ΠΠ°ΠΊ Π΄ΠΎΠΊΠ°Π·Π°ΡΡ ΡΡΠΎ ΠΌΠ°ΡΡΠΈΡΠ° ΡΠΈΠΌΠΌΠ΅ΡΡΠΈΡΠ½Π°
ΠΠΈΠ½Π΅ΠΉΠ½Π°Ρ Π°Π»Π³Π΅Π±ΡΠ° Π΄Π»Ρ ΠΈΡΡΠ»Π΅Π΄ΠΎΠ²Π°ΡΠ΅Π»Π΅ΠΉ Π΄Π°Π½Π½ΡΡ
Β«ΠΠ°ΡΠ° [ΠΡΠ²ΠΈΠ½Π³Π° ΠΠ°ΠΏΠ»Π°Π½ΡΠΊΠΎΠ³ΠΎ ΠΈ ΠΠΎΠ»Π° Π₯Π°Π»ΠΌΠΎΡΠ°] ΠΎΠ±ΡΠ°Ρ ΡΠΈΠ»ΠΎΡΠΎΡΠΈΡ Π² ΠΎΡΠ½ΠΎΡΠ΅Π½ΠΈΠΈ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎΠΉ Π°Π»Π³Π΅Π±ΡΡ ΡΠ°ΠΊΠΎΠ²Π°: ΠΌΡ Π΄ΡΠΌΠ°Π΅ΠΌ Π² Π±Π΅Π·Π±Π°Π·ΠΈΡΠ½ΡΡ ΡΠ΅ΡΠΌΠΈΠ½Π°Ρ , ΠΏΠΈΡΠ΅ΠΌ Π² Π±Π΅Π·Π±Π°Π·ΠΈΡΠ½ΡΡ ΡΠ΅ΡΠΌΠΈΠ½Π°Ρ , Π½ΠΎ ΠΊΠΎΠ³Π΄Π° Π΄ΠΎΡ ΠΎΠ΄ΠΈΡ Π΄ΠΎ ΡΠ΅ΡΡΠ΅Π·Π½ΠΎΠ³ΠΎ Π΄Π΅Π»Π°, ΠΌΡ Π·Π°ΠΏΠΈΡΠ°Π΅ΠΌΡΡ Π² ΠΎΡΠΈΡΠ΅ ΠΈ Π²ΠΎΠ²ΡΡ ΡΡΠΈΡΠ°Π΅ΠΌ Ρ ΠΏΠΎΠΌΠΎΡΡΡ ΠΌΠ°ΡΡΠΈΡΒ».
ΠΠ»Ρ ΠΌΠ½ΠΎΠ³ΠΈΡ Π½Π°ΡΠΈΠ½Π°ΡΡΠΈΡ ΠΈΡΡΠ»Π΅Π΄ΠΎΠ²Π°ΡΠ΅Π»Π΅ΠΉ Π΄Π°Π½Π½ΡΡ Π»ΠΈΠ½Π΅ΠΉΠ½Π°Ρ Π°Π»Π³Π΅Π±ΡΠ° ΡΡΠ°Π½ΠΎΠ²ΠΈΡΡΡ ΠΊΠ°ΠΌΠ½Π΅ΠΌ ΠΏΡΠ΅ΡΠΊΠ½ΠΎΠ²Π΅Π½ΠΈΡ Π½Π° ΠΏΡΡΠΈ ΠΊ Π΄ΠΎΡΡΠΈΠΆΠ΅Π½ΠΈΡ ΠΌΠ°ΡΡΠ΅ΡΡΡΠ²Π° Π² Π²ΡΠ±ΡΠ°Π½Π½ΠΎΠΉ ΠΈΠΌΠΈ ΠΏΡΠΎΡΠ΅ΡΡΠΈΠΈ.
kdnuggets
Π ΡΡΠΎΠΉ ΡΡΠ°ΡΡΠ΅ Ρ ΠΏΠΎΠΏΡΡΠ°Π»ΡΡ ΡΠΎΠ±ΡΠ°ΡΡ ΠΎΡΠ½ΠΎΠ²Ρ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎΠΉ Π°Π»Π³Π΅Π±ΡΡ, Π½Π΅ΠΎΠ±Ρ ΠΎΠ΄ΠΈΠΌΡΠ΅ Π² ΠΏΠΎΠ²ΡΠ΅Π΄Π½Π΅Π²Π½ΠΎΠΉ ΡΠ°Π±ΠΎΡΠ΅ ΡΠΏΠ΅ΡΠΈΠ°Π»ΠΈΡΡΠ°ΠΌ ΠΏΠΎ ΠΌΠ°ΡΠΈΠ½Π½ΠΎΠΌΡ ΠΎΠ±ΡΡΠ΅Π½ΠΈΡ ΠΈ Π°Π½Π°Π»ΠΈΠ·Ρ Π΄Π°Π½Π½ΡΡ .
ΠΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ²
ΠΠ»Ρ Π΄Π²ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x, y β ββΏ ΠΈΡ ΡΠΊΠ°Π»ΡΡΠ½ΡΠΌ ΠΈΠ»ΠΈ Π²Π½ΡΡΡΠ΅Π½Π½ΠΈΠΌ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΠ΅ΠΌ xα΅y
Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΡΠ»Π΅Π΄ΡΡΡΠ΅Π΅ Π²Π΅ΡΠ΅ΡΡΠ²Π΅Π½Π½ΠΎΠ΅ ΡΠΈΡΠ»ΠΎ:
ΠΠ°ΠΊ ΠΌΠΎΠΆΠ½ΠΎ Π²ΠΈΠ΄Π΅ΡΡ, ΡΠΊΠ°Π»ΡΡΠ½ΠΎΠ΅ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΠ΅ ΡΠ²Π»ΡΠ΅ΡΡΡ ΠΎΡΠΎΠ±ΡΠΌ ΡΠ°ΡΡΠ½ΡΠΌ ΡΠ»ΡΡΠ°Π΅ΠΌ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΡ ΠΌΠ°ΡΡΠΈΡ. Π’Π°ΠΊΠΆΠ΅ Π·Π°ΠΌΠ΅ΡΠΈΠΌ, ΡΡΠΎ Π²ΡΠ΅Π³Π΄Π° ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ ΡΠΎΠΆΠ΄Π΅ΡΡΠ²ΠΎ
ΠΠ»Ρ Π΄Π²ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x β βα΅, y β ββΏ (Π½Π΅ ΠΎΠ±ΡΠ·Π°ΡΠ΅Π»ΡΠ½ΠΎ ΠΎΠ΄Π½ΠΎΠΉ ΡΠ°Π·ΠΌΠ΅ΡΠ½ΠΎΡΡΠΈ) ΡΠ°ΠΊΠΆΠ΅ ΠΌΠΎΠΆΠ½ΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»ΠΈΡΡ Π²Π½Π΅ΡΠ½Π΅Π΅ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΠ΅ xyα΅ β βα΅Λ£βΏ. ΠΡΠΎ ΠΌΠ°ΡΡΠΈΡΠ°, Π·Π½Π°ΡΠ΅Π½ΠΈΡ ΡΠ»Π΅ΠΌΠ΅Π½ΡΠΎΠ² ΠΊΠΎΡΠΎΡΠΎΠΉ ΠΎΠΏΡΠ΅Π΄Π΅Π»ΡΡΡΡΡ ΡΠ»Π΅Π΄ΡΡΡΠΈΠΌ ΠΎΠ±ΡΠ°Π·ΠΎΠΌ: (xyα΅)α΅’β±Ό = xα΅’yβ±Ό, ΡΠΎ Π΅ΡΡΡ
Π‘Π»Π΅Π΄ΠΎΠΌ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β ββΏΛ£βΏ, ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°Π΅ΠΌΡΠΌ tr(A) (ΠΈΠ»ΠΈ ΠΏΡΠΎΡΡΠΎ trA), Π½Π°Π·ΡΠ²Π°ΡΡ ΡΡΠΌΠΌΡ ΡΠ»Π΅ΠΌΠ΅Π½ΡΠΎΠ² Π½Π° Π΅Π΅ Π³Π»Π°Π²Π½ΠΎΠΉ Π΄ΠΈΠ°Π³ΠΎΠ½Π°Π»ΠΈ:
Π‘Π»Π΅Π΄ ΠΎΠ±Π»Π°Π΄Π°Π΅Ρ ΡΠ»Π΅Π΄ΡΡΡΠΈΠΌΠΈ ΡΠ²ΠΎΠΉΡΡΠ²Π°ΠΌΠΈ:
ΠΠ»Ρ Π»ΡΠ±ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β ββΏΛ£βΏ: trA = trAα΅.
