• Title: Hölder’s Inequality

  • Series: Functional Analysis

  • YouTube-Title: Functional Analysis 19 | Hölder’s Inequality

  • Bright video: https://youtu.be/yIXahhfRbTc

  • Dark video: https://youtu.be/pjKPOxmGSUo

  • Ad-free video: Watch Vimeo video

  • PDF: Download PDF version of the bright video

  • Dark-PDF: Download PDF version of the dark video

  • Print-PDF: Download printable PDF version

  • Thumbnail (bright): Download PNG

  • Thumbnail (dark): Download PNG

  • Subtitle on GitHub: fa19_sub_eng.srt

  • Timestamps

    00:00 Introduction

    01:42 Hölder’s Inequality

    02:19 Young’s Inequality

    05:11 Proof

  • Subtitle in English

    1 00:00:00,550 –> 00:00:02,289 Hello and welcome back to

    2 00:00:02,299 –> 00:00:03,589 functional analysis

    3 00:00:03,710 –> 00:00:04,929 and as always, I want to

    4 00:00:04,940 –> 00:00:06,320 thank all the nice people

    5 00:00:06,329 –> 00:00:07,590 that support this channel

    6 00:00:07,599 –> 00:00:09,039 on Steady or PayPal.

    7 00:00:09,729 –> 00:00:11,689 In today’s part 19, we

    8 00:00:11,699 –> 00:00:13,420 will go back and prove Hölder’s

    9 00:00:13,430 –> 00:00:14,369 inequality.

    10 00:00:15,250 –> 00:00:16,530 There are a lot of different

    11 00:00:16,540 –> 00:00:17,930 inequalities named after

    12 00:00:17,940 –> 00:00:19,350 Hölder and indeed, they are

    13 00:00:19,360 –> 00:00:20,309 all related,

    14 00:00:20,319 –> 00:00:21,850 but here we look at the simplest

    15 00:00:21,860 –> 00:00:23,639 one. It’s the one that

    16 00:00:23,649 –> 00:00:24,979 holds in F^n

    17 00:00:24,989 –> 00:00:26,739 and for p greater than one.

    18 00:00:27,450 –> 00:00:29,190 For each such number p, there

    19 00:00:29,200 –> 00:00:30,870 is a Hölder conjugate which

    20 00:00:30,879 –> 00:00:32,119 we call p prime.

    21 00:00:32,909 –> 00:00:34,080 The definition is what you

    22 00:00:34,090 –> 00:00:35,479 should really memorize.

    23 00:00:35,490 –> 00:00:37,020 The reciprocals should add

    24 00:00:37,029 –> 00:00:37,610 to one.

    25 00:00:38,180 –> 00:00:39,729 This is an important relation,

    26 00:00:39,740 –> 00:00:41,130 because every time we use

    27 00:00:41,139 –> 00:00:42,750 p prime, this is what we

    28 00:00:42,759 –> 00:00:44,529 mean. Instead of p

    29 00:00:44,540 –> 00:00:46,099 prime, a lot of people just

    30 00:00:46,110 –> 00:00:46,830 use q.

    31 00:00:47,630 –> 00:00:48,979 However, here I want to use

    32 00:00:48,990 –> 00:00:50,790 q as another variable to

    33 00:00:50,799 –> 00:00:52,590 tell you again what the p-

    34 00:00:52,599 –> 00:00:53,380 norm was.

    35 00:00:54,020 –> 00:00:55,240 Or today, I would better

    36 00:00:55,250 –> 00:00:57,049 say the q-norm of a vector

    37 00:00:57,060 –> 00:00:57,450 x.

    38 00:00:58,250 –> 00:01:00,080 It’s defined by the sum over

    39 00:01:00,090 –> 00:01:01,569 the entries in the absolute

    40 00:01:01,580 –> 00:01:03,049 value to the power q

    41 00:01:03,709 –> 00:01:05,370 and then we take the q-th root

    42 00:01:05,379 –> 00:01:06,290 of the whole thing

    43 00:01:07,069 –> 00:01:08,470 and this is what we do for

    44 00:01:08,480 –> 00:01:09,980 all q between one and

    45 00:01:09,989 –> 00:01:11,769 infinity including one.

    46 00:01:12,470 –> 00:01:14,269 Now Hölder inequality will

    47 00:01:14,279 –> 00:01:15,849 connect the p-norm the p

    48 00:01:15,860 –> 00:01:17,550 prime-norm and the 1-

    49 00:01:17,559 –> 00:01:19,489 norm. Indeed, it’s one

    50 00:01:19,500 –> 00:01:20,839 nice short formula,

    51 00:01:20,849 –> 00:01:22,160 you always should remember.

