/
rational_to_socp.tex
276 lines (259 loc) · 10.5 KB
/
rational_to_socp.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
\documentclass[11pt]{article}
% \usepackage{statistics-macros}
\usepackage{hyperref}
\usepackage[numbers,square]{natbib}
\usepackage[top=2.54cm,bottom=2.54cm,left=2.54cm,right=2.54cm]{geometry}
\usepackage{color}
\usepackage{enumerate}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{amsfonts}
\usepackage{latexsym}
\usepackage{graphicx}
\usepackage{xifthen}
\usepackage{xspace}
\usepackage{bbm}
\newcommand{\R}{\mathbb{R}}
\newcommand{\N}{\mathbb{N}}
\newcommand{\half}{\frac{1}{2}}
\newcommand{\ceil}[1]{\left\lceil{#1}\right\rceil}
\newcommand{\norm}[1]{\left\|{#1}\right\|} % A norm with 1 argument
\newcommand{\lone}[1]{\norm{#1}_1} % l1 norm
\newcommand{\ltwo}[1]{\norm{#1}_2} % l2 norm
\newcommand{\linf}[1]{\norm{#1}_\infty} % l-infinity norm
\begin{document}
\begin{center}
{\LARGE $p$-norm cones and second-order cone representations} \\
\vspace{.2cm}
{\large John C.\ Duchi} \\
\end{center}
In this note, we review how a few power inequalities, and related $p$-norm
constraints for (rational) $p \in (1, \infty)$, may be represented as second
order cone programs. These results follow~\citet{AlizadehGo01}, who
enumerate the reductions below and several more such inequalities.
Our starting point is to note that given a vector $w \in \R^n$ and points
$x, y \ge 0$, the constraint
\begin{equation*}
\ltwo{w} \le xy
\end{equation*}
is equivalent to the constraint
\begin{equation}
\label{eqn:rotated-cone}
\ltwo{\left[\begin{matrix} 2w \\ x - y \end{matrix}\right]}
\le x + y,
~~ x \ge 0, y \ge 0,
\end{equation}
which is evident by squaring both sides and rearranging. Thus, we also
obtain that the set of $t, x, y$ such that $x, y \ge 0$ and $t^2 \le xy$ is
representable as a second-order cone. This insight allows us to represent
more complicated powers as second-order cones.
\subsection*{Products as second order cones}
Consider the set of $t \in \R, s \in \R^{2^k}_+$ such that
\begin{equation}
\label{eqn:two-k-set}
t^{2^k} \le s_1 s_2 \cdots s_{2^k},
~~ s_i \ge 0, ~ \mbox{all}~ s_i.
\end{equation}
We claim this set is SOCP representable. Indeed,
introduce variables
\begin{equation*}
\left\{u_{i,j} :
j \in \{1, \ldots, k-1\},
i \in \{1, \ldots, 2^j\},
u_{i,j} \ge 0\right\}.
\end{equation*}
Then it is clear that the constraint~\eqref{eqn:two-k-set}
is equivalent to
\begin{equation*}
t^{2^k} \le u_{1, k-1}^2 u_{2,k-1}^2 \cdots u_{2^{k-1}, k-1}^2,
~~ u_{i, k-1}^2 \le s_{2i - 1} s_{2i},
~~ \mbox{all~} i,
\end{equation*}
or
\begin{equation*}
t^{2^{k-1}} \le u_{1, k-1} u_{2,k-1} \cdots u_{2^{k-1}, k-1},
~~ u_{i, k-1}^2 \le s_{2i - 1} s_{2i},
~~ \mbox{all~} i \in \{1, \ldots, 2^{k-1}\}.
\end{equation*}
Recursively applying this construction through levels $j = k-2, \ldots,
1$, we obtain the set of inequalities
\begin{align}
t^2 \le u_{1,1} u_{2,1},
& ~~ u_{i,j-1}^2 \le u_{2i-1, j} u_{2i,j}
~ \mbox{for ~} j \in \{2, \ldots, k-2\}, i \in \{1, \ldots, 2^{j-1}\}
\nonumber \\
& ~~ u_{i, k-1}^2 \le s_{2i - 1} s_{2i}
~~ \mbox{for~} i \in \{1, \ldots, 2^{k-1}\}, \label{eqn:socp-two-k-set}
\end{align}
where all $u_{i,j}$ are non-negative. Each of these inequalities, by the
representation~\eqref{eqn:rotated-cone}, corresponds to a second-order
cone in $\R_+^3$, and we have introduced $2^k - 1$ such inequalities.