ΠΠ»Ρ Π»ΡΠ±ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β ββΏΛ£βΏ ΠΈ Π»ΡΠ±ΠΎΠ³ΠΎ ΡΠΈΡΠ»Π° t β β: tr(tA) = t trA.
ΠΠ»Ρ Π»ΡΠ±ΡΡ ΠΌΠ°ΡΡΠΈΡ A,B, ΡΠ°ΠΊΠΈΡ , ΡΡΠΎ ΠΈΡ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΠ΅ AB ΡΠ²Π»ΡΠ΅ΡΡΡ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΠ΅ΠΉ: trAB = trBA.
ΠΠ»Ρ Π»ΡΠ±ΡΡ ΠΌΠ°ΡΡΠΈΡ A,B,C, ΡΠ°ΠΊΠΈΡ , ΡΡΠΎ ΠΈΡ ΠΏΡΠΎΠΈΠ·Π²Π΅Π΄Π΅Π½ΠΈΠ΅ ABC ΡΠ²Π»ΡΠ΅ΡΡΡ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΠ΅ΠΉ: trABC = trBCA = trCAB (ΠΈ ΡΠ°ΠΊ Π΄Π°Π»Π΅Π΅ β Π΄Π°Π½Π½ΠΎΠ΅ ΡΠ²ΠΎΠΉΡΡΠ²ΠΎ ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ Π΄Π»Ρ Π»ΡΠ±ΠΎΠ³ΠΎ ΡΠΈΡΠ»Π° ΠΌΠ°ΡΡΠΈΡ).
TimoElliott
ΠΠΎΡΠΌΡ
ΠΠΎΡΠΌΡ β₯xβ₯ Π²Π΅ΠΊΡΠΎΡΠ° x ΠΌΠΎΠΆΠ½ΠΎ Π½Π΅ΡΠΎΡΠΌΠ°Π»ΡΠ½ΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»ΠΈΡΡ ΠΊΠ°ΠΊ ΠΌΠ΅ΡΡ Β«Π΄Π»ΠΈΠ½ΡΒ» Π²Π΅ΠΊΡΠΎΡΠ°. ΠΠ°ΠΏΡΠΈΠΌΠ΅Ρ, ΡΠ°ΡΡΠΎ ΠΈΡΠΏΠΎΠ»ΡΠ·ΡΠ΅ΡΡΡ Π΅Π²ΠΊΠ»ΠΈΠ΄ΠΎΠ²Π° Π½ΠΎΡΠΌΠ°, ΠΈΠ»ΠΈ Π½ΠΎΡΠΌΠ° lβ:
ΠΠΎΠ»Π΅Π΅ ΡΠΎΡΠΌΠ°Π»ΡΠ½ΠΎΠ΅ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½ΠΈΠ΅ ΡΠ°ΠΊΠΎΠ²ΠΎ: Π½ΠΎΡΠΌΠΎΠΉ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ Π»ΡΠ±Π°Ρ ΡΡΠ½ΠΊΡΠΈΡ f : βn β β, ΡΠ΄ΠΎΠ²Π»Π΅ΡΠ²ΠΎΡΡΡΡΠ°Ρ ΡΠ΅ΡΡΡΠ΅ΠΌ ΡΡΠ»ΠΎΠ²ΠΈΡΠΌ:
ΠΠ»Ρ Π²ΡΠ΅Ρ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x β ββΏ: f(x) β₯ 0 (Π½Π΅ΠΎΡΡΠΈΡΠ°ΡΠ΅Π»ΡΠ½ΠΎΡΡΡ).
f(x) = 0 ΡΠΎΠ³Π΄Π° ΠΈ ΡΠΎΠ»ΡΠΊΠΎ ΡΠΎΠ³Π΄Π°, ΠΊΠΎΠ³Π΄Π° x = 0 (ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½Π°Ρ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΡΡΡ).
ΠΠ»Ρ Π»ΡΠ±ΡΡ Π²Π΅ΠΊΡΠΎΡΠ° x β ββΏ ΠΈ ΡΠΈΡΠ»Π° t β β: f(tx) = |t|f(x) (ΠΎΠ΄Π½ΠΎΡΠΎΠ΄Π½ΠΎΡΡΡ).
