Introduction Linear constraints Conclusion Elimination of variables Presentation for Nonlinear optimisation, equations and least squares Anders Märak Leffler April 18, 2012 Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Introduction Talk on Nocedal & Wright ch 15.2-3 Motivation: reduce number of free variables Simpler problem. Complete characterisation (and all eq. constraints) gives unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Introduction Talk on Nocedal & Wright ch 15.2-3 Motivation: reduce number of free variables Simpler problem. Complete characterisation (and all eq. constraints) gives unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Introduction Talk on Nocedal & Wright ch 15.2-3 Motivation: reduce number of free variables Simpler problem. Complete characterisation (and all eq. constraints) gives unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Introduction Talk on Nocedal & Wright ch 15.2-3 Motivation: reduce number of free variables Simpler problem. Complete characterisation (and all eq. constraints) gives unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Introduction Talk on Nocedal & Wright ch 15.2-3 Motivation: reduce number of free variables Simpler problem. Complete characterisation (and all eq. constraints) gives unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Complications from inequality constraints Searching for candidates using necessary conditions looks ”simple”. Test out a working set W ⊆ I of active ineq-constraints and look at results. ci1 Ie find x: ∇f (x) = λT ∇c(x), c = ... , ij ∈ W . λ corresponding. cik Check that λ ≥ 0. Problem: which constraints are active? At most 2|I| such subsets. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion On inequality constraints A warning re nonlinear constraints A warning on nonlinear constraints One can reduce problems such as minx f (x) = f (x1 , x2 , x3 , x4 ) x1 + x32 − x4 x3 = 0, st −x2 + x4 + x32 = 0 to a two-variable problem by x1 = x4 x3 − x32 , x2 = x4 + x32 . In general: much harder to do mechanically. Consider min x 2 + y 2 , st(x − 1)3 = y 2 . [xy ] y 2 ≥ 0 and so x − 1 ≥ 0 is obvious only on inspection. Rest of talk (and ch 15.3): consider linear approximations. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation The problem We start with st min f (x) cj (x) = 0, j ∈ E Approximate with m linear constraints st min f (x) aiT x = bi , i = 1, . . . , m T With A = a1 . . . am , b = [bi ] min f (x), Anders Märak Leffler st Ax = b Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation The problem We start with st min f (x) cj (x) = 0, j ∈ E Approximate with m linear constraints st min f (x) aiT x = bi , i = 1, . . . , m T With A = a1 . . . am , b = [bi ] min f (x), Anders Märak Leffler st Ax = b Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation The problem We start with st min f (x) cj (x) = 0, j ∈ E Approximate with m linear constraints st min f (x) aiT x = bi , i = 1, . . . , m T With A = a1 . . . am , b = [bi ] min f (x), Anders Märak Leffler st Ax = b Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Simple elimination Assume A is m × n, full row rank (ie check feasibility, remove redundant constraints). Select basis variables, permute to front. AP = B|N , B basis for Rm . Similarly for variables. [xB T xN ] = P T x Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Simple elimination Assume A is m × n, full row rank (ie check feasibility, remove redundant constraints). Select basis variables, permute to front. AP = B|N , B basis for Rm . Similarly for variables. [xB T xN ] = P T x Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Simple elimination Assume A is m × n, full row rank (ie check feasibility, remove redundant constraints). Select basis variables, permute to front. AP = B|N , B basis for Rm . Similarly for variables. [xB T xN ] = P T x Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Simple elimination Assume A is m × n, full row rank (ie check feasibility, remove redundant constraints). Select basis variables, permute to front. AP = B|N , B basis for Rm . Similarly for variables. [xB T xN ] = P T x Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Simple elimination Assume A is m × n, full row rank (ie check feasibility, remove redundant constraints). Select basis variables, permute to front. AP = B|N , B basis for Rm . Similarly for variables. [xB T xN ] = P T x Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x b = Ax = A(PP T )x = (AP)(P T x) = [|{z} B | |{z} N ] B = BxB +NxN xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Then x B | |{z} N ] B = BxB +NxN b = Ax = A(PP T )x = (AP)(P T x) = [|{z} xN m×m m×(m−n) RHS first part has unique solution, b ∈ Rm . Thus xB = B −1 b − B −1 NxN Indirectly: parameter solution of Ax = b, select parameters. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation We had: min Now min f (x), s t Ax = b, x ∈ Rn −1 xB B b − B −1 NxN f P =f P xN xN Fewer variables (xN ∈ Rn−m ). Unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation We had: min Now min f (x), s t Ax = b, x ∈ Rn −1 xB B b − B −1 NxN f P =f P xN xN Fewer variables (xN ∈ Rn−m ). Unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation We had: min Now min f (x), s t Ax = b, x ∈ Rn −1 xB B b − B −1 NxN f P =f P xN xN Fewer variables (xN ∈ Rn−m ). Unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation We had: min Now min f (x), s t Ax = b, x ∈ Rn −1 xB B b − B −1 NxN f P =f P xN xN Fewer variables (xN ∈ Rn−m ). Unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation We had: min Now min f (x), s t Ax = b, x ∈ Rn −1 xB B b − B −1 NxN f P =f P xN xN Fewer variables (xN ∈ Rn−m ). Unconstrained problem. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation An example 8 3 | min f (x1 , x2 , x3 , x4 , x5 , x6 ), st −6 1 9 4 0 6 x= , x ∈ R6 2 0 −1 6 4 −4 {z } A Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation An example Pick x3 , x6 as basis variables. Permute to front. 1 0 | 8 −6 9 4 6 Px = 0 4 | 3 2 −1 6 −4 In old notation: 1 0 1 0 −1 B= ,B = , 0 4 0 14 8 −6 9 4 N= 3 2 −1 6 Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation An example Pick x3 , x6 as basis variables. Permute to front. 1 0 | 8 −6 9 4 6 Px = 0 4 | 3 2 −1 6 −4 In old notation: 1 0 1 0 −1 B= ,B = , 0 4 0 14 8 −6 9 4 N= 3 2 −1 6 Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation An example Pick x3 , x6 as basis variables. Permute to front. 1 0 | 8 −6 9 4 6 Px = 0 4 | 3 2 −1 6 −4 In old notation: 1 0 1 0 −1 B= ,B = , 0 4 0 14 8 −6 9 4 N= 3 2 −1 6 Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Thus x1 8 6 9 4 0 4 x2 − 3 1 1 3 1 −6 − 4 2 x4 4 4 2 {z } x5 B −1 b {z } | B −1 Nx N 6 − 8x1 − 6x2 − 9x4 − 4x5 = −1 − 34 x1 − 12 x2 + 14 x4 − 32 x5 1 x xB = 3 = x6 0 | Yielding min [x1 ,x2 ,x4 ,x5 ] f x1 , x2 , 6 − 8x1 . . . , x4 , x5 , −1 − Anders Märak Leffler 3 x1 . . . 4 Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Thus x1 0 8 6 9 4 4 x2 − 3 1 1 1 3 −6 − 4 2 x4 4 4 2 {z } x5 B −1 b {z } | B −1 Nx N 6 − 8x1 − 6x2 − 9x4 − 4x5 = −1 − 34 x1 − 12 x2 + 14 x4 − 32 x5 1 x xB = 3 = x6 0 | Yielding min [x1 ,x2 ,x4 ,x5 ] f x1 , x2 , 6 − 8x1 . . . , x4 , x5 , −1 − Anders Märak Leffler 3 x1 . . . 4 Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Thus x1 0 8 6 9 4 4 x2 − 3 1 1 1 3 −6 − 4 2 x4 4 4 2 {z } x5 B −1 b {z } | B −1 Nx N 6 − 8x1 − 6x2 − 9x4 − 4x5 = −1 − 34 x1 − 12 x2 + 14 x4 − 32 x5 1 x xB = 3 = x6 0 | Yielding min [x1 ,x2 ,x4 ,x5 ] f x1 , x2 , 6 − 8x1 . . . , x4 , x5 , −1 − Anders Märak Leffler 3 x1 . . . 4 Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Thus x1 0 8 6 9 4 4 x2 − 3 1 1 1 3 −6 − 4 2 x4 4 4 2 {z } x5 B −1 b {z } | B −1 Nx N 6 − 8x1 − 6x2 − 9x4 − 4x5 = −1 − 34 x1 − 12 x2 + 14 x4 − 32 x5 1 x xB = 3 = x6 0 | Yielding min [x1 ,x2 ,x4 ,x5 ] f x1 , x2 , 6 − 8x1 . . . , x4 , x5 , −1 − Anders Märak Leffler 3 x1 . . . 4 Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Linear algebra review Given a transformation x → Ax, A is m × n, rank (A) = p we say that Ran(A) = span{A1 , A2 , . . . , An } = span{Ai1 , Ai2 , . . . , Aip } ⊆ Rm where Aij linear independent. rank (A) = #linear independent cols/rows of A = dim Ran(A). dim Ran(A) + dim Null(A) = n (dimension theorem, transformation Rn → Rm ) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Linear algebra review Given a transformation x → Ax, A is m × n, rank (A) = p we say that Ran(A) = span{A1 , A2 , . . . , An } = span{Ai1 , Ai2 , . . . , Aip } ⊆ Rm where Aij linear independent. rank (A) = #linear independent cols/rows of A = dim Ran(A). dim Ran(A) + dim Null(A) = n (dimension