Next Article in Journal
Insight in Thermally Radiative Cilia-Driven Flow of Electrically Conducting Non-Newtonian Jeffrey Fluid under the Influence of Induced Magnetic Field
Next Article in Special Issue
Kinematic Modelling for Hyper-Redundant Robots—A Structured Guide
Previous Article in Journal
Supervised Classification of Healthcare Text Data Based on Context-Defined Categories
Previous Article in Special Issue
Predicate-Based Model of Problem-Solving for Robotic Actions Planning
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Towards the Sign Function Best Approximation for Secure Outsourced Computations and Control

by
Mikhail Babenko
1,2,
Andrei Tchernykh
2,3,4,*,
Bernardo Pulido-Gaytan
3,
Arutyun Avetisyan
2,
Sergio Nesmachnow
5,
Xinheng Wang
6 and
Fabrizio Granelli
7
1
North-Caucasus Center for Mathematical Research, North-Caucasus Federal University, 1 Pushkin Street, 355017 Stavropol, Russia
2
Control/Management and Applied Mathematics, Ivannikov Institute for System Programming, 109004 Moscow, Russia
3
Computer Science Department, CICESE Research Center, Ensenada 22800, Mexico
4
School of Electronic Engineering and Computer Science, South Ural State University, 454080 Chelyabinsk, Russia
5
Faculty of Engineering, Universidad de la República, Montevideo 11300, Uruguay
6
Department of Mechatronics and Robotics, Xi’an Jiaotong-Liverpool University, Suzhou 215123, China
7
Department of Information Engineering and Computer Science, University of Trento, 38150 Trento, Italy
*
Author to whom correspondence should be addressed.
Mathematics 2022, 10(12), 2006; https://doi.org/10.3390/math10122006
Submission received: 10 May 2022 / Revised: 7 June 2022 / Accepted: 7 June 2022 / Published: 10 June 2022
(This article belongs to the Special Issue Mathematics in Robot Control for Theoretical and Applied Problems)

Abstract

:
Homomorphic encryption with the ability to compute over encrypted data without access to the secret key provides benefits for the secure and powerful computation, storage, and communication of resources in the cloud. One of its important applications is fast-growing robot control systems for building lightweight, low-cost, smarter robots with intelligent brains consisting of data centers, knowledge bases, task planners, deep learning, information processing, environment models, communication support, synchronous map construction and positioning, etc. It enables robots to be endowed with secure, powerful capabilities while reducing sizes and costs. Processing encrypted information using homomorphic ciphers uses the sign function polynomial approximation, which is a widely studied research field with many practical results. State-of-the-art works are mainly focused on finding the polynomial of best approximation of the sign function (PBAS) with the improved errors on the union of the intervals [ 1 , ϵ ] [ ϵ , 1 ] . However, even though the existence of the single PBAS with the minimum deviation is well known, its construction method on the complete interval [ 1 , 1 ] is still an open problem. In this paper, we provide the PBAS construction method on the interval [ 1 , 1 ] , using as a norm the area between the sign function and the polynomial and showing that for a polynomial degree n 1 , there is (1) unique PBAS of the odd sign function, (2) no PBAS of the general form sign function if n is odd, and (3) an uncountable set of PBAS, if n is even.

1. Introduction

Comparing numbers in a homomorphic cipher causes the problem of finding the polynomial of best approximation of the sign function (PBAS). To approximate it, various approaches are used: rational functions [1], Bernstein polynomials [2], Chebyshev polynomials of the first kind [3,4], Fourier series expansions, artificial neural networks [5], least-squares [6,7,8,9], Newton–Raphson [10], etc. In these approaches, the noncontinuous sign function is replaced by a continuous function s ( x ) equal to:
s ( x ) = { 1     i f   x > ϵ ,         x ϵ ,       i f   x [ ϵ , ϵ ] . 1   o t h e r w i s e    
The main issue is that the approximation is considered on the union of two intervals [ 1 , ϵ ] [ ϵ , 1 ] . The smallest deviation of a polynomial from the sign function is used as a measure of quality. However, this measure has a maximum error close to 0.5 in the zero neighborhood regardless of the degree of the polynomial, which makes it inapplicable for approximating a polynomial on the complete interval [ 1 , 1 ] .
According to Chebyshev theory, there exists a single polynomial f ( x ) for continuous function s ( x ) with the minimum deviation min max x [ 1 , 1 ] | s ( x ) f ( x ) | [11], also known as minimax approximate polynomial or polynomial of best approximation.
The form of the minimax polynomial for the sign function approximation depends on ϵ . Various strategies for choosing ϵ for polynomial approximate s ( x ) are proposed. However, the problem of constructing PBAS remains open.
In this paper, we consider the classical definition of the sign function:
s i g n ( x ) = {       1   i f   x > 0 ,       0   i f   x = 0 , 1   i f   x < 0 .
To construct the PBAS, we use the norm as the area between the sign function and the polynomial f ( x ) , determined by the following formula.
f ( x ) = 1 0 | 1 f ( x ) | d x + 0 1 | 1 f ( x ) | d x = 1 0 | 1 + f ( x ) | d x + 0 1 | 1 f ( x ) | d x
This norm allows us to avoid dramatically increasing the least deviation of the polynomial from the sign function as a result in the zero neighborhood.
Let us formulate the problem of the PBAS construction.
It is required to find the polynomial Q n ( x ) = i = 0 n a i x i , where   i = 0 , n ¯ : a i x i is the i -th term, a i is a coefficient, x is a variable, and deg Q n ( x ) n .
It is formally defined as follows:
i = 0 n a i ( 0 ) x i = Δ = inf a 0 , a 1 , , a n i = 0 n a i x i
If Q n ( x ) exists, it is called the PBAS. In [11], p. 160, the theorem is proved that the PBAS exists. However, the number of PBAS and their form remains open. In this paper, we study these two problems.
The rest of the paper is organized as follows: Section 2 discusses the properties of the norm, which are then used in the proof. Section 3 discusses approximation of the sign function by Bernstein polynomials. It is shown that if n 1 and Q n ( x ) is the PBAS, then Q n ( x ) 1 . Section 4 discusses the PBAS properties. Section 5 discusses the number of the PBAS odd functions. Section 6 investigates the problem of the existence of the PBAS of general form. Section 7 contains a conclusion.

2. Norm and Its Properties

The section discusses the main properties of the norm used for the proof.
Property 1.
If f ( x ) is an even function, then f ( x ) 2 .
Proof. 
Since f ( x ) is an even function, then 1 0 | 1 + f ( x ) | d x = 0 1 | 1 + f ( x ) | d x ; therefore:
f ( x ) = 0 1 | 1 + f ( x ) | d x + 0 1 | 1 f ( x ) | d x = 0 1 | 1 + f ( x ) | + | 1 f ( x ) | d x
considering that x :   | 1 + f ( x ) | + | 1 f ( x ) | 2 , then
f ( x ) 0 1 2 d x = 2
The property is proven. □
Let us consider an example of calculating the norm for n = 0 .
Example 1.
(a) 
Calculate a 0 ; if | a 0 | 1 , then 0 1 | 1 + a 0 | + | 1 a 0 | d x = 2 .
(b) 
Calculate a 0 ; if | a 0 | > 1 , then 0 1 | 1 + a 0 | + | 1 a 0 | d x = 2 | a 0 | > 2 .
From the data presented in Example 1, we can conclude that for n = 0 , there is an uncountable number of PBAS, and they are given by f ( x ) = a 0 , where | a 0 | 1 .
Property 2.
If f ( x ) is an odd function, then f ( x ) = 2 0 1 | 1 f ( x ) | d x .
Proof. 
Since f ( x ) is an odd function, then 1 0 | 1 + f ( x ) | d x = 0 1 | 1 f ( x ) | d x ; therefore:
f ( x ) = 2 0 1 | 1 f ( x ) | d x
The property is proven. □
Property 3.
If f ( x ) = e ( x ) + o ( x ) is a general function, then f ( x ) o ( x )
, where e ( x ) is an even function and o ( x ) is an odd function.
Proof. 
f ( x ) = 1 0 | 1 + e ( x ) + o ( x ) | d x + 0 1 | 1 e ( x ) o ( x ) | d x
Let x = t , then:
1 0 | 1 + e ( x ) + o ( x ) | d x = 1 0 | 1 + e ( t ) + o ( t ) | d t = 0 1 | 1 + e ( t ) o ( t ) | d t
Therefore,
f ( x ) = 0 1 | 1 e ( x ) o ( x ) | + | 1 + e ( x ) o ( x ) | d x 0 1 | 2 2 o ( x ) | d x = 2 0 1 | 1 o ( x ) | d x
According to Property 2 o ( x ) = 2 0 1 | 1 o ( x ) | d x , we find:
f ( x ) o ( x )
The property is proven. □
Property 4.
  ϕ ( 0 , π 2 ) :  
f ( x ) + g ( x ) sin 2 ϕ 1 sin 2 ϕ f ( x ) + cos 2 ϕ 1 cos 2 ϕ   g ( x ) .
Proof. 
By the definition,
f ( x ) + g ( x ) = 1 0 | 1 + f ( x ) + g ( x ) | d x + 0 1 | 1 f ( x ) g ( x ) | d x
According to the basic trigonometric identity sin 2 ϕ + cos 2 ϕ = 1 , then
| 1 + f ( x ) + g ( x ) | = | sin 2 ϕ + f ( x ) + cos 2 ϕ + g ( x ) | | sin 2 ϕ + f ( x ) | + | cos 2 ϕ + g ( x ) | = sin 2 ϕ | 1 + 1 sin 2 ϕ f ( x ) | + cos 2 ϕ | 1 + 1 cos 2 ϕ g ( x ) |
| 1 f ( x ) g ( x ) | = | sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) | | sin 2 ϕ f ( x ) | + | cos 2 ϕ g ( x ) | = sin 2 ϕ | 1 1 sin 2 ϕ f ( x ) | + cos 2 ϕ | 1 1 cos 2 ϕ g ( x ) | .  
Therefore:
f ( x ) + g ( x ) sin 2 ϕ 1 sin 2 ϕ f ( x ) + cos 2 ϕ 1 cos 2 ϕ   g ( x )
The property is proven. □
Corollary 1.
  ϕ [ 0 , π 2 ] :
sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) .
Proof. 
According to Property 4   ϕ ( 0 , π 2 ) :   sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) . Let us show that the inequality holds in the case ϕ = 0 , then g ( x ) g ( x ) in the case ϕ = π 2 , then f ( x ) f ( x ) .
The corollary is proven. □
Corollary 2.
If f ( x ) = g ( x ) = a , then   ϕ [ 0 , π 2 ] :
sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) a
Proof. 
According to Corollary 1, we get:
sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) = a sin 2 ϕ + a cos 2 ϕ = a  
The corollary is proven. □
From Example 1, it follows that if n = 0 , then there are infinitely many PBAS of the zero degree. If f ( x ) = 1 and g ( x ) = 1 , then Q 0 ( x ) = sin 2 ϕ f ( x ) + cos 2 ϕ g ( x ) = cos 2 ϕ defines every PBAS of degree zero.
Let us investigate the problem of the number of PBAS of degrees greater than or equal to one.