% Suppose we have the convex constraint $x^p \le t$, where
\subsection*{Product inequalities as second order cones}
We now provide reductions of inequalities of the more restrictive form
\begin{equation}
\label{eqn:initial-power}
x^n \le t^{p_1} s^{p_2}, ~~ x, t, s \ge 0
\end{equation}
where $p_1 + p_2 = n$ and $p_1, p_2, n \in \N$, into a sequence of
second-order cone-represented sets. This is the building block for
representations of general $p$-norm inequalities to come. In particular, we
develop a procedure that takes inequalities of one of the three forms
\begin{subequations}
\label{eqn:power-recursions}
\begin{align}
\label{eqn:zero-power}
x^n & \le t^{p_1} s^{p_2} \\
\label{eqn:one-power}
x^n & \le t^{p_1} s^{p_2} u \\
\label{eqn:two-power}
x^n & \le t^{p_1} s^{p_2} u^2
\end{align}
\end{subequations}
and recurses with a power $\le (n + 1)/2$ on $x$ and one of the
forms~\eqref{eqn:power-recursions}. Note that if $n = 2$, then we already
have an immediate second order cone representation~\eqref{eqn:rotated-cone}.
\begin{enumerate}[(1)]
\item The case with $u^0$, inequality~\eqref{eqn:zero-power}.
We have two possibilities: either $n$ is even or $n$ is odd.
\begin{enumerate}[a.]
\item $n$ is even: we assume that $n \ge 4$, as otherwise we have
the inequality $x^2 \le ts$, which is trivial.
In this case, either both $p_1$ and $p_2$ are even or
both are is odd. If they are even, inequality~\eqref{eqn:zero-power} is
equivalent to $x^{n/2} \le t^{p_1/2} s^{p_2 / 2}$, which gives us a
recursive step. If $p_1$ is odd, then $p_2$ is odd, and
inequality~\eqref{eqn:zero-power} is equivalent to the pair
\begin{equation*}
x^n \le t^{p_1 - 1} s^{p_2 - 1} u^2, ~~ u^2 \le ts,
~~~ \mbox{or} ~~~
x^{n/2} \le t^{\frac{p_1 - 1}{2}} s^{\frac{p_2 - 1}{2}} u,
~~ u^2 \le ts,
\end{equation*}
which allows us to recurse to case~\eqref{eqn:one-power} with lower
powers on $s$, $t$, and $x$.
\item $n$ is odd: In this case, we have exactly one of $p_1$ and $p_2$ is
odd; assume w.l.o.g.\ that $p_1$ is odd. Then
inequality~\eqref{eqn:zero-power} is equivalent to
$x^{n + 1} \le t^{p_1 - 1} s^{p_2} x t$, and introducing the variable
$u \ge 0$, we have the equivalent representation
\begin{equation*}
x^{n + 1} \le t^{p_1 - 1} s^{p_2} u^2, ~~ u^2 \le xt,
~~~ \mbox{or} ~~~
x^{\frac{n + 1}{2}} \le t^{\frac{p_1 - 1}{2}} s^{\frac{p_2}{2}} u,
~~ u^2 \le xt.
\end{equation*}
This allows us to recurse to case~\eqref{eqn:one-power}.
\end{enumerate}
\item The case $u$, inequality~\eqref{eqn:one-power}. We again have
two possibilities: either $n$ is even or $n$ is odd.
\begin{enumerate}[a.]
\item $n$ is even: In this case, exactly one of $p_1$ and $p_2$ is odd;
assume w.l.o.g.\ that $p_1$ is odd. Then we have the equivalent
representation
\begin{equation*}
x^n \le t^{p_1 - 1} s^{p_2} w^2, ~~ w^2 \le tu,
~~~ \mbox{or} ~~~
x^{\frac{n}{2}} \le t^{\frac{p_1 - 1}{2}} s^{\frac{p_2}{2}} w,
~~ w^2 \le tu.
\end{equation*}
This is again an inequality of the form~\eqref{eqn:one-power}.
\item $n$ is odd: In this case, either both $p_1$ and $p_2$ are even
or they are both odd. If they are both even, then we have the equivalent
inequalities
\begin{equation*}
x^{n + 1} \le t^{p_1} s^{p_2} w^2, ~~ w^2 \le xu,
~~~ \mbox{or} ~~~
x^{\frac{n + 1}{2}} \le t^{\frac{p_1}{2}} s^{\frac{p_2}{2}} w,
~~ w^2 \le xu,
\end{equation*}
which is again of the form~\eqref{eqn:one-power}. If both $p_1$ and $p_2$
are odd, then we introduce a few more variables,
noting that inequality~\eqref{eqn:one-power} is equivalent to
\begin{equation*}
x^{n + 1} \le t^{p_1 - 1} s^{p_2 - 1} st ux,
~~~ \mbox{or} ~~~
x^{n + 1} \le t^{p_1 - 1} s^{p_2 - 1} y^4,
~~ y^2 \le w v, ~~ w^2 \le st, ~~ v^2 \le ux,
\end{equation*}
and raising the inequality involving $x^{n + 1}$ to the power $\half$
yields the four inequalities
\begin{equation*}
x^{\frac{n + 1}{2}} \le t^{\frac{p_1 - 1}{2}} s^{\frac{p_2 - 1}{2}} y^2,
~~ y^2 \le w v, ~~ w^2 \le st, ~~ v^2 \le ux,
\end{equation*}
which is of the form~\eqref{eqn:two-power}.