ΠΠ»Ρ Π»ΡΠ±ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x, y β ββΏ: f(x + y) β€ f(x) + f(y) (Π½Π΅ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ ΡΡΠ΅ΡΠ³ΠΎΠ»ΡΠ½ΠΈΠΊΠ°)
ΠΡΡΠ³ΠΈΠΌΠΈ ΠΏΡΠΈΠΌΠ΅ΡΠ°ΠΌΠΈ Π½ΠΎΡΠΌ ΡΠ²Π»ΡΡΡΡΡ Π½ΠΎΡΠΌΠ° lβ
ΠΡΠ΅ ΡΡΠΈ ΠΏΡΠ΅Π΄ΡΡΠ°Π²Π»Π΅Π½Π½ΡΠ΅ Π²ΡΡΠ΅ Π½ΠΎΡΠΌΡ ΡΠ²Π»ΡΡΡΡΡ ΠΏΡΠΈΠΌΠ΅ΡΠ°ΠΌΠΈ Π½ΠΎΡΠΌ ΡΠ΅ΠΌΠ΅ΠΉΡΡΠ²Π° lp, ΠΏΠ°ΡΠ°ΠΌΠ΅ΡΡΠΈΠ·ΡΠ΅ΠΌΡΡ Π²Π΅ΡΠ΅ΡΡΠ²Π΅Π½Π½ΡΠΌ ΡΠΈΡΠ»ΠΎΠΌ p β₯ 1 ΠΈ ΠΎΠΏΡΠ΅Π΄Π΅Π»ΡΠ΅ΠΌΡΡ ΠΊΠ°ΠΊ
ΠΠΎΡΠΌΡ ΡΠ°ΠΊΠΆΠ΅ ΠΌΠΎΠ³ΡΡ Π±ΡΡΡ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Ρ Π΄Π»Ρ ΠΌΠ°ΡΡΠΈΡ, Π½Π°ΠΏΡΠΈΠΌΠ΅Ρ Π½ΠΎΡΠΌΠ° Π€ΡΠΎΠ±Π΅Π½ΠΈΡΡΠ°:
ΠΠΈΠ½Π΅ΠΉΠ½Π°Ρ Π½Π΅Π·Π°Π²ΠΈΡΠΈΠΌΠΎΡΡΡ ΠΈ ΡΠ°Π½Π³
Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎ Π·Π°Π²ΠΈΡΠΈΠΌΡ, ΡΠ°ΠΊ ΠΊΠ°ΠΊ xβ = β2xβ + xβ.
Π‘ΡΠΎΠ»Π±ΡΠΎΠ²ΡΠΌ ΡΠ°Π½Π³ΠΎΠΌ ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ Π½Π°Π·ΡΠ²Π°ΡΡ ΡΠΈΡΠ»ΠΎ ΡΠ»Π΅ΠΌΠ΅Π½ΡΠΎΠ² Π² ΠΌΠ°ΠΊΡΠΈΠΌΠ°Π»ΡΠ½ΠΎΠΌ ΠΏΠΎΠ΄ΠΌΠ½ΠΎΠΆΠ΅ΡΡΠ²Π΅ Π΅Π΅ ΡΡΠΎΠ»Π±ΡΠΎΠ², ΡΠ²Π»ΡΡΡΠ΅ΠΌΡΡ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎ Π½Π΅Π·Π°Π²ΠΈΡΠΈΠΌΡΠΌ. Π£ΠΏΡΠΎΡΠ°Ρ, Π³ΠΎΠ²ΠΎΡΡΡ, ΡΡΠΎ ΡΡΠΎΠ»Π±ΡΠΎΠ²ΡΠΉ ΡΠ°Π½Π³ β ΡΡΠΎ ΡΠΈΡΠ»ΠΎ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎ Π½Π΅Π·Π°Π²ΠΈΡΠΈΠΌΡΡ ΡΡΠΎΠ»Π±ΡΠΎΠ² A. ΠΠ½Π°Π»ΠΎΠ³ΠΈΡΠ½ΠΎ ΡΡΡΠΎΡΠ½ΡΠΌ ΡΠ°Π½Π³ΠΎΠΌ ΠΌΠ°ΡΡΠΈΡΡ ΡΠ²Π»ΡΠ΅ΡΡΡ ΡΠΈΡΠ»ΠΎ Π΅Π΅ ΡΡΡΠΎΠΊ, ΡΠΎΡΡΠ°Π²Π»ΡΡΡΠΈΡ ΠΌΠ°ΠΊΡΠΈΠΌΠ°Π»ΡΠ½ΠΎΠ΅ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎ Π½Π΅Π·Π°Π²ΠΈΡΠΈΠΌΠΎΠ΅ ΠΌΠ½ΠΎΠΆΠ΅ΡΡΠ²ΠΎ.
ΠΠΊΠ°Π·ΡΠ²Π°Π΅ΡΡΡ (Π·Π΄Π΅ΡΡ ΠΌΡ Π½Π΅ Π±ΡΠ΄Π΅ΠΌ ΡΡΠΎ Π΄ΠΎΠΊΠ°Π·ΡΠ²Π°ΡΡ), ΡΡΠΎ Π΄Π»Ρ Π»ΡΠ±ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ ΡΡΠΎΠ»Π±ΡΠΎΠ²ΡΠΉ ΡΠ°Π½Π³ ΡΠ°Π²Π΅Π½ ΡΡΡΠΎΡΠ½ΠΎΠΌΡ, ΠΏΠΎΡΡΠΎΠΌΡ ΠΎΠ±Π° ΡΡΠΈΡ ΡΠΈΡΠ»Π° Π½Π°Π·ΡΠ²Π°ΡΡ ΠΏΡΠΎΡΡΠΎ ΡΠ°Π½Π³ΠΎΠΌ A ΠΈ ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°ΡΡ rank(A) ΠΈΠ»ΠΈ rk(A); Π²ΡΡΡΠ΅ΡΠ°ΡΡΡΡ ΡΠ°ΠΊΠΆΠ΅ ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ΅Π½ΠΈΡ rang(A), rg(A) ΠΈ ΠΏΡΠΎΡΡΠΎ r(A). ΠΠΎΡ Π½Π΅ΠΊΠΎΡΠΎΡΡΠ΅ ΠΎΡΠ½ΠΎΠ²Π½ΡΠ΅ ΡΠ²ΠΎΠΉΡΡΠ²Π° ΡΠ°Π½Π³Π°:
ΠΠ»Ρ Π»ΡΠ±ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ: rank(A) β€ min(m,n). ΠΡΠ»ΠΈ rank(A) = min(m,n), ΡΠΎ A Π½Π°Π·ΡΠ²Π°ΡΡ ΠΌΠ°ΡΡΠΈΡΠ΅ΠΉ ΠΏΠΎΠ»Π½ΠΎΠ³ΠΎ ΡΠ°Π½Π³Π°.
ΠΠ»Ρ Π»ΡΠ±ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ: rank(A) = rank(Aα΅).
ΠΠ»Ρ Π»ΡΠ±ΡΡ ΠΌΠ°ΡΡΠΈΡ A β βα΅Λ£βΏ, B β βnΓp: rank(AB) β€ min(rank(A),rank(B)).
ΠΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΡΠ΅ ΠΌΠ°ΡΡΠΈΡΡ
ΠΠ²Π° Π²Π΅ΠΊΡΠΎΡΠ° x, y β ββΏ Π½Π°Π·ΡΠ²Π°ΡΡΡΡ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΡΠΌΠΈ, Π΅ΡΠ»ΠΈ xα΅y = 0. ΠΠ΅ΠΊΡΠΎΡ x β ββΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ Π½ΠΎΡΠΌΠΈΡΠΎΠ²Π°Π½Π½ΡΠΌ, Π΅ΡΠ»ΠΈ ||x||β = 1. ΠΠ²Π°Π΄ΡΠ°ΡΠ½Π°Ρ ΠΌ
Π°ΡΡΠΈΡΠ° U β ββΏΛ£βΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΠΎΠΉ, Π΅ΡΠ»ΠΈ Π²ΡΠ΅ Π΅Π΅ ΡΡΠΎΠ»Π±ΡΡ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½Ρ Π΄ΡΡΠ³ Π΄ΡΡΠ³Ρ ΠΈ Π½ΠΎΡΠΌΠΈΡΠΎΠ²Π°Π½Ρ (Π² ΡΡΠΎΠΌ ΡΠ»ΡΡΠ°Π΅ ΡΡΠΎΠ»Π±ΡΡ Π½Π°Π·ΡΠ²Π°ΡΡ ΠΎΡΡΠΎΠ½ΠΎΡΠΌΠΈΡΠΎΠ²Π°Π½Π½ΡΠΌΠΈ). ΠΠ°ΠΌΠ΅ΡΠΈΠΌ, ΡΡΠΎ ΠΏΠΎΠ½ΡΡΠΈΠ΅ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΠΎΡΡΠΈ ΠΈΠΌΠ΅Π΅Ρ ΡΠ°Π·Π½ΡΠΉ ΡΠΌΡΡΠ» Π΄Π»Ρ Π²Π΅ΠΊΡΠΎΡΠΎΠ² ΠΈ ΠΌΠ°ΡΡΠΈΡ.
ΠΠ΅ΠΏΠΎΡΡΠ΅Π΄ΡΡΠ²Π΅Π½Π½ΠΎ ΠΈΠ· ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½ΠΈΠΉ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΠΎΡΡΠΈ ΠΈ Π½ΠΎΡΠΌΠΈΡΠΎΠ²Π°Π½Π½ΠΎΡΡΠΈ ΡΠ»Π΅Π΄ΡΠ΅Ρ, ΡΡΠΎ
ΠΡΡΠ³ΠΈΠΌΠΈ ΡΠ»ΠΎΠ²Π°ΠΌΠΈ, ΡΠ΅Π·ΡΠ»ΡΡΠ°ΡΠΎΠΌ ΡΡΠ°Π½ΡΠΏΠΎΠ½ΠΈΡΠΎΠ²Π°Π½ΠΈΡ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ ΡΠ²Π»ΡΠ΅ΡΡΡ ΠΌΠ°ΡΡΠΈΡΠ°, ΠΎΠ±ΡΠ°ΡΠ½Π°Ρ ΠΈΡΡ ΠΎΠ΄Π½ΠΎΠΉ. ΠΠ°ΠΌΠ΅ΡΠΈΠΌ, ΡΡΠΎ Π΅ΡΠ»ΠΈ U Π½Π΅ ΡΠ²Π»ΡΠ΅ΡΡΡ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΠ΅ΠΉ (U β βα΅Λ£βΏ, n
Π΄Π»Ρ Π»ΡΠ±ΡΡ Π²Π΅ΠΊΡΠΎΡΠ° x β ββΏ ΠΈ ΠΎΡΡΠΎΠ³ΠΎΠ½Π°Π»ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ U β ββΏΛ£βΏ.
TimoElliott
ΠΠ±Π»Π°ΡΡΡ Π·Π½Π°ΡΠ΅Π½ΠΈΠΉ ΠΈ Π½ΡΠ»Ρ-ΠΏΡΠΎΡΡΡΠ°Π½ΡΡΠ²ΠΎ ΠΌΠ°ΡΡΠΈΡΡ
ΠΠ±Π»Π°ΡΡΡΡ Π·Π½Π°ΡΠ΅Π½ΠΈΠΉ R(A) (ΠΈΠ»ΠΈ ΠΏΡΠΎΡΡΡΠ°Π½ΡΡΠ²ΠΎΠΌ ΡΡΠΎΠ»Π±ΡΠΎΠ²) ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ Π»ΠΈΠ½Π΅ΠΉΠ½Π°Ρ ΠΎΠ±ΠΎΠ»ΠΎΡΠΊΠ° Π΅Π΅ ΡΡΠΎΠ»Π±ΡΠΎΠ². ΠΡΡΠ³ΠΈΠΌΠΈ ΡΠ»ΠΎΠ²Π°ΠΌΠΈ,
ΠΡΠ»Ρ-ΠΏΡΠΎΡΡΡΠ°Π½ΡΡΠ²ΠΎΠΌ, ΠΈΠ»ΠΈ ΡΠ΄ΡΠΎΠΌ ΠΌΠ°ΡΡΠΈΡΡ A β βα΅Λ£βΏ (ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°Π΅ΠΌΡΠΌ N(A) ΠΈΠ»ΠΈ ker A), Π½Π°Π·ΡΠ²Π°ΡΡ ΠΌΠ½ΠΎΠΆΠ΅ΡΡΠ²ΠΎ Π²ΡΠ΅Ρ Π²Π΅ΠΊΡΠΎΡΠΎΠ², ΠΊΠΎΡΠΎΡΡΠ΅ ΠΏΡΠΈ ΡΠΌΠ½ΠΎΠΆΠ΅Π½ΠΈΠΈ Π½Π° A ΠΎΠ±ΡΠ°ΡΠ°ΡΡΡΡ Π² Π½ΡΠ»Ρ, ΡΠΎ Π΅ΡΡΡ
ΠΠ²Π°Π΄ΡΠ°ΡΠΈΡΠ½ΡΠ΅ ΡΠΎΡΠΌΡ ΠΈ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΏΠΎΠ»ΡΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΡΠ΅ ΠΌΠ°ΡΡΠΈΡΡ
ΠΠ»Ρ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β ββΏΛ£βΏ ΠΈ Π²Π΅ΠΊΡΠΎΡΠ° x β ββΏ ΠΊΠ²Π°Π΄ΡΠ°ΡΠΈΡΠ½ΠΎΠΉ ΡΠΎΡΠΌΠΎΠΉ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΡΠΊΠ°Π»ΡΡΠ½ΠΎΠ΅ Π·Π½Π°ΡΠ΅Π½ΠΈΠ΅ xα΅ Ax. Π Π°ΡΠΏΠΈΡΠ΅ΠΌ ΡΡΠΎ Π²ΡΡΠ°ΠΆΠ΅Π½ΠΈΠ΅ ΠΏΠΎΠ΄ΡΠΎΠ±Π½ΠΎ:
Π‘ΠΈΠΌΠΌΠ΅ΡΡΠΈΡΠ½Π°Ρ ΠΌΠ°ΡΡΠΈΡΠ° A β πβΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ, Π΅ΡΠ»ΠΈ Π΄Π»Ρ Π²ΡΠ΅Ρ Π½Π΅Π½ΡΠ»Π΅Π²ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x β ββΏ ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ Π½Π΅ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ xα΅Ax > 0. ΠΠ±ΡΡΠ½ΠΎ ΡΡΠΎ ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°Π΅ΡΡΡ ΠΊΠ°ΠΊ
(ΠΈΠ»ΠΈ ΠΏΡΠΎΡΡΠΎ A > 0), Π° ΠΌΠ½ΠΎΠΆΠ΅ΡΡΠ²ΠΎ Π²ΡΠ΅Ρ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΡΡ ΠΌΠ°ΡΡΠΈΡ ΡΠ°ΡΡΠΎ ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°ΡΡ