3. Approximation of the Sign Function by Bernstein Polynomials

Let us apply the Bernstein polynomials for an approximation of the sign function f n ( x ) .
f n ( x ) = 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 2 i + 1 ( n i ) x 2 i + 1
Since the function f n ( x ) is odd, using Property 2, we can calculate f n ( x ) using f n ( x ) = 2 0 1 | 1 f n ( x ) | d x . Let us calculate the value 0 1 | 1 f n ( x ) | d x , proving the following statement.
Statement 1.
  n + : 0 1 | 1 f n ( x ) | d x = 2 n + 1 ( 2 n + 2 ) 4 n ( 2 n n )
Proof. 
Since the Bernstein polynomials on the interval [ 1 , 1 ] have the property that   n + , x [ 1 , 1 ] : | f n ( x ) | 1 ,
0 1 | 1 f n ( x ) | d x = 0 1 1 f n ( x ) d x
Substituting—instead of f n ( x ) —expression (1), we find
0 1 | 1 f n ( x ) | d x = 0 1 1 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 2 i + 1 ( n i ) x 2 i + 1 d x = ( x 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 ( 2 i + 1 ) ( 2 i + 2 ) ( n i ) x 2 i + 2 ) | 0 1 = 1 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 ( 2 i + 1 ) ( 2 i + 2 ) ( n i )
We represent 1 ( 2 i + 1 ) ( 2 i + 2 ) in the form 1 ( 2 i + 1 ) ( 2 i + 2 ) = 1 2   i + 1 1 2 i + 2 , and we find:
0 1 | 1 f n ( x ) | d x = 1 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 2 i + 1 ( n i ) + 2 n + 1 4 n ( 2 n n ) i = 0 n ( 1 ) i 1 2 i + 2 ( n i )
Substitute
i = 0 n ( 1 ) i 1 2 i + 1 ( n i ) = 4 n ( 2 n + 1 ) ( 2 n n ) i = 0 n ( 1 ) i 1 2 i + 2 ( n i ) = 1 2 n + 2
Hence,
0 1 | 1 f n ( x ) | d x = 2 n + 1 ( 2 n + 2 ) 4 n ( 2 n n )
The statement is proven. □
Corollary 3.
  n + : f n ( x ) min ( 1 ,   2 3 n + 1 )
Proof. 
Since f n ( x ) is an odd function, according to Property 2:
f ( x ) = 2 0 1 | 1 f ( x ) | d x
Let n 1 :
f n ( x ) = 2 2 n + 1 ( 2 n + 2 ) 4 n ( 2 n n ) < 2 4 n ( 2 n n ) 2 3 n + 1 1
if n = 0 , then x = 1 ; therefore   n + : f n ( x ) 1 .
The corollary is proven. □
From Property 1 and Corollary 3, we can conclude that if n 1 , the PBAS is not an even function.