    52 00:01:22,989 –> 00:01:24,510 I will now use a possibly

    53 00:01:24,519 –> 00:01:26,110 strange notation to help

    54 00:01:26,120 –> 00:01:27,699 you to remember the important

    55 00:01:27,709 –> 00:01:29,610 inequality. Putting two

    56 00:01:29,620 –> 00:01:31,250 vectors together just

    57 00:01:31,260 –> 00:01:33,110 denotes a new vector which

    58 00:01:33,120 –> 00:01:34,949 has the product in the components.

    59 00:01:35,239 –> 00:01:36,930 So x_2, y_2 in the

    60 00:01:36,940 –> 00:01:38,559 second component and

    61 00:01:38,569 –> 00:01:39,849 x_n times y_n

    62 00:01:39,860 –> 00:01:41,580 in the last component.

    63 00:01:41,940 –> 00:01:43,760 Then for all vectors x and

    64 00:01:43,769 –> 00:01:45,650 y, we have that the 1-norm

    65 00:01:45,660 –> 00:01:47,440 of the vector xy

    66 00:01:47,940 –> 00:01:49,430 is less or equal than the

    67 00:01:49,440 –> 00:01:51,309 p-norm of x times the p

    68 00:01:51,319 –> 00:01:52,720 prime-norm of y

    69 00:01:53,379 –> 00:01:54,830 and that’s what we call

    70 00:01:54,839 –> 00:01:56,029 Hölder’s inequality.

    71 00:01:56,580 –> 00:01:57,720 Often you see it written

    72 00:01:57,730 –> 00:01:59,199 with the sums, but I think

    73 00:01:59,209 –> 00:02:00,910 it’s better to write it immediately

    74 00:02:00,919 –> 00:02:02,599 in this way. On the one

    75 00:02:02,610 –> 00:02:04,269 hand, it’s easier to remember

    76 00:02:04,360 –> 00:02:05,680 and on the other hand, we

    77 00:02:05,690 –> 00:02:06,919 will generalize the whole

    78 00:02:06,930 –> 00:02:07,860 thing later on.

    79 00:02:08,470 –> 00:02:09,899 Indeed, there we will have

    80 00:02:09,910 –> 00:02:11,199 the whole inequality for

    81 00:02:11,210 –> 00:02:13,160 functions defined on an abstract

    82 00:02:13,169 –> 00:02:13,929 measure space.

    83 00:02:14,429 –> 00:02:15,479 However, that’s not what

    84 00:02:15,490 –> 00:02:16,770 we do today. Today

    85 00:02:16,779 –> 00:02:18,149 we prove it for vectors

    86 00:02:18,160 –> 00:02:18,869 in F^n.

    87 00:02:19,649 –> 00:02:20,880 In order to prove Hölder’s

    88 00:02:20,889 –> 00:02:22,610 inequality, we first need

    89 00:02:22,619 –> 00:02:24,130 another inequality which

    90 00:02:24,139 –> 00:02:25,490 is known as Young’s

    91 00:02:25,500 –> 00:02:26,440 inequality.

    92 00:02:27,110 –> 00:02:28,649 In fact, this one is very

    93 00:02:28,660 –> 00:02:29,250 simple.

    94 00:02:29,259 –> 00:02:30,589 We just look at positive

    95 00:02:30,600 –> 00:02:32,550 numbers a, b and

    96 00:02:32,559 –> 00:02:33,850 then we conclude that the

    97 00:02:33,860 –> 00:02:35,529 product of ab is

    98 00:02:35,539 –> 00:02:37,210 always less or equal than

    99 00:02:37,220 –> 00:02:38,309 the following sum.

    100 00:02:38,470 –> 00:02:39,630 So what you should see here

    101 00:02:39,639 –> 00:02:41,160 is that p greater than one

    102 00:02:41,169 –> 00:02:42,850 goes in and the Hölder

    103 00:02:42,860 –> 00:02:43,610 conjugate.

    104 00:02:44,259 –> 00:02:45,940 So as before you can choose

    105 00:02:45,949 –> 00:02:47,669 p greater one as you want,

    106 00:02:47,690 –> 00:02:49,470 but then p prime is fixed.