Π‘ΠΈΠΌΠΌΠ΅ΡΡΠΈΡΠ½Π°Ρ ΠΌΠ°ΡΡΠΈΡΠ° A β πβΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΏΠΎΠ»ΡΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ, Π΅ΡΠ»ΠΈ Π΄Π»Ρ Π²ΡΠ΅Ρ Π²Π΅ΠΊΡΠΎΡΠΎΠ² ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ Π½Π΅ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ xα΅ Ax β₯ 0. ΠΡΠΎ Π·Π°ΠΏΠΈΡΡΠ²Π°Π΅ΡΡΡ ΠΊΠ°ΠΊ
(ΠΈΠ»ΠΈ ΠΏΡΠΎΡΡΠΎ A β₯ 0), Π° ΠΌΠ½ΠΎΠΆΠ΅ΡΡΠ²ΠΎ Π²ΡΠ΅Ρ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΏΠΎΠ»ΡΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΡΡ ΠΌΠ°ΡΡΠΈΡ ΡΠ°ΡΡΠΎ ΠΎΠ±ΠΎΠ·Π½Π°ΡΠ°ΡΡ
ΠΠ½Π°Π»ΠΎΠ³ΠΈΡΠ½ΠΎ ΡΠΈΠΌΠΌΠ΅ΡΡΠΈΡΠ½Π°Ρ ΠΌΠ°ΡΡΠΈΡΠ° A β πβΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ ΠΎΡΡΠΈΡΠ°ΡΠ΅Π»ΡΠ½ΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ
, Π΅ΡΠ»ΠΈ Π΄Π»Ρ Π²ΡΠ΅Ρ Π½Π΅Π½ΡΠ»Π΅Π²ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x β ββΏ ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ Π½Π΅ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ xα΅Ax
), Π΅ΡΠ»ΠΈ Π΄Π»Ρ Π²ΡΠ΅Ρ Π½Π΅Π½ΡΠ»Π΅Π²ΡΡ Π²Π΅ΠΊΡΠΎΡΠΎΠ² x β ββΏ ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ Π½Π΅ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ xα΅Ax β€ 0.
ΠΠ°ΠΊΠΎΠ½Π΅Ρ, ΡΠΈΠΌΠΌΠ΅ΡΡΠΈΡΠ½Π°Ρ ΠΌΠ°ΡΡΠΈΡΠ° A β πβΏ Π½Π°Π·ΡΠ²Π°Π΅ΡΡΡ Π½Π΅ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ, Π΅ΡΠ»ΠΈ ΠΎΠ½Π° Π½Π΅ ΡΠ²Π»ΡΠ΅ΡΡΡ Π½ΠΈ ΠΏΠΎΠ»ΠΎΠΆΠΈΡΠ΅Π»ΡΠ½ΠΎ ΠΏΠΎΠ»ΡΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ, Π½ΠΈ ΠΎΡΡΠΈΡΠ°ΡΠ΅Π»ΡΠ½ΠΎ ΠΏΠΎΠ»ΡΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½Π½ΠΎΠΉ, ΡΠΎ Π΅ΡΡΡ Π΅ΡΠ»ΠΈ ΡΡΡΠ΅ΡΡΠ²ΡΡΡ Π²Π΅ΠΊΡΠΎΡΡ xβ, xβ β ββΏ ΡΠ°ΠΊΠΈΠ΅, ΡΡΠΎ
Π‘ΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠ΅ Π·Π½Π°ΡΠ΅Π½ΠΈΡ ΠΈ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠ΅ Π²Π΅ΠΊΡΠΎΡΡ
ΠΠ»Ρ ΠΊΠ²Π°Π΄ΡΠ°ΡΠ½ΠΎΠΉ ΠΌΠ°ΡΡΠΈΡΡ A β ββΏΛ£βΏ ΠΊΠΎΠΌΠΏΠ»Π΅ΠΊΡΠ½ΠΎΠ΅ Π·Π½Π°ΡΠ΅Π½ΠΈΠ΅ Ξ» β β ΠΈ Π²Π΅ΠΊΡΠΎΡ x β ββΏ Π±ΡΠ΄ΡΡ ΡΠΎΠΎΡΠ²Π΅ΡΡΡΠ²Π΅Π½Π½ΠΎ ΡΠ²Π»ΡΡΡΡΡ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠΌ Π·Π½Π°ΡΠ΅Π½ΠΈΠ΅ΠΌ ΠΈ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠΌ Π²Π΅ΠΊΡΠΎΡΠΎΠΌ, Π΅ΡΠ»ΠΈ Π²ΡΠΏΠΎΠ»Π½ΡΠ΅ΡΡΡ ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ
ΠΠ° ΠΈΠ½ΡΡΠΈΡΠΈΠ²Π½ΠΎΠΌ ΡΡΠΎΠ²Π½Π΅ ΡΡΠΎ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½ΠΈΠ΅ ΠΎΠ·Π½Π°ΡΠ°Π΅Ρ, ΡΡΠΎ ΠΏΡΠΈ ΡΠΌΠ½ΠΎΠΆΠ΅Π½ΠΈΠΈ Π½Π° ΠΌΠ°ΡΡΠΈΡΡ A Π²Π΅ΠΊΡΠΎΡ x ΡΠΎΡ ΡΠ°Π½ΡΠ΅Ρ Π½Π°ΠΏΡΠ°Π²Π»Π΅Π½ΠΈΠ΅, Π½ΠΎ ΠΌΠ°ΡΡΡΠ°Π±ΠΈΡΡΠ΅ΡΡΡ Ρ ΠΊΠΎΡΡΡΠΈΡΠΈΠ΅Π½ΡΠΎΠΌ Ξ». ΠΠ°ΠΌΠ΅ΡΠΈΠΌ, ΡΡΠΎ Π΄Π»Ρ Π»ΡΠ±ΠΎΠ³ΠΎ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΠΎΠ³ΠΎ Π²Π΅ΠΊΡΠΎΡΠ° x β ββΏ ΠΈ ΡΠΊΠ°Π»ΡΡΠ½ΠΎΠ³ΠΎ Π·Π½Π°ΡΠ΅Π½ΠΈΡ Ρ β β ΡΠΏΡΠ°Π²Π΅Π΄Π»ΠΈΠ²ΠΎ ΡΠ°Π²Π΅Π½ΡΡΠ²ΠΎ A(cx) = cAx = cΞ»x = Ξ»(cx). Π’Π°ΠΊΠΈΠΌ ΠΎΠ±ΡΠ°Π·ΠΎΠΌ, cx ΡΠΎΠΆΠ΅ ΡΠ²Π»ΡΠ΅ΡΡΡ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠΌ Π²Π΅ΠΊΡΠΎΡΠΎΠΌ. ΠΠΎΡΡΠΎΠΌΡ, Π³ΠΎΠ²ΠΎΡΡ ΠΎ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΠΎΠΌ Π²Π΅ΠΊΡΠΎΡΠ΅, ΡΠΎΠΎΡΠ²Π΅ΡΡΡΠ²ΡΡΡΠ΅ΠΌ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΠΎΠΌΡ Π·Π½Π°ΡΠ΅Π½ΠΈΡ Ξ», ΠΌΡ ΠΎΠ±ΡΡΠ½ΠΎ ΠΈΠΌΠ΅Π΅ΠΌ Π² Π²ΠΈΠ΄Ρ Π½ΠΎΡΠΌΠ°Π»ΠΈΠ·ΠΎΠ²Π°Π½Π½ΡΠΉ Π²Π΅ΠΊΡΠΎΡ Ρ Π΄Π»ΠΈΠ½ΠΎΠΉ 1 (ΠΏΡΠΈ ΡΠ°ΠΊΠΎΠΌ ΠΎΠΏΡΠ΅Π΄Π΅Π»Π΅Π½ΠΈΠΈ Π²ΡΠ΅ ΡΠ°Π²Π½ΠΎ ΡΠΎΡ ΡΠ°Π½ΡΠ΅ΡΡΡ Π½Π΅ΠΊΠΎΡΠΎΡΠ°Ρ Π½Π΅ΠΎΠ΄Π½ΠΎΠ·Π½Π°ΡΠ½ΠΎΡΡΡ, ΡΠ°ΠΊ ΠΊΠ°ΠΊ ΡΠΎΠ±ΡΡΠ²Π΅Π½Π½ΡΠΌΠΈ Π²Π΅ΠΊΡΠΎΡΠ°ΠΌΠΈ Π±ΡΠ΄ΡΡ ΠΊΠ°ΠΊ x, ΡΠ°ΠΊ ΠΈ βx, Π½ΠΎ ΡΡΡ ΡΠΆ Π½ΠΈΡΠ΅Π³ΠΎ Π½Π΅ ΠΏΠΎΠ΄Π΅Π»Π°Π΅ΡΡ).