4. Properties of the PBAS

Since the polynomial Q n ( x ) is a continuous function on the interval [ 1 , 1 ] , according to the Weierstrass theorem, it is bounded by this in interval and reaches the minimum and maximum values—that is, there are x m , x M [ 1 , 1 ] such that   x [ 1 , 1 ] : Q n ( x m ) Q n ( x ) Q n ( x M ) . Let us denote m Q = Q n ( x m ) and M Q = Q n ( x M ) , and m Q M Q . Let us investigate the values of m Q and M Q for the PBAS Q n ( x ) . The result is presented in the form of the following lemma.
Lemma 1.
If n ≥ 1 and Q n ( x ) is the PBAS, then m Q 1 and M Q 1 .
Proof. 
We split the two-dimensional space 2 into subspaces using the curves m Q = ± 1 and M Q = ± 1 (see Figure 1).
In the following, we consider each subspace separately.
Subspace 1. Let us assume that PBAS Q n ( x ) satisfies the condition: m Q M Q 1 (see Figure 1, Subspace 1), then x [ 1 , 1 ] : Q n ( x ) 1 , 1 + Q n ( x ) 0 ,   1 Q n ( x ) 0 ; therefore:
Q n ( x ) = 1 0 | 1 + Q n ( x ) | d x + 0 1 | 1 Q n ( x ) | d x = 1 0 1 Q n ( x ) d x + 0 1 1 Q n ( x ) d x = 1 1 Q n ( x ) d x 1 1 1 d x = 2
From Corollary 3, it follows that for n 1 the PBAS has the property Q n ( x ) 1 . Therefore, we came to a contradiction and our assumption is not correct.
Subspace 2. Let us assume that the PBAS Q n ( x ) satisfies the condition: m Q < 1 and 1 < M Q < 1 (see Figure 1, Subspace 2), then x [ 1 , 1 ] :   1 Q n ( x ) 0 and 1 Q n ( x ) 0 :
Q n ( x ) = 1 0 | 1 + Q n ( x ) | d x + 0 1 1 Q n ( x ) d x
We calculate Q n ( x ) + 1 M Q and find
Q n ( x ) + 1 M Q = 1 0 | 2 + Q n ( x ) M Q | d x + 0 1 M Q Q n ( x ) d x
We subtract from Q n ( x ) the value Q n ( x ) + 1 M Q and find:
Q n ( x ) Q n ( x ) + 1 M Q = 1 0 | 1 + Q n ( x ) | | 2 + Q n ( x ) M Q | d x + 1 M Q = 1 0 | 1 + Q n ( x ) | | 2 + Q n ( x ) M Q | + 1 M Q d x
Considering that
x [ 1 , 0 ] : | 2 + Q n ( x ) M Q | | 1 + Q n ( x ) | + | 1 M Q | = | 1 + Q n ( x ) | + 1 M Q ,
then
| 1 + Q n ( x ) | | 2 + Q n ( x ) M Q | + 1 M Q 0
Therefore, Q n ( x ) Q n ( x ) + 1 M Q 0 . If Q n ( x ) Q n ( x ) + 1 M Q > 0 , then Q n ( x ) is not a PBAS, so Q n ( x ) Q n ( x ) + 1 M Q = 0 ; then, x [ 1 , 0 ] : 1 + Q n ( x ) 0 and
Q n ( x ) = 1 0 | 1 + Q n ( x ) | d x + 0 1 1 Q n ( x ) d x = 2 + 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Let λ = 2 1 + M Q > 1 , then
  x [ 1 , 0 ] :   λ Q n ( x ) + 1 M Q 1 + M Q + 1 = λ Q n ( x ) + λ = λ ( Q n ( x ) + 1 ) 0   x [ 1 , 0 ] : 1 λ Q n ( x ) 1 M Q 1 + M Q = λ M Q λ Q n ( x ) = λ ( M Q Q n ( x ) ) 0
Therefore,
λ Q n ( x ) + 1 M Q 1 + M Q = λ 1 0 1 + Q n ( x ) d x + λ 0 1 M Q Q n ( x ) d x = λ ( 1 + M Q ) + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) = 2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
since λ > 1 и 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 , then
λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Therefore,
λ Q n ( x ) + 1 M Q 1 + M Q < Q n ( x )
it means that Q n ( x ) is not a PBAS. We came to a contradiction.
Subspace 3. Let us assume that the PBAS Q n ( x ) satisfies the condition: 1 < m Q < M Q < 1 (see Figure 1, Subspace 3), then x [ 1 , 1 ] : 1 < Q n ( x ) < 1 , 1 + Q n ( x ) > 0 ,   1 Q n ( x ) > 0 . Let M = max ( | m Q | , | M Q | ) < 1 . If M = 0 , then Q n ( x ) = 0 and Q n ( x ) = 2 ; from the other side, n 1 : Q n ( x ) 1 . Therefore, we came to a contradiction and M 0 . Let λ = 1 M > 1 and x [ 1 , 1 ] : 1 λ Q n ( x ) 1 , 1 + λ Q n ( x ) > 0 ,   1 λ Q n ( x ) > 0 . We calculate the value Q n ( x ) and find:
Q n ( x ) = 1 0 1 + Q n ( x ) d x + 0 1 1 Q n ( x ) d x = 2 + ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
Since, according to the conditions of Theorem 1 and Corollary 3, n 1 and Q n ( x ) 1 . Hence,
1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 .
We calculate λ Q n ( x ) and find
λ Q n ( x ) = 1 0 1 + λ Q n ( x ) d x + 0 1 1 λ Q n ( x ) d x = 2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
Since λ > 1 и 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 ,
λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Therefore,
2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 2 + ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
and
λ Q n ( x ) < Q n ( x )
Therefore, we came to a contradiction and our assumption is not correct.
Subspace 4. Let us assume that the PBAS Q n ( x ) satisfies the condition: M Q > 1 and 1 < m Q < 1 (see Figure 1, Subspace 4).
Q n ( x ) = 1 0 1 + Q n ( x ) d x + 0 1 | 1 Q n ( x ) | d x
we calculate Q n ( x ) 1 m Q and get
Q n ( x ) 1 m Q = 1 0 Q n ( x ) m Q d x + 0 1 | 2 Q n ( x ) + m Q | d x
we subtract from Q n ( x ) the value Q n ( x ) 1 m Q and get:
Q n ( x ) Q n ( x ) 1 m Q = 1 + m Q + 0 1 | 1 Q n ( x ) | | 2 Q n ( x ) + m Q | d x = 0 1 | 1 Q n ( x ) | | 2 Q n ( x ) + m Q | + 1 + m Q d x
Considering that x [ 0 , 1 ] : | 2 Q n ( x ) + m Q | | 1 Q n ( x ) | + | 1 + m Q | = | 1 Q n ( x ) | + 1 + m Q , then x [ 0 , 1 ] : | 1 Q n ( x ) | | 2 Q n ( x ) + m Q | + 1 + m Q 0 ; therefore, Q n ( x ) Q n ( x ) 1 m Q 0 . If Q n ( x ) Q n ( x ) 1 m Q > 0 , then Q n ( x ) is not a PBAS, so Q n ( x ) Q n ( x ) 1 m Q = 0 and x [ 0 , 1 ] : 1 Q n ( x ) 0 and
Q n ( x ) = 2 + 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
let λ = 2 1 m Q > 1 , then
  x [ 1 , 0 ] :   λ Q n ( x ) 1 + m Q 1 m Q + 1 = λ Q n ( x ) λ m Q = λ ( Q n ( x ) m Q ) 0   x [ 0 , 1 ] :   1 λ Q n ( x ) + 1 + m Q 1 m Q = λ λ Q n ( x ) = λ ( 1 Q n ( x ) ) 0
Therefore,
λ Q n ( x ) 1 + m Q 1 m Q = λ 1 0 Q n ( x ) m Q d x + λ 0 1 1 Q n ( x ) d x = 2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
since λ > 1 and 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 , then
( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Therefore,
