n\|_\infty,\ \|B_n-C_n\|_\infty\le\delta,\qquad |\langle A_n\rangle-1|,\ |\langle B_n\rangle-1|, \ |\langle A_nB_n\rangle-1|\le\delta,\\ E_n=2+\sum_{j0,\quad (p_n)_s<0,\quad \|(p_n)_z\|_\infty\le1,\quad \|p_n-P_0\|_\infty\le1/64,\\ Z^n_0=0,\qquad |Z^n|\le1,\qquad Z^n_1\in\{-1,1\}. \end{gathered} \] The finite-variation part $V^n$ in \begin{equation}\label{eq:block-ito} p_n(Z^n_s,s)=p_n(0,0)+\int_0^s(p_n)_z(Z^n_r,r)\dd Z^n_r+V^n_s \end{equation} satisfies $\E W_n<\infty$ and $\E W_n\to\infty$, where $W_n=\Var_{[0,1]}V^n$. \end{lemma} \begin{proof} On $\overline Q$, set \[ \widehat A_n=A_n+7-4t,\qquad \widehat B_n=B_n+1-2x^2. \] The support condition gives $\widehat A_n>0$ on $\overline Q$, $\widehat B_n>0$ in $Q$ and $\widehat B_n(\pm1,t)=0$. For $-1\le t\le1$, define the mean-zero probability laws \[ \mu^n_t=\frac{\widehat A_n(x,t)}{48}\mathbf1_{(-1,1)}(x)\dd x +\frac{4+t}{12}(\delta_{-1}+\delta_1) \] and the coefficient $\mathfrak a_n=2\widehat B_n/\widehat A_n$ on $|x|<1$, zero elsewhere. The laws are narrowly continuous and the coefficient is bounded for each $n$. Since $\partial_t\widehat A_n=\partial_{xx}\widehat B_n$, $\widehat B_n(\pm1,t)=0$ and $\partial_x\widehat B_n(\pm1,t)=\mp4$, integration by parts gives, for $\varphi\in C_c^\infty(\R)$, \[ \frac{\mathrm d}{\mathrm dt}\int\varphi\dd\mu^n_t =\frac1{48}\int_{-1}^1\varphi\,\partial_{xx}\widehat B_n\dd x +\frac{\varphi(-1)+\varphi(1)}{12} =\frac1{48}\int_{-1}^1\widehat B_n\varphi''\dd x, \] that is, \[ \partial_t\mu^n_t=\tfrac12\partial_{xx}(\mathfrak a_n\mu^n_t). \] Trevisan's superposition theorem \cite[Theorem~2.5]{Trevisan} gives a continuous solution $Y^n$ of the martingale problem for $\frac12\mathfrak a_n\partial_{xx}$ with marginals $\mu_t^n$. The marginal supports imply $|Y^n|\le1$. Smooth compactly supported test functions equal to $x$ and $x^2$ on $[-1,1]$ show that $Y^n$ is a martingale with $\dd[Y^n]_t=\mathfrak a_n(Y^n_t,t)\dd t$. If $U^n$ is the finite-variation part of $P_n(Y^n_t,t)$, It\^o's formula gives \begin{equation}\label{eq:large-drift} \E\Var_{[-1,1]}U^n \ge\frac1{48}\int_Q|C_n\widehat B_n-D_n\widehat A_n| \ge\frac{J_n}{48}-1\longrightarrow\infty. \end{equation} Here we discard the nonnegative atomic contribution and use \eqref{eq:bounds}. Set \[ p_n(z,s)=P_0(z,s)+\tfrac14(P_n-P_0)(2z,4s-2). \] Then $(p_n)_{zz}=C_n$ and $(p_n)_s=-D_n$ at the rescaled point; convexity and $(p_n)_z(\pm1,s)=\pm1$ give the slope bound. On $[1/4,3/4]$, take $Z^n_s=Y^n_{4s-2}/2$. Prepend the continuous Brownian martingale $\E[h(B_1)\mid\mathcal F^B_{4s}]$, choosing $h(B_1)$ with the law of $Y^n_{-1}/2$. Given $h(B_1)$, attach $Y^n$ using its regular conditional law given $Y^n_{-1}=2h(B_1)$, independently of the prefix. These conditional laws are martingale laws for almost every initial value, so the join preserves the martingale property. Append the current endpoint plus an independent Brownian motion, stopped at $\{-1,1\}$ with clock $(s-3/4)/(1-s)$. It exits before $s=1$ almost surely. Thus $Z^n$ is a continuous martingale in its natural filtration. On the middle interval, $p_n(Z^n_s,s)=\frac14P_n(Y^n_{4s-2},4s-2)-\frac12$, so \eqref{eq:large-drift} gives $\E W_n\to\infty$. Finally, $\E[Z^n]_1=1$ and smoothness give $\E W_n\le\|(p_n)_s\|_\infty+\frac12\|(p_n)_{zz}\|_\infty<\infty$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:stochastic}] For each $m\ge1$, choose a block with $\E W>2\cdot4^m$. By the weak law, choose increasing integers $r_m$ so that $4^{r_m}$ independent copies have sample average exceeding $4^m$ with probability at least $1-2^{-m}$. Take the groups independent, scale group $m$ by $\epsilon=\sqrt3\,2^{-m-r_m}$, and enumerate the copies in increasing group order as $(p^k,Z^k,V^k,W^k,\epsilon_k)$. The scales satisfy \begin{equation}\label{eq:scales} \sum_k\epsilon_k^2=1,\qquad \epsilon_k/\epsilon_{k+1}\in\mathbb N,\qquad \sum_k\epsilon_k^2W^k=\infty\quad\text{a.s.} \end{equation} Each group has total squared scale $3\cdot4^{-m}$. By Borel--Cantelli, all but finitely many groups succeed, each contributing more than $3$. Following \cite{LowtherBlog}, put $T_0=0$, $T_k=\sum_{j\le k}\epsilon_j^2$ and concatenate: \[ M_t=\sum_{j0$ and $G_t<0$. In block $k$, \[ |G-P_0|\le\epsilon_k^2/64,\qquad |G_x-x|\le2\epsilon_k. \] Extend $G=P_0$ for $t\ge1$. Since $\epsilon_k\to0$, $G,G_x$ are continuous at $t=1$, and strict convexity and time decrease persist: $G(x,t)>\lim_{s\uparrow1}G(x,s)=P_0(x,1)$ for $t<1$. Stop $M$ at $\tau=\inf\{t:|M_t|\ge2\}$ and put $X=M^\tau$. The maximal inequality gives $\Prob(\tau>1)\ge3/4$. In the usual natural filtration of $X$, define on $[0,1]$ \[ \mathcal A_t=G(X_t,t)-G(0,0)-\int_0^tG_x(X_s,s)\dd X_s. \] The integral is a continuous martingale since $G_x$ is bounded on $[-2,2]\times[0,1]$. It\^o's formula gives, on each $[0,T_k]$, $\mathcal A_t=\int_0^tG_t(X_s,s)\dd s+\frac12\int_0^tG_{xx}(X_s,s)\dd[X]_s$. Since $|Z^k|\le1$, $M$ stays in its starting tile during block $k$. On $\{\tau>1\}$, $X=M$ and $\dd[M]=\epsilon_k^2\dd[Z^k]$ in rescaled time, so \eqref{eq:gluing} and \eqref{eq:block-ito} give drift increment $\epsilon_k^2V^k$ in block $k$. Consequently, \[ \Var_{[0,T_k]}\mathcal A=\sum_{j\le k}\epsilon_j^2W^j\longrightarrow\infty \quad\text{a.s. on }\{\tau>1\}. \] If $G(X_t,t)$ were a semimartingale on $[0,1]$, its continuous finite-variation part, starting at zero, would equal $\mathcal A$ on each $[0,T_k]$ by uniqueness. Its variation on $[0,1]$ would dominate all these sums, a contradiction. \end{proof} \begin{thebibliography}{9} \bibitem{LowtherMO} G.~Lowther, \emph{Do convex and decreasing functions preserve the semimartingale property?} \href{https://mathoverflow.net/questions/43319}{MathOverflow, Question~43319} (2010). \bibitem{LowtherBlog} G.~Lowther, \emph{Do convex and decreasing functions preserve the semimartingale property?} Almost Sure, 26 September 2016. \url{https://almostsuremath.com/2016/09/26/do-convex-and-decreasing-functions-preserve-the-semimartingale-property/}. \bibitem{Trevisan} D.~Trevisan, \emph{Well-posedness of multidimensional diffusion processes with weakly differentiable coefficients}, \href{https://doi.org/10.1214/16-EJP4453}{Electron. J. Probab. \textbf{21} (2016), no.~22, 1--41}. \end{thebibliography} \end{document}