ΠΠ΅ΡΠ΅Π²ΠΎΠ΄ ΡΡΠ°ΡΡΠΈ Π±ΡΠ» ΠΏΠΎΠ΄Π³ΠΎΡΠΎΠ²Π»Π΅Π½ Π² ΠΏΡΠ΅Π΄Π΄Π²Π΅ΡΠΈΠΈ ΡΡΠ°ΡΡΠ° ΠΊΡΡΡΠ° «ΠΠ°ΡΠ΅ΠΌΠ°ΡΠΈΠΊΠ° Π΄Π»Ρ Data Science». Π’Π°ΠΊΠΆΠ΅ ΠΏΡΠΈΠ³Π»Π°ΡΠ°Π΅ΠΌ Π²ΡΠ΅Ρ ΠΆΠ΅Π»Π°ΡΡΠΈΡ ΠΏΠΎΡΠ΅ΡΠΈΡΡ Π±Π΅ΡΠΏΠ»Π°ΡΠ½ΡΠΉ Π΄Π΅ΠΌΠΎΡΡΠΎΠΊ, Π² ΡΠ°ΠΌΠΊΠ°Ρ ΠΊΠΎΡΠΎΡΠΎΠ³ΠΎ ΡΠ°ΡΡΠΌΠΎΡΡΠΈΠΌ ΠΏΠΎΠ½ΡΡΠΈΠ΅ Π»ΠΈΠ½Π΅ΠΉΠ½ΠΎΠ³ΠΎ ΠΏΡΠΎΡΡΡΠ°Π½ΡΡΠ²Π° Π½Π° ΠΏΡΠΈΠΌΠ΅ΡΠ°Ρ , ΠΏΠΎΠ³ΠΎΠ²ΠΎΡΠΈΠΌ ΠΎ Π»ΠΈΠ½Π΅ΠΉΠ½ΡΡ ΠΎΡΠΎΠ±ΡΠ°ΠΆΠ΅Π½ΠΈΡΡ , ΠΈΡ ΡΠΎΠ»ΠΈ Π² Π°Π½Π°Π»ΠΈΠ·Π΅ Π΄Π°Π½Π½ΡΡ ΠΈ ΠΏΠΎΡΠ΅ΡΠ°Π΅ΠΌ Π·Π°Π΄Π°ΡΠΈ.
Symmetric Matrix
Then Aβ²A is a p Γ p symmetric matrix and is decomposed as Aβ²A = V Ξ Vβ², where V is an orthogonal matrix whose columns are eigenvectors of Aβ²A and Ξ is a diagonal matrix whose diagonal elements are eigenvalues of Aβ²A.
Related terms:
Vector and Matrix Operations for Multivariate Analysis
2.6.1 Symmetric Matrices
Figure 2.1 shows, in schematic form, various special matrices of interest to multivariate analysis. The first property for categorizing types of matrices concerns whether they are square (m = n) or rectangular. In turn, rectangular matrices can be either vertical (m > n) or horizontal (m
As we shall show in later chapters, square matrices play an important role in multivariate analysis. In particular, the notion of matrix symmetry is important. Earlier, a symmetric matrix was defined as a square matrix that satisfies the relation
That is, a symmetric matrix is a square matrix that is equal to its transpose. For example,
Symmetric matrices, such as correlation matrices and covariance matrices, are quite common in multivariate analysis, and we shall come across them repeatedly in later chapters. 7
A few properties related to symmetry in matrices are of interest to point out:
The product of any (not necessarily symmetric) matrix and its transpose is symmetric; that is, both AAβ² and Aβ²A are symmetric matrices.
If A is any square (not necessarily symmetric) matrix, then A + Aβ² is symmetric.
If A is symmetric and k is a scalar, then kA is a symmetric matrix.
The sum of any number of symmetric matrices is also symmetric.
The product of two symmetric matrices is not necessarily symmetric.
Later chapters will discuss still other characteristics of symmetric matrices and the special role that they play in such topics as matrix eigenstructures and quadratic forms.
Matrix Algebra
Spectral Decomposition
Any symmetric matrix A can be written as
where Ξ is a diagonal matrix of eigenvalues of A and V is an orthogonal matrix whose column vectors are normalized eigenvectors. This decomposition is called as spectral decomposition.
For example, eigenvalues of a symmetric matrix
are 50 and 25. The corresponding eigenvectors are (4/5, 3/5)β² and (β3/5, 4/5)β². Then, A can be written as
If A is a nonsingular symmetric matrix, A r = VΞ r Vβ². If A is a nonsingular symmetric idempotent matrix, eigenvalues of A should be 0 or 1 since A 2 = A leads to Ξ 2 = Ξ.
Symmetric Matrix
A matrix S is symmetric if it equals its transpose. Real matrices of the form AA T and A T A are symmetric and have nonnegative eigenvalues.