λ Q n ( x ) 1 + m Q 1 m Q < Q n ( x )
it means that Q n ( x ) is not a PBAS. We came to a contradiction.
Subspace 5. Let us assume that the PBAS Q n ( x ) satisfies the condition: M Q m Q 1 (see Figure 1, Subspace 5); therefore, x [ 1 , 1 ] : Q n ( x ) 1 , 1 + Q n ( x ) 0 ,   1 Q n ( x ) 0 means
Q n ( x ) = 1 0 | 1 + Q n ( x ) | d x + 0 1 | 1 Q n ( x ) | d x = 1 0 1 + Q n ( x ) d x + 0 1 Q n ( x ) 1 d x = 1 1 Q n ( x ) d x 1 1 1 d x = 2
From Corollary 3, it follows that for n 1 , the PBAS has the property Q n ( x ) 1 . This means that we have come to a contradiction and our assumption is not correct.
Subspace 6. Let us assume that the PBAS Q n ( x ) satisfies the condition: m Q = 1 and 1 < M Q < 1 (see Figure 1, Subspace 6), then
Q n ( x ) = 1 0 1 + Q n ( x ) d x + 0 1 1 Q n ( x ) d x = 2 + 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Let λ = 2 1 + M Q > 1 , then
  x [ 1 , 0 ] : λ Q n ( x ) + 1 M Q 1 + M Q + 1 = λ + λ Q n ( x ) = λ ( 1 + Q n ( x ) ) 0   x [ 0 , 1 ] :   1 λ Q n ( x ) 1 M Q 1 + M Q = λ M Q λ Q n ( x ) = λ ( M Q Q n ( x ) ) 0
Therefore,
λ Q n ( x ) + 1 M Q 1 + M Q = 1 0 λ + λ Q n ( x ) d x + 0 1 λ M Q λ Q n ( x ) d x = 2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
Since λ > 1 и 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 ,
λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Therefore,
λ Q n ( x ) + 1 M Q 1 + M Q < Q n ( x )
This means that Q n ( x ) is not a PBAS.
Subspace 7. Let us assume that Q n ( x ) satisfies the condition: M Q = 1 and 1 < m Q < 1 (see Figure 1, Subspace 7). Then,
Q n ( x ) = 1 0 1 + Q n ( x ) d x + 0 1 1 Q n ( x ) d x = 2 + 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Let λ = 2 1 m Q > 1 , then
  x [ 1 , 0 ] :   λ Q n ( x ) 1 + m Q 1 m Q + 1 = λ Q n ( x ) λ m Q = λ ( Q n ( x ) m Q ) 0   x [ 0 , 1 ] : 1 λ Q n ( x ) + 1 + m Q 1 m Q = λ λ Q n ( x ) = λ ( 1 Q n ( x ) )   0
Therefore,
λ Q n ( x ) 1 + m Q 1 m Q = λ 1 0 Q n ( x ) m Q d x + λ 0 1 1 Q n ( x ) d x = 2 + λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x )
Since λ > 1 and 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x 1 ,
λ ( 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x ) < 1 0 Q n ( x ) d x 0 1 Q n ( x ) d x
Therefore,
λ Q n ( x ) 1 + m Q 1 m Q < Q n ( x )
This means that Q n ( x ) is not a PBAS.
Subspace 8. Since in all seven cases we have come to a contradiction, if Q n ( x ) is a PBAS, it satisfies the boundary conditions defining Subspace 8 (See, Figure 1).
Lemma 1 is proven. □
Lemma 2.
For n 1 , there exists the PBAS odd function Q n 1 ( x ) .
Proof. 
The existence of the PBAS Q n ( x ) follows from Theorem [11] p. 160. Since for n 1 , the PBAS Q n ( x ) is not an even function, so Q n ( x ) is either a general function or an odd function.
Let us assume that Q n ( x ) is a general function; it can be represented in the form Q n ( x ) = Q n 0 ( x ) + Q n 1 ( x ) , where Q n 0 ( x ) is an even function and Q n 1 ( x ) is an odd function. It follows from Property 3 that Q n ( x ) Q n 1 ( x ) . Considering that Q n ( x ) is the PBAS, Q n ( x ) = Q n 1 ( x ) , so the odd function Q n 1 ( x ) is the PBAS. Therefore, for any n 1 , there is the PBAS Q n ( x ) , which is an odd function.
Lemma 2 is proven. □
Corollary 4.
Let n 1 , Q n 1 ( x ) be a PBAS odd function, M Q > 1 , and m Q < 1 .
Proof. 
We assume that PBAS is the odd function Q n 1 ( x ) and M Q = m Q = 1 .
Let us consider the function R ( x ) = λ Q n 1 ( x ) , where λ . Since Q n 1 ( x ) is an odd function, R ( x ) is also an odd function. We calculate Q n 1 ( x ) and R ( x ) using Property 2 and find:
Q n 1 ( x ) = 2 0 1 1 Q n 1 ( x ) d x R ( x ) = 2 0 1 | 1 λ Q n 1 ( x ) | d x
Let us show that there exists λ > 1 , for which the inequality Q n 1 ( x ) > R ( x ) is satisfied.
0 1 1 Q n 1 ( x ) d x > 0 1 | 1 λ Q n 1 ( x ) | d x 0 1 1 Q n 1 ( x ) | 1 λ Q n 1 ( x ) | d x > 0
We denote as G + a set of all x [ 0 , 1 ] for which the inequality 1 λ Q n 1 ( x ) 0 holds and G for which the inequality 1 λ Q n 1 ( x ) 0 holds. We then find:
0 1 1 Q n 1 ( x ) | 1 λ Q n 1 ( x ) | d x = G + λ Q n 1 ( x ) Q n 1 ( x ) d x + G 2 Q n 1 ( x ) λ Q n 1 ( x ) d x = ( λ 1 ) G + Q n 1 ( x ) d x + G 2 Q n 1 ( x ) λ Q n 1 ( x ) d x = ( λ 1 ) G + Q n 1 ( x ) d x + G 2 d x G Q n 1 ( x ) + λ Q n 1 ( x ) d x = ( λ 1 ) G + Q n 1 ( x ) d x + 2 | G | ( 1 + λ ) G Q n 1 ( x ) d x = λ ( G + Q n 1 ( x ) d x G Q n 1 ( x ) d x ) + 2 | G | 0 1 Q n 1 ( x ) d x = λ ( 0 1 Q n 1 ( x ) d x 2 G Q n 1 ( x ) d x ) + 2 | G | 0 1 Q n 1 ( x ) d x λ 0 1 Q n 1 ( x ) d x 2 λ | G | + 2 | G | 0 1 Q n 1 ( x ) d x = ( λ 1 ) ( 0 1 Q n 1 ( x ) d x 2 | G | )
where | G | is the length of the set G .
We denote g ( λ ) = { | G | | G = { x | 1 λ Q n 1 ( x ) 0 &   0 x 1 } } .
Since n 1 and x [ 0 , 1 ] :   Q n 1 ( x ) 1 , then g ( 1 ) = 0 and λ > 1 :   g ( λ ) < 1 .
Let us consider two cases.
Case 1: If x [ 0 , 1 ] :   Q n 1 ( x ) < 1 , then there is such a number x a [ 0 , 1 ] for which x [ 0 , 1 ] :   Q n 1 ( x ) Q n 1 ( x a ) = M Q a holds. If M Q a 0 , then 0 1 1 Q n 1 ( x ) d x 1 ; therefore, Q n 1 ( x ) 2 > 1 so Q n 1 ( x ) is not the PBAS. If M Q a > 0 , we choose as λ the value λ = 1 M Q a > 1 , for which the inequality Q n 1 ( x ) > R ( x ) holds, and Q n 1 ( x ) is not a PBAS. Therefore, we came to a contradiction.
Case 2: If M Q a = 1 , then g ( λ ) is an increasing function; that is, ξ > 1 , for which the inequality 0 1 Q n 1 ( x ) d x 2 | G | = 0 holds. Therefore, for any λ ( 1 , ξ ) , the following inequality holds:
0 1 1 Q n 1 ( x ) | 1 λ Q n 1 ( x ) | d x > 0
Therefore, we came to a contradiction. If M Q = 1 and m Q = 1 , then n 1 : Q n 1 ( x ) , which is not the PBAS.
The corollary is proven. □