    107 00:02:50,000 –> 00:02:51,610 For example, for p equals

    108 00:02:51,619 –> 00:02:53,570 to two, we have p prime equals

    109 00:02:53,580 –> 00:02:54,110 to two.

    110 00:02:54,600 –> 00:02:55,910 However, in this case, you

    111 00:02:55,919 –> 00:02:57,619 already know that the inequality

    112 00:02:57,630 –> 00:02:58,509 here is correct.

    113 00:02:59,300 –> 00:03:00,479 So please check that,

    114 00:03:00,490 –> 00:03:01,759 but for all other cases,

    115 00:03:01,770 –> 00:03:03,380 we have to write down a proof.

    116 00:03:04,190 –> 00:03:05,570 What we can use here is that

    117 00:03:05,580 –> 00:03:07,210 the common exponential function

    118 00:03:07,220 –> 00:03:08,929 is a so called convex

    119 00:03:08,940 –> 00:03:09,369 function.

    120 00:03:10,059 –> 00:03:11,190 This means that when you

    121 00:03:11,199 –> 00:03:12,279 look at the graph,

    122 00:03:12,970 –> 00:03:14,389 then you can choose any two

    123 00:03:14,399 –> 00:03:14,779 points

    124 00:03:14,789 –> 00:03:15,630 as you want.

    125 00:03:15,710 –> 00:03:16,929 The direct connection would

    126 00:03:16,940 –> 00:03:18,630 be always above the graph.

    127 00:03:19,110 –> 00:03:20,940 This property is called convex,

    128 00:03:20,949 –> 00:03:22,509 because the red line is the

    129 00:03:22,520 –> 00:03:24,100 convex combination of the

    130 00:03:24,110 –> 00:03:24,839 two points.

    131 00:03:25,440 –> 00:03:26,479 Therefore, we can easily

    132 00:03:26,490 –> 00:03:28,029 put that into a formula when

    133 00:03:28,039 –> 00:03:29,449 we call the function just

    134 00:03:29,460 –> 00:03:29,910 f.

    135 00:03:30,360 –> 00:03:31,990 Now, by denoting the two points

    136 00:03:32,000 –> 00:03:33,669 by x and y, we can form a

    137 00:03:33,679 –> 00:03:35,410 convex combination of x and

    138 00:03:35,419 –> 00:03:37,369 y and put that into the function

    139 00:03:37,380 –> 00:03:39,300 f. Which means we get out

    140 00:03:39,309 –> 00:03:40,220 the blue line here.

    141 00:03:40,589 –> 00:03:42,339 Please recall a convex

    142 00:03:42,350 –> 00:03:43,940 combination is just a special

    143 00:03:43,949 –> 00:03:45,179 linear combination, where

    144 00:03:45,190 –> 00:03:46,940 we only have one lambda which

    145 00:03:46,949 –> 00:03:48,029 comes from the interval

    146 00:03:48,039 –> 00:03:49,199 0 to 1.

    147 00:03:49,350 –> 00:03:50,660 Now with the red line, we

    148 00:03:50,669 –> 00:03:52,289 also get a convex combination

    149 00:03:52,300 –> 00:03:53,789 with the images. Which means

    150 00:03:53,800 –> 00:03:55,029 we have here instead of x

    151 00:03:55,039 –> 00:03:56,899 and y, just f(x)

    152 00:03:56,940 –> 00:03:57,899 and f(y).

    153 00:03:58,309 –> 00:03:59,770 Therefore being always above

    154 00:03:59,779 –> 00:04:01,199 the blue graph means we have

    155 00:04:01,210 –> 00:04:02,669 here our inequality.

    156 00:04:03,410 –> 00:04:04,949 Which we now want to use

    157 00:04:04,960 –> 00:04:06,229 for some special numbers.

    158 00:04:06,960 –> 00:04:08,339 Lambda should be one over

    159 00:04:08,350 –> 00:04:10,110 p and then one minus

    160 00:04:10,119 –> 00:04:11,509 lambda should be one over

    161 00:04:11,520 –> 00:04:12,240 p prime.