Illustration
A symmetric matrix
A symmetric matrix of the form S.Transpose[S]
MatrixForm [R1 = S.Transpose[S]]
β A symmetric matrix of the form Transpose[S].S
A symmetric matrix of the form (A + A T )
For every square matrix A, the matrix (A + A T ) is symmetric
SymmetricMatrixQ [A + Transpose [A]]
Elements of linear algebra
Theorem 2
A symmetric matrix A of order n is:
positive definite iff all the leading principal (upper-leftmost) minors are positive:
negative definite iff all the leading principal (upper-leftmost) minors alternate in sign, beginning with a negative minor;
positive semidefinite iff all the principal (not necessarily upper-leftmost) minors are β₯0, with the minor of order n equal to zero, i.e., |A| = 0;
negative semidefinite iff each one of the principal (not necessarily upper-leftmost) minors of even order is β₯ 0, each one of the principal (not necessarily upper-leftmost) minors of odd order is β€ 0, with the minor of order n equal to zero, i.e., |A| = 0;
indefinite in all the other cases.
Let us see now how we can implement a quadratic form in Excel through the following example.
Krylov Subspace Methods
21.9 The Minres Method
If a symmetric matrix is indefinite, the CG method does not apply. The minimum residual method (MINRES) is designed to apply in this case. In the same fashion as we developed the GMRES algorithm using the Arnoldi iteration, Algorithm 21.8 implements the MINRES method using the Lanczos iteration. In the resulting least-squares problem, the coefficient matrix is tridiagonal, and we compute the QR decomposition using Givens rotations.
MINRES
function minresb(A,b,x0,m,tol,maxiter)
% Solve Ax = b using the MINRES method
% Input: n Γ n matrix A, n Γ 1 vector b,
% initial approximation x0, integer m β€ n,
% error tolerance tol, and the maximum number of iterations, maxiter.
% Output: Approximate solution xm, associated residual r,
% and iter, the number of iterations required.
% iter = β1 if the tolerance was not satisfied.
while iter β€ maxiter do
Solve the (m + 1) Γ m least β squares problem T m Β― y m = Ξ² e 1
using Givens rotations that take advantage of the tridiagonal
if r
end if
end while
end function
MINRES does well when a symmetric matrix is well conditioned. The tridiagonal structure of Tkmakes MINRES vulnerable to rounding errors [69, pp. 84-86], [72]. It has been shown that the rounding errors propagate to the approximate solution as the square of ΞΊ (A). For GMRES, the errors propagate as a function of the ΞΊ (A). Thus, if A is badly conditioned, try mpregmres.
Example 21.11
Convergence
Like GMRES, there is no simple set of properties that guarantee convergence. A theorem that specifies some conditions under which convergence will occur can be found in Ref. [73, pp. 50-51].
Remark 21.6
If A is positive definite, one normally uses CG or preconditioned CG. If A is symmetric indefinite and ill-conditioned, it is not safe to use a symmetric preconditioner K with MINRES if K β1 A is not symmetric. Finding a preconditioner for a symmetric indefinite matrix is difficult, and in this case the use of GMRES is recommended.
Congruent Symmetric Matrices
Two symmetric matrices A and B are said to be congruent if there exists an orthogonal matrix Q for which A = Q T BQ.
Illustration
Two congruent symmetric matrices
Q.Transpose [Q] == IdentityMatrix [2]
Congruence and quadratic forms
The previous calculations show that following two matrices are congruent:
Decomposition of Matrix Transformations: Eigenstructures and Quadratic Forms
5.6.4 Recapitulation
At this point we have covered quite a bit of ground regarding eigenstructures and matrix rank. In the case of nonsymmetric matrices in general, we noted that even if a (square) matrix A could be diagonalized via
the eigenvalues and eigenvectors need not be all real valued. (Fortunately, in the types of matrices encountered in multivariate analysis, we shall always be dealing with real-valued eigenvalues and eigenvectors.)
In the case of symmetric matrices, any such matrix A is diagonalizable, and orthogonally so, via
where Tβ² = T β1 since T is orthogonal. 13 Moreover, all eigenvalues and eigenvectors are necessarily real. If the eigenvalues are not all distinct, an orthogonal basisβalbeit not a unique oneβcan still be constructed. Furthermore, eigenvectors associated with distinct eigenvalues are already orthogonal to begin with.
The rank of any matrix A, square or rectangular, can be found from its eigenstructure or that of its product moment matrices. If AnΓn is symmetric, we merely count up the number of nonzero eigenvalues k and note that r(A) = k β€ n. If A is nonsymmetric or rectangular, we can find its minor (or major) product moment and then compute the eigenstructure. In this case, all eigenvalues are real and nonnegative. To find the rank of A, we simply count up the number of positive eigenvalues k and observe that r(A) = k β€ min(m, n) if A is rectangular or r(A) = k β€ n if A is square.
Finally, if A is of product-moment form to begin with, or if A is symmetric with nonnegative eigenvalues, then it can be writtenβalthough not uniquely soβas A = Bβ²B. The lack of uniqueness of B was illustrated in the context of axis rotation in factor analysis.
Algorithmic Graph Theory and Perfect Graphs
2 Symmetric Matrices
all pivots are chosen along the main diagonal whose entries are each nonzero
correspond precisely to the perfect vertex elimination schemes of G(M). By Theorem 4.1 we obtain the equivalence of statements (ii) and (iii), first obtained by Rose [1970 ], in the following theorem. We present a generalization due to Golumbic [1978 ].
Theorem 12.1
Let M be a symmetric matrix with nonzero diagonal entries. The following conditions are equivalent:
M has a perfect elimination scheme;
M has a perfect elimination scheme under restriction (R);
G(M) is a triangulated graph.
Before proving the theorem, we must introduce a bipartite graph model of the elimination process. This model will be used here and throughout the chapter.
An edge e = xy of a bipartite graph H = (U, E) is bisimplicial if Adj(x) + Adj(y) induces a complete bipartite subgraph of H. Take note that the bisimpliciality of an edge is retained as a hereditary property in an induced subgraph. Let Ο = [e1, e2, β¦, ek] be a sequence of pairwise nonadjacent edges of H. Denote by Si the set of endpoints of the edges e1, β¦, ei and let S0 = β . We say that Ο is a perfect edge elimination scheme for H if each edge ei is bisimplicial in the remaining induced subgraph H U β S i β 1 and H U β S n has no edge. Thus, we regard the elimination of an edge as the removal of all edges adjacent to e. For example, the graph in Figure 12.5 has the perfect edge elimination scheme [x1y1, x2y2, x3y3, x4y4]. Notice that initially x2y2 is not bisimplicial. A subsequence Οβ² = [e1, e2, β¦, ek] of Ο(k β€ n) is called a partial scheme. The notations H β Οβ² and H U β S k will be used to indicate the same subgraph.
Proof of Theorem 12.1
We have already remarked that (ii) and (iii) are equivalent, and since (ii) trivially implies (i), it suffices to prove that (i) implies (iii). Let us assume that M is symmetric with nonzero diagonal entries, and let Ο be a perfect edge elimination scheme for B(M).