5. The Number of PBAS Odd Functions

In Lemma 2, it is proved that for n 1 , the PBAS is an odd function, but the question of their number remains open. The following theorem will answer this question.
Theorem 1.
If n 1 , then there is only one odd function Q n 1 ( x ) that is the PBAS. Depending on the n , the function Q n 1 ( x ) is determined as follows:
If n is odd, then
Q n 1 ( x ) = x i = 1 n + 1 2 1 sin i π n + 3 j = 1 ,   j i n + 1 2 x 2 sin 2 j π n + 3 sin 2 i π n + 3 sin 2 j π n + 3
and
Q n 1 ( x ) = 2 tan π 2 n + 6 ;
If n is even, then
Q n 1 ( x ) = x i = 1 n 2 1 sin i π n + 2 j = 1 ,   j i n 2 x 2 sin 2 j π n + 2 sin 2 i π n + 2 sin 2 j π n + 2
and
Q n 1 ( x ) = 2 tan π 2 n + 4 .
Proof. 
Let us consider two cases.
Case 1.  n is an odd number.
We consider the points 0 < x 1 < x 2 < < x u 1 such that i = 1 , u ¯ : Q n 1 ( x i ) = 1 . According to Corollary, 4 the value M Q a satisfies the condition M Q a > 1 . Considering that the function Q n 1 ( x ) is an odd continuous function, then at least one point x 1 [ 0 , 1 ] is such that Q n 1 ( x ) = 1 exists.
Let us consider the question of the number of zeroes of the function F ( x ) = d Q n 1 ( x i ) d x . Since the function Q n 1 ( x ) is an odd continuous function, then F ( x ) is an even function. The number of zeroes of F ( x ) is less or equal to n 1 , of which non-negative numbers are less than or equal to n 1 2 . Therefore, the number of solutions to the equation Q n 1 ( x ) = 1 satisfying the question x ( 0 , 1 ] is less than or equal to n 1 2 + 1 = n + 1 2 . That is, u n + 1 2 .
Let us consider the points 0 = y 0 < y 1 < y 2 < < y v < y v + 1 = 1 . In each of the points y 1 , y 2 , , y v the value of the function f ( x ) = 1 Q 2 v 1 1 ( x ) = 1 i = 0 v 1 a 2 i + 1 x 2 i + 1 changes its sign.
I v = Q 2 v 1 1 ( x ) 2 = 0 1 | 1 Q n 1 ( x ) | d x = i = 0 v ( 1 ) i y i y i + 1 f ( x ) d x = 2 i = 1 v ( 1 ) i + 1 F ( y i ) + ( 1 ) v F ( y v + 1 )
where F ( x ) = x i = 0 v 1 a 2 i + 1 2 i + 2 x 2 i + 2 .
We calculate the values of the partial derivatives i = 1 , v ¯ :
F ( y i ) y i = 1 i = 0 v 1 a 2 i + 1 y i 2 i + 1 i = 0 v 1 a 2 i + 1 y i y i 2 i + 2 2 i + 2
Since 1 i = 0 v 1 a 2 i + 1 y i 2 i + 1 = 0 by the definition, then:
F ( y i ) y i = i = 0 v 1 a 2 i + 1 y i y i 2 i + 2 2 i + 2  
We calculate the values of the partial derivatives i j :
F ( y i ) y j = i = 0 v 1 a 2 i + 1 y j y i 2 i + 2 2 i + 2  
The necessary condition for the value Q 2 v + 1 1 ( x ) to be minimal is: i = 1 , v ¯ :   I v y i = 0 ; therefore,
I v y i = 2 j = 1 v ( 1 ) j + 1 k = 0 v 1 a 2 k + 1 y i y j 2 k + 2 2 k + 2 ( 1 ) v k = 0 v 1 a 2 k + 1 y i 1 2 k + 2   = k = 0 v 1 a 2 k + 1 y i 1 2 k + 2 ( 2 j = 1 v ( 1 ) j + 1 y j 2 k + 2 + ( 1 ) v )
Solving the system I v y i = 0 [12], we find that k = 0 , v 1 ¯ :
2 j = 1 v ( 1 ) j + 1 y j 2 k + 2 + ( 1 ) v = 0
Considering that   i = 1 , v ¯ :   y i > 0 ; therefore,   i = 1 , v ¯ :   y i = sin i π 2 v + 2 [12].
Using the Lagrange interpolation formula, we calculate the value Q 2 v 1 1 ( x ) , and we find Q 2 v 1 1 ( x ) = i = 1 v l i ( x ) i = 1 v l ¯ i ( x ) , where
l i ( x ) = j = 1 v x + y j y i + y j j = 1 ,   j i v x y j y i y j   l ¯ i ( x ) = j = 1 ,   j i v x + y j y i y j j = 1 v x y j y i + y j  
Then,
Q 2 v 1 1 ( x ) = i = 1 v ( j = 1 v x + y j y i + y j j = 1 ,   j i v x y j y i y j + j = 1 ,   j i v x + y j y i y j j = 1 v x y j y i + y j   ) = i = 1 v j = 1 ,   j i v x y j y i y j j = 1 ,   j i v x + y j y i + y j   ( x + y i 2 y i + x y j 2 y i ) = x i = 1 v 1 y i j = 1 ,   j i v x y j y i y j j = 1 ,   j i v x + y j y i + y j   = x i = 1 v 1 y i j = 1 ,   j i v x 2 y j 2 y i 2 y j 2
Let F ( x ) = i = 1 v a i 2 i x 2 i and d F ( x ) d x = Q 2 v 1 1 ( x ) , so I v is equal to
I v = 0 1 | 1 Q 2 v 1 1 ( x ) | d x = i = 0 v ( 1 ) i sin i π 2 v + 2 sin ( i + 1 ) π 2 v + 2 1 Q 2 v 1 1 ( x ) d x = 2 i = 1 v ( 1 ) i + 1 sin i π 2 v + 2 + ( 1 ) v + 2 j = 1 v ( 1 ) j F ( sin j π 2 v + 2 ) + ( 1 ) v + 1 F ( 1 )
We calculate the value 2 j = 1 v ( 1 ) j F ( sin j π 2 v + 2 ) + ( 1 ) v + 1 F ( 1 ) , and we obtain:
2 j = 1 v ( 1 ) j F ( sin j π 2 v + 2 ) + ( 1 ) v + 1 F ( 1 ) = 2 j = 1 v ( 1 ) j i = 1 v a i 2 i sin 2 i j π 2 v + 2 + ( 1 ) v + 1 i = 1 v a i 2 i = i = 1 v a i 2 i ( 2 j = 1 v ( 1 ) j sin 2 i j π 2 v + 2 + ( 1 ) v + 1 )
Considering that (2) holds,   i = 1 , v ¯ :   2 j = 1 v ( 1 ) j sin 2 i j π 2 v + 2 + ( 1 ) v + 1 = ( 2 j = 1 v ( 1 ) j + 1 sin 2 i j π 2 v + 2 + ( 1 ) v ) = 0 , then
2 j = 1 v ( 1 ) j F ( sin j π 2 v + 2 ) + ( 1 ) v + 1 F ( x ) = 0
and
I v = 0 1 | 1 Q 2 v 1 1 ( x ) | d x = 2 i = 1 v ( 1 ) i + 1 sin i π 2 v + 2 + ( 1 ) v
If v is even, then
I v = 2 i = 1 v / 2 sin ( 2 i 1 ) π 2 v + 2 2 i = 1 v / 2 sin i π v + 1 + 1
Using the formula sin ( α β ) = sin α cos β sin β cos α , where α = 2 i π 2 v + 2 = i π v + 1 and β = π 2 v + 2 , we have
i = 1 v / 2 sin ( 2 i 1 ) π 2 v + 2 = i = 1 v / 2 ( sin i π v + 1 cos π 2 v + 2 sin π 2 v + 2 cos i π v + 1 ) = cos π 2 v + 2 i = 1 v / 2 sin i π v + 1 sin π 2 v + 2 i = 1 v / 2 cos i π v + 1
Since 1 2 + i = 1 n cos i x = sin ( n + 1 2 ) x 2 sin 1 2 x and i = 1 n sin i x = cos x 2 cos ( n + 1 2 ) x 2 sin 1 2 x ([13] p. 2), where n = v 2 and x = π v + 1 , we have:
i = 1 v / 2 cos i π v + 1 = sin ( v 2 + 1 2 ) π v + 1 2 sin π 2 v + 2 1 2 = 1 2 sin π 2 v + 2 1 2 i = 1 v / 2 sin i π v + 1 = cos π 2 v + 2 cos ( v 2 + 1 2 ) π v + 1 2 sin π 2 v + 2 = cos π 2 v + 2 2 sin π 2 v + 2
Therefore,
I v = 2 ( cos π 2 v + 2 cos π 2 v + 2 2 sin π 2 v + 2 sin π 2 v + 2 ( 1 2 sin π 2 v + 2 1 2 ) ) cos π 2 v + 2 sin π 2 v + 2 + 1 = cos 2 π 2 v + 2 + sin 2 π 2 v + 2 sin π 2 v + 2 cos π 2 v + 2 sin π 2 v + 2
Using the basic trigonometric identities cos 2 2 α + sin 2 2 α = 1 and 1 cos 2 α = 2 sin 2 α , sin 2 α = 2 sin α cos α , where α = π 4 v + 4 we obtain:
I v = 1 cos π 2 v + 2 sin π 2 v + 2 = 2 sin 2 π 4 v + 4 2 sin π 4 v + 4 cos π 4 v + 4 = tan π 4 v + 4
If v is odd, then
I v = 2 i = 1 v + 1 2 sin ( 2 i 1 ) π 2 v + 2 2 i = 1 v 1 2 sin i π v + 1 1
Using the formula sin ( α β ) = sin α cos β sin β cos α , where α = 2 i π 2 v + 2 = i π v + 1 and β = π 2 v + 2 , we have
i = 1 v + 1 2 sin ( 2 i 1 ) π 2 v + 2 = i = 1 v + 1 2 ( sin i π v + 1 cos π 2 v + 2 sin π 2 v + 2 cos i π v + 1 ) = cos π 2 v + 2 i = 1 v + 1 2 sin i π v + 1 sin π 2 v + 2 i = 1 v + 1 2 cos i π v + 1
Since i = 1 n sin i x = cos x 2 cos ( n + 1 2 ) x 2 sin 1 2 x [13] p. 2, where n = v + 1 2 and x = π v + 1 we find:
i = 1 v + 1 2 sin i π v + 1 = cos π 2 v + 2 cos ( v + 1 2 + 1 2 ) π v + 1 2 sin π 2 v + 2 = cos π 2 v + 2 cos ( π 2 + π 2 v + 2 ) 2 sin π 2 v + 2
According to the reduction formula cos ( π 2 + π 2 v + 2 ) = sin π 2 v + 2 , we have:
i = 1 v + 1 2 sin i π v + 1 = cos π 2 v + 2 + sin π 2 v + 2 2 sin π 2 v + 2
Using the formula 1 2 + i = 1 n cos i x = sin ( n + 1 2 ) x 2 sin 1 2 x [13] p. 2, where n = v + 1 2 and x = π v + 1 we find:
i = 1 v + 1 2 cos i π v + 1 = sin ( v + 1 2 + 1 2 ) π v + 1 2 sin π 2 v + 2 1 2 = sin ( π 2 + π 2 v + 2 ) 2 sin π 2 v + 2 1 2
According to the reduction formula sin ( π 2 + π 2 v + 2 ) = cos π 2 v + 2 , we obtain:
i = 1 v + 1 2 cos i π v + 1 = cos π 2 v + 2 2 sin π 2 v + 2 1 2
Since i = 1 n sin i x = cos x 2 cos ( n + 1 2 ) x 2 sin 1 2 x [13] p. 2, where n = v 1 2 and x = π v + 1 we find:
i = 1 v 1 2 sin i π v + 1 = cos π 2 v + 2 cos ( v 1 2 + 1 2 ) π v + 1 2 sin π 2 v + 2 = cos π 2 v + 2 cos ( π 2 π 2 v + 2 ) 2 sin π 2 v + 2
According to the reduction formula cos ( π 2 π 2 v + 2 ) = sin π 2 v + 2 , we find:
i = 1 v 1 2 sin i π v + 1 = cos π 2 v + 2 sin π 2 v + 2 2 sin π 2 v + 2
Therefore,
I v = 2 ( cos π 2 v + 2 cos π 2 v + 2 + sin π 2 v + 2 2 sin π 2 v + 2 sin π 2 v + 2 ( cos π 2 v + 2 2 sin π 2 v + 2 1 2 ) ) 2 cos π 2 v + 2 sin π 2 v + 2 2 sin π 2 v + 2 1 = 1 cos π 2 v + 2 sin π 2 v + 2 = tan π 4 v + 4
Therefore,   v N : Q 2 v 1 1 ( x ) = 2 I v = 2 tan π 4 v + 4 .
Since v N : I v 1 > I v , then the smallest value Q 2 v 1 1 ( x ) at the maximum v , considering that v u n + 1 2 , then v = n + 1 2 and 2 v + 2 = n + 3 .
Case 2. If n is an even number, then the result is obtained similarly to case 1, except v = n 2 and 2 v + 2 = n + 2 .
The theorem is proved. □
From Theorem 1, it follows that for n 1 , there is a unique odd function that is the PBAS, which is constructed using the Lagrange interpolation formula, and the interpolation nodes are an alternative to Chebyshev for Chebyshev polynomials of the second kind.
Example 2.
Construct the PBAS for n = 3 and n = 4 , which are odd functions.
Solution
If n = 3 , then, according to Theorem 1, the PBAS is given by the following formula:
Q 3 1 ( x ) = x i = 1 2 1 sin i π 6 j = 1 ,   j i 2 x 2 sin 2 j π 6 sin 2 i π 6 sin 2 j π 6 = x ( 1 sin π 6 x 2 sin 2 π 3 sin 2 π 6 sin 2 π 3 + 1 sin π 3 x 2 sin 2 π 6 sin 2 π 3 sin 2 π 6 ) = 2 x ( 2 x 2 + 3 2 + 2 3 3 x 2 3 6 ) = 4 3 12 3 x 3 + 9 3 3 x
If n = 4 , then, according to Theorem 1, the PBAS is given by the following formula:
Q 4 1 ( x ) = x i = 1 2 1 sin i π 6 j = 1 ,   j i 2 x 2 sin 2 j π 6 sin 2 i π 6 sin 2 j π 6 = 4 3 12 3 x 3 + 9 3 3 x
Let us pay attention to the fact that Q 3 1 ( x ) = Q 4 1 ( x ) . This fact can be generalized: if n is even and n 2 , then Q n 1 ( x ) = Q n 1 1 ( x ) .