    162 00:04:12,729 –> 00:04:13,789 Maybe that’s not so clear

    163 00:04:13,800 –> 00:04:15,360 but for x I want to put in

    164 00:04:15,369 –> 00:04:16,720 the natural logarithm of

    165 00:04:16,730 –> 00:04:18,108 a to the power p

    166 00:04:18,459 –> 00:04:19,640 and the similar thing for

    167 00:04:19,649 –> 00:04:21,559 y but now with b and p

    168 00:04:21,570 –> 00:04:23,540 prime. Of course, this

    169 00:04:23,549 –> 00:04:25,059 all fits together, because

    170 00:04:25,070 –> 00:04:26,329 by applying the logarithm

    171 00:04:26,339 –> 00:04:27,779 rules, we can bring this

    172 00:04:27,790 –> 00:04:29,559 power in front, where it cancels

    173 00:04:29,570 –> 00:04:30,779 out. Hence,

    174 00:04:30,790 –> 00:04:31,779 on the left-hand side, we

    175 00:04:31,790 –> 00:04:33,029 have the two logarithms in

    176 00:04:33,040 –> 00:04:33,940 the function f.

    177 00:04:34,320 –> 00:04:36,089 Now using the next logarithm

    178 00:04:36,100 –> 00:04:36,619 rule

    179 00:04:36,630 –> 00:04:38,239 and the fact that the exponential

    180 00:04:38,250 –> 00:04:39,820 function is the inverse function

    181 00:04:39,829 –> 00:04:41,179 of the logarithm, we have

    182 00:04:41,190 –> 00:04:42,859 just ab on the left-hand

    183 00:04:42,869 –> 00:04:43,290 side

    184 00:04:43,750 –> 00:04:44,980 and now you see the idea

    185 00:04:44,989 –> 00:04:46,390 of the whole proof. We want

    186 00:04:46,399 –> 00:04:48,140 to show this one and we’re

    187 00:04:48,149 –> 00:04:49,380 already finished with the

    188 00:04:49,390 –> 00:04:50,179 left part.

    189 00:04:50,920 –> 00:04:51,940 Regarding the right-hand

    190 00:04:51,950 –> 00:04:53,220 side, we just put in all

    191 00:04:53,230 –> 00:04:54,609 the numbers we already know.

    192 00:04:55,290 –> 00:04:56,869 Now this looks more complicated

    193 00:04:56,880 –> 00:04:58,230 than it really is, because

    194 00:04:58,239 –> 00:04:59,500 we have again the inverse

    195 00:04:59,510 –> 00:05:00,670 function of the logarithm

    196 00:05:00,679 –> 00:05:01,029 here.

    197 00:05:01,579 –> 00:05:03,170 Hence this all vanishes

    198 00:05:03,179 –> 00:05:04,309 and what remains is what

    199 00:05:04,320 –> 00:05:05,309 we wanted to show.

    200 00:05:05,799 –> 00:05:07,320 So Young’s inequality is

    201 00:05:07,329 –> 00:05:08,829 correct and we can use it

    202 00:05:08,839 –> 00:05:10,750 to prove Hölder’s inequality.

    203 00:05:11,260 –> 00:05:12,410 Starting with the proof,

    204 00:05:12,420 –> 00:05:13,480 let’s first consider the

    205 00:05:13,489 –> 00:05:14,420 simplest case.

    206 00:05:15,079 –> 00:05:16,320 What I mean by that is that

    207 00:05:16,329 –> 00:05:17,549 we look what happens when

    208 00:05:17,559 –> 00:05:19,239 x is the zero vector or

    209 00:05:19,250 –> 00:05:20,579 y is the zero vector,

    210 00:05:21,179 –> 00:05:22,459 but of course, this is not

    211 00:05:22,470 –> 00:05:23,899 a problem for us, because

    212 00:05:23,910 –> 00:05:25,269 by the definition of all

    213 00:05:25,279 –> 00:05:26,799 the norms here, we have that

    214 00:05:26,809 –> 00:05:28,250 the left-hand side is zero

    215 00:05:28,260 –> 00:05:29,459 and the right-hand side is

    216 00:05:29,470 –> 00:05:29,850 zero.

    217 00:05:29,859 –> 00:05:31,820 So the inequality is fulfilled.

    218 00:05:32,480 –> 00:05:33,739 Therefore, for the second

    219 00:05:33,750 –> 00:05:35,529 case, we can divide and bring

    220 00:05:35,540 –> 00:05:36,660 everything to the left-hand

    221 00:05:36,670 –> 00:05:37,070 side.

    222 00:05:37,700 –> 00:05:38,880 Of course, now we want to

    223 00:05:38,890 –> 00:05:40,609 bring this inside the norm

    224 00:05:40,619 –> 00:05:42,290 and divide x by its p-

    225 00:05:42,299 –> 00:05:44,250 norm and y by its p prime-

    226 00:05:44,260 –> 00:05:44,640 norm.