Corollary 12.2
A symmetric matrix with nonzero diagonal entries can be tested for possession of a perfect elimination scheme in time proportional to the number of nonzero entries.
Proof
Theorem 12.1 characterized perfect elimination for symmetric matrices. Moreover, it says that it suffices to consider only the diagonal entries. Haskins and Rose [1973 ] treat the nonsymmetric case under (R), and Kleitman [1974 ] settles some questions left open by Haskins and Rose. The unrestricted case was finally solved by Golumbic and Goss [1978 ], who introduced perfect elimination bipartite graphs. These graphs will be discussed in the next section. Additional background on these and other matrix elimination problems can be found in the following survey articles and their references: Tarjan [1976 ], George [1977 ], and Reid [1977 ]. A discussion of the complexity of algorithms which calculate minimal and minimum fill-in under (R) can be found in Ohtsuki [1976 ], Ohtsuki, Cheung, and Fujisawa [1976 ], Rose, Tarjan, and Lueker [1976 ], and Rose and Tarjan [1978 ].
Large Sparse Eigenvalue Problems
22.4 Eigenvalue Computation Using the Lanczos Process
As expected, a sparse symmetric matrix A has properties that will enable us to compute eigenvalues and eigenvectors more efficiently than we are able to do with a nonsymmetric sparse matrix. Also, much more is known about convergence properties for the eigenvalue computations. We begin with the following lemma and then use it to investigate approximate eigenpairs of A.
Lemma 22.1
Proof
Taking the norm of the equality and noting that the <pi> are orthonormal, we obtain
Note that β i = 1 n p i p i T = I (Problem 22.4). Let Ξ»k be the eigenvalue closest to ΞΈ, i.e., |Ξ»k β ΞΈ| β€ |Ξ»i β ΞΈ| for all i, and we have
Recall that the Lanczos process for a symmetric matrix discussed in Section 21.8 is the Arnoldi process for a symmetric matrix and takes the form
It follows from Lemma 22.1 that there is an eigenvalue Ξ» such that
Example 22.5
The very ill-conditioned (approximate condition number 2.5538Γ10 17 ) symmetric 60000Γ60000 matrix Andrews obtained from the Florida sparse matrix collection was used in a computer graphics/vision problem. As we know, even though the matrix is ill-conditioned, its eigenvalues are well conditioned ( Theorem 19.2 ). The MATLAB statements time the approximation of the six largest eigenvalues and corresponding eigenvectors using eigsymb. A call to residchk outputs the residuals.
γγ tic;[V, D] = eigssymb(Andrews, 6, 50);toc;
Elapsed time is 5.494509 seconds.
Eigenpair 1 residual = 3.59008e-08
Eigenpair 2 residual = 1.86217e-08
Eigenpair 3 residual = 2.31836e-08
Eigenpair 4 residual = 7.68169e-08
Eigenpair 5 residual = 4.10453e-08
Eigenpair 6 residual = 6.04127e-07
22.4.1 Mathematically Provable Properties
This section presents some theoretical properties that shed light on the use and convergence properties of the Lanczos method.
In Ref. [76, p. 245], Scott proves the following theorem:
Theorem 22.1
This theorem says that if the spectrum of A is such that Ξ΄A/4 is larger than some given convergence tolerance, then there exist poor starting vectors which delay convergence until the nth step. Thus, the starting vector is a critical component in the performance of the algorithm. If a good starting vector is not known, then use a random vector. We have used this approach in our implementation of eigssymb.
The expository paper by Meurant and Stratkos [70, Theorem 4.3, p. 504] supports the conclusion that orthogonality can be lost only in the direction of converged Ritz vectors. This result allows the development of sophisticated methods for maintaining orthogonality such as selective reorthogonalization [2, pp. 565-566], [6, pp. 116-123].
There are significant results concerning the rate of convergence of the Lanczos algorithm. Kaniel [77] began investigating these problems. Subsequently, the finite precision behavior of the Lanczos algorithm was analyzed in great depth by Chris Paige in his Ph.D. thesis [78]; see also Paige [79β81]. He described the effects of rounding errors in the Lanczos algorithm using rigorous and elegant mathematical theory. The results are beyond the scope of this book, so we will assume exact arithmetic in the following result concerning convergence proved in Saad [3, pp. 147-150].
Theorem 22.2
Let A be an n-by-n symmetric matrix. The difference between the ith exact and approximate eigenvalues Ξ»i and Ξ» i ( m ) satisfies the double inequality
Remark 22.3
Error bounds indicate that for many matrices and for relatively small m, several of the largest or smallest of the eigenvalues of A are well approximated by eigenvalues of the corresponding Lanczos matrices. In practice, it is not always the case that both ends of the spectrum of a symmetric matrix are approximated accurately. However, it is generally true that at least one end of the spectrum is approximated well.
Least-Squares Problems
16.2.1 Using the Normal Equations
Solve the Normal Equations Using the Cholesky Decomposition a. |
Find the Cholesky decomposition A T A = R T R.
Solve the system R T y = A T b using forward substitution.
Solve the system Rx = y using back substitution.
Example 16.4.
There are three mountains m1, m2, m3 that from one site have been measured as 2474 ft., 3882 ft., and 4834 ft. But from m1, m2 looks 1422 ft. taller and m3 looks 2354 ft. taller, and from m2, m3 looks 950 ft. taller. This data gives rise to an overdetermined set of linear equations for the height of each mountain.
In matrix form, the least-squares problem is
A T A = [ 3 β 1 β 1 β 1 3 β 1 β 1 β 1 3 ] has the Cholesky decomposition
[ 1.7321 β 0.5774 β 0.5774 0 1.6330 β 0.8165 0 0 1.4142 ] x = y
Algorithm 16.1 uses the normal equations to solve the least-squares problem.
Least-Squares Solution Using the Normal Equations
% Solve the overdetermined least-squares problem using the normal equations.
% Input: m Γ n full-rank matrix A, m β₯ n and an m Γ 1 column vector b.
% Output: the unique least-squares solution to Ax = b and the residual
Use the Cholesky decomposition to obtain A T A = R T R
Solve the lower triangular system R T y = c
Solve the upper triangular system Rx = y
return [ x β b β A x β 2 ]
NLALIB: The function normalsolve implements Algorithm 16.1 by using the functions cholesky and cholsolve.
The efficiency analysis is straightforward.
Cholesky decomposition of A T A : n 3 3 + n 2 + 5 n 3 flops
Forward and back substitution: 2 ( n 2 + n β 1 ) flops
The total is n 2 ( 2 m + n 3 ) + 2 m n + 3 n 2 + 11 3 n β 2 β n 2 ( 2 m + n 3 ) flops
It must be noted that, although relatively easy to understand and implement, it has serious problems in some cases. β’ |
There may be some loss of significant digits when computing A T A. It may even be singular to working precision.
We will see in Section 16.3 that the accuracy of the solution using the normal equations depends on the square of condition number of the matrix. If ΞΊ (A) is large, the results can be seriously in error.