6. The Number of PBAS of the Neither Function

Let us investigate the problem of the existence of PBAS Q n ( x ) .
Theorem 2.
If n ≥ 1, then the following statements are true:
1. 
If n is an odd number, then there is no PBAS Q n ( x ) .
2. 
If n is an even number, then there is an infinite number of PBAS Q n ( x ) .
Proof. 
From Theorem 1, it follows that there is a unique odd function Q n 1 ( x ) that is a PBAS. Let us show that there exists an even function Q n 0 ( x ) 0 , such that: Q n ( x ) = Q n 1 ( x ) . For this, we calculate Q n ( x ) Q n 1 ( x ) and find:
Q n ( x ) Q n 1 ( x ) = 0 1 | 1 Q n 0 ( x ) Q n 1 ( x ) | + | 1 + Q n 0 ( x ) Q n 1 ( x ) | 2 | 1 Q n 1 ( x ) | d x
where Q n ( x ) = Q n 0 ( x ) + Q n 1 ( x ) , Q n 0 ( x ) is an even function, and Q n 1 ( x ) is an odd function.
Q n ( x ) Q n 1 ( x ) is equal to zero only if the condition x [ 0 , 1 ] :   | 1 Q n 0 ( x ) Q n 1 ( x ) | + | 1 + Q n 0 ( x ) Q n 1 ( x ) | 2 | 1 Q n 1 ( x ) | = 0 holds, equivalent to:
x [ 0 , 1 ]   и   Q n 1 ( x ) 1 :   { 1 Q n 0 ( x ) Q n 1 ( x ) 0 , 1 + Q n 0 ( x ) Q n 1 ( x ) 0 ; Q n 1 ( x ) 1 Q n 0 ( x ) 1 Q n 1 ( x )
and
x [ 0 , 1 ]   и   Q n 1 ( x ) 1 :   { 1 Q n 0 ( x ) Q n 1 ( x ) 0 , 1 + Q n 0 ( x ) Q n 1 ( x ) 0 ; 1 Q n 1 ( x ) Q n 0 ( x ) Q n 1 ( x ) 1
Therefore: x [ 0 , 1 ] : | 1 Q n 1 ( x ) | Q n 0 ( x ) | 1 Q n 1 ( x ) | .
Since Q n 1 ( x i ) is an odd-function PBAS, it follows from Theorem 1 that there are points x 1 , x 2 , , x u ( 0 , 1 ] such that i = 1 , u ¯ : Q n 1 ( x i ) = 1 . Since Q n 1 ( x i ) is an odd-function PBAS, it follows from the proof of Theorem 1 that if n is an odd number, then u = n + 1 2 . Otherwise, u = n 2 .
Substituting x 1 , x 2 , , x u into the inequalities | 1 Q n 1 ( x ) | Q n 0 ( x ) | 1 Q n 1 ( x ) | we find i = 1 , u ¯ : 0 Q n 0 ( x i ) 0 ; therefore, the necessary condition is i = 1 , u ¯ : Q n 0 ( x i ) = 0 . Since the function Q n 0 ( x ) is an even function, i = 1 , u ¯ : Q n 0 ( x i ) = 0 ; therefore, Q n 0 ( x ) is divisible by the polynomial i = 1 u ( x 2 x i 2 ) and deg Q n 0 ( x ) 2 u . Let us consider two cases.
Case 1. If n is an odd number, then deg Q n 0 ( x ) 2 u = n + 1 . Therefore, there is no even polynomial satisfying the condition deg Q n 0 ( x ) n . Hence, if n is an odd number, there is no PBAS that is a function of general form.
Case 2. If n is an even number, then deg Q n 0 ( x ) 2 u = n . From the other side, deg Q n 0 ( x ) n ; therefore, deg Q n 0 ( x ) = n . To construct the polynomial Q n 0 ( x ) we consider the polynomial of the form:
Z n ( x ) = Q n 1 ( x ) 1 i = 1 n / 2 ( x x i )
where i = 1 , n 2 ¯ : x i = sin i π n + 2 .
We consider the equation Q n 1 ( x ) 1 = 0 , i = 1 , n 2 ¯ : Q n 1 ( x i ) 1 = 0 ; therefore, according to Rolle’s theorem, in each of the intervals ( x i , x i + 1 ) , at least one point ξ i ( x i , x i + 1 ) exists for which F ( ξ i ) = 0 , where F ( x ) = d ( Q n 1 ( x ) 1 ) d x = d Q n 1 ( x ) d x and i 1 , n 2 1 ¯ . Since Q n 1 ( x ) is an odd function, F ( x ) is an even function; therefore,   i 1 , n 2 1 ¯ : F ( ξ i ) = 0 . Considering that deg F ( x ) = n 2 , then, according to the main theorem of algebra, the equation F ( x ) = 0 over the field of real numbers can have at most n 2 roots—considering their multiplicity—so ± ξ i are roots of multiplicity one. Since ± ξ i are roots of multiplicity one, the function F ( x ) passing through ± ξ i changes its sign; therefore, ( , ξ n 2 1 ) , ( ξ n 2 1 , ξ n 2 2   ) ,…, ( ξ 2 , ξ 1   ) ,   ( ξ 1 , ξ 1   ) , ( ξ 1 , ξ 2   ) ,…,   ( ξ n 2 2 , ξ n 2 1   ) , ( ξ n 2 1 , +   ) are the intervals of the increase or decrease in the function Q n 1 ( x ) . Therefore, the equation Q n 1 ( x ) 1 = 0 has at most one solution for each of the intervals. Taking into account that the intervals ( ξ 1 , ξ 1   ) , ( ξ 1 , ξ 2   ) ,…,   ( ξ n 2 2 , ξ n 2 1   ) , ( ξ n 2 1 , +   ) , solutions of the equation Q n 1 ( x ) 1 = 0 are respectively x 1 , x 2 , , x n 2 ; therefore, ψ 0 does not exist, and i = 1 , n 2 ¯ : ψ x i and Q n 1 ( ψ ) 1 = 0 .
Let us show that x i is a root of multiplicity one of the equation Q n 1 ( x ) 1 = 0 . We suppose that there exists k , for which x k is a root of multiplicity greater than one of Q n 1 ( x ) 1 = 0 ; therefore, x k is also a root of the equation i = 1 , n 2 1 ¯ : ± ξ i and x k .
Provided that deg F ( x ) = n 2 , we have come to a contradiction. Therefore x i is a root of multiplicity one of the equation Q n 1 ( x ) 1 = 0 , so if there exists γ R for which the condition Z n ( γ ) = 0 is satisfied, then γ < 0 and one of the two conditions x 0 : Z n ( x ) > 0 or x 0 : Z n ( x ) < 0 hold.
Since Z n ( 0 ) = Q n 1 ( 0 ) 1 i = 1 n 2 ( x i ) = ( 1 ) n 2 + 1 i = 1 n 2 x i , then if n 2 is an even number, then x 0 : Z n ( x ) < 0 , otherwise x 0 : Z n ( x ) > 0 .
Let us consider the function R n ( x ) , given by the following formula:
R n ( x ) = Z n ( x ) j = 1 n 2 ( x + x j )
The function R n ( x ) is continuous on the interval [ 0 , 1 ] . According to the Weierstrass theorem, it is bounded; that is, there exist x m R , x M R [ 0 , 1 ] such that x [ 0 , 1 ] : R n ( x m R ) R ( x ) R n ( x M R ) . Considering that x [ 0 , 1 ] : j = 1 n 2 ( x + x j ) > 0 , we find that if n 2 is even number, then R n ( x m R ) < R n ( x M R ) < 0 . Otherwise, 0 < R n ( x m R ) < R n ( x M R ) . If n 2 is even number, τ = R n ( x M R ) ; otherwise, τ = R n ( x m R ) and we find the function Q n 0 ( x ) = τ i = 1 n / 2 ( x 2 x i 2 ) satisfying x [ 0 , 1 ] : | 1 Q n 1 ( x ) | Q n 0 ( x ) | 1 Q n 1 ( x ) | . Since Q n ( x ) = Q n 0 ( x ) + Q n 1 ( x ) , it follows from Corollary 2 that ϕ [ 0 , π 2 ] : sin 2 ϕ Q n ( x ) + cos 2 ϕ Q n 1 ( x ) = Q n 1 ( x ) , so Q ϕ , n ( x ) = sin 2 ϕ Q n ( x ) + cos 2 ϕ Q n 1 ( x ) is the PBAS and Q ϕ , n ( x ) = sin 2 ϕ Q n ( x ) + cos 2 ϕ Q n 1 ( x ) = sin 2 ϕ Q n 0 ( x ) + Q n 1 ( x ) . It is also worth noting that Q ¯ n ( x ) = Q n 0 ( x ) + Q n 1 ( x ) is a PBAS, so Q ¯ ϕ , n ( x ) = sin 2 ϕ Q ¯ n ( x ) + cos 2 ϕ Q n 1 ( x ) = sin 2 ϕ Q n 0 ( x ) + Q n 1 ( x ) is the PBAS.