    227 00:05:45,019 –> 00:05:46,380 However, since I use this

    228 00:05:46,390 –> 00:05:47,760 strange notation here, I

    229 00:05:47,769 –> 00:05:49,339 bring in now the whole sum

    230 00:05:49,350 –> 00:05:51,299 of the 1-norm. Here we have

    231 00:05:51,309 –> 00:05:52,649 it and we can pull in the

    232 00:05:52,660 –> 00:05:54,250 norm into the sum and then

    233 00:05:54,260 –> 00:05:55,299 into the absolute value

    234 00:05:55,989 –> 00:05:56,929 and there you should see

    235 00:05:56,940 –> 00:05:58,679 now, that we have two positive

    236 00:05:58,690 –> 00:06:00,149 numbers, which we could call

    237 00:06:00,160 –> 00:06:01,989 a and b and then apply

    238 00:06:02,000 –> 00:06:03,899 Young’s inequality inside

    239 00:06:03,910 –> 00:06:04,459 the sum.

    240 00:06:04,980 –> 00:06:06,380 So at this point, we get

    241 00:06:06,390 –> 00:06:07,690 the inequality sign in.

    242 00:06:08,399 –> 00:06:08,700 OK.

    243 00:06:08,709 –> 00:06:10,010 The first term is one over

    244 00:06:10,019 –> 00:06:11,940 p times x_j

    245 00:06:11,950 –> 00:06:13,739 to the power p divided by

    246 00:06:13,750 –> 00:06:15,450 the p-norm to the power p.

    247 00:06:15,950 –> 00:06:17,410 The second term looks similar

    248 00:06:17,420 –> 00:06:18,570 and I already distributed

    249 00:06:18,579 –> 00:06:20,089 the sum over both parts.

    250 00:06:20,739 –> 00:06:22,329 So here we just have p prime

    251 00:06:22,339 –> 00:06:23,799 and j instead of x.

    252 00:06:24,220 –> 00:06:25,769 However, that’s not so important.

    253 00:06:25,779 –> 00:06:26,869 The important thing you should

    254 00:06:26,880 –> 00:06:28,279 see is what we have in

    255 00:06:28,290 –> 00:06:29,600 the numerator together with

    256 00:06:29,609 –> 00:06:31,109 the sum, is the same as the

    257 00:06:31,119 –> 00:06:32,809 denominator and the

    258 00:06:32,820 –> 00:06:34,119 same in the second part.

    259 00:06:34,549 –> 00:06:35,760 Hence, the only things that

    260 00:06:35,769 –> 00:06:37,640 remain is one over p here

    261 00:06:37,649 –> 00:06:39,160 and one over p prime here

    262 00:06:39,630 –> 00:06:40,880 and by the definition of

    263 00:06:40,890 –> 00:06:42,470 the Hölder conjugate this

    264 00:06:42,480 –> 00:06:43,480 is simply one

    265 00:06:44,000 –> 00:06:45,540 and with this, we have proven

    266 00:06:45,549 –> 00:06:47,200 Hölder’s inequality, because

    267 00:06:47,209 –> 00:06:48,630 you can bring this one on

    268 00:06:48,640 –> 00:06:49,760 the right-hand side again,

    269 00:06:49,769 –> 00:06:50,440 if you want.

    270 00:06:51,130 –> 00:06:52,459 So you see the proof was

    271 00:06:52,470 –> 00:06:53,869 not so hard, but we will

    272 00:06:53,880 –> 00:06:55,630 need Hölder’s inequality to

    273 00:06:55,640 –> 00:06:57,309 prove another inequality

    274 00:06:57,739 –> 00:06:58,839 and this one will be the

    275 00:06:58,850 –> 00:07:00,149 so called Minkowski

    276 00:07:00,160 –> 00:07:01,010 inequality.

    277 00:07:01,320 –> 00:07:02,790 Essentially, it’s just the

    278 00:07:02,940 –> 00:07:04,869 triangle inequality for our l^p

    279 00:07:04,880 –> 00:07:05,470 space.

    280 00:07:06,010 –> 00:07:07,220 Therefore, I hope I see you

    281 00:07:07,230 –> 00:07:08,070 in the next video.

    282 00:07:08,079 –> 00:07:09,579 Thanks for listening and

    283 00:07:09,589 –> 00:07:10,410 see you then.

    284 00:07:10,420 –> 00:07:11,109 Bye.

  • Quiz Content (n/a)
  • Last update: 2024-10

  • Back to overview page


Do you search for another mathematical topic?