Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • Main

    Uploaded by

    VIVEK KOKATE
    4 views3 pages

    Original Title

    main

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    4 views3 pages

    Main

    Uploaded by

    VIVEK KOKATE

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    of 3

    A convex program for inferring a formula from GCMS

    steventeddy
    July 8, 2023

    1 Notation

    Table 1: Symbols
    Term Description
    [n] shorthand for {1, 2, . . . , n} Pn
    ∆n probability simplex over n dimensions, ∆n = {p ∈ Rn : 0 ≤ pi ≤ 1, i ∈ [n], i=1 pi = 1}
     elementwise inequality
    1n ones vector of length n
    0n zeros vector of length n
    In identity matrix of dimension n

    Table 2: Variables
    Term Description
    n # of aromachemicals present in GCMS
    m # of naturals
    y formula from GCMS, a vector in ∆n
    x our formula, with naturals, a vector in ∆n+m
    A a column stochastic matrix of size n × m representing our materials, each column ai ∈ ∆n

    2 Original Problem
    Assumptions:
    1. If the GCMS of a perfume leaves out some material, then that material is left out of the GCMS’ of all
    naturals.
    2. We know the exact composition of each natural used in the perfume.
    Goal:
    1. Approximate the aromachemicals in the GCMS, i.e. the vector y by some convex combination x of the
    materials.
    The result of the GCMS y ∈ ∆n is the combination of aromachemicals (and possibly) naturals. Denote
    the combination by x ∈ ∆n+m . Write
    x> = (xac xnat )>

    1
    where xac represents the amount of direct adds, and xnat is the indirect additions via naturals.
    Fix the order of aromachemicals in the GCMS and say linalool is the first aromachemical. This means
    that y1 is the amount of linalool in the perfume. Thus, each aromachemical is represented as a canonical
    basis vector, ei ∈ [0, 1]n where ei is 1 in the ith position and 0 elsewhere. The contribution of aromachemicals
    is In xac . I.e. one could get the GCMS result by directly adding every aromachemical.
    Let B ∈ [0, 1]n×m represent the column stochastic matrix of naturals. Let the columns of B be written
    as
    B = (b1 , b2 , . . . , bm ).
    B being column stochastic means bi ∈ ∆n for all i ∈ [m]. Say the first natural is lavender 40/42, and that
    linalool/linalyl acetate are the first two aromachemicals in that order. Then,

    b11 = 0.4 and b12 = 0.42,

    recording the amounts of linalool and linalyl acetate. Bxnat records the contributions to the n aromachem-
    icals from the naturals. I.e. take the column combination view of matrix vector multiplication.
    If we concatenate the matrices horizontally, we have A ∈ [0, 1]n×(n+m) column stochastic and defined as
    
    A = In B .

    The ordinary least squares problem is

    min kIn xac + Bxnat − yk22


    x∈Rn+m
    min kAx − yk22 s.t. 0n  xac  1n
    x
    or equivalently
    s.t. x ∈ ∆n+m 0m  xnat  1m
    xac> 1n + xnat> 1m = 1

    3 Letting the Compositions of Naturals be Unknown


    We want to drop assumption 2 from above. Rather than knowing the exact composition of natural i (say
    lavender EO) as bi ∈ ∆n , say we know upper and lower bounds for each of its constituents. Say

    b`i  bi  bui .

    Of course, the proportions of materials in a natural has to sum to 1, so the possible lavenders given the
    bounds are in the set
    {b ∈ ∆n : b`i  bi  bui }
    One approach is to formulate the new problem where we try to infer the composition of the naturals as

    min kxac + Bxnat − yk22


    x∈Rn+m
    B∈Rn×m

    s.t. b`i  bi  bui , ∀i ∈ [m]


    0n  bi  1n , ∀i ∈ [m]
    b>
    i 1n = 1, ∀i ∈ [m]
    0n  xac  1n
    0m  xnat  1m
    xac> 1n + xnat> 1m = 1

    Suppose instead that we write the B from above (which can vary) as B ref + P where P is a matrix of
    pertubations and is the same size as B ref . B ref can be the compositions of naturals you own. Of course,

    2
    we require that B ref + P is column stochastic. Letting pi be the ith column of P and putting arrows on
    vectors, the term in the objective is
    m
    X m
    X
    Bxnat = (B ref + P )xnat = ~bi xi + p~i xi
    i=1 i=1

    We want to write the perturbation term (the second sum) without reference to x. We will substitute ~νi for
    p~i xi and show how to change the constraints.
    Writing the constraints for column i, substituting ~bi = ~bref
    i + p~i ,

    ~b`  ~bi  ~bu ⇒ ~b`  ~bref + p~i  ~bu


    i i i i i

    0n  bi  1n ⇒ 0n  ~bref + p~i  1n
    i

    b>
    i 1n =1 ⇒ (~bref
    i + p~i )> 1n = 1

    Putting in ~νi /xi = p~i ,

    ~b`  ~bi  ~bu ⇒ ~b` − ~bref  ~νi  ~bu − ~bref ⇒ xi (~b`i − ~bref νi  xi (~bui − ~bref
    i i i i i i i )~ i )
    xi
    ~νi
    0 n  bi  1 n ⇒ −~bref
    i   1n − ~bref
    i ⇒ −xi~bref
    i  ~νi  xi (1n − ~bref
    i )
    xi
     >
    b> ⇒ ~bref + ~νi ⇒ ~νi> 1n = 0
    i 1n = 1 i 1n = 1
    xi

    where the last statement is due to ~bi ∈ ∆n by assumption.


    An equivalent formulation is
    m 2
    X
    min xac + B ref xnat + ~νi − y
    x∈Rn+m
    ~
    νi ,i∈[m] i=1 2

    s.t. xi (~b`i− ~bref


    i )~νi  xi (~bui − ~bref
    i ), ∀i ∈ [m]
    ref ref
    − xi~bi  ~νi  xi (1n − ~bi ), ∀i ∈ [m]
    ~νi> 1n = 0, ∀i ∈ [m]
    0n  xac  1n
    0m  xnat  1m
    xac> 1n + xnat> 1m = 1

    The program has a quadratic objective and linear constraints.

    You might also like