The theorem is proven. □
Example 3.
Construct the general form PBAS for n = 4 .
Solution follows from Example 2 that Q 4 1 ( x ) = 4 3 12 3 x 3 + 9 3 3 x . Calculating Z 4 ( x ) , we have
Z 4 ( x ) = Q 4 1 ( x ) 1 ( x 1 2 ) ( x 3 2 ) = 4 3 12 3 x 4 3 3
We calculate R 4 ( x ) and find:
R 4 ( x ) = Z 4 ( x ) ( x + 1 2 ) ( x + 3 2 ) = 4 3 12 3 x 4 3 3 ( x + 1 2 ) ( x + 3 2 ) = 4 3 3 ( x + 1 2 ) 4 ( x + 3 2 ) ( x + 1 2 ) ( x + 3 2 ) = 4 3 3 x + 3 2 4 x + 1 2
We calculate the derivative of the function R 4 ( x ) and find:
d R 4 ( x ) d x = 4 3 3 ( x + 3 2 ) 2 + 4 ( x + 1 2 ) 2  
Since there are no critical points on the segment [ 0 , 1 ] , the function R 4 ( x ) takes the maximum and minimum values at the ends of the segment. If we calculate R 4 ( 0 ) and R 4 ( 1 ) , respectively, we have: R 4 ( 0 ) = 16 3 and
R 4 ( 1 ) = 4 3 12 3 4 3 3 ( 1 + 1 2 ) ( 1 + 3 2 ) = 16 6 + 3 3
Therefore, τ = 16 6 + 3 3 and
Q 4 , μ 0 = μ ( x 2 1 4 ) ( x 2 3 4 ) = μ ( x 4 x 2 + 3 16 ) = μ U 5 ( x ) 32 x
where μ is any number satisfying the condition μ [ τ , τ ] , and U 5 ( x ) is a Chebyshev polynomial of the second kind. Thus, the PBAShas the form Q 4 , μ ( x ) = μ x 4 + 4 3 12 3 x 3 μ x 2 + 9 3 3 x + 3 16 μ .
Lemma 3.
If n is an even number, then
  i = 1 , n 2 ¯ : α i = j = 1 ,   j i n 2 ( sin 2 i π n + 2 sin 2 j π n + 2 ) = ( 1 ) n 2 i n + 2 2 n sin 2 2 i π n + 2
Proof. 
As x , y R : sin 2 x sin 2 y = sin ( x y ) sin ( x + y ) , then
α i = j = 1 , i j n / 2 sin ( i + j ) π n + 2 sin ( i j ) π n + 2
Consider two cases.
Case 1: If i = n 2 then
α n 2 = j = 1 n 2 1   sin ( n 2 + j ) π n + 2 sin ( n 2 j ) π n + 2 = j = 1 n 1 sin j π n + 2 sin n π 2 n + 4
Because j = 1 n + 1 sin j π n + 2 = n + 2 2 n + 1 , we have
α n 2 = n + 2 2 n sin 2 2 π n + 2
Case 2. If i n 2 then
α i = 1 sin 2 i π n + 2 sin i π n + 2 j = i n 2 1 sin j π n + 2 j = 1 i + n 2 sin j π n + 2
Because sin j π n + 2 = sin ( n + 2 + j ) π n + 2 , we obtain
α i = ( 1 ) n 2 i sin 2 i π n + 2 sin i π n + 2 j = i n 2 1 sin ( n + 2 + j ) π n + 2 j = 1 i + n 2 sin j π n + 2 = ( 1 ) n 2 i sin 2 i π n + 2 sin i π n + 2 j = n 2 + i + 2 n + 1 sin j π n + 2 j = 1 i + n 2 sin j π n + 2 = ( 1 ) n 2 i sin 2 i π n + 2 sin i π n + 2 sin ( n 2 + i + 1 ) π n + 2 j = 1 n + 1 sin j π n + 2
As j = 1 n + 1 sin j π n + 2 = n + 2 2 n + 1 , sin ( n 2 + i + 1 ) π n + 2 = cos i π n + 2 , and 2 cos i π n + 2 sin i π n + 2 = sin 2 i π n + 2 ,
α i = ( 1 ) n 2 i sin 2 2 i π n + 2 2 n n + 2
Lemma 3 is proven. □
Theorem 3.
If n is an even number, then PBAS is defined as
Q μ , n ( x ) = μ i = 1 n / 2 ( x 2 x i 2 ) + Q n 1 ( x ) ,  
where μ [ τ , τ ] , x i = sin i π n + 2 , and τ = 2 n + 1 n + 2 tan π 2 n + 4 .
Proof. 
Using the theorem on the expansion of rational functions in the case of different roots [14], we represent R n ( x ) as partial fraction decomposition:
R n ( x ) = Z n ( x ) j = 1 n / 2 ( x + x j ) = j = 1 n / 2 b j x + x j ,
where j = 1 , n 2 ¯ : b j R . Therefore, we have
Z n ( x ) = j = 1 n / 2 b j i = 1 , i j   n / 2 ( x + x i )
Calculating the values of Z n ( x ) at the point x = x j , we obtain:
Z n ( x j ) = b j i = 1 , i j   n / 2 ( x i x j )
On the other hand, Z n ( x ) = Q n 1 ( x ) 1 i = 1 n / 2 ( x x i ) , hence
Z n ( x j ) = 2 i = 1 n / 2 ( x j x i ) = ( 1 ) n 2 + 1 2 i = 1 n / 2 ( x j + x i )
Since Z n ( x j ) = b j i = 1 , i j   n / 2 ( x i x j ) = ( 1 ) n 2 + 1 2 i = 1 n / 2 ( x j + x i ) , it follows that
b j = ( 1 ) n 2 + 1 1 x j i = 1 , i j   n 2 ( x i 2 x j 2 ) = 1 x j i = 1 , i j   n 2 ( x j 2 x i 2 ) .
Using Lemma 3, we find
b j = ( 1 ) n 2 + j 2 n n + 2 ˙ x 2 j 2 x j .
Therefore,
R n ( x ) = ( 1 ) n 2 2 n n + 2 j = 1 n 2 ( 1 ) j x 2 j 2 x j 1 x + x j = ( 1 ) n 2 2 n + 2 n + 2 j = 1 n 2 ( 1 ) j x j x j 3 x + x j .
Calculating d R n ( x ) d x , we have
d R n ( x ) d x = j = 1 n 2 b j ( x + x j ) 2 .
Let us show that x [ 0 , 1 ] :   d R n ( x ) d x 0 . Using the corollary of the Cauchy–Schwarz inequality ( i = 1 n u i v i ) 2 ( i = 1 n v i ) ( i = 1 n u i 2 v i ) , we have
( b j > 0 b j x + x j   ) 2 ( b j > 0 b j ) ( b j > 0   b j ( x + x j ) 2 ) ( b j < 0 b j x + x j   ) 2 ( b j < 0 b j ) ( b j < 0   b j ( x + x j ) 2 )
Therefore,
( b j > 0 b j x + x j   ) 2 b j > 0 b j b j > 0   b j ( x + x j ) 2 ( b j < 0 b j x + x j   ) 2 b j < 0 b j b j < 0   b j ( x + x j ) 2
Let us add two inequalities:
( b j > 0 b j x + x j   ) 2 b j > 0 b j + ( b j < 0 b j x + x j   ) 2 b j < 0 b j j = 1 n / 2 b j ( x + x j ) 2
As x [ 0 , 1 ] :   ( b j > 0 b j x + x j   ) 2 b j > 0 b j > 0 and ( b j < 0 b j x + x j   ) 2 b j < 0 b j > 0 then j = 1 n / 2 b j ( x + x j ) 2 > 0 . Therefore, d R n ( x ) d x does not change sign on the interval [0, 1]. The minimum and maximum of the function R n ( x ) will be reached at the ends of the interval. Let us calculate the value of the function R n ( x ) at the points x = 0 and x = 1 :
R n ( 0 ) = ( 1 ) n 2 2 n + 2 n + 2 j = 1 n / 2 ( 1 ) j ( 1 x j 2 ) R n ( 1 ) = ( 1 ) n 2 2 n + 2 n + 2 j = 1 n / 2 ( 1 ) j ( x j x j 2 )
Considering that
j = 1 n / 2 ( 1 ) j = 1 + ( 1 ) n / 2 2 , j = 1 n / 2 ( 1 ) j x j = ( 1 ) n / 2 2 1 2 tan π 2 n + 4 , j = 1 n / 2 ( 1 ) j x j 2 = ( 1 ) n / 2 2 ,
we have
R n ( 0 ) = ( 1 ) n 2 + 1 2 n + 1 n + 2 R n ( 1 ) = ( 1 ) n 2 + 1 2 n + 1 n + 2 tan π 2 n + 4
As n 2 and | n | 2 = 0 : | R n ( 0 ) | > | R n ( 1 ) | , considering Theorem 2, we obtain
τ = | R n ( 1 ) | = 2 n + 1 n + 2 tan π 2 n + 4 .
The theorem is proven. □