theorem, transformation Rn → Rm ) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Linear algebra review Given a transformation x → Ax, A is m × n, rank (A) = p we say that Ran(A) = span{A1 , A2 , . . . , An } = span{Ai1 , Ai2 , . . . , Aip } ⊆ Rm where Aij linear independent. rank (A) = #linear independent cols/rows of A = dim Ran(A). dim Ran(A) + dim Null(A) = n (dimension theorem, transformation Rn → Rm ) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Linear algebra review Given a transformation x → Ax, A is m × n, rank (A) = p we say that Ran(A) = span{A1 , A2 , . . . , An } = span{Ai1 , Ai2 , . . . , Aip } ⊆ Rm where Aij linear independent. rank (A) = #linear independent cols/rows of A = dim Ran(A). dim Ran(A) + dim Null(A) = n (dimension theorem, transformation Rn → Rm ) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Linear algebra review Given a transformation x → Ax, A is m × n, rank (A) = p we say that Ran(A) = span{A1 , A2 , . . . , An } = span{Ai1 , Ai2 , . . . , Aip } ⊆ Rm where Aij linear independent. rank (A) = #linear independent cols/rows of A = dim Ran(A). dim Ran(A) + dim Null(A) = n (dimension theorem, transformation Rn → Rm ) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Null(A) = {x : Ax = 0}, a subspace with some basis {N1 , N2 , . . . N(n−p) } ⊆ Rn . Let Z = N1 N2 . . . N(n−p) AZ α = N1 N2 . . . N(n−p) α = 0 0 . . . 0 α = 0, α ∈ R(n−p) . Geometrically Z α ∈ Null(A), ∀α ∈ Rn−p , coordinate interpretation: α1 AZ α = AZ ... = A(α1 N1 + . . . + αn N(n−p) ) = 0 α(n−p) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Generalised strategy A as in simple elimination. Pick Y , n × m, Z , n × (n − m) such that [Y |Z ] (n × n) has full rank. All columns are linearly independent. AZ = 0, ie Ran(Z ) ⊆ Null(A) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Generalised strategy A as in simple elimination. Pick Y , n × m, Z , n × (n − m) such that [Y |Z ] (n × n) has full rank. All columns are linearly independent. AZ = 0, ie Ran(Z ) ⊆ Null(A) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Generalised strategy A as in simple elimination. Pick Y , n × m, Z , n × (n − m) such that [Y |Z ] (n × n) has full rank. All columns are linearly independent. AZ = 0, ie Ran(Z ) ⊆ Null(A) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Generalised strategy A as in simple elimination. Pick Y , n × m, Z , n × (n − m) such that [Y |Z ] (n × n) has full rank. All columns are linearly independent. AZ = 0, ie Ran(Z ) ⊆ Null(A) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Generalised strategy A as in simple elimination. Pick Y , n × m, Z , n × (n − m) such that [Y |Z ] (n × n) has full rank. All columns are linearly independent. AZ = 0, ie Ran(Z ) ⊆ Null(A) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation This has implications for Y and Z: rank (A) = m, full row rank, and Y has independent columns. rank (A[Y |Z ]) = rank ([AY |0]) = rank (AY ) = m ⇒ Ran(AY ) = Rm = Ran(A). Recall: A is m × n, So dim Null(A) + dim Ran(A) = n dim Null(A) = dim Rn − dim Ran(A) = n − rank (A) = n − m, so Null(A) = Ran(Z ) (Z has n − m linearly independent columns, all in Null(A)) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Express feasible x as x = YxY + ZxZ (ok, as Y has rank m). Ax = A (YxY + ZxZ ) = AYxY + 0 = b, xY = (AY )−1 b. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Express feasible x as x = YxY + ZxZ (ok, as Y has rank m). Ax = A (YxY + ZxZ ) = AYxY + 0 = b, xY = (AY )−1 b. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Express feasible x as x = YxY + ZxZ (ok, as Y has rank m). Ax = A (YxY + ZxZ ) = AYxY + 0 = b, xY = (AY )−1 b. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Express feasible x as x = YxY + ZxZ (ok, as Y has rank m). Ax = A (YxY + ZxZ ) = AYxY + 0 = b, xY = (AY )−1 b. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Transformed problem We had: minx f (x) s t Ax = b x ∈ Rn Now minxZ f Y (AY )−1 b + ZxZ xZ ∈ Rn−m where Y (AY )−1 b is constant; displacement. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Transformed problem We had: minx f (x) s t Ax = b x ∈ Rn Now minxZ f Y (AY )−1 b + ZxZ xZ ∈ Rn−m where Y (AY )−1 b is constant; displacement. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Transformed problem We had: minx f (x) s t Ax = b x ∈ Rn Now minxZ f Y (AY )−1 b + ZxZ xZ ∈ Rn−m where Y (AY )−1 b is constant; displacement. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Outline 1 Introduction On inequality constraints A warning re nonlinear constraints 2 Linear constraints The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation 3 Conclusion Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Motivation There are many ways to select Y , Z . We want well-conditioned problem (robust w r t perturbations) QR factorisation of AT (n × m) is one way. Bad for sparse A (may be expensive), otherwise nice. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Factorisation: AT Π = [Q1 Q2 ] R 0 Where Π - permutation matrix (n × n) [Q1 Q2 ] is (n × n), orthogonal. Q1 (n × m) and Q2 (n × (n − m)) have orthogonal columns. R is (m × m), upper triangular. pause Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Factorisation: AT Π = [Q1 Q2 ] R 0 Where Π - permutation matrix (n × n) [Q1 Q2 ] is (n × n), orthogonal. Q1 (n × m) and Q2 (n × (n − m)) have orthogonal columns. R is (m × m), upper triangular. pause Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Factorisation: AT Π = [Q1 Q2 ] R 0 Where Π - permutation matrix (n × n) [Q1 Q2 ] is (n × n), orthogonal. Q1 (n × m) and Q2 (n × (n − m)) have orthogonal columns. R is (m × m), upper triangular. pause Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Factorisation: AT Π = [Q1 Q2 ] R 0 Where Π - permutation matrix (n × n) [Q1 Q2 ] is (n × n), orthogonal. Q1 (n × m) and Q2 (n × (n − m)) have orthogonal columns. R is (m × m), upper triangular. pause Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Factorisation: AT Π = [Q1 Q2 ] R 0 Where Π - permutation matrix (n × n) [Q1 Q2 ] is (n × n), orthogonal. Q1 (n × m) and Q2 (n × (n − m)) have orthogonal columns. R is (m × m), upper triangular. pause Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Let Y = Q1 , Z = Q2 . Then AY = ΠR T , AZ = 0, follows earlier criteria for Y , Z . Explicitly: x = Q1 R −T ΠT b + Q2 xZ where displacement is easy to calculate. This means that x = AT (AAT )−1 b + Q2 xZ . Looks suspiciously similar to least-squares eq. Indeed, x solves min kxk22 subject to Ax = b. Why? Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Nocedal & Wright mainly relate QR to Cholesky. Naive algorithm for A=QR more intuitive. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Nocedal & Wright mainly relate QR to Cholesky. Naive algorithm for A=QR more intuitive. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Nocedal & Wright mainly relate QR to Cholesky. Naive algorithm for A=QR more intuitive. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Constructing Q For A n × m: If column a1 = 0, set Q:s column q1 = 0, otherwise q1 = a1 ka11 k . for k = 2, 3,. . . , m Pk −1 yk = ak − i=1 (qiT ak )qi if yk = 0 then qk = 0 else qk = T 1 1 yk (yk yk ) 2 R is upper-triangular. Adapted from Horn & Johnson - Matrix analysis (Cambridge, 1985) Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Geometry Gram-Schmidt orthonormalisation. Geometry (figure 15.4) of Nocedal & Wright. Cf projection onto Ax=0 and displacement. Ax = 0 define subspace, Ax = b affine. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Geometry Gram-Schmidt orthonormalisation. Geometry (figure 15.4) of Nocedal & Wright. Cf projection onto Ax=0 and displacement. Ax = 0 define subspace, Ax = b affine. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Geometry Gram-Schmidt orthonormalisation. Geometry (figure 15.4) of Nocedal & Wright. Cf projection onto Ax=0 and displacement. Ax = 0 define subspace, Ax = b affine. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion The problem Basis- and nonbasis variables Linear algebra review Generalised strategy QR factorisation Geometry Gram-Schmidt orthonormalisation. Geometry (figure 15.4) of Nocedal & Wright. Cf projection onto Ax=0 and displacement. Ax = 0 define subspace, Ax = b affine. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion Conclusion (Undergraduate) linear algebra is useful. Anders Märak Leffler Fundamentals - Elimination of variables Introduction Linear constraints Conclusion Conclusion (Undergraduate) linear algebra is useful. Anders Märak Leffler Fundamentals - Elimination of variables