\end{enumerate}
\item The case $u^2$, inequality~\eqref{eqn:two-power}. We show that
we can always reduce this to one of the cases~\eqref{eqn:power-recursions}.
\begin{enumerate}[a.]
\item $n$ is even: In this case, either both $p_1$ and $p_2$ are both even
or they are both odd. If $p_1$ and $p_2$ are even, then
inequality~\eqref{eqn:two-power} is equivalent to $x^{n/2} \le
t^{p_1/2} s^{p_2/2} u$, which is of type~\eqref{eqn:one-power}.
If $p_1$ and $p_2$ are odd, then we have the equivalent representation
\begin{equation*}
x^n \le t^{p_1 - 1} s^{p_2 - 1} u^2 st, ~~~ \mbox{or} ~~~
x^n \le t^{p_1 - 1} s^{p_2 - 1} y^4, ~~ y^4 \le u^2 w^2,
~~ w^2 \le st,
\end{equation*}
which (by inspection) is equivalent to the inequality of
type~\eqref{eqn:two-power} (along with an additional two SOCP inequalities)
\begin{equation*}
x^\frac{n}{2} \le t^\frac{p_1 - 1}{2} s^\frac{p_2 - 1}{2} y^2,
~~ y^2 \le u w,
~~ w^2 \le st.
\end{equation*}
\item $n$ is odd: In this case, exactly one of $p_1$ and $p_2$ is odd;
assume w.l.o.g.\ that $p_1$ is odd. Then inequality~\eqref{eqn:two-power}
is equivalent to
\begin{equation*}
x^{n + 1} \le t^{p_1 - 1} s^{p_2} u^2 t x,
~~~ \mbox{or} ~~~
x^\frac{n + 1}{2} \le t^\frac{p_1 - 1}{2} s^\frac{p_2}{2}
y^2,
~~ y^2 \le uw, ~~ w^2 \le tx.
\end{equation*}
\end{enumerate}
\end{enumerate}
The preceding enumerated steps suggest a recursive strategy, where at each
step, we check which of the cases we are in, then perform a reduction to
introduce at most three inequalities of the form $w^2 \le uv$, and reducing
the powers on all other variables by a factor of $2$. The recursion halts as
soon as we have an inequality of the form $x^n \le y^n$, or an inequality of
the form $x^2 \le uv$, which is second-order-cone representable.
Note that given an inequality of the form~\eqref{eqn:initial-power},
we introduce at most $O(1) \ceil{\log_2( p_1 \vee p_2)}$ new
inequalities via this recursion.
\subsection*{General (rational) $p$-norm cones as second-order cones}
Now we show how to represent the inequality
\begin{equation}
\label{eqn:p-norm-cone}
\norm{x}_p \le t,
~~~ \mbox{where~} p = \frac{n}{m}
~ \mbox{for~some~} n, m \in \N, n \ge m
\end{equation}
for $x \in \R^d$
as a collection of second-order cones and linear inequalities.
First, note that the inequality
\begin{equation*}
\bigg(\sum_{i=1}^d |x_i|^p \bigg)^{1/p} \le t
~~ \equiv ~~
\bigg(\sum_{i=1}^d |x_i|^{n/m} \bigg)^{m/n} \le t
~~ \equiv ~~
\sum_{i=1}^d t^{1 - n/m} |x_i|^{n/m} \le t
\end{equation*}
is equivalent to the collection of inequalities
$s^\top 1 \le t$, $|x_i|^{n/m} \le s_i t^{n/m - 1}$ for all $i$,
which in turn is equivalent to the set of inequalities
\begin{equation*}
v_i \ge |x_i|, ~ t \ge 0, ~ s_i \ge 0, ~ v_i^n \le s_i^m t^{n - m},
~ \sum_{i=1}^d s_i \le t.
\end{equation*}
We know this is SOCP representable by the
recursions~\eqref{eqn:power-recursions}.
\bibliographystyle{abbrvnat}
\bibliography{bib}
\end{document}