7. Conclusions

Homomorphic encryption enables the computing of encrypted data without access to the secret key. It has become a promising mechanism for the secure computation, storage, and communication of confidential data in cloud services [15]. Practical scenarios include robot control systems, machine learning models, image processing, and many others [6,7,8,9,10,16,17,18]. A challenge of processing encrypted information is finding a cryptographically compatible sign function approximation.
State-of-the-art works have mainly focused on constructing the polynomial of best approximation of the sign function (PBAS) on the union of the intervals [ 1 , ϵ ] [ ϵ , 1 ] . In this paper, we provide a construction of the PBAS on the complete interval [ 1 , 1 ] and prove that:
If n = 0 , then PBAS has the form Q n ( x ) = a 0 , where | a 0 | 1 .
If n 1 , then there is a unique PBAS odd function, which can be calculated using the zeros of the Chebyshev polynomial of the second kind.
If n 1 and n is an odd number, then there are no PBAS of the general form.
If n 1 and n is an even number, then there is an uncountable set of PBAS of the general form.
Future studies include assessing the accuracy and efficiency of PBAS on real systems, e.g., over privacy-preserving neural networks with homomorphic encryption, where the non-linear activation function is replaced with a PBAS to operate with encrypted data.

Author Contributions

Conceptualization, M.B., A.T., A.A. and F.G.; Data curation, B.P.-G. and F.G.; Formal analysis, M.B.; Investigation, M.B., A.T., B.P.-G., A.A., S.N., X.W. and F.G.; Methodology, A.A., S.N., X.W. and F.G.; Project administration, A.A.; Resources, S.N., X.W. and F.G.; Software, B.P.-G. and F.G.; Supervision, A.T. and A.A.; Validation, M.B., B.P.-G., S.N., X.W. and F.G.; Visualization, M.B. and B.P.-G.; Writing—original draft, M.B. and A.T.; Writing—review & editing, A.T. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Ministry of Education and Science of the Russian Federation (Project 075-15-2020-915).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Cheon, J.H.; Kim, D.; Kim, D.; Lee, H.H.; Lee, K. Numerical Method for Comparison on Homomorphically Encrypted Numbers. Lect. Notes Comput. Sci. 2019, 11922, 415–445. [Google Scholar] [CrossRef]
  2. Cheon, J.H.; Kim, D.; Kim, D. Efficient Homomorphic Comparison Methods with Optimal Complexity. Lect. Notes Comput. Sci. 2020, 12492, 221–256. [Google Scholar]
  3. Chen, H.; Chillotti, I.; Song, Y. Improved Bootstrapping for Approximate Homomorphic Encryption. Lect. Notes Comput. Sci. 2019, 11477, 34–54. [Google Scholar]
  4. Han, K.; Ki, D. Better bootstrapping for approximate homomorphic encryption. Lect. Notes Comput. Sci. 2020, 12006, 364–390. [Google Scholar] [CrossRef]
  5. Boura, C.; Gama, N.; Georgieva, M. Chimera: A unified framework for B/FV, TFHE and HEAAN fully homomorphic encryption and predictions for deep learning. IACR Cryptol. ePrint Arch. 2018, 2018, 758. [Google Scholar]
  6. Gilad-Bachrach, R.; Dowlin, N.; Laine, K.; Lauter, K.; Naehrig, M.; Wernsing, J. Cryptonets: Applying neural networks to encrypted data with high throughput and accuracy. In Proceedings of the 33rd International Conference on Machine Learning, New York, NY, USA, 19–24 June 2016; Volume 48, pp. 201–210. Available online: http://proceedings.mlr.press/v48/gilad-bachrach16.pdf (accessed on 10 April 2022).
  7. Kim, M.; Song, Y.; Wang, S.; Xia, Y.; Jiang, X. Secure logistic regression based on homomorphic encryption: Design and evaluation. JMIR Med. Inform. 2018, 6, e19. [Google Scholar] [CrossRef] [PubMed]
  8. Bonte, C.; Vercauteren, F. Privacy-preserving logistic regression training. BMC Med. Genom. 2018, 11, 86. [Google Scholar] [CrossRef] [PubMed]
  9. Kim, A.; Song, Y.; Kim, M.; Lee, K.; Cheon, J.H. Logistic regression model training based on the approximate homomorphic encryption. BMC Med. Genom. 2018, 11, 83. [Google Scholar] [CrossRef] [PubMed]
  10. Bajard, J.-C.; Martins, P.; Sousa, L.; Zucca, V. Improving the Efficiency of SVM Classification with FHE. IEEE Trans. Inf. Forensics Secur. 2019, 15, 1709–1722. [Google Scholar] [CrossRef]
  11. Bakhvalov, N.S.; Zhidkov, N.P.; Kobelkov, G.M. Numerical Methods; Nauka: Moscow, Russia, 1987; p. 600. [Google Scholar]
  12. Korkine, A.; Zolotareff, G. Sur un certain minimum. Nouv. Ann. Mathématiques J. Candidats Écoles Polytech. Norm. 1873, 12, 337–355. Available online: http://www.numdam.org/item/NAM_1873_2_12__337_0/ (accessed on 7 November 2019).
  13. Zygmund, A. Trigonometric Series; Cambridge University Press: New York, NY, USA, 2002; Volume 1, p. 375. [Google Scholar]
  14. Graham, R.L.; Knuth, D.E.; Patashnik, O. Concrete Mathematics; Addison-Wesley Publishing Company: Reading, MA, USA, 1994; p. 625. [Google Scholar]
  15. Tchernykh, A.; Babenko, M.; Chervyakov, N.; Miranda-Lopez, V.; Avetisyan, A.; Drozdov, A.Y.; Rivera-Rodriguez, R.; Radchenko, G.; Du, Z. Scalable Data Storage Design for Nonstationary IoT Environment with Adaptive Security and Reliability. IEEE Internet Things J. 2020, 7, 10171–10188. [Google Scholar] [CrossRef]
  16. Pulido-Gaytan, B.; Tchernykh, A.; Cortés-Mendoza, J.M.; Babenko, M.; Radchenko, G.; Avetisyan, A.; Drozdov, A.Y. Privacy-preserving neural networks with Homomorphic encryption: Challenges and opportunities. Peer-to-Peer Netw. Appl. 2021, 14, 1666–1691. [Google Scholar] [CrossRef]
  17. Babenko, M.; Tchernykh, A.; Chervyakov, N.; Kuchukov, V.; Miranda-López, V.; Rivera-Rodriguez, R.; Du, Z.; Talbi, E.-G. Positional Characteristics for Efficient Number Comparison over the Homomorphic Encryption. Program. Comput. Softw. 2019, 45, 532–543. [Google Scholar] [CrossRef]
  18. Cortes-Mendoza, J.M.; Radchenko, G.; Tchernykh, A.; Pulido-Gaytan, B.; Babenko, M.; Avetisyan, A.; Bouvry, P.; Zomaya, A. LR-GD-RNS: Enhanced Privacy-Preserving Logistic Regression Algorithms for Secure Deployment in Untrusted Environments. In Proceedings of the 2021 IEEE/ACM 21st International Symposium on Cluster, Cloud and Internet Computing (CCGrid), Melbourne, Australia, 10–13 May 2021; IEEE: Melbourne, Australia, 2021; pp. 770–775. [Google Scholar]
Figure 1. The set of possible values m Q and M Q .
Figure 1. The set of possible values m Q and M Q .
Mathematics 10 02006 g001
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Babenko, M.; Tchernykh, A.; Pulido-Gaytan, B.; Avetisyan, A.; Nesmachnow, S.; Wang, X.; Granelli, F. Towards the Sign Function Best Approximation for Secure Outsourced Computations and Control. Mathematics 2022, 10, 2006. https://doi.org/10.3390/math10122006

AMA Style

Babenko M, Tchernykh A, Pulido-Gaytan B, Avetisyan A, Nesmachnow S, Wang X, Granelli F. Towards the Sign Function Best Approximation for Secure Outsourced Computations and Control. Mathematics. 2022; 10(12):2006. https://doi.org/10.3390/math10122006

Chicago/Turabian Style

Babenko, Mikhail, Andrei Tchernykh, Bernardo Pulido-Gaytan, Arutyun Avetisyan, Sergio Nesmachnow, Xinheng Wang, and Fabrizio Granelli. 2022. "Towards the Sign Function Best Approximation for Secure Outsourced Computations and Control" Mathematics 10, no. 12: 2006. https://doi.org/10.3390/math10122006

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop